Datasets:

emozilla

/

dolma-v1_7-arxiv

like 1

\section{Introduction} \IEEEPARstart{T}{he} goal of the field of reinforcement learning (RL) is to develop learning algorithms that can effectively deal with the complexities of the real world. Games are a structured form of interactions between one or more players in an environment, making them ideal for the study of reinforcement learning. Much of research in artificial intelligence has focused on games which emulate different challenges of the real world. In Go \cite{silver2017mastering}, the agent has to discover complex strategies in a large search space. In card games like Poker \cite{moravvcik2017deepstack, brown2018superhuman, brown2019superhuman}, the agent has to deal with the imperfect-information, such as the unknown cards of the opponent. In StarCraft~II~\cite{vinyals2019grandmaster} and Dota~2~\cite{berner2019dota}, the agent has to compete with other agents who take simultaneous actions from a large action space. \color{diffcolor} In this work, we consider the problem of learning to play games with a novel set of challenges: imperfect-information multi-agent games with simultaneous moves and large state-action spaces. We consider two such games as learning environments: Clash Royale (a popular multiplayer real-time strategy game) and Pommerman \cite{DBLP:journals/corr/abs-1809-07124}. Clash Royale is a unique game combining elements of different genres such as MOBA (multiplayer online battle arena), collective-card games, and tower defense games. The complexity in learning to play Clash Royale comes from the presence of cyclic strategies, partial observability, and exploration in large dynamic action spaces (more details in Section~\ref{sec:cr}). Pommerman is a popular multi-agent RL benchmark which is difficult due to the need for opponent modelling and therefore a large branching factor as decisions are made in the combinatorial action space. In this paper, we introduce a new algorithm for efficient learning in large imperfect-information games\footnote{See \url{https://sites.google.com/view/l2p-clash-royale} for an explanation video of our method, in the context of Clash Royale.}, which does not require modifying the core game implementation. Our approach (illustrated in Fig.~\ref{fig:approach}) consists of two separate components: an oracle planner and a follower agent. The oracle planner has access to the full state of the environment and performs self-play tree search to compute effective (oracle) actions for each player. The oracle planner by itself can be used to implement a cheating AI for game implementations that do not support randomizing hidden information. A follower agent that can play the imperfect-information game is obtained by training a neural network to predict the oracle actions from partial observations using supervised learning. Planning is non-trivial in imperfect-information games \cite{brown2017safe}. The classical solution is to use Monte Carlo tree search (MCTS) with \emph{determinization} of the hidden information during search to account for the lack of the fully observed state of the environment \cite{frank1998search, ginsberg2001gib, bjarnason2009lower, cowling2012information}. However, this approach cannot be directly used in practice for many games as most existing simulators do not support the possibility of varying the hidden information. \color{black} Simultaneous moves with large action spaces makes model-based planning exceptionally challenging. Conventional MCTS can easily get stuck at creating new nodes corresponding to untried actions in a combinatorial action space. In this paper, we propose to build an oracle planner based on fixed-depth tree search (FDTS) with use of decoupled Thompson sampling for action selection. Our experiments show that FDTS can discover efficient strategies via self-play in the two challenging games that we consider in the paper. \emph{Contributions.} 1)~We introduce a new algorithm for efficient planning and learning in large imperfect-information games with implementations that do not support varying of hidden information. 2)~We demonstrate that naive Monte Carlo tree search can be problematic in large action spaces and introduce fixed-depth tree search to improve the quality of planning. 3)~We demonstrate the effectiveness of the algorithm in the novel setting of Clash Royale and the popular multi-agent RL benchmark of Pommerman. \begin{figure}[t] \centering \includegraphics[width=\linewidth, trim=0 0 70 0, clip]{high_level.PNG} \caption{\color{diffcolor} Proposed approach to solving imperfect-information games.} \label{fig:approach} \end{figure} \lstset{basicstyle=\scriptsize\ttfamily,breaklines=true} \section{Imperfect-Information Games} We formalize imperfect-information games as partially-observable stochastic games (POSG) \cite{hansen2004dynamic}. In POSG, a game is played by a set of $N$ players and each game begins in an initial state $s_0$ sampled from an initial state distribution. In any state $s$, observation functions $O_i(s)$ yield observations $o_i = O_i(s)$ for each player $i$. After receiving observation $o_i$, each player $i$ chooses an action $a_i \in A_i(s)$, where $A_i(s)$ is the set of actions available to player $i$ in state $s$. Once all players choose actions $a = (a_1, \ldots, a_N)$, the game transitions to a new state $s'$ as defined by a transition function $s' = f(s, a)$. Thus, the joint action space is $A(s) = A_1(s) \times \ldots \times A_N(s)$. The end of a game is defined by a set of terminal states $Z$. Once the game reaches a terminal state $z \in Z$, all players receive a ternary reward of 1 (win), 0 (draw) or -1 (loss) as defined by a reward function $R_i(z)$. A player does not have access to the true initial state distribution or the transition function but can sample from them by playing games. We now introduce the two games studied in this paper. \subsection{Clash Royale} \label{sec:cr} Clash Royale is a multiplayer real-time strategy game consisting of short battles lasting a few minutes. We focus on the two-player mode of Clash Royale. Before a battle, each player picks a \emph{deck} of eight different cards that is not revealed to the opponent. The game has nearly 100 \emph{cards} that represents playable troops, buildings or spells that will be used in battles. As the game begins, each player is dealt a random subset of four different cards (\emph{hand}) from their deck. Solving the whole game of Clash Royale involves solving the meta-game of choosing the right deck. In this paper, we focus on a fixed beginner deck consisting of \emph{Knight, Giant, Archer, Arrows, Minions, Fireball, Musketeer,} and \emph{Baby Dragon}. Battles in Clash Royale are played on a visually immersive $18~\times~32$ board initially consisting of a king tower and two princess towers for each player (see Fig.~\ref{fig:cr-screenshot}). The gameplay primarily consists of players deploying cards from their hand onto the battle arena to destroy the towers of the opponent. Each card has an \emph{Elixir} cost associated with it and a card can only be deployed if the player has enough Elixir. Once a card is deployed in a specific location, it creates a troop or building or spell in the battle arena that follows predefined behaviours, and the player is dealt a new card from the deck. A battle ends instantaneously if a king tower is destroyed. If not, the player with the highest number of towers after three minutes wins. Otherwise, the battle extends for an overtime of two minutes and the first player to destroy an enemy tower wins. Otherwise, the battle results in a draw. \begin{figure}[!t] \centering \includegraphics[height=50mm]{screenshot-min.png} \hfil \includegraphics[height=50mm, trim=5 0 95 0, clip]{pommerman_frame.jpg} \caption{\color{diffcolor} Screenshot of Clash Royale (left) and Pommerman (right).} \label{fig:cr-screenshot} \end{figure} \begin{table}[t] \caption{State space of a battle in Clash Royale} \label{t:state} \centering \begin{tabular}{lll} \toprule & Feature & Description \\ \midrule \multirow{5}{*}{Cards} & Decks & \textnormal{Sets of $8$ cards chosen by both players for} \\ & & \textnormal{this battle.} \\ & Level & \textnormal{Levels of all cards in the chosen decks.} \\ & Cost & \textnormal{Elixir cost of all cards in the chosen decks.} \\ & Hand & \textnormal{Sets of $4$ cards currently available to both} \\ & & \textnormal{players.} \\ & Next card & \textnormal{Next card to be dealt to each player.} \\ \midrule \multirow{5}{*}{Progress} & Elixir & \textnormal{Elixir currently available to each player.} \\ & Time & \textnormal{Time elapsed since beginning of battle.} \\ & 2x Elixir & \textnormal{Is battle in double elixir mode?} \\ & Overtime & \textnormal{Is battle in sudden death mode?} \\ & Past actions & \textnormal{List of actions by the player} \\ & & \textnormal{in the past 10 steps.} \\ \midrule \multirow{5}{*}{Objects} & Type & \textnormal{Types of all objects in the battle arena.} \\ & Position & \textnormal{$x$, $y$ coordinates of all objects in the arena.} \\ & Level & \textnormal{Levels of all objects in the arena.} \\ & Health & \textnormal{Health of all objects in the arena.} \\ & Color & \textnormal{Object belongs to blue or red player?} \\ \bottomrule \end{tabular} \end{table} \color{diffcolor} The state $s$ of Clash Royale is comprehensively defined in Table~\ref{t:state}. \color{black} Each player observes the state of the battle arena, battle progress, the player's own hand and the next card. Information about the cards of the other player is not visible. At any game state, player $i$ can choose either to deploy a legal card (a card that costs less than or equal to available elixir) or to wait for one time step. \color{diffcolor} In this paper, an agent interacts with the Clash Royale game engine such that one time step corresponds to 0.5 seconds. \color{black} The action $a_i$ of deploying card $c$ by player $i$ can be represented as a tuple $(c, x, y)$ where $c$ is a card identifier and $(x, y)$ is the deploy position in the discrete $18 \times 32$ battle arena. The action of waiting is represented with a special \emph{Wait} card. Additionally, we augment the action space with cards in the hand that are illegal (with not enough Elixir). Choosing an illegal card forces the agent to intentionally wait until that card becomes available, after which it can choose to deploy any legal card or wait further. The action space augmented in this way aids uniform exploration of all cards in the game and we use this in all our experiments. Although the rules of Clash Royale are easy to learn, the game has great depth coming from predicting your opponent's moves, including their predictions of yours, which makes it hard to master. Playing Clash Royale effectively requires a well coordinated combination of attacks and defenses and fast adaptation to the opponents' deck and style of play. Further, because of limited Elixir resources and hidden information, waiting for a good deploy time is an important part of strategy. Below, we describe the various scientific challenges in learning to play Clash Royale: \begin{itemize} \item \textbf{Cyclic strategies.} The cards in the game are designed such that each card can be countered effectively with another card (that is, Clash Royale is a \emph{non-transitive game}). Like the game of rock-paper-scissors, there is no single best deterministic strategy. \item \textbf{Partial observability.} Cards of the opponent are hidden and are only revealed throughout the opponent's deploys. Players can deceive their opponents by choosing to hide cards (not deploy) until later in the game (akin to bluffing). \item \textbf{Exploration.} At any time during a battle in Clash Royale, only \emph{legal cards} (cards with costs less than the currently available Elixir) can be deployed. Naive exploration methods that choose random actions at each step leads to a greedy strategy of almost always deploying the card with the lowest cost (and thereby depleting Elixir). Good exploration strategies have to intentionally wait for the costlier cards. \item \textbf{Dynamic, large, and discrete action space.} Clash Royale has a large discrete action space with the possibility to deploy any of 100 cards in the $18 \times 32$ arena ($\sim$60,000 discrete actions). However, at a particular time in a battle, it is only possible to deploy from the legal cards in the hand. \end{itemize} \subsection{Pommerman} Pommerman is a popular multi-agent RL benchmark based on the classic Nintendo game Bomberman. Battles in Pommerman are played on a $11 \times 11$ board initialized randomly with rigid walls and wooden walls (that may contain some power-ups) and four players near each corner (see Fig.~\ref{fig:cr-screenshot}). The players can move in horizontal or vertical directions (that are not blocked by walls or bombs), collect power-ups or lay bombs in their current locations. A player dies when they are on a tile affected by a bomb blast and effective gameplay requires strategic laying of bombs to knock down all of the opponents. Hidden information in Pommerman consists of power-ups hidden inside wooden walls and the power-ups collected by other players. The Pommerman benchmark consists of different scenarios and we consider the Free-For-All (FFA) variant in this paper. The goal of each agent in the FFA mode is to be the last agent to stay alive within a fixed-length episode of 800 timesteps. The challenges in performing tree search on Pommerman involves: 1) the large branching factor (upto 1296) caused by four players simultaneously choosing from six actions, 2) the difficulty in credit assignment due to the presence of four players, and, 3) the common noisy rewards caused by suicides. \color{diffcolor} To assist learning, we mask out actions that immediately leads players into walls or flames (suicide). We use a Cython implementation of the Pommerman environment based on \cite{matiisen2018pommerman}. For clarity of our experimental setup and ease of reproducibility, we open source the code for our Pommerman experiments here: \url{https://github.com/rinuboney/l2p-pommerman}. \color{black} \section{Oracle Planner With Full Observability} \label{s:oracle} In our approach, we first build an oracle planner which has access to the full game state. The goal of planning is to discover the optimal sequence of actions that maximize expected rewards. A dynamic programming approach to the planning problem involves estimating expected rewards for every legal action in each state, after which one can act greedily by choosing the action with the largest expected reward. A policy $\pi_i$ of player $i$ is a distribution over actions available in state $s$ for player $i$, that is, $a_i \sim \pi_i(a_i | s)$. Let $\pi(a|s)=\pi_1(a_1|s)\pi_2(a_2|s)$ be the joint policy followed by players $i \in \{1, 2\}$. Let $z \sim p(z|s, \pi)$ be the probability distribution over the set of all terminal states induced by following policy $\pi$ from state $s$. The state value function $V_i(s)$ is the mean reward of player $i$ while players follow policy $\pi$ from state $s$: \begin{equation} V_i(s) = \E_{z \sim p(z|s, \pi)}[R_i(z)] \label{eq:v} \end{equation} The state-action value function $Q_i(s, a)$ is the mean reward of player $i$ while players first take actions $a = (a_1, a_2)$ and then follow policy $\pi$ from state $s$: \begin{align} Q_i(s, a) = \E_{z \sim p(z|s, a, \pi)}[R_i(z)] \end{align} A possible way to do planning is to estimate $Q_i(s, a)$ for each player and choose the action for each player which maximises its expected reward. One problem with this approach is that one has to consider all combinations of actions $(a_1, a_2)$, which is prohibitive in games like Clash Royale where each player chooses from tens of thousands of actions. In this paper, we take a different approach. We assume that the actions $a_1$ and $a_2$ are chosen independently, that is, we estimate $Q_i(s, a_i)$ taking an expectation over the opponent policy: \begin{equation} Q_i(s, a_i) = \E_{z \sim p(z|s, a_i, \pi)}[R_i(z)] \,. \label{eq:q_i} \end{equation} With this approximation, the problem formulation can be seen as a Partially Observable Markov Decision Process (POMDP) from the perspective of each player, where the opponent is subsumed into the stochastic environment. At the end of planning, each player independently chooses the action that maximises the estimated Q values: \[ a_i = \mathop{\mathrm{argmax}}_{a_i \in A_i(s)} Q_i(s, a_i) \,. \] \subsection{Monte Carlo Search (MCS)} Monte Carlo search (MCS) \cite{anthony2019policy} is a simple search method where $Q_i(s, a_i)$ is estimated for all actions $a_i \in A_i(s)$ by performing several iterations of random \emph{rollouts} from state $s$. That is, both players estimate $Q_i(s, a_i)$ assuming that policies $\pi_1$ and $\pi_2$ are uniform distributions over the legal actions in every state. In practice, we perform random rollouts for a fixed number of steps and then use a value function estimate $V$ to evaluate the final state. In each iteration of MCS from state $s$, both players independently and randomly choose actions $a_i \in A_i(s)$ and continue to do so for a fixed number of steps (planning horizon), to reach state $\tilde{s}$. At the end of an iteration, the estimate of $Q_i(s, a_i)$ is updated based on the value estimate $V(\tilde{s})$. \color{diffcolor} \subsection{Multi-Armed Bandits (MAB)} \label{sec:mab} Monte Carlo search can be improved by exploring more promising actions more often. This can be achieved by viewing action selection as a multi-armed bandit (MAB) problem: In the current state $s$, player $i$ has to choose an action $a_i \in A_i(s)$ with maximum expected reward. There are $|A_i(s)|$ arms and player $i$ can explore new actions or exploit actions with highest value estimates. When MCS is enhanced by MAB, the MAB selection is done at the current state $s$ and the value estimates $Q_i(s, a_i)$ are obtained as in MCS by performing random rollouts. In this paper, we use a \emph{decoupled} approach to action selection: each player independently chooses an action $a_i \in A_i(s)$ using its own instance of an MAB, thus the opponents are subsumed into the stochastic environment. We consider two popular MAB algorithms: the Upper Confidence Bound (UCB) and Thompson sampling. \subsubsection{Upper Confidence Bound} UCB algorithms estimate the upper confidence bound that any given action is optimal \cite{browne2012survey, silver2017mastering}. While there exist different variations of UCB, we consider the commonly used UCB1 variant introduced in \cite{auer2002finite}. Each player $i$ independently estimates the upper confidence bound $\UCB_i(s, a_i)$ for each action $a_i \in A_i(s)$ as: \begin{align} \UCB_i(s, a_i) = Q_i(s, a_i) + c \sqrt{\frac{\log N}{n_{a_i}}} \,, \label{eq:ucb} \end{align} where the $c$ hyperparameter controls the exploration-exploitation trade-off, $n_{a_i}$ is the visit counts of action $a_i$ and $N = \sum_{a_i \in A_i(s)} n_{a_i}$. In each iteration, the action with the highest UCB value is chosen deterministically. At the end of planning, normalized visit counts define a probability distribution over actions. The final action can be chosen stochastically by sampling from this distribution of by deterministically choosing the action with the highest visit count. \subsubsection{Thompson Sampling (TS)} Thompson sampling \cite{thompson1933likelihood} maintains probability distributions of cumulative rewards for each action and chooses actions according to the probability that they are optimal. Since the rewards in Clash Royale and Pommerman are binary, the probability that taking action $a_i$ will lead to a win can be modeled using the Bernoulli distribution. The mean parameter $\theta_{a_i}$ of the Bernoulli distribution can be modeled with a Beta distribution which is the conjugate prior distribution for the Bernoulli likelihood. The parameters of the Beta ditribution can be updated by maintaining win and loss counts ($S_{a_i}$ and $F_{a_i}$ respectively) for each action. Note that this posterior update assumes independent samples from a Bernoulli distribution, even though this is not true in a multi-agent setting. During each iteration of planning, the action is chosen as \begin{align*} a_i &= \mathop{\mathrm{argmax}}_{a_i \in A_i(s)} \theta_{a_i} \\ \theta_{a_i} &\sim \Beta(S_{a_i} + \alpha, F_{a_i} + \beta) \end{align*} In all the experiments in the paper, we set $\alpha=\beta=1$ and do not tune these hyperparameters. At the end of planning, the final action can be chosen stochastically in a similar manner or deterministically based on the estimated means of the Beta distributions. \begin{figure*}[t] \centering\scriptsize \includegraphics[width=.8\linewidth]{trees.png} \\ \hspace{-27mm}MCTS\hspace{65mm}FDTS \caption{\color{diffcolor} Illustration of how a search tree is modified in one planning iteration in MCTS (left) and FDTS (right).} \label{fig:fdts} \end{figure*} \subsection{Monte Carlo Tree Search (MCTS)} MCS described previously has several limitations: 1) it only plans actions for the current state and hence cannot discover effective action combinations, 2) it discards all information about future states and actions traversed during rollouts and plans from scratch in each step, and 3) the rollout policy is random and hence the estimated $Q$ values are under the assumption that both players will act randomly in the future. MCTS builds upon MCS by considering action selection in all states encountered during rollouts as a multi-armed bandit (MAB) problem. MCTS is a best-first tree search algorithm and begins from a root node corresponding to current state $s$. We start with the most common variant of MCTS in which each MCTS iteration from current state $s$ consists of the following steps: \begin{enumerate} \item \textbf{Selection-expansion.} Starting at the root node (which corresponds to the current state of the game), a \emph{tree policy} is used to descend through the tree until a new state $s'$ is reached. In the case of two players acting simultaneously, the tree policy can be implemented by both players independently choosing actions $a_i \in A_i(s)$ using one of the MAB algorithms discussed in Section~\ref{sec:mab}. \item \textbf{Evaluation.} The value $V(s')$ of the new state $s'$ is evaluated, which can be done in different ways: 1)~by applying a handcrafted or a learned value function to $s'$, 2)~by random rollout(s) from state $s'$ until a terminal state $z$ and using $R(z)$ as a Monte Carlo estimate of the value, or 3)~by a fixed length rollout and applying a value function to the reached state. \item \textbf{Backup.} The values $Q_i(s, a_i)$ for all the ancestors of node $s'$ are updated using the estimate $V(s')$ and the visit counts $n_{a_i}$ are incremented by one. \end{enumerate} See Fig.~\ref{fig:fdts} for a simplified illustration of one MCTS iteration. After several planning iterations, both players independently choose their best actions and the search tree built by MCTS is re-used for planning in subsequent states by moving the root node to the child node corresponding to the chosen joint action. MCTS allows for discovery of effective sequence of actions, reuse of statistics computed from previous states and iterative improvement of the rollout policy. A potential problem with MCTS is that the selection-expansion step may stop very early in the tree. This is likely to happen in the games with a large branching factor of the search tree. It is very probable that the tree policy will encounter a novel game state in one of the upper levels of the tree, after which the state is evaluated. This can limit the effective planning horizon of MCTS and makes it problematic to properly evaluate long-term plans. \begin{figure*} \centering \includegraphics[width=\linewidth, trim=0 12 0 12, clip]{revisits.pdf} \caption{\color{diffcolor} Comparison of different planners based on the reuse of information stored in the search tree, in a game of Pommerman. A Pommerman game can last for a maximum of 800 steps and in each step we execute the planning procedure for 100 iterations (and a fixed horizon of 20 in the case of FDTS). We plot the (low-pass filtered) ratio of state revisits during planning at each game step (that is, in all of the times the planner visits a state in depth $d$ of the search tree, the ratio of states that it has previously visited.) We use this to measure the effectiveness of use of information stored in the search tree. The best-performing FDTS+TS planner frequently reuses information, even up to the maximum depth of 20. \color{black} } \label{fig:revisits} \end{figure*} \subsection{Fixed-Depth Tree Search (FDTS)} We propose to improve MCTS by encouraging planning at least several steps ahead from the current state. The proposed algorithm, that we call \emph{fixed-depth tree search (FDTS)}, consists of the following steps: \begin{enumerate} \item \textbf{Selection-expansion-rollout.} Starting at the root node, an MAB tree policy is applied exactly $k$ times to descend through the tree. If the game reaches a novel state at a particular level, a new node is added to the tree and the tree policy continues action selection from that node until a desired depth level $k$ is reached. This steps results in creating a new branch with a leaf node with state $s'$ at a particular depth level. \item \textbf{Evaluation.} The value of the node state $s'$ reached at depth $k$ is evaluated. In our experiments, the evaluation step is done by applying a handcrafted value function without performing random rollouts. \item \textbf{Backup.} The values $Q_i(s, a_i)$ for all the ancestors of node $s'$ are updated using the estimate $V(s')$. \end{enumerate} One iteration of FDTS is illustrated in Fig.~\ref{fig:fdts} and the Python pseudocode for FDTS can be found in Listing~\ref{code:fdts}. The proposed algorithm can be viewed as combining in one step the selection-expansion step and the fixed-length rollout part of the evaluation step of classical MCTS. After a novel state is reached, the MAB algorithm is recursively used to expand that node into a branch that reaches a fixed tree depth $k$. This is essentially equivalent to a random rollout. The important difference is that we add nodes to the tree for all the states encountered during the random rollout. Keeping the trajectories encountered during random rollouts may seem wasteful, especially for problems with a large branching factor. However, this turns out to work well in the games considered in this paper because the MAB selection process systematically re-visits nodes existing in the tree despite of the large branching factor. In Fig.~\ref{fig:revisits}, we demonstrate that the FDTS equipped with UCB and especially with TS re-uses information collected in the previous planning steps. The increased percentage of re-visited nodes in FDTS compared to MCTS suggests that storing the rollout trajectories in the search tree is indeed beneficial. The same figure shows that Thompson sampling tends to re-visit existing nodes more often than UCB and this further improves the quality of planning, which is supported by our experimental results. \subsection{Memory and Computation Requirements} MCS is simple to implement and has minimal memory and computation requirements. MCS only stores statistics of legal actions in the current state. MCTS and FDTS requires storing statistics of legal actions in all previously visited states of an episode. The main computation in MCS is the stepping forward of the game state using the game engine. MCTS and FDTS further requires more computation at every state for action selection using an MAB algorithm. \color{black} \section{Experiments on Planning With the Oracle} In this section, we evaluate the proposed planning algorithms on the games of Pommerman and Clash Royale. Although optimal policies in multiplayer games are stochastic, similar to \cite{perick2012comparison}, we observe that deterministic policies perform better in practice. In all the experiments presented in this paper, we deterministically choose the action with the highest value at the end of planning. \subsection{Pommerman} Since Pommerman FFA is a four player game, we compare different planning algorithms by pitting them against three copies of the strong rule-based agent that is provided along with the Pommerman environment. \color{diffcolor} It is important to note that the proposed algorithms perform planning in the self-play mode using decoupled action selection for each player, that is they are not aware of the policy of the rule-based agents. Planning against known agents would be a much easier task. In Pommerman, the number of legal actions for each player can vary from 1 to 6, that is, the branching factor of the search tree can vary from 1 to 1296. In all our experiments, we perform 100 simulations of the planning algorithm at every time-step and use a planning depth of $k=20$ (in MCS and FDTS). \color{black} In the evaluation step of tree search, we simply use the reward function of the Pommerman as the value function \cite{matiisen2018pommerman}. \color{diffcolor} In Table~\ref{t:pommerman_oracle}, we report the number of wins, draws and losses in 400 games for different settings. We consider three planning algorithms: MCS, MCTS and FDTS, and two alternative ways for actions selection: Thompson sampling and UCB1 with $c=2$. For a fair comparison to MCS and FDTS, we use MCTS with random rollouts (at the end of the expansion step in an MCTS iteration, we perform random rollouts till a fixed depth of 20 and use that state for evaluation), which is similar to FDTS except that we do not add the nodes visited during the random rollouts to the search tree. A comparison of MCTS performance with and without this random rollouts is reported in Table~\ref{t:mcts-random}. The best results are obtained with FDTS+TS which attains a win-rate of $51.3\%$ with no reward shaping. A similar setup of self-play planning on a Java implementation of the Pommerman environment was considered in \cite{perez2019analysis} who reported win rates of $46.5\%$ for MCTS and $33.0\%$ for Rolling Horizon Evolutionary Algorithm \cite{perez2013rolling} using shaped rewards. \color{black} \begin{table}[!t] \caption{A four-player Pommerman FFA: Percentages of wins, draws and losses against three rule-based opponents computed after 400 games. The first column is evaluation of the rule-based agent against three copies of itself. } \label{t:pommerman_oracle} \centering \begin{tabular}{l|c|cc|cc|cc} \toprule & \multirow{2}{*}{Rule-based} & \multicolumn{2}{c|}{MCS} & \multicolumn{2}{c|}{\color{diffcolor} MCTS} & \multicolumn{2}{c}{\textbf{FDTS}} \\ & & UCB & TS & {\color{diffcolor} UCB} & {\color{diffcolor} TS} & UCB & \textbf{TS} \\ \midrule Wins & $19.2$ & $35.8$ & $36.3$ & {\color{diffcolor} $37.3$} & {\color{diffcolor} $40.8$} & $42.3$ & $\mathbf{51.3}$ \\ Draws & $23.4$ & $41.0$ & $37.5$ & {\color{diffcolor} $32.4$} & {\color{diffcolor} $36.1$} & $32.5$ & $\mathbf{29.0}$ \\ Losses & $57.4$ & $23.2$ & $26.2$ & {\color{diffcolor} $30.3$} & {\color{diffcolor} $23.1$} & $25.2$ & $\mathbf{18.2}$ \\ \bottomrule \end{tabular} \end{table} \subsection{Clash Royale} In Clash Royale, the number of discrete actions $A_i(s)$ is very large but the actions are correlated: deploying a card on nearby positions tend to produce the same outcome. To approximate a good policy, we sample a random set of 64 positions from the space of legal positions for every legal card. A sufficiently large random set would include the optimal deploy positions. \color{diffcolor} With this approximation, in Clash Royale, there are two players and the legal actions for each player (with the random sampling of deploy positions) can vary from 1 to 257. That is, the branching factor of the search tree can vary from 1 to 66049. \color{black} In our experiments, we use simple handcrafted value functions for oracle planning: we compute $V(s)$ by doing a rollout from state $s$ assuming that both players do not deploy any more cards. Since the consequences of already deployed cards have predefined behaviour, we can reach state $s'$ where the battle arena only contains towers. Then, we evaluate $V(s)$ using the terminal reward function $R(s')$. We compare UCB1 with $c=1$, Thompson sampling and simple random sampling using Monte Carlo search, by pitting one MAB algorithm against another. For example, to compare Thompson sampling with UCB, Player~1 performs planning using Thompson sampling for action selection of both players and Player~2 independently performs planning using UCB for action selection of both players. We compute the win rate of of an algorithm against another for 400 games in this setting. The results are shown in Table~\ref{t:cr-mab}. Both Thompson sampling and UCB clearly outperform random sampling. Thompson sampling performs the best of all. \color{diffcolor} While UCB could potentially be fine-tuned to work better with a more comprehensive search over the hyperparameters or using a different UCB variants such as UCB1-Tuned \cite{auer2002finite}, we found Thompson sampling to robustly work well in most settings. To test the robustness of Thompson sampling, we compare it against UCB with different value of exploration hyperparameter $c$ and planning horizon/depth. The win rates of the comparison in Clash Royale is shown in Table~\ref{t:ucb}, where Thompson sampling clearly outperforms UCB in most tested settings. \color{black} We therefore use Thompson sampling as the MAB algorithm in all our further experiments. \begin{table}[t] \caption{Clash Royale: Comparison of MCS, MCTS and FDTS. The planning horizon is $k=50$. Shown are win rates and $95\%$ confidence intervals. } \label{t:cr-oracle} \centering \begin{tabular}{lc} \toprule & Win rate \\ \midrule MCTS vs MCS & $41.3\pm4.5\%$ \\ \textbf{FDTS} vs MCTS & $96.5\pm1.8\%$ \\ \textbf{FDTS} vs MCS & $80.3\pm3.9\%$\\ \bottomrule \end{tabular} \end{table} \begin{table}[t] \caption{Clash Royale: Comparison of MCS with MCTS and FDTS based on their win rates by evaluating each pair on 40 games. } \label{t:mcts-ablation} \centering \begin{tabular}{ccc} \toprule Horizon & MCTS vs MCS &\textbf{FDTS} vs MCS \\ \midrule 10 & $43.7\%$ & $\mathbf{68.1\%}$ \\ 25 & $19.4\%$ & $\mathbf{70.5\%}$ \\ 50 & $41.3\%$ & $\mathbf{80.3\%}$ \\ \bottomrule \end{tabular} \end{table} We compare MCS, MCTS and FDTS in Clash Royale by pitting one algorithm against another for 400 games, where each player independently performs planning using the assigned algorithm. The results of our experiments are shown in Table~\ref{t:cr-oracle}. The proposed FDTS planning achieves the best performance. For further comparison of MCTS and MCS, we pit the two variations of MCTS against MCS for different planning horizons. The win rates on 40 games of Clash Royale are shown in Table~\ref{t:mcts-ablation}. FDTS outperforms MCS on all planning horizons, with an increased difference for deeper search. These results suggest that FDTS is able to discover better combinations of actions and re-uses statistical information (as demonstrated in Fig.~\ref{fig:revisits}) to outperform MCS for all planning horizons, with an improved performance as the planning horizon increases. \color{black} \section{Training Follower Policy with Partial Observability} \label{s:follower} Planning enables competitive play with generalization to unseen states. However, the oracle planner has two limitations: 1)~It performs many rollouts to make decisions in every state, requiring a game implementation that must run much faster than real-time, to be able to act in a real-time battle. 2)~The oracle planner cheats by having access to the full game state: private information like the deck and hand of the opponent in Clash Royale and hidden power-ups in Pommerman becomes visible during future states of planning rollouts. This could be avoided by randomizing hidden information during planning but the game engines of these games do not support this. In our approach, we propose to use imitation learning to train a \emph{follower} policy network to perform similarly to the oracle planner but under real-time computation and partial observability. One straightforward way of doing this would be via cloning of the oracle behavior: one can collect trajectories generated by the oracle planner with self-play and use that data to train the follower policy. However, this approach results in a relatively poor performance (see Table~\ref{t:pommerman-follower}). \color{diffcolor} We instead use the DAgger algorithm \cite{ross2011reduction} for better performance. In DAgger, the follower policy makes decisions during self-play and the oracle planner is used to compute better actions in the states encountered by the follower. Since the follower is initially random, self-play encounters diverse states and the oracle planner provides stable training targets, leading to an efficient improvement of the follower. The algorithm for training follower networks is listed in Algorithm~\ref{alg:main}. \color{black} \begin{algorithm}[H] \caption{Learning to play imperfect-information games by imitating an oracle planner with DAgger} \begin{algorithmic}[1] \STATE Initialize follower policy $\pi_f$ and replay buffer $\mathbb{D}$. \FOR{each episode} \STATE Initialize oracle planner $\pi_o$. \FOR{time $t$ until the episode is over} \color{diffcolor} \STATE Compute follower actions $a_f = \pi_f(o_t)$ from partial observations $o_t$, and apply them to the game. \STATE Compute oracle actions $a_o \sim \pi_o(s_t)$ using (self-play) tree search, with access to full state $s_t$. \STATE Add data $(o_t, a_o)$ to replay buffer $\mathbb{D}$ and train follower policy $\pi_f$ using $\mathbb{D}$ to predict the oracle actions $a_o$ from partial observations $o_t$. \color{black} \ENDFOR \ENDFOR \end{algorithmic} \label{alg:main} \end{algorithm} \section{Experiments on Training the Follower} We train follower networks to imitate the oracle planner by predicting the oracle action from partial observations $o_i$. The oracle is chosen to be the best performing fixed-depth tree search (FDTS) with Thompson sampling (TS). \subsection{Pommerman} In Pommerman, we train a follower network to imitate the oracle planner on 500 battles. We use the same network architecture as \cite{DBLP:journals/corr/abs-1809-07124}. The observations are represented in a $11 \times 11$ spatial representation (corresponding to the $11\times11$ board in the game), with 14 feature maps. The features represent presence and positions of 10 different objects in the board, bomb blast positions and lifetime and the power-ups collected by the agent. The network architecture consists of four convolutional layers with 32 channels (with ReLU activations) and a final linear layer that predicts the softmax probabilities of the six discrete actions. We used random search to tune the hyperparameters of the oracle planner and the follower policy. All hyperparameters along with their search range and final values for Pommerman are reported in Table~\ref{t:pommerman-hp}. \begin{table}[!t] \caption{A four player Pommerman FFA: Comparison of follower agents against three rule-based baseline opponents on 400 games. Evaluation of the rule-based agent against three copies of itself is reported in the first column for reference. } \label{t:pommerman-follower} \centering \begin{tabular}{l|c|cc} \toprule & \multirow{2}{*}{Rule-based} & \multicolumn{2}{c}{\textbf{Follower}} \\ & & \textbf{DAgger} & Oracle-behavioral cloning \\ \midrule Wins & $19.2\%$ & $\mathbf{23.3\%}$ & $17.4\%$\\ Draws & $23.4\%$ & $\mathbf{22.5\%}$ & $19.2\%$ \\ Losses & $57.4\%$ & $\mathbf{54.2\%}$ & $63.4\%$ \\ \bottomrule \end{tabular} \end{table} \begin{table*}[t] \caption{Clash Royale: Comparison of all agents based on win rates (and $95\%$ confidence intervals) of each pair evaluated on 100 games. Win rates of Q-MC and Follower are averaged over win rates of networks trained used 5 different seeds.} \label{t:cr-follower} \centering \begin{tabular}{lccccc} \toprule & Random & Q-MC & Human-BC & Follower & Oracle \\ \midrule Random & - & $20.9\pm7.9\%$ & $1.0\pm1.9\%$ & $1.4\pm2.3\%$ & $0.0\pm0.0\%$ \\ Q-MC & $79.1\pm7.9\%$ & - & $14.9\pm6.9\%$ & $8.6\pm4.3\%$ & $0.0\pm0.0\%$ \\ Human-BC & $99.0\pm1.9\%$ & $85.1\pm6.9\%$ & - & $28.6\pm8.8\%$ & $6.1\pm4.7\%$ \\ \textbf{Follower} & $\mathbf{98.6\pm2.3\%}$ & $\mathbf{91.4\pm4.3\%}$ & $\mathbf{71.4\pm8.8\%}$ & - & $\mathbf{6.6\pm4.8\%}$ \\ Oracle & $100.0\pm0.0\%$ & $100.0\pm0.0\%$ & $93.9\pm4.7\%$ & $93.4\pm4.8\%$ & - \\ \bottomrule \end{tabular} \end{table*} We evaluate the Pommerman follower against three rule-based opponents and the results are shown in Table~\ref{t:pommerman-follower}. The follower agent trained with DAgger is able to achieve a win rate 23.3\%, outperforming the rule-based agent and even the MCTS oracle planner. Note that previous works have achieved high win-rates against the rule-based opponent by directly training against it \cite{matiisen2018pommerman, gao2019skynet}. Instead, we learn purely from self-play, which yields an agent that is able to compete with different kinds of opponents. \subsection{Clash Royale} In Clash Royale, we train a follower network to imitate the oracle planner on 300 battles. The follower is a convolutional neural network that predicts the $Q$ values (means of the Beta distributions) of the spatial deploy positions for all legal cards. During self-play and evaluation, the follower network deterministically chooses the action with the largest $Q$ value. In Clash Royale, the objects in the battle arena, battle progress, current cards and the past 10 actions are represented using $18\times32$ spatial feature maps (corresponding to the $18\times32$ battle arena in the game). Card types and object types in the battle arena are represented using learnable embeddings. The follower network predicts the $Q$ values of all deploy positions of all legal cards based in these spatial features. All hyperparameters of the follower network along with the random search range and final values are reported in Table~\ref{t:cr-hp}. We evaluate the follower against three baseline agents: 1) \emph{Random}: a simple uniform random policy, 2) \emph{Q-MC}: a model-free agent trained with Monte Carlo value targets \cite{adkb2018td}, and 3)~\emph{Human-BC}: a strong agent trained to imitate human actions. DQN \cite{mnih2015human} was not included in the comparison because it was unstable, most likely due to the large action space and delayed actions. \color{diffcolor} \textsc{Human-BC} is a very strong baseline: it is a mature agent that has been in production for over a year. That agent was trained using behavioural cloning (supervised learning) to imitate human actions from 76 million frames of human replay data from Clash Royale. These replays consisted of games played by humans with a good skill level, all from 4000 trophies and above, and played with a diverse set of decks. The architecture of \textsc{Human-BC} and the training parameters were tuned for metrics like prediction accuracy of deployed cards and their deploy positions. The \textsc{Human-BC} agent consists of two feature extraction networks and an action prediction network. A battle arena feature extraction network embeds the objects (along with their features) in the battle arena in a spatial grid based on their positions and extracts features from the spatial inputs using residual blocks. A battle context feature extraction network extracts battle context features based on cards and battle progress, similar to the follower network architecture, but with a larger network consisting of residual blocks. The battle arena and battle context features are combined using a sum operation and an action prediction network consisting of residual blocks predicts: 1)~when to deploy, 2)~card to be deployed, 3)~deploy position, and 4)~value of current state (auxiliary task). The predicted card is deployed onto the predicted deploy position only if the policy predicts that it should be deployed in the current step. The win rates of all pairs of agents are presented in Table~\ref{t:cr-follower}. The Q-MC agent does not perform very well as it is able to beat only the random agent. By analyzing its playing style, one can notice that it tends to learn a particular strategy that is easily predictable by human players. The Human-BC agent is very competitive, the analysis of its gameplays suggests that is able to use strategies which are common for human players. \color{black} The oracle planner beats the other agents almost always, which is natural because it has access to more information. By analyzing its gameplays, we observed that the oracle planner was able to discover effective strategies commonly used by human players.\footnote{\color{diffcolor} See \url{https://sites.google.com/view/l2p-clash-royale} for our supplementary video including the gameplay videos.} Some of the discovered strategies are \begin{enumerate} \item \textbf{Groups of troops.} The planner is consistently playing high-hitpoint ``tank'' troops like \emph{Giants}, \emph{Knights}, or \emph{Baby Dragons} in the front, and support units like \emph{Musketeers} or \emph{Archers} behind the tank. This is a key strategy for successful attacks that requires coordinating deploys across several timesteps. \item \textbf{Defense against tanks.} When attacked by a single tank unit without support units, the planner deploys high DPS (damage per second) troops like \emph{Musketeer} or \emph{Minions} to directly and efficiently remove the tank. However, if there are support units behind the tank, then the defending planner typically tries to destroy the support units first, to minimize potential tower damage from such more threatening attacks. \item \textbf{Hedging.} Clash Royale games often have pivotal moments where one of the players must decide between two high level strategies: trying to defend against an oncoming attack, or hedging bets by skipping defense and launching a similarly powerful attack on the other lane. The planning agent is able to decide to forgo defense and respond with an attack against the other tower. \item \textbf{Slowing down attacks.} If an attack is approaching but there are no good defense cards in the hand, the planner is able to deflect a threatening attack by deploying a tank like \emph{Giant} to slow down the attack and thus rotating more suitable cards to the hand. \item \textbf{Race against time.} In the end of the game, when both players are equally close to winning, it's essential to damage the opponent's king tower quicker than the opponent damages yours. In these scenarios, the planner is coordinating all deploys at the king tower, using even weak damage from spells like \emph{Arrows}. \end{enumerate} \color{diffcolor} Training the follower network with the oracle supervision resulted in a Follower agent which outperforms the very strong \textsc{Human-BC} baseline. Although the Follower does not have access to the full game state, it successfully uses the strategies discovered by the oracle, which we observed by analyzing its playing style. \color{black} \section{Related Work} Previous works on RL in games with high-dimensional state-action spaces such as StarCraft~II \cite{vinyals2019grandmaster}, Dota~2 \cite{berner2019dota} and Honor of Kings \cite{ye2019mastering} have used model-free RL algorithms \cite{schulman2017proximal, espeholt2018impala}, requiring a large amount of data to learn. We take a model-based planning approach to learn to play imperfect-information games. Previous works have found MCTS to be an effective planning algorithm in various simultaneous-move games with low-dimensional state-action spaces \cite{shafieicomparing, teytaud2011lemmas, bovsansky2016algorithms}, even though it does not have any theoretical guarantees on achieving optimal play in simultaneous-move or imperfect-information games and can be exploited by regret minimization algorithms \cite{shafieicomparing}. MCTS has been used for planning in imperfect-information games essentially by \emph{determinization} of the hidden information \cite{frank1998search, ginsberg2001gib, bjarnason2009lower}, also known as Perfect Information Monte Carlo (PIMC) \cite{long2010understanding}. The determinization technique involves performing several instances of the MCTS procedure with different randomizations of the hidden information and average across the resulting policies. Information Set MCTS (IS-MCTS) \cite{cowling2012information} involves determinization of hidden information in each MCTS iterations to construct a search tree of information sets. MCTS algorithms that use determinization \cite{frank1998search, ginsberg2001gib, bjarnason2009lower, cowling2012information} are not applicable to complex games or real-world problems, where it is not possible to randomize hidden information. In this paper, we introduce an algorithm for efficient planning and learning in imperfect-information games by using a function approximator to average across the resulting policies produced by an oracle planner that has access to the hidden information. Even though averaging across different actions computed by the oracle in different states are not optimal, similar to previous works \cite{frank1998search, ginsberg2001gib, schafer2008uct, buro2009improving, bjarnason2009lower}, we found it effective in learning strong policies. Learning to play card-based real-time strategy (RTS) games was previously considered in \cite{liu2019} using DQN to learn to select cards and computing the deploy positions in a post-hoc manner using an attention mechanism, which is suboptimal as the deploy positions are never trained. Guo et al. \cite{guo2014deep} used imitation learning of an MCTS planner in the simpler single-player setting of Atari games, with full obervability and a small number of discrete actions. We show that the naive MCTS used in \cite{guo2014deep} is problematic in imperfect-information simultaneous-move games with large action spaces and introduce fixed-depth tree search with Thompson sampling for better planning. \color{diffcolor} Combinatorial multi-armed bandit (CMAB) algorithms can be applied in settings where the action space of each player consists of combinations of multiple variables \cite{gai2010learning, chen2013combinatorial, ontanon2013combinatorial}. For example, in Clash Royale, an action consists of a card and the ($x$ and $y$) deploy position of the card. In this work, we resort to use of MAB algorithms as the combinations of 4 cards and a random sample of 64 deploy positions are limited to only 256 arms. Alternatively, CMAB algorithms can be used for a proper treatment of combinatorial action spaces with very large branching factors \cite{ontanon2017combinatorial}. \color{black} \section{Discussion} \color{diffcolor} We demonstrate good performance on learning to play in the novel setting of Clash Royale and the challenging multi-agent RL benchmark of Pommerman. Our approach consists of an oracle planner that has access to the full state of the environment and a follower agent which is trained to play the imperfect-information game by imitating the actions of the oracle from partial observations. We demonstrate that naive MCTS is problematic in high-dimensional action spaces. We show that fixed-depth tree search (FDTS) and Thomspon sampling overcome these problems to discover efficient playing strategies in Clash Royale and Pommerman. The follower policy learns to implement them from scratch by training on a handful of battles. Our two-step approach can be combined in an iterative fashion by improving the oracle planner using $Q$ estimates from the follower policy. Potential directions of future work include exploration of regret minimization algorithms used in Poker \cite{zinkevich2008regret, moravvcik2017deepstack}. \color{black} While Clash Royale serves as a novel setting of reinforcement learning research, learned agents also have several use cases in game design. For example: 1)~agents can do automated testing of new game content, such as new cards or levels, 2)~agents can be used as practice opponents, 3)~new single player games can be designed where humans play against computer agents, and 4)~agents can provide assistance to new players during tutorial or unlocking of new cards. \section{Societal Impacts} The research presented in this paper can have an impact on the gaming industry. On the positive side, self-play algorithms can replace handcrafted rules which are currently most widely used for: 1)~designing bots that play a game in the place of a human, 2)~producing game content like \emph{boss levels} (fights against a strong computer-controlled enemy). Designing rule-based bots which are game-specific and difficult to maintain is an expensive component of game development, replacing this component with a general self-play algorithm can have strong impact on the industry. Self-play bots can also be easily retrained and used to reduce manual work for game testing, which involves finding bugs and assessing the difficulty levels of a game. On the negative side, in the wrong hands, skillful bots can be used for cheating in the game, which is a major issue in video games, especially in online games \cite{blackburn2014cheating, zuo2016bad, paay2018motivations}. Bots can be used to cheat by providing unfair advantage to a player during gameplay. If players cannot know for sure that they are playing against other human opponents on equal grounding, it can erode the trust of the player community towards the game as a whole. Similarly to any other RL algorithm, our research results alone are not enough to enable cheating in games in general, because the model would have to be first trained against a specific game environment, and then integrated to the game software, both of which require low level access to the game engine. Overall, further research in data-efficient RL will increase the risk of bot misuse in games, but dealing with that is a line of future work. \color{diffcolor} \subsection{Decoupled Thompson sampling} The most common action selection algorithm used in simultaneous-move MCTS is Decoupled UCB (upper confidence bound) \cite{auer2002using}. In Decoupled UCB, each node in the tree consists of independent instances of UCB for each player. Each player $i$ independently estimates the upper confidence bound for each action $a_i \in A_i(s)$ as \[ \UCB_i(s, a_i) = Q_i(s, a_i) + c \sqrt{\frac{\log N}{n_{a_i}}} \,, \] where the hyperparameter $c$ controls the exploration-exploitation trade-off, $n_{a_i}$ is the visit count of action $a_i$ and $N = \sum_{a_i \in A_i(s)} n_{a_i}$. However, we found it difficult to tune the exploration parameter $c$ of UCB for games with large and dynamic action spaces (such as Clash Royale), in which the number of actions available in each state varies a lot. We therefore propose to use decoupled Thompson sampling as a solution which naturally handles the exploration-exploitation trade-off without any hyperparameter tuning. Thompson sampling \cite{thompson1933likelihood} maintains probability distributions of cumulative rewards for each action and chooses actions proportional to the probability that they are optimal. Since the rewards in Clash Royale and Pommerman are binary, the probability that each action $a_i$ is optimal is modeled using Bernoulli distributions with mean $\theta^*_{a_i}$. The means $\theta^*_{a_i}$ can be modeled using conjugate Beta distributions, which is implemented by maintaining win and loss counts ($S_{a_i}$ and $F_{a_i}$ respectively) for each action. Note that this posterior update assumes independent samples from a Bernoulli distribution, even though this is not true in a multi-agent setting. During each iteration of planning, the action is chosen as \[ a_i = \mathop{\mathrm{argmax}}_{a_i \in A_i(s)} \theta_{a_i} \,, \qquad \theta_{a_i} \sim \Beta(S_{a_i} + \alpha, F_{a_i} + \beta) \,. \] In all the experiments in the paper, we set $\alpha=\beta=1$ and do not tune these hyperparameters. At the end of planning, the final action can be chosen stochastically in a similar manner by sampling or deterministically based on the estimated means of the Beta distributions. \begin{enumerate} \item \textbf{Selection.} In state $s$, both players independently choose actions $a_i \in A_i(s)$ based on a MAB algorithm which leads the battle to a new state $s'$. Action selection in the new state $s'$ is also performed similarly using the same MAB algorithm. Nodes are added to the search tree corresponding to all encountered states with the edges corresponding to joint action taken to reach the state. Each node consists of an instance of an MAB algorithm used to perform action selection in that state/node. \item \textbf{Expansion.} From root state $s$, the selection step continues until it reaches a newly encountered state $\tilde{s}$ (that is, a state with no corresponding node in the search tree). A new leaf node is added to the search tree to represent the new state $\tilde{s}$. \item \textbf{Evaluation.} This step involves estimating the value $V(\tilde{s})$ of state $\tilde{s}$ \eqref{eq:v}. This can be performed by: 1) random rollout(s) from state $\tilde{s}$ until the battle reaches a terminal state $z$ (the outcome $R(z)$ is a Monte Carlo estimate of the value), 2) a handcrafted value function, or 3) a learned value function (one can train a function approximator to predict the expected outcome of the game from any state). {\bf discuss how this is done in CR and P.} \item \textbf{Backup.} Instances of MAB algorithms corresponding to all states encountered in an iteration are updated based on the value estimate $V(\tilde{s})$ to make better choices in future iterations. \end{enumerate} \subsection{Fixed-Depth Tree Search (FDTS)} MCS described previously has several limitations: 1) it only plans actions for the current state and hence cannot discover effective action combinations, 2) it discards all information about future states and actions traversed during rollouts and plans from scratch in each step, and 3) the rollout policy is random and hence the estimated $Q$ values are under the assumption that both players will act randomly in the future. The problems can be effectively mitigated by using tree search, that is, building a tree of all states visited during rollouts and considering action selection in these states as a multi-armed bandit problem. We propose to perform fixed-depth rollouts in each iteration and refer to this algorithm as Fixed-Depth Tree Search (FDTS). Each iteration of FDTS begins from the root node corresponding to current state $s$ and proceeds with the following three steps: \begin{enumerate} \item \textbf{Selection.} In state $s$, both players independently choose actions $a_i \in A_i(s)$ based on a MAB algorithm which leads the battle to a new state $s'$. Action selection in the new state $s'$ is also performed similarly using the same MAB algorithm. Nodes are added to the search tree corresponding to all encountered states with the edges corresponding to joint action taken to reach the state. Each node consists of an instance of an MAB algorithm used to perform action selection in that state/node. This is repeated for a fixed number of times to reach state $\tilde{s}$, that is, till we reach a fixed depth in the search tree. \item \textbf{Evaluation.} This step involves estimating the value $V(\tilde{s})$ of state $\tilde{s}$ (Equation~\ref{eq:v}). This can be performed by: 1) random rollout(s) from state $\tilde{s}$ until the battle reaches a terminal state $z$ (the outcome $R(z)$ is a Monte Carlo estimate of the value), 2) a handcrafted value function, or 3) a learned value function (one can train a function approximator to predict the expected outcome of the game from any state). \item \textbf{Backup.} Instances of MAB algorithms corresponding to all states encountered in an iteration are updated based on the value estimate $V(\tilde{s})$ to make better choices in future iterations. \end{enumerate} After several planning iterations, both players independently choose their best actions and the search tree built by FDTS is re-used for planning in subsequent states by moving the root node to the child node corresponding to the chosen joint action. FDTS allows for discovery of effective sequence of actions, reuse of statistics computed from previous states and iterative improvement of the rollout policy.

8b3eee431a960388d51da4b8f29a2deb2497c533

\section{Introduction} With the recent discoveries of gravitational waves and black holes, General Relativity (GR) has been one of the most successful theories in physics. However, several problems have not been covered by GR, these include dark matter, dark energy, and quantum gravity. Attempts to solve these problems have led to some motivation in modifying the original General Relativity, proposed by Einstein in 1915. Different types of modification had flourished with rapid progression; for a complete review on theories of modified gravity, one could consult \cite{Oikonomou}. A specific way to achieve a modified theory is by adding higher-order terms on curvature to the Einstein-Hilbert action. In $f(\mathcal{R})$-gravity, the higher-order terms are restricted to the power series of the Ricci scalar curvature \cite{Oikonomou,Sotiriou}. Another way to modify GR is by considering a general affine connection with torsion, and even non-metricity, hence the connection is not necessarily Levi-Civita as in the standard GR. Different theories of gravity do not use curvature to describe the effect of gravity, they instead use torsion in Teleparallel Gravity \cite{Pereira}, or non-metricity in Symmetric Teleparallel Gravity \cite{Nester,Jimenez}. There exists even a more general theory that uses both torsion and non-metricity (but zero curvature) in General Teleparallel Quadratic Gravity \cite{Jimenez3}. On the other hand, the Metric-Affine-Gravity (MAG) is a curvature theory of gravity that includes both non-metricity and torsion for the connection \cite{Hehl,Iosifidis}. The choice of action for MAG could vary greatly, for example, the Ricci scalar $\mathcal{R}$, the power series of Ricci scalar $f\left(\mathcal{R}\right)$, and other exotic actions as discussed in \cite{Iosifidis}. In this article, we only consider Metric-Affine General Relativity (MAGR), a class of MAG with Ricci scalar $\mathcal{R}$ as the curvature part of action. It could also be considered as a special case of the most general version of the $f(\mathcal{R})$-gravity, the Metric-Affine $f(\mathcal{R})$-gravity, that is a modified theory of gravity with $f(\mathcal{R})$-action and general affine connection \cite{Sotiriou}. The existence of torsion in the connection is important for the coupling of gravity with fermions \cite{Hehl2}, while the non-metricity part is important if one considers certain kinds of matter fields that contain a non-vanishing symmetric part of the hypermomentum \cite{Sotiriou,Hehl3,Hehl4,Hehl5}. To obtain the dynamics of a covariant gravitational theory, one needs to decompose the covariant fields into their temporal and spatial parts via the (3+1) decomposition. It was first considered by Darmois in \cite{Hehl}, Lichnerowicz in \cite{Lichnerowicz1,Lichnerowicz2,Lichnerowicz3}, then by Choquet-Bruhat in \cite{Choquet1,Choquet2}. However, it gained a lot of attention after the work by Arnowitt, Deser, and Misner in \cite{ADM}. The (3+1) ADM decomposition is a formulation of GR which split the spacetime and the covariant fields into the temporal and spatial components, thus allowing one to obtain the dynamics of the variables on the 3D hypersurface via the equation of motions. At the core of the ADM formulation, lies the Gauss-Codazzi-Mainardi equations, which relates the intrinsic (Riemann) curvatures of a manifold with the extrinsic and intrinsic curvature of its submanifold. The result of the formulation is the (3+1) Einstein Field Equations (EFE): 4 constraint equations and 6 dynamical ones. Attempts to apply the (3+1) formulation to $f(\mathcal{R})$-gravity had been done for special cases; in $f(\mathcal{R})$-gravity with Levi-Civita connection \cite{Nathalie,Paschalidis} and in teleparallel-gravity with torsion \cite{Shapiro}. Even more, the Hamiltonian analysis and possible quantization procedure of $f(\mathcal{R})$-gravity had been done in \cite{Olmo2,Ma1,Ma2}. Little attention had been given to the (3+1) formulation of gravity with non-metricity, with the latest work is concerning the derivation of the (1+3) Raychaudhuri equation \cite{Iosifidis,Iosifidis2}. However, for various reasons, it is interesting to perform the (3+1) decomposition for MAG and $f(\mathcal{R})$-gravity which includes non-metricity, particularly, if one considers matter fields with a general hypermomentum. With this motivation in mind, in this article, we apply the (3+1) ADM formulation for Metric-Affine General Relativity (MAGR) with general affine connection that includes torsion and non-metricity, as a special case to MAG and $f(\mathcal{R})$-gravity. The article is organized as follows: In Section II, we derive the generalized Gauss-Codazzi-Mainardi (GCM) equations for a general affine connection with torsion and non-metricity. A similar treatment is also applied to torsion, resulting in a torsion (3+1) decomposition. These results are valid for manifold with any dimension. For the derivation, we use 2 different definitions of extrinsic curvature that exist in the literature; both differ by the quantity $\nabla_{X}g$ (with $\left(\nabla,g\right)$ are the connection and the metric on the manifold, and $X$ is a vector on the hypersurface) if the connection is non-metric. The generalized GCM equations are written in terms of these quantities, together with the definition of the acceleration tensor. In Section III, we show that the metric compatibility and the torsionless condition of a connection $\nabla$ on a manifold are inherited to the connection $\,^{3}\nabla$ on the hypersurface. This was already a well-known result and was used to derive the original GCM equation for the Levi-Civita connection. Since the generalized Riemann curvature tensor of an affine connection is lacking the symmetries possessed by the Levi-Civita connection, we need to define three different types of independent contraction of the Riemann tensor, following the definition introduced in \cite{Iosifidis3,Jimenez2,Iosifidis4}. These contractions are important for the derivation of the Einstein tensor in the next section. Section IV consists of the physical application of the generalized GCM and the torsion decomposition. Minimization of the $f(\mathcal{R})$-action with respect to the variation of metric, connection, and a Lagrange multiplier results in 2 Euler-Lagrange equations and 1 constraint equation. However, considering the scope of this article, our focus will be on the generalized stress-energy-momentum equation (EFE): the one that comes from the variation of the metric. We perform the ADM formulation to the generalized EFE using the generalized GCM equation we derive in Section II, resulting in 3 parts of the equation: the energy, momentum, and stress-energy equation. With the introduction of some additional variables on the hypersurface, we could write the energy, momentum, and stress-energy part of the EFE in terms of tensor fields defined only on the hypersurface, motivated by the concept of geometrodynamics in \cite{Wheeler}. An interesting result here is, unlike the standard GR case, there is no constraint originating from the generalized EFE; all the 3 equations contain dynamics, i.e, contain the derivative of a quantity with respect to the time coordinate. However, with the Levi-Civita condition, one could recover the Hamiltonian and momentum (or diffeomorphism) constraint, together with the standard dynamics of GR. We discuss some subtleties on the results in Subsection VA, and for the completeness of the analysis in this article, we present a special case of (3+1) decomposition of the 2 remaining equations of motion: the hypermomentum and the projective constraint equation. This is done in Subsection VB. The detailed derivation and general treatment of these equations will be presented in our companion article. Finally, we conclude our works in Subsection VC. \section{The Generalized Gauss-Codazzi-Mainardi Equations} \subsubsection*{The Metric and Connection on the Hypersurface} Let $\mathcal{M}$ be an $n$-dimensional manifold, equipped with a (pseudo)-Riemannian metric $g$ and an affine connection $\nabla.$ In the most general case, $g$ and $\nabla$ are independent of one another. Let $\Sigma$ be the hypersurface, namely, an ($n$-1)-dimensional submanifold embedded on $\mathcal{M}.$ One could define the unit vector normal to $\Sigma$ at each point $p\in\Sigma$, namely, $\hat{n}_{p}$, satisfying $g_{p}\left(\hat{n}_{p},\hat{n}_{p}\right)=\pm1,$ where the $-$ sign (the one on the upper side) and the $+$ sign (the one on the lower side) depends on whether $\left(M,g\right)$ is Riemannian or Lorentzian, respectively. From this point up to the rest of the article, the signature on the upper side of the equation is the one valid for the Riemannian case, while the lower side is for the Lorentzian's. The existence of metric $g$ on $\mathcal{M}$ induces a metric on the hypersurface $\Sigma$ as follows: \begin{equation} \,^{3}q=g\mp\hat{n}^{*}\otimes\hat{n}^{*},\label{eq:1a} \end{equation} with $\hat{n}^{*}\in T_{p}^{*}\mathcal{M}$ is the covariant vector to $\hat{n},$ satisfying $\hat{n}^{*}=g\left(\hat{n},\cdot\right)$ (here, we omit the label $p$ for simplicity). Let $V,X\in T_{p}\Sigma$, then one could construct a new vector $\nabla_{V}X$ $\in T_{p}\mathcal{M},$ not necessarily an element of $T_{p}\Sigma$, since $\nabla$ is defined on $\mathcal{M}$. The decomposition of $\nabla_{V}X$ into the components parallel and perpendicular to $\Sigma$ is the following: \begin{equation} \nabla_{V}X=\underset{\left(\nabla_{V}X\right)_{\parallel}}{\underbrace{\nabla_{V}X\mp g\left(\nabla_{V}X,\hat{n}\right)\hat{n}}}\pm\underset{\left(\nabla_{V}X\right)_{\perp}}{\underbrace{g\left(\nabla_{V}X,\hat{n}\right)\hat{n}}.}\label{eq:x} \end{equation} The labels $\parallel$ and $\perp$ denote, respectively, the parallel part of $\nabla_{V}X$ which lies on $\Sigma$ and the part normal to $\Sigma.$ The covariant derivative $\nabla$ on the quantity $\left(\nabla_{V}X\right)_{\parallel}$ associates $V,X\in T_{p}\Sigma$ to a transformed vector $\left(\nabla_{V}X\right)_{\parallel}\in T_{p}\Sigma$, hence, one defines: \begin{equation} \left(\nabla_{V}X\right)_{\parallel}:=\,^{3}\nabla_{V}X=\nabla_{V}X\mp g\left(\nabla_{V}X,\hat{n}\right)\hat{n},\label{eq:2} \end{equation} where $\,^{3}\nabla$ is the covariant derivative on $\Sigma$. It could be easily shown that $\,^{3}\nabla$ is an affine connection. One could define an entirely different affine connection $^{3}\nabla$ on $\Sigma$, which in general, is not induced from $\nabla$. However, since the objective of the work is to obtain the generalized Gauss-Codazzi relation, we only consider the case where $^{3}\nabla$ is induced by $\nabla$ as in (\ref{eq:x}). \subsubsection*{Extrinsic Curvature of the First Kind} The component of the perpendicular part of (\ref{eq:x}) is defined as the extrinsic curvature of $\Sigma$ in the direction of $\left(V,X\right)$ as follows \cite{Baez}: \begin{equation} \left(\nabla_{V}X\right)_{\perp}:=K\left(V,X\right)=\pm g\left(\nabla_{V}X,\hat{n}\right).\label{eq:3} \end{equation} Written in terms of components, one could show that the extrinsic curvature of the first kind satisfies: \begin{equation} K\left(V,X\right)=\pm V^{\mu}X^{\beta}\omega_{\mu\,\:\,\beta}^{\,\:\,\alpha}n_{\alpha},\label{eq:z} \end{equation} using the fact that $n_{\alpha}V\left(X^{\alpha}\right)=0$, since $X,V\in T_{p}\Sigma$ are perpendicular to $\hat{n}$. The quantity $\omega_{\mu\,\:\,\beta}^{\,\:\,\alpha}\partial_{\alpha}=\nabla_{\mu}\partial_{\beta}$ is the vector potential of $\nabla$. Notice that in general, $\omega_{\mu\,\:\,\beta}^{\,\:\,\alpha}$ of an affine connection does not possess any symmetry in the indices $\left(\mu,\alpha,\beta\right),$ therefore, one needs to be careful not to switch the position of these indices. With definitions (\ref{eq:2}) and (\ref{eq:3}), (\ref{eq:x}) could be written as: \begin{equation} \nabla_{V}X=\,^{3}\nabla_{V}X+K\left(V,X\right)\hat{n}.\label{eq:y} \end{equation} There exists another definition of extrinsic curvature that will be explored in the next sections. \subsubsection*{The Torsion Tensor} The torsion tensor on $\mathcal{M}$ is defined as: \begin{equation} T\left(X,Y\right)=\nabla_{X}Y-\nabla_{Y}X-\left[X,Y\right],\qquad X,Y\in T_{p}\mathcal{M}.\label{eq:torsion} \end{equation} For the case where $X,Y\in T_{p}\Sigma,$ using the parallel and perpendicular decomposition of $\nabla_{V}X$ as in (\ref{eq:y}), one could obtain: \begin{equation} T\left(X,Y\right)=\,^{3}T\left(X,Y\right)+\left(K\left(X,Y\right)-K\left(Y,X\right)\right)\hat{n},\label{eq:torsion-1} \end{equation} with: \[ \,^{3}T\left(X,Y\right)=\,^{3}\nabla_{X}Y-\,^{3}\nabla_{Y}X-\left[X,Y\right], \] is the torsion on the hypersurface $\Sigma$. Moreover, using (\ref{eq:z}), one could obtain: \[ K\left(X,Y\right)-K\left(Y,X\right)=\pm g\left(T\left(X,Y\right),\hat{n}\right), \] and hence (\ref{eq:torsion-1}) could be regarded as the torsion decomposition into the parallel and normal part of the hypersurface. The torsion tensor is antisymmetric in the first and second argument: \[ T\left(X,Y\right)=-T\left(Y,X\right),\qquad\forall\;X,T\,\in T_{p}\mathcal{M}. \] \subsubsection*{The Generalized (Riemann) Intrinsic Curvature} The original Riemann curvature is defined only for the curvature of a Levi-Civita connection in the context of Riemannian geometry. In this article, we use a similar definition to describe the intrinsic curvature of a general affine connection in non-Riemannian geometry: \begin{equation} \boldsymbol{R}\left(U,V\right)X=\nabla_{U}\nabla_{V}X-\nabla_{V}\nabla_{U}X-\nabla_{\left[U,V\right]}X,\qquad U,V,X\in T_{p}\mathcal{M}.\label{eq:Riemann} \end{equation} To obtain the intrinsic curvature on $\Sigma$, let $U,V,X\in T_{p}\Sigma$, then with the definitions in (\ref{eq:2}) and (\ref{eq:3}), one could obtain the double derivative of $X$: \begin{equation} \nabla_{U}\nabla_{V}X=\,^{3}\nabla_{U}\,^{3}\nabla_{V}X+K\left(V,X\right)\nabla_{U}\hat{n}+\left(U\left[K\left(V,X\right)\right]+K\left(U,\,^{3}\nabla_{V}X\right)\right)\hat{n},\label{eq:4} \end{equation} from a direct calculation. Notice here that $U\left[f\left(x\right)\right]$ denotes a vector $U$ acting on a scalar function $f\left(x\right)$ as $U^{\mu}\partial_{\mu}f\left(x\right)$. Moreover one could have the following quantity: \begin{align} \nabla_{V}\nabla_{U}X & =\,^{3}\nabla_{V}\,^{3}\nabla_{U}X+K\left(U,X\right)\nabla_{V}\hat{n}+\left(V\left[K\left(U,X\right)\right]+K\left(V,\,^{3}\nabla_{U}X\right)\right)\hat{n},\label{eq:5}\\ \nabla_{\left[U,V\right]}X & =\,^{3}\nabla_{\left[U,V\right]}X+K\left(\left[U,V\right],X\right)\hat{n},\label{eq:6} \end{align} using the fact that $\left[U,V\right]\in T_{p}\Sigma$ for $U,V\in T_{p}\Sigma$. By a little tensor algebra, one could show that the following relation is valid: \begin{align} K\left(U,\,^{3}\nabla_{V}X\right)-K\left(V,\,^{3}\nabla_{U}X\right)-K\left(\left[U,V\right],X\right)\label{eq:7}\\ +U\left[K\left(V,X\right)\right]-V\left[K\left(U,X\right)\right] & =\left(\,^{3}\nabla_{U}K\right)\left(V,X\right)-\left(\,^{3}\nabla_{V}K\right)\left(U,X\right)+K\left(\,^{3}T\left(U,V\right),X\right),\nonumber \end{align} with $\,^{3}T\left(U,V\right)$ is the torsion tensor on $\Sigma$. Inserting (\ref{eq:4}), (\ref{eq:5}), and (\ref{eq:6}) to (\ref{eq:Riemann}) and then using relation (\ref{eq:7}), one obtains the decomposition of intrinsic curvature on $\mathcal{M}$ into the intrinsic and extrinsic curvature of $\Sigma$: \begin{align} \boldsymbol{R}\left(U,V\right)X= & \,^{3}\boldsymbol{R}\left(U,V\right)X+K\left(V,X\right)\nabla_{U}\hat{n}-K\left(U,X\right)\nabla_{V}\hat{n}\label{eq:8}\\ & +\left(\left(\,^{3}\nabla_{U}K\right)\left(V,X\right)-\left(\,^{3}\nabla_{V}K\right)\left(U,X\right)+K\left(\,^{3}T\left(U,V\right),X\right)\right)\hat{n}.\nonumber \end{align} $\,^{3}\boldsymbol{R}$ is the generalized Riemann curvature of the connection $\,^{3}\nabla$ on the hypersurface $\Sigma$. The only symmetry possessed by the generalized Riemann curvature (\ref{eq:Riemann}) is: \[ \boldsymbol{R}\left(U,V\right)X=-\boldsymbol{R}\left(V,U\right)X. \] As a comment concerning the terminology, the definition in (\ref{eq:Riemann}) is defined originally for the Levi-Civita connection, and so does the term Riemann curvature tensor. However, in this article, we use a similar term and definition (i.e., (\ref{eq:Riemann})) for the intrinsic curvature of a generalized affine connection. From the decomposition of generalized Riemann tensor (\ref{eq:8}), the Gauss-Codazzi-Mainardi equations could be obtained. \subsubsection*{The Acceleration Tensor} In the Eulerian frame, one could regard the normal $\hat{n}$ as the 4-velocity of an event $p\in\mathcal{M}$ \cite{Eric}. Let us choose the spatial part $x^{i}$ of $x^{\mu}=\left(x^{0},x^{i}\right)$ as an 'adapted' local ($n$--1) coordinate on $\varSigma$ (notice that in general, the spatial part $x^{i}$ of $x^{\mu}=\left(x^{0},x^{i}\right)$ is not necessarily a local coordinate on $\varSigma$). The vector basis related to this coordinate is $\partial_{i}$. Together with $\hat{n}$, they define a complete basis on $T_{p}\mathcal{M},$ namely $\left(\hat{n},\partial_{i}\right)$. Let us define the acceleration tensor $\boldsymbol{a}$ in the coordinate basis $\left(\hat{n},\partial_{i}\right)$ as follows: \begin{equation} \boldsymbol{a}=\boldsymbol{a}_{\nu}=\left(\boldsymbol{a}_{\hat{n}},\boldsymbol{a}_{i}\right):=\left(\nabla_{\hat{n}}\hat{n},\nabla_{i}\hat{n}\right).\label{eq:accell} \end{equation} where: \[ \boldsymbol{a}_{\nu}=\boldsymbol{a}_{\nu}^{\mu}\partial_{\mu}\;\;\in\:T_{p}\mathcal{M},\qquad\nu=\left(\hat{n},i\right),\:i=1,2,..,n-1. \] Hence, for an arbitrary vector $U\in T_{p}\Sigma$, one has: \[ \nabla_{U}\hat{n}=U^{i}\nabla_{i}\hat{n}=U^{i}\boldsymbol{a}_{i}=\boldsymbol{a}\left(U\right). \] The quantity $\boldsymbol{a}_{\hat{n}}:=\alpha$ is known as the 4-acceleration. It is a well-known fact that in the original General Relativity with Levi-Civita connection, the 4-acceleration is orthogonal to the 4-velocity, namely $g\left(\alpha,\hat{n}\right)=0$. This is a consequence of the metric compatibility $\nabla_{X}g=0$, $\forall\:X\in T_{p}\mathcal{M}$. With the same condition, one could show that the acceleration tensor is orthogonal to the 4-velocity as well: \[ g\left(\boldsymbol{a}_{\nu},\hat{n}\right)=g\left(\nabla_{\nu}\hat{n},\hat{n}\right)=0,\qquad\nu=\left(\hat{n},i\right),\:i=1,2,..,n-1, \] using the fact that: \[ \underset{0}{\underbrace{\nabla_{\nu}g\left(\hat{n},\hat{n}\right)}}=\underset{0}{\underbrace{\left(\nabla_{\nu}g\right)}}\left(\hat{n},\hat{n}\right)+g\left(\nabla_{\nu}\hat{n},\hat{n}\right)+g\left(\hat{n},\nabla_{\nu}\hat{n}\right), \] and the symmetricity of $g$. Notice that if the metricity condition is violated, one has: \[ g\left(\boldsymbol{a}_{\nu},\hat{n}\right)=-\frac{1}{2}\left(\nabla_{\nu}g\right)\left(\hat{n},\hat{n}\right), \] which defines the 'angle' between $\boldsymbol{a}_{\nu}$ and $\hat{n}.$ Let us label the angle as: \begin{equation} \Theta_{\nu}=\pm g\left(\boldsymbol{a}_{\nu},\hat{n}\right).\label{eq:angle} \end{equation} For an arbitrary vector $U\in T_{p}\Sigma$, one could write $\Theta\left(U\right)=\pm g\left(\boldsymbol{a}\left(U\right),\hat{n}\right)$, that is, the angle between the acceleration in the spatial direction $U$, with the normal $\hat{n}.$ The definition of the angle $\Theta$ will be useful for the next sections. \subsubsection*{Extrinsic Curvature of the Second Kind} As we had mentioned in the previous sections, besides the extrinsic curvature defined in (\ref{eq:3}), there exists another definition of the extrinsic curvature \cite{Eric}: \begin{equation} \mathcal{K}\left(U,V\right)=g\left(\nabla_{U}\hat{n},V\right),\qquad U,V\in T_{p}\Sigma.\label{eq:extrinsic2} \end{equation} For the original General Relativity where the connection is Levi-Civita, the first and second extrinsic curvature coincide (by the factor of $\mp1$ for Riemannian/ Lorentzian case): \[ \underset{0}{\underbrace{\nabla_{U}g\left(V,\hat{n}\right)}}=\underset{0}{\underbrace{\left(\nabla_{U}g\right)}}\left(V,\hat{n}\right)+\underset{\pm K\left(U,V\right)}{\underbrace{g\left(\nabla_{U}V,\hat{n}\right)}}+\underset{\mathcal{K}\left(U,V\right)}{\underbrace{g\left(V,\nabla_{U}\hat{n}\right)}}. \] Notice that without the metric compatibility, they differ by $\nabla_{U}g\left(V,\hat{n}\right)$. Written in terms of components, one could show that the extrinsic curvature of the second kind satisfies: \begin{equation} \mathcal{K}\left(U,V\right)=g\left(\nabla_{U}\hat{n},V\right)=U^{\mu}V^{\nu}g\left(\nabla_{\mu}\hat{n},\partial_{\nu}\right)=\underset{0}{\underbrace{V_{\alpha}U^{\mu}\partial_{\mu}n^{\alpha}}}+U^{\mu}V_{\alpha}n^{\beta}\omega_{\mu\,\:\,\beta}^{\,\:\,\alpha},\label{eq:ex2index} \end{equation} where the first term is zero since $U^{\mu}\partial_{\mu}V_{\alpha}n^{\alpha}$ and $n^{\alpha}U^{\mu}\partial_{\mu}V_{\alpha}$ are zero for $U,V\in T_{p}\Sigma$. The geometrical interpretation of $\mathcal{K}\left(\partial_{\mu},\partial_{\nu}\right)$ differs from $K\left(\partial_{\mu},\partial_{\nu}\right)$: it describes the $\nu^{\mathrm{th}}$-component of the change of the normal in the direction of $\partial_{\mu}$. Before we proceed to the main calculation, it is convenient to fix the notations and conventions we use concerning the metric and the inner-product. First, the covector $V^{*}\in T_{p}^{*}\mathcal{M}$ is defined as follows: \begin{equation} g\left(V,W\right)=\left\langle V^{*},W\right\rangle ,\qquad\forall\;\;W\in T_{p}\mathcal{M},\label{eq:covector} \end{equation} with $g=g_{\mu\nu}dx^{\mu}\otimes dx^{\nu}$ is the metric on $\mathcal{M},$ $V\in T_{p}\mathcal{M}$, and $\left\langle \cdot,\cdot\right\rangle $ is an inner product on $T_{p}\mathcal{M}.$ With this definition, then (\ref{eq:extrinsic2}) could be written as $\mathcal{K}\left(U,V\right)=\left\langle V^{*},\nabla_{U}\hat{n}\right\rangle $, for example. Second, as we have mentioned at the beginning of this chapter, one could write (\ref{eq:covector}) as $g\left(V,\cdot\right)=g\left(\cdot,V\right)=V^{*}$. Moreover, given $V^{*},W^{*}\in T_{p}^{*}\mathcal{M},$ (\ref{eq:covector}) could be written as: \begin{equation} g\left(V,W\right)=g^{*}\left(V^{*},W^{*}\right),\label{eq:convention} \end{equation} with $g^{*}=g^{\mu\nu}\partial_{\mu}\otimes\partial_{\nu}$ is the dual of metric $g,$ such that $g\left(g^{*}\right)=\mathbb{I}$ (or in indices: $g^{\mu\sigma}g_{\sigma\nu}=\delta_{\nu}^{\mu}$). However, these relations should be used carefully, for example $g\left(\partial_{\mu},\partial_{\nu}\right)=g_{\mu\nu}$, is not equal to either $\left\langle dx^{\mu},\partial_{\nu}\right\rangle =\delta_{\nu}^{\mu}$ or $g^{*}\left(dx^{\mu},dx^{\nu}\right)=g^{\mu\nu}$. As another example, even if $g\left(V,\cdot\right)=g\left(\cdot,V\right)=V^{*}$, this is not valid for basis vectors and covectors, as $g\left(\partial_{\mu}\right)\neq dx^{\mu}$. The correct relation is: \begin{equation} dx^{\mu}=g^{\mu\nu}g\left(\partial_{\nu}\right).\label{eq:pentingx} \end{equation} \subsubsection*{The Torsion Decomposition, Gauss, and Codazzi-Mainardi Equations} The results (\ref{eq:torsion-1}) and (\ref{eq:8}) will be contracted with vectors $\hat{n}$ and $Y\in T_{p}\Sigma$ to obtain the temporal and spatial parts of the equations. As a first step, we need to decompose the quantity $\nabla_{U}\hat{n}$ into the normal and parallel parts. Let us consider the following quantity: \begin{equation} \nabla_{U}\left(g\left(X,\hat{n}\right)\right)=\left(\nabla_{U}g\right)\left(X,\hat{n}\right)+g\left(\nabla_{U}X,\hat{n}\right)+g\left(X,\nabla_{U}\hat{n}\right),\qquad X\in T_{p}\Sigma.\label{eq:b} \end{equation} The LHS of (\ref{eq:b}) could be written as $\nabla_{U}\left(g\left(X,\hat{n}\right)\right)=U\left[g\left(X,\hat{n}\right)\right]=0,$ since $X\in T_{p}\Sigma$ is perpendicular to $\hat{n}$. One could do the following decomposition on the quantity $\nabla_{U}\hat{n}$: \begin{equation} \nabla_{U}\hat{n}=\underset{\left(\nabla_{U}\hat{n}\right)_{\perp\Sigma}}{\underbrace{\pm\left\langle \hat{n}^{*},\nabla_{U}\hat{n}\right\rangle \hat{n}}}+\underset{\left(\nabla_{U}\hat{n}\right)_{\parallel\Sigma}}{\underbrace{\left\langle \,^{3}dx^{i},\nabla_{U}\hat{n}\right\rangle \partial_{i}}}.\label{eq:d} \end{equation} with $\partial_{i}$ and $\,^{3}dx^{i}=\,^{3}q^{ij}q\left(\partial_{j}\right)$ are the vector and covector basis corresponding to $x^{i},$ the adapted coordinate on $\varSigma$. With the definitions in (\ref{eq:angle}) and (\ref{eq:extrinsic2}), one could obtain the following relations: \begin{align*} \left(\nabla_{U}\hat{n}\right)_{\perp\Sigma} & =\pm\left\langle \hat{n}^{*},\nabla_{U}\hat{n}\right\rangle \hat{n}=\pm g\left(\nabla_{U}\hat{n},\hat{n}\right)\hat{n}=\Theta\left(U\right)\hat{n},\\ \left(\nabla_{U}\hat{n}\right)_{\parallel\Sigma} & =\left\langle \,^{3}dx^{i},\nabla_{U}\hat{n}\right\rangle \partial_{i}=\,^{3}q^{ij}g\left(\partial_{j},\nabla_{U}\hat{n}\right)\partial_{i}=\,^{3}q^{ij}\mathcal{K}\left(U,\partial_{j}\right)\partial_{i}, \end{align*} using equation (\ref{eq:1a}), (\ref{eq:extrinsic2}), and (\ref{eq:pentingx}). Inserting (\ref{eq:d}) to (\ref{eq:8}) gives: \begin{align} \boldsymbol{R}\left(U,V\right)X & =\,^{3}\boldsymbol{R}\left(U,V\right)X+\,^{3}q^{ij}\left(K\left(V,X\right)\mathcal{K}\left(U,\partial_{j}\right)-K\left(U,X\right)\mathcal{K}\left(V,\partial_{j}\right)\right)\partial_{i}\label{eq:e}\\ & +\left(\left(\,^{3}\nabla_{U}K\right)\left(V,X\right)-\left(\,^{3}\nabla_{V}K\right)\left(U,X\right)+K\left(\,^{3}T\left(U,V\right),X\right)+K\left(\Theta\left(U\right)V-\Theta\left(V\right)U,X\right)\right)\hat{n}.\nonumber \end{align} For the next step, by contracting (\ref{eq:e}) with $\hat{n}$, we obtain the generalized Codazzi-Mainardi equation: \begin{align} g\left(\hat{n},\boldsymbol{R}\left(U,V\right)X\right)= & \pm\left(\left(\,^{3}\nabla_{U}K\right)\left(V,X\right)-\left(\,^{3}\nabla_{V}K\right)\left(U,X\right)+K\left(\,^{3}T\left(U,V\right),X\right)+K\left(\Theta\left(U\right)V-\Theta\left(V\right)U,X\right)\right).\label{eq:CM} \end{align} Moreover, contracting (\ref{eq:e}) by $Y\in T_{p}\Sigma$ gives the generalized Gauss equation: \begin{eqnarray} g\left(Y,\boldsymbol{R}\left(U,V\right)X\right) & = & g\left(Y,\,^{3}\boldsymbol{R}\left(U,V\right)X\right)+K\left(V,X\right)\mathcal{K}\left(U,Y\right)-K\left(U,X\right)\mathcal{K}\left(V,Y\right),\label{eq:i} \end{eqnarray} where $Y$ could be written in the adapted coordinate as $Y=Y^{i}\partial_{i}$, with $g\left(Y^{j}\partial_{j},\partial_{i}\right)=\,^{3}Y_{i}$, and $Y^{*}=\,^{3}Y_{i}\,^{3}dx^{i}\in T_{p}^{*}\Sigma$. Equation (\ref{eq:torsion-1}) could be equally contracted with $\hat{n}$ and $Z\in T_{p}\Sigma$ to give the torsion decomposition relations: \begin{eqnarray} g\left(\hat{n},T\left(X,Y\right)\right) & = & \left(K\left(X,Y\right)-K\left(Y,X\right)\right),\label{eq:torsion1}\\ g\left(Z,T\left(X,Y\right)\right) & = & g\left(Z,\,^{3}T\left(X,Y\right)\right).\label{eq:torsion2} \end{eqnarray} These four equations are the main results in this article and will be heavily used in the next sections. Notice that if the connection is metric and torsionless, then $K\left(U,V\right)=\mp\mathcal{K}\left(U,V\right)$, $\Theta\left(U\right)=0$, and $T\left(U,V\right)=0,$ hence the generalized GCM returns to the standard original GCM. \section{Metric, Torsionless, and Levi-Civita Cases} In this section, we will discuss some special cases of the affine connection, in particular, metric, torsionless, and Levi-Civita connection. We will show that the conditions applied on the connection $\nabla$ on $\mathcal{M}$ are inherited to the connection $\,^{3}\nabla$ on the hypersurface $\Sigma$. To prevent any confusion, we need to clarify the conventions we use to write the notations indices of some geometrical objects. This is important because of the different conventions used in the literature. First, we write, using indices, the (vector potential of the) connection as $\nabla_{\mu}\partial_{\beta}=\omega_{\mu\,\:\,\beta}^{\,\:\,\alpha}\partial_{\alpha}$, the torsion tensor as $\boldsymbol{T}\left(\partial_{\mu},\partial_{\beta}\right)=T_{\mu\,\:\,\beta}^{\:\:\alpha}\partial_{\alpha}$, and the generalized Riemann tensor as $\boldsymbol{R}\left(\partial_{\mu},\partial_{\nu}\right)\partial_{\beta}=R_{\mu\nu\,\:\,\,\beta}^{\:\:\:\:\,\,\alpha}\partial_{\alpha}.$ These conventions are different from the ones used in, for examples \cite{Sotiriou,Olmo}. \subsubsection*{Metric Compatibility} A metric connection is a connection that satisfies the metric compatibility condition: \begin{equation} \nabla_{X}g=X^{\mu}\nabla_{\mu}\left(g_{\alpha\beta}dx^{\alpha}dx^{\beta}\right)=0,\qquad\forall\,X\,\in T_{p}\mathcal{M}.\label{eq:metric} \end{equation} The condition could be written in components, with abuse of notation as follows: \begin{equation} \nabla_{\mu}g_{\alpha\beta}=\partial_{\mu}g_{\alpha\beta}-\omega_{\mu\beta\alpha}-\omega_{\mu\alpha\beta}=0.\label{eq:z00} \end{equation} It describes the failure of the last two indices of $\omega$ (borrowing the term from \cite{Miller}, the 'rotation' bivector indices) to be antisymmetric, differing with the quantity $\partial_{\mu}g_{\alpha\beta}$. In the next paragraph, we will show that the metric compatibility of a connection in $\mathcal{M}$ will induce also a metric connection in $\Sigma$. Let us remember that the metric $g$ in $\mathcal{M}$ could be decomposed into the perpendicular and parallel parts with respect to the embedded hypersurface $\Sigma,$ as in (\ref{eq:1a}): \begin{equation} g=\underset{\,^{3}q}{\underbrace{g_{\parallel}}}\pm\underset{\hat{n}^{*}\otimes\hat{n}^{*}}{\underbrace{g_{\perp}}}.\label{eq:z-0} \end{equation} With this, we could obtain the following relation: \begin{equation} \nabla_{X}g=\underset{\nabla_{X}\,^{3}q}{\underbrace{\nabla_{X}g_{\parallel}}}\pm\nabla_{X}g_{\perp}=0,\qquad X\in T_{p}\Sigma,\label{eq:z-1} \end{equation} from the metric compatibility. On the other hand, we could directly decompose $\nabla_{X}g$ as: \[ \nabla_{X}g=\left(\nabla_{X}g\right)_{\parallel}\pm\left(\nabla_{X}g\right)_{\perp}=0,\quad\implies\quad\left(\nabla_{X}g\right)_{\parallel}=0,\quad\left(\nabla_{X}g\right)_{\perp}=0. \] Notice that in general, $\left(\nabla_{X}g\right)_{\parallel}\neq\nabla_{X}g_{\parallel}$ and $\left(\nabla_{X}g\right)_{\perp}\neq\nabla_{X}g_{\perp}$. Now let us consider the following quantity: \begin{equation} \nabla_{X}\,^{3}q=\left(\nabla_{X}\,^{3}q\right)_{\parallel}\pm\left(\nabla_{X}\,^{3}q\right)_{\perp}.\label{eq:z-2} \end{equation} Since $X\in T_{p}\varSigma$ and $\left(\nabla_{X}\,^{3}q\right)_{\parallel}$ live in the hypersurface $\Sigma$, we could define: \begin{equation} \left(\nabla_{X}\,^{3}q\right)_{\parallel}:=\,^{3}\nabla_{X}\,^{3}q,\label{eq:z-3a} \end{equation} with $\,^{3}\nabla$ is the induced connection on hypersurface $\Sigma$. With the definition (\ref{eq:z-3a}) and (\ref{eq:z-1}), we obtain: \[ \,^{3}\nabla_{X}\,^{3}q=\mp\nabla_{X}g_{\perp}\pm\left(\nabla_{X}g_{\perp}\right)_{\perp}. \] Decomposing $\nabla_{X}g_{\perp}$ in its parallel and perpendicular part (and also using (\ref{eq:z-1})), we have: \[ \,^{3}\nabla_{X}\,^{3}q=\mp\left(\nabla_{X}g_{\perp}\right)_{\parallel}. \] Using (\ref{eq:z-0}), then one obtains: \begin{equation} \,^{3}\nabla_{X}\,^{3}q=\mp\left(\nabla_{X}\hat{n}^{*}\otimes\hat{n}^{*}\right)_{\parallel}=\mp\left(\left(\nabla_{X}\hat{n}^{*}\right)\otimes\hat{n}^{*}+\hat{n}^{*}\otimes\nabla_{X}\hat{n}^{*}\right)_{\parallel}.\label{eq:135} \end{equation} Notice that the RHS of (\ref{eq:135}) could be written as a projection of such quantity to $\Sigma$ using a projection operator $p_{\Sigma}=g^{\mu\lambda}\,^{3}q_{\lambda\nu}\partial_{\mu}\otimes dx^{\nu}$, namely: \[ \left(\left(\nabla_{X}\hat{n}^{*}\right)\otimes\hat{n}^{*}+\hat{n}^{*}\otimes\nabla_{X}\hat{n}^{*}\right)_{\parallel}=p_{\Sigma}\cdot\left(\left(\nabla_{X}\hat{n}^{*}\right)\otimes\hat{n}^{*}+\hat{n}^{*}\otimes\nabla_{X}\hat{n}^{*}\right)\cdot p_{\Sigma}^{T}, \] with $\cdot$ denotes the inner product (matrix multiplication) between matrices. One could write: \begin{align*} p_{\Sigma}\cdot\left(\hat{n}^{*}\otimes\nabla_{X}\hat{n}^{*}\right) & =p_{\Sigma}\left(\hat{n}^{*}\right)\otimes\nabla_{X}\hat{n}^{*},\\ \left(\left(\nabla_{X}\hat{n}^{*}\right)\otimes\hat{n}^{*}\right)\cdot p_{\Sigma}^{T} & =\left(\nabla_{X}\hat{n}^{*}\right)\otimes p_{\Sigma}\left(\hat{n}^{*}\right). \end{align*} Since $p_{\Sigma}\left(\hat{n}^{*}\right)=0$ ($\,^{3}q$ lives in $\Sigma$ and $\hat{n}$ is perpendicular to $\Sigma$ ), we have: \[ \,^{3}\nabla_{X}\,^{3}q=0. \] Therefore, we have shown that the metric compatibility of a connection $\nabla$ on $\left(\mathcal{M},g\right),$ induces a connection $\,^{3}\nabla$ on the hypersurface $\left(\Sigma,\,^{3}q\right),$ which also satisfies the metric compatibility with respect to $\,^{3}q$. \subsubsection*{Torsionless Condition} Let us perform a similar derivation for a special condition on the torsion tensor. The torsionless condition is a case where the connection has zero torsion: \begin{equation} T\left(X,Y\right)=0,\qquad\forall\,X,Y\,\in T_{p}\mathcal{M}.\label{eq:torsionles} \end{equation} With the definition of torsion in (\ref{eq:torsion}), a direct consequence is: \begin{equation} T\left(\partial_{\alpha},\partial_{\beta}\right)=\underset{T_{\alpha\:\:\beta}^{\,\:\,\sigma}}{\underbrace{\left(\omega_{\alpha\,\:\beta}^{\,\:\:\sigma}-\omega_{\beta\,\:\alpha}^{\:\:\:\sigma}\right)}}\partial_{\sigma}=0,\qquad\omega_{\alpha\,\:\beta}^{\,\:\:\sigma}=\omega_{\beta\,\:\alpha}^{\:\:\:\sigma},\label{eq:z-3} \end{equation} which results in the symmetricity of the first and third index of the connection. Notice here that we do not use factor $2$ on the definition of torsion, as in \cite{Sotiriou,Olmo}. Using the torsion decomposition (\ref{eq:torsion-1}), the torsionless condition becomes: \[ \,^{3}T\left(X,Y\right)+\left(K\left(X,Y\right)-K\left(Y,X\right)\right)\hat{n}=0. \] From the definition of the first extrinsic curvature (\ref{eq:z}), and then using (\ref{eq:z-3}), we have: \[ \,^{3}T\left(X,Y\right)=0. \] Hence, the torsionless connection $\nabla$ on $\mathcal{M}$ induced a torsionless connection $\,^{3}\nabla$ on $\Sigma$. \subsubsection*{Levi-Civita Connection} Since the metric compatibility and torsionless condition are independent of one another, one could define a connection that satisfies both of the conditions. This connection is unique and known as the Levi-Civita connection. A metric compatible and torsionless connection $\nabla$ on $\left(\mathcal{M},g\right)$ induces a connection $\,^{3}\nabla$ on $\left(\Sigma,\,^{3}q\right)$ which is also metric (with respect to $\,^{3}q$) and torsionless. The Levi-Civita connection could be derived as follows. With the metric compatibility condition (\ref{eq:z00}), one considers a certain combination of the vector potential satisfies equality as follows: \begin{equation} \omega_{\mu\alpha\beta}+\omega_{\mu\beta\alpha}-\omega_{\alpha\beta\mu}-\omega_{\alpha\mu\beta}+\omega_{\beta\mu\alpha}+\omega_{\beta\alpha\mu}=\partial_{\mu}g_{\alpha\beta}-\partial_{\alpha}g_{\beta\mu}+\partial_{\beta}g_{\mu\alpha}.\label{eq:quantity} \end{equation} Using the torsionless condition (\ref{eq:z-3}) on the LHS, one could obtain, and then define the (vector potential of the) Levi-Civita connection: \begin{equation} \omega_{\mu\alpha\beta}:=\Gamma_{\mu\alpha\beta}=\frac{1}{2}\left(\partial_{\mu}g_{\alpha\beta}-\partial_{\alpha}g_{\beta\mu}+\partial_{\beta}g_{\mu\alpha}\right),\label{eq:LevCiv} \end{equation} also known as the Christoffel symbol. It is easy to show that the Levi-Civita connection is symmetric on the first and third indices, namely $\Gamma_{\mu\alpha\beta}=\Gamma_{\beta\alpha\mu}.$ \subsubsection*{The Symmetric and Antisymmetric Parts of Connections, Torsion, and Curvatures} A general affine connection could be decomposed into its Levi-Civita, torsion, and non-metricity part. The derivation could be found in classic literature, for example, in \cite{Schouten}. Let us consider the previous combinations of $\omega$ as in (\ref{eq:quantity}), but now without the metric compatibility. With (\ref{eq:metric}), one has: \[ \omega_{\mu\alpha\beta}+\omega_{\mu\beta\alpha}-\omega_{\alpha\beta\mu}-\omega_{\alpha\mu\beta}+\omega_{\beta\mu\alpha}+\omega_{\beta\alpha\mu}=\partial_{\mu}g_{\alpha\beta}-\nabla_{\mu}g_{\alpha\beta}-\partial_{\alpha}g_{\beta\mu}+\nabla_{\alpha}g_{\beta\mu}+\partial_{\beta}g_{\mu\alpha}-\nabla_{\beta}g_{\mu\alpha}. \] Notice here that we do not use any requirement for the equation. Using (\ref{eq:torsion}) to rewrite the LHS, one obtains: \[ 2\omega_{\mu\alpha\beta}+T_{\mu\beta\alpha}+T_{\beta\mu\alpha}+T_{\beta\alpha\mu}=\partial_{\mu}g_{\alpha\beta}-\partial_{\alpha}g_{\beta\mu}+\partial_{\beta}g_{\mu\alpha}-\left(\nabla_{\mu}g_{\alpha\beta}-\nabla_{\alpha}g_{\beta\mu}+\nabla_{\beta}g_{\mu\alpha}\right), \] that could be written as: \begin{equation} \omega_{\mu\alpha\beta}=\Gamma_{\mu\alpha\beta}+C_{\mu\alpha\beta}+Q_{\mu\alpha\beta},\label{eq:Decomposition} \end{equation} with the Levi-Civita connection $\Gamma$ satisfies (\ref{eq:LevCiv}), and: \begin{align} Q_{\mu\alpha\beta} & =-\frac{1}{2}\left(\nabla_{\mu}g_{\alpha\beta}-\nabla_{\alpha}g_{\beta\mu}+\nabla_{\beta}g_{\mu\alpha}\right),\label{eq:metricpart}\\ C_{\mu\alpha\beta} & =\frac{1}{2}\left(T_{\mu\alpha\beta}+T_{\alpha\beta\mu}-T_{\beta\mu\alpha}\right).\label{eq:connectionpart} \end{align} $Q$ and $C$ are the non-metricity and the torsion parts, usually known respectively, as the disformation/deflection and the contorsion tensor. If the disformation $Q$ is zero, the connection is metric, and if the contorsion $C$ is zero, the connection is torsionless. If both of them are zero, the connection is Levi-Civita. Let $\left(\alpha\right|...\left|\beta\right)$ and $\left[\alpha\right|...\left|\beta\right]$ denote, respectively, the symmetric and antisymmetric part of an arbitrary tensor with indices $\alpha$ and $\beta$. With these notations, one could write the symmetries of $\Gamma$, $Q$ and $C$ as follows: \begin{center} \begin{tabular}{lclcl} $\Gamma_{\left(\mu\alpha\right)\beta}=\frac{1}{2}\partial_{\beta}g_{\mu\alpha},$ & $\qquad$ & $Q_{\left(\mu\alpha\right)\beta}=-\frac{1}{2}\nabla_{\beta}g_{\mu\alpha},$ & $\qquad$ & $C_{\left(\mu\alpha\right)\beta}=\frac{1}{2}\left(T_{\mu\alpha\beta}-T_{\beta\mu\alpha}\right)$\tabularnewline $\Gamma_{\left[\mu\alpha\right]\beta}=\partial_{\left[\mu\right.}g_{\left.\alpha\right]\beta},$ & & $Q_{\left[\mu\alpha\right]\beta}=-\nabla_{\left[\mu\right.}g_{\left.\alpha\right]\beta},$ & & $C_{\left[\mu\alpha\right]\beta}=\frac{1}{2}T_{\alpha\beta\mu},$\tabularnewline $\Gamma_{\left(\mu\right|\alpha\left|\beta\right)}=\Gamma_{\mu\alpha\beta},$ & & $Q_{\left(\mu\right|\alpha\left|\beta\right)}=Q_{\mu\alpha\beta},$ & & $C_{\left(\mu\right|\alpha\left|\beta\right)}=\frac{1}{2}\left(T_{\alpha\beta\mu}-T_{\beta\mu\alpha}\right),$\tabularnewline $\Gamma_{\left[\mu\right|\alpha\left|\beta\right]}=0,$ & & $Q_{\left[\mu\right|\alpha\left|\beta\right]}=0,$ & & $C_{\left[\mu\right|\alpha\left|\beta\right]}=\frac{1}{2}T_{\mu\alpha\beta},$\tabularnewline $\Gamma_{\mu\left(\alpha\beta\right)}=\frac{1}{2}\partial_{\mu}g_{\alpha\beta},$ & & $Q_{\mu\left(\alpha\beta\right)}=-\frac{1}{2}\nabla_{\mu}g_{\alpha\beta},$ & & $C_{\mu\left(\alpha\beta\right)}=0,$\tabularnewline $\Gamma_{\mu\left[\alpha\beta\right]}=-\partial_{\left[\alpha\right.}g_{\left.\beta\right]\mu},$ & & $Q_{\mu\left[\alpha\beta\right]}=\nabla_{\left[\alpha\right.}g_{\left.\beta\right]\mu},$ & & $C_{\mu\left[\alpha\beta\right]}=C_{\mu\alpha\beta}.$\tabularnewline \end{tabular} \par\end{center} From these relations, it is clear that $\Gamma$ and $Q$ have similar symmetries, in particular, they are symmetric in the first and third indices. On the other hand, the contorsion $C$ is antisymmetric in the second and third indices. One should keep in mind that other authors use different conventions for the order of the indices, in particular \cite{Sotiriou,Olmo}, but these do not alter the physical symmetries of the connection. With the decomposition of connection (\ref{eq:Decomposition}), one could obtain directly the decomposition of extrinsic curvature of the first kind: \[ K\left(X,Y\right)=\underset{K_{\Gamma}\left(X,Y\right)}{\underbrace{\pm X^{\mu}Y^{\beta}\Gamma_{\mu\,\:\beta}^{\:\:\alpha}n_{\alpha}}}+\underset{K_{Q}\left(X,Y\right)}{\underbrace{\pm X^{\mu}Y^{\beta}Q_{\mu\,\:\,\beta}^{\:\:\,\alpha}n_{\alpha}}}+\underset{K_{C}\left(X,Y\right)}{\underbrace{\pm X^{\mu}Y^{\beta}C_{\mu\,\:\,\beta}^{\:\:\,\alpha}n_{\alpha}}}, \] where $K_{\Gamma}$ and $K_{Q}$ are symmetric on the first and second argument, hence, the torsionless condition $C=0$, gives a symmetric extrinsic curvature. On the other hand, the second extrinsic curvature satisfies the following decomposition: \[ \mathcal{K}\left(X,Y\right)=\underset{\mathcal{K}_{\Gamma}\left(X,Y\right)}{\underbrace{X^{\mu}Y_{\alpha}n^{\beta}\Gamma_{\mu\,\:\,\beta}^{\,\:\,\alpha}}}+\underset{\mathcal{K}_{Q}\left(X,Y\right)}{\underbrace{X^{\mu}Y_{\alpha}n^{\beta}Q_{\mu\,\:\,\beta}^{\,\:\,\alpha}}}+\underset{\mathcal{K}_{C}\left(X,Y\right)}{\underbrace{X^{\mu}Y_{\alpha}n^{\beta}C_{\mu\,\:\,\beta}^{\,\:\,\alpha}}}. \] Notice that now $\mathcal{K}_{\Gamma}$ and $\mathcal{K}_{Q}$ are \textit{not} symmetric on the first and second argument. However, for a metric connection, the extrinsic curvatures could be written as: \[ K\left(X,Y\right)=\mp\left(\underset{0}{\underbrace{\left(\nabla_{X}g\right)}}\left(Y,n\right)+g\left(Y,\nabla_{X}n\right)\right)=\mp\mathcal{K}\left(X,Y\right), \] which comes from relation (\ref{eq:b}). Hence, the extrinsic curvature of the first kind coincides with the second kind. The torsion could be written as: \begin{equation} T\left(X,Y\right)=\underset{0}{\underbrace{T_{\Gamma}\left(X,Y\right)}}+T_{C}\left(X,Y\right)+\underset{0}{\underbrace{T_{Q}\left(X,Y\right)}}=X^{\mu}Y^{\beta}\left(C_{\mu\,\:\beta}^{\:\:\alpha}-C_{\beta\,\:\mu}^{\:\:\alpha}\right)\partial_{\alpha},\label{eq:Decomposition3} \end{equation} from the symmetries of the parts of the affine connection. The Riemann curvature (\ref{eq:Riemann}) could be rewritten in indices as follows: \begin{align} \boldsymbol{R}\left(\partial_{\mu},\partial_{\mu}\right)\partial_{\beta}= & \nabla_{\mu}\nabla_{\nu}\partial_{\beta}-\nabla_{\nu}\nabla_{\mu}\partial_{\beta}=\underset{R_{\mu\nu\,\:\,\beta}^{\:\:\:\:\,\,\alpha}}{\underbrace{\left(\partial_{\mu}\omega_{\nu\,\:\beta}^{\:\:\alpha}-\partial_{\nu}\omega_{\mu\,\:\beta}^{\:\:\alpha}+\omega_{\mu\,\:\sigma}^{\:\:\alpha}\omega_{\nu\,\:\beta}^{\:\:\sigma}-\omega_{\nu\,\:\sigma}^{\:\:\alpha}\omega_{\mu\,\:\beta}^{\:\:\sigma}\right)}\partial_{\alpha}.}\label{eq:Rindices} \end{align} One could insert (\ref{eq:Decomposition}) to (\ref{eq:Rindices}), to obtain the decomposition of the Riemann tensor in terms of $\Gamma$, $Q$, and $C$, but it is not necessary for the derivation in this article. Unlike the first extrinsic curvature and the torsion tensor, the Riemann tensor could not be cleanly decomposed because of the existence of the non-linear terms in (\ref{eq:Rindices}). It is convenient to lower all the indices of the Riemann tensor to obtain clearer symmetries: \begin{equation} R_{\mu\nu\alpha\beta}=\partial_{\mu}\omega_{\nu\alpha\beta}-\partial_{\nu}\omega_{\mu\alpha\beta}+\omega_{\mu\,\:\beta}^{\,\:\:\lambda}\omega_{\nu\lambda\alpha}-\omega_{\nu\,\:\beta}^{\,\:\:\lambda}\omega_{\mu\lambda\alpha}+\omega_{\mu\,\:\beta}^{\,\:\:\lambda}\nabla_{\nu}g_{\lambda\alpha}-\omega_{\nu\,\:\beta}^{\,\:\:\lambda}\nabla_{\mu}g_{\lambda\alpha}.\label{eq:b-2} \end{equation} Notice that the generalized Riemann tensor (\ref{eq:Rindices}) only satisfies the antisymmetricity on the first and second index, namely, on the loop orientation indices \cite{Miller}: $R_{\mu\nu\alpha\beta}=-R_{\nu\mu\alpha\beta}.$ Only if the connection is metric; $Q=0$, one could show that the curvature satisfies an additional symmetry, namely $R_{\mu\nu\alpha\beta}=-R_{\mu\nu\beta\alpha}.$ Moreover, if the metric connection is also torsionless, it is symmetric by the switching between the loop orientation indices $\left(\mu,\nu\right)$ and the 'rotation' bivector indices $\left(\alpha,\beta\right)$, thus recovering the usual symmetries of the Riemannian tensor of the Levi-Civita connection: $R_{\mu\nu\alpha\beta}=R_{\alpha\beta\mu\nu},$ where the loop orientation and the rotation bivector indices are symmetric. \subsubsection*{Bianchi Identities} The generalized Riemann tensor satisfies the second Bianchi Identity: For any vector $X,Y,Z\in T_{p}\mathcal{M},$ the second Bianchi Identity is: \begin{equation} \left(\nabla_{X}\boldsymbol{R}\right)\left(Y,Z\right)+\left(\nabla_{Y}\boldsymbol{R}\right)\left(Z,X\right)+\left(\nabla_{Z}\boldsymbol{R}\right)\left(X,Y\right)=0.\label{eq:Bianchi2} \end{equation} It is a consequence of the orthogonality of $\omega$ and $\boldsymbol{R}$, namely, $\mathrm{d}_{\nabla}\boldsymbol{R}=\mathrm{d}_{\nabla}^{2}\omega=\boldsymbol{R}\wedge\omega=0$, with $\mathrm{d}_{\nabla}$ is the exterior covariant derivative of the connection $\nabla$, see \cite{Baez}. The Bianchi Identity (\ref{eq:Bianchi2}) could be written in terms of torsion as follows: \begin{align*} \left(\left(\nabla_{X}\boldsymbol{R}\right)\left(Y,Z\right)+\left(\nabla_{Y}\boldsymbol{R}\right)\left(Z,X\right)+\left(\nabla_{Z}\boldsymbol{R}\right)\left(X,Y\right)\right)= & \quad\nabla_{\left[X,\left[Y,Z\right]\right]+\left[Y,\left[Z,X\right]\right]+\left[Z,\left[X,Y\right]\right]}W\\ & +\left(\boldsymbol{R}\left(X,T\left(Y,Z\right)\right)+\boldsymbol{R}\left(Y,T\left(Z,X\right)\right)+\boldsymbol{R}\left(Z,T\left(X,Y\right)\right)\right)W. \end{align*} The first term contains the Jacobi identity, which is zero: $\left[X,\left[Y,Z\right]\right]+\left[Y,\left[Z,X\right]\right]+\left[Z,\left[X,Y\right]\right]=0.$ Hence, we have the second Bianchi Identity in terms of torsion: \[ \boldsymbol{R}\left(X,T\left(Y,Z\right)\right)+\boldsymbol{R}\left(Y,T\left(Z,X\right)\right)+\boldsymbol{R}\left(Z,T\left(X,Y\right)\right)=0. \] A torsionless connection satisfies another similar identity, the first Bianchi Identity. One could show that the following relation is true: \begin{align*} R\left(X,Y\right)Z+R\left(Y,Z\right)X+R\left(Z,X\right)Y= & \nabla_{X}T\left(Y,Z\right)+\nabla_{Y}T\left(Z,X\right)+\nabla_{Z}T\left(X,Y\right)+T\left(X,\left[Y,Z\right]\right)\\ & +T\left(Y,\left[Z,X\right]\right)+T\left(Z,\left[X,Y\right]\right)-\left[X,\left[Y,Z\right]\right]-\left[Y,\left[Z,X\right]\right]-\left[Z,\left[X,Y\right]\right]. \end{align*} With the Jacobi Identity and the torsionless condition, one obtain the first Bianchi Identity: \[ R\left(X,Y\right)Z+R\left(Y,Z\right)X+R\left(Z,X\right)Y=0. \] \subsubsection*{Contractions of the Generalized Riemann Tensor} To understand the significance of Gauss-Codazzi equations for General Relativity, it is convenient to introduce the contractions of Riemann tensors, the Ricci tensor $\mathbf{Ric}$ and Ricci scalar $\mathcal{R}$ as follows: \begin{align} \mathbf{Ric} & =\left\langle dx^{\mu},\boldsymbol{R}\left(\partial_{\mu},\partial_{\nu}\right)\partial_{\beta}\right\rangle dx^{\nu}\otimes dx^{\beta}=\underset{R_{\nu\beta}}{\underbrace{\delta_{\alpha}^{\mu}R_{\mu\nu\,\:\,\beta}^{\:\:\:\:\,\,\alpha}}}dx^{\nu}\otimes dx^{\beta},\label{eq:ricci}\\ \mathcal{R} & =\mathrm{tr}\,\mathbf{Ric}=g^{\nu\beta}R_{\nu\beta}.\label{eq:scalar} \end{align} The lack of symmetries of the generalized Riemann tensor results in the ambiguities in defining the Ricci tensor. Contracting the index $\nu$ with $\alpha$ of the Riemann tensor $R_{\mu\nu\,\:\,\beta}^{\:\:\:\:\:\alpha}$ will give the minus of (\ref{eq:ricci}), but contracting either $\mu,$ or $\nu$ with $\beta$ instead of $\alpha$ will give a different quantity due to the lack of symmetry in the indices $\left(\alpha,\beta\right)$. This quantity is defined as the co-Ricci tensor \cite{Jimenez2}: \begin{equation} \mathbf{\overline{Ric}}=\underset{R_{\nu\alpha}}{\underbrace{g^{\mu\beta}R_{\mu\nu\alpha\beta}}}dx^{\nu}\otimes dx^{\alpha}.\label{eq:coRic} \end{equation} Another type of contraction also exists, known as the homothetic curvature \cite{Iosifidis3,Jimenez2,Iosifidis4}: \[ \mathrm{tr}\boldsymbol{R}\left(\partial_{\mu},\partial_{\nu}\right)=\delta_{\alpha}^{\beta}R_{\mu\nu\,\:\,\beta}^{\:\:\:\:\:\alpha}, \] this, again, is not zero as in the Levi-Civita case, due to the non-metricity nature of the general affine connection. Moreover, the Ricci tensor is no longer symmetric in the indices $\left(\nu,\beta\right)$. This will cause problem in constructing the Einstein tensor $\boldsymbol{G}$, that needs to be symmetric since it is proportional to the stress-energy tensor $\mathcal{T}$ via the Einstein-Field Equation $\boldsymbol{G}=\kappa\mathcal{T}.$ However, in Metric-Affine and $f\left(\mathcal{R}\right)$-gravity, only the symmetric part of the Ricci tensor enters the equation of motion \cite{Sotiriou}. \section{Application to Gravity} \subsection{The Modified Theories of Gravity} \subsubsection*{Second Order Formulation: The Standard GR and $f\left(\mathcal{R}\right)$-Gravity} The original derivation of Einstein Field Equation (EFE) from a variational principle was first considered by Einstein and Hilbert, where the action is $S_{EH}\left[g\right]=\mathcal{R}\textrm{vol}+S_{\mathrm{matter}}\left[g\right]$, as a functional of the metric $g.$ The connection is assumed \textit{a priori} to be Levi-Civita and therefore is a function of the metric. Minimizing the action with respect to the variation of $g$ gives the standard EFE. This is known as the second-order formulation of gravity. $f\left(\mathcal{R}\right)$-gravity is one of the modification of General Relativity via a more general choice of action: \begin{equation} S\left[g\right]=\intop_{\mathcal{M}}f\left(\mathcal{R}\right)\mathrm{vol}+S_{\mathrm{matter}}\left[g\right],\label{eq:fR} \end{equation} where $f\left(\mathcal{R}\right)$ is the power series function of Ricci scalar with the form as follows \cite{Oikonomou,Sotiriou}: \[ f\left(\mathcal{R}\right)=..+\frac{c_{2}}{\mathcal{R}^{2}}+\frac{c_{1}}{\mathcal{R}}-2\Lambda+\mathcal{R}+\frac{\mathcal{R}^{2}}{k_{2}}+\frac{\mathcal{R}^{3}}{k_{3}}+... \] Minimizing the action (\ref{eq:fR}) with respect to the metric gives the $f\left(\mathcal{R}\right)$-EFE as follows: \begin{equation} f'\left(\mathcal{R}\right)R_{\left(\alpha\beta\right)}-\frac{1}{2}f\left(\mathcal{R}\right)g_{\alpha\beta}=\kappa\mathcal{T}_{\alpha\beta}.\label{eq:EL1} \end{equation} $\mathcal{T}_{\alpha\beta}$ is the stress-energy-momentum tensor, obtained from the variation of $g$ on $S_{\mathrm{matter}}$: \begin{equation} \delta_{g}S_{\mathrm{matter}}\left[g\right]=-\intop_{\mathcal{M}}\kappa\mathcal{T}_{\alpha\beta}\delta g^{\alpha\beta}\mathrm{vol}.\label{eq:tiga} \end{equation} Standard GR is a special case of $f\left(\mathcal{R}\right)$-gravity, where $f\left(\mathcal{R}\right)=\mathcal{R}$. \subsubsection*{First Order Formulation: The Palatini Formalism} In a more general approach known as the Palatini formalism (or the first-order formulation of gravity), the connection is not assumed to be Levi-Civita, hence it is independent of the metric on $\mathcal{M}$ \cite{Olmo}. The Einstein-Hilbert action is now a functional over the metric $g$ and connection $\omega$: \begin{equation} S_{EH}\left[g,\omega\right]=\intop_{\mathcal{M}}\mathcal{R}\mathrm{vol}+S_{\mathrm{matter}}\left[g\right].\label{eq:Palatini} \end{equation} $\omega$ is a general affine connection defined in (\ref{eq:Decomposition}). Minimizing the action with respect to the variation of $g$ gives an equation of motion similar to the EFE: \begin{equation} \underset{G_{\alpha\beta}}{\underbrace{R_{\left(\alpha\beta\right)}-\frac{1}{2}\mathcal{R}g_{\alpha\beta}}}=\kappa\mathcal{T}_{\alpha\beta},\label{eq:EL1-1} \end{equation} but with the torsional and non-metricity contributions inside $R_{\left[\alpha\beta\right]}$ and $\mathcal{R}$. Minimizing the action (\ref{eq:Palatini}) with respect to the connection $\omega$ results in another equation of motion: \begin{equation} \underset{P_{\mu}^{\:\:\alpha\beta}}{\underbrace{\left(\underset{A_{\nu}}{\underbrace{T_{\lambda\,\:\nu}^{\:\:\lambda}-\frac{1}{2}g^{\sigma\lambda}\nabla_{\nu}g_{\sigma\lambda}}}-\nabla_{\nu}\right)\left(2\delta_{\mu}^{\:\left[\nu\right.}g^{\left.\alpha\right]\beta}\right)+T_{\mu}^{\:\:\alpha\beta}}}=0.\label{eq:EL2} \end{equation} $G_{\alpha\beta}$ and $P_{\mu}^{\:\:\alpha\beta}$ are known, respectively, as the Einstein and Palatini tensor \cite{Iosifidis4}. Although (\ref{eq:EL2}) is different from the result found in the literature \cite{Sotiriou,Iosifidis4,Olmo}, it could be shown that they are equal, up to the factor 2 in the third term. This is due to the difference in the definition of torsion in (\ref{eq:torsion}) which does not include the factor $\frac{1}{2},$ as defined in literatures, in particular \cite{Sotiriou,Iosifidis4,Olmo}. As clearly derived in \cite{Iosifidis4}, the solution to (\ref{eq:EL2}) contains unspecified vector degrees of freedom on the connection. However, since the Einstein-Hilbert action (\ref{eq:Palatini}) is invariant under a projective transformation, these vector degrees of freedom could be eliminated, which results in a condition that the connection must be metric compatible and torsionless, hence Levi-Civita \cite{Sotiriou,Olmo}. This will be discussed in the next subsection. With the Levi-Civita connection, (\ref{eq:EL1-1}) returns to the original EFE. Therefore, by considering the projective transformation, the Palatini formalism gives a similar dynamics with the second-order formulation \cite{Iosifidis4}. \subsubsection*{The Projective Invariance Problem} The projective transformation is defined as follows \cite{Sotiriou,Iosifidis,Iosifidis4}: \begin{equation} \omega_{\mu\,\:\beta}^{\:\:\alpha}\rightarrow\omega_{\mu\,\:\beta}^{\:\:\alpha}+\delta_{\beta}^{\alpha}\xi_{\mu}.\label{eq:projectivetranfs} \end{equation} By a direct calculation, one could show that the Ricci scalar $\mathcal{R}$ is invariant under (\ref{eq:projectivetranfs}). This property could be utilized to obtain the Levi-Civita condition from (\ref{eq:EL2}) as follows. First, one could show that the Palatini tensor is traceless: $P_{\mu}^{\:\:\alpha\mu}=0$ \cite{Iosifidis,Iosifidis4}. Therefore, in 4-dimension, one could only obtain 60 independent equations from (\ref{eq:EL2}). These equations are not enough to determine completely the connection, as it will need 64 independent equations. Instead, the maximal condition on the connection one could obtain from (\ref{eq:EL2}) is: \begin{equation} \omega_{\mu\,\:\beta}^{\:\:\alpha}=\Gamma_{\mu\,\:\beta}^{\:\:\alpha}-\frac{1}{(n-1)}\underset{T_{\mu}}{\underbrace{T_{\lambda\,\:\mu}^{\:\:\lambda}}}\delta_{\beta}^{\alpha}=\Gamma_{\mu\,\:\beta}^{\:\:\alpha}+\frac{1}{n}\underset{Q_{\mu}}{\underbrace{\left(-\frac{1}{2}g^{\sigma\lambda}\nabla_{\mu}g_{\sigma\lambda}\right)}}\delta_{\beta}^{\alpha},\label{eq:solutionwithgauge} \end{equation} with $n$ is the dimension of $\mathcal{M}$, as shown in a detailed derivation carried in \cite{Iosifidis,Iosifidis4}. $\Gamma_{\mu\,\:\beta}^{\:\:\alpha}$ is the Levi-Civita connection, while $T_{\mu}$ and $Q_{\mu},$ are respectively, the torsion and disformation vectorial degrees of freedom ($Q_{\mu}$ is also known as the Weyl vector). Note that there exist differences in the factors in front of $T_{\mu}$ and $Q_{\mu}$ with \cite{Iosifidis,Iosifidis4}; these are due to the different definition on torsion tensor (\ref{eq:torsion}). However, since the action (\ref{eq:Palatini}) is invariant under (\ref{eq:projectivetranfs}), one could eliminate the vectorial degrees of freedom using the transformation (\ref{eq:projectivetranfs}), by setting an appropriate value of $\xi_{\mu}.$ In this article, we only consider the case where $\xi_{\mu}=\frac{1}{3}T_{\mu}$, with $n=4$. For the case where $\xi_{\mu}=-\frac{1}{n}Q_{\mu}$, or moreover, a linear combination of $T_{\mu}$ and $Q_{\mu}$, one could consult \cite{Iosifidis4,Smalley}. In the end, by performing the projective transformation (\ref{eq:projectivetranfs}) on (\ref{eq:solutionwithgauge}), one could fix the 4 vectorial degrees of freedom on $\omega_{\mu\,\:\beta}^{\:\:\alpha}$ such that (\ref{eq:solutionwithgauge}) becomes Levi-Civita: \[ \omega_{\mu\,\:\beta}^{\:\:\alpha}=\Gamma_{\mu\,\:\beta}^{\:\:\alpha}, \] with the condition: \begin{equation} T_{\mu}=T_{\lambda\,\:\mu}^{\:\:\lambda}=C_{\lambda\,\:\mu}^{\:\:\lambda}=0.\label{eq:projectcon} \end{equation} (\ref{eq:projectcon}) is known as the traceless torsion constraint. It needs to be kept in mind that the tracelessness of the torsion (\ref{eq:projectcon}) is introduced at the kinematical level; it does not result from the dynamics, i.e. the equation of motion. To obtain the Levi-Civita connection from the solution of the equation of motion, one needs to add a new term corresponding to the constraint into the action \cite{Sotiriou}: \begin{equation} S_{EH}\left[g,\omega,\chi\right]=\intop_{\mathcal{M}}\mathcal{R}\left[\omega\right]\mathrm{vol}\left[g\right]+S_{\mathrm{matter}}\left[g\right]+S_{LM}\left[\chi\right],\label{eq:PalatiniplusLM} \end{equation} where $S_{LM}=\intop_{\mathcal{M}}\chi^{\mu}T_{\mu}\textrm{vol }$and $\chi^{\mu}$ is the Lagrange multiplier corresponding to $T_{\mu}$. Minimizing (\ref{eq:PalatiniplusLM}) with respect to $g,$ $\omega,$ and $\chi$ gives: \begin{align} R_{\left(\alpha\beta\right)}-\frac{1}{2}\left(\mathcal{R}+\chi^{\mu}T_{\lambda\,\:\mu}^{\:\:\lambda}\right)g_{\alpha\beta} & =\mathcal{\kappa T}_{\alpha\beta},\label{eq:ELPalatini1}\\ \left(T_{\lambda\,\:\nu}^{\:\:\lambda}+Q_{\lambda\,\:\nu}^{\:\:\lambda}-\nabla_{\nu}\right)\left(2\delta_{\mu}^{\:\left[\nu\right.}g^{\left.\alpha\right]\beta}\right)+T_{\mu}^{\:\:\alpha\beta} & =\chi^{\alpha}\delta_{\mu}^{\beta}-\chi^{\beta}\delta_{\mu}^{\alpha},\label{eq:ELPalatini2}\\ T_{\lambda\,\:\mu}^{\:\:\lambda} & =0,\label{eq:ELPalatini3} \end{align} that could be solved to obtain $\chi^{\alpha}=0,$ hence, giving the Levi-Cita condition and the standard EFE. One could conclude that the Palatini formalism of gravity \textit{with} the traceless torsion constraint is equivalent with GR \cite{Iosifidis,Iosifidis4}. \subsubsection*{Metric-Affine Gravity, Metric-Affine $f\left(\mathcal{R}\right)$-Gravity, and Metric-Affine GR} Metric-Affine-Gravity (MAG) is a large class of theories based on the first-order formalism with a general affine connection that includes torsion and non-metricity. The choice of action for MAG could vary greatly: the Ricci scalar $\mathcal{R}$ for the Metric-Affine General Relativity (or Generalized Palatini), the power series of Ricci scalar $f\left(\mathcal{R}\right)$ for the Metric-Affine $f\left(\mathcal{R}\right)$ Gravity, and other exotic actions such as $f\left(\mathcal{R},R_{\mu\nu}R^{\mu\nu}\right)$ and $\mathcal{L}\left(g_{\mu\nu},R_{\mu\nu\;\beta}^{\quad\alpha}\right)$ \cite{Iosifidis4}. In this article, we only consider Metric-Affine $f\left(\mathcal{R}\right)$ Gravity and Metric-Affine General Relativity. The action of Metric-Affine $f\left(\mathcal{R}\right)$-gravity is defined as: \begin{equation} S\left[g,\omega,\chi\right]=\intop_{\mathcal{M}}f\left(\mathcal{R}\left[\omega\right]\right)\mathrm{vol}+S_{\mathrm{matter}}\left[g,\omega\right]+S_{\mathrm{LM}}\left[\chi\right].\label{eq:MAfR} \end{equation} Notice that now the matter action $S_{\mathrm{matter}}$ is also a functional of the affine connection $\omega$. The variation of $S_{\mathrm{matter}}$ with respect to $\omega$ gives the hypermomentum tensor $\mathcal{H}$: \begin{equation} \delta_{\omega}S_{\mathrm{matter}}\left[g,\omega\right]=-\kappa\mathcal{H}_{\:\:\mu}^{\alpha\,\:\beta}\delta\omega_{\alpha\,\:\beta}^{\:\:\mu}\mathrm{vol},\label{eq:hypemom} \end{equation} and therefore, minimizing the action (\ref{eq:MAfR}) with respect to the $g,$ $\omega,$ and $\chi$, results in three equations of motion: \begin{align} f'\left(\mathcal{R}\right)R_{\left(\alpha\beta\right)}-\frac{1}{2}\left(f\left(\mathcal{R}\right)+\chi^{\mu}T_{\lambda\,\:\mu}^{\:\:\lambda}\right)g_{\alpha\beta} & =\mathcal{\kappa T}_{\alpha\beta},\label{eq:ELmafr1}\\ \left(\left(T_{\lambda\,\:\nu}^{\:\:\lambda}+Q_{\lambda\,\:\nu}^{\:\:\lambda}-\nabla_{\nu}\right)\left(2\delta_{\mu}^{\:\left[\nu\right.}g^{\left.\alpha\right]\beta}\right)+T_{\mu}^{\:\:\alpha\beta}\right)f'\left(\mathcal{R}\right) & =\kappa\mathcal{H}_{\:\:\mu}^{\alpha\,\:\beta}+2\chi^{\left[\alpha\right.}\delta_{\mu}^{\left.\beta\right]},\label{eq:ELmafr2}\\ T_{\lambda\,\:\mu}^{\:\:\lambda} & =0.\label{eq:ELmaffr3} \end{align} The last equation is the constraint equation. Notice that the hypermomentum only exists if the matter term $S_{\mathrm{matter}}$ is a functional of the connection, hence in the Palatini formalism, $\mathcal{H}=0$. One could show that the torsion enters the dynamics through the antisymmetric part of the second and third indices $\mathcal{H}^{\alpha\left[\mu\beta\right]},$ while the non-metricity enters through the symmetric part of the first and third indices $\mathcal{H}^{\left(\alpha\right|\mu\left|\beta\right)},$ and if $\mathcal{H}_{\:\:\mu}^{\alpha\,\:\beta}=0,$ (\ref{eq:ELmafr2}) and (\ref{eq:ELmaffr3}) give the requirements of Levi-Civita connection, hence the theory coincides with the original GR, for $f\left(\mathcal{R}\right)=\mathcal{R}$. The equations of motions (\ref{eq:ELmafr2}) and (\ref{eq:ELmaffr3}) are the crucial results in the $f\left(\mathcal{R}\right)$-theory of gravity. However, due to the scope of this article, our focus will be on the stress-energy-momentum equation (\ref{eq:ELmafr1}). The theory of Metric-Affine General Relativity (MAGR) or Generalized Palatini theory could be obtained from Metric-Affine $f\left(\mathcal{R}\right)$-gravity by setting $f\left(\mathcal{R}\right)=\mathcal{R}$. With this requirement, (\ref{eq:ELmafr1})-(\ref{eq:ELmaffr3}) becomes: \begin{align} R_{\left(\alpha\beta\right)}-\frac{1}{2}\left(\mathcal{R}+\chi^{\mu}T_{\lambda\,\:\mu}^{\:\:\lambda}\right)g_{\alpha\beta} & =\mathcal{\kappa T}_{\alpha\beta},\label{eq:ELmagr1}\\ \left(T_{\lambda\,\:\nu}^{\:\:\lambda}+Q_{\lambda\,\:\nu}^{\:\:\lambda}-\nabla_{\nu}\right)\left(2\delta_{\mu}^{\:\left[\nu\right.}g^{\left.\alpha\right]\beta}\right)+T_{\mu}^{\:\:\alpha\beta} & =\kappa\mathcal{H}_{\:\:\mu}^{\alpha\,\:\beta}+2\chi^{\left[\alpha\right.}\delta_{\mu}^{\left.\beta\right]}\label{eq:ELmagr2}\\ T_{\lambda\,\:\mu}^{\:\:\lambda} & =0.\label{eq:ELmagr31} \end{align} which could be simplified to obtain: \begin{align} R_{\left(\alpha\beta\right)}-\frac{1}{2}\mathcal{R}g_{\alpha\beta} & =\mathcal{\kappa T}_{\alpha\beta},\label{eq:EoM1}\\ \left(Q_{\lambda\,\:\nu}^{\:\:\lambda}-\nabla_{\nu}\right)\left(2\delta_{\mu}^{\:\left[\nu\right.}g^{\left.\alpha\right]\beta}\right)+T_{\mu}^{\:\:\alpha\beta} & =\kappa\left(\mathcal{H}_{\:\:\mu}^{\alpha\,\:\beta}-\frac{2}{3}\mathcal{H}_{\:\:\;\;\,\sigma}^{\left[\alpha\right|\,\:\sigma}\delta_{\mu}^{\left|\beta\right]}\right),\label{eq:EoM2}\\ T_{\lambda\,\:\mu}^{\:\:\lambda} & =0,\label{eq:EoM3} \end{align} by inserting (\ref{eq:ELmagr31}) to (\ref{eq:ELmagr1})-(\ref{eq:ELmagr2}) and solving $\chi$ from (\ref{eq:ELmagr2}). Let us first focus only on the stress-energy-momentum equation (\ref{eq:EoM1}). The quantity in the LHS of (\ref{eq:EoM1}) is known as the generalized Einstein tensor: \[ G_{\alpha\beta}=R_{\left(\alpha\beta\right)}-\frac{1}{2}\mathcal{R}g_{\alpha\beta}, \] which is symmetric on the $\left(\alpha,\beta\right)$-indices. The stress-energy-momentum equation (\ref{eq:EoM1}) could be written in an equivalent form as follows \cite{Eric}: \begin{equation} R_{\left(\alpha\beta\right)}=\kappa\mathcal{T}_{\alpha\beta}-\frac{1}{2}\mathcal{T}g_{\alpha\beta},\label{eq:efenice} \end{equation} with $\mathcal{T}=g^{\alpha\beta}\mathcal{T}_{\alpha\beta}$ is the trace of the stress-energy momentum tensor (\ref{eq:tiga}). In the next part of this section, (\ref{eq:EoM1}) will be decomposed into its temporal and spatial part using the GCM equations derived in (\ref{eq:e}). \subsection{(3+1) Decomposition for MAGR} \subsubsection*{The Adapted Coordinate, Lapse Function, and Shift Vector} Let $\mathcal{M}$ be a globally hyperbolic Lorentzian manifold, and let $x^{\mu}=\left\{ x^{0},x^{i}\right\} $ be a general local coordinate on $\mathcal{M}.$ The coordinate vector basis on $T_{p}\mathcal{M}$ is $\partial_{\mu}=\left\{ \partial_{0},\partial_{i}\right\} $. Let $\partial_{0}$ be the temporal component of $\partial_{\mu},$ that could be decomposed according to the hypersurface $\Sigma$ as follows: \begin{equation} \partial_{0}=\underset{\boldsymbol{N}=N^{\mu}\partial_{\mu}}{\underbrace{\partial_{0}+g\left(\partial_{0},\hat{n}\right)\hat{n}}}+\underset{N}{\underbrace{-g\left(\partial_{0},\hat{n}\right)}\hat{n}},\label{eq:lapseshift} \end{equation} (and hence we take the lower part of the $\pm$ signature in (\ref{eq:lapseshift})). The scalar $N$ is the lapse function describing the normal component of $\partial_{0}$, while the vector $\boldsymbol{N}=N^{\mu}\partial_{\mu}$ is the shift vector describing the parallel part of $\partial_{0}$. Let the metric $g$ be written in terms of the component in the local coordinate $\partial_{\mu},$ using (\ref{eq:lapseshift}): \begin{align} g\left(\partial_{0},\partial_{0}\right) & =g_{00}=N^{\mu}N_{\mu}-N^{2},\label{eq:g1}\\ g\left(\partial_{0},\partial_{i}\right) & =g_{0i}=N_{i}+Nn_{i},\label{eq:g2}\\ g\left(\partial_{i},\partial_{j}\right) & =g_{ij}=\,^{3}q_{ij}-n_{i}n_{j}.\label{eq:g3} \end{align} Comparing with (\ref{eq:1a}), one could obtain $q_{00}=N_{0}=N^{\mu}N_{\mu},$ $q_{0i}=N_{i},$ and $n_{0}=-N$. Let us consider the adapted coordinate on $\mathcal{M}$ (it had been mentioned on the previous sections) where $\left\{ x^{i}\right\} $, $i=1,2,3,$ is a local coordinate on $\Sigma$. Notice that in this special coordinate, $n\perp\partial_{i}$, so that $n_{i}=g\left(n,\partial_{i}\right)=0$. Moreover, in this coordinate, the shift $\boldsymbol{N}\in T_{p}\Sigma$ does not have a temporal component, namely $N^{0}=0.$ Using the adapted coordinate, the spatial and the temporal part of $\Sigma$ could be cleanly separated. The components of metric $g$ could be written in the adapted coordinate as follows; (\ref{eq:g1})-(\ref{eq:g3}) becomes: \begin{align*} g\left(\partial_{0},\partial_{0}\right) & =g_{00}=N^{i}N_{i}-N^{2},\\ g\left(\partial_{0},\partial_{i}\right) & =g_{0i}=N_{i},\\ g\left(\partial_{i},\partial_{j}\right) & =g_{ij}=\,^{3}q_{ij}, \end{align*} while the components of the inverse metric $g^{-1},$ using the convention in (\ref{eq:convention}), are: \begin{align*} g^{*}\left(dx^{0},dx^{0}\right) & =g^{00}=-N^{-2},\\ g^{*}\left(dx^{0},dx^{i}\right) & =g^{0i}=N^{i}N^{-2},\\ g^{*}\left(dx^{i},dx^{j}\right) & =g^{ij}=\,^{3}q^{ij}-\left(N^{i}N^{j}\right)N^{-2}, \end{align*} with the coordinate basis vector on $T_{p}^{*}\mathcal{M}$ satisfies $dx^{0}=-\hat{n}^{*}N^{-1}$ and $dx^{i}=\hat{n}^{*}N^{i}N^{-1}+\,^{3}dx^{i}$. Notice that $dx^{i}$ is not necessarily equal to $\,^{3}dx^{i}=\,^{3}q^{ij}\,^{3}q\left(\partial_{j}\right)$ in (\ref{eq:d}). In the adapted coordinate, one could check that the following relations are true: \begin{align*} n & =n^{0}\partial_{0}+n^{i}\partial_{i}=N^{-1}\left(\partial_{0}-N^{i}\partial_{i}\right),\\ n^{*} & =n_{0}dx^{0}+n_{i}dx^{i}=-Ndx^{0},\\ \boldsymbol{N} & =N^{0}\partial_{0}+N^{i}\partial_{i}=-Nn^{i}\partial_{i},\\ \boldsymbol{N}^{*} & =N_{0}dx^{0}+N_{i}dx^{i}=N^{i}N_{i}dx^{0}+\,^{3}q_{ij}N^{j}dx^{i}. \end{align*} These relations will be useful for the following derivation. Applying the adapted coordinate to the Riemann curvature and torsion decomposition, namely (\ref{eq:e}) and (\ref{eq:torsion1})-(\ref{eq:torsion2}), one obtains: \begin{align} R_{ij\,\:\,k}^{\:\:\:\:0}= & \left(\,^{3}\nabla_{i}K_{jk}-\,^{3}\nabla_{j}K_{ik}+\,^{3}T_{i\,\:\,j}^{\:\:l}K_{lk}+\Theta_{i}K_{jk}-\Theta_{j}K_{ik}\right)n^{0},\label{eq:gcm1}\\ R_{ij\,\:\,l}^{\:\:\:\:k}= & \,^{3}R_{ij\,\:\,l}^{\:\:\:\:k}+K_{jl}\mathcal{K}_{i}^{\:\:k}-K_{il}\mathcal{K}_{j}^{\:\:k}+\underset{\frac{1}{n^{0}}R_{ij\,\:\,l}^{\:\:\:\:\:0}}{\underbrace{\left(\,^{3}\nabla_{i}K_{jl}-\,^{3}\nabla_{j}K_{il}+\,^{3}T_{i\,\:\:\:j}^{\,\:m}K_{ml}+\Theta_{i}K_{jl}-\Theta_{j}K_{il}\right)}}n^{k},\label{eq:gcm2} \end{align} which are, respectively, the components of Codazzi-Mainardi and Gauss equations, and: \begin{eqnarray} \hat{n}_{\mu}T_{i\,\:\:j}^{\:\:\mu} & = & K_{ij}-K_{ji},\label{eq:T1}\\ g_{k\mu}T_{i\,\:\:j}^{\:\:\mu} & = & \,^{3}q_{kl}\,^{3}T_{i\,\:\,j}^{\:\:l},\label{eq:T2} \end{eqnarray} which are the components of torsion decomposition. Here, the indices of the extrinsic curvatures are raised with the 3-metric, for example, $\mathcal{K}_{i}^{\:\:j}=\,^{3}q^{jk}\,^{3}\mathcal{K}_{ik}$. \subsubsection*{The Energy Part} Let us consider the purely-time part of the Einstein tensor, where the $\alpha,\beta$ indices of $G_{\alpha\beta}$ is contracted with the normal $\hat{n}$ to give the following scalar quantity: \begin{equation} G_{\alpha\beta}n^{\alpha}n^{\beta}=G\left(\hat{n},\hat{n}\right)=\mathbf{Ric}\left(\hat{n},\hat{n}\right)-\frac{1}{2}g\left(\hat{n},\hat{n}\right)\mathcal{R}.\label{eq:ham1} \end{equation} By a direct calculation, one could show that the generalized Ricci scalar could be decomposed into: \begin{equation} \mathcal{R}=\,^{3}\mathcal{R}-\mathrm{tr}\left(K\mathcal{K}\right)+\left(\mathrm{tr}K\right)\left(\mathrm{tr}\mathcal{K}\right)-\mathrm{\mathbf{Ric}}\left(\hat{n},\hat{n}\right)+\overline{\mathrm{\mathbf{Ric}}}\left(\hat{n},\hat{n}\right),\label{eq:Riccidec} \end{equation} where $\overline{\mathrm{\mathbf{Ric}}}$ is the co-Ricci tensor satisfying (\ref{eq:coRic}). Inserting (\ref{eq:Riccidec}) to (\ref{eq:ham1}) gives: \[ G\left(\hat{n},\hat{n}\right)=\frac{1}{2}\left(\mathrm{\mathbf{Ric}}\left(\hat{n},\hat{n}\right)+\overline{\mathrm{\mathbf{Ric}}}\left(\hat{n},\hat{n}\right)\right)+\frac{1}{2}\left(\,^{3}\mathcal{R}-\mathrm{tr}\left(K\mathcal{K}\right)+\left(\mathrm{tr}K\right)\left(\mathrm{tr}\mathcal{K}\right)\right). \] The first term is the additional part due to the non-metricity and torsion, this will be clear in the next subsections. The normal part of the stress-energy-momentum tensor (\ref{eq:tiga}) is the energy density, namely $\mathcal{T}\left(\hat{n},\hat{n}\right)=E,$ and the energy equation in MAG is: \begin{equation} \frac{1}{2}\left(\mathrm{\mathbf{Ric}}\left(\hat{n},\hat{n}\right)+\overline{\mathrm{\mathbf{Ric}}}\left(\hat{n},\hat{n}\right)\right)+\frac{1}{2}\left(\,^{3}\mathcal{R}-\mathrm{tr}\left(K\mathcal{K}\right)+\left(\mathrm{tr}K\right)\left(\mathrm{tr}\mathcal{K}\right)\right)=\kappa E.\label{eq:Hamconstraint} \end{equation} For metric connections, $\mathrm{\mathbf{Ric}}=-\overline{\mathrm{\mathbf{Ric}}}$ and $K=\mathcal{K},$ hence, for Levi-Civita connection, (\ref{eq:Hamconstraint}) returns to the original form, the Hamiltonian constraint. Notice that (\ref{eq:Hamconstraint}) is also valid for a metric connection with torsion, since the antisymmetric part of $\,^{3}\mathcal{R}$ and $K$, resulting from a non-vanishing torsion, do not contribute to the energy equation. \subsubsection*{The Momentum Part} The momentum part of the Einstein tensor is a mixture between the time and spatial parts. In the adapted coordinate, it could be written as follows: \begin{align} G_{i\mu}n^{\mu}=G\left(\partial_{i},\hat{n}\right)=G\left(\hat{n},\partial_{i}\right)= & \frac{1}{2}\left(\mathrm{\mathbf{Ric}}\left(\partial_{i},\hat{n}\right)+\mathrm{\mathbf{Ric}}\left(\hat{n},\partial_{i}\right)\right).\label{eq:dff} \end{align} where the term containing the Ricci scalar is zero due to the fact that in the adapted coordinate, $g\left(\partial_{i},\hat{n}\right)=n_{i}=0.$ From a direct calculation, one could show that the terms containing the 3-covariant derivative of the extrinsic curvature in (\ref{eq:gcm1}) comes from the co-Ricci tensor, instead of the Ricci tensor: \begin{equation} \mathrm{\overline{\mathbf{Ric}}}\left(\partial_{i},\hat{n}\right)=q^{jk}\left(\,^{3}\nabla_{i}K_{jk}-\,^{3}\nabla_{j}K_{ik}+\,^{3}T_{i\,\:\,j}^{\:\:l}K_{lk}+\Theta_{i}K_{jk}-\Theta_{j}K_{ik}\right)-g\left(\hat{n},\boldsymbol{R}\left(\hat{n},\partial_{i}\right)\hat{n}\right).\label{eq:aneeh} \end{equation} For a metric connection, $\mathrm{\mathbf{Ric}}=-\overline{\mathrm{\mathbf{Ric}}}$, so one could immediately insert (\ref{eq:aneeh}) to (\ref{eq:dff}). This is not the case for a general connection. Hence, in MAGR, there is no direct way to write the momentum equation in terms of the intrinsic curvature of the first kind $K$. Let us postponed this problem for the moment and write the momentum equation as follows: \begin{align} \frac{1}{2}\left(\mathrm{\mathbf{Ric}}\left(\partial_{i},\hat{n}\right)+\mathrm{\mathbf{Ric}}\left(\hat{n},\partial_{i}\right)\right) & =\kappa p_{i},\label{eq:diffeom} \end{align} with $p_{i}$ is the 'mixed' part of the stress-energy-momentum tensor (\ref{eq:tiga}), namely, the (relativistic) momentum $\mathcal{T}\left(\partial_{i},\hat{n}\right)=\mathcal{T}\left(\hat{n},\partial_{i}\right)=p_{i}.$ For a metric connection with torsion, the momentum equation becomes: \begin{equation} \,^{3}\nabla_{j}K_{i}^{\:\:j}-\,^{3}\nabla_{i}K_{j}^{\:\:j}-\,^{3}T_{i\,\:\,j}^{\:\:k}K_{k}^{\:\:j}+\frac{1}{2}n^{\nu}\left(\nabla_{\mu}T_{\nu\,\:i}^{\:\:\mu}\right)=\kappa p_{i},\label{eq:torsiondiff} \end{equation} where in general, the extrinsic curvature $K$ contains an antisymmetric part. For the Levi-Civita case, the torsion vanishes, and (\ref{eq:diffeom}) returns to the original momentum (or diffeomorphism) constraint, with symmetric $K$. \subsubsection*{The Stress-Energy Part} The purely-spatial part of the stress-energy-momentum equation are the following set of equations: \begin{equation} G_{ij}=G\left(\partial_{i},\partial_{j}\right)=\frac{1}{2}\left(\mathbf{Ric}\left(\partial_{i},\partial_{j}\right)+\mathbf{Ric}\left(\partial_{j},\partial_{i}\right)\right)-\frac{1}{2}g\left(\partial_{i},\partial_{j}\right)\mathcal{R}.\label{eq:dyns2} \end{equation} The first term, namely, the spatial part of the Ricci tensor, could be decomposed as follows: \begin{equation} \mathbf{Ric}\left(\partial_{i},\partial_{j}\right)=\,^{3}\mathbf{Ric}\left(\partial_{i},\partial_{j}\right)+K_{ij}\mathrm{tr}\mathcal{K}-\mathcal{K}_{i}^{\:\:k}K_{kj}-g\left(\hat{n},\boldsymbol{R}\left(\hat{n},\partial_{i}\right)\partial_{j}\right).\label{eq:ricdec} \end{equation} Inserting (\ref{eq:ricdec}) and the Ricci scalar (\ref{eq:Riccidec}) to (\ref{eq:dyns2}) gives immediately the spatial part of $G$: \begin{align*} G\left(\partial_{i},\partial_{j}\right)= & \,^{3}G\left(\partial_{i},\partial_{j}\right)-g\left(\hat{n},\boldsymbol{R}\left(\hat{n},\partial_{\left(i\right.}\right)\partial_{\left.j\right)}\right)+K_{\left(ij\right)}\mathrm{tr}\mathcal{K}-\mathcal{K}_{\left(i\right|}^{\,\:\:k}K_{k\left|j\right)}\\ & \qquad\qquad\quad-\frac{1}{2}\,^{3}q_{ij}\left(\left(\mathrm{tr}K\right)\left(\mathrm{tr}\mathcal{K}\right)-\mathrm{tr}\left(K\mathcal{K}\right)-\mathbf{Ric}\left(\hat{n},\hat{n}\right)+\overline{\mathbf{Ric}}\left(\hat{n},\hat{n}\right)\right), \end{align*} with: \[ \,^{3}G\left(\partial_{i},\partial_{j}\right)=\frac{1}{2}\left(\,^{3}\mathbf{Ric}\left(\partial_{i},\partial_{j}\right)+\,^{3}\mathbf{Ric}\left(\partial_{j},\partial_{i}\right)-\,^{3}q_{ij}\,^{3}\mathcal{R}\right) \] is the Einstein tensor on $\Sigma.$ The spatial part of the stress-energy-momentum tensor (\ref{eq:tiga}) is the stress tensor $\mathcal{T}\left(\partial_{i},\partial_{j}\right)=\mathcal{S}\left(\partial_{i},\partial_{j}\right)=\mathcal{S}\left(\partial_{j},\partial_{i}\right),$ hence the stress-energy equations of EFE is: \begin{align} \,^{3}G\left(\partial_{i},\partial_{j}\right)-g\left(\hat{n},\boldsymbol{R}\left(\hat{n},\partial_{\left(i\right.}\right)\partial_{\left.j\right)}\right)+K_{\left(ij\right)}\mathrm{tr}\mathcal{K}-\mathcal{K}_{\left(i\right|}^{\,\:\:k}K_{k\left|j\right)}\qquad\qquad\quad\label{eq:stress}\\ -\frac{1}{2}\,^{3}q_{ij}\left(\left(\mathrm{tr}K\right)\left(\mathrm{tr}\mathcal{K}\right)-\mathrm{tr}\left(K\mathcal{K}\right)-\mathbf{Ric}\left(\hat{n},\hat{n}\right)+\overline{\mathbf{Ric}}\left(\hat{n},\hat{n}\right)\right) & =\kappa\mathcal{S}\left(\partial_{i},\partial_{j}\right).\nonumber \end{align} One could write (\ref{eq:stress}) in a more convenient way as follows. Projecting (\ref{eq:efenice}) to the hypersurface $\Sigma$ gives: \[ \frac{1}{2}\left(\mathbf{Ric}\left(\partial_{i},\partial_{j}\right)+\mathbf{Ric}\left(\partial_{j},\partial_{i}\right)\right)=\kappa\mathcal{S}\left(\partial_{i},\partial_{j}\right)-\frac{1}{2}g\left(\partial_{i},\partial_{j}\right)\kappa\mathcal{T}. \] Using (\ref{eq:ricdec}), one could obtain: \begin{equation} \,^{3}R_{\left(ij\right)}+K_{\left(ij\right)}\mathrm{tr}\mathcal{K}-\mathcal{K}_{\left(i\right|}^{\,\:\:k}K_{k\left|j\right)}-g\left(\hat{n},\boldsymbol{R}\left(\hat{n},\partial_{\left(i\right.}\right)\partial_{\left.j\right)}\right)=\kappa\mathcal{S}_{ij}-\frac{1}{2}\,^{3}q_{ij}\kappa\mathcal{T}.\label{eq:stress2} \end{equation} In the original GR, the purely-spatial part of the EFE is the only part that contains the dynamics of the system, i.e., the equations which contain the time derivative of the 3-metric. In (\ref{eq:stress2}), the time derivative is hidden such that it is contained implicitly in the covariant term $g\left(\hat{n},\boldsymbol{R}\left(\hat{n},\partial_{\left(i\right.}\right)\partial_{\left.j\right)}\right).$ We will show that this is indeed the case in the following subsections. \subsubsection*{The Additional Variables on the Hypersurface} In the geometrodynamics concept introduced by Wheeler \cite{Wheeler}, the system of GR could be equivalently described using only fields in $\Sigma$ which evolve in time. In this perspective, one does not need to refer to the 4-manifold $\mathcal{M}$ explicitly. However, the EFE (\ref{eq:Hamconstraint}), (\ref{eq:diffeom}), and (\ref{eq:stress2}) contain some terms that are still covariant, therefore, these terms need to be decomposed into the temporal and spatial parts. Let us regard $\hat{n}$, the unit normal to $\Sigma$, as a 4-velocity in a Eulerian frame. The 4-acceleration, defined as $\nabla_{\hat{n}}\hat{n}=\alpha=\alpha^{\mu}\partial_{\mu}$ in (\ref{eq:accell}), has time and spatial components as follows: \begin{align} \left\langle \hat{n}^{*},\nabla_{\hat{n}}\hat{n}\right\rangle & =g\left(\nabla_{\hat{n}}\hat{n},\hat{n}\right)=g\left(\alpha,\hat{n}\right):=-\Theta\left(\hat{n}\right),\label{eq:1a-1}\\ \left\langle \,^{3}dx^{i},\nabla_{\hat{n}}\hat{n}\right\rangle & =\left\langle \,^{3}dx^{i},\alpha\right\rangle :=\alpha^{i}.\label{eq:2a} \end{align} Equation (\ref{eq:1a-1}), already defined in (\ref{eq:angle}), is the angle between the 4-velocity with the 4-acceleration which is zero for a metric connection. (\ref{eq:2a}) is the components of the (relativistic) 3-acceleration, $\,^{3}\alpha=\alpha^{i}\partial_{i}$ (not to be confused with $\boldsymbol{a}_{i}=\nabla_{i}\hat{n}$ in (\ref{eq:accell})). The acceleration $\alpha$ is a part of the acceleration tensor $\boldsymbol{a}$ defined in (\ref{eq:accell}), the other part of the tensor are the following: \begin{align} \left\langle \hat{n}^{*},\nabla_{i}\hat{n}\right\rangle =g\left(\nabla_{i}\hat{n},\hat{n}\right) & =g\left(\boldsymbol{a}_{i},\hat{n}\right):=-\Theta_{i},\label{eq:3a} \end{align} $\boldsymbol{a}_{i}$, together with $\boldsymbol{a}_{n}=\alpha$, define the acceleration tensor $\boldsymbol{a}=\left(\boldsymbol{a}_{n},\boldsymbol{a}_{i}\right).$ $\boldsymbol{a}_{i}$ is the rate of change of the 4-velocity in the spatial direction $\partial_{i}$; therefore $\Theta_{i}$ is the temporal components of $\boldsymbol{a}_{i}$ or the angle between $\alpha_{i}$ and $\hat{n}$. One could similarly obtain the spatial components of $\boldsymbol{a}_{i}$, but one could prove using equation (\ref{eq:1a}), (\ref{eq:extrinsic2}), and (\ref{eq:pentingx}), that this is only the extrinsic curvature of the second kind: \begin{equation} \left\langle \,^{3}dx^{j},\nabla_{i}\hat{n}\right\rangle =\,^{3}q^{jk}g\left(\nabla_{i}\hat{n},\partial_{k}\right)=\,^{3}q^{jk}g\left(\boldsymbol{a}_{i},\partial_{k}\right)=\,^{3}q^{jk}\mathcal{K}_{ik}=\mathcal{K}_{i}^{\:\:j},\label{eq:4a} \end{equation} which had been defined in (\ref{eq:extrinsic2}). Having defined all the covariant derivative of $\hat{n}$ in all directions and its components, now we could proceed to define the covariant derivative of the spatial direction. The first two are the components of $\nabla_{i}\partial_{j}$, the rate of change of $\partial_{j}$ in the direction $\partial_{i}$: \begin{align} \left\langle \hat{n}^{*},\nabla_{i}\partial_{j}\right\rangle & =g\left(\nabla_{i}\partial_{j},\hat{n}\right)=g\left(\omega_{i\:\:\:j}^{\,\,\:\mu}\partial_{\mu},\hat{n}\right)=-K_{ij},\label{eq:5a}\\ \left\langle \,^{3}dx^{k},\nabla_{i}\partial_{j}\right\rangle & =\left\langle \,^{3}dx^{k},\omega_{i\:\:\:j}^{\,\,\:\mu}\partial_{\mu}\right\rangle =\,^{3}\omega_{i\:\,\:j}^{\:\:k}.\label{eq:6a} \end{align} (\ref{eq:5a}) is the time component of $\nabla_{i}\partial_{j}$ which is exactly the extrinsic curvature of the first kind defined in (\ref{eq:3}). The spatial component of $\nabla_{i}\partial_{j}$ is exactly the induced 3-connection (\ref{eq:6a}) on $\Sigma$. The last quantities are the components of $\nabla_{\hat{n}}\partial_{i}$, which define the rate of change of $\partial_{i}$ in the direction of $\hat{n}$, i.e., the evolution of $\partial_{i}$ in time: \begin{align} \left\langle \hat{n}^{*},\nabla_{\hat{n}}\partial_{i}\right\rangle & =g\left(\nabla_{\hat{n}}\partial_{i},\hat{n}\right)=g\left(n^{\alpha}\omega_{\alpha\:\,\:i}^{\,\:\:\mu}\partial_{\mu},\hat{n}\right)=\omega\left(\hat{n}\right)_{\;\:i}^{\mu}n_{\mu}:=-\Delta_{i},\label{eq:7a}\\ \left\langle \,^{3}dx^{j},\nabla_{\hat{n}}\partial_{i}\right\rangle & =\left\langle \,^{3}dx^{j},n^{\alpha}\omega_{\alpha\:\,\:i}^{\,\:\:\mu}\partial_{\mu}\right\rangle =\omega\left(\hat{n}\right)_{\;\:i}^{j}:=\Delta_{\;\:i}^{j}.\label{eq:8a} \end{align} One could think of $\Delta_{i}$ and $\Delta_{\;\:i}^{j}$ as the generator of the evolution of $\partial_{i}$. (\ref{eq:7a}) is the temporal component of the evolution, which drags $\partial_{i}$ along the normal direction, while (\ref{eq:8a}) are the spatial components of the evolution, which moves $\partial_{i}$ along $\Sigma.$ Since $\nabla$ is not only a differentiation $\partial$, but it also rotates and shears vectors by $\omega$, the existence of the spatial components (\ref{eq:8a}) is understandable. Notice that in the original GR, the EFE could be written as functions of the following variables: the extrinsic curvature, the 3-connection (which is a function of metric) in terms of 3-Ricci scalar and 3-Einstein tensor, the 3-acceleration in terms of the lapse $N$, and the evolution part $\Delta_{\;\:i}^{j}$, which contains the shift $\boldsymbol{N}$. These variables are not independent of one another. For MAGR (and hence MAG), we have 8 different (with 4 additional) variables: $K,$ $\mathcal{K}$, $\,^{3}\alpha$, $\,^{3}\omega,$ $\Theta\left(\hat{n}\right),$ $\Theta_{i}$, $\Delta_{i}$, and $\Delta_{\;\:i}^{j}$. With the full variables on $\Sigma$, one could write the decompositions of the derivatives of $\hat{n}$ and $\partial_{i}$: \begin{align} \nabla_{\hat{n}}\hat{n} & =\Theta\left(\hat{n}\right)\hat{n}+\alpha^{i}\partial_{i},\label{eq:aa}\\ \nabla_{i}\hat{n} & =\Theta_{i}\hat{n}+\mathcal{K}_{i}^{\:\:j}\partial_{j},\label{eq:bb}\\ \nabla_{i}\partial_{j} & =K_{ij}\hat{n}+\,^{3}\omega_{i\:\,\:j}^{\:\:k}\partial_{k},\label{eq:cc}\\ \nabla_{\hat{n}}\partial_{i} & =\Delta_{i}\hat{n}+\Delta_{\;\:i}^{j}\partial_{j},\label{eq:dd} \end{align} to rederive the torsion and the non-metricity factor in terms of the additional variables as follows. The decomposition of the torsion tensor is: \begin{align} T\left(n,n\right) & =0,\label{eq:t1}\\ T\left(n,\partial_{i}\right)=-T\left(\partial_{i},n\right) & =\left(\Delta_{i}-\Theta_{i}-g\left(\partial_{i}\hat{n},\hat{n}\right)\right)\hat{n}+\left(\Delta_{\;\:i}^{j}-\mathcal{K}_{i}^{\:\:j}+\left\langle \,^{3}dx^{j},\partial_{i}\hat{n}\right\rangle \right)\partial_{j},\label{eq:t2}\\ T\left(\partial_{i},\partial_{j}\right) & =\,^{3}T\left(\partial_{i},\partial_{j}\right)+\left(K\left(\partial_{i},\partial_{j}\right)-K\left(\partial_{j},\partial_{i}\right)\right)\hat{n},\label{eq:t3} \end{align} while the decomposition of the non-metricity factor is: \begin{align} \nabla_{n}g^{*} & =-2\Theta\left(\hat{n}\right)\hat{n}\otimes\hat{n}+\left(\Delta^{i}-\alpha^{i}\right)\left(\hat{n}\otimes\partial_{i}+\partial_{i}\otimes\hat{n}\right)+\,^{3}\nabla_{n}\,^{3}q^{*}\label{eq:q1}\\ \nabla_{i}g^{*} & =-2\Theta_{i}\hat{n}\otimes\hat{n}+\left(K_{i}^{\;j}-\mathcal{K}_{i}^{\:\:j}\right)\left(\hat{n}\otimes\partial_{j}+\partial_{j}\otimes\hat{n}\right)+\,^{3}\nabla_{i}\,^{3}q^{*}\label{eq:q2} \end{align} where: \begin{align} \,^{3}\nabla_{n}\,^{3}q^{*}=\left(\,^{3}\nabla_{n}\,^{3}q^{ij}\right)\partial_{i}\otimes\partial_{j}= & \left(n\left[\,^{3}q^{ij}\right]+\Delta_{\;\:k}^{i}\,^{3}q^{kj}+\Delta_{\;\:k}^{j}\,^{3}q^{ki}\right)\partial_{i}\otimes\partial_{j},\label{eq:q3}\\ \,^{3}\nabla_{i}\,^{3}q^{*}=\left(\,^{3}\nabla_{i}\,^{3}q^{jk}\right)\partial_{i}\otimes\partial_{k}= & \left(\partial_{i}\,^{3}q^{jk}+\omega_{i}^{\;jk}+\omega_{i}^{\;kj}\right)\partial_{i}\otimes\partial_{k}.\label{eq:q4} \end{align} Therefore, the torsionless condition $T=0$ is equivalent to: \begin{equation} \begin{array}{ccccc} \,^{3}\omega_{i\:\,\:j}^{\:\:k}=\,^{3}\omega_{j\:\,\:i}^{\:\:k}, & \qquad & \Delta_{i} & = & g\left(\partial_{i}\hat{n},\hat{n}\right)+\Theta_{i},\\ K_{ij}=K_{ji}, & & \Delta_{\;\:i}^{j} & = & \mathcal{K}_{i}^{\:\:j}-\left\langle \,^{3}dx^{j},\partial_{i}\hat{n}\right\rangle , \end{array}\label{eq:tor1} \end{equation} with: \begin{align} g\left(\hat{n},\partial_{i}\hat{n}\right) & =\partial_{i}\ln N,\label{eq:N}\\ \left\langle dx^{j},\partial_{i}\hat{n}\right\rangle & =\frac{1}{N}\left(N^{j}\partial_{i}\ln N-\partial_{i}N^{j}\right),\label{eq:Nj}\\ \left\langle \,^{3}dx^{j},\partial_{i}\hat{n}\right\rangle & =-\frac{1}{N}\partial_{i}N^{j}, \end{align} while the metric compatibility $\nabla_{\mu}g^{*}=0$ is equivalent to: \begin{equation} \begin{array}{cccc} \Theta\left(\hat{n}\right) & =0, & \qquad & \,^{3}\alpha_{i}=\Delta_{i},\\ \Theta_{i} & =0, & & \mathcal{K}_{ij}=K_{ij}, \end{array}\label{eq:met1} \end{equation} together with: \begin{align} \,^{3}\nabla_{n}\,^{3}q^{*}= & 0,\label{eq:met}\\ \,^{3}\nabla_{i}\,^{3}q^{*}= & 0.\nonumber \end{align} The relation (\ref{eq:t1})-(\ref{eq:q2}) could be used to confirm the theorems we proved in Section III: The metricity condition $\nabla_{\mu}g=0$ will cause $\,^{3}\nabla_{i}\,^{3}q^{*}=0,$ and $T\left(\partial_{\mu},\partial_{\nu}\right)=0$ will cause $\,^{3}T\left(\partial_{i},\partial_{j}\right),$ but these relations are not valid vice-versa. The Levi-Civita connection must satisfy the metric compatibility and torsionless condition, and hence, the 8 additional variables are constrained by (\ref{eq:tor1})-(\ref{eq:met1}). Applying these constraints, the variables in the stress-energy-momentum equation for the Levi-Civita connection reduce to 4: $K=\mathcal{K}$, $\,^{3}\omega,$ $\,^{3}\alpha_{i}=\Delta_{i}=\partial_{i}\ln N$, and $\Delta_{\;\:i}^{j}=\mathcal{K}_{i}^{\:\:j}+\frac{1}{N}\partial_{i}N^{j}$. \subsubsection*{The Results} Relations (\ref{eq:aa})-(\ref{eq:dd}) are used to split the covariant parts in (\ref{eq:Hamconstraint}), (\ref{eq:diffeom}), and (\ref{eq:stress2}) into (3+1) forms, in particular: \begin{align*} \mathrm{\mathbf{Ric}}\left(\hat{n},\hat{n}\right)= & \Theta\left(\hat{n}\right)\mathrm{tr}\mathcal{K}-\Theta_{i}\alpha^{i}-\hat{n}\left[\mathrm{tr}\mathcal{K}\right]-\mathcal{K}_{i}^{\:\:j}\Delta_{\;\:j}^{i}+\,^{3}\nabla_{i}\alpha^{i}+\alpha^{i}g\left(\hat{n},\partial_{i}\hat{n}\right)-\mathcal{K}_{j}^{\:\:i}\left\langle ^{3}dx^{j},\partial_{i}\hat{n}\right\rangle ,\\ \overline{\mathrm{\mathbf{Ric}}}\left(\hat{n},\hat{n}\right)= & \,^{3}q^{ij}\left(\Theta\left(\hat{n}\right)K_{ij}-\Theta_{i}\Delta_{j}+\hat{n}\left[K_{ij}\right]-K_{ik}\Delta_{\;\:j}^{k}-\,^{3}\nabla_{i}\Delta_{j}-\Delta_{j}g\left(\hat{n},\partial_{i}\hat{n}\right)+K_{kj}\left\langle ^{3}dx^{k},\partial_{i}\hat{n}\right\rangle \right), \end{align*} \begin{align*} \mathrm{\mathbf{Ric}}\left(\hat{n},\partial_{i}\right)= & -\alpha^{j}K_{ji}-\hat{n}\left[\,^{3}\omega_{j\:\,\:i}^{\:\:j}\right]+\Delta_{i}\mathrm{tr}\mathcal{K}+\,^{3}\nabla_{j}\Delta_{\:\;i}^{j}+\Delta_{\:\;i}^{j}g\left(\hat{n},\partial_{j}\hat{n}\right)-\,^{3}\omega_{k\:\,\:i}^{\:\:j}\left\langle ^{3}dx^{k},\partial_{j}\hat{n}\right\rangle ,\\ \mathrm{\mathbf{Ric}}\left(\partial_{i},\hat{n}\right)= & \,^{3}\nabla_{j}\mathcal{K}_{i}^{\:\:j}-\,^{3}\nabla_{i}\mathcal{K}_{j}^{\:\:j}+\,^{3}T_{j\,\:\,i}^{\:\:k}\mathcal{K}_{k}^{\:\:j}+\Theta_{i}\mathcal{K}_{j}^{\:\:j}-\Theta_{j}\mathcal{K}_{i}^{\:\:j}\\ & +\mathcal{K}_{i}^{\:\:j}\Delta_{j}-K_{ij}\alpha^{j}-\partial_{i}\Theta\left(\hat{n}\right)+\hat{n}\left[\Theta_{i}\right]-\Theta\left(\hat{n}\right)g\left(\hat{n},\partial_{i}\hat{n}\right)+\Theta_{j}\left\langle ^{3}dx^{j},\partial_{i}\hat{n}\right\rangle , \end{align*} and: \begin{align*} g\left(\hat{n},\boldsymbol{R}\left(\hat{n},\partial_{i}\right)\partial_{j}\right)= & \,^{3}\nabla_{i}\Delta_{j}+\Theta_{i}\Delta_{j}-\Theta\left(\hat{n}\right)K_{ij}+K_{ik}\Delta_{\;\:j}^{k}-\hat{n}\left[K_{ij}\right]+\Delta_{j}g\left(\hat{n},\partial_{i}\hat{n}\right)-K_{kj}\left\langle \,^{3}dx^{k},\partial_{i}\hat{n}\right\rangle . \end{align*} One could observe that the Ricci tensor $\mathrm{\mathbf{Ric}}\left(\partial_{i},\hat{n}\right)$ contains the 3-covariant derivative of $\mathcal{K}$ instead of $K$ as in (\ref{eq:gcm1}). Together with (\ref{eq:N}) and (\ref{eq:Nj}), the (3+1) field equations for MAG could be written in partial differential equations containing only 3-dimensional variables on the hypersurface $\Sigma$: \begin{align} & \frac{1}{2}\left(\,^{3}\mathcal{R}-\mathrm{tr}\left(K\mathcal{K}\right)+\left(\mathrm{tr}K\right)\left(\mathrm{tr}\mathcal{K}\right)-\left(\mathcal{K}_{i}^{\:\:j}+K_{\:\:i}^{j}\right)\Delta_{\;\:j}^{i}+\,^{3}\nabla_{i}\alpha^{i}-\,^{3}q^{ij}\,^{3}\nabla_{i}\Delta_{j}+\,^{3}q^{ij}\hat{n}\left[K_{ij}\right]\right.\label{eq:satu}\\ & \left.\quad-\hat{n}\left[\mathrm{tr}\mathcal{K}\right]-\Theta_{i}\left(\alpha^{i}+\,^{3}q^{ij}\Delta_{j}\right)+\Theta\left(\hat{n}\right)\left(\mathrm{tr}\mathcal{K}+\mathrm{tr}K\right)+\left(\alpha^{i}-\,^{3}q^{ij}\Delta_{j}\right)\partial_{i}\ln N+\frac{1}{N}\left(\mathcal{K}_{j}^{\:\:i}-K_{j}^{\,\:i}\right)\partial_{i}N^{j}\right)=\kappa E.\nonumber \end{align} \begin{align} & \frac{1}{2}\left(\,^{3}\nabla_{j}\mathcal{K}_{i}^{\:\:j}-\,^{3}\nabla_{i}\mathcal{K}_{j}^{\:\:j}+\,^{3}T_{j\,\:\,i}^{\:\:k}\mathcal{K}_{k}^{\:\:j}-\alpha^{j}\left(K_{ij}+K_{ji}\right)+\,^{3}\nabla_{j}\Delta_{\;i}^{j}-\hat{n}\left[\,^{3}\omega_{j\:\,\:i}^{\:\:j}\right]+\hat{n}\left[\Theta_{i}\right]-\partial_{i}\Theta\left(\hat{n}\right)\right.\label{eq:dua}\\ & \quad\left.+\left(\Delta_{i}+\Theta_{i}\right)\mathcal{K}_{j}^{\:\:j}+\left(\Delta_{j}-\Theta_{j}\right)\mathcal{K}_{i}^{\:\:j}+\Delta_{\;\:i}^{j}\partial_{j}\ln N-\Theta\left(\hat{n}\right)\partial_{i}\ln N+\frac{1}{N}\left(\,^{3}\omega_{k\:\,\:i}^{\:\:j}\partial_{j}N^{k}-\Theta_{j}\partial_{i}N^{j}\right)\right)=\kappa p_{i},\nonumber \end{align} \begin{align} & \hat{n}\left[K_{\left(ij\right)}\right]-\frac{1}{N}K_{k\left(i\right.}\partial_{\left.j\right)}N^{k}-\left(\partial_{\left(i\right|}\ln N+\Theta_{\left(i\right|}-\,^{3}\nabla_{\left(i\right|}\right)\Delta_{\left|j\right)}\label{eq:tiga-1}\\ & \quad\;\;\;+\,^{3}R_{\left(ij\right)}+\left(\mathrm{tr}\mathcal{K}+\Theta\left(\hat{n}\right)\right)K_{\left(ij\right)}-\mathcal{K}_{\left(i\right|}^{\,\:\:k}K_{k\left|j\right)}-K_{\left(i\right|k}\Delta_{\;\:\left|j\right)}^{k}=\kappa\mathcal{S}_{ij}-\frac{1}{2}\,^{3}q_{ij}\kappa\left(\mathcal{S}-E\right),\nonumber \end{align} where we use the fact that $\mathcal{T=\mathcal{S}}-E$, with $\mathcal{S}=g^{ij}\mathcal{S}_{ij}$. (\ref{eq:satu})-(\ref{eq:tiga-1}), are respectively, the energy, momentum, and stress-energy equations for MAGR. Notice the existence of the additional variables. One could show that by inserting the Levi-Civita condition (\ref{eq:tor1}) and (\ref{eq:met1})-(\ref{eq:met}), they return to the original (3+1) Einstein field equation. \section{Discussions and Conclusions} \subsection{The Stress-Energy-Momentum Equation} Equation (\ref{eq:satu})-(\ref{eq:tiga-1}) are the (3+1) decomposition of the first Euler-Lagrange equation (\ref{eq:EoM1}); it comes from the variation of action (\ref{eq:MAfR}) (for $f\left(\mathcal{R}\right)=\mathcal{R}$) with respect to metric $g$. One could see that they provide 1+3+6=10 differential equations. For simplicity, let us take the time gauge (or the Gauss normal coordinate \cite{Eric}), namely $N=1$, and $\boldsymbol{N}=0$. Hence, equation (\ref{eq:satu})-(\ref{eq:tiga-1}) becomes: \begin{align} & \frac{1}{2}\left(\,^{3}\mathcal{R}-\mathrm{tr}\left(K\mathcal{K}\right)+\left(\mathrm{tr}K\right)\left(\mathrm{tr}\mathcal{K}\right)-\left(\mathcal{K}_{i}^{\,\:j}+K_{\,\:i}^{j}\right)\Delta_{\;j}^{i}-\hat{n}\left[\mathrm{tr}\mathcal{K}\right]\right.\label{eq:satua}\\ & \left.\quad+\,^{3}q^{ij}\hat{n}\left[K_{ij}\right]+\,^{3}\nabla_{i}\alpha^{i}-\,^{3}q^{ij}\,^{3}\nabla_{i}\Delta_{j}-\Theta_{i}\left(\alpha^{i}+\,^{3}q^{ij}\Delta_{j}\right)+\Theta\left(\hat{n}\right)\left(\mathrm{tr}\mathcal{K}+\mathrm{tr}K\right)\right)=\kappa E,\nonumber \end{align} \begin{align} & \frac{1}{2}\left(\:\,^{3}\nabla_{j}\mathcal{K}_{i}^{\,\:j}-\,^{3}\nabla_{i}\mathcal{K}_{j}^{\,\:j}+\,^{3}T_{j\,\:i}^{\,\:\:k}\mathcal{K}_{k}^{\,\:j}-\alpha^{j}\left(K_{ij}+K_{ji}\right)\right.\label{eq:duaa}\\ & \quad\;\;\;\left.+\,^{3}\nabla_{j}\Delta_{\;i}^{j}-\hat{n}\left[\,^{3}\omega_{j\,\:i}^{\,\:\:j}\right]+\left(\Delta_{i}+\Theta_{i}\right)\mathcal{K}_{j}^{\,\:j}+\left(\Delta_{j}-\Theta_{j}\right)\mathcal{K}_{i}^{\,\:j}-\partial_{i}\Theta\left(\hat{n}\right)+\hat{n}\left[\Theta_{i}\right]\right)=\kappa p_{i},\nonumber \end{align} \begin{align} \hat{n}\left[K_{\left(ij\right)}\right]-\Theta_{\left(i\right.}\Delta_{\left.j\right)}-\,^{3}\nabla_{\left(i\right.}\Delta_{\left.j\right)}+\,^{3}R_{\left(ij\right)}+\left(\mathrm{tr}\mathcal{K}+\Theta\left(\hat{n}\right)\right)K_{\left(ij\right)}-\mathcal{K}_{\left(i\right|}^{\,\:\:k}K_{k\left|j\right)}-K_{\left(i\right|k}\Delta_{\;\:\left|j\right)}^{k} & =\kappa\mathcal{S}_{ij}-\frac{1}{2}\,^{3}q_{ij}\kappa\left(\mathcal{S}-E\right),\label{eq:tigaa} \end{align} but the physical interpretation is invariant under the change of coordinate. The energy equation (\ref{eq:satua}), contains time evolution (and hence the dynamics) from the terms $\hat{n}\left[\mathrm{tr}\mathcal{K}\right]$ and $\hat{n}\left[K_{ij}\right]$. These originate from the term $\mathrm{\mathbf{Ric}}\left(\hat{n},\hat{n}\right)+\overline{\mathrm{\mathbf{Ric}}}\left(\hat{n},\hat{n}\right),$ which is not zero due to the non-metricity and torsion. The momentum equation (\ref{eq:duaa}) also contains dynamics from $\hat{n}\left[\,^{3}\omega_{j\,\:i}^{\,\:\:j}\right]$ and $\hat{n}\left[\Theta_{i}\right]$. The first term contains the change of 3-connection in time, while the second is the change of the angle between $\boldsymbol{a}_{i}$ and $\hat{n}$. The second term will vanish for the Levi-Civita connection. For the first term, it enters the equation of motion because the connection is treated as an independent variable in an equal footing with metric, in the Palatini formulation. For the Levi-Civita case, one could show that $\hat{n}\left[\,^{3}\omega_{j\:\,\:i}^{\:\:j}\right]=\,^{3}\nabla_{i}K_{j}^{\:\:j}.$ Notice also the explicit existence of the 3-torsion in the momentum equation. The last equation, the stress-energy part, contains dynamics via $\hat{n}\left[K_{\left[ij\right]}\right]$ as in the energy equation. The torsion only enters the momentum equation explicitly, but it is contained implicitly in all the equations, for example, in the antisymmetric part of the extrinsic curvatures $K$ and $\mathcal{K}$. All these three equations are dynamical. One might ask where are the terms containing the time derivatives of the 3-metric in equation (\ref{eq:satua})-(\ref{eq:tigaa}). In the original EFE, equation (\ref{eq:satua}) and (\ref{eq:duaa}) become, respectively, the Hamiltonian and momentum (or diffeomorphism) constraint with Levi-Civita condition (\ref{eq:tor1}) and (\ref{eq:met1})-(\ref{eq:met}), there is no term containing the derivative with respect to time in such equations because the additional parts cancel with each other: \begin{align} \frac{1}{2}\left(\,^{3}\mathcal{R}-\mathrm{tr}\left(K^{2}\right)+\left(\mathrm{tr}K\right)^{2}\right) & =\kappa E,\label{eq:nic1}\\ \,^{3}\nabla_{j}K_{i}^{\,\:j}-\,^{3}\nabla_{i}K_{j}^{\,\:j} & =\kappa p_{i}.\label{eq:nic2} \end{align} On the other hand, equation (\ref{eq:tigaa}), in the standard GR, becomes: \begin{equation} \mathcal{L}_{\hat{n}}K_{ij}+\,^{3}R_{ij}+\left(\mathrm{tr}K\right)K_{ij}-2K_{i}^{\,\:\:k}K_{kj}=\kappa\mathcal{S}_{ij}-\frac{1}{2}\,^{3}q_{ij}\kappa\left(\mathcal{S}-E\right),\label{eq:nic} \end{equation} with $\mathcal{L}_{\hat{n}}K_{ij}$ is the Lie derivative of $K$ in the direction $\hat{n}$: \[ \mathcal{L}_{\hat{n}}K_{ij}=\hat{n}\left[K_{ij}\right]+K_{ik}\partial_{j}n^{k}+K_{kj}\partial_{i}n^{k}, \] where the last two terms in the RHS are zero in the normal coordinate. The dynamics in (\ref{eq:tigaa}) is contained in the term $\hat{n}\left[K_{\left[ij\right]}\right]$, however, the 3-metric and $K$ are related by the following equation: \begin{equation} \hat{n}\left[\,^{3}q_{ij}\right]=\left(\nabla_{\hat{n}}g\right)\left(\partial_{i},\partial_{j}\right)+K_{ij}+K_{ji},\label{eq:notnice} \end{equation} where for the Levi-Civita case becomes: \begin{equation} \hat{n}\left[\,^{3}q_{ij}\right]=2K_{ij},\label{eq:niceonemaan} \end{equation} hence, one could insert (\ref{eq:niceonemaan}) to (\ref{eq:nic}) to obtain the terms containing the double derivative of the 3-metric with respect to the time coordinate $\partial_{0}^{2}\,^{3}q_{ij},$ giving the standard dynamics of GR. However, for the general affine connection, $\nabla_{\hat{n}}g\neq0$, but: \[ \left(\nabla_{\hat{n}}g\right)\left(\partial_{i},\partial_{j}\right)=\hat{n}\left[\,^{3}q_{ij}\right]-\Delta_{ij}-\Delta_{ji}, \] causing the $\hat{n}\left[\,^{3}q_{ij}\right]$'s in (\ref{eq:notnice}) to cancel each other, while leaving $K_{ij}+K_{ji}=\Delta_{ij}+\Delta_{ji},$ free from $\,^{3}q_{ij}$. Therefore, the momentum and stress-energy equation (\ref{eq:duaa})-(\ref{eq:tigaa}) are free from the time derivative of $\,^{3}q_{ij}$. However, the dynamics of the 3-metric enters the energy equation (\ref{eq:satua}) from the term $\hat{n}\left[\mathrm{tr}\mathcal{K}\right]=\hat{n}\left[\,^{3}q_{ij}\mathcal{K}^{ij}\right]$. One needs to keep in mind that at this stage, we are only working with the first equation of motion (\ref{eq:EoM1}); there still exists another equation of motion, namely, the one obtained from the variation of the action with respect to the connection (\ref{eq:EoM2}). The first equation of motion (\ref{eq:EoM1}) provides only 10 differential equations, whereas the unknown variables are 74 (10 from the metric, 64 from the connections, assuming the theory does not have constraint). With the additional variables on the hypersurface, we have introduced 64 unknown variables (1 for $\Theta\left(\hat{n}\right),$ 3 for each $\Theta_{i}$, $\Delta_{i}$, $\,^{3}\alpha_{i}$, 9 for $K,$ $\mathcal{K}$, $\Delta_{\;i}^{j}$, and 27 for $\,^{3}\omega$), and 10 more unknowns should come from the metric $g_{\mu\nu},$ in terms of the 3-metric $\,^{3}q_{ij},$ the lapse $N$, and the shift $\boldsymbol{N}.$ The second Euler-Lagrange equation (\ref{eq:EoM2}) will provide 60 more differential equations, leaving 4 vectorial degrees of freedom on the connection. The last 4 equations come from the traceless torsion constraint (\ref{eq:EoM3}), by taking the projective invariance transformation (\ref{eq:projectivetranfs}) into account. Without the (3+1) decomposition of these equations of motions as well, it is impossible to do a complete analysis of the (3+1) MAGR theory. \subsection{The Hypermomentum Equation and Traceless Torsion Constraint} For the completeness of the discussion in this article, we add some of the results from our companion paper. Here, we only present the (3+1) decomposition of the hypermomentum equation in the normal (Gauss) coordinate. The general treatment and the detailed derivation of the result will be discussed in our companion paper. \subsubsection*{(3+1) Decomposition of Hypermomentum} The hypermomentum (\ref{eq:hypemom}) could be written with the spacetime index $\alpha$ is hidden as $\mathcal{H}\left(\partial_{\mu},dx^{\beta}\right)=\mathcal{H}_{\:\:\mu}^{\alpha\,\:\beta}\partial_{\alpha}.$ Using this notation, one could decompose the hypermomentum by its 'internal' indices into the normal and parallel parts, with respect to the hypersurface $\Sigma$: \begin{align} \mathcal{H}\left(\hat{n},\hat{n}^{*}\right) & =\kappa n^{\mu}n_{\beta}\mathcal{H}_{\:\:\mu}^{\alpha\,\:\beta}\partial_{\alpha}=-\left\langle \hat{n}^{*},\mathcal{H}\left(\hat{n},\hat{n}^{*}\right)\right\rangle \hat{n}+\left\langle \,^{3}dx^{i},\mathcal{H}\left(\hat{n},\hat{n}^{*}\right)\right\rangle \partial_{i},\nonumber \\ \mathcal{H}\left(\hat{n},\,^{3}dx^{i}\right) & =\kappa n^{\mu}\mathcal{H}_{\:\:\mu}^{\alpha\,\:i}\partial_{\alpha}=-\left\langle \hat{n}^{*},\mathcal{H}\left(\hat{n},\,^{3}dx^{i}\right)\right\rangle \hat{n}+\left\langle \,^{3}dx^{j},\mathcal{H}\left(\hat{n},\,^{3}dx^{i}\right)\right\rangle \partial_{j},\nonumber \\ \mathcal{H}\left(\partial_{i},\hat{n}^{*}\right) & =\kappa n_{\beta}\mathcal{H}_{\:\:i}^{\alpha\,\:\beta}\partial_{\alpha}=-\left\langle \hat{n}^{*},\mathcal{H}\left(\partial_{i},\hat{n}^{*}\right)\right\rangle \hat{n}+\left\langle \,^{3}dx^{j},\mathcal{H}\left(\partial_{i},\hat{n}^{*}\right)\right\rangle \partial_{j},\label{eq:h1}\\ \mathcal{H}\left(\partial_{i},\,^{3}dx^{j}\right) & =\kappa\mathcal{H}_{\:\:i}^{\alpha\,\:j}\partial_{\alpha}=-\left\langle \hat{n}^{*},\mathcal{H}\left(\partial_{i},\,^{3}dx^{j}\right)\right\rangle \hat{n}+\left\langle \,^{3}dx^{k},\mathcal{H}\left(\partial_{i},\,^{3}dx^{j}\right)\right\rangle \partial_{k}.\nonumber \end{align} Following the decomposition of $\mathcal{T}$ into its 3 components, i.e., the energy $E$, the momentum $p_{i}$, and the stress $\mathcal{S}_{ij}$, one could apply the same procedure to the hypermomentum $\mathcal{H},$ where the decomposition is based on the split of the 'spacetime' into space and time. This should not be confused with the split of $\mathcal{H}$ based on the symmetricity of the indices which gives the spin, shear, and dilation parts as in \cite{Hehl}. \subsubsection*{The (3+1) Hypermomentum Equation in Normal Coordinate} With the (3+1) decomposition in (\ref{eq:h1}), the hypermomentum equation (\ref{eq:EoM2}) could be split into 4 equations as follows: \begin{align} & n^{\mu}n^{\lambda}T_{\mu\;\;\;\lambda}^{\:\:\alpha}\partial_{\alpha}-n_{\beta}\left(\nabla_{\hat{n}}g^{\alpha\beta}\right)\partial_{\alpha}+n_{\beta}\left(\nabla_{\nu}g^{\nu\beta}\right)\hat{n}=\kappa n^{\mu}n_{\beta}\mathcal{H}_{\:\:\mu}^{\alpha\,\:\beta}\partial_{\alpha}-\frac{2}{3}n^{\mu}n_{\beta}\kappa\mathcal{H}_{\:\:\;\;\,\sigma}^{\left[\alpha\right|\,\:\sigma}\delta_{\mu}^{\left|\beta\right]}\partial_{\alpha},\label{eq:gokil}\\ & n^{\mu}g^{i\lambda}T_{\mu\;\;\;\lambda}^{\:\:\alpha}\partial_{\alpha}+\frac{1}{2}\left(g_{\sigma\lambda}\nabla_{\nu}g^{\sigma\lambda}\right)\left(n^{\nu}g^{\alpha i}\partial_{\alpha}-g^{\nu i}\hat{n}\right)-\left(\nabla_{\hat{n}}g^{\alpha i}\right)\partial_{\alpha}+\left(\nabla_{\nu}g^{\nu i}\right)\hat{n}=\kappa n^{\mu}\mathcal{H}_{\:\:\mu}^{\alpha\,\:i}\partial_{\alpha}-n^{\mu}\delta_{\beta}^{i}\frac{2}{3}\kappa\mathcal{H}_{\:\:\;\;\,\sigma}^{\left[\alpha\right|\,\:\sigma}\delta_{\mu}^{\left|\beta\right]}\partial_{\alpha},\nonumber \\ & n^{\lambda}T_{i\;\;\;\lambda}^{\:\:\alpha}\partial_{\alpha}+\frac{1}{2}\left(g_{\sigma\lambda}\nabla_{\nu}g^{\sigma\lambda}\right)\left(\delta_{i}^{\:\nu}\hat{n}-n^{\nu}\partial_{i}\right)-n_{\beta}\left(\nabla_{i}g^{\alpha\beta}\right)\partial_{\alpha}+n_{\beta}\left(\nabla_{\nu}g^{\nu\beta}\right)\partial_{i}=\kappa n_{\beta}\mathcal{H}_{\:\:i}^{\alpha\,\:\beta}\partial_{\alpha}-\delta_{i}^{\mu}n_{\beta}\frac{2}{3}\kappa\mathcal{H}_{\:\:\;\;\,\sigma}^{\left[\alpha\right|\,\:\sigma}\delta_{i}^{\left|\beta\right]}\partial_{\alpha},\nonumber \\ & g^{j\lambda}T_{i\;\;\;\lambda}^{\:\:\alpha}\partial_{\alpha}+\frac{1}{2}\left(g_{\sigma\lambda}\nabla_{\nu}g^{\sigma\lambda}\right)\left(\delta_{i}^{\:\nu}g^{\alpha j}\partial_{\alpha}-g^{\nu j}\partial_{i}\right)-\left(\nabla_{i}g^{\alpha j}\right)\partial_{\alpha}+\left(\nabla_{\nu}g^{\nu j}\right)\partial_{i}=\kappa\mathcal{H}_{\:\:i}^{\alpha\,\:j}\partial_{\alpha}-\delta_{i}^{\mu}\delta_{\beta}^{j}\frac{2}{3}\kappa\mathcal{H}_{\:\:\;\;\,\sigma}^{\left[\alpha\right|\,\:\sigma}\delta_{\mu}^{\left|\beta\right]}\partial_{\alpha}.\nonumber \end{align} Using equation (\ref{eq:h1}) together with torsion and non-metricity decomposition in (\ref{eq:t1})-(\ref{eq:q2}), one could rewrite the four equations (\ref{eq:gokil}) in terms of the additional variables as follows: \begin{align} \left(\mathcal{K}_{i}^{\:\:i}-K_{i}^{\;i}\right)\hat{n}+\left(\Delta^{i}-\alpha^{i}\right)\partial_{i} & =\kappa\left(\mathcal{H}\left(\hat{n},\hat{n}^{*}\right)+\frac{1}{3}\left(\mathrm{tr}\mathcal{H}+\left\langle \hat{n},\mathrm{tr}\mathcal{H}\right\rangle \hat{n}\right)\right),\label{eq:hyp1} \end{align} \begin{align} \left(\,^{3}q^{ij}\left(\Theta\left(\hat{n}\right)+\frac{1}{2}\,^{3}q_{kl}\left(n\left[\,^{3}q^{kl}\right]+\Delta^{kl}+\Delta^{lk}\right)\right)-\mathcal{K}^{ij}-n\left[\,^{3}q^{ij}\right]-\Delta^{ij}\right)\partial_{j}\label{eq:hyp2}\\ +\left(\Delta^{i}-2\Theta^{i}+\,^{3}\nabla_{j}\,^{3}q^{ji}-\frac{1}{2}\,^{3}q_{jk}\,^{3}\nabla^{i}\,^{3}q^{jk}\right)\hat{n} & =\kappa\left(\mathcal{H}\left(\hat{n},dx^{i}\right)+\frac{1}{3}\left\langle dx^{i},\mathrm{tr}\mathcal{H}\right\rangle \hat{n}\right),\nonumber \end{align} \begin{align} \left(\mathcal{K}_{j}^{\:\:j}-K_{j}^{\;j}+\Theta\left(\hat{n}\right)-\frac{1}{2}\,^{3}q_{kj}\left(n\left[\,^{3}q^{kj}\right]+\Delta^{kj}+\Delta^{jk}\right)\right)\partial_{i}\label{eq:hyp3}\\ +\left(K_{i}^{\;j}-\Delta_{\;\:i}^{j}\right)\partial_{j}+\left(\frac{1}{2}\,^{3}q_{jk}\,^{3}\nabla_{i}\,^{3}q^{jk}-\Delta_{i}\right)\hat{n} & =\kappa\left(\mathcal{H}\left(\partial_{i},\hat{n}^{*}\right)+\frac{1}{3}\left\langle \hat{n},\mathrm{tr}\mathcal{H}\right\rangle \partial_{i}\right),\nonumber \end{align} \begin{align} \left(\mathcal{K}_{i}^{\:\:j}-K_{\;i}^{j}\right)\hat{n}+\left(\Delta^{j}-\alpha^{j}-\Theta^{j}+\,^{3}\nabla_{k}\,^{3}q^{jk}-\frac{1}{2}\,^{3}q_{kl}\,^{3}\nabla^{j}\,^{3}q^{kl}\right)\partial_{i}\label{eq:hyp4}\\ +\left(\left(\Theta_{i}+\frac{1}{2}\,^{3}q_{lm}\,^{3}\nabla_{i}\,^{3}q^{lm}\right)\,^{3}q^{jk}-\,^{3}\nabla_{i}\,^{3}q^{jk}\right)\partial_{k}+\,^{3}q^{jk}\,^{3}T\left(\partial_{i},\partial_{k}\right) & =\kappa\left(\mathcal{H}\left(\partial_{i},dx^{j}\right)-\frac{1}{3}\left(\delta_{i}^{j}\mathrm{tr}\mathcal{H}-\left\langle dx^{j},\mathrm{tr}\mathcal{H}\right\rangle \partial_{i}\right)\right),\nonumber \end{align} with $\mathrm{tr}\mathcal{H}=\mathcal{H}\left(\partial_{\mu},dx^{\mu}\right)=\mathcal{H}_{\:\:\mu}^{\alpha\,\:\mu}\partial_{\alpha}$. These are the (3+1) hypermomentum equations in normal coordinates. Notice that the quantity $\mathcal{H}\left(U,V^{*}\right)\in T_{p}\mathcal{M}$ is a vector and each equation still has the normal and parallel parts with respect to $\Sigma$. Moreover, one could contract (\ref{eq:hyp1})-(\ref{eq:hyp4}) with $\hat{n}^{*}$ and $dx^{i}$ to obtain 8 equations. \subsubsection*{The Traceless Torsion Constraint Decomposition} As explained in the previous sections, the hypermomentum equation (\ref{eq:EoM2}) only provides 60 equations; one needs 4 more equations to reduce the vectorial degrees of freedom in the connection. These are provided by the traceless torsion constraint (\ref{eq:EoM3}), which could be decomposed (in normal coordinate) as follows. Notice that $\boldsymbol{T}=T_{\nu}dx^{\nu}\in T_{p}^{*}\mathcal{M}$ is a 1-form: \[ T_{\nu}=T_{\mu\;\nu}^{\;\mu}=g\left(dx^{\mu},T\left(\partial_{\mu},\partial_{\nu}\right)\right)=0. \] By contracting $\boldsymbol{T}$ with the normal $\hat{n}$ and $\partial_{i}$, then using torsion decomposition (\ref{eq:t1})-(\ref{eq:t3}), one obtains (in normal coordinate): \begin{align} & \left\langle \boldsymbol{T},\hat{n}\right\rangle =n^{\nu}T_{\nu}=n^{\nu}T_{\mu\;\nu}^{\;\mu}=\Delta_{\;\:i}^{i}-\mathcal{K}_{i}^{\:\:i}=0,\label{eq:proj1}\\ & \left\langle \boldsymbol{T},\partial_{i}\right\rangle =T_{i}=T_{\mu\;i}^{\;\mu}=\Delta_{i}-\Theta_{i}+\,^{3}T_{j\;i}^{\;j}=0.\label{eq:proj2} \end{align} \subsubsection*{The Zero Hypermomentum Case} When the hypermomentum is zero, the hypermomentum equation (\ref{eq:EoM2}), together with the traceless torsion constraint (\ref{eq:EoM3}) must give the condition for the Levi-Civita connection. Let us check if the (3+1) equations (\ref{eq:hyp1})-(\ref{eq:hyp4}) agree with this fact. Setting $\mathcal{H}=0$ and then contracting (\ref{eq:hyp1})-(\ref{eq:hyp4}) with $\hat{n}^{*}$ and $dx^{i}$, we obtain 8 equations, which include 1 scalar equation: \begin{equation} \mathcal{K}_{i}^{\:\:i}-K_{i}^{\;i}=0,\label{eq:sclr} \end{equation} 3 vector equations: \begin{align} & \Delta^{i}-\alpha^{i}=0,\label{eq:vect1}\\ & \Delta^{i}-2\Theta^{i}+\,^{3}\nabla_{j}\,^{3}q^{ji}-\frac{1}{2}\,^{3}q_{jk}\,^{3}\nabla^{i}\,^{3}q^{jk}=0,\label{eq:vect2}\\ & \frac{1}{2}\,^{3}q_{jk}\,^{3}\nabla_{i}\,^{3}q^{jk}-\Delta_{i}=0,\label{eq:vect3} \end{align} 3 matrix equations: \begin{align} & \,^{3}q^{ij}\left(\Theta\left(\hat{n}\right)+\frac{1}{2}\,^{3}q_{kl}\left(n\left[\,^{3}q^{kl}\right]+\Delta^{kl}+\Delta^{lk}\right)\right)-\mathcal{K}^{ij}-n\left[\,^{3}q^{ij}\right]-\Delta^{ij}=0,\label{eq:mat1}\\ & K_{i}^{\;j}-\Delta_{\;\:i}^{j}+\left(\mathcal{K}_{k}^{\:\:k}-K_{k}^{\;k}+\Theta\left(\hat{n}\right)-\frac{1}{2}\,^{3}q_{kl}\left(n\left[\,^{3}q^{kl}\right]+\Delta^{kl}+\Delta^{lk}\right)\right)\delta_{i}^{j}=0,\label{eq:mat2}\\ & \mathcal{K}_{i}^{\:\:j}-K_{\;i}^{j}=0,\label{eq:mat3} \end{align} and 1 tensor equation of order $\tbinom{2}{1}$: \begin{align} \left(\Delta^{j}-\alpha^{j}-\Theta^{j}+\,^{3}\nabla_{l}\,^{3}q^{jl}-\frac{1}{2}\,^{3}q_{lm}\,^{3}\nabla^{j}\,^{3}q^{lm}\right)\delta_{i}^{k}+\,^{3}q^{jl}\,^{3}T_{i\;\:l}^{\:k}\label{eq:tensor}\\ +\left(\Theta_{i}+\frac{1}{2}\,^{3}q_{lm}\,^{3}\nabla_{i}\,^{3}q^{lm}\right)\,^{3}q^{jk}-\,^{3}\nabla_{i}\,^{3}q^{jk} & =0.\nonumber \end{align} The traceless torsion constraint (\ref{eq:EoM3}) provides 1 scalar equation (\ref{eq:proj1}) and 1 vector equation (\ref{eq:proj2}). These equations could be simplified as follows. Solving the vector equations (\ref{eq:vect2})-(\ref{eq:vect3}) for $\Theta^{i}$ and $\Delta^{i}$ gives: \begin{align} \Theta^{i} & =\frac{1}{2}\,^{3}\nabla_{j}\,^{3}q^{ji},\label{eq:vect2a}\\ \Delta^{i} & =\frac{1}{2}\,^{3}q_{jk}\,^{3}\nabla^{i}\,^{3}q^{jk},\label{eq:vect3a} \end{align} while solving the matrix equations (\ref{eq:mat1})-(\ref{eq:mat2}) gives: \begin{align} & \Theta\left(\hat{n}\right)=\frac{1}{6}\,^{3}q_{kl}\nabla_{n}\,^{3}q^{kl},\label{eq:mata}\\ & K^{ij}-\mathcal{K}^{ij}=\nabla_{n}\,^{3}q^{ij}-\frac{1}{3}\,^{3}q^{ij}\,^{3}q_{kl}\nabla_{n}\,^{3}q^{kl},\label{eq:matb} \end{align} with the help of the scalar equation (\ref{eq:sclr}). At last, using the vector equations (\ref{eq:vect1}), (\ref{eq:vect2a}), and (\ref{eq:vect3a}) to (\ref{eq:tensor}), then decomposing this tensor equation into the symmetric and antisymmetric parts of the $\left(i,j\right)$-indices, gives: \begin{align} & \,^{3}\nabla^{\left(i\right.}\,^{3}q^{\left.j\right)k}-\,^{3}q^{k\left(i\right.}\,^{3}\nabla_{l}\,^{3}q^{\left.j\right)l}=0,\label{eq:tensorsym}\\ & \,^{3}T^{ikj}=\,^{3}q^{k\left[i\right.}\,^{3}q_{lm}\,^{3}\nabla^{\left.j\right]}\,^{3}q^{lm}+\,^{3}\nabla^{\left[i\right.}\,^{3}q^{\left.j\right]k}.\label{eq:tensorantsym} \end{align} Let us solve these simplified equations. The easiest way is to start from (\ref{eq:tensorsym}), which is satisfied if: \begin{equation} \,^{3}\nabla^{i}\,^{3}q^{jk}=0.\label{eq:1} \end{equation} Inserting (\ref{eq:1}) to (\ref{eq:tensorantsym}) gives: \begin{equation} \,^{3}T^{ikj}=0,\label{eq:tors1} \end{equation} while inserting (\ref{eq:1}) to vector equation (\ref{eq:vect2a}) gives: \begin{equation} \Theta^{i}=0.\label{eq:2-1} \end{equation} Inserting (\ref{eq:tors1}) to (\ref{eq:proj2}) gives: \begin{equation} \Delta_{i}-\Theta_{i}=0.\label{eq:tors2} \end{equation} Now let us focus on the matrix equation (\ref{eq:mat3}) and (\ref{eq:matb}). Applying the symmetries decomposition to the indices $(i,j)$ gives the following conditions: \begin{align} & K^{\left(ij\right)}-\mathcal{K}^{\left(ij\right)}=0,\label{eq:p}\\ & K^{\left[ij\right]}+\mathcal{K}^{\left[ij\right]}=0,\label{eq:q}\\ & K^{\left(ij\right)}-\mathcal{K}^{\left(ij\right)}=\nabla_{n}\,^{3}q^{ij}-\frac{1}{3}\,^{3}q^{ij}\,^{3}q_{kl}\nabla_{n}\,^{3}q^{kl},\label{eq:r}\\ & K^{\left[ij\right]}-\mathcal{K}^{\left[ij\right]}=0,\label{eq:s} \end{align} where (\ref{eq:p})-(\ref{eq:q}) comes from (\ref{eq:mat3}), while (\ref{eq:r})-(\ref{eq:s}) comes from (\ref{eq:matb}). Comparing (\ref{eq:q}) and (\ref{eq:s}), then $K^{\left[ij\right]}=\mathcal{K}^{\left[ij\right]}=0$, which means that both the extrinsic curvatures do not possess antisymmetric parts. On the other hand, from (\ref{eq:p}) and (\ref{eq:r}), we obtain $\mathcal{K}^{\left(ij\right)}=K^{\left(ij\right)},$ that are equivalent with the following conditions: \begin{align} K^{ij} & =\mathcal{K}^{ij},\label{eq:6-1}\\ K^{ij} & =K^{ji},\label{eq:7-1} \end{align} and: \begin{align*} \nabla_{n}\,^{3}q^{ij}-\frac{1}{3}\,^{3}q^{ij}\,^{3}q_{kl}\nabla_{n}\,^{3}q^{kl}=0, \end{align*} which is satisfied if: \begin{equation} \nabla_{n}\,^{3}q^{ij}=0.\label{eq:8-1} \end{equation} Inserting (\ref{eq:8-1}) to (\ref{eq:mata}) gives: \begin{equation} \Theta\left(\hat{n}\right)=0.\label{eq:9-1} \end{equation} Finally, using (\ref{eq:q3}) and inserting (\ref{eq:sclr}), (\ref{eq:8-1}), and (\ref{eq:9-1}) to (\ref{eq:mat2}) gives: \begin{equation} \mathcal{K}_{i}^{\;j}-\Delta_{\;\:i}^{j}=0.\label{eq:10} \end{equation} The conditions obtained from solving the trivial hypermomentum equation could be classified as follows: \{(\ref{eq:tors1}),(\ref{eq:tors2}), (\ref{eq:7-1}), (\ref{eq:10})\} and \{(\ref{eq:vect1}), (\ref{eq:1}), (\ref{eq:2-1}), (\ref{eq:6-1}), (\ref{eq:8-1}), (\ref{eq:9-1})\} are, respectively, the torsionless and metric condition (\ref{eq:tor1}) and (\ref{eq:met1})-(\ref{eq:met}) in the normal (Gauss) coordinate where $N=1$ and $\boldsymbol{N}=0$. Hence, for zero hypermomentum, the affine connection becomes Levi-Civita, and the (3+1) stress-energy-momentum equations for MAGR (\ref{eq:satua})-(\ref{eq:tigaa}) return to the original EFE (\ref{eq:nic1})-(\ref{eq:nic}). Notice that without the traceless torsion constraint (\ref{eq:proj1})-(\ref{eq:proj2}), it is not possible to retrieve the metric compatibility and torsionless condition, in the absence of hypermomentum. A more detailed analysis of the hypermomentum equation will be discussed in our companion article. \subsection{Conclusions and Further Remarks} In this subsection, we summarize the results achieved from our work. First, we have derived the generalized Gauss-Codazzi-Mainardi equations and torsion (3+1) decomposition, which are valid for manifold with any dimension. This is done in Section II with equation (\ref{eq:e}) and (\ref{eq:torsion1})-(\ref{eq:torsion2}) (or equation (\ref{eq:gcm1})-(\ref{eq:gcm2}) and (\ref{eq:T1})-(\ref{eq:T2}), written in terms of components) as the results. Second, we have shown that a metric connection $\nabla$ on $\left(\mathcal{M},g\right),$ induce a connection $\,^{3}\nabla$ on the hypersurface $\left(\Sigma,\,^{3}q\right),$ which also satisfies the metric compatibility with respect to $\,^{3}q$. A similar statement is also valid for a torsionless connection. These had been proven in Section III. Third, having the previous results in hand, we have performed the ADM formulation to the generalized EFE, $\mathbf{Ric}-\frac{1}{2}\boldsymbol{g}\mathcal{R}=\kappa\mathcal{\boldsymbol{T}},$ for Metric-Affine General Relativity, resulting in 3 parts of the equation: the energy equation (\ref{eq:satu}), momentum equation (\ref{eq:dua}), and stress-energy equation (\ref{eq:tiga-1}), with the help of the additional variables on the hypersurface introduced in Section IV. Unlike the standard GR case, the energy and momentum part of the Einstein field equations are not constraint equations, since all of the 3 equations contain the derivative of quantities with respect to the time coordinate. We have also shown in the discussion that by applying metric compatibility (\ref{eq:met1}) and torsionless condition (\ref{eq:tor1}) to (\ref{eq:satu})-(\ref{eq:tiga-1}), one will recover the Hamiltonian and momentum (or diffeomorphism) constraint, together with the standard dynamics of GR. For the completeness of the discussion in this article, we present some results of our companion paper, which is the (3+1) decomposition (in normal coordinates) of the hypermomentum equation (\ref{eq:hyp1})-(\ref{eq:hyp4}) and the traceless torsion constraint (\ref{eq:proj1})-(\ref{eq:proj2}). We have shown that in the (3+1) perspective, setting hypermomentum to zero, one will recover the metric compatibility and torsionless condition, hence, the connection must be Levi-Civita. These conditions, in MAGR (and MAG) is not \textit{a priori} given but as a consequence of the absence of hypermomentum. However, the tracelessness of torsion in (\ref{eq:ELmagr31}) does not result from the equation of motion, it is introduced in the kinematical level, via the projective transformation. If one expects to derive the tracelessness of torsion from the dynamics, the action must be added with new terms at least quadratic in torsion \cite{Percacci}. This will be an interesting subject to pursue. Another remark is concerning the way to break the projective invariance. As we had mentioned at the beginning of Section IV, fixing the torsion to be traceless is only one of the possibilities to break the invariance. It is also possible to choose the non-metricity factor, or moreover, a linear combination of the non-metricity factor and torsion to be traceless, as discussed in \cite{Iosifidis4,Smalley}. These choices might result in different (3+1) decompositions, in particular, the form of equation (\ref{eq:proj1})-(\ref{eq:proj2}), and will also be an interesting subject for further works. As we had already mentioned in the discussion, in this article, the focus is on the first equation of motion (\ref{eq:EoM1}); we are writing another article that focuses on the hypermomentum equation and the projective invariant constraint, as a companion to this article. With a complete ADM decomposition to these equations of motions as well, it will be possible to do a complete analysis on the (3+1) formulation of MAGR, in particular, the Cauchy problem and/or the Hamiltonian analysis of the theory. As further works, it will be interesting to perform the (3+1) ADM formulation to more general theories such as Metric-Affine $f\left(\mathcal{R}\right)$-Gravity. We expect our preliminary work in this article could motivate more research in this direction. \section*{Acknowledgement} This research was funded by the P3MI-ITB and PDUPT DIKTI programs.

b82f866b90ee4ddd48e9beff8ac46e766b6f6cdf

\section*{Introduction} The definition of the Lipschitz saturation of an ideal appears in \cite{G2}, in the context of bi-Lipschitz equisingularity. The study of bi-Lipschitz equisingularity was started by Zariski \cite{Z}, Pham and Teissier \cite{PT,P1}, and was further developed by Lipman \cite{L}, Mostowski \cite{M1,M2}, Parusinski \cite{PA1,PA2}, Birbrair \cite{B} and others. The Lipschitz Saturation and the double of an ideal $I$, denoted $I_S$ and $I_D$, respectively, were defined in \cite{G2}, where $I$ is a sheaf of ideals of ${\mathcal O}_X$, the analytic local ring of an analytic variety $X$. The ideal $I_S$ consists of elements in ${\mathcal O}_X$ whose quotient of its pullback by the blowup-map, with a local generator of the pullback of $I$ is Lipschitz. The double $I_D$ is the submodule of ${\mathcal O}_{X\times X}^2$ generated by $(h\circ\pi_1,h\circ\pi_2)$, $h\in I$, where $\pi_1,\pi_2:X\times X\rightarrow X$ are the projections. Theorem 2.3 of \cite{G2} gives an equivalence between $I_S$ and the integral closure of $I_D$. In \cite{SG} the authors generalize the notion of the double for a sheaf of modules ${\mathcal M}$, and the Genericity Theorem is generalized for an arbitrary family of analytic varieties. They also get an interesting decomposition of the stalk of ${\mathcal M}_D$ at $(x,x')$, as the direct sum of the stalks ${\mathcal M}_{x}$ and ${\mathcal M}_{x'}$, as long $x\neq x'$. If ${\mathcal M}$ is the jacobian module of a family of analytic varieties, the stalks at $x$ and $x'$ determine the tangent hyperplanes at these two points. Since, to control the Lipschitz behavior of the tangent hyperplanes to $X$, it is natural to look for a sheaf on $X\times X$ whose stalks determine the tangent hyperplanes at each pair of distinct points, it is natural to consider the double of the jacobian module. This paper is organized as follows: In section \ref{sec0} we recall some basic background material. In section \ref{sec3} we present some definitions for the Lipschitz Saturation of a module, inspired by Theorem 2.3 of \cite{G2}, and also motivated by characterizations of the integral closure of modules coming from the integral closure of ideals. Despite all these definitions of Lipschitz saturation agree on the ideal case, we show these notions are not the same anymore in the module case. Nevertheless, we prove all the definitions of Lipschitz saturation of modules are generically equivalent. In section \ref{sec4} we prove the Necessity theorem, and we relate the infinitesimal Lipschitz condition A coming from the first Lipschitz saturation with the strongly bi-Lipschitz equisingularity and the Mostowski's conditions for a Lipschitz stratification. A new proof of the Genericity Theorem arises using the Necessity Theorem and the Mostowski's results about Lipschitz stratifications. \section*{Acknowledgements} The authors are grateful to Nivaldo Grulha Jr. for his careful reading of this work and to Maria Aparecida Soares Ruas for the valuable conversations about the subject of this work, especially in the generic equivalence among the Lipschitz saturations defined here. The first author was supported in part by PVE-CNPq, grant 401565/2014-9. The second author was supported by Funda\c{c}\~ao de Amparo \`a Pesquisa do Estado de S\~ao Paulo - FAPESP, Brazil, grant 2013/22411-2. \section{Background on Lipschitz Saturation of Ideals and Integral Closure of Modules}\label{sec0} The Lipschitz saturation of a local ring was defined by Pham and Teissier in \cite{PT}. \begin{definition} Let $I$ be an ideal of ${\mathcal O}_{X,x}$, $SB_I(X)$ the saturation of the blow-up and $\pi_S: SB_I(X)\rightarrow X$ the projection map. The {\bf Lipschitz saturation} of the ideal $I$ is denoted $I_S$, and is the ideal $I_S:=\{h\in{\mathcal O}_{X,x}\mid\pi_S^*(h)\in\pi_S^*(I)\}$. \end{definition} Since the normalization of a local ring $A$ contains the Lipschitz Saturation of $A$, it follows that $I\subseteq I_S\subseteq\overline{I}$. In particular, if $I$ is integrally closed then $I_S=\overline{I}$. This definition can be given an equivalent statement using the theory of integral closure of modules. Since Lipschitz conditions depend on controlling functions at two different points as the points come together, we should look for a sheaf defined on $X\times X$. We describe a way of moving from a sheaf of ideals on $X$ to a sheaf on $X\times X$. Let $\pi_1,\pi_2: X\times X\rightarrow X$ be the projections to the i-th factor, and let $h\in{\mathcal O}_{X,x}$. Define $h_D\in{\mathcal O}_{X\times X,(x,x)}^{2}$ as $(h\circ\pi_1,h\circ\pi_2)$, called the double of $h$. We define the double of the ideal $I$, denoted $I_D$, as the submodule of ${\mathcal O}_{X\times X,(x,x)}^2$ generated by $h_D$, where $h$ is an element of $I$. We can see in \cite{G2}, the following result gives a link between Lipschitz saturation and integral closure of modules. \begin{theorem}[\cite{G2}, Theorem 2.3] Suppose $(X,x)$ is a complex analytic set germ, $I\subseteq{\mathcal O}_{X,x}$ and $h\in{\mathcal O}_{X,x}$. \noindent Then $h\in I_S$ if, and only if, $h_D\in\overline{I_D}$. \end{theorem} Using the Lipschitz saturation of ideals (and doubles), in \cite{G1}, we defined the infinitesimal Lipschitz conditions for hypersurfaces. Let us recall two results about the integral closure of modules which will inspire good definitions for Lipschitz saturation of modules. \vspace{0,5cm} {\it The ideal sheaf $\rho({\mathcal M})$ on $X\times {\mathbb{P}}^{p-1}$ associated to a submodule sheaf ${\mathcal M}$ of ${\cO_X^p}$ (see \cite{GK})}: Given $h=(h_1,...,h_p)\in{\cO_X^p}$ and $(x,[T_1,...,T_p])\in X\times{\mathbb{P}}^{p-1}$, with $T_i\neq0$, we define $\rho(h)$ as the germ of the analytic map given by $\sum\limits_{j=1}^p h_j(z)\frac{T_j}{T_i}$ which is well-defined on a Zariski open subset of $X\times{\mathbb{P}}^{p-1}$ that contains the point $(x,[T_1,...,T_p])$. We define $\rho({\mathcal M})$ as the ideal generated by $\{\rho(h)\mid h\in {\mathcal M}\}$. The next result, proved in \cite{GK}, gives a strong relation between the integral closure of modules and ideals. \begin{proposition}[\cite{GK}, Proposition 3.4]\label{proposition GK} Let $h\in{\mathcal O}_{X,x}^p$. Then $h\in\overline{{\mathcal M}}$ at $x$ if, and only if, $\rho(h)\in\overline{\rho({\mathcal M})}$ at all point $(x,[T_1,...,T_p])\in V(\rho({\mathcal M}))$. \end{proposition} In \cite{G3} there is another way to make a link between the integral closure of modules and ideals, using minors of a matrix of generators of ${\mathcal M}$. Let ${\mathcal M}$ be a submodule sheaf of ${\cO_X^p}$, and $[{\mathcal M}]$ a matrix of generators of ${\mathcal M}$. For each $k$, let $J_k({\mathcal M})$ denote the ideal of ${\mathcal O}_X$ generated by the $k\times k$ minors of $[{\mathcal M}]$. If $h\in{\cO_X^p}$, let $(h,{\mathcal M})$ be the submodule generated by $h$ and ${\mathcal M}$. \begin{proposition}[\cite{G3}, Corollary 1.8]\label{proposition G} Suppose $(X,x)$ is a complex analytic germ with irreducible components $\{V_i\}$. Then, $h\in\overline{{\mathcal M}}$ at $x$ if, and only if, $J_{k_i}((h,{\mathcal M}_i))\subseteq \overline {J_{k_i}({\mathcal M}_i)}$ at $x$, where ${\mathcal M}_i$ is the submodule of ${\mathcal O}_{V_i,x}^p$ induced from ${\mathcal M}$ and $k_i$ is the generic rank of $(h,{\mathcal M}_i)$ on $V_i$. \end{proposition} Now we recall some basic results about the double of a module developed in \cite{SG}. Let $X \subseteq \mathbb{C}^n$ be an analytic space, and let ${\mathcal M}$ be an ${\mathcal O}_X $-submodule of ${\mathcal O}_X^p$. Consider the projection maps $\pi_1,\pi_2: X \times X \rightarrow X$. We assume that ${\mathcal M}$ is finitely generated by global sections. \begin{definition} Let $h\in{\cO_X^p}$. The double of $h$ is defined as the element \noindent $h_D:=(h\circ\pi_1,h\circ\pi_2) \in{\mathcal O}_{X\times X}^{2p}$. The double of ${\mathcal M}$ is denoted by ${\mathcal M}_D$, and is defined as the ${\mathcal O}_{X\times X}$-submodule of ${\mathcal O}_{X\times X}^{2p}$ generated by $\{h_D \mid h\in {\mathcal M}\}$. \end{definition} Consider $z_1,...,z_n$ the coordinates on $\mathbb{C}^n$. The next lemma is a useful tool to deal with the double. \begin{lemma}[\cite{SG}, Lemma 2.2]\label{L1} Suppose $\alpha\in{\mathcal O}_X$ and $h\in{\cO_X^p}$. Then: \begin{enumerate} \item $(\alpha h)_D=-(0,(\alpha\circ\pi_1-\alpha\circ\pi_2)(h\circ\pi_2))+(\alpha\circ\pi_1)h_D$; \item $(0,(\alpha\circ\pi_1-\alpha\circ\pi_2)(h\circ\pi_2))\in {\mathcal M}_D$, for all $h\in {\mathcal M}$ and $\alpha\in{\mathcal O}_X$; \item $\alpha\circ\pi_1-\alpha\circ\pi_2 \in I(\Delta(X))=(z_1\circ\pi_1-z_1\circ\pi_2,\ldots, z_n\circ\pi_1-z_n\circ\pi_2)$, for all $\alpha\in{\mathcal O}_X$; \item $(g+h)_D=g_D+h_D$, for all $g,h\in{\cO_X^p}$. \end{enumerate} \end{lemma} In \cite{SG} we see that it is possible to obtain a set of generators for ${\mathcal M}_D$ from a set of generators of ${\mathcal M}$. \begin{proposition}[\cite{SG}, Proposition 2.3]\label{P2} Suppose that ${\mathcal M}$ is generated by $\{h_1,\ldots,h_r \}$. Then, the following sets are generators of ${\mathcal M}_D$: \begin{enumerate} \item ${\mathcal B}=\{(h_1)_D,\ldots,(h_r)_D\} \cup \{(0_{{\mathcal O}_{X\times X}^{p}},(z_i\circ\pi_1-z_i\circ\pi_2)(h_j\circ\pi_2))$ $|$ $i\in\{1,\ldots,n\}$ and $j\in\{1,\ldots,r\}\}$. \item ${\mathcal B}'=\{(h_1)_D,\ldots,(h_r)_D\} \cup \{((z_i\circ\pi_1-z_i\circ\pi_2)(h_j\circ\pi_1),0_{{\mathcal O}_{X\times X}^{p}}))$ $|$ $i\in\{1,\ldots,n\}$ and $j\in\{1,\ldots,r\}\}$. \item ${\mathcal B}''=\{(h_1)_D,\ldots,(h_r)_D\} \cup \{(z_ih_j)_D$ $|$ $i\in\{1,\ldots,n\}$ and $j\in\{1,\ldots,r\}\}$. \end{enumerate} \end{proposition} The following proposition gives a link between the integral closure of the double of a module and the integral closure of the original module. \begin{proposition}[\cite{SG}, Proposition 2.9]\label{P12} Let $h\in{\cO_X^p}$. \begin{enumerate} \item If $h_D\in\overline{{\mathcal M}_D}$ at $(x,x')$ then $h\in\overline{{\mathcal M}}$ at $x$ and $x'$. \item If $h_D\in ({\mathcal M}_D)^{\dagger}$ at $(x,x')$ then $h\in {\mathcal M}^{\dagger}$ at $x$ and $x'$. \end{enumerate} \end{proposition} In the next theorem we compute the generic rank of the double of a module. \begin{theorem}[\cite{SG}, Proposition 2.5]\label{T2.9} Let $(X,x)$ be an irreducible analytic complex germ of dimension $d\ge 1$, and ${\mathcal M}\subseteq{\mathcal O}_{X,x}^p$ a submodule of generic rank $k$. Then ${\mathcal M}_D$ has generic rank $2k$ at $(x,x)$. \end{theorem} \begin{corollary}[\cite{SG}, Corollary 2.6]\label{C2.10} Let $\{V_i\}$ be the irreducible components of $(X,x)$. For each $i$, if ${\mathcal M}$ has generic rank $k_i$ on $V_i$ then ${\mathcal M}_D$ has generic rank $2k_i$ on $V_i\times V_i$. In particular, if ${\mathcal M}$ has generic rank $k$ on each component of $X$ then ${\mathcal M}_D$ has generic rank $2k$ on each component of $X\times X$. \end{corollary} We end this section recalling an important result which states that the double of a sheaf of modules ${\mathcal M}$ carries all the information at $(x,x')$ as the stalks of ${\mathcal M}$ do at $x$ and $x'$, as long $x\neq x'$. \begin{proposition}[\cite{SG}, Proposition 2.11]\label{P2.13} Let ${\mathcal M}\subseteq{\mathcal O}_X^p$ be a sheaf of submodules. Consider $(x,x')\in X\times X$ with $x\neq x'$. Then: $${\mathcal M}_D=({\mathcal M}_x\circ\pi_1)\oplus ({\mathcal M}_{x'}\circ\pi_2)$$ at $(x,x')$. \end{proposition} Proposition \ref{P2.13} provides additional motivation for the idea of the double: In order to control the Lipschitz behavior of pairs of tangent planes at two different points $x$ and $x'$ of a family ${\mathcal X}$, it is helpful to have each module which determines the tangent hyperplanes at each point as part of the construction. Furthermore, this proposition shows that $JM({\mathcal X})_D$ at $(x,x')$ contains both $JM({\mathcal X})_x$ and $JM({\mathcal X})_{x'}$. \section{The Lipschitz Saturation of a Module}\label{sec3} We want extend to modules the notion of Lipchitz saturation that was defined in \cite{G2} for ideals. Motivated by some equivalent descriptions that the Lipschitz saturation has in the ideal case, we define corresponding versions of the Lipschitz saturation for modules, and explore the extent to which they are equivalent. The main motivation of the next definition is Theorem 2.3 of \cite{G2}. Let $X\subseteq {\mathbb{C}}^n$ be an analytic set and ${\mathcal M}$ be an ${\mathcal O}_X$-submodule of ${\cO_X^p}$. \begin{definition} The {\bf 1-Lipschitz saturation of ${\mathcal M}$ at $x\in X$} is denoted by $({\mathcal M}_{S_1})_x$, and is defined by $$({\mathcal M}_{S_1})_x:=\{h\in{\mathcal O}_{X,x}^{p}\mid h_D\in\overline{{\mathcal M}_D}\mbox{ at }(x,x)\}.$$ In the family case, the {\bf 1-Lipschitz saturation of ${\mathcal M}$ at $x\in X$ relative to $Y$} is defined as above taking the double relative to $Y$. \end{definition} \begin{proposition}\label{P15} Let ${\mathcal M}$ be a sheaf of ${\mathcal O}_X$submodules of ${\cO_X^p}$. \begin{enumerate} \item ${\mathcal M}_{S_1}$ is an ${\mathcal O}_X$-submodule of ${\cO_X^p}$; \item ${\mathcal M}\subseteq {\mathcal M}_{S_1} \subseteq \overline{{\mathcal M}}$. In particular, ${\mathcal M}$ is a reduction of ${\mathcal M}_{S_1}$ and $e({\mathcal M},{\mathcal M}_{S_1})=0$. \end{enumerate} The same result holds in the family case. \end{proposition} \begin{proof} Let $x\in X$ be an arbitrary point. (1) Let $\alpha\in{\mathcal O}_{X,x}$ and $h,h'\in {\mathcal M}_{S_1}$ at $x$. Since $h_D\in\overline{{\mathcal M}_D}$ then by Proposition \ref{P12} (1) we have that $h\in\overline{{\mathcal M}}$ at $x$. Thus, $(0,(\alpha\circ\pi_1-\alpha\circ\pi_2)(h\circ\pi_2))\in\overline{{\mathcal M}_D}$. Hence $(\alpha h+h')_D=(\alpha\circ\pi_1)h_D+(0,(\alpha\circ\pi_1-\alpha\circ\pi_2)(h\circ\pi_2))+h'_D\in\overline{{\mathcal M}_D}$. (2) If $h\in {\mathcal M}$ at $x$ then $h_D\in {\mathcal M}_D\subseteq\overline{{\mathcal M}_D}$ at $(x,x)$, so $h\in {\mathcal M}_{S_1}$ at $x$. Therefore, ${\mathcal M}\subseteq {\mathcal M}_{S_1}$. Furthermore, if $h\in {\mathcal M}_{S_1}$ at $x$ then $h_D\in\overline{{\mathcal M}_D}$ at $(x,x)$, and by Proposition \ref{P12} (1) we have that $h\in\overline{{\mathcal M}}$ at $x$. Therefore, ${\mathcal M}_{S_1}\subseteq\overline{{\mathcal M}}$. \end{proof} In order to define the second Lipschitz saturation, let us fix some notations. For each $\psi:X\rightarrow \mbox{Hom}({\mathbb{C}}^p,{\mathbb{C}})$, $\psi=(\psi_1,...,\psi_p)$ and $h=(h_1,...,h_p)\in{\cO_X^p}$, we define $\psi\cdot h\in{\mathcal O}_X$ given by $(\psi\cdot h)(z):=\sum\limits_{i=1}^{p}\psi_i(z)h_i(z)$. We define $\psi\cdot {\mathcal M}$ as the ideal of ${\mathcal O}_X$ generated $\{\psi\cdot h \mid h\in {\mathcal M}\}$. \begin{lemma}\label{L16} \begin{enumerate} \item $\psi\cdot(\alpha g+h)=\alpha(\psi\cdot g)+(\psi\cdot h)$, $\forall g,h\in {\cO_X^p}$ and $\alpha\in{\mathcal O}_X$. \item If ${\mathcal M}$ is generated by $\{h_1,...,h_r\}$ then $\psi\cdot {\mathcal M}$ is generated by $\{\psi\cdot h_1,...,\psi\cdot h_r\}$. \end{enumerate} \end{lemma} \begin{proof} It is easy to see that (2) is a straightforward consequence of (1). Now, write $g=(g_1,...,g_p)$ and $h=(h_1,...,h_p)$. Then, for every $z$ we have $(\psi\cdot(\alpha g+h))(z)=\sum\limits_{i=1} ^{p}\psi_i(z)(\alpha(z)g_i(z)+h_i(z))=\alpha(z)\sum\limits_{i=1}^{p}\psi_i(z)g_i(z)+\sum\limits_{i=1}^{p}\psi_i(z)h_i(z)=(\alpha(\psi\cdot g)+(\psi\cdot h))(z)$. \end{proof} Let us fix some notation for the minors of a matrix. Let $k\in{\mathbb{N}}$ and let $A$ be a matrix. If $I=(i_1,...,i_k)$ and $J=(j_1,...,j_k)$ are $k$-indexes, $A_{IJ}$ is defined as the $k\times k$ submatrix of $A$ formed by the rows $i_1,...,i_k$ and columns $j_1,...,j_k$ of $A$. We denote $J_{IJ}(A):=\det(A_{IJ})$. \begin{lemma}\label{L21} Suppose that ${\mathcal M}$ has generic rank $k$ in each component of $X$. If $I=(i_1,...,i_k)$ and $J=(j_1,...,j_k)$ are indexes with $j_1=1$ then there exists $\psi: X\rightarrow \mbox{Hom}({\mathbb{C}}^p,{\mathbb{C}})$ such that: \begin{enumerate} \item $\psi\cdot h=J_{IJ}(h,{\mathcal M})$, $\forall h\in{\mathcal O}_{X,x}^p$; \item $\psi\cdot {\mathcal M}\subseteq J_k({\mathcal M})$. \end{enumerate} \end{lemma} \begin{proof} Let us fix a matrix of generators of ${\mathcal M}$, $[{\mathcal M}]=$ $ \begin{bmatrix} | & & | \\ g_1 & \ldots & g_r \\ | & & | \end{bmatrix}$. We have $J_{IJ}(h,{\mathcal M})=\det [h,{\mathcal M}]_{IJ}$, where $$[h,{\mathcal M}]_{IJ}=\begin{bmatrix} h_{i_1} & g_{i_1,j_2-1} & \ldots & g_{i_1,j_k-1} \\ \vdots & \vdots & & \vdots \\ h_{i_k} & g_{i_k,j_2-1} & \ldots & g_{i_k,j_k-1} \end{bmatrix}$$ for all $h=(h_1,...,h_p)$. Let $G_{i_l,j_s-1}$ be the $(l,s)$-cofactor of $[h,{\mathcal M}]_{IJ}$, for all $l,s\in\{1,...,k\}$. Notice that the $(l,1)$-cofactors $G_{i_l,0}$ do not depend of $h$. Then, $J_{IJ}(h,{\mathcal M})=\det[h,{\mathcal M}]_{IJ}=\sum\limits_{l=1}^{k}G_{i_l,0}\cdot h_{i_l}$, for all $h$. Take $\psi: X\rightarrow \mbox{Hom}({\mathbb{C}}^p,{\mathbb{C}})$ given by $(\psi_1,...,\psi_p)$ where $\psi_{i_l}=G_{i_l,0}$, for all $l\in\{1,...k\}$, and $\psi_j=0$, for every index $j$ off $I$. Thus, for all $h\in{\mathcal O}_{X,x}^p$ we get $J_{IJ}(h,{\mathcal M})=\psi_I\cdot h_I=\psi\cdot h$. Now, let $g\in {\mathcal M}$ arbitrary. Then $(g,{\mathcal M})={\mathcal M}$. By (1) we have $\psi\cdot g=J_{I,J}(g,{\mathcal M})\in J_k((g,{\mathcal M}))=J_k({\mathcal M})$, hence $\psi\cdot {\mathcal M}\subseteq J_k({\mathcal M})$. \end{proof} {\bf Remark:} It is enough work with indexes $I=(i_1,...,i_k)$ and $J=(j_1,...,j_k)$ such that $i_1<...<i_k$ and $j_1<...<j_k$. Propositions \ref{proposition GK} and \ref{proposition G} are characterizations for the integral closure of modules using integral closure of ideals. Notice the next proposition is another characterization of the same type. \begin{proposition}\label{P18} Let $h\in{\mathcal O}_{X,x}^p$ and suppose ${\mathcal M}$ has generic rank $k$ on each component of $X$. Then, $h\in\overline{{\mathcal M}}$ at $x$ if, and only if, $\psi\cdot h\in\overline{\psi\cdot {\mathcal M}}$, at $x$, $\forall \psi:X\rightarrow \mbox{Hom}({\mathbb{C}}^p,{\mathbb{C}})$. \end{proposition} \begin{proof} $(\Longrightarrow)$ It is a straightforward consequence of the curve criterion. $(\Longleftarrow)$ The proof now use 1.7 and 1.8 of \cite{G3}. By these results it is enough to check that $J_{IJ}(h,{\mathcal M})\in\overline{J_k({\mathcal M})}$, for all indexes $I$ and $J$. Write $I=(i_1,...,i_k)$ and $J=(j_1,...,j_k)$. Let $[{\mathcal M}]=$ $ \begin{bmatrix} | & & | \\ g_1 & \ldots & g_r \\ | & & | \end{bmatrix}$ be a matrix of generators of ${\mathcal M}$. Write $h=(h_1,...,h_p)$. Then, $[h,{\mathcal M}]=$ $ \begin{bmatrix} | & | & & | \\ h & g_1 & \ldots & g_r \\ | & | & & | \end{bmatrix}$. If $j_1>1$ then $J_{IJ}(h,{\mathcal M})$ is a $k\times k$ minor taken only among the generators of ${\mathcal M}$, hence $J_{IJ}(h,{\mathcal M})\in J_k({\mathcal M})\subseteq\overline{J_k({\mathcal M})}$. Now suppose that $j_1=1$. By Lemma \ref{L21} there exists $\psi:X\rightarrow \mbox{Hom}({\mathbb{C}}^p,{\mathbb{C}})$ such that $\psi\cdot h=J_{I,J}(h,{\mathcal M})$ and $\psi\cdot {\mathcal M}\subseteq J_k({\mathcal M})$. By the hypothesis we have $\psi\cdot h\in\overline{\psi\cdot {\mathcal M}}$, hence $J_{I,J}(h,{\mathcal M})=\psi\cdot h\in\overline{\psi\cdot {\mathcal M}}\subseteq\overline{J_k({\mathcal M})}$. \end{proof} Given $h\in{\mathcal O}_{X,x}^p$ and ${\mathcal M}\subseteq{\mathcal O}_{X,x}^p$ submodule, we can ask if $h\in\overline{{\mathcal M}}$ or $\rho(h)\in\overline{\rho({\mathcal M})}$. For integral closure these questions have the same answer. Then we can ask the analogous questions for the Lipschitz saturation of ${\mathcal M}$. The next version of a Lipschitz saturation is motivated by working with $\rho(h)$ and $\rho({\mathcal M})$. \begin{definition} The {\bf 2-Lipschitz saturation of ${\mathcal M}$ at $x\in X$} is denoted by $({\mathcal M}_{S_2})_x$, and is defined by $$({\mathcal M}_{S_2})_x:=\{h\in{\mathcal O}_{X,x}^{p}\mid\psi\cdot h\in(\psi\cdot {\mathcal M})_S\mbox{ at }x\mbox{, }\forall \psi:X\rightarrow \textrm{Hom}({\mathbb{C}}^p,{\mathbb{C}})\}.$$ In the family case, the {\bf 2-Lipschitz saturation of ${\mathcal M}$ at $x\in X$ relative to $Y$} is defined as above taking the Lipschitz saturation of $\psi\cdot {\mathcal M}$ relative to $Y$. \end{definition} \begin{proposition}\label{P19} Let ${\mathcal M}$ be a sheaf of ${\mathcal O}_X$submodules of ${\cO_X^p}$. \begin{enumerate} \item ${\mathcal M}_{S_2}$ is an ${\mathcal O}_X$-submodule of ${\cO_X^p}$; \item ${\mathcal M}\subseteq {\mathcal M}_{S_2} \subseteq \overline{{\mathcal M}}$. In particular, ${\mathcal M}$ is a reduction of ${\mathcal M}_{S_2}$ and $e({\mathcal M},{\mathcal M}_{S_2})=0$. \end{enumerate} The same result holds in the family case. \end{proposition} \begin{proof} Let $x\in X$ be an arbitrary point. (1) Let $h,h'\in {\mathcal M}_{S_2}$ and $\alpha\in{\mathcal O}_X$ at $x$, and let $\psi:X\rightarrow \mbox{Hom}({\mathbb{C}}^p,{\mathbb{C}})\}$ arbitrary. Then, $\psi\cdot h,\psi\cdot h'\in(\psi\cdot {\mathcal M})_S$ at $x$, and by Lemma \ref{L16} (1) we have $\psi\cdot(\alpha h+h')=\alpha(\psi\cdot h)+\psi\cdot h'\in(\psi\cdot {\mathcal M})_S$ at $x$. Therefore, $\alpha h+h'\in {\mathcal M}_{S_2}$ at $x$. (2) If $h\in {\mathcal M}$ then $\psi\cdot h\in\psi\cdot {\mathcal M}\subseteq(\psi\cdot {\mathcal M})_S$, so $h\in {\mathcal M}_{S_2}$ and ${\mathcal M}\subseteq {\mathcal M}_{S_2}$. Now, let $h\in {\mathcal M}_{S_2}$. Then, $\psi\cdot h\in(\psi\cdot {\mathcal M})_S\subseteq\overline{\psi\cdot {\mathcal M}}$, $\forall\psi:X\rightarrow \mbox{Hom}({\mathbb{C}}^p,{\mathbb{C}})$. By Proposition \ref{P18} we conclude that $h\in\overline{{\mathcal M}}$. \end{proof} Next, we begin to compare the first and second Lipschitz Saturations. \begin{proposition}\label{P23} ${\mathcal M}_{S_1}\subseteq {\mathcal M}_{S_2}$ at every point $x\in X$. \end{proposition} \begin{proof} Let $h=(h_1,...,h_p)\in {\mathcal M}_{S_1}$ at $x$. Then $h_D\in \overline{{\mathcal M}_D}$ at $(x,x)$. We need to check that $(\psi\cdot h)_D\in\overline{(\psi\cdot {\mathcal M})_D}$, for all $\psi:X\rightarrow \mbox{Hom}({\mathbb{C}}^p,{\mathbb{C}})$. Let $\phi=(\phi_1,\phi_2):({\mathbb{C}},0)\rightarrow(X\times X,(x,x))$ be an arbitrary analytic curve. Since $h_D\in\overline{{\mathcal M}_D}$ then we can write $$h_D\circ\phi=\sum\limits_{j}\alpha_j\phi^*((g_j)_D)$$ with $g_j=(g_{1j},...,g_{pj})\in {\mathcal M}$ and $\alpha_j\in{\mathcal O}_{{\mathbb{C}},0}$ for all $j$. Looking to the above equation and comparing the $2p$ coordinates we conclude that $h_i\circ\phi_1=\sum\limits_{j}\alpha_j(g_{ij}\circ\phi_1)$ and $h_i\circ\phi_2=\sum\limits_{j}\alpha_j(g_{ij}\circ\phi_2)$, for all $i\in\{1,...,p\}$. Let $\psi_1,...,\psi_p$ be the coordinate functions of $\psi$. Thus: $(\psi\cdot h)_D\circ\phi=(\sum\limits_{i}(\psi_i\circ\phi_1)\cdot(h_i\circ\phi_1),\sum\limits_{i}(\psi_i\circ\phi_2)\cdot(h_i\circ\phi_2))\\=(\sum\limits_{i,j}(\psi_i\circ\phi_1)\alpha_j(g_{ij}\circ\phi_1),\sum\limits_{i,j}(\psi_i\circ\phi_2)\alpha_j(g_{ij}\circ\phi_2))\\=\sum\limits_{j}\alpha_j((\psi\cdot g_j)_D\circ\phi)\in(\psi\cdot {\mathcal M})_D\circ\phi$. \end{proof} The next definition of Lipschitz saturation is motivated by Proposition \ref{proposition G} which relates $\overline{{\mathcal M}}$ and $\overline{J_k({\mathcal M})}$. \begin{definition} Suppose that ${\mathcal M}\subseteq{\mathcal O}_X^p$ is an ${\mathcal O}_{X}$-submodule of generic rank $k$ on each component of $X$. The \textbf{3-Lipschitz saturation of ${\mathcal M}$ at $x\in X$} is denoted by $({\mathcal M}_{S_3})_x$, and is defined by $$({\mathcal M}_{S_3})_x:=\{h\in{\mathcal O}_{X,x}^{p}\mid J_k(h,{\mathcal M})\subseteq(J_k({\mathcal M}))_S\mbox{ at }x\}.$$ In the family case, the \textbf{3-Lipschitz saturation of ${\mathcal M}$ at $x\in X$ relative to $Y$} is defined as above taking the Lipschitz saturation of $J_k({\mathcal M})$ relative to $Y$. \end{definition} The next proposition allows us to see ${\mathcal M}_{S_3}$ as a sheaf of ${\mathcal O}_X$-submodules of ${\cO_X^p}$. \begin{proposition} Suppose that ${\mathcal M}\subseteq{\mathcal O}_X^p$ is an ${\mathcal O}_{X}$-submodule of generic rank $k$ on each component of $X$. Then: \begin{enumerate} \item ${\mathcal M}_{S_3}$ is an ${\mathcal O}_X$-submodule of ${\cO_X^p}$; \item ${\mathcal M}\subseteq {\mathcal M}_{S_3} \subseteq \overline{{\mathcal M}}$. In particular, ${\mathcal M}$ is a reduction of ${\mathcal M}_{S_3}$ and $e({\mathcal M},{\mathcal M}_{S_3})=0$. \end{enumerate} The same result holds in the family case. \end{proposition} \begin{proof} (1) Let $g,h\in {\mathcal M}_{S_3}$ at $x\in X$ and $\alpha\in{\mathcal O}_{X,x}$. By the basic properties of determinants we have that $$J_k(\alpha g+h,{\mathcal M})\subseteq \alpha J_k(g,{\mathcal M})+J_k(h,{\mathcal M})\subseteq(J_k({\mathcal M}))_S.$$ Hence, $\alpha g+h\in {\mathcal M}_{S_3}$ at $x$. (2) Since $J_k({\mathcal M})\subseteq (J_k({\mathcal M}))_S$ then ${\mathcal M}\subseteq {\mathcal M}_{S_3}$. Now, let $h\in {\mathcal M}_{S_3}$ at $x$. Thus, $J_k(h,{\mathcal M})\subseteq (J_k({\mathcal M}))_S\subseteq\overline{J_k({\mathcal M})}$ which implies that $h\in\overline{{\mathcal M}}$. \end{proof} \begin{proposition}\label{P22} Suppose that ${\mathcal M}$ has generic rank $k$ on each component of $X$. Then ${\mathcal M}_{S_2}\subseteq{\mathcal M}_{S_3}$. \end{proposition} \begin{proof} Suppose $h\in{\mathcal M}_{S_2}$ at $x$. By Theorem 2.3 of \cite{G2} it is enough to prove that $(J_k(h,{\mathcal M}))_D\subseteq\overline{(J_k({\mathcal M}))_D}$ at $(x,x)$. So, we have to prove that all the generators $(J_{IJ}(h,{\mathcal M}))_D$ are in $\overline{(J_k({\mathcal M}))_D}$, for all $i\in\{1,...n\}$ and for all indexes $I$ and $J$. Write $I=(i_1,...,i_k)$ and $J=(j_1,...,j_k)$. We have two cases. Suppose $j_1>1$. Then, as we saw before, $J_{IJ}(h,{\mathcal M})\in J_k({\mathcal M})$, so $(J_{IJ}(h,{\mathcal M}))_D\in (J_k({\mathcal M}))_D$, in particular, $(J_{IJ}(h,{\mathcal M}))_D\in \overline{(J_k({\mathcal M}))_D}$. Suppose $j_1=1$. By Lemma \ref{L21} there exists $\psi:X\rightarrow \mbox{Hom}({\mathbb{C}}^p,{\mathbb{C}})$ such that $\psi\cdot h=J_{IJ}(h,{\mathcal M})$ and $\psi\cdot {\mathcal M}\subseteq J_k({\mathcal M})$. Since $h$ is in the 2-Lipschitz Saturation of ${\mathcal M}$ at $x$ then $\psi\cdot h\in(\psi\cdot {\mathcal M})_S$, which is equivalent to $(\psi\cdot h)_D\in\overline{(\psi\cdot {\mathcal M})_D}$. Thus, $(J_{I,J}(h,{\mathcal M}))_D=(\psi\cdot h)_D\in\overline{(\psi\cdot {\mathcal M})_D}\subseteq\overline{(J_k({\mathcal M}))_D}$. \end{proof} The next result shows the third saturation is closely related to the pullback of the saturated blow-up map. \begin{proposition}\label{P20} Suppose that ${\mathcal M}$ has generic rank $k$ on every component of $X$. Let $\pi: SB_{J_{k}({\mathcal M})}(X)\rightarrow B_{J_{k}({\mathcal M})}(X)$, $p: B_{J_{k}({\mathcal M})}(X)\rightarrow X$ and $\pi_S=p\circ\pi$ be the projections maps. Let $h\in{\mathcal O}_{X,x}^p$. Then: \begin{center} $h\in{\mathcal M}_{S_3}$ if and only if $\pi_S^{*}(h)\in\pi_S^{*}({\mathcal M})$.\end{center} \end{proposition} \begin{proof} Fix a set of generators $\{g_1,...,g_r\}$ of ${\mathcal M}$. We work at $x'\in E$, $E=\pi^{-1}(E_B)$, $E_B$ the exceptional divisor of $B_{J_{k}({\mathcal M})}(X)$. Suppose first that $J_k(h,{\mathcal M})\subseteq (J_k({\mathcal M}))_S$. Let $\pi_S^{*}(J_{I,J}({\mathcal M}))$ be a local generator of the principal ideal $\pi_S^{*}(J_k({\mathcal M}))$. Then, by Cramer's rule we can write $$(J_{I,J}({\mathcal M})\circ\pi_S)(h_I\circ\pi_S)=\sum\limits_{j\in J}(J_{I,j}(h_I,{\mathcal M}_I)\circ\pi_S)(m_{I,j}\circ\pi_S) \eqno(1)$$ with $m_j\in {\mathcal M}$. We claim in fact that $$(J_{I,J}({\mathcal M})\circ\pi_S)(h\circ\pi_S)=\sum\limits_{j\in J}(J_{I,j}(h_I,{\mathcal M}_I)\circ\pi_S)(m_{j}\circ\pi_S).$$ To see this pick a curve $\phi:({\mathbb{C}},0)\rightarrow(SB_{J_k({\mathcal M})}(X),x')$ and choose $\phi$ so that the rank of $\pi_S^*({\mathcal M})|_{\phi}$ is generically $k$. Since by hypothesis $h\in\overline{{\mathcal M}}$ then $h\circ\pi_S\circ\phi\in\phi^*(\pi_S^*({\mathcal M}))$. So, the element $$\star=h\circ\pi_S\circ\phi-\sum\limits_{j\in J}(\frac{J_{I,j}(h_I,{\mathcal M}_I)\circ\pi_S}{J_{I,J}({\mathcal M})\circ\pi_S}\circ\phi)(m_{j}\circ\pi_S\circ\phi)\in\phi^*(\pi_S^*({\mathcal M}))$$ because the above quotients by hypothesis are regular functions on $SB_{J_k({\mathcal M})}(X)$. By equation (1), the above element has $0$ for the entries indexed by $I$. So, if $\star$ is not zero then $\phi^*(\pi_S^*({\mathcal M}))$ has rank at least $k+1$, which is a contradiction. Since the images of $\phi$ fill up a Z-open set it follows that $\star$ is locally zero. Therefore, $$h\circ\pi_S=\sum\limits_{j\in J}(\frac{J_{I,j}(h_I,{\mathcal M}_I)\circ\pi_S}{J_{I,J}({\mathcal M})\circ\pi_S})(m_{j}\circ\pi_S)\in\pi_S^*({\mathcal M}).$$ Conversely, suppose that $\pi_S^{*}(h)\in\pi_S^{*}({\mathcal M})$. Then we can write $$h\circ\pi_S=\sum\limits_{j=1}^{r}\alpha_j(g_j\circ\pi_S) \eqno(2)$$ where $\alpha_j\in\widetilde{{\mathcal O}}_{B_{J_k({\mathcal M})}(X),x'}$, $\forall j\in\{1,...,r\}$. \noindent Let $c$ be an arbitrary generator of $J_k(h,{\mathcal M})$. Then we can write $c=\det[h,{\mathcal M}]_{IJ}$ for some $k$-indexes $I$ and $J$. If the $k$-index $J$ does not pick the first column of $[h,{\mathcal M}]$, then $c$ is a $k\times k$ minor of the matrix $[{\mathcal M}]$, hence $c\in J_k({\mathcal M})\subseteq (J_k({\mathcal M}))_S$ and we are done. Thus we may assume that $j_1=1$. Then, we can write $$c=\det\left[\begin{matrix} h_I & (g_{j_1})_I &... &(g_{j_{k-1}})_I \end{matrix}\right].$$ Applying the pullback of the projection map we have $$\pi_S^*(c)=\det\left[\begin{matrix} h_I\circ\pi_S & (g_{j_1})_I\circ\pi_S &... &(g_{j_{k-1}})_I\circ\pi_S \end{matrix}\right].$$ \noindent By the equation (2), if we look only to the entries of the $k$-index $I$, we get $$h_I\circ\pi_S=\sum\limits_{j=1}^{r}\alpha_j((g_j)_I\circ\pi_S).$$ \noindent Using the linearity of the determinant in the first column, we conclude that $\pi_S^*(c)=\sum\limits_{j=1}^{r}\alpha_j(v_j\circ\pi_S)$, where $v_j:=\det\left[\begin{matrix} (g_j)_I & (g_{j_1})_I &... &(g_{j_{k-1}})_I \end{matrix}\right],$ for all $j\in\{1,..,r\}$. Clearly, $v_j\in J_k({\mathcal M})$, $\forall j\in\{1,...,r\}$. Hence, $\pi_S^*(c)\in\pi_S^*(J_k({\mathcal M}))$ and by the definition of the Lipschitz saturation of an ideal, we conclude that $c\in(J_k({\mathcal M}))_S$. \end{proof} \begin{corollary}\label{C21} Suppose that ${\mathcal M}$ has generic rank $k$ on every component of $X$. If $h\in{\mathcal M}_{S_3}$ then there exists an open covering of $X$ such that $h$ can be written on each $U$ of the covering as an element of ${\mathcal M}$ by using Lipschitz functions (Lipschitz with respect to $(z_i)$ and $(\frac{J_{I',J'}({\mathcal M})}{J_{I,J}({\mathcal M})})$, where $J_{I,J}({\mathcal M})$ is the minor which gives the local generator on the preimage of $U$ in $SB_{J_k({\mathcal M})}(X)$) . \end{corollary} Now we look for conditions so that the above notions of Lipschitz saturation are equivalent. The next proposition gives us such condition. \begin{proposition}\label{T4.17} Let ${\mathcal M}$ be an ${\mathcal O}_{X}$-submodule of ${\mathcal O}_{X}^p$ with generic rank $k$ on each component of $X$, $x\in X$ and $h\in{\mathcal O}_{X,x}^p$. Suppose that there exists an ideal $I$ of ${\mathcal O}_{X\times X,(x,x)}^{2p}$ such that $$ IJ_2((J_k({\mathcal M}))_D)\subseteq\overline{J_{2k}({\mathcal M}_D)}$$ $$J_{2k}(h_D,{\mathcal M}_D)\subseteq\overline{IJ_2((J_k(h,{\mathcal M}))_D)}$$ at $(x,x)$. Then: $$h\in({\mathcal M}_{S_1})_x \Leftrightarrow h\in({\mathcal M}_{S_2})_x \Leftrightarrow h\in({\mathcal M}_{S_3})_x $$ \end{proposition} \begin{proof} The first implication follows from Proposition \ref{P23}. The second one follows from Proposition \ref{P22}. Suppose $h\in{\mathcal M}_{S_3}$ at $x$. Then, $(J_k(h,{\mathcal M}))_D\subseteq\overline{(J_k({\mathcal M}))_D}$ and so $\overline{(J_k(h,{\mathcal M}))_D}=\overline{(J_k({\mathcal M}))_D}$. Let us prove that $J_{2k}(h_D,{\mathcal M}_D)\subseteq\overline{J_{2k}({\mathcal M}_D)}$. In fact, let $\phi:({\mathbb{C}},0)\rightarrow(X\times X,(x,x))$ be an arbitrary analytic curve. Since $\overline{(J_k(h,{\mathcal M}))_D}=\overline{(J_k({\mathcal M}))_D}$ then $\phi^*((J_k(h,{\mathcal M}))_D)=\phi^*((J_k({\mathcal M}))_D)$. Using the inclusions of the hypothesis and the curve criterion for the integral closure of modules we have $\phi^*(J_{2k}(h_D,{\mathcal M}_D))\subseteq\phi^*(I\cdot J_2((J_k(h,{\mathcal M}))_D))\\=\phi^*(I)\cdot J_2(\phi^*((J_k({\mathcal M}))_D))=\phi^*(I\cdot J_2((J_k({\mathcal M}))_D))\subseteq\phi^*(J_{2k}({\mathcal M}_D))$. By Theorem 2.9 of \cite{SG} we have that ${\mathcal M}_D$ has generic rank $2k$ on each component of $X\times X$. Therefore, by Proposition \ref{proposition G} we conclude that $h_D\in\overline{{\mathcal M}_D}$. \end{proof} In order to use the above criterion, we prove some useful lemmas that allow us to get the desired equivalence. \textbf{Notation:} For each object $A$, we denote $A=A\circ\pi_1$ and $A'=A'\circ\pi_2$. For $k$-indexes $I,J$, we denote by ${\mathcal M}_{IJ}$ the submatrix of $[{\mathcal M}]$ defined by these $k$-indexes. \begin{lemma} Let ${\mathcal M}\subseteq{\mathcal O}_{X,x}^p$ be a submodule and $k\in{\mathbb{N}}$. Then $$(z_{t_1}-z_{t_1}')...(z_{t_k}-z_{t_k}')\det({\mathcal M}_{IJ})\det({\mathcal M}_{KL}')\in J_{2k}({\mathcal M}_D) \mbox{ at }(x,x)$$ for every $t_1,...,t_k\in\{1,...,n\}$ and $k$-indexes $I,J,K,L$. \end{lemma} \begin{proof} Write ${\mathcal M}_{KL}=(g_{rs})$. The $2k\times 2k$ matrix $$\left[\begin{matrix} {\mathcal M}_{IJ} & | & 0_{k\times k} \\ --------& & ------------\\ {\mathcal M}_{IJ}' & | & \begin{matrix} (z_{t_1}-z_{t_1}')g'_{11} & ... & (z_{t_k}-z_{t_k}')g'_{1k}\\ \vdots & & \vdots \\ (z_{t_1}-z_{t_1}')g'_{k1} & ... & (z_{t_k}-z_{t_k}')g'_{kk} \end{matrix} \end{matrix}\right]$$ is a submatrix of $[{\mathcal M}_D]$. So, its determinant, which is $(z_{t_1}-z_{t_1}')...(z_{t_k}-z_{t_k}')\det({\mathcal M}_{IJ})\det({\mathcal M}_{KL}')$, belogns to $J_{2k}({\mathcal M}_D)$. \end{proof} The next result gives us the first condition of Proposition \ref{T4.17} in terms of the ideal coming from the diagonal. \begin{lemma}\label{L4.20} Let ${\mathcal M}\subseteq{\mathcal O}_{X,x}^p$ be a submodule and $k\in{\mathbb{N}}$. \begin{itemize} \item[a)] $I_{\Delta}^{k}J_2((J_k({\mathcal M}))_D)\subseteq J_{2k}({\mathcal M}_D)$ at $(x,x)$; \item[b)] If $J_k({\mathcal M})$ is principal then $I_{\Delta}^{k-1}J_2((J_k({\mathcal M}))_D)\subseteq J_{2k}({\mathcal M}_D)$ at $(x,x)$. \end{itemize} \end{lemma} \begin{proof} (a) We have that $$[(J_k({\mathcal M}))_D] =\left[\begin{matrix} \det({\mathcal M}_{IJ}) & ... & 0 \\ \det({\mathcal M}_{IJ}') & ... & (z_i-z_i')\det({\mathcal M}_{KL}')\\ \end{matrix}\right]$$ varying the $k$-indexes $I,J,K,L$ and $i\in\{1,...,n\}$. Thus, the desired inclusion is a straightforward consequence of the previous lemma. (b) Since $J_k({\mathcal M})$ is principal then there exist $k$-indexes $I,J$ such that $g=\det({\mathcal M}_{IJ})$ and $J_k({\mathcal M})$ is generated by $\{g\}$.Thus we can write $$[(J_k({\mathcal M}))_D] =\left[\begin{matrix} g & 0 \\ g' & (z_i-z_i')g'\\ \end{matrix}\right]$$ varying $i\in\{1,...n\}$. So, in this case $J_2((J_k({\mathcal M}))_D)$ is generated by $\{g.g'(z_i-z_i')\mid i\in\{1,...,n\}\}$. By previous lemma we have that $$[(z_{t_1}-z_{t_1}')...(z_{t_{k-1}}-z_{t_{k-1}}')].[g.g'(z_i-z_i')]\in J_{2k}({\mathcal M}_D),$$ for all $t_1,...,t_{k-1},i\in\{1,...,n\}$. Therefore, $I_{\Delta}^{k-1}J_2((J_k({\mathcal M}))_D)\subseteq J_{2k}({\mathcal M}_D)$ at $(x,x)$. \end{proof} Now, we start to get conditions in order to obtain the second condition of Proposition \ref{T4.17}. \begin{lemma} Let ${\mathcal M}$ be an ${\mathcal O}_{X,x}$-submodule of ${\mathcal O}_{X,x}^p$ and $k\in{\mathbb{N}}$. Denote by $[\tilde{{\mathcal M}}]$ the matrix such that $$[{\mathcal M}_D]=\left[\begin{matrix} [{\mathcal M}] & 0 \\ [{\mathcal M}]' & [\tilde{{\mathcal M}}] \end{matrix}\right].$$ Let ${\mathcal J}_{2k}({\mathcal M}_D)$ be the subideal of $J_{2k}({\mathcal M}_D)$ generated by $$\{\det({\mathcal M}_{IJ})\det(\tilde{{\mathcal M}}_{KL})\mid I,J,K,L\mbox { are }k\mbox{-indexes}\}.$$ Then, $${\mathcal J}_{2k}({\mathcal M}_D)\subseteq I_{\Delta}^{k-1}J_2((J_k({\mathcal M}))_D)\mbox { at }(x,x) .$$ \end{lemma} \begin{proof} It suffices to prove that each generator of ${\mathcal J}_{2k}({\mathcal M}_D)$ belongs to $I_{\Delta}^{k-1}J_2((J_k({\mathcal M}))_D)$. The columns of $\tilde{{\mathcal M}}_{KL}$ are columns of $[{\mathcal M}]'$ possibly in a different order with terms of type $z_i-z_i'$ multiplied on each column. If the $k$-indexes $K,L$ pick repeated columns then $\det(\tilde{{\mathcal M}}_{KL})=0$ and we are done. If this does not occur, then $$\det(\tilde{{\mathcal M}}_{KL})=(z_{i_1}-z_{i_1}')...(z_{i_k}-z_{i_k}')(\pm \det({\mathcal M}'_{K'L'}))$$ for some reorganization $k$-indexes $K',L'$, where $i_1,...,i_k\in\{1,...,n\}$ are the indexes which $z_{i_1}-z_{i_1}',...,z_{i_k}-z_{i_k}'$ appear on each of the $k$ columns of $\tilde{{\mathcal M}}_{KL}$. Since $\pm(z_{i_1}-z_{i_1}')...(z_{i_{k-1}}-z_{i_{k-1}}')\in I_{\Delta}^{k-1}$ and \\$\det({\mathcal M}_{IJ})(z_{i_k}-z_{i_k}')\det({\mathcal M}'_{KL})\in J_2((J_k({\mathcal M}))_D)$ then $$\det({\mathcal M}_{IJ})\det(\tilde{{\mathcal M}}_{KL})\in I_{\Delta}^{k-1}J_2((J_k({\mathcal M}))_D).$$ \end{proof} The next result ensures that any free submodule satisfies the second condition of Proposition \ref{T4.17} for a suitable power of the ideal coming from the diagonal. \begin{lemma} Let ${\mathcal M}$ be a free ${\mathcal O}_{X,x}$-submodule of ${\mathcal O}_{X,x}^p$ of rank $k$. Then $$J_{2k}({\mathcal M}_D)\subseteq I_{\Delta}^{k-1}J_2((J_k({\mathcal M}))_D)\mbox{ at }(x,x).$$ \end{lemma} \begin{proof} We have that $$[{\mathcal M}_D]=\left[ \begin{matrix} [{\mathcal M}]_{p\times k} & | & 0 \\ ------ & | & ------- \\ [{\mathcal M}]'_{p\times k} & | & (z_i-z_i')[{\mathcal M}]' \end{matrix} \right] $$ varying $i\in\{1,...,n\}$. Let $d\in J_{2k}({\mathcal M}_D)$ at $(x,x)$ be an arbitrary generator. Then $d=\det N$ where $N$ is a $2k\times 2k$ submatrix of $[{\mathcal M}_D]$. (i) Suppose first that there are $k+t$ columns of $N$ taken on the part $$\left[\begin{matrix} 0 \\ (z_i-z_i')[{\mathcal M}]' \end{matrix}\right]$$ of $[{\mathcal M}_D]$, with $1\leq t \leq k$. Then we can write $$N=\left[\begin{matrix} ({\mathcal M}_{IJ})_{k\times(k-t)} & | & 0_{k\times({k+t})}\\ -------- & | & --------\\ ({\mathcal M}_{IJ})'_{k\times(k-t)} & | & (\tilde{{\mathcal M}})_{KL} \end{matrix}\right]$$ $$=\left[\begin{matrix} {\mathcal M}_{IJ} & 0_{k\times t} & | & 0_{k\times k}\\ --- & --- & | & ------\\ {\mathcal M}_{IJ}' & * & | & ** \end{matrix}\right]$$ for some matrices * and **, where ** is a square matrix of size $k\times k$. So in this case we have that $d=\det N=\det\left[\begin{matrix} {\mathcal M}_{IJ} & 0_{k\times t} \end{matrix}\right].\det(**)=0\in I_{\Delta}^{k-1}J_2((J_k({\mathcal M}))_D)$. (ii) Now suppose that we have exactly $k$ columns of $N$ taken on the part $$\left[\begin{matrix} 0 \\ (z_i-z_i')[{\mathcal M}]' \end{matrix}\right]$$ of $[{\mathcal M}_D]$. Then we can write $$N=\left[\begin{matrix} {\mathcal M}_{IJ} & | & 0_{k\times k} \\ ------ & | & ------ \\ {\mathcal M}_{IJ}' & | & \tilde{{\mathcal M}}_{KL} \end{matrix}\right].$$ Thus, $d=\det({\mathcal M}_{IJ})\det(\tilde{{\mathcal M}}_{KL})\in{\mathcal J}_{2k}({\mathcal M}_D)$ and by previous lemma we conclude that $d\in I_{\Delta}^{k-1}J_2((J_k({\mathcal M}))_D)$. \end{proof} As a consequence of the previous lemma, we have a result which states the second condition of Proposition \ref{T4.17} holds locally in a dense Zariski open subset of $X$ in a sheaf of submodules of ${\cO_X^p}$. \begin{lemma}\label{L4.23} Let ${\mathcal M}$ be an ${\mathcal O}_X$-submodule of ${\cO_X^p}$ of generic rank $k$ on each component of $X$. Then there exists a dense Zariski open subset $U$ of $X$ such that $U\cap V$ is a dense Zariski open subset of $V$, $\forall V$ component of $X$, $J_k({\mathcal M})={\mathcal O}_{X,x}$ and $$J_{2k}({\mathcal M}_D)\subseteq I_{\Delta}^{k-1}J_2((J_k({\mathcal M}))_D)\mbox{ at }(x,x),$$ for all $x\in U$. \end{lemma} \begin{proof} Consider $[{\mathcal M}]$ a matrix of generators of ${\mathcal M}$. Using the Cramer's rule, we can choose $k$ ${\mathcal O}_X$-linear independents columns of $[{\mathcal M}]$ such that these columns generates ${\mathcal M}$ in a such dense Zariski open subset $U$ of $X$ and $J_k({\mathcal M})={\mathcal O}_X$ along $U$. Let ${\mathcal M}_k$ be the ${\mathcal O}_X$-submodule of ${\cO_X^p}$ generated by the columns chosen above. Thus, given $x\in U$, we have that ${\mathcal M}_x=({\mathcal M}_k)_x$ is a free ${\mathcal O}_{X,x}$-submodule of ${\mathcal O}_{X,x}^p$ of rank $k$ and the desired inclusion is a consequence of the previous lemma. \end{proof} Before we state the main theorem of this section, we prove that the generic rank of $(h,{\mathcal M})$ is the same as the generic rank of ${\mathcal M}$, provided $h\in\overline{{\mathcal M}}$. \begin{lemma} Let ${\mathcal M}$ be an ${\mathcal O}_{X,x}$-submodule of ${\mathcal O}_{X,x}^p$ of generic rank $k$ on each component of $X$. If $h\in\overline{{\mathcal M}}$ then $(h,{\mathcal M})$ also has generic rank $k$ on each component of $X$. \end{lemma} \begin{proof} Since $h\in\overline{{\mathcal M}_x}$ then $h\in\overline{{\mathcal M}_{x'}}$, for all $x'$ in an open neighborhood $U$ of $x$ in $X$. Suppose the rank of $(h,{\mathcal M})$ is $k+1$. Then $h(x')\notin {\mathcal M}(x')$. But, $x'$ is a constant curve, so $h\notin\overline{{\mathcal M}_{x'}}$, contradiction. \end{proof} The next theorem gives us the desired equivalence. \begin{theorem} Let ${\mathcal M}$ be an ${\mathcal O}_X$-submodule of ${\mathcal O}_X^p$ of generic rank $k$ on each component of $X$. Then there exists a dense Zariski open subset $U$ of $X$ such that $$ {\mathcal M}_{S_1}={\mathcal M}_{S_2}={\mathcal M}_{S_3}$$ along $U$. \end{theorem} \begin{proof} Since $X\subseteq{\mathbb{C}}^n$ is an analytic complex variety then ${\mathcal O}_X$ is a noetherian sheaf of rings, hence ${\mathcal O}_X^p$ is a noetherian sheaf of ${\mathcal O}_X$-modules. Since ${\mathcal M}_{S_3}$ is a sheaf of ${\mathcal O}_X$-modules then ${\mathcal M}_{S_3}$ is finitely generated by global sections $h_1,...,h_r$. Since $h_i\in\overline{{\mathcal M}}$, $\forall i\in\{1,...,r\}$ then by previous lemma $(h_i,{\mathcal M})$ also has generic rank $k$ on each component of $X$, $\forall i\in\{1,...,r\}$. By Lemma \ref{L4.23}, for each $i\in\{1,...,r\}$ there exists a dense Zariski open subset $U_i$ of $X$ such that $U_i\cap V$ is a dense Zariski open subset of $V$, for all $V$ component of $X$, and $$J_{2k}((h_i)_D,{\mathcal M}_D)\subseteq I_{\Delta}^{k-1}J_2((J_k(h_i,{\mathcal M}))_D)\mbox{ at }(x,x),$$ $\forall x\in U_i$. Also by Lemma \ref{L4.23} there exists a dense Zariski open subset $U_0$ of $X$ such that $U_0\cap V$ is a dense Zariski open subset of $V$, for all component $V$ of $X$, and $J_k({\mathcal M})={\mathcal O}_{X,x}$ which is principal at $x$, $\forall x\in U_0$. From Lemma \ref{L4.20}(b), we conclude that $$I_{\Delta}^{k-1}J_2((J_k({\mathcal M}))_D)\subseteq J_{2k}({\mathcal M}_D)\mbox{ at }(x,x),$$ $\forall x\in U_0$. Take $U:=\bigcap_{j=0}^r U_j$. Then $U$ is a dense Zariski open subset of $X$ and $U\cap V$ is a dense Zariski open subset of $V$, for all $V$ component of $X$. We already know that ${\mathcal M}_{S_1}\subseteq {\mathcal M}_{S_2}\subseteq {\mathcal M}_{S_3}$ at any point of $X$, in particular, along $U$. Let us prove that ${\mathcal M}_{S_3}\subseteq {\mathcal M}_{S_1}$ along $U$. In fact, let $x\in U$. Given an arbitrary $i\in\{1,...,r\}$, since $x\in U_0$ and $x\in U_i$ then: \begin{itemize} \item $I_{\Delta}^{k-1}J_2((J_k({\mathcal M}))_D)\subseteq J_{2k}({\mathcal M}_D)\mbox{ at }(x,x)$; \item $J_{2k}((h_i)_D,{\mathcal M}_D)\subseteq I_{\Delta}^{k-1}J_2((J_k(h_i,{\mathcal M}))_D)\mbox{ at }(x,x)$. \end{itemize} Since $h_i\in {\mathcal M}_{S_3}$ at $x$ then by Proposition \ref{T4.17} we have that $h_i \in {\mathcal M}_{S_1}$ at $x$, $\forall i\in\{1,...,r\}$. Since ${\mathcal M}_{S_3}$ at $x$ is generated by $h_1,...,h_r$ at $x$, then ${\mathcal M}_{S_3}\subseteq {\mathcal M}_{S_1}$ at $x$. \end{proof} In the next example we show that in general ${\mathcal M}_{S_3}$ is not a submodule of ${\mathcal M}_{S_1}$. \begin{example} Consider the submodule ${\mathcal M}$ of ${\mathcal O}_2^2$ given by \\$[{\mathcal M}]=\left[\begin{matrix} x & 0 & y\\ y & x & 0 \end{matrix}\right]$ and let $h=(x,3y)$. Thus, $J_2({\mathcal M})=\langle x^2,xy,y^2\rangle=J_2(h,{\mathcal M})$. In particular, $h\in{\mathcal M}_{S_3}$. Consider the curve $\Phi:({\mathbb{C}},0)\rightarrow({\mathbb{C}}^2\times{\mathbb{C}}^2,(0,0))$ given by $\Phi(t)=(t,\alpha t, t, \beta t)$. We have that $$[\Phi^*({\mathcal M}_D)]=\left[\begin{matrix} t & 0 & \alpha t & 0 & 0 & 0 \\ \alpha t & t & 0 & 0 & 0 & 0\\ t & 0 & \beta t & (\alpha-\beta)t^2 & 0 & \beta(\alpha-\beta)t^2\\ \beta t & t & 0 & \beta(\alpha-\beta)t^2 & (\alpha-\beta)t^2 & 0 \end{matrix}\right]$$ \noindent and $[\Phi^*(h_D)]=\left[\begin{matrix} t\\ 3\alpha t \\ t\\ 3\beta t \end{matrix}\right]$. Let $D$ be the first $4\times 3$ submatrix of $[\Phi^*(M_D)]$. The complementar submatrix has degree 2, so we can ignore its terms. Notice that: $[\Phi^*(h_D)]-D\cdot e_1=\left[\begin{matrix} 0\\ 2\alpha t \\ 0\\ 2\beta t \end{matrix}\right]$ and $[\Phi^*(h_D)]-D\cdot (2\alpha e_2)=\left[\begin{matrix} 0\\ 0 \\ 0\\ (2\beta-2\alpha) t \end{matrix}\right]$. However, the first $3$ vectors of $D$ generically have rank $3$. Therefore, $h\notin {\mathcal M}_{S_1}$. \end{example} \section{The Necessity Theorem}\label{sec4} In this section we prove the Necessity theorem, and we relate the infinitesimal Lipschitz condition coming from the first Lipschitz saturation with the strongly bi-Lipschitz equisingularity and the Mostowski's conditions for a Lipschitz stratification. Let $X\subseteq{\mathbb{C}}^n$ be an analytic variety. Let ${\mathcal L}_X$ be the sheaf of locally Lipschitz functions on $X$. Clearly ${\mathcal O}_X$ is a subsheaf of ${\mathcal L}_X$. \begin{definition} Let ${\mathcal M}\subseteq{\cO_X^p}$ be an ${\mathcal O}_X$-submodule. We define ${\mathcal M}_{{\mathcal L}}$ as the ${\mathcal O}_X$-submodule of ${\mathcal O}_X^p$ given by ${\mathcal M}_{{\mathcal L}}:={\mathcal L}_X{\mathcal M}\cap{\cO_X^p}$. The module ${\mathcal M}_{{\mathcal L}}$ is called \textbf{strong Lipschitz saturation of $\mathbf{{\mathcal M}}$}. \end{definition} \begin{theorem}[Necessity Theorem]\label{T2.1} Let ${\mathcal M}\subseteq{\mathcal O}_{X}^p$ be an ${\mathcal O}_X$-submodule of ${\cO_X^p}$ generated by global sections $\{h_1,...,h_r\}$, $x\in X$. If $h\in {\mathcal M}_{{\mathcal L}}$ at $x$ then $h_D\in \overline{{\mathcal M}_D}$ at $(x,x')$, for any $x'\in X$. In particular, ${\mathcal M}_{{\mathcal L}}\subseteq{\mathcal M}_{S_1}$. \end{theorem} \begin{proof} The generators of generators of ${\mathcal M}_D$ are given by the set $$B=\{(h_i)_D\}_{i=1}^r\cup\{(0,(z_j\circ\pi_1-z_j\circ\pi_2)(h_i\circ\pi_2))\}_{i,j=1}^{r,n} .$$ We will use the criterion for integral closure of modules obtained in Proposition 1.11 of \cite{G3}. For each $i\in\{1,...,r\}$ we can write \\$h_{i}=(h_{i1},...,h_{ip})$. Let $\varphi:X\times X\rightarrow \mbox{Hom}({\mathbb{C}}^{2p},{\mathbb{C}})$ be an arbitrary global section of $\Gamma(\mbox{Hom}({\mathbb{C}}^{2p},{\mathbb{C}}),X\times X)$ with components $\varphi_1,...,\varphi_{2p}$. For each $(z,z')\in X\times X$ let $$M(z,z'):=\sup\{{\parallel} \varphi(z,z')\cdot s(z,z'){\parallel}\mid s\in B \}.$$ Since $h\in{\cO_X^p}$ then $h_D\in{\mathcal O}_{X\times X}^{2p}$. By hypothesis, for each $i\in\{1,...,r\}$ there exists $C_i>0$ such that $${\parallel} \alpha_i(z)-\alpha_i(z'){\parallel} \leq C_i\underset{j\in\{1,...,n\}}{\sup}\{{\parallel} z_j-z'_j{\parallel}\}$$ \noindent $\forall (z,z')\in X\times X$. Take $C_0:=\max\{C_1,...,C_r\}$. For each $i\in\{1,...,r\}$, $\alpha_i$ is locally bounded, so there exists $D_i>0$ and $U_i$ neighborhood of $x$ in $X$ such that ${\parallel} \alpha_i(z){\parallel}\leq D_i$, $\forall z\in U_i$. Take $U:=\bigcap\limits_{i=1}^r U_i$ and $D:=\max\{D_1,...,D_r\}$. Let $(z,z')\in U\times X$ be arbitrary. Then: ${\parallel} \varphi(z,z')\cdot h_D(z,z'){\parallel} \leq \sum\limits_{i=1}^{r}{\parallel} \alpha_i(z){\parallel} {\parallel} \varphi(z,z')\cdot (h_i)_D(z,z'){\parallel} + \sum\limits_{i=1}^{r}{\parallel} \alpha_i(z)-\alpha_i(z'){\parallel} {\parallel} \sum\limits_{\ell=1}^{p}\varphi_{p+\ell}(z,z')h_{i\ell}(z,z'){\parallel} \\ \leq DrM(z,z')+C_0r\underset{\begin{smallmatrix} i\in\{1,...,r\}\\ j\in\{1,...,n\} \end{smallmatrix}}{\sup}\{{\parallel} \sum\limits_{\ell=1}^{p}\varphi_{p+\ell}(z,z')(z_j-z'_j)h_{i\ell}(z'){\parallel}\} \\ \leq(Dr+C_0r)M(z,z')$. Taking $C:=C_0r+Dr>0$ we just proved that $${\parallel} \varphi(z,z')\cdot h_D(z,z'){\parallel}\leq C\underset{s\in B}{\sup}\{{\parallel} \varphi(z,z')\cdot s(z,z'){\parallel} \} $$ \noindent for all $(z,z')\in U\times X$, where $B$ is a generator set of $M_D$ and $U\times X$ is a neighborhood of $(x,x')$ in $X\times X$. Therefore, $h_D\in\overline{M_D}$ at $(x,x')$. \end{proof} \begin{remark} If the elements of ${\mathcal L}_X$ were meromorphic functions the above theorem would follow easily. However we assume nothing about them other than that they are Lipschitz. \end{remark} \begin{remark} The Necessity Theorem still works for the real analytic case. One can use the Proposition 4.2 of \cite{G3} and the proof is the same. \end{remark} \begin{corollary} Let ${\mathcal M}\subseteq{\mathcal O}_{X}^p$ be an ${\mathcal O}_X$-submodule of ${\cO_X^p}$. Then, ${\mathcal M}\subseteq {\mathcal M}_{{\mathcal L}}\subseteq \overline{{\mathcal M}}$ and, in particular, $e({\mathcal M},{\mathcal M}_{{\mathcal L}})=0$. \end{corollary} In \cite{FR2} the authors defined the notion of \textit{strongly bi-Lipschitz equisingularity} for a family of analytic complex functions. \begin{definition} Let $U$ be a connected open subset of ${\mathbb{C}}$ and let $(X,0)\subseteq({\mathbb{C}}^n,0)$ be a germ of an analytic variety. Consider a family of analytic functions $F:X\times U\rightarrow {\mathbb{C}}$ with $F(0,y)=0$ and $f_y(z)=F(z,y)$. We say that $F$ is \textbf{strongly bi-Lipschitz equisingular (sLe)} if there exists a continuous map $v:X\times{\mathbb{C}}\rightarrow{\mathbb{C}}^n$, $v_y(z):=v(z,y)$, such that each vector field $v_y:{\mathbb{C}}^n\rightarrow {\mathbb{C}}^n$ is Lipschitz, $\forall y\in U$ and $$\frac{\partial f_y}{\partial y}(z)=D(f_y)_z(v_y(z))$$ \noindent for every $y\in U$ and $z$ near to the origin in $X$. Notice the last equation can be rewritten as $$\frac{\partial F}{\partial y}=\sum\limits_{i=1}^{n}v_i\frac{\partial F}{\partial z_i}$$ \noindent where $v_1,...,v_n$ are the components of $v$. \end{definition} So the sLe condition is equivalent to asking that the partial derivative of $F$ is an element of the strong Lipschitz saturation of the jacobian ideal $J(F)$. In \cite{G4} the author defined a condition for a family of functions analogous to the $iL_A$ condition for sets (see \cite{SG,G1}). \begin{definition} Let $F:{\mathbb{C}}^n\times {\mathbb{C}}\rightarrow {\mathbb{C}}$ be a family of analytic fucntions with $F(0,y)=0$ and $f_y(z)=F(z,y)$. Denote the parameter space $Y=0\times {\mathbb{C}}$ and assume that $Y$ is the singular locus of $F$. Here we work with the double relative to $Y$, i.e, defined by the projections $\pi_1,\pi_2$ restricted to the fiber product $({\mathbb{C}}^n\times{\mathbb{C}})\underset{Y}{\times}({\mathbb{C}}^n\times{\mathbb{C}})$. We say that $F$ is \textbf{infinitesimally Lipschitz equisingular (iLe)} at a point of the fiber product if $$(JM(F)_Y)_D\subseteq\overline{(JM_z(F))_D}$$ at this point. \end{definition} Using the Necessity Theorem, we get the following \begin{corollary}\label{T2.5} If $F:{\mathbb{C}}^n\times{\mathbb{C}}\rightarrow {\mathbb{C}}$ is a strongly bi-Lipschitz equisingular family of functions then $F$ is infinitesimally Lipschitz equisingular. \end{corollary} \begin{corollary} Let $X\subseteq{\mathbb{C}}^n$ be an analytic variety. If $S$ is a stratum of the Mostowski stratification of $X$ then the $iL_A$ condition holds along $S$. \end{corollary} \begin{example}[Fernandes and Ruas example] Consider the family of functions given by $F(x,y,t)=\frac{1}{3}x^3-t^2xy^{3n-2}+y^{3n}$. In \cite{FR} the authors proved this family is not strongly bi-Lipschitz trivial. Here we use our criterion to get the same conclusion. In fact, we have: $\frac{\partial F}{\partial t}=-2txy^{3n-2}$ $\frac{\partial F}{\partial x}=x^2-t^2y^{3n-2}$ $\frac{\partial F}{\partial y}=(2-3n)t^2xy^{3n-3}+3ny^{3n-1}$. Let us prove that $iL_A$ does not hold for the family $F$. Consider the curve $\phi(t)=(ty^{\frac{3n-2}{2}},y,t,-ty^{\frac{3n-2}{2}},y,t)$, $y=t^2$, which is formed by the two parametrizations $\phi_1$ and $\phi_2$ of the two branches of the polar curve defined by $\frac{\partial F}{\partial x}=0$. Since $(\frac{\partial F}{\partial x})_D\circ\phi=(0,0)$ and $y\circ\phi_1-y\circ\phi_2=0$ then $\phi^*((J_z(F))_D)$ has two generators $(\frac{\partial F}{\partial y}\circ\phi_1,\frac{\partial F}{\partial y}\circ\phi_2)$ and $(0,(x\circ\phi_1-x\circ\phi_2)(\frac{\partial F}{\partial y}\circ\phi_2))$. Thus, if $F$ satisfies $iL_A$ then we nust have \\ $\frac{\partial F}{\partial t}\circ\phi_2-\left((\frac{\partial F}{\partial t}\circ\phi_1)\left(\frac{\frac{\partial F}{\partial y}\circ\phi_2}{\frac{\partial F}{\partial y}\circ\phi_1}\right)\right)\in((\frac{\partial F}{\partial y}\circ\phi_2)(x\circ\phi_1-x\circ\phi_2))$. This implies that $$2t^2y^{\frac{9n-6}{2}}\left(1+\frac{3n+(3n-2)t^3y^{\frac{3n-6}{2}}}{3n+(2-3n)t^3y^{\frac{3n-6}{2}}}\right)\in(ty^{\frac{9n-4}{2}})=(ty^{\frac{9n-6}{2}}.y)$$ \noindent which is a contradiction. Hence, $F$ does not satisfy $iL_A$ condition. Hence, $F$ is not strongly bi-Lipschitz trivial. \end{example} \begin{remark} Since Mostowski showed that any complex analytic has a Mostowski stratification, when coupled with our necessity result, this gives another proof of the genericity of the $iL_A$ condition. \end{remark}

28bcfe3c8919e5f6c467a7557084c2c7717caa00

\section*{Introduction}\label{introduction}} \addcontentsline{toc}{section}{Introduction} In their 2010 paper ``Computing in the Statistics Curriculum'', Deborah Nolan and Duncan Temple Lang noted that ``computational literacy and programming are as fundamental to statistical practice and research as mathematics'' and that ``these changes necessitate re-evaluation of the training and education practices in statistics'' \citep{nolan_templelang_2010}. We couldn't agree more about the fundamental role of computing and the need for change at all educational levels. Over the last decade we've seen the role of computing in the statistics curriculum change and grow. The tools have become better, computing is now more established in almost every classroom, and arguably most importantly, the development and success of modern statistics has been enhanced by ideas of computational thinking. Before introducing the articles in this special issue, we reflect on the questions originally posed by Nolan and Temple Lang: \begin{enumerate} \def\arabic{enumi}.{\arabic{enumi}.} \tightlist \item When they graduate, what ought our students be able to do computationally, and are we preparing them adequately in this regard? \item Do we provide students the essential skills needed to engage in statistical problem solving and keep abreast of new technologies as they evolve? \item Do our students build the confidence needed to overcome computational challenges to, for example, reliably design and run a synthetic experiment or carry out a comprehensive data analysis? \item Overall, are we doing a good job preparing students who are ready to engage in and succeed at statistical inquiry? \end{enumerate} Nolan and Temple Lang also provided a damning critique of the status quo at the time: \begin{quote} Many statisticians advocate-or at least practice-the approach in which students are told to learn how to program by themselves, from each other, or from their teaching assistant in a two-week ``crash course'' in basic syntax at the start of a course. Let us reflect on how effective this approach has been. Can our students compute confidently, reliably, and efficiently? We find that this do-it-yourself `lite' approach sends a strong signal that the material is not of intellectual importance relative to the material covered in lectures. In addition, students pick up bad habits, misunderstandings, and, more importantly, the wrong concepts. They learn just enough to get what they need done, but they do not learn the simple ways to do things nor take the time to abstract what they have learned and assimilate these generalities. Their initial knowledge shapes the way they think in the future and typically severely limits them, making some tasks impossible. (page 100) \end{quote} We concur that such an approach to computation is insufficient and at times counterproductive. What has happened in the intervening decade? We believe that there is a growing consensus on the importance of computational literacy and computing in the statistics and data science curriculum. The American Statistical Associations updated Guidelines for Undergraduate Programs in Statistics \citep{asa_guidelines_2014}, the revised GAISE (Guidelines for Assessment and Instruction in Statistics Education) College report \citep{gaisecollege}, and the National Academies of Science, Engineering, and Medicine's consensus study on ``Data Science for Undergraduates: Opportunities and Options'' \citep{nasem_2018} provide detailed rationales for the fundamental role computing plays in statistical thinking. More pointedly, George \citet{cobb_2015} noted a convergence of mathematics, computation, and context in statistics education and called for a deep-rethinking of the curriculum from the ground up. The ``Mere Renovation is Too Little Too Late'' paper sparked 19 spirited responses and a provocative rejoinder (more on the ``tear-down'' metaphor) that challenged the community in a number of fundamental ways \citep{cobb_2015_discuss}. We envisioned this special issue as a way both to highlight innovations and approaches that have helped move the profession forward, as well as to identify places where future work is needed. Many of these papers work to answer the questions posed by Nolan and Temple Lang as well as ones they had not anticipated in 2010. The set of articles included in the special issue can be organized into three non-mutually exclusive clusters that take different approaches to address the questions laid out by Nolan and Temple Lang. The first approach features \textbf{creative structures} for changing how we integrate computing into the learning of statistics. The second approach focuses on \textbf{novel or technical data science skills and habits}. The third reflects that, more and more, statistics educators are embracing and teaching ideas of \textbf{computational thinking}. \hypertarget{creative-structures}{% \subsubsection*{Creative structures}\label{creative-structures}} \addcontentsline{toc}{subsubsection}{Creative structures} Restructuring how we conceive of a syllabus and how we teach particular material is never a small task. However, as different individuals modernize their own courses, we can all learn from their experiences. Both \citet{mine} and \citet{ellis} describe creative and modern data science courses that fold together aspects of statistical inference with vital computational skills. \citet{schwab-mccoy} report on a study describing the emerging consensus of the elements of a data science course. \citet{kim} present some of the technical aspects vital to getting a solid computational course up and running. A less technical approach is described by \citet{burckhardt} using the suite of materials implemented by their Integrated Statistics Learning Environment (ISLE). An immersive data science living and learning community is presented by \citet{ward}. Finally, \citet{theobold} describe an alternative to course learning through a series of workshops. \hypertarget{novel-or-technical-data-science-skills-and-habits}{% \subsubsection*{Novel or technical data science skills and habits}\label{novel-or-technical-data-science-skills-and-habits}} \addcontentsline{toc}{subsubsection}{Novel or technical data science skills and habits} The world of data science is rapidly changing, and it can be incredibly difficult to keep up. Many of the papers in this special issue focus on new, important, and exciting skills and tools that are important for students if they want to contribute in today's data-centric world. \citet{boehm} and \citet{mine} discuss the full cycle of iterating a data science project. \citet{kim-hardin} take it one step further and describe the importance of iterating on the full cycle. A few specific skills are laid out in detail: \citet{dogucu} describe web scraping and \citet{cummiskey} explore techniques for working with multivariate data. \citet{horton} compare ways of incorporating Git in the statistical classroom so that students have the skills to hit the ground running in jobs and in their own data projects, reinforcing the value of reproducible workflows as a foundation for reproducible research. \hypertarget{computational-thinking}{% \subsubsection*{Computational thinking}\label{computational-thinking}} \addcontentsline{toc}{subsubsection}{Computational thinking} The last approach may be the most difficult for statistics and data science educators to embrace and implement in their own classes. The value of bringing in ideas of software engineering or computational thinking is that they help create a mindset that empowers students to simultaneously think both statistically and computationally. \citet{wing} describes how computing can impact a field, for example, ``Computer science's contribution to biology goes beyond the ability to search through vast amounts of sequence data looking for patterns. The hope is that data structures and algorithms---our computational abstractions and methods---can represent the structure of proteins in ways that elucidate their function. Computational biology is changing the way biologists think.'' As a discipline, we are embracing the many ways that computing is changing how statisticians think. \citet{woodard} report on a study where students spoke through their thought process as they performed computational tasks; their results are, somewhat unsurprisingly, that computing is difficult and not intuitive. \citet{schwab-mccoy} describe the challenge in front of us to teach computational thinking effectively. \citet{ellis}, who describe debugging, and \citet{theobold}, who discuss the teaching of iteration (a fundamental component of algorithmic thinking), speak to integrating small pieces of computational thinking within the data science curriculum. \citet{reinhart} describe an entire course that is a cross between software engineering and statistics, providing insight into the types of skills that many statisticians (at all levels) need for success. \hypertarget{what-would-deb-and-duncan-say}{% \subsubsection*{What would Deb and Duncan say?}\label{what-would-deb-and-duncan-say}} \addcontentsline{toc}{subsubsection}{What would Deb and Duncan say?} As we worked on the special issue, we thought that there would be value in asking the authors of \citet{nolan_templelang_2010} to share their thoughts about the paper, what they saw as most valuable, and to peer into the future. Their incisive and provocative retrospective leads off the special issue \citep{nolan-tl-response}. \hypertarget{conclusion}{% \subsubsection*{Conclusion}\label{conclusion}} \addcontentsline{toc}{subsubsection}{Conclusion} The articles in the special issue encourage us to redouble our efforts to embrace computing in the classroom, to constantly push ourselves to learn more tools, and to let computational thinking make our own work better. We believe that the leading thinkers of the next decade will be those who seamlessly knit together tools from both statistics and computing and that how we think about statistics will be informed by complementary computational thinking. To forge ahead we need to cultivate computing foundations throughout the statistics paradigm. It is our hope that the papers in this special issue initiate a new way of thinking for your and your students. \pagebreak \bibliographystyle{agsm}

77b427e30f226a30865ed09c96b4d48296e55a33

\section{Introduction} In studies of child development, investigators often administer a variety of different tests since cognition is a complex and multi-faceted quantity that is difficult to measure directly. In the application that motivates our work, pregnant mothers were interviewed with respect to their alcohol consumption during pregnancy, then the children were followed from birth and assessed with a variety of cognitive and neurodevelopmental tests. Data were combined from six cohort studies in order to boost the statistical power to characterize the dose response relationship between prenatal alcohol exposure and child cognition. This paper introduces a novel method to analyze these multi-cohorts, multiple outcomes data using Structural Equation Modeling. Structural Equation Modeling (SEM) is a popular analysis tool in such settings since it facilitates the fitting of models that specify a structural relationship between exposures and other predictive factors of interest and latent score that reflecting the observable, measured outcomes \citep{kaplan2008structural}. In principle, SEMs can be fitted to a collection of studies or to data obtained from different samples \citep{kaplan2008structural}. For example, Meta-Analysis SEM (MASEM) can be used to combine information from different studies into a full structural model. However, this technique typically requires a large number of studies, each using the same outcome measures, in order for the result to be reliable \citep{cheung2016special}. MASEM for data from studies with diverse outcome measures is typically performed using a two-stage approach \citep{cheung2005meta,cheung2016special}. Sample correlation matrices are computed for each study and then combined using a SEM or used to create a pooled matrix that then provides the input for a SEM (an approach known as two-stage SEM). The disadvantage of these approaches is that they require a large number of studies and also require at least one study that collects all endpoint measures \citep{lv2019evaluation}. When the data can be clustered into groups, multi-group SEM is typically used to test for invariance of factor loadings and mean differences between groups \citep{kaplan2008structural}. Multi-group SEM assigns different parameters for each groups but does not assume a common structure for the factor loadings. Moreover, there is no information in the literature about multi-group SEM for data sets where the groups have different observed variables. The latter restricts the application of SEM in many scenarios. Classical maximum likelihood estimation for SEMs can be obtained via software such as Mplus, OpenMX or the R package \textit{lavaan}. Bayesian inference for SEM (BSEM) has recently received more attention as the framework allows more flexible models and works better with small data sets. Recently, estimation procedures for BSEM have also been added to Mplus and the Bayesian version of \textit{lavaan}, which is known as \textit{blavaan} \citep{merkle2015blavaan}, making analysis more convenient. Although relatively simple to use, software implementations of SEM and BSEM are typically restricted by the normality assumption of the latent factors. They also cannot fit multi-group models in which different sets of variables are available for different groups. In the application that motivates our work, there is an additional complication that limits the suitability of standard SEM software. In particular, a central aim for the investigators was to explore the nature of the dose response relationship in order to assess whether pre-natal alcohol effects persist at all levels of exposure, or whether there might be a minimum level of PAE that results in minimal cognitive deficit. Answering this question is of critical clinical importance in terms of providing guidance when diagnosing children who may have been alcohol affected in utero, and hence in need of intervention. In this paper we introduce a Bayesian SEM approach that allows pooling of information across multiple data sets with different sets of outcome measures. Our model is built upon a second-order Confirmatory Factor Analysis (CFA) model, which is one of the basic SEMs, where each cohort is assigned a different set of SEM parameters. This part of the model links the various outcomes in each cohorts with the latent cognition variable. The model is then extended with a piecewise regression component, where the regression change point is also treated as random. Parameters in this dose response part of the model are shared by all cohorts, thus facilitating an integrative analysis that allows us to effectively explore the relationship between prenatal alcohol exposure (PAE) and child cognition. This setting allows combining the information in all cohorts to estimate the dose-response curve between PAE and child cognitive function even if the set of outcomes in the cohorts is not entirely the same; and also facilitate more flexible functional forms for the dose-response curve. This model is not straightforward to estimate in any standard software and therefore we implement it in a Bayesian framework, which allows estimating all parts of this sophisticated model simultaneously. Bayesian inference is also suitable for our application with a large number of outcome measures and relatively small number of individuals per cohort. In Section \ref{sec:Data} we provide some details about the application that motivates our work and describe the challenges in constructing a suitable model for our data. In Section \ref{sec:Method} we present our Bayesian multi-group model with regression changepoint and provide the details for Bayesian implementation. In Section \ref{sec:simulation} we demonstrate the proposed method in a simulation example. Section \ref{sec:RealApplication} revisits our motivating application and illustrates the method using our data. Section \ref{sec:discussion} concludes the paper. \section{Motivating Application}\label{sec:Data} It is well known that high levels of prenatal alcohol exposure (PAE) can result in a distinct pattern of craniofacial anomalies, growth restriction, and cognitive and behavioral deficits, known as Fetal Alcohol Syndrome (FAS), which is the most severe of a continuum of fetal alcohol syndrome disorders (FASD) \citep{carter2016fetal,hoyme2005practical, hoyme2016updated,jacobson2004maternal,jacobson2008impaired,mattson2019fetal}. However, some individuals with PAE exhibit cognitive and/or behavioural impairment without the characteristic craniofacial dysmorphology, a condition known as Alcohol-Related Neurodevelopmental Disorder (ARND). Although a confirmed history of maternal alcohol consumption may suggest the presence of ARND, the levels of PAE associated with an increased risk of clinically significant adverse effects are not known. In addition, the effects at low levels of exposure are less well understood, thus the full extent of the dose-response curve between PAE and child cognitive function remains unclear. In order to boost the power needed to study the dose response effect, investigators combined data that had been collected from six longitudinal cohort studies conducted in the United States. In this paper, the cohorts are referred to as Detroit, Seattle, Atlanta 1, Atlanta 2, Pittsburgh 1 and Pittsburgh 2, based on the location where the studies were conducted. In these studies, the mothers were interviewed prenatally or shortly after delivery about their drinking habits during pregnancy, and the children were followed longitudinally to assess their IQ, academic achievement in reading and arithmetic, learning and memory abilities and executive function. Together these tests provide a very comprehensive assessment of the child's cognitive function. IQ and each of these cognitive domains were assessed using several tests. The final data set consists of data from more than 2200 children. Their mothers were interviewed about consumption of beer, wine, and liquor during pregnancy, and these data were summarized in terms of ounces of absolute alcohol (AA) consumed per day (1 ounce AA equals 2 standard drinks of alcohol). The number of prenatal maternal interviews varied across the cohorts. Table \ref{tab:dataDescribe} presents information about the six studies. This rich data set is used to analyze the effect of PAE on child cognitive function and to determine the levels of PAE that are associated with higher risk of adverse effect. It is important to emphasize that cognitive function cannot be observed directly and thus we use the children's neuropsychological test results. A SEM is therefore a suitable model to link the observed outcomes and the unobservable measures (IQ, learning and memory and executive function), and to link these measures with unobserved cognition latent variable. The goal is to analyze the effect of PAE on child cognition and determine the shape of the dose-response curve. Since the number of participants in each cohort is small compared to the number of observed outcome measures, the information from all six cohorts is combined in one model instead of analyzing the cohorts separately. This is particularly challenging since different tests were often used in different cohorts. The number of tests also differs across studies, therefore it is not possible to use the method in \cite{ke2019bayesian} for this data set. If all the outcomes were combined and used to construct a large correlation matrix, the proportion of missing data would be too large to be accommodated by MASEM. We will show how we can use a Bayesian multi-group SEM to analyze these data in Section \ref{sec:RealApplication}. \section{Methodology}\label{sec:Method} In this Section we present our BSEM model for multiple studies and the details of Bayesian estimation of the model. \subsection{Multi-group BSEM model}\label{subsec:Model} Let $\mathbf{Y}_{ci}$ be a $K_c \times1$ vector that denotes the $K_c$ outcomes observed for individual $i$ in cohort $c$, $ i=1,\dots,n_c,\quad c= 1,\dots,C$. In our application $C=6$. Let $X_{ci}$ denote the exposure variable for individual $i$ in cohort $c$. We next define the three component regression models in terms of the latent factors comprising our structural equation model. The full model is given as \begin{eqnarray} \mathbf{Y}_{ci} &=& \bm{\nu}_c + \bm{\Lambda}_c \bm{\xi}_{ci} + \bm{\epsilon}_{ci},\quad \bm{\epsilon}_{ci}\sim MN(0,\bm{\Psi}_c), \label{eq:modelFull1}\\ \bm{\xi}_{ci}& =& \bm{\gamma}_c\eta_{ci} + \bm{u}_{ci}, \quad \bm{u}_{ci} \sim MN(0,\bm{\Phi}_c).\label{eq:modelFull2} \\ \eta_{ci} &= &\beta_1 X_{ci} + \beta_2 (X_{ci} - X^{bp})_+ + \alpha_c P_{ci} + e_{ci},\quad e_{ci} \sim N(0,\sigma_c^2), \label{eq:modelFull3} \end{eqnarray} where $\bm{\nu}_c$ is the vector of intercepts for cohort $c$. The $K_c\times 1$ errors $\bm{\epsilon}_{ci}$ are independent and normally distributed with mean 0 and covariance matrix $\bm{\Psi_c}$. We now take some time to discuss the meaning of the different terms in the models beginning with those in \eqref{eq:modelFull1} and \eqref{eq:modelFull2}. The $\bm{\xi}_{ci}$ term is a $d \times 1$ vector $(d\leq K_c)$ of latent factors with distinct elements corresponding to the different subdomains of cognition for individual $i$ in cohort $c$. The $K_c \times d$ loading matrix $\bm{\Lambda}_c$ is sparse with only $K_c$ non-zero elements: in column $j$ only the rows corresponding to the $j$th latent variable (i.e. the $j$th element of $\bm{\xi}_{ci}$) are non-zero, $j=1,\ldots, d$. Note that by having cohort-specific loading matrices $\bm{\Lambda}_c$, $c=1,\ldots, C$, we accommodate different numbers and types of outcome variables between the cohorts. The overall measure of \textit{cognition} for individual $i$ in cohort $c$ is represented by the latent variable $\eta_{ci}$ which is related to the subdomain-specific latent variables in $\bm{\xi_{ci}}$ via \eqref{eq:modelFull2}. The first elements of the $d\times1$ loading vectors $\bm{\gamma}_c = (1, \gamma_{c,1},\dots,\gamma_{c,d-1} )^\prime$ are fixed at 1 to ensure identifiability. Finally, the $\bm{u}_{ci}$ terms in \eqref{eq:modelFull2} are independent and identically distributed (i.i.d.) error terms and the covariance matrices $\bm{\Phi}_c$ is constrained to be diagonal. Equation \eqref{eq:modelFull3} relates the latent cognition variable to PAE. We aim to investigate whether there is a level of PAE above which the effect of PAE on child cognition becomes stronger, and we denote this threshold ``break-point" by $X^{bp}$, and define this as a parameter to be estimated. In the piecewise linear model of \eqref{eq:modelFull3} we write $(\textit{X}_{ci} - X^{bp})_+$ to represent$(\textit{X}_{ci} - X^{bp})\mathbf{I}(\textit{X}_{ci} >X^{bp})$. The coefficient $\beta_1$ of \eqref{eq:modelFull3} thus represents the effect of PAE on cognition at doses below $X^{bp}$, while the coefficient $\beta_2$ represents the change in the slope after the break-point $X^{bp}$. To adjust for confounders associated with both alcohol exposure and cognition we incorporate a propensity score $P_{ci}$ in the linear predictor; for details on the covariates included in the propensity score and how it is constructed we refer readers to \cite{hocagil2020pscore}. The error terms $e_{ci}$ in \eqref{eq:modelFull3} are i.i.d. with variance $\sigma_c^2$ for $c= 1,\dots, C$, $i=1,\dots, n_c$. There have been other multi-group Bayesian SEMs that allow group-specific factor loadings. For example, \cite{ke2019bayesian} propose a multilevel Bayesian SEM in which the factor loadings $\bm{\Lambda_c}$ may vary across cohorts. The unique parameters in these matrices all come from a multivariate normal distribution of which parameters will be estimated. However their approach utilizes the observed correlation matrices observed from each study, and hence requires that all studies collect the same outcome variables. While being computationally efficient, their approach also requires a large number of studies and thus is not appropriate for the application we are considering. The diagram in Figure \ref{fig:sempath1} shows the structural relation between the variables in Equations \eqref{eq:modelFull1}-- \eqref{eq:modelFull3}. \subsection{Bayesian estimation of a SEM} There is limited literature on multi-group SEM with shared parameters across groups. It is also not straightforward to estimate and make inferences about the break-point in a piecewise-linear regression. However, the model in Section \ref{subsec:Model} fits well into the Bayesian framework, and we outline how to construct and fit a Bayesian model in the next Section. In Bayesian statistics, we make inferences about the \textit{posterior distribution} of the parameter vector $\boldsymbol{\theta} \in \Theta \subset \mathbb{R}^{d} $ given the data $\bm{Y}$, $$p(\boldsymbol{\theta}|\bm{Y}) = \frac{p_{\Theta}(\boldsymbol{\theta})p(\bm{Y}|\boldsymbol{\theta})}{\int_{\Theta} p_{\Theta}(\boldsymbol{\theta})p(\bm{Y}|\boldsymbol{\theta})d\boldsymbol{\theta}},$$ where $p(\bm{Y}|\boldsymbol{\theta})$ is the likelihood, $p_{\Theta}(\boldsymbol{\theta})$ is the \textit{prior distribution} which encodes our prior knowledge about the parameters $\boldsymbol{\theta}$ (for example we know that the variance of the cognition latent variable should be positive and near 100). Bayesian inference can be viewed as a way to use the data to update our existing knowledge of the parameters. \cite{ke2019bayesian} suggest that Bayesian estimation of SEMs is less sensitive to missing data but typically suffers from slow convergence due to the large number of free parameters. In addition, the functional form of $p(\boldsymbol{\theta}|\bm{Y})$ typically does not correspond to a known distribution, so simulation-based techniques are often used to make inference about the posterior distribution. Bayesian SEMs can be estimated by Markov Chain Monte Carlo (MCMC) \citep{lee2007structural,merkle2015blavaan,muthen2012bayesian}. MCMC is a collection of simulation based methods to approximately sample from the posterior distribution of the parameters. More details about Bayesian computation can be found in Part III of \cite{gelman2013bayesian}. The parameters $\boldsymbol{\theta}$ in this model consists of $\bm\nu_c$, $\boldsymbol{\Lambda_c}$ , $\bm{\gamma_c}$, $\bm{\xi}_{ci}$, $\eta_{ci}$, $\bm{\Psi_c}$, $\bm{\Phi_c}$, $\sigma^2_c$ and $\alpha_c$ for $ i = 1,\dots, n_c; c = 1,\dots, C,$ the break-point $X^{bp}$ and the 2 common slopes $\beta_1, \beta_2$. An advantage of the Bayesian approach is that we can make inference about the latent factor scores $\eta_{ci}$ by analyzing their posterior distribution. Analyzing the maximum likelihood estimates of the factor scores is challenging because of their complicated asymptotic behavior \citep{lee2007structural}. Bayesian estimation of SEMs has been implemented in standard software such as \textit{Mplus} \citep{muthen2012bayesian} and \textit{blavaan} \citep{merkle2015blavaan}. The estimation procedure by \cite{muthen2012bayesian} and the BUGS implementation of \textit{blavaan} \citep{merkle2015blavaan} update the parameters in blocks by a Metropolis-within-Gibbs algorithm. A more efficient implementation of Bayesian SEM using Hamiltonian Monte Carlo (HMC) \citep{betancourt2017conceptual,neal2011mcmc} is used in the more recent version of \textit{blavaan} \citep{merkle2015blavaan}. However, it is very costly to run HMC with such a large number of parameters and hence the package estimates a ``reduced model'' obtained by integrating out the latent variables $\bm{\xi}_{ci}$ and $\eta_{ci}$. This relies on the normality assumption of the data, and does not provide inference on the latent variables. We attempted to fit the full model, including sampling the latent variables using HMC for this paper, but it was too computationally expensive. We then tried to speed up calculation by making use of the normal assumption and integrate out the latent variable $\boldsymbol{\xi_{ci}}$. This allows more flexibility in the distribution of the error terms associated with the outcome variables while making the computation faster. Details are provided in Section \ref{subsec:semiMarginalModel}. Our simulation example and application are implemented in STAN \citep{carpenter2017stan}. We also would like to note that software such as \textit{blavaan} cannot handle models with unknown break-point such as those in this paper. \subsection{Handling missing data}\label{subsec:missingdata} Because the data for each cohort were derived from longitudinal studies conducted over a long period of time, there are missing test results for some children. In this PAE application, some cohorts had more than 70\% of the participants with at least one unobserved outcome. Therefore listwise deletion is not possible. For the frequentist approach, the Full Information Maximum Likelihood (FIML) technique that uses case-wise likelihood is often used when the data are assumed to be missing at random \citep{arbuckle1996full,finkbeiner1979estimation}. For Bayesian SEM, it is straightforward to treat the missing outcomes as parameters and sample them together with the rest of the parameters. This approach is often used because it allows estimation of the missing data and therefore multiple imputation is not needed. However, for this application, the number of missing cells in the data table is very large and this might significantly slow down posterior sampling. Therefore we propose a similar approach to FIML to define the likelihood of the observed data and use that for the Bayesian estimation We follow the notation and formulation from \cite{finkbeiner1979estimation} and define a fixed matrix $W_{ci}$ for individual $i$ in cohort $c$ to collapse the missing elements of $\mathbf{Y}_{ci}$. Let $m_{ci}$ be the number of observed outcomes for case $i$ in cohort $c$. $W_{ci}$ is an $m_{ci}\times K_c$ matrix that is formed by removing the rows corresponding to the missing outcomes from an $K_c\times K_c$ identity matrix. Then we define the likelihood in terms of the observed data $\mathbf{Y}_{ci}^o= W_{ci}\mathbf{Y}_{ci}$ instead of $\mathbf{Y}_{ci} $. Let $\bm{\theta}$ denotes all the parameters in the model and let $\mathbf{Y}_{ci}^m$ be the missing element of $\mathbf{Y}_{ci}$, we have \begin{eqnarray*} p(\mathbf{Y}_{ci},W_{ci}|X_{ci},P_{ci},\bm{\theta}) &= & p(\mathbf{Y}_{ci}^o,\mathbf{Y}_{ci}^m,W_{ci}|X_{ci},P_{ci},\bm{\theta})\\ & = & p(\mathbf{Y}_{ci}^o|X_{ci},P_{ci},\bm{\theta} )p(\mathbf{Y}_{ci}^m|\mathbf{Y}_{ci}^o,X_{ci},P_{ci},\bm{\theta})\\ & & \times p(W_{ci}|\mathbf{Y}_{ci}^m,\mathbf{Y}_{ci}^o,X_{ci},P_{ci},\bm{\theta}). \end{eqnarray*} If the data are missing at random and the missing data process is non-informative, then we can omit $p(W_{ci}|\mathbf{Y}_{ci}^m,\mathbf{Y}_{ci}^o,X_{ci},P_{ci},\bm{\theta})$ and $$p(\mathbf{Y}_{ci},W_{ci}|X_{ci},P_{ci},\bm{\theta}) \propto p(\mathbf{Y}_{ci}^o|X_{ci},P_{ci},\bm{\theta} )p(\mathbf{Y}_{ci}^m|\mathbf{Y}_{ci}^o,X_{ci},P_{ci},\bm{\theta}). $$ The likelihood of the observed data is then \begin{eqnarray*} p(\mathbf{Y}_{ci}^o,W_{ci}|X_{ci},P_{ci},\bm{\theta}) &= & \int_{ \mathbf{Y}_{ci}^m} p(\mathbf{Y}_{ci}^o|X_{ci},P_{ci},\bm{\theta} )p(\mathbf{Y}_{ci}^m|\mathbf{Y}_{ci}^o,X_{ci},P_{ci},\bm{\theta}) d \mathbf{Y}_{ci}^m. \end{eqnarray*} Notice that conditioning on the latent factors, $(\mathbf{Y}_{ci}^o,\mathbf{Y}_{ci}^m)$ are jointly normal and hence after integrating out $\mathbf{Y}_{ci}^m $, the model in Equations \eqref{eq:modelFull1} $-$ \eqref{eq:modelFull3} becomes \begin{eqnarray}\label{eq:modelwMiss} W_{ci}\mathbf{Y}_{ci} &=& W_{ci} \left(\boldsymbol{\nu_c} + \bm{\Lambda}_c \bm{\xi}_{ci}\right) + \bm{\epsilon}_{ci},\quad \bm{\epsilon}_{ci}\sim MN(0,W_{ci}\bm{\Psi}_c W_{ci}^\prime), \label{eq:modelFullMissing1}\\ \bm{\xi}_{ci}& =& \bm{\gamma}_c\eta_{ci} + \bm{u}_{ci}, \quad \bm{u}_{ci} \sim MN(0,\bm{\Phi}_c), \label{eq:modelFullMissing2}\\ \eta_{ci} &= &\beta_1 X_{ci} + \beta_2 (X_{ci} - X^{bp})_+ + \alpha_c P_{ci} + e_{ci},\quad e_{ci} \sim N(0,\sigma_c^2) \label{eq:modelFullMissing3}. \end{eqnarray} \subsection{Bayesian model evaluation and model selection}\label{subsec:GoodnessOfFit} In this Section, we discuss the two quantities that we use to evaluate a model's performance and to do model selection. \subsubsection{Information criteria} For a Bayesian model with the set of parameters $\bm{\theta}$ and data $\mathbf{Y} = \{\mathbf{Y}_{ci}:i = 1,\dots,n_c; c = 1,\dots,C\} $, a measure of its predictive accuracy for the data points taken one at a time is the \textit{expected log pointwise predictive density for a new data set (eldp)} $$eldp = \sum_{c=1}^C \sum_{i = 1}^{n_c}\int p_t(\mathbf{\tilde{Y}}_{ci})\log p(\mathbf{\tilde{Y}}_{ci}|\mathbf{Y})d\mathbf{\tilde{Y}}_{ci}, $$ where $p_t(\mathbf{\tilde{Y}}_{ci})$ is the distribution representing the true data-generating process for the new data $\mathbf{\tilde{Y}}_{ci}$ and $C$ is the number of cohorts. A helpful quantity is the log pointwise predictive density ($lpd$) $$lpd = \sum_{c=1}^C \sum_{i=1}^{n_c} \log p(\mathbf{Y}_{ci}|\mathbf{Y}) = \sum_{c=1}^C \sum_{i=1}^{n_c} \log \int p(\mathbf{Y}_{ci}|\bm{\theta})p(\bm{\theta}|\mathbf{Y})d\bm{\theta}.$$ In practice, $lpd$ can be estimated from $S$ MCMC posterior draws by $$\widehat{lpd} = \sum_{c=1}^C \sum_{i=1}^{n_c} \log(\frac{1}{S}\sum_{i=1}^S p(\mathbf{Y}_{ci}|\bm{\theta}^s)). $$ The $lpd$ of observed data $\mathbf{Y}$ is an overestimate of the $elpd$ for future data, because it is evaluated on the data from which the model was fitted. For our analysis, we use the Watanabe-Akaike Information Criterion (WAIC). WAIC \citep{watanabe2010asymptotic} is a more fully Bayesian approach to estimate "out-of-sample" expectation than the Deviance Information Criterion (DIC) \citep{spiegelhalter2002bayesian}. It is constructed by computing the log pointwise posterior predictive density then correcting for the effective number of parameters $$\widehat{eldp}_{\text{WAIC}} = \widehat{lpd} - p_{\text{WAIC}}, $$ where $$p_{\text{WAIC}} = \sum_{c=1}^C \sum_{i=1}^{n_c} V_{s=1}^S(\log p(\mathbf{Y}_{ci}|\bm{\theta}^s)),$$ with $V_{s=1}^S$ represents the sample variance. We use the \textit{loo} package in R to compute $p_{\text{WAIC}}$ and $\widehat{lpd}$ from the MCMC output. The WAIC is then $$ \text{WAIC} = -2 \widehat{lpd} + 2p_{\text{WAIC}}. $$ When comparing a set of models, the model with lower WAIC is preferred. Unlike the Akaike Information Criterion \citep{akaike1973second} and DIC, WAIC averages over the posterior distribution and is asymptotically equivalent to Leave-one-out cross validation (LOO-CV), however using LOO-CV in the application with real data is not straightforward because there are incomplete observations and the data are clustered into cohorts. \subsubsection{Bayes Factor via bridge sampling} Bayes Factor \citep{berger2013statistical,kass1995bayes} is an important statistics for model comparison. Suppose we are choosing between two models $\mathcal{M}_1$ and $\mathcal{M}_2$, then the Bayes Factor for $\mathcal{M}_1$ versus $\mathcal{M}_2$ is the ratio of their marginal likelihood \begin{eqnarray*} \text{BF}_{12} = \frac{p(\mathbf{Y}|\mathcal{M}_1)}{p(\mathbf{Y}|\mathcal{M}_2)}. \end{eqnarray*} The marginal likelihood is $p(\mathbf{Y}|\mathcal{M}_a) = \int p(\mathbf{Y}|\bm{\theta},\mathcal{M}_a)p(\bm{\theta}|\mathcal{M}_a)d\bm{\theta}$ where $\bm{\theta}$ denotes the vector of parameters for each model and $\mathbf{Y}$ is the observed data. \cite{jeffreys1998theory} suggest that $\text{BF}_{12}>10^2$ is considered decisive evidence in favor of $\mathcal{M}_1$, and $10^{3/2}<\text{BF}_{12}<10^2$ is strong evidence supporting $\mathcal{M}_1$. The most challenging step in computing the Bayes Factor is the need to evaluate the marginal likelihood of the model, which is typically intractable. There are several methods to estimate the marginal likelihood and Bayes Factor, but we use bridge sampling \citep{meng1996simulating,gronau2017tutorial} as it only requires running the MCMC once; this method is more feasible than path sampling \citep{gelman1998simulating} and more straightforward than the method by \cite{chib2001marginal}. Bridge sampling is implemented in the R package \textit{bridgesampling} by \cite{gronau2017bridgesampling} and the computation can be done conveniently with some simple modification to the STAN code. In the SEM literature, the posterior predictive p-value (ppp) is often used to evaluate whether a SEM structure fits the data well \citep{kaplan2012handbook}. However for our application we would like to test whether the piecewise-linear regression equation is appropriate, compared to the linear regression model. Moreover \cite{asparouhov2010bayesian} show that, when using ppp, the rejection rate increases with sample size. Therefore we chose not to use ppp in our analysis. \subsection{Computational efficiency gained from a reduced model}\label{subsec:semiMarginalModel} For the model in Section \ref{subsec:Model}, all the latent variables $\eta_{ci}$ and $\bm{\xi_{ci}}$ will be sampled. This means the number of parameters is very large and the model is computationally intensive to estimate. Software such as \textit{lavaan} estimates a reduced form of the model instead by integrating out the latent variables and estimating the other parameters by maximum likelihood estimation. The package \textit{blavaan} estimates the latent variables in their BUGS implementation but does not do that in their STAN implementation. We are also interested in estimating the latent cognition score $\eta_{ci}$, however the latent variables of the subdomains are not needed, therefore we propose to integrate out only $\xi_{ci}$. Because $\mathbf{Y}_{ci}, \bm{\xi_{ci}}, \eta_{ci}$ are jointly normal, we have that \begin{eqnarray*} \mathrm{E}(\mathbf{Y}_{ci}|\eta_{ci}) &=& \bm \nu_c + \boldsymbol{\Lambda_c}(\boldsymbol{\gamma_c}\eta_{ci} ),\\ \mathrm{Cov}(\mathbf {Y}_{ci} |\eta_{ci}) &=& \boldsymbol{\Psi_c} + \boldsymbol{\Lambda_c\Phi_c\Lambda_c^\prime}. \end{eqnarray*} The reduced model is then \begin{eqnarray} \mathbf{Y}_{ci} & \sim & MN(\mathrm{E}(\mathbf{Y}_{ci}|\eta_{ci}), \mathrm{Cov}(\mathbf{Y}_{ci}) ), \label{eq:reducedModel1 } \\ \eta_{ci} &= &\beta_1 X_{ci} + \beta_2 (X_{ci} - X^{bp})_+ + \alpha_c P_{ci} + e_{ci},\quad e_{ci} \sim N(0,\sigma_c^2). \label{eq:reducedModel2} \end{eqnarray} The regression equation of the model remains the same as before. In our application with the data from six cohorts, running the full model takes 12 hours, but the reduced model only takes less than 4 hours to run. \section{Simulation example}\label{sec:simulation} We now illustrate how the model works for a simulated data set. In this experiment, we generate a data set consisting of six cohorts, each with 400 participants, according to the model in Section \ref{subsec:Model}. The covariate $X_{ci}$ is generated from $U(0,2.6)$, the propensity score $P_{ci}$ is generated from $N(0,1)$, and the true underlying cognition factor $\eta_{ci}$ is generated as $$\eta_{ci} = -2 \textit{X}_{ci} - 3(\textit{X}_{ci} - 1.3)_+ + \alpha_c P_{ci} + e_{ci},\quad e_{ci} \sim N(0,\sigma_c^2). $$ The loadings $\bm{\gamma_c}$ are the same for all cohorts. For each of the 3 lower level factors, we generate 5 outcome variables. The loadings corresponding to these outcomes vary across cohorts. The ``observed'' outcomes for each cohort are chosen randomly from these 15 outcomes, making sure that there are at least 3 variables corresponding to each factor. The cohort-specific variances $\sigma_c^2$ take values in $(2.5,1.5,2,3,2.5,1.8)$ and the intercepts $\boldsymbol{\nu_c}$ are generated from $N(0,3^2)$ for 6 cohorts. The $\alpha_c$ are taken randomly from $N(0,0.5^2)$ for $c = 1,\dots,6$. This data set is generated to reflect the condition of the real data where the tests used in the cohorts are different. The sample size is also chosen to be similar to that of the real data application. For the prior distributions, the non-zero elements of the loadings $\bm{\Lambda_c}$ are assigned prior $N(1,1)$ truncated to be positive to avoid indeterminacy; the non-zero elements of $\bm{\gamma_c}$ have prior $N(1,1)$ truncated to be positive, for $c = 1,\dots, 6$. The mean $\boldsymbol{\nu_c}$ are jointly normal $N(0,50)$ for all cohort $c$. The diagonal elements of $\bm{\Psi_c}$ are given prior $N(100,50)$, the square root of the diagonal elements of $\bm{\Phi_c}$ are assigned Inverse Gamma prior IG(2,3). We choose this parameterization because modeling the standard errors is more stable than modeling the variance of the $\bm{u}_{ci}$ directly. We use a log-normal prior $\text{Lognormal}(0,0.5)$ for the break-point. The variances $\sigma^2_c$ have a normal prior $N(3,1)$ for all cohorts $c$. We implement this model in the STAN probabilistic programming language \citep{carpenter2017stan}, and run 3 MCMC chains, each with $10,000$ iterations. The final result uses the last 5000 iterations in each chain. The estimates of the slopes and break-point are shown in Table \ref{tab:simmodel}, where the log of the break-point is reported. The model estimates the slopes and the break-point quite accurately, given a moderate sample size. Figure \ref{fig:simfactorest} plots the estimated cognition variable against its true value for each cohort and shows that the model is able to recover the true factor scores. \section{Application: Modelling the effect of prenatal alcohol exposure on cognitive and behavioural deficits in children} \label{sec:RealApplication} \subsection{Model setup} We now fit the multi-cohort model described in Section \ref{subsec:missingdata} to the six-cohort data set in Section \ref{sec:Data}. Each study collected a large number of observed variables, however we only include outcomes that are known to be associated with child cognition. Since the tests used varied across cohort, we set up the model by trying to match as many outcome variables across cohorts as possible. This means that we try to include the same or similar tests for all cohorts. Note that these are longitudinal studies; therefore, some tests were repeated. To avoid additional complications due to serial correlation, we only use results from one administration of those tests (mostly the first one). Table \ref{tab:datasum} summarizes the number of outcome variables for each cognition measure for each cohort. In this analysis, PAE is measured by the mother's average daily dose of absolute alcohol consumed during pregnancy (AA/day). However the distribution of alcohol exposure is positively skewed, therefore we take the natural log transformation of AA/day and use that in our model. More precisely we compute $\log(\text{AA/day} + 1)$ because the minimum level of AA/day is 0. $\bm{\Psi_c}, c= 1,\dots,6$ are modeled such that most of the diagonal elements are 0, except for a few entries where we assume there are additional correlations between the output, which cannot be explained by the SEM structure. The selection is done by first fitting the model in \textit{lavaan} with diagonal $\bm{\Psi_c}$, and computing the residuals. We then add correlation for pairs where the unexplained correlation from the residuals are large, and re-estimate the model. This procedure is repeated until there is no remaining large correlation or when the fit indices (for example RMSEA) are reasonably good. To ensure proper scaling, we rescale all outcome variables beforehand to have standard deviation 15. Similar to the simulation example, the non-zero elements of the loadings $\bm{\Lambda_c}$ are assigned prior $N(1,1)$ truncated to be positive; the non-zero elements of $\bm{\gamma_c}$ have prior $N(1,1)$ and truncated to be positive, for $c = 1,\dots, 6$. The mean $\boldsymbol{\nu_c}$ are jointly normal $N(0,50)$ for all cohorts $c$. The diagonal elements of $\bm{\Psi_c}$ are given prior $N(100,50)$, the square root of the diagonal elements of $\bm{\Phi_c}$ are assigned Inverse Gamma prior IG(6,20), based on the result from one run of \textit{lavaan}. The variances $\sigma^2_c$ all have normal prior $N(150,50)$ for all cohort $c$. Let $J_c$ be the number of residual correlation coefficients to be estimated for $\bm{\Psi_c}$. The $j^{th}$ correlation coefficient in cohort $c$, $\rho_{cj}$, is modeled via a $Beta(1,1)$ prior for $0.5(\rho_{cj} +1)$, for $j = 1,\dots, J_c,\quad c = 1,\dots 6$. We use a log-normal prior $\text{Lognormal}(0,0.5)$ for the break-point. We implement this model in STAN, running 3 chains of 8000 iterations and remove the first 3000 iterations in each chain as burn-in. To test the significance of a change point, we also fit a similar model, but this model assumes a linear effect of PAE on cognition $\eta_{ci}$ instead. We use the same priors as those for the piecewise-linear model in Equations \eqref{eq:modelFullMissing1}-\eqref{eq:modelFullMissing3}. This means that equation \eqref{eq:modelFullMissing3} becomes $$\eta_{ci} = \beta_1 X_{ci} + \alpha_c P_{ci} + e_{ci},\quad e_{ci} \sim N(0,\sigma_c^2). \label{eq:linearModel} $$ We refer to this model as the ``linear model'' in our discussion. We note that the distribution of many of the outcome variables deviate from normal, though not severely. The literature on Bayesian SEM with non-normal variables is limited; however, it is well known that maximum likelihood estimation provides unbiased and consistent parameter estimates even for non-normal data \citep{west1995structural}. Bayesian estimate of SEM is hence likely to be robust to non-normality. \subsection{Model assessment}\label{subsec:comparemodel} In Section \ref{subsec:GoodnessOfFit} we discuss using WAIC and Bayes Factor to compare candidate models. A challenge of computing these statistics with Bayesian SEM models is that the number of parameters grows with increasing number of participants. Therefore, the estimates of the log marginal likelihood from the R package \textit{bridgesampling}, which is needed for computing the Bayes Factor, become very unstable and unreliable. The package also requires running the MCMC for an unnecessarily large number of iterations because of the large number of parameters. For this analysis, we compare between using a piecewise-linear with a linear model for the relationship between PAE and child cognitive function. The structural relationship between the observed outcome variables and the latent variables is the same for both models. Therefore we choose to compute the marginal likelihood from a reduced model where the latent variable $\eta_{ci}$ is also integrated out. This does not change the inference of the slopes $\beta_1$ and $\beta_2$ but significantly reduces the number of parameters in the model and hence bridge sampling works much better. We still report the estimated parameters and discuss the result from the model in Section \ref{subsec:semiMarginalModel} because the reduced model does not provide inference on the cognition scores $\eta_{ci}$. \subsection{Results} Table \ref{tab:fullmodel} shows the posterior estimates of the 2 slopes $\beta_1$, $\beta_2$ and log of the break-point $\log(X^{bp})$. It is clear from the results that there is not much information about the location of the break-point, as its posterior does not differ significantly from the prior. This can also be seen clearly in Figure \ref{fig:bp}, which shows the kernel density estimates of the marginal posterior densities of $\log(X^{bp})$ and $X^{bp}$. The posterior distribution of $\log(X^{bp})$ does not differ significantly from its prior, a sign that the data are not informative enough to identify a break-point. Figure \ref{fig:betas} plots the kernel density estimates of the marginal posterior densities of the slope coefficients $\beta_1$ and $\beta_2$. The 95\% credible interval of $\beta_1$ does not contain 0, implies that there is clear evidence of a negative effect of log alcohol exposure on child cognition, even at lower doses. However, the increment in the effect after the break-point is not very clear: The posterior of $\beta_2$, clearly deviates from 0, but 0 is not too far from the mean of the distribution, and the posterior variance is still large. This result is consistent with the fact that there is not much data at the higher levels of PAE to detect the break-point. In addition, the total number of participants in this study is not large enough for such a complicated model. Figure \ref{fig:bpmodelfit} shows the posterior mean of the cognition score $\eta_{ci}$ for all individuals in all six cohorts together with the estimated regression line and uncertainty interval. Because the break-point $X^{bp}$ was sampled with the rest of the parameters, the estimated regression line is the average over the posterior of $X^{bp}$. The result dose-response curve (in red) therefore does not show a clear ``break'' but instead changes smoothly between $0.5$ and $1.5$. If the data provided more support for the piecewise-linear model, the posterior variance of $X^{bp}$ would be smaller, and we would expect a more obvious ``bend'' in the regression curve. We note that the cognition score varies widely at a given level of PAE, implying that the effect of PAE on cognition is weak. Figure \ref{fig:cogposterior} shows the posterior distribution of $\eta_{ci}$ for a few randomly chosen individuals. The traceplot of the all parameters indicate good mixing of the MCMC chains. Table \ref{tab:fullmodel} also shows the posterior estimate of the slope $\beta_1$ in the linear model. The result suggests a clear negative effect of PAE on cognition, which can be seen in Figure \ref{fig:pwvslinear}. The effect size is larger in magnitude than the first slope in the piecewise-linear model, and the posterior standard error is smaller. This may be due to the participants with very high levels of PAE, who also have lower mean cognition scores, as shown in Figure \ref{fig:pwvslinear}. These observations pull the regression line down and result in a steeper slope, compared to the piecewise-linear model where the effect is assumed to change at higher doses. The WAIC of the linear model is 263430.9, which is slightly lower than that of the piecewise-linear model (263443.4). However, given the large standard error of the estimates (both are approximately 2067), the two models are essentially the same in term of WAIC. The Bayes Factor is 2.32435, which means that the piecewise-linear model is slightly better than the linear model but the difference between the 2 models is negligible. One possible explanation of this result is that the PAE variable is an averaged measure of daily consumption, therefore it cannot capture the effect on those who drinks occasionally but have high dose per drinking occasion. \section{Discussion}\label{sec:discussion} In this paper we examine a new Bayesian SEM to analyze data from multiple studies with different sets of outcomes. The Bayesian framework allows for pooling the information across cohorts to analyze the relationship between the latent variable and the observed covariates. The model is flexible and can be used in different settings with different sets of outcomes and predictors. The framework provides considerable flexibility in the overall dose-response curve through changing the functional form of the regression component. We test the model in a simulation example and show that it can recover the true effect. We then use the model to analyze the dose-response curve between PAE and child cognition function, using data from six longitudinal cohort studies in the United States. We consider two different dose-response functions, a linear and a piecewise-linear function, in order to test for the significance of a change point in the dose response curve. While our method has a number of strengths, there are some limitations. Firstly, the Bayesian model is computationally expensive due to the large number of parameters and the large number of incomplete observations. We are able to reduce the computational time by integrating parts of the latent variables, however the MCMC sampling still takes a long time. Fast approximate inference, such as Variational Bayes (VB) methods \citep{attias2000variational} could potentially be used instead. However, VB methods are not currently available for SEM and deriving them is out of the scope of this paper. Secondly, the current estimation procedure depends on the assumption that the outcome variables are normally distributed, and so the model might not capture highly non-normal data very well. However relaxing the normality assumption will prevent us from using the reduced models and will likely increase the computational cost. Lastly, the model currently uses a pre-computed propensity score without incorporating the estimation errors into the model. Estimating the propensity score as a part of the model is likely to make it more computationally intensive, and therefore we leave it for future research. \section*{Acknowledgments} Khue-Dung Dang was supported by Australian Research Council Centre of Excellence for Mathematical and Statistical Frontiers (ACEMS) grant CE140100049. Richard J. Cook was supported by the Natural Sciences and Engineering Research Council of Canada through grants RGPIN 155849 and RGPIN 04207. Louise M. Ryan, Sandra W. Jacobson, Joseph L. Jacobson and Tugba Akkaya-Hocagil were supported by the National Institute on Alcohol Abuse and Alcoholism grant R01 AA025095 and the Lycaki-Young Fund from the State of Michigan. {\it Conflict of Interest}: None declared. \newpage \DeclareRobustCommand{\disambiguate}[3]{#3} \bibliographystyle{apalike}

c93b4823b7d89f6fb698e96e208b26070e901da4

\section{Introduction} In early 2020 the world has been taken by a very aggressive global pandemic, the covid-19, which spread around the entire planet at unprecedented speed in mankind history. As we speak, the pandemic, originated in Wuhan, China, allegedly in January 2020 has spread out over 100 countries, with over ten million contagion cases and over 500,000 casualties worldwide, as of end June 2020 \cite{covidnumbers}, giving rise to the trade-off between strong confinement measures to stave off devastating effects on health systems \cite{Armocida2020,Legido-Quigley2020} and economic, social and psychological terms \cite{Chen2020a, Duan2020, Nicola2020, Toda2020}. A distinctive feature of the covid-19 pandemia is the heterogeneity of the viral infection in space; in many countries a large fraction of the overall infection counts originated from very specific hotspots, such as Lombardy in Italy and NYC in the USA. This strong inhomogeneity calls, among others, for a proper modelling of the mechanism by which the infection propagates in space and time \cite{Lloyd1996}. In this paper we address this issue by isolating a toy-problem, namely the interaction between two infectious hotspots sitting at two separate locations in space. Special attention is paid to the spatial interference \cite{Dietz1979} between the two hotspots, in particular the way that the presence of the first affects the viral evolution in the second, depending on the mobility and infection rates in the two hotspots. To this purpose, spatial mobility is described by a simple advection-diffusion SIR (ADSIR) model, in which diffusion encodes small-distance mobility (say walking), while advection stands for mid-range mobility (say train or car driving). { Rather than being an exact model of these real-world mobility schemes -- for which network models would certainly be a more accurate choice -- diffusion and advection have the general purpose to realize two different mobility mechanisms with a different scaling in time and space. } Even though human mobility proceeds by more complex mechanisms than AD, typically encoded by mobility networks \cite{Ball2010,Lang2018,VespiCovid1,VespiCovid2}, the present ADSIR model exposes nonetheless a number of interesting qualitative features related the spatio-temporal coupling between the two infectious hotspots. \section{Mathematical formulation} We describe the covid-19 pandemic by means of a standard SIR model \cite{Kermack1927} -- which is the starting point for numerous interesting models in epidemiology \cite{Kaushal2020,Kaxiras2020,Pathak2010,Ji2011,Lang2018,Roda2020, Toda2020,Calafiore2020} -- coupled in space via an advection-diffusion mechanism: \begin{eqnarray} \label{ADSIR} \partial_t s = \nabla (-U s + D \nabla s) -\beta si\\ \partial_t i = \nabla (-U i + D \nabla i) +\beta si -\gamma i\\ \partial_t r = \nabla (-U r + D \nabla r) +\gamma i\ \end{eqnarray} where $sir(x,y,t)$ is the population of Susceptible, Infected and Recovered individuals at position $x,y$ and time $t$, respectively. The coefficients $\beta,\gamma$ correspond to infection and recovery rate, respectively. In the above equations $U$ is the wind speed, which we take aligned with the x-axis without loss of generality and $D$ is the diffusivity. The total number of individuals of species $k=s,i,r$ at a given time $t$ is thus given by integral over the entire domain of the corresponding densities: $ N_{k,h}(t) = \int_{H_i} n_k(x,y,t) dxdy, \;k=s,i,r, \;i=1,2 $ where the integral runs over the hotspot regions HS1 and HS2. \begin{figure}[t!] \centering \includegraphics[width=\linewidth]{domain.pdf} \caption{{The simulation domain.} Schematically shown is the simulation domain of length $L$. The boxes indicate the hotspots HS1 and HS2 with different contagion rate $\beta_i=f_i\times \beta_0$, where $\beta_0$ is the basic contagion rate in the normal domain. The hotspots at distance $d$ have width $w$ in both directions. The homogeneous wind $U$ is indicated by the black solid arrow and the red area indicates the small outbreak. } \label{fig:domain} \end{figure} The speed $U$ will be compared to an important intrinsic reference velocity:\\ The infected population $i(t)$ grows in HS1 as $i(t)\approx A\cdot e^{(s(t)/N\cdot f_1\cdot\beta_0-\gamma)t}$. Whenever the wind exceeds a critical speed $U_c=w\cdot (s(t)/N\cdot f_1\cdot\beta-\gamma)$ mitigation of epidemics is expected, due to the removal of infected individuals from the hotspot. For $U/U_c>1$, we expect the first hot-spot to become transparent and have little influence on the second one. This of course largely depends on the a priori unknown parameters $s(t)/N$ and the variable parameter $f_1$. For a reference, we take $f_1=f_2=50$ and $s(t)/N\approx0.5$. Hence, we define the reference speed $U_r=w\cdot(\frac{f_2\beta}{2}-\gamma)$. The main independent (dimensionless) parameters are then defined as follows: The contagion rate is $\beta_0=0.2$ in the normal domain and $\beta_i=f_i\cdot \beta_0$ in the hotspots $i=1,2$. The recovery rate is homogeneous $\gamma=0.15$, such that the reproduction factor is $R_i=f_i\cdot \frac{\beta_0}{\gamma}>1$. We fix $f_2=50$ while $f_1$ ranges from $1$ to $500$, so that the relative infectivity ratio $f=f_1/f_2\in\left(0.02,10\right)$. The diffusion coefficient is $D=5$. The width of the hotspots is $w=10$ and their distance is $d=50$. The wind speed is measured in units of the reference speed, i.e. $u=U/U_r$. The grid spacing is $1\ \text{km}$ and time is measured in days, corresponding to a reference speed $U_r=50\ \text{km}/\text{day}$ and a diffusivity $D=5\ \text{km}^2/\text{day}$. These are plausible scales for human mobility. We then study the solution of the ADSIR problem above as a function of the parameters $u$ and $f$. In particular, we wish to assess under what conditions the presence of HS1 causes an increase of infections in HS2 in terms of both peak intensity and duration. \section{Simulation setup and results} We set up two hotspots HS1 and HS2 of width $w$ at position $x_1=L/4$ and $x_2=x_1+d+w$ respectively. The domain is a grid of size $64\times 1024$. We place a small Gaussian outbreak at $x=0$ and $y=W/2$ with a cutoff at $x=10$ in order to ensure that there are no initial infections in the hot-spots. The boundary conditions are chosen to be fully periodic. Fig. \ref{fig:domain} shows the geometric set-up of the hot-spots in the domain. In Fig. \ref{fig:times} we show the typical evolution of the infected species in a single hotspot. The starting time of the hot-spot is defined as the time $t_s$ at which the infected population reaches $i(t_s)=0.002\cdot N$ and the end time $t_e$ at which the infected population has dropped to $i(t_e)=0.05\cdot i_{max}$. The time difference $\Delta t=t_e-t_s$ can be defined as the duration of the epidemic in this region. \\ \begin{figure}[t!] \includegraphics[width=\linewidth]{timeevolution.pdf} \caption{{Time evolution of the SIR populations in a single hotspot.} The baseline parameters are: $f_1=f_2=5$, $\beta=0.5$, $U=0$, $D=0.05$. The position of the hot-spot is at $x=L/4$. The red curve shows the infected population in the hotspot, the black x's mark the start-up time, peak time and the decay time of the local epidemic, respectively. } \label{fig:times} \end{figure} In Figure \ref{fig:profile}, we report a typical spatial interference pattern between HS1 on HS2. The figure clearly shows that the infected population generated in HS1 reaches up to HS2 and increases the local infection rate, thereby increasing the peak and possibly the duration as well. This is the typical scenario that HS2 policy makers endeavour to combat via lock-down measures. \begin{figure}[t!] \centering \includegraphics[width=\linewidth]{profile.pdf} \caption{{Spatial distribution of the infected population.} The y-integrated population with respect to the $x$-position. The hotspots are marked with vertical lines. The parameters are $f_1=f_2=50$, $\beta=0.2$, $\gamma=0.15$, and $u=0.28$, $d=50$, $w=10$, $D=0.05$. The snapshots are taken at the start-up time of both hotspots and the peak time of the second one. The development of a spatial interference between the two hotspots is clearly visible. The HS2 peak for the homogeneous case is $0.99 \times 10 = 9.9$, against an observed one of about 12, showing a 20 percent increase due to spatial interference from HS2.} \label{fig:profile} \end{figure} \textit{Peak and duration of the epidemics.}--In Fig. \ref{fig:close_wind}, we summarize the effect of the wind and HS1 infectivity on the peak intensity of HS2. \begin{figure}[t!] \includegraphics[width=\linewidth]{small_wind_frac.pdf} \caption{{Peak intensity in HS1 as a function of the wind speed at varying the infection rate in HS2} The simulation parameters are the same as in figure 3. We clearly observe the emergence of a non-monotonic wind speed regime in the range $1<f<5$, followed by a loss of any beneficial wind effect above $f \sim 5$. } \label{fig:close_wind} \end{figure} We measure the dimensionless peak value $i_{max}/N$ in the second (downstream) hotspot as a function of $u$ and $f$, where $N$ is the total number of individuals in the hotspot and $i_{max}$ is the peak value of the infected population. \\\ A few comments are in order. First, we see that the peak intensity is a fast decreasing function of the wind speed for all HS1 infection ratios well below the HS2 values. This is expected, since the infected in HS2 get replaced by less infected from HS1. However, upon increasing $f_1$ in the vicinity and then above $f_2$, a shoulder appears at intermediate wind speeds, indicating that a highly contagious mobile population from HS1 is capable of spoiling the beneficial effect of the wind. This is also a plausible result, since the infected removed by the wind in HS2 are quickly replaced by even more infected transmitted by HS1. This is the typical scenario dreaded by southern Italy towards the "stampede" from northern regions in the early stage of the Italian epidemics. Our simple model shows that such fears were indeed justified, an infectivity ratio $f=2$ is already capable of producing a secondary peak in the curve and raising $f$ only makes the situation worst, with the emergence of a whole range of wind speeds in which the peak intensity grows instead of decaying, almost reaching up to the value of the windless case. Since this strongly reminds of the unstable region of a non-ideal equation of state, in which pressure goes down upon increasing density (condensation), we dub this effect "epidemic condensation". This is the main result of this paper, as it highlights the existence of an optimal wind speed $u_{min} \sim 0.5$ which minimises the HS2 peak, and a second, higher, characteristic speed $u_{max}$, beyond which the beneficial effects of the wind are restored. By further increasing the relative infectivity of HS1, between five and ten, no decay of the peak intensity at increasing wind speed above $u_{min}$ is observed anymore {in the simulated window of the wind speed $u$}, indicating that the presence of HS1 completely cancels any benefit of the wind speed above the optimal value $u_{min}$. {However, for a very large wind speed $u\to\infty$ we expect again a decrease of the infected ratio, since for infinite wind speed the hot-spots become transparent again, and infectivity will have no impact.} In Fig. \ref{fig:dur_u} we report the duration of the epidemics as a function of $u$ and $f$. A major peak is observed at low-wind, corresponding to the fact that the infected are convected away at very low rates. As expected, high infectivity goes with high peaks and short durations, the dreaded scenario for intensive care departments. As the wind speed increases, the local infected are efficiently removed and the epidemic duration shortens. However, starting from comparatively low infectivity ratios, $f=0.2$, further satellite peaks appear, indicating the existence of a sequence of wind speeds such that the duration grows back, if only mildly. This is again interpreted as a spatial interference effect, although we must caution that such measurement is very sensitive to small changes of the duration threshold, hence should be taken with great caution. \begin{figure}[t!] \includegraphics[width=\linewidth]{duration_close_u.pdf} \caption{{Duration of the epidemic in HS2.} The curve shows a peak at very low wind speeds $u \sim 0.01$, followed by a sequence of secondary peaks at higher speeds, all well below $u_{min}$. By and large, wind speeds above $u=0.1$ are consistently beneficial in shortening the epidemic duration. } \label{fig:dur_u} \end{figure} \section{Qualitative scenario and discussion} The ADSIR model presented in this paper focusses on the effects of spatial coupling, advection and diffusion, on epidemic growth as dictated by local infection rates. It is well known that in the presence of random heterogeneities, such coupling can lead to highly nontrivial behaviour, such as the formation of striated infection highways \cite{Chotibut2017}. Here we take a simpler model problem, namely the effect of a primary hotspot (HS1) on the epidemic growth on a secondary hotspot (HS2) downstream HS1. In particular, we focus on the effect of a uniform "wind" at speed $u$, mimicking a uniform human mobility across the two hotspots. In the absence of any wind, $u=0$, and discounting diffusion, the two hotspots evolve independently based on their corresponding infection rates. As soon as the wind is switched on, a beneficial effect is expected for both HSs because the wind sweeps infected individuals away into the "country side", where the chance to infect is much lower and healing can proceed nearly undisturbed. This is certainly true as soon as the wind speed exceeds the infection speed, namely the size of the hotspot divided by the typical infection timescale (reference speed), because, under such conditions, the wind blows susceptible individuals away before they have time to get significantly infected. So, the baseline expectation is that "wind is good", as it gives no time for infection to develop substantially. This is true for HS1, but not necessarily for HS2, which is exposed to the incoming flux of infected individuals from HS1. The quantitative question is whether, from the HS2 perspective, there exists an optimal wind speed which corresponds to a local minimum of the infection peak . In the following, we shall present evidence that the answer is in the affirmative. In particular, it is shown that as soon as HS1 is more infectious than HS2, the peak intensity in HS2 develops a much slower decay with the wind speed, and when HS1 is significantly more infectious than HS2, the HS2 peak increases at increasing wind speed, before it starts to decay again in the strong wind regime. In other words, the HS2 peak develops a non-monotonic dependence on the wind speed, with a local minimum, $u_{min}$ at about half the reference speed and a local maximum $u_{max}$ about twice as large. Such non-monotonic dependence bears an intriguing resemblance to a non-ideal equation of state, with the unstable branch in the wind speed region $u_{min} \le u \le u_{max}$. Because of this close resemblance to equation of state of non-ideal gas, and most notably to the unstable region where a density increase leads to a pressure decrease (condensation), we dub this effect {\it epidemic condensation}. We also monitor the duration of the epidemics as a function of the wind speed and infection rates. Note that while the peak intensity is the prime concern for health capacity issues, the duration bears directly on the mid-long term policies towards social and economic impact (many countries insisted on "curve flattening" policies). Again, we find that wind increase above a very low threshold is generally beneficial, although at increasing HS2 infectivity, the duration increases and shows repeated small-amplitude "sawtooth" oscillations. Such oscillations are yet another signature of spatial coupling, although their specific nature remains to be fully ascertained. \section{Effect of the hotspot distance and the diffusivity} We also inspected the effect of the hotspot distance on epidemic condensation. To this purpose, we ran a series of simulations at different wind speeds and distances in the range $50 \le d \le 200$, keeping a fixed value $f=4$. As expected, the local maximum observed in the condensation decreases with the distance and, less expectedly, so does the local minimum. \begin{figure}[t!] \centering \includegraphics[width=\linewidth]{dist_minimum.pdf} \caption{The local minimum and maximum of the condensation curve. The infectivity ratio is set to be $f=4$, all other parameters are chosen as in Fig. \ref{fig:close_wind}. Shown are the local minimum and maximum of the curve of the infected ratio in HS 2. The power-law fit is performed with a small exponent, $d^{-0.32}$, hence the correlation effects decay slower than the distance.} \label{fig:dist} \end{figure} {Fig. \ref{fig:dist} shows that} both quantities decay according to an inverse power law $d^{-\alpha}$, with $\alpha \sim 1/3$, indicating that the correlation between the two hotspots decays much more slowly than their inverse distance. To assess the effect of the diffusivity, we computed the condensation curve -- similar to Fig. \ref{fig:close_wind} with fixed $f=4$, for different values of the diffusion constants $D$. \begin{figure}[t] \centering \includegraphics[width=\linewidth]{diff.pdf} \caption{Effect of diffusion. We change the diffusivity and inspect its effect on the condensation curve for an intermediate high infectivity ratio $f=4$. For small $D$, we observe no qualitative change of the curve. $D=5\,km^2/day$ is the value we used throughout the rest of this paper. When the Fischer speed $U_f=\sqrt{Df_1 \beta_0}$ approaches the reference wind speed $U_r$, we expect additional interference effects which may lead a revival of infectivity in HS2. A detailed study of these phenomena is left to a future study.} \label{fig:diff} \end{figure} In Fig. \ref{fig:diff} we observe a quantitative effect of the diffusion parameter on the condensation curve, due to the fact that diffusion smears out sharp spatial changes in population number, such as those observed at the hot-spot boundaries. Hence, the effect of increasing the diffusivity $D$ is similar to lowering the infectivity ratio $f$, as long as the diffusivity remains sufficiently small enough, meaning by this that the Fisher speed $U_f=\sqrt{D f_1\beta_0}$ remains well below the reference wind speed. In the simulations carried out here, $U_f$ is always significantly smaller than $U_r$, hence these effects do not play any role. Whenever the condition $U_f \ll U_r$ is violated, non trivial interference effects are expected, which may eventually lead to a revival of infectivity in the second hotspot. A detailed analysis of these effects warrants a separate study on its own, hence it is deferred to future investigations. \section{Conclusions} Summarizing, we have evidenced a non-monotonic relation between the wind speed and the peak intensity on the downstream hotspot as a function of the infectivity ratio with respect to the upstream one. Despite its drastic simplification of the mechanism of human mobility, it is hoped that the non-monotonic "constitutive relation" revealed by the present ADSIR model, as shown in Figures 4 and 5, may offer useful qualitative clues on the effects of spatial interference between infected hotspots. \section*{Acknowledgements} SS kindly acknowledges funding from the European Research Council under the European Union’s Horizon 2020 Framework Programme (No. FP/2014-2020)/ERC Grant Agreement No. 739964 (COPMAT). JD was supported by the ERASMUS program and Physics Advanced program of the Elite Network Bavaria (University of Regensburg, Germany). One of the authors would like to acknowledge useful discussions with Prof E. Marinari and G. Parisi in an early stage of the project. \bibliographystyle{ws-ijmpc}

be31ff1cb1164e823d96194fbe0d8b40b7d8d8c7

\section{Introduction} Gravitational Waves (GW), first introduced by Einstein in 1916 \cite{einstein1,einstein2} as a linear regime of the field equations of General Relativity, have been directly detected for the first time in 2015 by the LIGO/Virgo collaboration~\cite{LIGO}. These ground-based detectors use laser interferometry techniques, but there also exists other detection strategies. Einstein's Equivalence Principle implies that all types of energies produce and experience gravity in the same way. The energy of electromagnetic (EM) radiation must therefore source gravity just like compact objects do and, the other way around, gravity (e.g. gravitational waves) can manifest itself in the physical characteristics of electromagnetic radiation. This is the basic principle behind the EM detection (or emission) of GWs. The use of EM fields to both generate and detect GWs has been actually considered for decades, for instance based on the (inverse) Gertsenshtein effect~\cite{gertsenshtein} relying on the coupling between GWs and EM waves in the presence of a strong static magnetic field. However, the weakness of this coupling makes any GW detection extremely challenging. Therefore experimental efforts have mostly concentrated on ground-based detectors based on laser interferometry, probing a rather low frequency range (typically $1-10\, {\rm kHz}$). But this technique is not suited for the detection of high-frequency GWs (HFGWs), at the opposite of EM-based detectors. The Section~\ref{sec:hist} presents a short historic review of HFGW detectors that have been build or proposed. The interested reader will find more details about detection strategies and potential high-frequency sources in a recent review~\cite{Aggarwal:2020olq}. In this paper, we propose two novel patented~\cite{patent} experimental designs of resonant high-frequency EM detectors, based on the inverse Gerstenshtein effect, operating at MHz or GHz frequencies and feasible with current technology. We compute numerically the EM signal produced by passing HFGWs and consider planetary-mass primordial black holes (PBHs) as their potential sources. The detectors are constituted by either a waveguide or a cavity immersed into a transverse static magnetic field. This outer magnetic field deserves two purposes in our design. First, a proper (transverse) orientation of this field is mandatory to convert HFGWs into EM waves through the inverse Gerstenshtein mechanism. Second, the external magnetic field boosts the output signal through a resonance mechanism on specific radiation modes that are excited by the passing HFGWs. There exists a broad range of hypothetical astrophysical or cosmological sources of HFGWs with frequencies above kHz (see~\cite{Aggarwal:2020olq} and references therein), such as exotic compact objects, PBHs, inflation, reheating, oscillons, cosmic strings and other topological defects, large curvature fluctuations and phase transitions in the early Universe. Some sources produce transient signals, like the merging of compact objects, while cosmological sources typically induce a permanent stochastic GW background. In particular, HFGW detectors operating at MHz and GHz frequencies might probe new Physics up to the scale of the Grand Unified Theories. In this paper, we focus on the particular case of transient signals from the merging of light (planetary-mass scale) PBHs. Their existence is still hypothetical but motivated by several recent observations, such as microlensing events towards the galactic bulge, recently detected by OGLE~\cite{Niikura:2019kqi,Mroz2017} and suggesting that such PBHs may constitute between 1\% and 10\% of the Dark Matter (DM). LIGO/Virgo observations provide additional motivations for the existence of PBHs in the stellar range~\cite{Bird:2016dcv,Clesse:2016vqa,Sasaki:2016jop} and a unified scenario with a wide mass distribution imprinted by the known thermal history of the Universe has been presented in~\cite{Carr:2019kxo}. PBHs may originate from the gravitational collapse of primordial inhomogeneities in the early Universe. Above a mass $m_{\rm PBH} \approx10^{11}$ kg, their evaporation time through the Hawking-Bekenstein mechanism is much larger than the age of the Universe. Beyond these motivations, the GW waveform from PBH mergers is well-known, which allows us to simulate the exact detector response. Our work therefore aims at paving the road towards an experimental realization of a HFGW detector with a good enough strain sensitivity, of order $h\sim 10^{-30}$, to detect various GW sources. The paper is organized as follows: After reviewing previous proposals of EM-based HFGW detectors in Section~\ref{sec:hist}, we introduce in Section~\ref{sec:EM-G} the theory behind the Einstein-Maxwell system and the Gertsenshtein effects. Section~\ref{sec:det} applies the theory to the case of two detector designs based on a resonant waveguide and a cavity, into an external static magnetic field. Section~\ref{sec:PBH} is devoted to the computation of the expected rate and signal from planetary-mass PBH mergers. In Section~\ref{sec:det-PBH} we gather these calculations to compute the expected signal in our detectors, in terms of frequency response and induced EM power, for different realistic designs and different PBH masses. Finally, we compute forecasted limits on the possible PBH abundance, for two possible binary formation channels (primordial binaries and tidal capture in clusters). We discuss our results and present some perspectives in the conclusion (Section~\ref{sec:conc}). \section{Brief History of Electromagnetic and HFGW detectors} \label{sec:hist} The use of EM fields to both generate and detect GWs has been actually considered for decades. While Weber~\cite{weber} envisioned the importance of both the generation and detection of GWs as early as 1960, Gertsenshtein~\cite{gertsenshtein} discovered in 1962 a resonance mechanism allowing to produce GWs from EM waves in the presence of a strong static magnetic field. Later in~\cite{boccaletti,delogi}, this mechanism was studied in greater details using Einstein's equations, scattering theory and Feynman perturbation techniques. Gertsenshtein's mechanism was applied to astrophysics by Zeldovich~\cite{zeldovich}. Grishchuk and Sazhin then introduced in~\cite{grishchuk1,grishchuk2,grishchuk2bis} purely electromagnetic generators of GWs, using transverse magnetic/electric (TM/TE) resonant cavities. They envisioned GW emission-reception laboratory experiments and concluded that they might be experimentally feasible~\cite{grishchuk2,grishchuk2bis}, which would open the road to futuristic technologies based on GW physics~\cite{grishchuk3}. Resonant cavities and EM waveguides were then considered as possible detectors of gravitational radiation, either emitted by natural or artificial sources~\cite{braginskii1,braginskii2,grishchuk2,grishchuk2bis,grishchuk3,pegoraro1,pegoraro2,caves,gerlach}. EM detectors of GWs would allow exploring higher frequency range than with laser interferometry, typically from kHz to 100 GHz when using radio-frequencies or from 100GHz to THz when using microwaves. Investigations of electromagnetic GW detectors began in the 1970's with the works~\cite{braginskii1,braginskii2,grishchuk2,grishchuk2bis,boccaletti,pegoraro1,pegoraro2,caves,cruise}. Those detectors either make use of the conversion of GWs to photons \cite{boccaletti}, the excitation or modification of resonant modes of EM cavities and waveguides~\cite{braginskii1,braginskii2,grishchuk2,grishchuk2bis,pegoraro1,pegoraro2,caves,ballantini}, the change of polarization plane of an EM wave due to the passing GWs~\cite{cruise}, or induced birefringence of the interior of the cavity~\cite{gerlach}. More recently, Ejlli, Cruise et al. in~\cite{Ejlli2019} used available data from experiments designed for the detection of weakly interacting slim particles to set limits on the stochastic GW background through the graviton to photon conversion in the ultra-high frequency band (above 1 THz). There has been also a method proposed by Bentley et al.~\cite{bentley} to reduce the noise of interferometric GW detectors at high-frequency (kHz). Other types of HFGW detectors have been recently proposed~\cite{gwdetect1,Aggarwal:2020umq,Ito2020,Aggarwal:2020umq,Goryachev2014}, using optically trapped dielectric microspheres in a cavity, resonance between a graviton and a magnon that is based on the Dirac equation in a curved spacetime or high frequency phonon trapping acoustic cavities. \section{Einstein-Maxwell system and the Gertsenshtein effects} \label{sec:EM-G} The Einstein-Maxwell system models the interplay of gravitation and electromagnetism by the coupling of their respective field equations (in S.I. units\footnote{The relevant fundamental constants of the Einstein-Maxwell system are $G$ as Newton's constant, $c$ as the speed of light and $\mu_0$ as the (vacuum) magnetic permeability.}) \begin{eqnarray} R_{\mu\nu}&=&\frac{8\pi G}{c^4} T^{\left(\rm em\right)}_{\mu\nu} \; \label{einstein}\\ \nabla_\mu F^{\mu\nu}&=&-\mu_0 J^\nu\; \label{maxwell} \end{eqnarray} where \begin{equation} T^{\left(\rm em\right)}_{\mu\nu}=\frac{1}{\mu_0}\left(g^{\alpha\beta}F_{\mu\alpha}F_{\nu\beta}-\frac{1}{4}g_{\mu\nu}F_{\alpha\beta}F^{\alpha\beta}\right)\label{tmunu_EM} \end{equation} is the Maxwell stress-energy tensor, $F_{\mu\nu}=\partial_\mu A_\nu-\partial_\nu A_\mu$ being the Faraday tensor of the electromagnetic field, $g_{\mu\nu}$ and $A_\mu$ are the metric and the four-vector potential, $\nabla_\mu$ is the covariant derivative with respect to $g_{\mu\nu}$, $R_{\mu\nu}$ is the Ricci tensor and $J^\nu$ the four-current density. In general relativity, the interplay between gravitation and electromagnetism is twofold: first, spacetime is curved by the energy of the electromagnetic field as ruled by Eq.~(\ref{einstein}) while, at the same time, the propagation of the electromagnetic field $F_{\mu\nu}$ is governed by the covariant Maxwell equations in curved spacetime, Eq.~(\ref{maxwell}). For a small electromagnetic compactness $G E_{\rm EM}/(c^4 L)\ll 1$, where $E_{\rm EM}$ is the EM energy stored in some physical system of length $L$, the gravitational sector of the system can be safely treated in the weak-field limit~\cite{fuzfa}. First, one can consider the gravitational perturbations arising from EM sources in the Einstein field equations. Considering EM configurations consisting in a superposition of a static field $F^{(s)}_{\mu\nu}$ and a varying one $F^{(v)}_{\mu\nu}$, the quadratic terms in Eq.~(\ref{tmunu_EM}) yield to three general classes of electromagnetically-induced gravitational perturbations around a Minkowski background \begin{equation} g_{\mu\nu}=\eta_{\mu\nu}+c_{\mu\nu}+w_{\mu\nu}+h_{\mu\nu}\label{decomp} \end{equation} where $\eta_{\mu\nu}=\rm diag(-1,+1,+1,+1)$ is the Minkowski metric and $c_{\mu\nu},w_{\mu\nu},h_{\mu\nu}\ll 1$ represent the metric perturbations. In the above equation, (i) $c_{\mu\nu}$ represents the static gravitational field generated by some external static magnetic or electric field (from a coil or a capacitor) and arises from the quadratic term in $F^{(s)}_{\mu\nu}$ in~(\ref{tmunu_EM}), while (ii) $w_{\mu\nu}$ is a gravitational wave generated by the varying EM [for which the source is the quadratic term in $F^{(v)}_{\mu\nu}$ in Eq.~(\ref{tmunu_EM})] and (iii) $h_{\mu\nu}$ is another gravitational wave generated by the coupling between the external static field $F^{(s)}_{\mu\nu}$ and some EM wave $F^{(v)}_{\mu\nu}$ [crossed terms in Eq.~(\ref{tmunu_EM})]. The case (i) has been studied in~\cite{fuzfa} but does not give rise to a GW since the outer EM field is static. Case (ii) has been studied for light propagating pulses in~\cite{tolman,ratzel} and in~\cite{grishchuk1,grishchuk2,grishchuk2bis,tang} for EM waves in resonant hollow cavities\footnote{In~\cite{grishchuk1}, a special case of hollow spherical cavity in an outer radial magnetic field is briefly considered, giving rise to an admixture of terms $w_{\mu\nu}$ (case (ii)) and $h_{\mu\nu}$ (resonance - case (iii)) which were not identified as such nor exploited by the authors.}. Case (iii) actually corresponds to what is called the direct Gertsenshtein effect~\cite{gertsenshtein,zeldovich}. This Gertsenshtein effect is a wave resonance mechanism in which light passing through a region of uniform magnetic field, perpendicular to the direction of light propagation, produces GWs. A monochromatic EM wave leads to an outgoing GW of same frequency. Electromagnetic generation of GWs is a very faint process, due to the extreme weakness of gravitational coupling. Indeed, the metric perturbations $h_{\mu\nu}$ produced through the Gerstenshtein mechanism have an amplitude of order \begin{equation} h_{\mu \nu} \sim \frac{4 G B_0 E_0 L^2}{c^5\mu_0}~, \end{equation} where $L$ is the size of the region in which the magnetic field and the EM wave interact, and where $B_0$ and $E_0$ are the amplitudes of the static magnetic and varying electric fields respectively. To give an idea, in order to generate a strain $h\approx 10^{-21}$ with $B_0\approx 10$ T and $E_0\approx 1 \, {\rm MV/m}$, one needs a truly astronomical size of the interacting region, $L\approx 10^6\; {\rm km}$. Therefore, while the direct Gertsenshtein effect can be used to build electromagnetic GW generators, its practical application constitutes an extreme experimental challenge. Second, ripples in spacetime can also interact with a static magnetic field to produce an outgoing EM wave. This inverse Gertsenshtein effect is described by the Maxwell equations, Eq.~(\ref{maxwell}), on a perturbed background. An obvious application of this effect is the detection of GWs passing into a transverse static magnetic field, yielding to their conversion into EM waves. One possible way to derive the equations governing the inverse Gertsenshtein effect is to develop the covariant derivative in Eq.~(\ref{maxwell}) at first order in metric perturbations, and to treat these ones as an effective current density. We follow here a different approach, based on a covariant generalization of the EM wave equations\footnote{These are obtained from the two groups of covariant Maxwell equations $ \nabla_\mu F^{\mu \nu}= 0$ and $\nabla_\kappa F_{\mu \nu} + \nabla_\mu F_{\nu \kappa} + \nabla_\nu F_{\kappa \mu}=0$ which can be combined to retrieve Eq.~(\ref{wavenormal}).} \begin{equation} \label{wavenormal} g^{\alpha\beta}\nabla_\alpha\nabla_\beta \tensd{F}{\mu \nu} + \tensd{R}{\mu \nu \alpha \beta} \tensu{F}{\alpha \beta} + \tensud{R}{\alpha}{\mu} \tensd{F}{\nu \alpha} + \tensud{R}{\alpha}{\nu} \tensd{F}{ \alpha \mu} = 0, \end{equation} where $\tensd{R}{\mu \nu \alpha \beta}$ is the Riemann tensor. This set of equations describes the propagation of EM waves on a curved spacetime. In the case of a flat Minkowski spacetime, the wave equation Eq.~(\ref{wavenormal}) reduces to the classical wave equation $ g^{\alpha\beta}\nabla_\alpha\nabla_\beta \tensd{F}{\mu \nu}=0$. Let us now consider a small perturbation $h_{\mu \nu}$ propagating on a Minkowski background, such that the metric is given by $g_{\mu \nu}=\eta_{\mu \nu} + h_{\mu \nu}$ at first order ($h_{\mu \nu} \ll 1$). If this perturbation satisfies the Lorenz gauge condition, $\partial_{\mu} h^{\mu \alpha} =0$, one gets from Eq~(\ref{wavenormal}) the linearized wave equation for the Faraday tensor (see also~\cite{grishchuk2bis}), \begin{widetext} \begin{align} \begin{split} g^{\alpha\beta}\nabla_\alpha\nabla_\beta \tensud{F}{(1)}{\mu \nu} &= \tensu{h}{\alpha \kappa} \nabla^{(\eta)}_\alpha \nabla^{(\eta)}_\kappa \tensud{F}{(0)}{\mu \nu} - \partial_\rho\left(\partial_\mu \tensd{h}{\alpha \nu}-\partial_\nu \tensd{h}{\alpha \mu}\right) \tensu{F}{(0) \, \rho \alpha} \eol -\left(\partial^\gamma \tensu{h}{\alpha \beta} + \partial^\alpha \tensu{h}{ \beta \gamma} - \partial^\beta \tensu{h}{\gamma \alpha} \right) \left(\tensd{\eta}{\alpha \mu} \nabla^{(\eta)}_\gamma \tensud{F}{(0)}{\nu \beta}-\tensd{\eta}{\alpha \nu} \nabla^{(\eta)}_\gamma \tensud{F}{(0)}{\mu \beta}\right)=S_{\mu \nu}. \label{lineq} \end{split} \end{align} \end{widetext} In the above equation, we have assumed that the total EM field $F_{\mu \nu}$ is the superposition of some background EM field $\tensud{F}{(0)}{\mu \nu}$ with which the metric perturbation $h_{\mu\nu}$ interacts to produce an EM perturbation $\tensud{F}{(1)}{\mu \nu}$. In the following, we focus on the magnetic conversion of GWs into photons, i.e. the interaction between a passing GW and an external static magnetic field ($\tensud{F}{(0)}{\mu \nu}$ is purely magnetic) which results in an EM wave emission. Eq.~(\ref{lineq}) governs the inverse Gertsenshtein effect. Under the assumption of a uniform static magnetic field, the first and third source terms in Eq. (\ref{lineq}), both including $\nabla^{(\eta)}_\gamma \tensud{F}{(0)}{\alpha \beta}$, identically vanish, leaving as the only source term the one with second derivatives of the metric perturbations, \begin{equation} S_{\mu\nu}= -\partial_\alpha \left(\partial_\mu \tensd{h}{\beta \nu}-\partial_\nu \tensd{h}{\beta \mu}\right) \tensu{F}{(0) \, \alpha \beta}\cdot\label{Smunu} \end{equation} We can now apply this theory and conceive a specific experiment for the detection of HFGWs, produced e.g. by inspiralling PBHs. \section{Resonant Electromagnetic detectors of HFGWs} \label{sec:det} In this section, we describe two detector designs that are based on the patents~\cite{patent}. The detection principle is based on the inverse Gertsenshtein effect, thus a passing GW interacts with an intense static magnetic field. If the direction of the incoming GW is not colinear with the magnetic field, faint transverse EM waves are generated and these can be further amplified by EM resonators. The experimental set-up consists of either a waveguide or a cavity whose axis of symmetry is orthogonal to the magnetic field. Such a set-up is similar to haloscopes that are used for the search of axions, like the ADMX experiment~\cite{ADMX,ADMX2}, except for the orientation of outer magnetic field. As shown below, it is mandatory that it is orthogonal to the axis of the cavity/waveguide in order to detect GWs. A theorem by Choquet-Bruhat~\cite{choquet} establishes that both direct and inverse Gertsenshtein effects require the condition of orthogonality between the external EM field and the direction of GW propagation. This theorem starts from the hypothesis that incoming or generated GWs have a Wentzel-Kramers-Brillouin (WKB) form and are of high-frequencies. The equation of propagation in~\cite{choquet} is then obtained after a development in frequency. We propose below a variant of Choquet-Bruhat's demonstration with a development in amplitude instead of frequency. We first assume that the incoming GW is a plane wave, $h_{\mu \nu} = a_{\mu \nu}\,\,e^{i\omega\Phi},$ with a general varying phase, $\Phi=\Phi(x^\alpha)$. The goal is to show that the constant EM field must be orthogonal to the direction of propagation of the incoming plane wave, in order to produce an EM wave. In other words, no EM wave can be generated from the interaction between a constant EM field and a GW, unless the first one is orthogonal to the direction of propagation of the second one. In order to show this, we demonstrate that a vanishing source term $S_{\mu \nu} = 0$ in Eqs.~(\ref{lineq}) is equivalent to the condition $ \Phi_\alpha F^{\alpha \mu\, (0)} = q\,\Phi^\mu$ with $q$ a real constant and $\Phi_\alpha = \partial_\alpha \ \Phi $. Since we assume that $E^{(0)}$ and $B^{(0)}$ are constants in our problem, the source term is given by Eq.(\ref{Smunu}). Therefore, the non-zero part of $S_{\mu \nu}$ is the exterior derivative of an effective 4-current density: $J^\mathrm{eff}_{\mu}=\partial_\alpha\,h_{\beta \mu} \, F^{\alpha \beta\,(0)}$. We can thus rewrite our source term as \begin{subequations} \nonumber \begin{align} S_{\mu \nu} &= \partial_\nu\,J^\mathrm{eff}_{\mu} - \partial_\mu\,J^\mathrm{eff}_{\nu}\\ &= (\Phi_\nu\,J^\mathrm{eff}_{\mu} - \Phi_\mu\,J^\mathrm{eff}_{\nu})' \quad \mathrm{with} \quad (\cdot)'\equiv \partial(\cdot)/\partial\Phi \, . \end{align} \end{subequations} The last line is obtained as a result of the plane wave approximation (the amplitude $a_{\mu\nu}$ above is constant), which implies that the partial derivative $ \partial_\nu $ is equivalent to $\Phi_\nu\,\frac{\partial}{\partial \Phi}$. Because a GW verifies the eikonal $\Phi_\mu \Phi^\mu = 0$ (if not, it is inconsistent and vanishes with a change of coordinate \cite{choquet}), multiplying the above expression by $\Phi^\mu$ gives that $S_{\mu \nu} = 0 \Leftrightarrow \Phi^\mu J^\mathrm{eff}_{\mu} = 0$, and thus $\Phi_\nu$ is orthogonal to $J^\mathrm{eff}_{\mu}$. Since $\Phi_\mu$ does not vanish, this yields to the equivalence $S_{\mu \nu} = 0 \Leftrightarrow J^\mathrm{eff}_{\mu} = 0$. Let us now show the central result \begin{equation} S_{\mu \nu} = 0 \Leftrightarrow J^\mathrm{eff}_{\mu} = 0 \Leftrightarrow \Phi_\alpha F^{\alpha \mu\, (0)} = q\,\Phi^\mu\label{thm}~. \end{equation} For plane waves $h_{\mu \nu} = a_{\mu \nu}\,\,e^{i\omega\Phi(x^\alpha)}$, we have that $\partial_{\alpha} h_{\lambda \nu}= \Phi_{\alpha} h'_{\lambda \nu}$. Therefore, the effective four-current can be simplified to \begin{subequations} \nonumber \begin{align} J^\mathrm{eff}_{\ \nu} = F^{\alpha \lambda\, (0)}\, \Phi_{\alpha} \, h'_{\lambda \nu} \end{align} \end{subequations} If $\Phi_\alpha F^{\alpha \mu \, (0)} = q\,\Phi^\mu$ then $J^\mathrm{eff}_{\ \nu} = q\,\Phi^{\lambda}\,{h'}_{\lambda \nu}.$ In the meantime, the Lorenz gauge condition, $\partial^\mu \,h_{\mu \nu} = 0$, is equivalent to $\Phi^{\mu}\,{h'}_{\mu \nu} = 0$ in the plane wave approximation. So we can conclude that the effective 4-current density $J^\mathrm{eff}_{\ \nu}$ vanishes . Now, let us now prove the implication in the reverse way. Let us assume $J^\mathrm{eff}_{\ \nu} = 0$ and move to radiative coordinates, that is to say that $\Phi = x^0$ so we have directly $\Phi_0 = 1$ and $\Phi_i = 0$. The GW obey to the eikonal, thus $\eta^{00} = 0$ and $\eta^{0i} = \Phi^i$. In such a comobile coordinate system, the significant component of the wave are the $h_{ij}$. The source term becomes \begin{subequations} \nonumber \begin{align} J^\mathrm{eff}_{\ \nu} &= \Phi_0\,F^{0\lambda \, (0)}\,h'_{\lambda \nu} + \Phi_i\,F^{i\lambda \, (0)}\,h'_{\lambda \nu}\\ &= F^{0\lambda \, (0)}\,h'_{\lambda \nu}\\ &= F^{0j \, (0)}\,h'_{ij} \end{align} \end{subequations} So the fact that $J^\mathrm{eff}_{\ \nu} = 0$ implies that $F^{0j \, (0)}\,h'_{ij} = 0$. At any point of space time, we can choose a spatial coordinate system such that $F^{0j\,(0)} = A \, \delta^j_1$. Then we have $F^{0j \, (0)}\,h'_{ij} = 0,$ that leads to $A\,{h'}_{i1} = 0 $ so $A = 0$ since the GW is non-zero. So $F^{0j \, (0)} = 0 =q\,\Phi^j $, because $\Phi^j = \eta^{0j} = 0$. Knowing that $\Phi_0 = 1$ and $\Phi_i = 0$, we can show that $F^{0j \, (0)}=\Phi_\alpha \, F^{\alpha j \, (0)}$. We can also consider that $\Phi^0 = 0 $ and $F^{\alpha 0} = - F^{0 \alpha}$ to conclude that $\Phi_\alpha F^{\alpha \mu\, (0)} = q\,\Phi^\mu$. This completes our demonstration. Let us now particularise the final result $\Phi_\alpha F^{\alpha \mu\, (0)} = q\,\Phi^\mu$ in a illustrating case. In cartesian coordinates, we can write down the null vector $\Phi_\alpha = (k,0,0,k)$ with $k$ the wave vector of the incident GW which is therefore propagating along the $z$-direction. The above-mentioned condition Eq.(\ref{thm}) now leads now to two constraints on the EM field : $E^x + B^y = 0$ and $E^y - B^x = 0$. If one considers the case when there is no electric field, then this condition implies that the components of the outer magnetic field that are transverse to the direction of GW propagation vanish: $B^x=B^y=0\cdot$ In other words, any longitudinal magnetic field $B^z$ does not produce any EM wave by inverse Gertsenshtein effect (since $S_{\mu\nu}= 0$). To produce GW by this mechanism, one needs $S_{\mu\nu}\ne 0$ or, equivalently, a non vanishing magnetic field in the direction transverse to the GW propagation ($B_\perp\ne 0$). This is the reason why experiments like ADMX \cite{ADMX} do not have the right configuration to detect GW. Indeed, they use a longitudinal outer magnetic field which can therefore only interact with GW propagating transversely to it. This interaction can only produce EM waves that are in the same direction as the constant magnetic field but this is forbidden in the TM cavity they are using (since only transverse excitation modes are allowed, not longitudinal ones). To turn a haloscope into a HFGW detector, then one simply needs to rotate the outer magnetic field by a quarter of turn. Let us now present our proposed experimental set-ups. One can either consider the resonance of the induced EM waves inside a cylindrical cavity of radius $R$ or inside a waveguide made of two (or more) concentric open cylinders with inner radius $R_1$ and outer radius $R_2$. We denote by $L$ the length of the resonators and by $B^{(0)}_{\rm ext}$ the external magnetic field, assumed of constant magnitude for simplicity. A schematic representation of our cavities can be found in the figure~\ref{schema}. \begin{figure}[ht!] \centering \includegraphics[clip,scale=0.15]{schema.png} \caption{Schematic representation of the experimental designs: a cylindrical TM cavity (top) and TEM waveguide (bottom), into an external static and transverse magnetic field. } \label{schema} \end{figure} We briefly present here the responses of these cavities to an incoming GW signal. The interested reader will find the details of the computations at the end of this paper, in the Appendix. In the following, we will assume $c=1$. The starting point is the induction of EM waves when the GW passes perpendicularly to the static magnetic field, as described by Eq. (\ref{lineq}). Considering the outer magnetic field along the x-direction : $\vec{B^{(0)}_{\rm ext}}=B^{(0)}_{\rm ext}\vec{e_x}\cdot$, we obtain the following wave equation for the induced magnetic field $\vec{B}^{(1)}$ \begin{equation} \left(-\frac{\partial^2}{\partial t^2}+\vec{\Delta}\right) \vec{B}^{(1)}=B^{(0)}_{\rm ext} \begin{pmatrix} \frac{\partial^2 h_{+}}{\partial z^2} \cos(\phi)+\frac{\partial^2 h_{\times}}{\partial z^2 } \sin(\phi)\\ -\frac{\partial^2 h_{+}}{\partial z^2} \sin(\phi)+\frac{\partial^2 h_{\times}}{\partial z^2 } \cos(\phi)\\ 0 \end{pmatrix} \label{waveB} \end{equation} where $h_{+}$ and $h_{\times}$ are the usual polarizations of the incoming GW in the traceless-transverse gauge. Although there is also an induced electric field $\vec{E}^{(1)}$, the response of the detector is dominated by the induced magnetic field, as we shall see below. We can then project this Eq.(\ref{waveB}) on the proper functions of the laplacian operator in cylindrical coordinates. This spectral decomposition is given by \begin{equation} B^{(1)}_{r,\phi}(t,r,\phi,z)\approx \sum_{k,m,n} \hat{b}^{r,\phi}_{k,m,n}(t) \cdot\psi^{r,\phi}_{kmn}(r,\phi,z) \end{equation} where $\psi^{r,\phi}_{kmn}(r,\phi,z)$ are the cylindrical harmonics that satisfy the boundary conditions of our EM cavities. The result of this spectral decomposition allows reducing the above wave equation Eq.(\ref{waveB}) to an ordinary differential equation describing a forced harmonic oscillator for each spectral mode $\hat{b}^{r,\phi}_{k,m,n}$ in our cavity: \begin{equation} \frac{d^2\hat{b}^{r,\phi}_{k,m,n}}{dt^2}+\Omega^2_{kn}\hat{b}^{r,\phi}_{k,m,n}=\hat{s}^{r,\phi}_{k,m,n}(t) \label{osc} \end{equation} where $\Omega^2_{kn}$ are the proper frequencies of the resonant cavities and $\hat{s}^{r,\phi}_{k,m,n}(t)$ are the spectral coefficients of the source of the wave equation Eq.(\ref{waveB}). The energy variation $\Delta \mathcal{E}$ inside the cavity is given by, at the leading order (see our Appendix for details), \begin{equation} \Delta \mathcal{E}\approx \frac{2\pi B_0 }{\mu_0} \cdot \sum_{k} \mathcal{I}_k\hat{b}_{k,1,0}(t) \label{deltaE} \end{equation} where the coefficients $\mathcal{I}_k$ arise from the spatial averaging in the transverse directions (these adimensional quantities depends on the cavity geometry only). Let us recall that the numbers $k$, $m$ label the transverse decomposition in the radial and azimuthal direction, while number $n$ labels the different longitudinal modes. We can see that only the $(k,1,0)$ modes, which are constant in the longitudinal direction $z$ (since $n=0$), are contributing to the energy variation at first order in $B^{(1)}$. Since these $(k,1,0)$ modes do not propagate along $z$, one can see that there is no spatial phase shift with this energy variation at first order. These $(k,1,0)$ modes of the induced magnetic field are sourced by \begin{equation} \hat{s}_{k,1,0}^{r,\phi}(z,t)=\pi B_0 L^2 \mathcal{I}_k\int_{-L/2}^{L/2} \frac{\partial^2 h_{+}(z,t)}{\partial z^2 }dz \label{shat} \end{equation} The detailed computations from Eq.(\ref{lineq}) to Eqs.(\ref{osc}-\ref{shat}) is available at the end of this paper, in the Appendix. A dimensional analysis of Eq.(\ref{deltaE}) leads to the following estimation for the order of magnitude of the induced energy variation inside the resonator \begin{equation} \Delta \mathcal{E}\approx \frac{2\pi B_0^2 L^3 }{\mu_0}\mathcal{H}_{\rm GW}\mathcal{F} \label{estim} \end{equation} where $\mathcal{H}_{\rm GW}$ is the (dimensionless) amplitude of the strain of the GW and where $\mathcal{F}$ is a adimensional geometrical factor\footnote{In other words, we extract all the dimensional factors in the equation (\ref{deltaE}), letting just one adimensional expression that depends on the geometry of the detector and its frequency sensitivity.} accounting for the shape of the resonator (length and diameter). This factor $\mathcal{F}$ is of the order of unity when the spectrum of the incoming GW matches the resonance bandwidth. In the other cases, we must consider a frequency dependant geometrical factor $\mathcal{F}(\omega)$. Please note that with our model there is no temporal phase shift in the conversion process. Indeed there is no imaginary part in the Fourier transform of $\Delta E$ and so the complex argument is null. This is mainly due to the fact that Eq.(\ref{osc}) is purely harmonic, without any dissipation, and therefore no phase shift can happen. Instead of a derivation in the time domain, one could also use a frequency approach of the cavities responses to demonstrate this, but this goes beyond the scope of this paper. However, the ohmic losses of energy in the sidewalls of the cavities could be represented by a dissipative term in equation (\ref{osc}), and so a phase shift could appear in the conversion process for resistive cavities. Although ohmic losses will lower the efficiency of the resonance mechanism investigated here, this can be avoided by working with superconducting cavities. \section{HFGWs from planetary-mass primordial black hole mergers} \label{sec:PBH} PBH binaries may have formed through two different channels. First, in the early Universe, when two PBHs form sufficiently close to each other for their dynamics to decouple from the expansion of the Universe, before the matter-radiation equality~\cite{Nakamura:1997sm,Sasaki:2016jop}. Second, by tidal capture in dense environments~\cite{Bird:2016dcv,Clesse:2016vqa}, such as ultra-faint dwarf galaxies. In this section, we review the motivations to consider planetary-mass PBHs binaries and we estimate their expected merging rate and gravitational-wave signal, for those two channels. We then calculate the astrophysical range of resonant electromagnetic detectors as a function of their strain sensitivity. Finally, we compute for each formation channel the limits that could be set on the abundance of planetary-mass PBHs. \subsection{Motivations} The progenitor masses and low effective spins of the black hole mergers detected by LIGO/Virgo have revived the interest for PBHs in the $ [1-100] M_\odot$ range~\cite{Bird:2016dcv,Clesse:2016vqa,Sasaki:2016jop,Kashlinsky:2016sdv}. However it is debated if PBHs could constitute only a small fraction, or up to the totality of the DM in the Universe. In this context, detecting a sub-solar black hole would almost clearly point to a primordial origin\footnote{See however~\cite{Kouvaris:2018wnh,PhysRevLett.126.141105} for another subsolar black hole formation channel, with different spin predictions, in a specific dark matter scenario.}. Going beyond the simplest but unrealistic assumption of a monochromatic mass function, the distribution of PBHs could span several decades of masses, as it is the case if curvature fluctuations at the origin of PBH formation are nearly scale invariant -- a generic prediction of inflation -- or come from a broad peak in their power spectrum. Then the known thermal history of the Universe, in particular the QCD transition at $\sim 100$ MeV and the electroweak epoch at $\sim 100$ GeV, should have left imprints in the PBH mass function~\cite{Byrnes:2018clq,Carr:2019kxo}, whatever is the mechanism at the origin of those curvature fluctuations. These features take the form of a high peak at the solar mass scale and two bumps at $\sim 30 M_\odot$ and $\sim 10^{-5} M_\odot $. Such an extended mass function could explain a series of puzzling observations (see~\cite{Clesse:2017bsw,Carr:2019kxo} and references therein) such as unexpected microlensing events, LIGO/Virgo black hole mergers, some properties of ultra-faint dwarf galaxies, unexpected correlations in X-ray and infrared cosmic backgrounds, and super-massive black holes at high redshifts. The bump in the planetary-mass range is consistent with recent detections of star and quasar microlensing events~\cite{Niikura:2019kqi,Hawkins:2020zie,Bhatiani2019}, which suggest a DM fraction of $f_{\rm PBH} \sim 0.01$ made of compact objects, quite much than one can expect for floating planets, but that could be expected for PBHs in the unified scenario presented in~\cite{Carr:2019kxo}. Recently, the possible detection of a stochastic GW background at nanoHerz frequencies by NANOGrav~\cite{Arzoumanian:2020vkk} may also hint at the existence of PBHs with planetary~\cite{Domenech:2020ers} or stellar~\cite{DeLuca:2020agl,Vaskonen:2020lbd,Kohri:2020qqd} masses, and Wang et al.~\cite{Wang2019,Wang2020} puts some constraints on the PBH abundance for the current detectors to probe a stochastic GW background made of PBHs. However all these observations could have another origin and the derived limits are still subject to large astrophysical uncertainties. In the future, it is therefore important to find complementary ways to probe the existence of such objects, and to distinguish their nature and origin. As we show in this paper, HFGW detectors will have the ability to detect or set new limits on the abundance of light, subsolar PBHs, of mass $m_{\rm PBH} \sim 10^{-5} M_\odot$. HFGWs are indeed produced during the merging phase of such light PBHs. The frequency associated to the innermost stable circular orbit (ISCO), when the GW emission is close to maximal, is given by \begin{equation} f_{\rm ISCO} = \frac{ 4400\, {\rm Hz} } {(m_1 + m_2) / M_\odot} ~. \label{fisco} \end{equation} with $m_1$ and $m_2$ the masses of the two binary components. A frequency of $200$ MHz thus corresponds to a PBH mass of $10^{-5} M_\odot$, the same order than the mass of the lenses at the origin of the microlensing events reported in~\cite{Niikura:2019kqi}. Nevertheless, for being an interested HFGW signal, one needs to investigate if the merging rate of such PBHs can lead to at least $\mathcal O(1) $ mergers per year within the HFGW detector range. PBHs therefore constitute a target of much interest for our experimental concept of EM detection of HFGWs. From the amplitude and spectral response of the resonant detectors, we will characterize the expected signals from PBHs mergers for a large range of progenitor masses in the interval $[10^{-8};10^{-3}] \, M_\odot$, located at $1$ Gpc distance. Then, for a given detector sensitivity, we will compute the expected limits on the PBH abundance. \subsection{Gravitational-waves from inspiraling binaries} A good estimation of the GW strain produced at a given frequency $f_{\rm GW}$ during the inspiralling phase of a black hole binary is provided by the Post-Newtonian approximation~\cite{Antelis:2018sfj}, \begin{equation} \label{eq:strainPBHs} h \approx \frac{2}{D} \left( \frac{G \mathcal M }{c^2} \right)^{5/3} \left( \frac{ \pi f_{\rm GW}}{c} \right)^{2/3} ~, \end{equation} where $\mathcal M \equiv (m_1 m_2)^{3/5}/(m_1+m_2)^{1/5} $ is the binary chirp mass and $D$ is the distance to the observer. The GW emission is close to maximal at the ISCO frequency and, for a given chirp mass, for equal mass binaries, $m_1 = m_2 = m_{\rm PBH}$. For an experiment with a detector strain sensitivity $h_{\rm det}$ at this maximal frequency, the corresponding astrophysical reach $D_{\rm max}$ is given by \begin{equation} \label{eq:Dmax} D_{\rm max} \approx 1.6 \times \frac{(m_{\rm PBH}/M_\odot) }{h_{\rm det} \times 10^{20}} {\rm Mpc}. \end{equation} For instance, for a strain sensitivity of $h_{\rm det} \sim 10^{-25}$ and $m_{\rm PBH} \sim 10^{-5} M_\odot$, eventual mergers towards the galactic center, in the Milky Way DM halo, or in satellite ultra-faint dwarf galaxies, could be detected. For a better sensitivity down to $h_{\rm det} \sim 10^{-30}$, corresponding to the optimal sensitivity of the proposed designs of resonant EM detectors, then one would probe planetary-mass PBH mergers in more distant galaxies. \subsection{Merging rate of primordial binaries} If PBHs are spatially randomly distributed at formation, it happens that two PBHs form so close to each other that their gravitational attraction overpasses the effect of the Hubble-Lema\^itre expansion at some point before matter radiation equality. In such a case, they directly form a binary whose orbital parameters and lifetime do not only depend on the two black hole masses but also on the mass and distance of the nearest PBHs. Eventually, it takes of the order of the age of the Universe for the PBH binary to merge. The merging rates $\tau $ today associated with this binary formation channel and an arbitrary mass function have been evaluated in ~\cite{Kocsis_2018,Raidal:2018bbj,Gow:2019pok,Liu2019} as \begin{eqnarray} R^{\rm prim} (m_1, m_2) & \equiv & \frac{{\rm d} \tau}{{\rm d} \ln m_1 {\rm d} \ln m_2} \nonumber \\ & \approx & \frac{1.6 \times 10^6}{\rm Gpc^3 yr} f_{\rm PBH}^2 f(m_1) f(m_2) f_{\rm sup} \nonumber \\ & \times & \left(\frac{m_1 + m_2}{M_\odot}\right)^{-\frac{32}{37}} \left[\frac{m_1 m_2}{(m_1+m_2)^2}\right]^{-{\frac{34}{37}}} \end{eqnarray} where $f(m)$ is the today density distribution of PBHs normalized to one ($\int f(m) {\rm d} \ln m = 1$) and $f_{\rm PBH}$ is the integrated DM fraction made of PBHs. We also define an effective parameter \begin{equation} \tilde f_{\rm PBH} (m_{\rm PBH}) \equiv f_{\rm PBH} f(m_{\rm PBH}) f_{\rm sup}^{1/2} \end{equation} that includes a rate suppression factor ($f_{\rm sup}$) to take into account the possible rate suppression due to binary disruption by early-forming clusters, an effect put in evidence by N-body simulations when $f_{\rm PBH} \gtrsim 0.1$~\cite{Vaskonen:2019jpv}. In such a case, one can recover the LIGO/Virgo merging rates inferred from the recent detections of GW190425, GW190521 and GW190814 involving at least one BH in the mass gaps, with $f_{\rm PBH} = 1$ and $f_{\rm sup} \simeq 0.0025$~\cite{Clesse:2020ghq}. In the opposite case, $f_{\rm sup}= 1$ and $\tilde f_{\rm PBH}$ simply represents the DM density fraction made of PBHs at a given mass and within a unit logarithmic mass interval. If one considers the merging rates of equal-mass binaries that produce the largest strain signal, one gets \begin{eqnarray} R^{\rm prim}(m_{\rm PBH}) &\approx& \frac{3.1 \times 10^6}{\rm Gpc^{3} yr} \tilde f_{\rm PBH}^2 \left( \frac{m_{\rm PBH}}{M_\odot}\right)^{-0.86}. \end{eqnarray} In turn, one can determine the radius of the sphere in which one expects one event per year, \begin{equation} D^{\rm prim}_1 = \left( \frac{4\pi}{3 } R^{\rm prim}\right)^{-1/3} \approx 4.2 \, {\rm Mpc} \times \tilde f_{\rm PBH}^{-2/3} \left( \frac{m_{\rm PBH}}{M_\odot}\right)^{0.29}~. \end{equation} For simplicity we neglected the effects of redshift that are anyway insignificant for most of the considered cases. For instance, in the scenario of~\cite{Carr:2019kxo,Clesse:2020ghq} with $f_{\rm PBH} =1$, $f_{\rm sup}=0.0025$ and $f(10^{-5} M_\odot) \simeq 10^{-2}$, one gets $D_1^{\rm prim} (10^{-5} M_\odot) \approx 23 \, {\rm Mpc}$. Using Eqs.~(\ref{eq:strainPBHs}) and~(\ref{eq:Dmax}), one then obtains the required GW strain sensitivity to detect one of these merger events per year, \begin{eqnarray} h_{1}^{\rm prim} & \approx & 3.8 \times 10^{-21} \tilde f_{\rm PBH} ^{2/3} \left(\frac{m_{\rm PBH}}{M_\odot}\right)^{0.7} \\ & \approx & 8.3 \times 10^{-19} \tilde f_{\rm PBH} ^{2/3} \left( \frac{\rm Hz}{f} \right)^{0.7} ~, \end{eqnarray} which can be typically targeted by GW experiments operating at frequencies from kHz up to GHz. This relation can be inverted to obtain a limit on the DM fraction at a given mass (if $f_{\rm PBH} <0.1$) in case of null detection, as a function of the strain sensitivity, \begin{eqnarray} \tilde f_{\rm PBH} \lesssim 9.1 \left[ \frac{h_{\rm det}}{10^{-20}}\right]^{3/2} \left( \frac{m_{\rm PBH}}{M_\odot} \right)^{-1.07} . \end{eqnarray} However, the strain sensitivity of the detector depends on the waveform and signal duration, which depend on the PBH mass. It is thus more adequate to compute a limit on the PBH abundance taking these effects into account and instead assuming an EM power sensitivity, which we do in the next section. For instance, we obtain with the two proposed experimental designs and a power sensitivty of $10^{-10}$W, limits that are competitive with the current microlensing limits at the same mass scale. These are represented in our final Figure~\ref{fig_fDM}. Finally, we point out that when our analysis was being finalized, the authors of~\cite{Boehm:2020jwd} have claimed that the rates from primordial binaires are highly suppressed compared to previous calculation. The reason is a subtle general relativistic effect arising when one considers geodesics in black hole exterior spacetime metrics that are FLRW asymptotic. If this claim is correct, primordial binaires are by far outside the reach of EM detectors, but one can nevertheless consider the merging rates inside PBH clusters, which we detail thereafter. \subsection{Merging rate from PBH clusters} The second binary formation channel is through dynamical capture in dense PBH halos. As any other DM candidate, PBHs are expected to form halos during the cosmic history, and their clustering properties determine the overall merging rate. For instance, for a monochromatic mass spectrum and a standard Press-Schechter halo mass function, one gets a rate~\cite{Bird:2016dcv} \begin{equation} \label{eq:capturerate} R^{\rm capt} \sim f_{\rm PBH}^2 \times \mathcal O(1-100) \, {\rm yr^{-1} \, Mpc}^{-3} \end{equation} that is independent of the PBH mass. For more realistic extended mass functions, the abundance, size and evolution of DM halos, partially or entirely made of PBHs, is impacted by several effects, see e.g.~\cite{Chisholm:2005vm,Chisholm:2011kn,Belotsky:2018wph,Carr:2018rid,Suyama:2019cst,MoradinezhadDizgah:2019wjf,Matsubara:2019qzv,Young:2019gfc,Padilla:2020xlo,DeLuca:2020jug,Jedamzik:2020ypm}. Let us mention a Poissonian noise from the discrete nature of PBHs, a seeding effect from heavy PBHs, the enhancement of the primordial power spectrum at the origin of PBH formation, the dynamical heating and evaporation of clusters, etc. These effects can either boost or suppress the merging rates from clusters and make the whole clustering dynamics a rather complex and model-dependent process, subject to large uncertainties. Invoking clustering is also crucial to evade microlensing limits on stellar-masses~\cite{Garcia-Bellido:2017xvr,Calcino:2018mwh} in scenarios with $f_{\rm PBH}=1$. As an alternative of using uncertain theoretical predictions, on can instead infer an upper limit on the PBH merging rate from LIGO/Virgo observations, see e.g~\cite{Clesse:2020ghq}. The merging rate from tidal capture in PBH clusters for an arbitary mass function is given by~\cite{Clesse:2016vqa,Clesse:2020ghq} \begin{eqnarray} R^{\rm capt} (m_1,m_2) &\equiv& \frac{{\rm d}\tau}{{\rm d} \ln m_1 {\rm d} \ln m_2 } \nonumber \\ &\approx& R_{\rm clust} f_{\rm PBH}^2 \times f(m_1) f(m_2) \nonumber \\ & \times & \frac{(m_1 + m_2)^{10/7}}{(m_1 m_2)^{5/7}} \rm{yr^{-1}Gpc^{-3}}, \label{eq:ratescatpure2} \end{eqnarray} where $ R_{\rm clust} $ is an effective parameter encompassing the clustering properties. For $f_{\rm PBH} =1$ and $R_{\rm clust} \approx 450$, these rates are consistent with the latest LIGO/Virgo observations, for a broad PBH mass function impacted by the transient reduction of the critical threshold of PBH formation at the QCD epoch. As already mentioned, this effect is unavoidable and may have induced a peak around $2.5 M_\odot$ and a bump around $30 \, M_\odot$ in the PBH mass function. In this scenario, around one percent of the DM could be made of planetary-mass PBHs around $10^{-5} M_\odot$. Like for primordial binaries, we define an effective parameter \begin{equation} \tilde f_{\rm PBH} \equiv \left( \frac{R_{\rm clust}}{450}\right) \times f_{\rm PBH} f(m_{\rm PBH}) \end{equation} representing the DM density fraction at a given mass and per logarithmic mass interval, in the above mentioned scenario. One then obtains the merging rate for equal-mass binaries, \begin{equation} R^{\rm capt} (m_{\rm PBH}) = 1.2 \times 10^3 \tilde f_{\rm PBH}^2~, \end{equation} the corresponding source distance $D_1^{\rm capt} $, \begin{equation} D_1^{\rm capt} \approx 58 \, {\rm Mpc} \times \tilde f_{\rm PBH}^{-2/3}, \end{equation} and the required experimental strain sensitivity to detect one event per year, \begin{equation} h_{\rm 1}^{\rm capt} \approx 2.7 \times 10^{-22} \tilde f_{\rm PBH}^{2/3} \left( \frac{m_{\rm PBH}}{M_\odot}\right) . \end{equation} This GW signal is therefore typically lower than for PBH binaires on planetary-mass scales. Finally, like for the case of primordial binaries, we have derived the expected limits on $\tilde f_{\rm PBH}$ for a given experimental strain sensitivity, \begin{eqnarray} \tilde f_{\rm PBH} \lesssim 2.5 \times 10^{4} \left[ \frac{h_{\rm det}}{10^{-20}}\right]^{3/2} \left( \frac{m_{\rm PBH}}{M_\odot} \right)^{-3/2}. \end{eqnarray} As an example, if $h_{\rm det} = 10^{-30}$ and $m_{\rm PBH} = 10^{-5}$, one gets $\tilde f_{\rm PBH} \lesssim 7 \times 10^{-4}$, which is better than the current microlensing limits. Again, it is more accurate to assume a power sensitivity rather than a strain sensitivity. Doing so, the corresponding limits on $\tilde f_{\rm PBH}$ have been represented in our final Figure~\ref{fig_fDM}. \section{Probing PBH mergers with resonant EM detectors of HFGWs}\label{sec:det-PBH} Let us consider GW trains produced during the final inspiraling phase of PBHs binaries of different masses and passing through the resonator and let us analyze the induced EM radiation. For simplicity, we consider a GW propagation colinear with the longitudinal axis of the EM resonator and perpendicular to the outer magnetic field. In this case, the GW can be approximated by a plane wave, i.e. $h_{+,\times}=h_{+,\times}(z,t)$ since the radius of the detector is much smaller than the incoming wavefront for a distant source. Some inclination of the direction of the incoming GW would result in a signal of lower amplitude, since only the component of the GW in the direction of the outer magnetic field contributes to the inverse Gertsenshtein effect. Yet, this ideal case allows to illustrate the physical process and to estimate the expected response of the detector and the output signal. We have computed the typical EM signals for resonators of various shapes. Post-Newtonian time-domain GW waveforms are generated by using the \texttt{LALSuite} library \cite{lalsuite}, assuming a 4PN approximation. For all the simulations, we choose a signal sampling frequency that is four time the ISCO frequency [Eq.~(\ref{fisco})]. The initial frequency of the signal is set to $f_{\rm ISCO}/25$. On Fig.~\ref{fig1}, we provide an example of the GW waveform produced by the merging of a PBH binary with component masses of $10^{-5} M_\odot$ and located at a distance of $1\,\rm Gpc$. The amplitude of the GW strain at reception, denoted $\mathcal{H}_{\rm GW}$, is of order $10^{-28}$ and the signal duration is of order $10^{-5}s\cdot$. \begin{figure}[ht!] \centering \includegraphics[clip,scale=0.26]{signal.png} \caption{Left: GW waveform $h_+$ for the final inspiraling phase of a PBH binary with component masses of $10^{-5} M_\odot$, at a distance of $1\,\rm Gpc$, in the post-newtonian (4PN) approximation. Right: corresponding GW signal power spectrum. } \label{fig1} \end{figure} \begin{figure}[ht!] \begin{tabular}{c} \includegraphics[clip,scale=0.26]{plotP.png}\\ \includegraphics[clip,scale=0.26]{powerP1e-5.png} \end{tabular} \caption{Top: time evolution of the induced EM power in the two resonators. Bottom: GW signal power spectrum (blue) and induced power spectrum (green). The detectors are 1-meter long, in a 5\,T constant magnetic field. The outer radius is 5\,m and for TEM coaxial case, with an inner radius of 10\,cm. The peaks in the power spectrum correspond to the resonance frequencies, represented by the red dots. } \label{figres} \end{figure} Then, by using Eq.~(\ref{deltaE}) and solving the forced harmonic oscillator equations, Eqs.~(\ref{osc}), we compute the induced EM power for the two designs of resonators. Our results are displayed on Fig. \ref{figres}, for 1-meter long resonators in a 5\,T constant magnetic field. The radius of the resonators is set to 5\,m and for the TEM coaxial design, we consider a 10 cm inner radius. As expected, the induced power increases near the merger. In each case, we have also computed the frequency spectrum of the induced power, which exhibits peaks corresponding to the resonance frequencies, coming from Eqs.~(\ref{freq1}) and (\ref{freq2}) in the Appendix. The power rms values are $1.00\times 10^{-10}$ W for the TEM resonator and $1.03 \times 10^{-10}$~W for the TM resonator. \begin{figure}[ht!] \includegraphics[clip,scale=0.45]{proper.png} \caption{The first five proper frequencies of the cavity (TM resonator, straight lines) and the waveguide (TEM resonator, dashed line) as a function of their outer radius (assuming a length of $L=1$\,m and an inner radius of $R_1=0.1$\,m for the waveguide). } \label{figgeo} \end{figure} Fig.~\ref{figgeo} shows the first five proper frequencies of the TEM and TM cavities as a function of their outer radius, within the range $10$\,cm to $10$\,m, for a detector of one meter length. For the TM resonator, these frequencies range from 10\,MHz to GHz, while the TEM waveguide can reach higher frequencies for the same detector size. In the case of an outer radius almost equal to the inner one (thin case), the resonance frequencies can be one order of magnitude larger than for a TM resonator. We point out that it would even be possible to use several concentric TEM waveguides to further extend the range of resonance frequencies. Fig~\ref{figgeo} therefore illustrates how one could play with the geometry of the experimental set-up in order to optimize the detector response to some motivated GW signal. The PBH mass is another parameter to consider. For the same detectors as in Fig.~\ref{figres}, we have simulated the detection of the signal for two additional PBH masses, of $10^{-3}$ and $10^{-7} M_\odot$, the corresponding power spectra are shown in Fig.~\ref{figres2}. For $10^{-3}\, M_\odot$, the resonance frequencies are higher than the ISCO frequency. The resulting EM power spectrum in the detector exhibits a continuous set of frequencies that matches the incoming GW power spectrum followed by the excitation of the detector resonance frequencies that constitutes the highest values in the induced radiation power spectrum. The continuous set of induced frequencies comes from the particular solution of the forced oscillator Eq. (\ref{osc}), while the excitation of the detector resonance frequencies comes from the general solution. In the $10^{-7}\, M_\odot$ case, the detector resonance frequencies are lower than the main GW signal frequencies. Therefore we have a clear separation between the EM power spectrum that is induced by the resonance frequencies of the detector and the one induced by a range of frequencies inherited from the incoming GW signal. In Fig.~\ref{figres2}, the part of the power spectrum induced by the resonance frequencies is overestimated. Indeed the numerical FFT of the incoming GW leads to a plateau at low frequencies as shown in Fig.~\ref{fig1}. As a result this plateau will be excited by the proper frequencies of the cavity leading to an overestimation of this part of the induced power spectrum. We show below how to avoid this numerical limitation. \begin{figure}[ht!] \centering \begin{tabular}{c} \includegraphics[clip,scale=0.26]{powerP1e-3.png}\\ \includegraphics[clip,scale=0.26]{powerP1e-7.png}\\ \end{tabular} \caption{Same as the bottom panel of Fig.~\ref{figres}, but for a PBH mass of $10^{-3}\, M_\odot$ (top pannel) and $10^{-7}\, M_\odot$ (bottom panel). In the first/second case, the GW frequency is below/above the resonance frequencies (represented by the red points) of the detector. } \label{figres2} \end{figure} \begin{table}[hb] \centering \begin{adjustbox}{width=\columnwidth,center} \begin{tabular}{|c|c|c|c|c|c|} \hline $m_{\text{PBH}}$ ($M_\odot$) & time (s) & $\mathcal{H}_{\rm GW}$ & $f_{\text{ISCO}}$ (Hz) & $P_{\text{RMS}}$ TEM (W) & $P_{\text{RMS}}$ TM (W) \\ \hline $10^{-3}$ & 5.43e-03 & 1.15e-26 & 2.20e+06 & 2.37e-14 & 3.19e-14 \\ \hline $10^{-4}$ & 5.43e-04 & 1.37e-27 & 2.20e+07 & 2.98e-12 & 4.96e-12 \\ \hline $10^{-5}$ & 5.43e-05 & 1.40e-28 & 2.20e+08 & 1.00e-10 & 1.03e-10 \\ \hline $10^{-6}$ & 5.43e-06 & 1.17e-29 & 2.20e+09 & 1.51e-11 & 6.31e-12 \\ \hline \end{tabular} \end{adjustbox} \caption{GW signal duration, maximal strain, ISCO frequency and the corresponding induced power in the resonant cavity or waveguide, for different values of the PBH mass. The detectors are 1~m long for an outer radius of 5~m. The inner radius in the TEM case is 10\,cm. The transverse static magnetic field is set to $5$~T. The minimal frequency of the GW waveform signal is set to $f_{\rm ISCO} /25$. The bandwidth of the resonant frequencies considered is $\left[2\cdot10^{7},2\cdot10^{8}\right]$ Hz } \label{table} \end{table} \begin{figure*} \includegraphics[clip,scale=0.42]{PEmass.png} \caption{RMS values of the induced power (left) and energy variation (right) in the TEM (blue) and TM (red) detectors as in Table~\ref{table}, as a function of the PBH mass, for mergers located at a distance of one Gpc. The excitation of resonant frequencies boosts the signal in the range between $10^{-6}$ and $10^{-4} M_\odot$. Below $10^{-6}M_{\odot}$ we plot in dashed lines the expected behaviour due to numerical limitation in our simulations. Asymptotic behaviour coming from our analysis has been plotted in black dotted lines. } \label{figmass} \end{figure*} Finally, we have computed the power in a 1-meter long resonator of 5-meter radius in a 5T magnetic field, induced by the merging of PBHs with a broad range of masses. Our results are summarized in Table~\ref{table}. We find that the induced power is maximal for a PBH mass around $10^{-5}\, M_{\odot}$, i.e. when the corresponding GW frequency lies within the range of the resonance frequencies. We have shown in Fig. \ref{figmass} the rms power as a function of the PBH mass, for the same experimental configuration. In order to do so, we have considered 41 simulated signals of PBH mergers for 61 different PBH masses (i.e., 41 signals per discretization point in mass, and each of these signals corresponding to different initial frequency ranging from $f_{\rm ISCO}/30$ to $f_{\rm ISCO}/10$). Because of the numerical problem at masses beyond $10^{-6}M_{\odot}$(poor resolution of the FFT of the GW signal at low frequencies), we extrapolate our data with the expected behaviour that we will develop further in this section. When the GW spectrum covers the fundamental resonance modes of the detector, its response is maximal and this occurs for PBH masses covering two orders of magnitude around $10^{-5}M_{\odot}$. In addition, we have also shown on Fig.~\ref{figmass} the rms energy variation inside the cavity, as a function of the PBH mass. The induced energy typically corresponds to millions of photons. It is well above the sensitivity of the ADMX experiment, which has achieved an effective instrumental noise temperature of order $\mathcal O(1)$K at similar frequencies~\cite{ADMX,ADMX2} and with a similar value of the magnetic field. Further work is however needed in order to quantify more accurately the energy or power sensitivity that could be achieved with a similar technology. For masses larger than $10^{-4}\, M_\odot$, the frequencies associated to the GWs are smaller than the proper frequencies of the cavity and so the amplitude of the total solution of (\ref{osc}) is proportional to the amplitude of the source term, divided by the square of the proper frequencies of the cavity. So the behavior of the particular solution is the same as the source term. From Eq.(\ref{waveB}), we know that the source term is proportional to the second time derivative of the strain. On one hand, the GW strain is proportional to PBH mass (see Eq.(\ref{eq:strainPBHs})). On the other hand, the maximal GW frequency is inversely proportional to the PBH mass. As a result, the energy released in the detector goes like $1/m_{\rm PBH}$, which explains the linear decreases observed at large mass. As mentionned earlier, because of a numerical limitation we do not plot the induced energy and RMS power for masses smaller than $10^{-6}\, M_\odot$. Nonetheless we can extrapolate the expected behaviour of the mass dependence. The analysis is the same as in the large mass regime except that this time it is the proper frequencies of the cavity that are smaller than the frequencies associated to the GWs. Thus the amplitude of the total solution of (\ref{osc}) is now proportional to the amplitude of the source term, divided by the square of the GW frequencies. As these frequencies are inversely proportional to the PBH mass, the behaviour of the energy released in the detector is $m_{\rm PBH}^2$ times what it was in the previous case. So the energy variation goes like $m_{\rm PBH}$. We will show this in a forthcoming paper by using a frequency domain approach. \iffalse At masses lower than $10^{-6}\, M_\odot$, the induced power reaches a plateau of about $10^{-12}\,$W. Indeed, in such a case the GW frequency is by far larger than the proper frequencies of the cavity, so the amplitude of the particular solution of Eq.~(\ref{osc}) vanishes and only the harmonic solution remains. For masses larger than $10^{-4}\, M_\odot$, the induced power decreases linearly as the mass increases. In addition, we have shown on Fig.~\ref{figmass} the rms energy variation inside the cavity, as a function of the PBH mass. For comparison, we have also shown the estimated energy directly computed from Eq.(\ref{estim}). The induced energy typically corresponds to millions of photons. It is well above the sensitivity of the ADMX experiment, which has achieved an effective instrumental noise temperature of order $\mathcal O(1)$K at similar frequencies~\cite{ADMX,ADMX2} and with a similar value of the magnetic field. Further work is however needed in order to quantify more accurately the energy or power sensitivity that could be achieved with a similar technology. Therefore, in the low mass regime, the detected variation of energy only depends on the characteristics of the detector itself and not on the PBH mass. In the large mass regime, the frequencies associated to the GWs are smaller than the proper frequencies of the cavity and so the amplitude of the total solution is proportional to the amplitude of the source term, divided by the square of the proper frequencies of the cavity. The behavior of the particular solution is the same as the source term and from Eqs.(\ref{E1par}) and(\ref{B1perp}), we know that the source term is proportional to the second time derivative of the strain. On one hand, the GW strain is proportional to PBH mass (see Eq.(\ref{eq:strainPBHs})). On the other hand, the maximal GW frequency is inversely proportional to the PBH mass. As a result, the energy released in the detector goes like $1/m_{\rm PBH}$, which explains the linear decreases at large mass. \fi Finally, we observe that there is no significant difference between the TEM and the TM detectors, highlighting some robustness with respect to the choice of the experimental set-up. As a final step, we have combined our calculations of the PBH merging rates and of the induced power as a function of the PBH mass for mergers at a fixed distance of 1 Gpc, in order to forecast limits on the PBH abundance for a fixed detector sensitivity of $10^{-10}$W and a survey of one year. For doing so, we have used the calculated values of $D_1$ for the two binary formation channels. Given that the GW strain is inversely proportional to the distance of the source, the EM power released in the detector is also inversely proportional to the distance, see Eq.~(\ref{estim}). Our final results are displayed in Fig.~\ref{fig_fDM}. We can clearly see that the resonance plays a role by boosting the detection limits to be as low as $\tilde f_{\rm PBH} \lesssim 10^{-8}$ for primordial binaries, and $\tilde f_{\rm PBH} \lesssim 10^{-4}$ for tidal capture in clusters. . Resonant EM-based HFGW detectors could therefore set unprecedented and independent limits on the abundance of planetary-mass PBHs. % \begin{figure}[htb!] \includegraphics[clip,scale=0.6]{PlotfDMv2.png} \caption{Expected limits on the effective parameter $\tilde f_{\rm PBH}$ corresponding to the dark matter density fraction in PBHs at a given mass and per logarithmic mass interval in the two models described in the text: primordial binaries (blue limit) and tidal capture in halos (red limit). These limits are computed for the proposed experimental designs of EM resonant HFGW detectors, assuming a power sensitivity of $10^{-10}$W, achievable with current technology. The orange curve shows the possible abundance of planetary-mass PBHs inferred form recent microlensing observations towards the galactic bulge~\cite{Niikura:2019kqi}.} \label{fig_fDM} \end{figure} \section{Conclusion} \label{sec:conc} The detection of dozens of black hole or neutron star mergers by LIGO/Virgo and, more recently, the plausible observation of a stochastic GW background with pulsar timing arrays (PTAs) by NANOGrav, have revealed the bright future of GW astronomy. LIGO/Virgo probe GW frequencies between $10$ and $10^4$ Hz and PTAs of $10^{-9}$ Hz. In the future, ground-based and space-detectors like Einstein Telescope, Cosmic Explorer and LISA will fill the gap between these two ranges and probe other GW sources like intermediate mass black hole binaries. On the other side of the GW spectrum, HFGW detectors might be equally interesting and probe cosmological stochastic backgrounds or exotic planetary-mass compact objects like PBHs. Motivated by recent claims that PBHs with a wide mass distribution could explain, among others, the masses and spins of LIGO/Virgo black holes, the NANOGrav observation and the dark matter in the Universe, we envision their detectability with two designs of resonant EM detectors based on the inverse Gertsenshtein effect and operating at frequencies of order $100$ MHz. The released energy in a EM resonator has been calculated for fourth-order post-Newtonian GW waveforms corresponding to the final inspiraling phase of light PBH binaries. The experimental apparatus either consists in a one-meter long cylindrical resonant cavity (TM) or in a conducting waveguide (TEM), together with a static 5 Tesla magnetic field orthogonal to their axis of symmetry. For a $5$ meters radius resonator or a waveguide of $0.1$ meter inner radius, we find a typical power variation of order of $10^{-10}$ W, detectable with present technology. Similar resonant cavities have been already built for axion searches, but their collinear magnetic field prevents the detection of HFGWs. We have also studied the expected signal and its power spectrum for different binary masses. Their merging rates have been estimated for two PBH binary formation channels, at formation in the early Universe or due to tidal capture in dense clusters. The optimal sensitivity is obtained for PBH masses of order $10^{-5} M_\odot$. For the two considered models, we forecast the stringent limits on the PBH abundance that could be set with such experiments, of order $\tilde f_{\rm PBH} \lesssim 10^{-8}$ and $\tilde f_{\rm PBH} \lesssim 10^{-4}$, respectively. Finally, one should keep in mind that we can tune the parameters of the detector to match the frequency range of highest interest. Therefore HFGW therefore have the ability to set new constraints on the fraction of dark matter made of light PBHs that would be complementary to microlensing limits. In particular, they could be used to distinguish between PBH and planetary origins of recently detected microlensing events towards the galactic center, suggesting that about one percent of the DM could be made of such objects. HFGW detectors with similar designs could also be used to detect cosmological stochastic GW backgrounds. Frequencies around $100$ MHz are of particular interest in this context, because these are characteristic to GW sources at energies close to the GUT scale, like (p)reheating, oscillons, phase transitions or evaporating PBHs. In particular, we point out that a strain sensitivity down to $h \sim 10^{-30}$ for achievable TM and TEM detectors is of order of the typical amplitude of such cosmological backgrounds. Ongoing research focuses on using a frequency analysis of the resonance mechanim to get direct information about the frequency sensitivity of our detector. Another ongoing work is computing the power variation in the detectors generated by a stochastic, almost isotropic GW background. Other future research topic is to characterize the expected experimental noise with such detectors. The results for axion haloscope detection \cite{ADMX,ADMX2} could be an interesting starting point. In summary, resonant electromagnetic HFGW detectors are ideal to probe various aspects of fundamental physics, from early Universe cosmology to exotic compact objects like planetary-mass primordial black holes. Our work contributes to pave the way in this direction and provides strong motivations to start the experimental development of such detectors, with currently available technology.\\\\ \noindent \textit{Acknowledgments}\\ This research used resources of the "Plateforme Technologique de Calcul Intensif (PTCI)" (\url{http://www.ptci.unamur.be}) located at the University of Namur, Belgium, which is supported by the F.R.S.-FNRS under the convention No. 2.5020.11. The PTCI is member of the "Consortium des Équipements de Calcul Intensif (CÉCI)" (\url{http://www.ceci-hpc.be}).

30307373394117de5370a20ef10ea67e556bdb44

\subsubsection{Structural break identification} \label{sec:analysis_structural} We use a Wald test to look for a single structural break in social distancing decisions over time. This involves performing a simple linear regression of social distancing decisions (left-hand side) on the round number (right-hand side), and then testing whether the coefficient on the time variable exhibits a structural break. The test assumes there is at most one structural break but is agnostic as to at what point it may occur. We perform this test on the data at three levels of aggregation. First, we consider all observations together. Second, we split by the rate of contagion and the policy intervention. This gives four equal-sized group: (1) $15\%$ contagion, fine; (2) $15\%$ contagion, nudge; (3) $65\%$ contagion, fine; and (4) $65\%$ contagion, nudge. Finally, we subdivide each of the four groups above by network type (i.e. complete vs star). If the policy intervention is effective then we expect the test would find a structural break at round 21. With all observations together, the Wald test finds the structural break at round 21 ($p < 0.001$). When considering the dis-aggregated data, the Wald test finds a structural break at round 21 for each of the subsets, except for complete networks with the nudge treatment. \begin{table}[ht] \centering \caption{Results from Structural Break analysis using the Wald Test} \begin{tabular}{c|c} \textbf{Low contagion: $\alpha = 0.15$} & \textbf{High contagion $\alpha = 0.65$} \\ \hline \begin{tabular}{llll} \textbf{Treatment} & \textbf{Network} & \textbf{Break} & \textbf{p-value} \\ fine & all & $21$ & $< 0.001$ \\ fine & complete & $21$ & $< 0.001$ \\ fine & star & $21$ & $< 0.001$ \\ nudge & all & $21$ & $< 0.001$ \\ nudge & complete & $7$ & $< 0.001$ \\ nudge & star & $21$ & $< 0.001$ \\ \end{tabular} & \begin{tabular}{llll} \textbf{Treatment} & \textbf{Network} & \textbf{Break} & \textbf{p-value} \\ fine & all & $21$ & $< 0.001$ \\ fine & complete & $21$ & $< 0.001$ \\ fine & star & $21$ & $< 0.001$ \\ nudge & all & $21$ & $< 0.001$ \\ nudge & complete & $16$ & $< 0.001$ \\ nudge & star & $21$ & $0.009$ \\ \end{tabular} \end{tabular} \end{table} \subsubsection{Aggregate level} \label{sec:analysis_aggregate} In this section, we focus on the aggregate analysis of the decision data and conduct a set of nonparametric tests. Since individual observations are correlated at the group level, our unit of observation is a group of 5 participants. Recall that we have 80 groups in total, equally distributed across 8 treatments. For each group, we calculate the average proportion of participants that practice social distancing in the last 10 rounds of both parts of the experiment. Below, we refer to this average as `distancing levels'. We discard the first 10 rounds of both parts of the experiment because our participants display clear convergence behavior. In particular, by round 11, in both Baseline and Intervention, at least 80\% of participants converge to a particular strategy in all treatments. Further details on convergence behavior is in Section \ref{sec:robustness_convergence}. Section \ref{sec:robustness_aggregate} shows that the key results below are robust to including all 20 rounds of both parts of the experiment. \textbf{Results.} \Cref{tab:nonparametric_analysis} presents the results of the non-parametric tests -- Mann-Whitney U-test (MW)~\citeplatex{MW_1947_a} for unmatched samples and Wilcoxon Signed-Rank test (WSR)~\citeplatex{W_1992a} for matched samples. Samples are matched when they are generated by the same set of groups (e.g. comparing distancing levels in Baseline with those in Intervention after the introduction of a fine). On the other hand, when we make comparisons across groups from different treatments (e.g. those that were exposed to the low rate of contagion versus the high rate) then samples are independent or unmatched. Section \ref{sec:robustness_aggregate} shows that all of the findings of this section are robust to using parametric analogs of these non-parametric tests. In our aggregate analysis, we test the set of hypotheses in Section \ref{sec:model}. Note that we cannot test \Cref{theor:risk_aversion,,theor:social_values} here, which are instead covered in Section \ref{sec:analysis_individual}. Part 1 of \Cref{tab:nonparametric_analysis} shows that the data from the Baseline from fine and nudge treatments can be pulled together as these samples are statistically identical at all conventional significance levels. This result is robust to various specifications. \input{tables/nonpar_analysis} In Part 2 of \Cref{tab:nonparametric_analysis}, we examine the effect of introducing a fine in Intervention. From the table, it is evident that a fine has a positive and statistically significant effect on distancing levels compared to the Baseline with no fine in all specifications considered. Most results in this part are statistically significant at least at the 5\% level. The effect of the fine is also sizable -- it results in an increase in mean distancing levels of 5\% to 9\% depending on the specification. The following result summarizes the above observations. Notice that the numbering of all results in this section corresponds to the numbering of Hypotheses in Section \ref{sec:model}. \begin{findingA} The introduction of the fine increases the level of social distancing. The effect is observed for all three positions as well as for the two rates of contagion. The effect is both statistically significant (WSR, $p<0.05$ in all but one specification where $p=0.09$) and sizable in magnitude. \end{findingA} In Part 3 of \Cref{tab:nonparametric_analysis} we repeat the analysis from Part 2, but now looking at the nudge Intervention rather than the fine. The tests find a statistically significant effect of the nudge in only 4 of the 6 specifications. Generally, we can see that the nudge has a smaller effect which is not as robust as that of the fine. In particular, the effect of the nudge is very significant and large in magnitude (7\%) under high contagion, but disappears completely both statistically and in terms of economic magnitude under low contagion. Also, the nudge seems to have no effect on the superspreader. Note also that the effects of the nudge are not robust to using data from all rounds of Baseline and Intervention (see Section \ref{sec:robustness_aggregate} for more details). \begin{findingA} The introduction of the nudge generally increases the levels of social distancing. The effect, however, is not robust to alternative specifications and is smaller in magnitude than that of the fine. In particular, the effect is statistically significant (WSR, $p=0.0003$) and sizeable in magnitude under high contagion, but not low contagion (WSR, $p=0.5$). Further, a statistically significant effect is observed for the close-knit ($p=0.05$) and peripheral (WSR, $p=0.06$) participants but not for superspreaders. \end{findingA} Part 4 of \Cref{tab:nonparametric_analysis} shows that the data from the Intervention in the second part of the experiment is not statistically different for the fine and nudge treatments. In particular, even though the mean level of distancing in Intervention in fine treatments is higher than that in nudge treatments, our non-parametric testing fails to find any difference between them in 4 out of 6 specifications. We do, however, find evidence that the fine is more effective than the nudge (1) when data is pulled across treatments, and (2) for the superspreader. The effect is substantial in magnitude (7-10\%) and significant at 10\% level. \begin{findingA} There is limited evidence that the fine is more effective than the nudge, and the effect is not robust. In particular, the fine increases distancing levels by more than the nudge (1) when all data is pulled together (MW, $p=0.08$), and (2) separately for the superspreader (MW, $p=0.06$). The effect, however, is not statistically significant for other specifications. \end{findingA} In Part 5 of \Cref{tab:nonparametric_analysis}, we compare distancing levels in the three positions -- close-knit, superspreader, and peripheral -- in Baseline. From the table, we can see that (1) distancing levels are higher in the superspreader position relative to both the close-knit and the peripheral, and (2) higher in the close-knit relative to the peripheral. We run hypothesis tests both (a) aggregating over the rate of contagion, and (b) separately by two levels of the rate of contagion. All results in this part of the table are significant at the 5\% level, with most also being significant at 1\%. The differences between mean distancing levels in different positions are also large in magnitude -- between 10\% and 27\% depending on the specification. Part 6 of \Cref{tab:nonparametric_analysis} repeats the analysis in Part 5, but now focusing on Intervention, rather than Baseline. The conclusions here are the same as above, with all results being significant at the 5\% and most also at the 1\% level. \begin{findingA} Superspreaders practice more social distancing than close-knit participants, who, in turn, practice more distancing than peripheral participants. The differences are both statistically significant (MW, $p<0.05$) and large in magnitude in all specifications. \end{findingA} In Parts 7 and 8 of \Cref{tab:nonparametric_analysis}, we investigate the effects on the rate of contagion on distancing behavior, separately for Baseline and Intervention. We find that there is generally significantly more social distancing under 65\% rate of contagion relative to 15\%. The test is not statistically significant only for the superspreader in Baseline. The effect is also large in magnitude -- an average of 8-22\% depending on the specification. Note that the effect on superspreaders is smallest, but is probably explained, at least in part, by ceiling effects. In particular, in all treatments, the mean distancing levels in the last 10 rounds in the superspreader position are at least 70\%, so there is limited room for a further increase. \begin{findingA} Subjects in the experiment generally practice more social distancing in high contagion environments relative to low contagion environments. The effect is statistically significant for close-knit (MW, $p=0.9$ in Baseline and $p=0.009$ in Intervention) and peripheral (MW, $p=0.03$ and $p=0.002$ in Baseline and Interventon respectively) subjects. For superspreaders, the effect is relatively smaller, and less robust to specifications (MW, $p=0.16$ in Baseline and $p=0.05$ in Intervention). \end{findingA} Finally, in Parts 9 and 10 of the table, we focus on testing \Cref{theor:complete_network,,theor:star_network}. Specifically, we compare the outcomes in the complete and the star networks to (1) predictions of the Nash equilibrium, and (2) efficient outcomes. We find that the levels of social distancing are not in line with theoretical predictions, with all tests being statistically significant at least at the 5\% level. In particular, distancing levels in the complete network are well above those predicted by Nash equilibrium. The difference is particularly large in the low contagion environment, where the equilibrium prediction is no social distancing, but the actual average level of social distancing in the last 10 rounds of Baseline stands at 62\%. Distancing levels are also higher than those predicted by efficiency under low contagion. On the other hand, in the high contagion environment, the actual distancing levels are below efficiency requirements. \begin{findingA} Prior to intervention, the observed levels of social distancing in the complete differ from equilibrium and efficiency predictions. In particular, more social distancing is observed than predicted by Nash equilibrium. When it comes to efficiency, actual levels of distancing are above efficient under low contagion, but below efficient under high contagion. All results are statistically significant (MW, $p\leq0.01$). \end{findingA} In the star network, the amount of social distancing done by the peripheral participants is well above both equilibrium predictions and efficiency requirements. On the other hand, the levels of social distancing observed for the superspreaders is above the equilibrium but below the efficient level for low contagion, and below both the equilibrium and efficient levels for high contagion. \begin{findingA} Prior to intervention, the observed levels of social distancing in the star network differ from equilibrium and efficiency predictions. The result is true for both positions and both rates of contagion. In low contagion, participants in both positions practice more distancing than predicted by equilibrium analysis, while in high contagion superspreaders practice less and peripheral participants practice more than predicted. When it comes to efficiency, peripheral agents practice more social distancing and superspreaders practice less social distancing in both high and low contagion. All results are statistically significant (MW, $p<0.0001$). \end{findingA} \textbf{Perception and Behavior.} As part of the Post-experimental Questionnaire, subjects are asked whether they agree that fines and nudges are effective in promoting social distancing. These two questions are both on a 5-point Likert scale, with higher values indicating greater disagreement with the statements. \Cref{tab:attitudes} summarizes individual self-reported perceptions of fines and nudges in our sample. The majority of subjects believe that nudges are effective in promoting social distancing, while attitudes to fines appear to be polarized. \begin{table}[ht] \centering \caption{Self-reported perception of fines and nudges, $n = 400$.} \label{tab:attitudes} \begin{tabular}{@{}lccccc@{}} \toprule & strongly agree & agree & \begin{tabular}[c]{@{}c@{}}neither agree\\ nor disagree\end{tabular} & disagree & strongly disagree \\ \midrule fines are effective & 11.75\% & 35.5\% & 21.25\% & 20.25\% & 9.25\% \\ nudges are effective & 31\% & 50.5\% & 13.75\% & 3.5\% & 1.25\% \\ \bottomrule \end{tabular} \end{table} We construct a group-level index for the perception of fines, equal to one if at least three subjects in the group believe that fines are effective, i.e. they (strongly) agree with the corresponding statement, and an identical index for the perception of nudges. We then construct another index that captures changes in the observed behavior. This index is equal to one if the average level of social distancing in the last 10 rounds of Intervention is higher than in the last 10 rounds of Baseline. We then compare the perception indices to the behavior index. \begin{figure}[ht] \centering \includegraphics[width=0.5\textwidth]{images/perception_behavior.pdf} \caption{Perception of effectiveness of fines and nudges versus observed effectiveness.} \label{fig:perception_behavior} \end{figure} \Cref{fig:perception_behavior} summarizes how group-level perception of fines and nudges compares to behavior. Two patterns emerge. First, subjects seem to underestimate the effectiveness of fines: only 42\% of groups in the fine treatment (55\% in the nudge treatment) believe in the effectiveness of fines, whereas 70\% of groups subjected to fines show an increase in average distancing levels. Second, participants seem to overestimate the effectiveness of nudges -- 95\% of the groups in the nudge treatment (98\% in the fine treatment) believe that nudges are effective, while only 65\% of the groups actually increase their average distancing levels when subjected to nudges. To test this formally, we use non-parametric analysis with the three group-level indices. We find that irrespective of treatment they were placed in, subjects' perception of fines matches their (fines') measured effectiveness (WSR and MW, $p=0.99$ and $p > 0.92$ respectively). On the other hand, participants overestimate the effectiveness of nudges (WSR and MW, $p=0.001$ $p<0.0001$ respectively). \subsubsection{Individual level} \label{sec:analysis_individual} We now turn to individual behavior, and examine the determinants of individual social distancing decisions. To do this, we use a Random Effects Logit model. This implicitly models the following Random Utility framework for binary choice: \begin{align} y_{it} = 1 \iff x_{it} \beta + v_{i} + \epsilon_{it} > 0. \end{align} According to this model, subjects choose to practice social distancing if and only if they receive greater utility from doing so than from not doing so. Note that (implicitly) we have normalized the utility from not distancing to zero -- this is without loss of generality. This is a flexible and tractable way to model binary decisions. The model assumes that the subject-specific random effect, $v_{i}$, is normally distributed and that $\epsilon_{it}$ follows the logistic distribution. In addition, we only use data from the final 10 rounds of each part of the experiment, because it takes time for subjects' behavior to converge as shown in the convergence analysis in Section \ref{sec:robustness_convergence}. Section \ref{sec:robustness_individual} shows that the results are not sensitive to either using the Logit modeling framework or restricting the data to the final 10 rounds of each part. Having set out the econometric framework, we now need a set of control variables that might plausibly explain social distancing decisions. Clearly, it is neither possible nor desirable to control for every conceivable covariate, but we have five categories of controls that collectively cover a wide variety of factors. First, and most obviously, we control for the experimental treatments: the fine, nudge, rate of contagion, and network position (model 1 of \cref{tab:logit_main}). As we randomly assign subjects to these treatments, we can be confident that their effects are causal. Next, we add controls for social demographics (in model 2), geographic and institutional factors (in model 3), social and risk preferences (in model 4), and finally, ideology (in model 5). To complete our model, we also control for interactions between ideology and the policy interventions (fine/nudge). This is because we find that subjects' responsiveness to the fine depends on their ideology -- more conservative subjects are \emph{less responsive} to the fine. \\ \noindent \textbf{Results.} Ex ante, it seems plausible that each of the controls could be related to social distancing decisions. However, as we can see in model 6, only some of them are.\footnote{The p-values in all results relate to model 6.} All experimental treatments have a significant effect -- both statistically and in terms of economic relevance -- and their direction is intuitive. Statistical tests are t-test on coefficients in the Random Effects Logit regression (REL hereafter), or t-test on coefficients in the instrumental variables Random Effects Logit (IV, hereafter). First, the fine significantly increases social distancing (\Cref{theor:fine_intervention}), while the nudge marginally increases social distancing (\Cref{theor:nudge_intervention}). Qualitatively, the fine is more effective (the statistical significance of the nudge is also not robust), but in our preferred model, the difference between the two effects is not statistically significant (\Cref{theor:fine_vs_nudge}). Note that the numbering of the results corresponds exactly to the number of the hypotheses in Section \ref{sec:model}, and the ``I'' prefix denotes individual-level results. \begin{findingI} The fine increases the probability that an individual practices social distancing (REL, $p=0.001$). \end{findingI} \begin{findingI} The nudge marginally increases the probability that an individual practices social distancing (REL, $p=0.09$). The effect is approximately half that of the fine and is not robust to changes in the regression specification. \end{findingI} \begin{findingI} The effect of the fine is larger than the effect of the nudge, but the difference is not statistically significant (REL, $p=0.11$). \end{findingI} Second, superspreaders distance more than `close-knit' agents, even though they have the same number of links in the network (\Cref{theor:position_effect}). This suggests that when subjects know they are relatively highly connected, they take some action to compensate. Note that our experimental design cannot identify \emph{why} they do this -- whether to protect themselves or to protect others. Peripheral agents distance significantly less -- being less exposed to community contagion in the first place reduces the likelihood that agents take protective action. Further, subjects do more social distancing in a high contagion environment than in a low contagion environment, irrespective of their network position (\Cref{theor:contagion_effect}). \begin{findingI} Superspreaders practice social distancing with a greater probability than close-knits subjects (REL, $p=0.05$), who in turn, practice social distancing with a greater probability than peripheral subjects (REL, $p<0.0001$). \end{findingI} \begin{findingI} Subjects practice social distancing with a greater probability in the high contagion setting than in the low contagion setting (REL, $p<0.0001$). \end{findingI} Beyond the experimental treatments, we find significant effects for age, gender, race, social and risk preferences, and ideology. Older, female, and non-white subjects all practice more social distancing. More risk-averse agents and prosocial agents (as classified by the SVO task) also practice more social distancing (\Cref{theor:risk_aversion} and \Cref{theor:social_values}). \stepcounter{findingI} \stepcounter{findingI} \begin{findingI} More risk-averse subjects practice social distancing with a higher probability (REL, $p=0.001$). \end{findingI} \begin{findingI} Subjects who indicate greater concern for others' well-being practice social distancing with a higher probability (REL, $p<0.0001$). \end{findingI} Perhaps the most interesting non-treatment effect is the subjects' political ideology. Subjects with stronger Conservative ideology practice less social distancing, and are marginally \emph{less responsive} to fines. Recall that our ideology index is constructed from subjects' responses to three questions from the post-experiment questionnaire. They ask about subjects' support for President Donald Trump's handling of the COVID-19 pandemic, their support for universal healthcare, and their belief that social distancing measures impose unjustified economic costs. \begin{findingI} More conservative subjects both practice distancing with a lower probability (iv, $p=0.04$) and are marginally less responsive to the fine (IV, $p=0.08$). \end{findingI} \input{tables/main_regressions.tex} It is important to bear in mind that subjects were not randomly allocated to their ideology (as they were to the experimental treatments) -- clearly, that would not be feasible. As such, the association we have found between ideology and social distancing decisions may not be causal -- it is possible that ideology is \emph{endogeneous}. That is, it could be correlated with some unobserved factor that affects social distancing decisions. Therefore, we use an instrumental variables approach to deal with possible endogeneity. We use a measure of subjects' skepticism of global warming as the instrument. As a partisan issue in the United States, this is strongly correlated with political ideology \citeplatex{mccright2011politicization}. However, as it is unrelated to COVID-19 and the types of decisions that we ask subjects to make in this experiment, it should not have a \emph{separate} effect on social distancing decisions. Mathematically, this means that after we have controlled for ideology, then skepticism of global warming is uncorrelated with the error term, $\epsilon$, in the regression. This property is required for the instrument to be \emph{valid}, and so to give consistent estimates. Note that it is not possible to test this -- it is an assumption.\footnote{See, for example, \citeplatex{wooldridge2010econometric}.} As we are using a Logit specification, it is not possible to do standard 2-Stage Least Squares. Instead we use a Control Function method \citeplatex{train2009discrete}. The Control Function method takes predicted residuals from the first stage regression, and adds them into the Logit regression. This is in contrast to 2-Stage Least Squares, which takes predicted values from the first stage, and uses them instead of the (potentially) endogenous variable in the second stage. Model 7 of \cref{tab:logit_main} reports the preferred regression specification with the instrumental variables method. Using the instrument does not have a large impact on the point estimate for the ideology variable. This suggests that there may in fact be a causal relationship between ideology and social distancing decisions -- conservatives practice less social distancing. \textbf{Partial Effects.} Using our preferred Logit specification -- model 6 of \cref{tab:logit_main} -- we can predict the probability that an agent chooses to practice social distancing. With estimated regression coefficients $\hat{\beta}$ and a set of subject characteristics $x$, then: \begin{align} Pr(y_{it}=1 | x_{it}) = \frac{ e^{ \hat{\beta}^{'} x } }{1 + e^{ \hat{\beta}^{'} x } }. \end{align} However, calculating partial effects with a Logit model is far from straightforward. In a non-linear model, the partial effect of one variable depends on the full set of an individual's characteristics, and so is highly heterogeneous. For example, the estimated partial effect of age depends on an agent's gender, race, religion, degree of risk aversion -- and \emph{all other} variables in the model. Note that this even includes variables that are not statistically significant. Given this, we calculate a variant on Average Partial Effects (APE). This is more useful than a Partial Effect at the Average; especially given our extensive use of binary variables -- a Partial Effect at the Average would give us the partial effect for a subject who is 47\% female, 85\% white, and at a node position that simply does not exist (50\% `close-knit', 40\% peripheral, and 10\% superspreader). To calculate our variant on an APE for a variable, for example, gender, we predicted the probability of social distancing for each of our 400 subjects, first assuming that they are all male, and then second assuming that they are all female. This gives us an individual partial effect of gender for each subject -- taking an average across all subjects yields the variant on an APE. We assume that subjects are not exposed to a policy intervention (fine/nudge) when calculating APEs for all (other) variables. When the variable is continuous, we first assume that all subjects are at the 25th percentile of that variable (based on the actual distribution in the experiment), and second that they are at the 75th percentile. \begin{figure}[ht] \centering \caption{Average partial effects. Note that all variables from model 6 of \Cref{tab:logit_main} are used in calculations, but we only report APEs for statistically significant variables.} \label{fig:ape} \includegraphics[width=0.9\textwidth]{images/fig_3_app.png} \end{figure} We use this variant -- only looking at the 400 subjects in our experiment -- due to data constraints. While data is readily available on the population-wide distributions of most of the individual variables, they are only available as marginal distributions, not as joint distributions. That is, one can easily find an age distribution, a gender distribution, and an education distribution for the US population; but they are only available separately. A joint distribution -- especially over \emph{all} of the variables in our preferred specification is not available. Therefore, our variant allows us to calculate partial effects for individuals who actually exist, and so makes our APE meaningful. Figure \ref{fig:ape} shows a box plot Partial Effects for variables that are statistically significant. The colored bars cover the 25th to 75th percentiles of estimated partial effects. The vertical line in each bar shows the 50th percentile, and the black dot shows the mean (the APE). \subsubsection{Design} \label{sec:experiment_design} \textbf{Calibration.} Throughout the experiment, we set the cost of social distancing at $c = 35$ points, and the benefit of not getting infected at $b = 100$ points. We also set the probability of patient zero becoming infected if she practices social distancing at $\gamma = 0.5$. Further, our experiment has three dimensions which are varied in our treatments. These are: \begin{itemize} \item two 5-node \textbf{networks}: (1) a complete network, where everyone is connected to everyone else, and (2) a star network, where one node is connected to all other nodes which are not connected between themselves; \item two levels of the \textbf{rate of contagion}: (1) low ($\alpha = 15\%$), and (2) high ($\alpha = 65\%$). \item two types of \textbf{intervention}: (1) a fine for not practicing social distancing $f = 15$ points, and (2) a behavioral nudge highlighting the harm caused to others by not practicing social distancing. \end{itemize} Consequently, we have a $2 \times 2 \times 2$ full-factorial design, with a total of 8 treatments. We run each treatment 10 times, resulting in a final sample of 400 participants. Further details on the resulting sample are in Section \ref{sec:experiment_data}. In the remainder of this section, we set out the flow of the experiment. For expository purposes, we focus on the star network, 65\% rate of contagion treatment. We also explain both fine and nudge interventions. \begin{figure}[t] \centering \includegraphics[width=\textwidth]{images/schema_draft_v7.pdf} \caption{Flow of a typical round of baseline and intervention. In the experiment, we use the following parameterization: $f = 0$ or $15$ points (fine intervention only); $c = 35$ and $b = 100$ points; $\alpha = 0.15$ (low contagion) or $0.65$ (high contagion). Final payoffs for the round are a combination of individual social distancing choice and infection status. For example, a participant who practices social distancing and is healthy, receives $(-c+b) = 65$ points. In the figure, the chosen network architecture is the star. In the experiment, half of the treatments had the star network architecture and the other half had the complete.} \label{fig:round_flow} \end{figure} Once a participant joins the session, she enters her ID and starts working on instructions for the first part of the experiment, which we refer to as the Baseline. To qualify for the experiment she must read the instructions and pass an understanding quiz. The instructions together with the quiz take an average of 8.8 minutes (s.d. 3.3 minutes) to complete. As part of the instructions, the participant is explicitly primed to think about COVID-19. In particular, we explicitly describe the main symptoms of COVID-19 using the guidelines by Centers for Disease Control and Prevention. Full instructions are in Section \ref{sec:instructions_main_experiment}. Upon completing the quiz, the subject joins a waiting room where she waits to be matched with four others. Rather than being allocated into groups on a first-come-first-served basis, group allocation is randomized. Once a group is formed, participants proceed to the Baseline where they play 20 rounds of the same game. \Cref{fig:round_flow} presents the flow of a typical round. At the beginning of the round, participants are randomly allocated to the five positions on the network (top-left corner of \Cref{fig:round_flow}) which is fixed throughout the whole experiment. In this example, the network is a star. The participant is asked to privately make her social distancing decision (top-center part of \Cref{fig:round_flow}). Practicing social distancing costs 35 points, while not practicing social distancing is free. The participant has 20 seconds to make her decision, otherwise, she receives a penalty of 200 points and her decision is automatically recorded as a ``No''. A participant who fails to submit her decisions in three consecutive rounds is disqualified from the experiment and does not receive any compensation. Once everyone in the group has made their decisions, the computer randomly selects exactly one participant to be \emph{patient zero} (top-right corner of \Cref{fig:round_flow}). If patient zero is not practicing social distancing she becomes infected with COVID-19 with certainty. If patient zero is practicing social distancing, however, she becomes infected with probability 50\%. In the example in \Cref{fig:round_flow}, patient zero decided not to practice social distancing, and therefore she becomes infected. As explained in Section \ref{sec:model}, since patient zero is infected and does not practice social distancing, she may pass COVID-19 to other participants. In our example, patient zero only interacts with the superspreader -- who is not practicing social distancing. Therefore, the superspreader becomes infected with probability $\alpha = 0.65$. In our example, contagion is successful, and the superspreader becomes infected (bottom-right corner of \Cref{fig:round_flow}). Next, the superspreader can potentially infect two out of three other participants in peripheral positions, since one of them is practicing social distancing and so is protected from contagion. In our example, one of the participants in the peripheral position becomes infected, which happens with probability 65\%, and the other remains healthy, which occurs with probability 35\%, as shown in the bottom-center part of \Cref{fig:round_flow}. At the end of the contagion process, payoffs for the round are determined based on participants' social distancing decisions and infection status. Healthy participants receive a bonus of $b = 100$ points while infected ones get 0 points, minus the cost of social distancing $c = 35$ points if applicable. In our example, three participants end up with a payoff of zero points, one gets 65 points, and one 100 points (bottom left corner of \Cref{fig:round_flow}). At the end of each round, a participant is informed of her health outcome and the number of points earned and is also reminded about her position within the network and social distancing choice. This information is presented for the last five rounds, and the participant has 15 seconds to review this information. Notice that the participant is \emph{not} informed about decisions and outcomes of other members of her group at any point during the experiment. Once participants complete the Baseline, they are taken to the instructions for the second part of the experiment, which we refer to as Intervention. The instructions for the Intervention differ depending on the treatment. For the fine treatment, participants are shown textual instructions explaining how the fine works. If instead, the intervention is the nudge, participants are shown a three-minute video highlighting the harm to others caused by not practicing social distancing. Further details and instructions for the two types of intervention are in Section \ref{sec:instructions_main_experiment}. Upon completing the instructions, participants must pass a one-question understanding quiz. Once all five participants pass the quiz, they proceed to the Intervention. In the Intervention, participants play 20 rounds of the same game as in the Baseline. In particular, the network and the rate of contagion remain unchanged. For the nudge treatment, payments also remain unchanged. In the fine treatment, participants who decide not to practice social distancing receive a fine of $f=15$ points, irrespective of their health outcome. That is, a participant who does not practice social distancing in a round can receive either 85 points or -15 points depending on her health outcome. Once participants have completed the Intervention, the interactive part of the experiment is over, and they proceed to the Post-experimental Questionnaire. Here, we ask the same set of demographics questions as in the recruitment survey. In addition, we ask a set of knowledge and attitudes questions on a range of topics including the COVID-19 pandemic, social distancing, religion, global warming, ideology, and political affiliation. Finally, a participant completes a Bonus Task which is the Bomb Risk Elicitation Task (BRET) \citeplatex{CF_2013, HP_2016} to elicit risk preferences. The participant is presented with 100 boxes arranged in a 10 $\times$ 10 matrix. Of these, one randomly chosen box contains a bomb, but the location of the bomb is unknown. The participant is asked to choose how many boxes she wants to collect. Boxes are collected from the top-left corner of the matrix, left to right, at a rate of one box per second. The participant must decide when to stop collecting boxes. The value of each box is 2 cents. After she is done with collecting boxes, the subject opens them. If one of her collected boxes contains the bomb, it explodes and reduces the participant's earnings for the Bonus Task to zero. If the bomb is not in the collected boxes, the participant earns 2 cents for each collected box. Assuming a power utility function\footnote{Utility of payoff $x$ is defined as $u(x)=x^r$, where $r$ is the risk aversion coefficient.}, a risk-neutral participant opens 50 boxes. If a participant opens fewer than 50 boxes, she is considered risk-averse, and if she opens more than 50 boxes she is considered risk-seeking. Instructions for the BRET together with screenshots of the interface are in Section \ref{sec:instructions_bret}. Once the participant completes the Bonus Task, she is taken to the Payment Page, where she can see her earnings for the experiment. She can also browse her history of play in the Baseline and Intervention. All participants are paid a fixed fee of 1 USD. Additionally, participants earn a performance-based bonus for the interactive part of the experiment, as well as the Bonus Task. In particular, to reduce wealth effects \citeplatex{CGH_2016}, subjects are paid for 4 randomly chosen rounds in Baseline and Intervention. \subsubsection{Implementation} \label{sec:experiment_implementation} We conducted the experiment on Amazon Mechanical Turk (henceforth, Mturk; \url{www.mturk.com}) between May 4th and May 29th, 2020. The experiment involves two separate processes -- recruitment and the main experiment. We used Qualtrics (\url{www.qualtrics.com}) for recruitment, and the main experiment was programmed in o-Tree (\url{www.otree.org}) \citeplatex{CSW_2016} with a server deployed on Heroku (\url{www.heroku.com}). For recruitment, we set qualifications on MTurk to make our survey visible to US residents who have completed at least 500 Human Intelligence Tasks (HITs) on the platform and have an approval rate of at least 96\%. We also utilize location and age qualifications provided by MTurk to get a sample that is broadly representative of the US population. The recruitment survey takes an average of 5 minutes (s.d. 3 minutes) to complete and pays a fixed reward of 1 USD. As part of the survey, we collect information about participants' age, gender, experience with decision-making experiments, and self-reported attitudes to risk \citeplatex{WBB_2002}. Additionally, we inform participants about the upcoming interactive experiment and collect their consent for participation. Further, the survey contains several questions that test understanding of basic concepts of probability. Only those participants who correctly answer the qualifying questions and give their consent for participation are subsequently invited to the main experiment. Participants who correctly answer the qualifying questions take part in a bonus task for a chance to win an amount in the 0.6-4.0 USD range. The bonus task is the 6-item Social Value Orientation (SVO) task \citeplatex{MAH_2011}. The underlying idea of the SVO framework is that people vary in terms of their motivations when evaluating different allocations of resources between themselves and others. Consequently, the SVO scale identifies four different types of preferences: individualistic, competitive, prosocial, and altruistic. On a practical level, for each of the 6 items of the SVO scale, participants are asked to choose between nine different allocations of money between themselves and another person. These preferences are then used to determine subjects' types. Further details on the distribution of types in our sample are in Section \ref{sec:experiment_data}. The instructions for the SVO scale along with screenshot of the interface from the recruitment survey are in Section \ref{sec:instructions_svo}. We use the U.S. Census Bureau 2018 estimates of the demographics composition of US states to get the age-gender-state distribution of the adult American population \citeplatex{census_2019b}. With respect to age, we use 6 categories that match those provided by MTurk qualifications: 18-25, 25-30, 30-35, 35-45, 45-55, and 55+. We also group US states into 10 Standard Federal Regions as defined by the Office of Management and Budget.\footnote{Circular A-105, ``Standard Federal Regions,'' April, 1974.} \Cref{fig:distribution_location} shows the map of the US divided into these 10 regions. The resulting distribution is presented in \Cref{tab:population_distribution}. Note that we do not consider US territories. During recruitment, we use the distribution in \Cref{tab:population_distribution} to generate a diverse standing panel of qualified participants. We then email a random sample from the panel, inviting them to the experiment. In the invitation email, we ask participants to search for the experiment on MTurk at the specified time and click on the link provided in the description of the HIT. Doing this transfers them to our application on Heroku. For a typical session of the experiment, we invite a representative sample of 150-200 participants to fill 30-35 places on a first-come-first-served basis. We rely on the distribution in \Cref{tab:population_distribution} to determine the demographics of the invitees and then draw a random sample with the required characteristics from our standing panel. Throughout the experiment, we keep track of the demographics of those participants who have completed the experiment to minimize over- and under-sampling. Finally, to ensure random assignment into treatments, we randomize the order of experimental sessions. It is important to note, however, that it is impossible to control precisely the demographics of the participants who actually (a) show up, and (b) complete all parts of the experiment. Nevertheless, the final sample is very diverse and near-representative in terms of age, gender, and geographic location. Further details on the sample and its characteristics and representativeness of the adult US population are in Section \ref{sec:experiment_data}. On average, the experiment takes 30 minutes (s.d. 6 minutes) to complete and pays 5.81 USD (s.d. 1.1 USD) including a 1 USD fixed fee. Upon completing the experiment, participants receive a code to submit on MTurk. Throughout the experiment, we keep track of the IP addresses of those participants who already took the experiment (or have seen the instructions), and exclude participants with duplicate IP addresses from our standing panel. \subsubsection{Data} \label{sec:experiment_data} The experimental dataset contains decisions of 400 participants. Each participant took part in one session, so she took part in one of the eight treatments. For all treatments, we collected data for 10 groups of 5 participants each. In each treatment, participants interacted for a total of 40 rounds, so we have 16,000 decisions in total. We also match the experimental data with data from the recruitment survey. To check there are no biases in assigning subjects to treatments, we run a chi-squared test on participants' assignment to treatment (a categorical variable with 8 options) and gender, age category, and geographical location variables. Results indicate that subjects in all eight experimental treatments are not significantly different from each other (two-sided $\chi^2$, $p>0.05$). This suggests that our randomization procedures outlined in Section \ref{sec:experiment_implementation} were effective. Apart from data on subjects' decisions in the experiment, we collect data on a set of variables, which can be broadly categorized as follows: demographic controls, preference controls, and location-based controls. \Cref{tab:controls} presents summary statistics for some of these controls. \input{tables/data_summary_stats} \textbf{Demographic controls.} All participants are residents of the US, 47\% are female, and the mean age is just under 44 years. 84\% of our subjects are white, and 70\% are employed (either full- or part-time). Further, we estimate the number of years of education received based on educational attainment.\footnote{To do this we assume that all subjects took the standard number of years to complete each qualification, and undertook no education that did not lead to a qualification.} Based on this, the average subject in our sample has 15.3 years of education. Finally, 51\% of our participants consider themselves religious, with 39\% reporting Christianity as their religion. \begin{figure}[t] \centering \includegraphics[width=\linewidth]{images/distribution_svo_bret.pdf} \caption{Bomb Risk Elicitation Task (BRET) and Social Value Orientation (SVO) Scale distributions for the sample. We draw a vertical line on the subplots for each subject whose score in BRET/SVO is of the corresponding value. More intense line color indicates that more subjects are concentrated at that value.} \label{fig:distribution_bret_svo} \end{figure} \textbf{Preference controls.} A total of 92.5\% of subjects reported that they try to stay at home as much as possible because of the COVID-19 pandemic. Of these, 61.6\% (or 57\% of the full sample) report that the desire to `protect others' is one of the main reasons behind this decision. To capture subjects' political ideology, we construct an index from three questions in the Post-experimental Questionnaire. These questions ask about: 1) subjects' support for President Donald Trump's handling of the COVID-19 pandemic, 2) their support for universal healthcare, and 3) their belief that social distancing measures impose unjustified economic costs. All are on a 5-point Likert scale \citeplatex{likert1932technique}, and so yield a 13-point index (0-12). Responses that indicate support for President Trump, opposition to universal healthcare, and belief in the unjustified economic costs of social distancing measures are scored positively. Higher scores, therefore, indicate a more Conservative ideology. In our sample, the mean ideology score is 3.1 (s.d. 3.1), and over 75\% of subjects have a score of 5 or less. We also collect information on participants' attitudes to climate change. Specifically, our post-experimental survey includes three 5-point Likert scale questions asking whether subjects believe that global warming is (1) happening, (2) caused mostly by human activity, and (3) affecting weather in the United States \citeplatex{HMML_2015a}. The resulting climate change attitudes index is on a 0-12 scale, with higher scores indicating greater skepticism towards climate change. The mean score in our sample is 1.99. Further, as explained in the previous sections, we collect information on subjects' social value (SVO) and risk (BRET) preferences. \Cref{fig:distribution_bret_svo} presents the distributions of these preferences for our sample. The average subject in the sample is moderately risk-averse, with a BRET score of 34.73 boxes. When it comes to social values, the majority of our subjects (58\%) are classified as prosocial (with SVO angle $\geq 22.45\degree$ and $<57.15\degree$). The second-largest category in the sample are individualists (the angle is $\geq 12.04\degree$ and $<22.45\degree$). Notice that we only have 1 subject who is classified as competitive (angle $<-12.04\degree$), and no subjects classified as altruistic (angle $\geq57.1\degree$). \textbf{Location-based controls.} As part of both recruitment and the experiment, we collect subjects' IP-addresses. We use these to infer their geographic location, down to the county-level. \Cref{fig:distribution_location} shows the region-level distribution of our sample. \begin{table}[H] \centering \begin{threeparttable} \caption{Age-gender-location distribution of the US adult population.} \label{tab:population_distribution} \begin{tabular}{@{}rl|adadadadadad@{}} \toprule \multirow{2}{*}{Region} & &\multicolumn{2}{c}{18-25} &\multicolumn{2}{c}{25-30} &\multicolumn{2}{c}{30-35} &\multicolumn{2}{c}{35-45} &\multicolumn{2}{c}{45-55} &\multicolumn{2}{c}{55+} \\ & &\multicolumn{1}{c}{M} &\multicolumn{1}{c}{F} &\multicolumn{1}{c}{M} &\multicolumn{1}{c}{F} &\multicolumn{1}{c}{M} &\multicolumn{1}{c}{F} &\multicolumn{1}{c}{M} &\multicolumn{1}{c}{F} &\multicolumn{1}{c}{M} &\multicolumn{1}{c}{F} &\multicolumn{1}{c}{M} &\multicolumn{1}{c}{F} \\ \midrule I & & 0.29 & 0.29 & 0.20 & 0.19 & 0.19 & 0.19 & 0.34 & 0.35 & 0.38 & 0.41 & 0.85 & 1.00 \\ II & & 0.51 & 0.50 & 0.41 & 0.40 & 0.39 & 0.39 & 0.70 & 0.71 & 0.73 & 0.77 & 1.51 & 1.83 \\ III & & 0.56 & 0.54 & 0.43 & 0.42 & 0.41 & 0.41 & 0.74 & 0.76 & 0.78 & 0.81 & 1.70 & 2.01 \\ IV & & 1.21 & 1.15 & 0.92 & 0.91 & 0.83 & 0.84 & 1.58 & 1.65 & 1.66 & 1.74 & 3.64 & 4.33 \\ V & & 0.99 & 0.95 & 0.73 & 0.70 & 0.67 & 0.66 & 1.27 & 1.27 & 1.31 & 1.33 & 2.88 & 3.33 \\ VI & & 0.83 & 0.78 & 0.63 & 0.61 & 0.60 & 0.58 & 1.10 & 1.10 & 1.01 & 1.03 & 1.97 & 2.29 \\ VII & & 0.28 & 0.26 & 0.19 & 0.18 & 0.18 & 0.18 & 0.34 & 0.34 & 0.33 & 0.33 & 0.77 & 0.89 \\ VIII & & 0.25 & 0.23 & 0.19 & 0.18 & 0.18 & 0.17 & 0.33 & 0.31 & 0.28 & 0.27 & 0.60 & 0.65 \\ IX & & 0.98 & 0.92 & 0.81 & 0.76 & 0.75 & 0.71 & 1.34 & 1.31 & 1.27 & 1.28 & 2.54 & 2.92 \\ X & & 0.26 & 0.24 & 0.22 & 0.2 & 0.21 & 0.20 & 0.38 & 0.36 & 0.34 & 0.34 & 0.76 & 0.85 \\ \midrule \multicolumn{13}{r}{\textbf{Total:}} & 100\% \\ \bottomrule \end{tabular} \begin{tablenotes} \item All numbers are in percentage terms; `18-25' etc. -- age categories; `M' -- males, `F' -- females; we divide states into 10 Standard Federal Regions (Office of Management and Budget) defined as follows: I -- CT, ME, MA, NH, RI, VT; II -- NJ, NY; III -- DE, DC, MD, PA, VA, WV; IV -- AL, FL, GA, KY, MS, NC, SC, TN; V -- IL, IN, MI, MN, OH, WI; VI -- AR, LA, NM, OK, TX; VII -- IA, KS, MO, NE; VIII -- CO, MT, ND, SD, UT, WY; IX -- AZ, CA, HI, NV; X -- AK, ID, OR, WA. Note that we exclude US territories. \end{tablenotes} \end{threeparttable} \vspace*{1cm} \includegraphics[width=\linewidth]{images/distribution_location.pdf} \captionof{figure}{Location distribution, created with www.mapchart.net. States are divided into 10 Standard Federal Regions, using the same definition as \Cref{tab:population_distribution}. X/Y indicate participants counts, where X indicates the actual count in our sample, and Y -- the target representative count.} \label{fig:distribution_location} \end{table} Further, using location information, we create a set of location-based control variables. First, for each subject, we record population density at the county level. The average density in the sample is 2.4k per square mil ~\citeplatex{census_2011,census_2019a}. Second, we use location data together with data on the development of the COVID-19 pandemic in the US. In particular, for each subject, we record the number of cumulative COVID-19 cases (average 63.4k) and new COVID-19 related deaths (average 59) in the state on the day of the experiment. We also record whether stay-at-home orders were in place in the participant's state on the day of the experiment. In our sample, 62\% of the subjects had stay-at-home orders in place on the day they participated in our experiment. \textbf{Sample representativeness.} We check the representativeness of our sample along the three dimensions as defined in Section \ref{sec:experiment_design} -- i.e. age, gender, location. \Cref{tab:population_distribution} presents the distribution of the US adult population. The gender distribution of our sample is not statistically different from that of the adult US population (two-sided t-test, $p=0.11$). Further, \Cref{fig:distribution_location} presents a map of the US divided into the 10 Standard Federal Region. For each region, we indicate the target number of people and the realized count in our sample. The difference between the observed and target location distributions is not significant (two-sided $\chi^2$, $p=0.08$). When it comes to age, our distribution is unfortunately not quite representative of the adult US distribution (two-sided $\chi^2$, $p<0.0001$). In particular, the 35-45 and 45-55 age categories are over-represented, while the 55+ is under-represented. \subsection{Main Experiment} \label{sec:instructions_main_experiment} This section contains instructions for the Baseline and Intervention parts of the experiment. For the Baseline (Part 1), the rate of contagion is set at 65\%. Note that the type of network and intervention do not feature in this part of instructions. For the Intervention (Part 2), we show instructions for the fine. Instructions for the nudge are similar, except that instead of explaining how the fine is implemented, participants are asked to watch a 3-minute video. The video can be accessed online at \url{https://youtu.be/tyf6EpSMeGs}. The section concludes with \Cref{fig:experiment_interface}, which shows the decision and results interfaces from the main experiment. \textbf{Part 1: Instructions (page 1/6)} Welcome to this interactive experiment! This experiment consists of two parts. You will be paid a \textbf{fixed reward of \$1} for completing all parts of the experiment. Additionally, you can earn points for your choices in Parts 1 and 2, which will be converted into \$ at the end of this experiment. There may also be a Bonus Task at the end of the experiment. In Parts 1 and 2 of the experiment, you will \textbf{interact with a group of other real people recruited through MTurk}. Recall that you will never learn the identities of other people and no one will learn about your identity. The expected duration of the experiment is \textbf{30-40 minutes} and your average expected total earnings will be \textbf{\$3-6 excluding the Bonus Task}. Note that you can earn less or more than this amount depending on your choices and the choices of others in your group. At the beginning of Part 1 of the experiment, you will be randomly allocated to a \textbf{group of 5} and you will \textbf{remain in this group} for the duration of Parts 1 and 2 of the experiment. Note that you might have to wait while we are matching you with 4 other people, but we will compensate you for the time you wait. Since this experiment is interactive, it is important that you \textbf{remain continuously attentive}, otherwise you may slow down others and may even be disqualified from the experiment. In Part 1 of the experiment, you will be asked to play a game with the other members of your group. The instructions on the next 5 pages explain the rules of the game. You will receive more information about Part 2 after you complete Part 1. All participants are given the \textbf{same} instructions. It is important that you read these instructions carefully. Note that there is no deception in this experiment. Once you read the instructions, you will be required to pass a short understanding Quiz. \textbf{If you fail the Quiz, you will not be allowed to take part in the experiment and will not receive the fixed reward.} To continue to the instructions, press the Next button below. \textbf{Part 1: Instructions (page 2/6)} In Part 1 of the experiment you are asked to play a game with the other members of your group. In what follows, you and the other members of your group are referred to as participants. At the start of the game, you are presented with a diagram with 5 circles labeled by capital letters (P, E, C, M, Q) and lines between them. Each circle represents a \textbf{position} -- one for each participant. At the start of the game, each participant is \textbf{randomly allocated} to one of these positions in the diagram. Your position is the one colored in blue. The lines between positions indicate the structure of interactions between participants in these positions. These lines \textbf{indicate which participants interact} with one another in the game. An example is the diagram below. Here, you are in position M and directly interact with participants in positions P and E, but you do not interact directly with C and Q. \vspace{0.3cm} \centerline{\includegraphics[scale=.5]{example_network.png}} The next page of the instructions explains the choice you need to make in the game. To continue to the next page, press the Next button below. To go to the previous page, press the Back button. \textbf{Part 1: Instructions (page 3/6)} In the game, you and the other participants face a \textbf{risk of getting infected with COVID-19} (commonly known as coronavirus). The main symptoms of COVID-19 are shortness of breath, a high fever and a new, continuous cough. Most patients experience mild symptoms and recover in 1-2 weeks, but cases can progress to pneumonia and organ failure in the most vulnerable individuals. After you learn your position in the diagram, your need to \textbf{choose whether to practice social distancing}. Below you can see what the buttons to make your choice look like. \vspace{0.3cm} \centerline{\includegraphics[scale=.75]{images/interface_buttons.png}} You have \textbf{20 seconds} to make your choice and the timer is displayed at the top of the interface throughout the experiment. If you do not make a choice within the allowed time, your choice is automatically recorded as a ‘No’. After everyone in your group has made their social distancing choice, the computer randomly chooses \textbf{one and only one participant to contract COVID-19 directly}. If a participant does not practice social distancing and is randomly picked by the computer then s/he gets infected for sure. In other words, \textbf{if you do not practice social distancing there is a 20\% chance you get infected with COVID-19 directly}. Social distancing gives you some level of protection against COVID-19. If you practice social distancing and you are the participant randomly chosen by the computer to be infected then the computer flips a fair coin. If the coin flip is Head you are infected with COVID-19. If the coin flip is Tail then you are not infected with COVID-19. In other words, \textbf{if you practice social distancing there is a 10\% chance you get infected with COVID-19 directly}. A participant who practices social distancing cannot pass COVID-19 to other participants and cannot be infected with COVID-19 by another participant. On the other hand, an infected participant who does not practice social distancing may infect other participants through contagion. In particular, other participants who do not practice social distancing face a risk of getting infected because COVID-19 may spread through \textbf{interactions} between the participants. The next page of the instructions explains how COVID-19 spreads through interactions between participants. To continue to the next page, press the Next button below. To go to the previous page, press the Back button. \textbf{Part 1: Instructions (page 4/6)} A healthy participant who does not practice social distancing may get infected with COVID-19 through \textbf{contagion} by interacting with infected participants who do not practice social distancing. The probability a participant contracts COVID-19 through interaction with another participant is referred to as the \textbf{rate of contagiousness} of COVID-19. \textbf{Throughout Part 1 of the experiment the rate of contagiousness of COVID-19 is fixed at 65\%.} Consider again the example diagram of interactions and, as an example, suppose that: \vspace{0.3cm} \centerline{\includegraphics[scale=.5]{example_network.png}} \begin{itemize} \item You (participant M) do not practice social distancing. \item Participants in position E and Q practice social distancing, while participants in positions P and C do not. \end{itemize} In this example, suppose C is randomly picked by the computer to contract COVID-19 directly. First, C has chosen not to practice social distancing so s/he gets infected for sure. Moreover, C can pass COVID-19 to other participants because s/he does not practice social distancing. Second, E and Q cannot get infected with COVID-19 through contagion because they practice social distancing. Next, P may get infected through contagion because s/he does not practice social distancing and interacts with C. This can happen with probability 65\% -- the rate of contagiousness. Finally, you may also become infected by contagion through your interaction with P. Specifically, there is a 65\% probability you might get infected through your interaction with P if s/he becomes infected. However, if P remains healthy you will also remain healthy for the duration of the game. It follows that in this example if 1) you do not practice social distancing, 2) E and Q practice social distancing whereas P and C do not, and 3) C contracts COVID-19 directly, then you may become infected with probability 42.25\%. The diagram below shows how this percentage is computed. \vspace{0.3cm} \centerline{\includegraphics[scale=.75]{images/65_percent.png}} Note that you will not be informed of the choices of other participants at any point during the experiment. The next page of the instructions explains how you earn points in Part 1 of the experiment. To continue to the next page, press the Next button below. To go to the previous page, press the Back button. \textbf{Part 1: Instructions (page 5/6)} At the end of the game you earn the following points depending on your social distancing choice and infection status: \begin{itemize} \item \textbf{100 points:} if you did not practice social distancing and did not get infected; \item \textbf{65 points:} if you practiced social distancing and did not get infected (100 points for being healthy minus 35 points cost of social distancing); \item \textbf{0 points:} if you did not practice social distancing and got infected; \item \textbf{-35 points:} if you practiced social distancing and got infected (0 points for being infected minus 35 points cost of social distancing). \end{itemize} Note that \textbf{if you fail to submit} your social distancing choice then you will receive a \textbf{penalty of 200 points}. The information about the points you can earn, the rate of contagiousness of COVID-19, the structure of interactions between participants and your position are always displayed on the screen when you make your social distancing choice. Below you can see an example of how this part of the interface looks like. The diagram of interactions is on the right, while the textual information on the left reminds you of the rate of contagiousness and the possible outcomes. \vspace{0.3cm} \centerline{\includegraphics[scale=.75]{images/interface_decision.png}} At the end of the game, you are reminded of the structure of interactions between participants, your position within that structure, and your social distancing choice. You are also informed of your infection status and the number of points earned. Below you can see an example of how this part of the interface looks like. \vspace{0.3cm} \centerline{\includegraphics[scale=.75]{images/interface_results.png}} You have 15 seconds to review this information and the timer is always displayed at the top of the interface. The next page of the instructions explains how points are converted into your earnings in \$. To continue to the next page, press the Next button below. To go to the previous page, press the Back button. \textbf{Part 1: Instructions (page 6/6)} Part 1 of the experiment has \textbf{20 separate games} as described in these instructions. \textbf{The choice you make in one game has no effect on other games.} The only variation between games is the random reassignment of the positions of all the participants (including you) in the diagram of interactions. The participants assigned to your group, the structure of interactions between positions, the probability of contracting COVID-19 directly, the rate of contagiousness of COVID-19 and the number of points earned depending on social distancing choice and infection status remain unchanged. At the end of each game, you can review the history of your choices and outcomes for the last 5 games. The table below shows an example of how your history after 9 games might look like. \vspace{0.3cm} \centerline{\includegraphics[scale=.75]{images/interface_history.png}} It is important that you make a choice in every game. If you fail to make a choice for \textbf{3 consecutive games}, you will be \textbf{disqualified from the experiment}. In this case, you will not receive any payment for this experiment. At the end of this experiment, the computer randomly picks \textbf{4 out of 20 games} to determine your earnings for Part 1 of the experiment. \textbf{The points are converted to \$ at a rate of 115 points per \$1.} Suppose that you earn 260 points in the 4 randomly drawn games. Then, your total earnings for Part 1 are \$2.26. To continue to the short Quiz, press the Next button below. To go to the previous page of the instructions, press the Back button. \textbf{Part 2: Instructions} You have completed Part 1 of the experiment and will now proceed to Part 2. Below are the instructions for Part 2 of the experiment. It is important that you read these instructions carefully. This part of the experiment also has 20 games, and you are assigned to the same group of 5 people as in Part 1. The structure of interactions between participants, the probability of contracting COVID-19 directly and the rate of contagiousness of COVID-19 are the same as in Part 1. \textbf{The single difference from Part 1 is that in Part 2 of the experiment you will receive a fine of 15 points in any game in which you do not practice social distancing.} Hence, in Part 2 of the experiment the points you earn at the end of the game are: \begin{itemize} \item \textbf{85 points:} if you did not practice social distancing and did not get infected (100 points for being healthy minus 15 points fine); \item \textbf{65 points:} if you practiced social distancing and did not get infected (100 points for being healthy minus 35 points cost of social distancing); \item \textbf{-15 points:} if you did not practice social distancing and got infected (0 points for being infected minus 15 points fine); \item \textbf{-35 points:} if you practiced social distancing and got infected (0 points for being infected minus 35 points cost of social distancing). \end{itemize} Note that if you fail to submit your social distancing choice then you will receive a penalty of 200 points. Your earnings for Part 2 are computed in the same way as in Part 1. At the end of this experiment, the computer randomly picks 4 out of 20 games to determine your earnings for Part 2 of the experiment. As in Part 1, the points are converted to \$ at a rate of \textbf{115 points per \$1.} Suppose that you earn 300 points in the 4 randomly drawn games. Then, your total earnings for Part 2 are \$2.61. Before you can start Part 2 of the experiment, you must answer a \textbf{Quiz question on the instructions above}. If you fail to answer the question correctly, you will not be able to continue to Part 2 of the experiment. To continue to the short Quiz, press the Next button below. \textbf{The experiment: Interface} \Cref{fig:experiment_interface} shows the decision and results interface from the main experiment. We focus on game 5 of the Intervention part of the experiment, where the rate of contagion is set at 65\%, the network is the star, and fine is the intervention. In this game, our example participant is assigned to position P. She decides to practice social distancing, and does not get infected. \begin{figure}[!ht] \begin{minipage}{\linewidth} \centering \subfloat[Decision interface]{\label{fig:decision_interface} \includegraphics[width=0.95\textwidth]{images/interface_decision_full.png} } \end{minipage} \begin{minipage}{\linewidth} \centering \subfloat[Results interface]{\label{fig:results_interface} \includegraphics[width=0.95\textwidth]{images/interface_results_full.png} } \end{minipage} \caption{Game 5 of the Intervention: interface} \label{fig:experiment_interface} \end{figure} \subsection{Social Value Orientation (SVO) Slider Measure} \label{sec:instructions_svo} This section contains instructions from and interface of the Social Value Orientation (SVO) task which participants complete as part of the recruitment survey. \textbf{SVO: Instructions} You have answered the qualifying questions correctly and are now in the Bonus Task. In this Bonus Task you will be making a series of decisions about allocating money between you and another anonymous Turker (hereafter, Other). All of your decisions will be completely confidential. There is a total of 6 decisions to make which are independent of each other. For each decision, you are asked to pick the distribution of money between yourself and Other that you prefer, all values are stated in cents. After you have made your decision, select the resulting distribution of money by clicking on the button below your choice. As you will see, your choices will influence both the amount of money you receive as well as the amount of money the Other receives. There are no right or wrong answers, this is all about personal preferences. Every time 50 Turkers complete the task, we will randomly pick two of them and pay them for the Bonus Task as follows. We will pick one of the 6 decisions, and randomly implement the decision of one of the two chosen Turkers. For example, suppose we randomly chose Turkers X and Y out of those 50 Turkers, and that we further randomly chose to implement decision 3 of Turker Y. Suppose that in decision 3 Turker Y allocated 150 cents to themselves, and 140 cents to Other. Therefore, Turkers X and Y will be paid 140 and 150 cents respectively. \textbf{SVO: Interface} \Cref{fig:svo_interface} presents part of the interface of the SVO task. We focus on one of the six decisions. In this example, a participant decided to allocate 288 cents to self and 188 cents to other. \begin{figure}[ht] \centering \includegraphics[width=0.9\textwidth]{images/svo_choice.png} \caption{Social Value Orientation (SVO) Slider Measure: Decision interface} \label{fig:svo_interface} \end{figure} \subsection{Bomb Risk Elicitation Task (BRET)} \label{sec:instructions_bret} This section contains instructions from and interface of the Bomb Risk Elicitation Task (BRET) which participants complete after the main experiment. \textbf{BRET: Instructions} Thank you very much for taking part in the experiment! You are now in the Bonus Task in which you have an opportunity to earn an extra payment. On the next page, you will see 100 boxes. As soon as you start the task by pressing the \textbf{`Start'} button, one box is collected per second starting from the top left corner. Once collected, the box is marked by a tick symbol. \textbf{For each box collected, you earn 2 cents.} One of the 100 boxes contains a \textbf{bomb} that destroys all of your earnings. \textbf{You do not know where the bomb is located.} You only know that the bomb can be in any box with equal probability. Your task is to choose when to stop collecting the boxes and open those you have collected. You stop collecting boxes by pressing \textbf{`Stop'} at any time. After that you open the boxes you have collected by pressing the \textbf{`Open'} button. Note that once you press `Stop' you cannot restart collecting boxes. A dollar or a bomb symbol will be shown on each of the boxes you have collected. If the bomb symbol does not appear, that means that you have \textbf{not collected the box with the bomb}. In this case, you \textbf{earn the amount accumulated by the boxes you have collected}. If the bomb symbol appears, that means that you \textbf{have collected the box with the bomb}. In this case, you \textbf{earn zero} for the Bonus Task. To proceed to the Bonus Task, press the button below. \textbf{BRET: Interface} \Cref{fig:bret_interface} presents the interface of the BRET. \Cref{fig:bret_choice} shows an example of the interface after a participant has stopped collecting boxes. We can see that she opened 56 boxes. \Cref{fig:bret_result} shows the interface after she opened the boxes. We can see that the bomb is among the 56 collected boxes. Consequently, in this example the participant earns zero for the BRET. \begin{figure}[ht] \begin{minipage}{.5\linewidth} \centering \subfloat[Decision interface]{\label{fig:bret_choice} \includegraphics[scale=0.6]{images/bret_choice.png} } \end{minipage} \begin{minipage}{.5\linewidth} \centering \subfloat[Results interface]{\label{fig:bret_result} \includegraphics[scale=0.6]{images/bret_result.png} } \end{minipage} \caption{Bomb Risk Elicitation Task (BRET) interface} \label{fig:bret_interface} \end{figure} \section{Experimental Design} This section describes the game, the workflow of the experiment, and treatments. \textbf{Game.} \Cref{fig:round_flow_main} illustrates the social distancing game which corresponds to a round of the experiment. At the beginning of a round, participants are randomly allocated to one of the positions in an unweighted and undirected network (top left). The structure of the network is common knowledge. Subjects must simultaneously decide whether to practice social distancing at a known cost $c>0$ (top center). After the decisions are made, one and only one participant -- \emph{patient zero} -- is randomly picked to contract COVID-19 directly. In the example in \Cref{fig:round_flow_main}, patient zero is the participant highlighted in red in the top right panel. \begin{figure}[ht] \centering \includegraphics[width=\textwidth]{images/schema_draft_v7.pdf} \caption{Flow of a typical round of baseline and intervention. In the experiment, we use the following parameterization: $f = 0$ or $15$ points (fine intervention only); $c = 35$ and $b = 100$ points; $\alpha = 0.15$ (low contagion) or $0.65$ (high contagion). Final payoffs for the round are a combination of individual social distancing choice and infection status. For example, a participant who practices social distancing and is healthy, receives $(-c+b) = 65$ points. In the figure, the chosen network architecture is the star. In the experiment, half of the treatments had the star network architecture and the other half had the complete.} \label{fig:round_flow_main} \end{figure} There are two benefits to practicing social distancing. First, if the individual is randomly picked to be patient zero, then she becomes infected with $50\%$ probability, rather than for sure. Second, any individual who practices distancing cannot transmit COVID-19 to others or contract it from infected individuals. In the example in \Cref{fig:round_flow_main}, patient zero decided to not practice social distancing, hence she becomes infected. COVID-19 then spreads through the network from infected to healthy individuals who do not practice social distancing through contagion with a known probability $\alpha \in (0,1)$. The bottom panels of \Cref{fig:round_flow_main} show a possible spread in the example with three individuals infected and two healthy ones at the end of the contagion process. At the end of the round, healthy individuals receive a benefit of $b=100$ points, whereas those infected receive zero benefit. Any individual who chose to practice social distancing pays the cost $c=35$ points, irrespective of their final infection status. Assuming self-interested fully rational agents, the model predicts that the pure strategy Nash equilibrium outcomes will be inefficient for a wide range of parameter values. In fact, given an arbitrary network, it is possible to fully characterize both equilibrium and efficient outcomes.\footnote{See Online Appendix for the full characterization.} In particular, we show that, depending on the network structure, the equilibrium number of agents practicing social distancing can be below the efficient number. In this model, social distancing is essentially a merit good so it exhibits positive externalities -- welfare loss may therefore occur in the absence of intervention. \textbf{Workflow.} In the experiment participants first play 20 rounds of the baseline game above in fixed groups of five subjects, with positions on the network being randomly reallocated in each round. Notice that the participants are not informed about the decisions and outcomes of other members of their group at any point during or after the experiment. After the initial 20 rounds, their group is treated with either the fine or the behavioral nudge policy intervention. In the fine treatment, whenever a participant decides not to practice social distancing she receives a fine of $f = 15$ points, independent of her infection status at the end of the round. In the nudge treatment, every participant must watch a 3-minute video explaining how failing to practice social distancing may harm others. The same group of 5 participants then play 20 rounds of the game under one of the policy interventions. Hence we investigate the impact of each policy on social distancing behavior using a within-subjects design and the relative effectiveness of the two policies with a between-subjects design. Once participants have completed the interactive part of the experiment, they proceed to the post-experimental questionnaire, with basic demographics questions and a set of knowledge and attitudes questions on a range of topics including the COVID-19 pandemic, social distancing, religion, global warming, ideology, and political affiliation. Finally, participants complete the Bomb Risk Elicitation Task (BRET) \citep{CF_2013} to elicit risk preferences. Subjects are explicitly primed to think about COVID-19 in the instructions by naming the disease and describing the main symptoms according to the guidelines by Centers for Disease Control and Prevention. \textbf{Treatments.} The focus of our experiment is on three treatment dimensions. First, we consider two 5-node network architectures for the structure of interactions among participants: (1) a complete network, where everyone is connected to everyone else, and (2) a star network, where one node is connected to all other nodes which are not connected between themselves. Next, we vary the level of contagiousness of COVID-19 to be either low ($\alpha = 15\%$) or high ($\alpha = 65\%$). Finally, we focus on two intervention methods: (1) a fine for not practicing social distancing $f = 15$ points, or (2) a behavioral nudge in the form of an informational video highlighting the harm caused to others by not practicing social distancing. Consequently, we have a $2 \times 2 \times 2$ full-factorial design, with a total of 8 treatments. \section{Data} In this section, we summarize our data collection methods and present the resulting dataset. The section also includes a description of the key demographics of the final sample and a convergence analysis of the decision data from the experiment. Following standard practices in interactive online experiments \citep{GY_2015, SW_2011}, we first recruit a standing panel of subjects using a short survey on MTurk. During recruitment, we use the 2018 US Census data to generate a representative panel of the adult US population in terms of age, gender, and geographical location \citep{census_2019}. As part of recruitment, we also collect subjects' social preferences using the Social Value Orientation (SVO) task \citep{MAH_2011}. The SVO task classifies individuals into four categories -- prosocial, individualistic, competitive, and altruistic -- although a typical sample is predominantly composed of individualistic and prosocial individuals only. For each of the experimental sessions, we invite a random representative sample of subjects from our standing panel.\footnote{We restrict participation to US-resident participants who have completed at least 500 tasks and have an overall ranking of at least 96\%. We also ensure that no two subjects with the same IP address participate in the experiment. Finally, in line with recent contributions \citep{rand2011dynamic,gallo2015effects}, multiple participation is not allowed even by participants who have only seen the instructions.} Despite the possibility of selection bias in participating in the experimental session, we obtain a near-representative sample of $n=400$ participants completing the experiment. To check there are no biases in assigning subjects to treatments, we run a chi-squared test on participants’ assignment to treatment (a categorical variable with 8 options) with respect to gender, age category, and geographical location variables. Results indicate that subjects in all eight experimental treatments are not significantly different from each other (two-sided $\chi^2$, $p >0.05$). The experiment took place between May 4th and 29th, 2020. Each session had between one and five independent groups of five subjects playing simultaneously, with the exact numbers depending on turnout. The experiment lasted on average 33 minutes, and subjects earned an average of 5.81 USD (including a fixed participation fee of 1 USD). Subjects remained completely anonymous throughout the experiment, and informed consent was obtained from subjects before participation. All 400 subjects in our sample are residents of the US, 47\% are female, and the mean age is just under 44 years. The average subject in our sample has 15.3 years of education and 70\% are employed (either full or part-time). In terms of religious attitudes, 51\% of our participants consider themselves religious, with 39\% reporting Christianity as their primary religion. The average subject in the sample is moderately risk-averse, with a CRRA score of $0.53-0.55$. With respect to social values, our sample mostly consists of prosocials (58\%) and individualists (41.75\%), and only one subject is classified as competitive.\footnote{See Online Appendix for a more detailed description of the demographics of the sample.} Overall, the final dataset contains 40 decisions for each of the 400 subjects for a total of 16,000 observations. \Cref{fig:1}A shows the evolution of average individual propensity to social distance aggregated over contagiousness and network treatments. The figure shows a downward trend in social distancing behavior in the first rounds. Such behavior is typical for public goods experiments whereby contributions to the public good decline as the session progresses \citep{ledyard1994public}. In this paper, we are interested in the limiting outcomes of this convergence process, and therefore we would like to restrict our attention to those rounds where the majority of subjects have settled on some stable strategy. In particular, we define that a participant has converged to a stable strategy by a certain round if she used this strategy for the previous four rounds, and in all subsequent rounds, the number of consecutive deviations from the chosen strategy does not exceed two.\footnote{We consider three types of convergence strategies. In both networks, we look at the strategy where the subject always chooses the same action. For the star network, we also consider two extra strategies. In one strategy the participant always chooses the same action when she is in the superspreader position and the complement action when she is peripheral. In the other strategy, she always chooses the same action when she is the superspreader and alternates between the actions when peripheral.} Given the convergence definition above, we find that at least $80\%$ of participants converge to a stable strategy in every treatment after the first 10 rounds. Our main analysis, therefore, focuses on the last 10 rounds of the baseline and intervention in order to control for the time trend due to learning and experimentation by subjects. The results are robust to including all of the data.\footnote{Our analysis is robust to using a more/less conservative definition with one or three consecutive deviations allowed. See Online Appendix for a complete description of our robustness checks.} \section{Analysis}\label{sec:results} This section investigates the determinants of social distancing decisions using both non-parametric and parametric methods. It includes a discussion of the impact of political ideology and the instrumental variable approach we use to identify its causal effect. The non-parametric analysis uses the Mann-Whitney U-test (MW hereafter) \citep{MW_1947} for unmatched samples and the Wilcoxon Signed-Rank test (WSR hereafter) \citep{W_1992} for matched samples. Samples are aggregated at the group level and matched when they are generated by the same groups, and they are unmatched when making comparisons across groups from different treatments.\footnote{All results using these non-parametric tests are robust to using their parametric analogs, i.e. t-test for unpaired and paired samples.} We use a Random Effects Logit (henceforth, REL) to model the binary choice of whether to practice social distancing. Formally, it models a random utility model with individual-specific random effects: \begin{align} y_{it} = 1 \iff x_{it} \beta + v_{i} + \epsilon_{it} > 0 \end{align} where $y_{it}$ is the decision to practice social distancing, $x_{it}$ is the set of controls, $v_{i}$ are individual specific random-effect and $\epsilon_{it}$ is an error term with a logistic distribution. It allows individual-specific propensity to practice social distancing, but assumes that the impact on utility of a variable $x$ is not individual-specific. We cluster errors at the group-level. Our main specification (column 5 in \Cref{tab:main text table}) includes the following five categories of independent variables $x_{it}$: (1) dummy variables for the experimental treatments (fine intervention, nudge intervention, contagion level, and node types), (2) demographics controls, (3) location and institutional controls, (4) preference controls, and (5) ideology controls. P-values throughout this section are presented for a single specification, but, as \Cref{tab:main text table} shows, results are robust across all specifications.\footnote{Further, the results are robust to using a Linear Probability Model or a Probit model. We discuss this further in the Online Appendix.} A preliminary question is whether the policy intervention has an impact on the aggregate level of social distancing. We conduct a Wald test to identify a structural break. This test assumes that there is at most one structural break in the data, but is agnostic as to when it occurs and whether it occurs at all. In the fine intervention, a change in the evolution of social distancing occurs at round 21, immediately after the implementation of the fine (Wald test, $p<0.001$ for all data and subdivided by the other treatments). The same break at round 21 occurs in the nudge intervention in the aggregate data (Wald, $p<0.001$), but the result is not robust once we subdivide by the other treatments. The introduction of the fine therefore has an immediate impact on social distancing behavior, while the impact of the nudge is less clear-cut. \textbf{Results.} Fines increase the level of social distancing, and the effect is significant both in the non-parametric (WSR, $p=0.0008$) and in the regression (Specification 1, S1 hereafter, $p=0.001$) analyses. The increase in social distancing behavior is present both in low (WSR, $p=0.04$) and high (WSR, $p=0.003$) contagion environments. In the last 10 rounds of the fine intervention, there is an $8\%$ increase in social distancing compared to the last 10 rounds of the pooled baseline so the effect of the fine is sizeable and long-lasting. In contrast, the impact of the nudge intervention is marginal and not robust to all specifications (WSR, $p=0.008$; S1, $p=0.03$). Interestingly, narrowing the focus to a specific contagion environment, the nudge significantly increases distancing with high contagion (WSR, $p=0.0003$) but it has no effect with low contagion (WSR, $p=0.5$). Comparing the two policies, the impact of the fine is marginally higher than the nudge (MW, $p=0.08$). Overall, the increase in social distancing in the last 10 rounds due to the nudge intervention is only $2\%$. \begin{figure}[h] \centering \includegraphics[width=0.9\linewidth,keepaspectratio]{images/fig1_consolidated.jpg} \caption{Policy and network position. (A) Evolution of aggregate probability of distancing split by policy intervention. (B) Probability of distancing split by network position and policy intervention. Node size is proportional to distancing probability. Width of dotted band around the circumference of each circle indicates 95$\%$ confidence intervals.} \label{fig:1} \end{figure} A crucial driver in the spread of COVID-19 is the presence of superspreaders -- individuals who go on to infect a much larger number of others than the average \citep{AWWLTCLC_2020, LGTBNL_2020}. A primary determinant of being a superspreader is biological -- something that cannot be varied experimentally and is outside of the scope of this study. However, there is also a social element driven by the wide variations in the number of social interactions across individuals \citep{jackson2007meeting}. An important difference between biological and social drivers of being a superspreader is that an individual is aware of the latter, and therefore the tendency to social distance may vary with position in the network. \Cref{fig:1}B illustrates the three positions in our experiment drawn with the node size proportional to the amount of social distancing in that position for each policy treatment. On the left side, there is the complete network in which all participants are in what we dub a close-knit position. On the right side, there is the star network with one participant -- the superspreader -- interacting with all others in what we dub a peripheral position. Superspreaders practice more social distancing than peripheral participants both in the baseline and policy interventions (MW and S1, $p<0.0001$ for all). Moreover, superspreaders also practice social distancing more than close-knit participants (MW, $p=0.004$; S1, $p=0.03$) despite having the same number of interactions. This may be because they are aware of their central role in spreading COVID-19 and want to protect the group, or because they realize peripheral participants are less likely to distance due to their isolated position. High contagiousness leads to significantly more social distancing (MW, $p=0.01$, S1, $p < 0.0001$), and the size of the effect is similar to the difference between being in the close-knit and peripheral positions. Social distancing is $10\%$ and $15\%$ higher in the high contagion setting than in the low contagion setting in baseline and intervention respectively. Social preferences play an important role in the decision to social distance. First, an analysis of equilibrium decisions with self-interested individuals predicts levels of social distancing that are significantly lower than participants' decisions in all treatments (MW, $p<0.0001$). Second, using the SVO task, we classify subjects into individualistic and prosocial. As \Cref{fig:2}A illustrates, prosocial participants are $20\%$ more likely to choose distancing (S2, $p<0.0001$). In terms of sociodemographics, females (S2, $p=0.004$) and older individuals (S2, $p<0.0001$) are more likely to distance. From the data, being female translates to a $12.6\%$ increase in the probability of distancing, and there are no significant interactions between gender and the effect of each policy intervention. Social distancing is increasing with age -- an average 60-year-old is $40.1\%$ more likely to distance than an otherwise identical 20-year-old. Interestingly, older subjects are relatively more responsive to the nudge (S3, $p=0.02$) and relatively less responsive to the fine (S3, $p < 0.0001$). Whites are less likely to distance (S2, $p=0.04$), but the effect is not robust. Education, religion, and employment status are not associated with distancing behavior. These associations between sociodemographic characteristics and distancing are broadly consistent with evidence from sociomobility \citep{BMB_2020} and survey-based data \citep{MSBAKMHLW_2020, GPPBBF_2020, PF_2020, PZB_2020}. We elicit participants' risk preferences using BRET. As expected, risk-seeking individuals are less likely to practice social distancing (S2, $p=0.001$), and the effect is about half the size of being prosocial in the SVO task. The post-experimental survey includes a question asking why subjects choose to stay at home -- participants who list ``protecting others'' as one of the stated reasons are more likely to social distance (S2, $p = 0.007$) and the effect is similar in size to being risk-averse.\\ \input{tables/main_text_table} There is no association between distancing and the geographical evolution of the pandemic as captured by new cases, cumulative cases, and total deaths at the state level on the day of participation in the experiment. Similarly, institutional decisions such as the timing of stay-at-home orders in different states are unrelated to social distancing decisions.\footnote{The Online Appendix contains more details on the non-significant covariates.} \begin{figure}[t] \includegraphics[width=\linewidth, keepaspectratio]{images/fig2_consolidated_v3.jpg} \caption{Ideology and social preferences. (A) Probability of distancing split by social preference and policy intervention. (B) Probability of distancing and (C) Percentage change in probability of distancing split by policy intervention and quintile of ideology index. Vertical bars indicate 95\% confidence intervals.} \label{fig:2} \end{figure} \textbf{Ideology.} We find that self-reported Democrats are significantly more likely to practice distancing than Republicans (S4, $p=0.02$). However, this binary classification is a very coarse measure. To obtain a more fine-grained picture of participants' political leanings, we construct an ideology index, with a score for each subject based on responses to questions about (1) support for President Trump’s handling of the COVID-19 pandemic, (2) support for universal healthcare, and (3) belief that social distancing measures impose unjustified economic costs. Questions are on a 5-point Likert scale, so we obtain a 0-12 index with 0 and 12 indicating extremely progressive and extremely conservative participants respectively. \Cref{fig:2}B shows that the probability of practicing distancing decreases the more conservative a participant is -- an increase in ideology index from 1 to 5 (25th to 75th percentile among our subjects) corresponds to a $14.6\%$ decrease in the probability of distancing. The ideology index is a highly significant correlate of social distancing decisions (S5, $p<0.0001$), in line with recent sociomobility and survey-based studies in the US \citep{ABCGTY_2020, BH_2020, SSDB_2020}. Our experimental design exogenously varies policy intervention, contagiousness, and the network, allowing us to make causal inferences on these dimensions. The same is obviously not possible with ideology. We investigate the causal effect of ideology with an instrumental variable approach (IV hereafter), using a measure of participants' skepticism of global warming as the instrument. As a partisan issue in the US, attitudes toward global warming are strongly correlated with political ideology \citep{HB_2016, PEW_2019}. The exclusion restriction our IV identification relies on is that global warming attitudes do not affect social distancing decisions separately from their association with ideology. Specification 6 (S6) in \Cref{tab:main text table} is identical to our main specification (S5), except that it uses an instrument for ideology. The instrument is an index of climate change attitudes -- constructed from subjects' beliefs that global warming is (1) happening, (2) caused mostly by humans, and (3) affecting weather in the USA.\footnote{We used the questions from \cite{HMML_2015}.} Political ideology is a significant causal determinant of social distancing, with conservatives less likely to practice distancing (S6, $p=0.04$). \Cref{fig:2}C illustrates an interesting further interaction between political ideology and the type of policy intervention. While the nudge intervention has no differential impact on subjects with different ideologies, conservative participants are marginally less responsive to the fine (S6, $p=0.08$). Note that using the ideology index itself (as in specification 5) instead of the IV approach (as in specification 6) has very little effect on the results. \textbf{Interpreting the results.} The nonlinearity of the REL model makes interpreting the point estimates in \Cref{tab:main text table} difficult. The partial effect of any given variable depends on both the initial value for that variable and on the values of all variables. For a given individual, the partial effect of changing a variable from $x$ to $x'$ is not the same as that of changing it from $x'$ to $x''$. As a consequence, the estimated partial effect is different for every person. To meaningfully compare the effect of different factors, we compute the Average Partial Effects (APE) calibrated on the characteristics and/or decisions of the subjects in our experiment. To do this, we calculate the partial effect for each participant in the experiment, and then take the simple mean over these individual-specific partial effects. For the individual-specific partial effect, we compare the estimated probability of distancing for each subject for two different values of the variable of interest while keeping all other variables at their observed value for that individual. For binary variables, we compare $0$ and $1$. For continuous variables, we compare the $25$th and $75$th percentiles (based on the experimental data). \begin{figure}[b!] \includegraphics[width=0.9\linewidth, keepaspectratio]{images/fig_3.png} \caption{Average Partial Effects (APE) for variables that are significant in the REL main specification. Bar boundaries indicate 25th and 75th percentiles, vertical line inside bars indicate 50th percentile, and black dots show the mean. Blue bars indicate treatment variables. Orange and yellow bars indicate effects from social preferences and demographics respectively. Purple bar is the ideology variable. Variables not included because they are not significant include: education, religion, labor force participation, case numbers, and geography controls.} \label{fig:3} \end{figure} \Cref{fig:3} shows the mean partial effects (APEs) for all variables that are statistically significant in our main specification, as well as their 25th, 50th, and 75th percentiles. Blue bars indicate the impact of our treatment variables. A high contagion setting and being in the peripheral network position have the largest impact with $18\%$ and $-22\%$ respectively. The fine intervention and being in the superspreader position have a moderate impact of about $8\%$. The smallest impact is the nudge intervention with only a $4\%$ effect. Yellow bars indicate sociodemographics correlates. Increasing age from the 25th to the 75th percentile in our data has a $23\%$ effect on distancing. Gender and race have a moderate effect of around $10\%$. Orange bars display the effect of correlates related to preferences. Being prosocial is associated with a large $20\%$ effect on distancing, while risk preferences and the desire to protect others have a moderate impact of about $10\%$. Finally, the purple bar shows that political leanings have a moderate $12\%$ effect, which we argue is causal in our analysis. Omitted from \Cref{fig:3} are a set of demographic, institutional, and geographic factors that are not statistically significant. A subject's level of education, religion, and labor force participation are all unrelated to their social distancing decisions. Also insignificant are their geographic location (as measured by Federal Region), the number of COVID-19 cases and deaths in their state, and population density. Finally, \Cref{fig:3} clearly shows how the partial effect of any given variable is highly heterogeneous across subjects. In fact, the range in partial effects from the 25th to 75th percentiles \emph{exceeds} the mean in many cases. \section{Concluding remarks} The sudden disruption brought about by the COVID-19 pandemic demanded a quick response from policymakers who had to implement novel social distancing policies whose success determined the fate of tens of thousands of lives. Our work provides some of the first experimental evidence on how to enforce these policies and how their efficacy varies across social and individual characteristics. Additionally, it shows how the novel methodology of interactive web-based experiments can play a crucial role to inform policymakers, and complement data obtained using sociomobility and survey-based studies. Our first main finding is that fines are effective at increasing social distancing, while the impact of nudges is marginal. A limitation of our design is that we examine one specific form of nudge -- a video that emphasizes the harms to others of failing to practice social distancing. However, the efficacy of nudges is sensitive to the content, media, and cultural contexts. For instance, \cite{NBERw27496} show that nudges in the form of 2.5-minute information videos are effective in encouraging social distancing in the Indian context. Despite their low impact in our study, future research should continue exploring the efficacy of nudges in promoting social distancing because their ease of deployment makes them a particularly attractive tool for policymakers. The second main finding is that conservative-leaning individuals are less likely to practice social distancing in the US. While this tendency has been documented in other empirical studies \citep{gollwitzer2020, ABCGTY_2020}, we provide evidence of a causal relationship using an IV approach. An open question is whether this relationship is peculiar to the US context or it extends to other countries with a less stark partisan divide. In terms of methodology, this study highlights the important role interactive web-based experiments can play in providing timely information to policymakers to test the effectiveness of policies. Experiments are the gold standard to test the causal impact of policy interventions, but lab and/or field experiments were severely constrained during the COVID-19 pandemic, depriving researchers of the standard tools to identify causal effects \citep{HM_2020}. Web-based experiments, however, were unaffected. Moreover, they can be deployed quickly, scaled up at minimal cost, and easily reach a diverse sample and/or specific populations of particular interest. We believe they should be part of the standard policymaker toolkit to test the impact of novel policies before their implementation to the general population. Evidence regarding behavioral responses to social distancing policies has so far been mostly based on either sociomobility data or survey studies based on self-reported claims about hypothetical behavior. Aside from a clean identification of causality, web-based interactive experiments complement these methodologies and present some further distinct advantages. Mobility data provides detailed information on behavior in real-life settings that can verify the external validity of web-based experiments. For instance, recent studies validate our finding that conservative-leaning individuals practice less social distancing \citep{ABCGTY_2020, BH_2020}. However, it is usually stripped of any personal information, including sociodemographics, so that inferences about the effects of these variables can only be made for geographic units rather than individuals. Moreover, sociomobility data is unable to shed light on the effectiveness of counterfactual policies, while policymakers frequently need information in advance of implementation. Surveys are a standard methodology for fast, large-scale data collection, but our study raises some question marks about their accuracy in gauging actual behavioral responses. In our post-experimental survey, we ask participants to estimate how effective fines and nudges are in promoting social distancing. Participants significantly overestimate the efficacy of nudges compared to their actual choices in the experiment independently on whether they are in the fine (MW, $p<0.0001$) or nudge (WSR, $p=0.001$) treatment. This difference between survey-based self-reported intentions and actual behavior is consistent with recent evidence on the short-lived efficacy of nudges \citep{BFL_2017}. In other words, picking a policy based on self-reported measures may lead policymakers to opt for the wrong solution. These results caution against excessive reliance on survey-based self-reported data to gauge behavioral responses. \subsubsection{Convergence}\label{sec:robustness_convergence} The analysis in Sections \ref{sec:analysis_aggregate} and \ref{sec:analysis_individual} focuses on the last 10 rounds of Baseline and Intervention. In this section, we validate this choice by showing that the majority of participants exhibit clear convergence behavior after the first 10 rounds in both parts of the experiment. We define individual convergence as follows: \begin{definition} A participant converges to a strategy $s$ by round $n$ if (i) she used this strategy for the last $k$ rounds (including $n$), and (ii) in all subsequent rounds $[n+1, 20]$ the number of consecutive deviations from the chosen strategy does not exceed $a$. \end{definition} We consider three types of convergence strategies as follows: \begin{enumerate} \item The subject always chooses the same action. We define this strategy for both the complete and the star networks. \item The participant always chooses the same action when she is the superspreader, and the complement action when she is peripheral. We define this strategy for the star network only. \item The subject always chooses the same action when she is the superspreader and alternates between the two actions when she is peripheral. We define this strategy for the star network only. \end{enumerate} \input{tables/convergence} We set $k = 4$ and $a=2$. This means that the earliest a participant can be considered to converge to a particular strategy is by round 4 if she has not deviated from this strategy in the last four rounds, and in all subsequent rounds, she never performs more than two consecutive deviations. \Cref{tab:convergence} summarizes convergence analysis using the above definition. From the table, we can see that by round 11 at least 80\% of our participants have converged to a certain strategy in all parameterizations. In fact, the lowest and highest convergence rates are 82.5\% and 92\% respectively, while the weighted mean convergence rate is 85.9\% across the two parts of the experiment. Convergence is higher (a) in the complete network and (b) with the fine intervention. Note that we did not discretize by the rate of contagion, although the findings are robust to this. However, with this discretization, the lowest observed convergence rate falls to 79\% (Baseline with the star network under 15\% rate of contagion). \begin{figure}[t] \centering \includegraphics[width=\textwidth]{images/convergence.pdf} \caption{Evolution of the share of converged participants throughout the xxperiment separately for Baseline and Intervention, $a \in [1,3]$. Note: we exclude the first 3 rounds in both parts since $k=4$.} \label{fig:convergence} \end{figure} As a robustness check, we also consider $a=1$ and $a=3$, allowing for one and three consecutive deviations respectively. \Cref{fig:convergence} plots the share of converged participants for each round separately by parts for $a \in [1,3]$. We can see that the share of converged subjects does not change much when we allow for a more/less conservative definition. In particular, with $a=1$ the share of subjects who converge by round 11 in Baseline drops to 79.8\% while in Intervention it reaches 83.5\%. With $a=3$ the share of subjects who converge by round 11 in Baseline and Intervention stands at 86.3\% and 90\% respectively. The above analysis suggests that it is reasonable to claim that the absolute majority of subjects converge to a particular strategy by round 11 in both parts of the experiment. \subsubsection{Aggregate analysis} \label{sec:robustness_aggregate} We test the robustness of findings in Section \ref{sec:analysis_aggregate} in two ways: (1) using data from all 20 rounds of both parts of the experiment, and (2) employing parametric rather than non-parametric tests on the data from the last 10 rounds of both parts. \Cref{tab:nonparametric_robustness} summarizes the results of these robustness checks. In the columns under `Robustness \#1', we present the results of our hypothesis tests when using decision data from all rounds. From the tables, we can see that most of the results from Section \ref{sec:analysis_aggregate} are robust to using data from all 20 rather than the last 10 rounds of both parts of the experiment. Further, recall that we observed previously that the nudge effects appear to be small and sensitive to specification. Part 3 of \Cref{tab:nonparametric_robustness} provides further evidence strengthening this observation -- we see that the effect of the nudge on distancing levels is now insignificant in most of the specifications. In our second robustness check, we use the standard unpaired t-test (UT) on unmatched samples instead of the MW test, and the paired version of the t-test (PT) on matched samples in place of the WSR test. In the columns under `Robustness \#2' of \Cref{tab:nonparametric_robustness}, we repeat the analysis from Section \ref{sec:analysis_aggregate} using these parametric tests. Comparing the tables, we can see that all significant effects are still present if we use parametric rather than non-parametric tests. Overall, the findings of this section suggest that the observations made in Section \ref{sec:analysis_aggregate} are robust to changes in the distributional assumptions of the tests, and inclusion of all observations. \input{tables/nonpar_robustness} \subsubsection{Individual analysis} \label{sec:robustness_individual} We now test the robustness of the parametric analysis in Section \ref{sec:analysis_individual}. Model 1 of Table \ref{tab:logit_robustness} restates our preferred specification to aid comparison. Models 2 and 3 vary the type of model: they use a Linear Probability Model (LPM) and a Probit model, respectively. Both include random effects, to allow for individual heterogeneity. Changing the type of model used does not materially affect the results. Note that the coefficients are not directly comparable across Logit, LPM, and Probit. Figure \ref{fig:ape_robust} shows the average partial effects under each type of model -- as we can see, the differences are small. Model 4 shows that our results are not sensitive to excluding the period before subjects converge -- they are little affected by including all rounds. The results could also be driven by subjects who either might not be paying attention or who might have a poor understanding of the game. To test for this, we exclude any subjects who chose to practice social distancing in all 40 rounds (there are 122 subjects who do this) in model 5. Then in model 6, we exclude subjects failing the quiz between the two parts of the experiment more than once, or who failed the post-experiment attention check.\footnote{A question in the post-experiment survey asked subjects to select a particular option to demonstrate that they were paying attention.} Finally, Model 7 tests an alternative specification of the instrument -- only using subjects' responses to the question of whether global warming is caused by humans.\footnote{Using the final question on its own also gives similar results in terms of magnitude and statistical significance. The question regarding whether global warming is happening gives non-significant estimates -- likely driven by the fact that there is very little variation in the responses to this question. 353 (out of 394) subjects respond that global warming is happening.} While clearly not definitive, these suggest that our findings are robust, and in particular are not driven by the choices we made on the econometric model and which, if any, observations to exclude. In our main analysis, we used an ideology index, constructed from subjects' responses to questions immediately asked after the experiment. \Cref{tab:political_party_robustness} shows that these results hold even if we use political party affiliation -- a much coarser, but much more widely available, measure of ideology. Apart from the ideology control, all variables are the same as in the main model, so we suppress them in the table for convenience. Model 1 shows the main Random Effects Logit model for reference. Model 2 uses dummies for self-identified Republicans, and `Other' (leaving self-identified Democrats as the base category). Model 3 omits the `Other' category -- leaving a direct comparison of Republicans and Democrats. Model 4 uses the instrumental variables method. Note that we only have one instrument, so it is not possible to instrument for both Republicans and Independents in the same regression. Our preferred specification only contains a single type of heterogeneous treatment effect -- the effect of the fine varies with subjects' ideology. There also appear to be some other interaction effects present. \Cref{tab:logit_heterogenous} summarizes heterogeneous treatment effects. Older subjects are \emph{relatively less} responsive to the fine (REL, $p<0.0001$); although the fine does still increase their probability of social distancing (REL, $p<0.0001$). Prosocial subjects are also marginally less responsive to the fine (REL, $p=0.07$). Subjects in a high contagion setting, and older subjects are \emph{relatively more} responsive to the nudge (REL, $p=0.006$ and $p=0.02$ respectively). This suggests that the impact of policy interventions may vary across groups, and perhaps more importantly may vary with how contagious the disease is. \clearpage \input{tables/robustness_checks_1.tex} \clearpage \input{tables/political_party_logit.tex} \clearpage \input{tables/heterogeneous_regs} \section{Online Appendix} This Appendix contains additional information. Section \ref{sec:model} presents the theoretical framework and derives the hypotheses that we investigate in the experiment. Section \ref{sec:design} summarizes our experimental design along with implementation and describes the resulting dataset. Section \ref{sec:analysis} contains details of the analysis and is followed by Section \ref{sec:further} which describes our robustness checks. \subsection{Theory} \label{sec:model} \input{chapters/model.tex} \subsection{Experiment} \label{sec:design} \input{chapters/design.tex} \subsection{Statistical analysis} \label{sec:analysis} \input{chapters/analysis.tex} \subsection{Further analysis} \label{sec:further} \input{chapters/robustness.tex} \clearpage \bibliographystylelatex{chicago} \bibliographylatex{appendix_bib.bib} \clearpage \section{Experimental Instructions} \label{sec:instructions} \setlength{\parskip}{6pt} \setlength\parindent{0pt} \renewcommand{\thefigure}{B\arabic{figure}} \renewcommand{\thetable}{B\arabic{table}} \setcounter{figure}{0} \input{chapters/instructions.tex} \end{document}

7baa98c20373229ce23b2b602407b5b91d7fe6b9

\section{INTRODUCTION} \label{sec:intro} The \gls{WFIRST}\cite{spergel_wide-field_2015} \gls{CGI} is a technology demonstration\cite{noecker_coronagraph_2016,kasdin_wfirst_2018,douglas_wfirst_2018,bailey_wfirst_2019-1} that will use high-contrast imaging and spectroscopy (coronagraphy), wavefront sensing, and wavefront control\cite{shi_low_2016,sidick_wfirst_2018}, to image planets in reflected light \cite{kasdin_wfirst_2018}. Formerly known as the Wide-Field Infrared Survey Telescope, \gls{WFIRST} is orders of magnitude more sensitive than \gls{HST} or ground-based observatories\cite{bailey_wfirst_2019}. Two primary coronagraph technologies will be tested, a \gls{HLC} coronagraph\cite{trauger_hybrid_2016} for high contrast small-\gls{FOV} ($<$0.5 as) imaging, a bow-tie \gls{SPC}\cite{balasubramanian_wfirst-afta_2015} for spectroscopy and a wide-\gls{FOV} \gls{SPC}. The design and preparation for \gls{CGI} has spawned many new modeling tools to accurately predict performance. This work attempts to summarizes the most common of these tools, both to serve as a roadmap for future potential users of \gls{CGI} and to aid other missions which may benefit from reuse of the many open source tools which have been shared by \gls{CGI} science and engineering teams. This review cannot completely capture the simulation work that has gone into developing Roman CGI but it is intended to serve as a starting reference for the community to engage with the technical and science capabilities of the instrument. Additionally, to better understand the context of these tools, we suggest previous works, as well as several papers in these proceedings, that provide detailed descriptions of the design reference mission and typical observing modes\cite{kasdin_wfirst_2018,bailey_wfirst_2019,poberezhskiy_cgi_2020,kasdin_wfirst_2020}. \begin{table}[ht] \renewcommand*{\arraystretch}{1.5} \centering \caption{Partial Listing of Open Source CGI Software Packages and Libraries} \label{tab:my_label} \footnotesize \begin{tabular}{L{3.75cm}|L{3.75cm}|c|p{6.25cm}} \hline Package & Application & References & URL\\ \hline Observing Scenarios & Complete Observation Simulation & \citenum{krist_numerical_2015}& \url{https://roman.ipac.caltech.edu/sims/Coronagraph_public_images.html}\\ PROPER & Diffraction Simulation & \citenum{krist_proper:_2007,krist_wfirst_2017} & \url{https://github.com/ajeldorado/proper-models/tree/master/wfirst_cgi}\\ \verb+FALCO+ & Coronagraph Simulation & \citenum{riggs2018falco1} & \url{https://github.com/ajeldorado/falco-matlab}, \url{https://github.com/ajeldorado/falco-python} \\ Lightweight Coronagraph Simulator & Coronagraph Simulation & \citenum{pogorelyuk_effects_2020}& \url{https://github.com/leonidprinceton/LightweightSpaceCoronagraphSimulator}\\ CZT-based Optical Propagation & Diffraction Simulation && \url{https://github.com/ARCExoplanetTechnologies/ACED}\\ \verb+WebbPSF+ & Diffraction Simulation & \citenum{perrin_simulating_2012,perrin_updated_2014} & \url{https://github.com/spacetelescope/webbpsf}\\ MSWC & Binary Star Simulation & \citenum{dsirbu2017mswc,dsirbu2018RomanMSWC} & \url{https://github.com/ARCExoplanetTechnologies/MSWC}\\ \verb+EXOSIMS+ & Mission Simulation & \citenum{savransky_exosims_2017,savransky_wfirst-afta_2016} &\url{https://github.com/dsavransky/EXOSIMS}\\ Imaging Mission Database&Mission Planning&\citenum{savransky_exploration_2019} & \url{https://plandb.sioslab.com}\\ Coronagraph convolved Debris Disks & Disk Simulation & \citenum{mennesson_wfirst_2018}& \url{https://roman.ipac.caltech.edu/sims/Circumstellar_Disk_Sims.html} \\ Known debris disk simulated scenes & Disk Simulation &\citenum{chen_wfirst_nodate}& \url{https://roman.ipac.caltech.edu/sims/Chen_WPS.html}\\ direct-imaging-sims + model spectra & Target simulation &\citenum{lacy_characterization_2019, lacy_prospects_2020} & \url{https://github.com/blacy/direct-imaging-sims}, \url{https://www.astro.princeton.edu/~burrows/wfirst/index.html}\\ Giant planet albedo spectra & Target simulation & \citenum{cahoy_exoplanet_2010} & \url{https://wfirst.ipac.caltech.edu/sims/Exoplanet_Characterization.html}\\ \hline \end{tabular} \end{table} \section{Coronagraph Simulators} \label{sec:packages} \subsection{Observing Scenarios The most detailed and physically realistic simulations of \gls{CGI} observations released to date have taken the form of numbered observing scenarios. These are referred to as OS$n$, e.g., OS9 for the ninth scenario. Physical optics simulations of the instrument including optical surfaces, coronagraphs, and wavefront sensing and control are generated using PROPER\cite{krist2007proper,krist_wfirst_2018}. The scenarios include speckle time series, derived from wavefront maps produced using \gls{STOP} modeling of the Roman observatory\cite{saini_impipeline_2017}. The OS model outputs include noisy and noiseless datacubes of coronagraphic intensity images versus time, with injected planets and realistic observing scenarios that include reference stars, and off-axis \gls{PSF}s for injecting additional targets. These scenario files are available to the public from IPAC \begin{figure}[ht] \centering \includegraphics{os9_target_roll_-11deg_processed_image_with_planets.pdf} \caption{Images of 47 Uma c generated from publicly available OS9 simulations of the \gls{HLC} including time dependent speckle and detector effects. Left: simulated image with injected planets. Right: the same scene after reference subtraction, showing injected planets at a much higher \gls{SNR}.} \label{fig:coronagraphs} \end{figure} \subsection{FALCO The FALCO\cite{riggs2018falco1} library provides a framework for running wavefront sensing and control algorithms in MATLAB and Python 3. FALCO can use a PROPER\cite{krist2007proper} model as its truth model in simulations, and examples using the CGI PROPER model are included in the publicly available FALCO repository. Example status window updates for the hybrid Lyot coronagraph are shown in Fig.~\ref{fig:FALCO}. \begin{figure}[ht] \centering \subfigure[] { \label{fig:before_wfsc} \includegraphics[width = .4\columnwidth,trim=.05in 0in .05in 0in]{fig_falco_before.png}} \subfigure[] { \label{fig:after_wfsc} \includegraphics[width = .4\columnwidth,trim=.05in 0in .05in 0in]{fig_falco_after.png}} \caption{Status report windows from FALCO before (a) and after (b) performing wavefront control using the CGI PROPER model as the truth model. The large starting shapes on the deformable mirrors are the HLC design settings combined with the settings for flattening the starting wavefront errors.} \label{fig:FALCO} \end{figure} \subsection{Lightweight Space Coronagraph Simulator} Based on FALCO\cite{riggs2018falco1}, the ``Lightweight Space Coronagraph Simulator' computes small linear perturbations about the nominal dark hole instead of propagating the full optical model. It allows quickly simulating observation scenarios with time evolving \gls{WFE}, \gls{DM} drift, \gls{LOWFSC} residual jitter and \gls{HOWFSC} \cite{pogorelyuk_effects_2020}. The sensitivities to DM commands (the Jacobian) and to \gls{WFE} were computed in 6 wavelengths using FALCO and remain valid in the linear regime (up to 10 nm phase perturbations). This allows specification of DM voltages, \gls{WFE} Zernikes, LOWFS residual jitter, detector noise and switching between broadband and narrowband modes. The Python code is designed to run fast and only requires NumPy\cite{numpy}. Fig. \ref{fig:leonid} shows an example dark hole time series (left) and dark hole image (right). \begin{figure}[ht] \centering \includegraphics[width=0.9\textwidth]{leonid_thumbnail_LSCS.png} \caption{An example of contrast evolution in presence of various wavefront instabilities in a linearized model of Roman-CGI. Left: Closing the loop allows maintaining the contrast throughout the observation. The drift rate is exaggerated to illustrate dark hole maintenance in the worst-case scenario of random walk of Zernike coefficients. Right: A broadband photon-counts image of a single exposure (used to close the contrast loop). See Pogorelyuk et al.\cite{pogorelyuk_effects_2020} for details.} \label{fig:leonid} \end{figure} \subsection{CZT-based Optical Propagation} The Chirp Z-Transform (CZT) based optical propagation library, developed at Princeton and NASA Ames, implements 2D Fraunhofer and Fresnel diffraction propagation on arbitrarily sampled input and output planes. It provides the same functionality and answer (to within numerical precision) as the more commonly used MFT technique (Matrix Fourier Transform), but is asymptotically faster. \subsection{MSWC} Binary stars represent a special challenge for a diffraction simulation due to the angular separation between the on-axis target and its off-axis companion. The off-axis companion can contribute stellar leakage due to optical aberrations at high-frequencies for every component in the optical train. The Multi-Star Wavefront Control (MSWC) package can be used to simulate binary stars, predict the expected leakage for different binary star imaging scenarios, and determine if the contrast leakage due to the off-axis companion introduces a background contrast floor for a particular Roman CGI imaging mode. This can flag known binaries as benign or requiring suppression. Depending on the Roman CGI imaging mode, wavefront control techniques may remove the binary companion's leakage \cite{sthomas15snwc, dsirbu2017mswc,dsirbu2018RomanMSWC}. \subsection{WebbPSF WebbPSF\cite{perrin_simulating_2012,douglas_accelerated_2018-1} is a \gls{PSF} simulation tool originally developed for \gls{jwst} in Python. Basic \gls{SPC} modes are currently included in WebbPSF\cite{perrin_poppy_2016}\footnote{\url{https://www.stsci.edu/jwst/science-planning/proposal-planning-toolbox/psf-simulation-tool}} and are in the process of being updated to match current filters and mask designs. \section{Yield} \subsection{Yield Calculator} Nemati\cite{nemati_sensitivity_2017,nemati_method_2020} developed an analytic model of instrument sensitivity that calculates the time-to-\gls{SNR} for known \gls{RV} exoplanets. This model has been widely used by the project team as an Excel spreadsheet and is now publicly available as part of EXOSIMS (see below). \subsection{EXOSIMS EXOSIMS\cite{savransky_exosims_2017} is an open source, full exoplanet imaging mission simulation tool, which generates a survey ensemble of possible exoplanet detections given an underlying universe (e.g. exoplanet phase curves and occurrence rates) and observatory properties such as orbit, optical system performance, and background sources. \subsection{Imaging Mission Database The online Cornell Space Imaging and Optical Systems Lab \textit{Imaging Mission Database uses stellar and \gls{RV} exoplanet physical properties\cite{batalha_color_2018}. As shown in Fig. \ref{fig:plandb}, these properties can be combined in joint distributions and compared to the instrumental sensitivity floor (blue curve) to assess the frequency of detection in an observation. Similarly, the Imaging Mission Database generates depth of search maps\cite{garrett_2017} using EXOSIMS for blind-search targets such as the EXOCAT database\cite{turnbull_exocat_2015}. \begin{figure}[ht] \centering \includegraphics[width=0.7\textwidth]{plandb_47umac_completeness.png} \caption{Map of the joint distribution of planet projected separation and $\Delta$mag for 47 Uma c generated using the Imaging Mission Database.} \label{fig:plandb} \end{figure} \section{Science Targets} \subsection{Light from Young Giant Exoplanets Some of the substellar companions and young giant planets discovered by ground-based direct imaging surveys will be viable targets for Roman-CGI spectroscopy and photometry. This will present the first opportunity to make observations of such objects at wavelengths shorter than 0.95 $\mu$m. At the time of writing, we have identified HD 984 B, $\beta$-Pic b, HD 206893b, HR 8799e, and 51 Eri b as possible targets for Roman-CGI bandpasses 1, 2, and 3 which include spectroscopy and photometry. HR 2562b, HR 8799d, HR 8799c, $\kappa$-And b, HD 95086b, HR 3549b, HD 1160b, and HIP 65426b are possible targets for photometry in bandpass 4. We used \textit{coolTLUSTY} to generate self-consistent 1D radiative-convective equilibrium atmosphere models for each possible target, assuming values for effective temperature, radius, and surface gravity taken from the literature and that atmospheres are clear with solar abundances. For some targets, we also compare to models with higher metallicities and with forsterite clouds. These are discussed in detail in Lacy \& Burrows 2020\cite{lacy_prospects_2020} and are available publicly For all these objects, the dominant spectral features are the pressure broadened Na and K resonant doublets around 0.59 $\mu$m and 0.77 $\mu$m respectively, especially for the cooler objects where there is stronger pressure broadening at the photosphere. For the hotter objects (above $\sim$1300 K), metal hydrides like CaH and MgH have not yet condensed and rained out, so their absorption features are present in the spectra. For intermediate temperature objects and higher metallicity hot objects, absorption features from TiO and VO also appear in the spectra. When forsterite clouds are included they increase the continuum flux ratio in the optical range and weaken gaseous absorption features. Combining optical spectroscopy and photometry with NIR and MIR observations of these objects from other instruments will aid efforts to characterize these young planets since changes to cloud properties and metallicity have different effects in the optical than at longer wavelengths. Ideally, Roman-CGI spectroscopy will measure the strength of the $\sim$0.73 $\mu$m methane absorption feature for a reflected-light target, and constrain parameterized models incorporating cloud properties, a temperature-pressure profile, and mixing ratios of major gas-phase absorbers\cite{lupu_developing_2016,nayak_atmospheric_2017,Damiano2020}. Whether this task is achieved will depend on the nature of the planets observed, the quality of data Roman-CGI collects, and the quality of auxiliary measurements like mass and orbital separation that are available from other sources. In the event that Roman-CGI performance is insufficient to constrain detailed models, Lacy et al.\cite{lacy_characterization_2019} put forward a set of models suitable for addressing a simpler task: assessing how cool giant exoplanets compare to the cool gas giants and ice giants in our own solar system. Saturn and Jupiter have higher cloud layers and lower metallicities than Uranus and Neptune. Lower metallicities decrease the amount of methane, and higher clouds make for a smaller amount of gas above the cloud layer. Together these effects weaken Jupiter and Saturn's methane absorption features compared to Neptune and Uranus. At shorter wavelengths, from 0.4-0.6 microns, Saturn and Jupiter exhibit absorption from an unidentified chromophore which reddens their appearance. This is similar to the mix of hydrocarbons, commonly termed Tholins, that give Titan its yellow-orange appearance. Retrievals on simulated observations showed that Roman-CGI observations should be able to fit a model consisting of a two-part linear combination of Jupiter and Neptune's reflective properties. One parameter sets the short wavelength weighting towards Jupiter versus Neptune and essentially depends on whether a chromophore is present or not. A second parameter sets the longer wavelength weighting towards Jupiter versus Neptune and mainly reflects the amount of methane present above the cloud layer. This approach supposes that Jupiter and Neptune represent two bounds of cool giant planet atmosphere behavior and that those observed will fall somewhere in between. Of course, this framework is over-simplified, but, in the absence of high resolution high SNR spectra, it could provide a useful starting point. A small grid of geometric albedo models in this form is also available\footnote{\url{https://www.astro.princeton.edu/~burrows/wfirst/index.html}}, along with a GUI for calculating the light curve through out a planet's orbit and a set of models representing Jupiter's geometric albedo with constant cloud properties but varying metallicity. \begin{figure}[ht] \centering \includegraphics[width=0.9\textwidth]{all_self_lum_targets.png} \caption{Model spectra for potential self-luminous targets including imaging and spectra. All models assume solar metallicity and an atmosphere free of clouds. Other model parameters are listed in table 2 of Lacy \& Burrows 2020\cite{lacy_prospects_2020}. Lines for different model spectra are colored in the order of the assumed effective temperature for each object. The shaded regions indicate the Roman-CGI bandpasses. The horizontal dashed lines mark the required contrast for technology demonstration success and the engineer's best estimate of the contrast that Roman-CGI can achieve. Note that the spectral resolution of model spectra shown here are higher than the Roman-CGI spectral resolution.} \label{fig:self_lum_models} \end{figure} \subsection{Reflected Light from Cold Gas Giants Giant Exoplanet Albedo Spectra and colors as a function of planet phase, separation, and metallicity from Cahoy et al\cite{cahoy_exoplanet_2010} have been released along with a reference solar spectrum. \subsection{Debris Disks} The unprecedented point source sensitivity of CGI is expected to also lead to many new scattered light detections of debris disks and exozodiacal dust. Simulations of dusty systems, particularly 1 Vir, Eps Eri, HD 10647, HD 69830, HD 95086, HR 8799, and Tau Cet were prepared by a the Preparatory Science Project: The Circumstellar Environments of Exoplanet Host Stars by Chen et al\cite{chen_wfirst_nodate,chen_circumstellar_2014} and have been publicly released and include injected giant planets. General libraries of disks\cite{mennesson_wfirst_2018} with varied morphology and geometry have convolved with coronagraph transmission functions have also been released publicly. \subsection{Fast Extended Source Simulation with a PSF Library} The effect of the spatially varying coronagraph transmission functions on complex scenes, such as exozodiacal dust models can be simulated using public \glspl{PSF} libraries\cite{douglas_simulating_2019}. These simulations are efficiently performed by generating an over-sampled scene model and applying a coronagraphic transfer function via matrix multiplication. Because of the field-dependent evolution of the PSF, a PSF must be generated for every angular offset of the pixel coordinates of the scene model. These PSFs are generated via interpolation of the PSFs available from IPAC and stored as matrices in memory. For example, an exozodiacal model can be flattened into a vector rather than a 2D array and multiplied by the matrix of interpolated PSFs since the coronagraph is still assumed to be a linear system. For details and examples of disk generation, see Milani et al. \cite{milani_faster_2020}. \subsection{PSF subtraction} Post-processing via \gls{KLIP} subtraction of residual speckles has been extensively explored for Roman\cite{ygouf_data_2015-1,ygouf_data_2016} and exoplanet extraction has been one of the data challenge topics\cite{mandell_wfirst_2019}. pyKLIP\cite{wang_pyklip:_2015} supports generic data and the documentation currently includes an example of \gls{KLIP} run on an observing scenario. \section{Conclusions} Roman CGI has has stimulated and nurtured a wide array of coronagraph and mission simulation tools. In addition to the simulation software describe above, Data Challenges\cite{hildebrandt_wfirst_2018,mandell_wfirst_2019,girard_2019_2020} have allowed the community to engage and develop additional tools\footnote{\url{https://roman.ipac.caltech.edu/sims/Exoplanet_Data_Challenges.html}}. Integral field spectrograph modes were considered for Roman and simulated in Coronagraph and Rapid Imaging Spectrograph in Python (\verb+crispy+)\cite{rizzo_simulating_2017} which may also prove useful for other missions\footnote{\url{https://github.com/mjrfringes/crispy}}. The majority of the tools described above have been developed in Python; however, other tools have been developed for estimating coronagraph noise\cite{robinson_characterizing_2015} in languages such as IDL\footnote{\url{https://github.com/tdrobinson/coronagraph_noise}}. It is hoped that the open-source availability of all these tools will allow the community to more rapidly develop science questions with Roman CGI and provide springboards for future coronagraphic missions. \acknowledgments This research has made use of the Imaging Mission Database, which is operated by the Space Imaging and Optical Systems Lab at Cornell University. The database includes content from the NASA Exoplanet Archive, which is operated by the California Institute of Technology, under contract with the National Aeronautics and Space Administration under the Exoplanet Exploration Program, and from the SIMBAD database, operated at CDS, Strasbourg, France\cite{wenger_simbad_2000}. Portions of this work were supported by the Roman/WFIRST Science Investigation team prime award \#NNG16PJ24C. Portions of this work were supported by the Arizona Board of Regents Technology Research Initiative Fund (TRIF).

e852ebd1f9fab1ca42553d61b58a1303ac99b9ba

\section{Introduction} The observed accelerated expansion of the Universe at late times is one of the most challenging enigmas in theoretical physics. Observations over the last two decades or so, using Type Ia Supernovae, the cosmic microwave background (CMB) anisotropies and large scale structure (LSS) data imply the existence of a repulsive force that dominates over gravity on cosmological (large) scales. In the context of General Relativity (GR), this suggests the presence of an energy-momentum component with a negative equation of state, referred to as Dark Energy (DE). So far, with the current available data, the best phenomenological explanation for this accelerated expansion is the cosmological constant $\Lambda$ \cite{Aghanim:2018eyx}, which behaves as a uniform vacuum energy spread over all space. However, the cosmological constant introduces extreme fine-tuning since its value is much smaller than what is predicted by quantum field theories \cite{Weinberg:1988cp,Carroll:2000fy} and also suffers from the coincidence problem. This has given rise to a plethora of different DE models based on ad-hoc ideal fluids or new scalar fields that mediate the force between particles ranging from canonical scalar fields \cite{Ratra:1987rm,Wetterich:1987fm,Caldwell:1997ii}, scalar fields with a generalized kinetic terms \cite{ArmendarizPicon:2000dh,ArmendarizPicon:2000ah}, non-minimal couplings \cite{Uzan:1999ch,Perrotta:1999am,Riazuelo:2001mg} or coupled DE models \cite{Dent:2008vd} in addition to GR. One of the more well known candidates for DE is quintessence, which is described by a slowly rolling scalar field leading to an accelerated expansion \cite{Amendola:2015ksp}. Thus, the scalar field can control the fate of the early and late Universe by dominating its energy density and posing as a source of DE respectively, see for example Ref.~\cite{Tsujikawa:2013fta} for a review. Another viable alternative to DE models can be found through covariant modifications of GR, known as Modified Gravity (MG) theories. These are motivated by high energy physics (Quantum Gravity and String Theory) and have specific signatures, e.g. a time and scale dependent Newton's constant, a different evolution of the matter density perturbations or specific patterns in the emission of gravitational waves, compared to GR. From this point of view, GR is seen as an effective low-energy theory that, as we increase the energy scale, higher order corrections are needed \cite{tHooft:1974toh}. In view of the plethora of DE and MG models, there has been an effort to provide a unified framework which encloses several of them like the Effective Field Theory (EFT) approach \cite{Gubitosi:2012hu,Hu:2013twa} or the Effective Fluid approach (EFA) \cite{Arjona:2018jhh,Arjona:2019rfn,Arjona:2020gtm}. Despite this large variety of models, current Bayesian analyses of astrophysical measurements indicate that the standard cosmological model which contains the cosmological constant $\Lambda$ and a cold dark matter component (CDM) \cite{Peebles:2002gy} outperforms all other models \cite{Heavens:2017hkr}. For about a century theoretical physicists have been on the quest to develop a theory of quantum gravity which could encompass the assumptions of Einstein's theory of GR with those of quantum field theory. Although GR has demonstrated to have a very high predictive power below the Planck scale, its quantization its troublesome since it is renormalizable only at one loop \cite{Birrell:1982ix}, thus, it is believed that GR could be the low energy limit of the more fundamental higher energy theory. There is an ongoing search to distinguish effective quantum field theories that can potentially arise within UV-complete quantum gravity theories (the Landscape) from those that cannot (the Swampland). In this regard, although not rigorously proven in string theory, some conjectures have been considered to discern the Swampland from the landscape. The two proposed Swampland criteria that we will consider (which we will define as SC1 and SC2) refer to the constraints on the field range of a scalar field $\phi$ defined by an effective field theory and to the slope of the potential of such fields respectively. In reduced Planck units these conjectures are defined as \begin{enumerate} \item SC1: The scalar field net excursion has to satisfy $\frac{|\Delta \phi|}{M_\textrm{pl}}<\Delta \sim O(1)$ \cite{Ooguri:2006in}. \item SC2: There is a lower bound for the gradient of the scalar field potential $M_\textrm{pl}\left|\nabla_{\phi} V\right| / V>c \sim \mathcal{O}(1)$ in any consistent theory of gravity when $V>0$ \cite{Obied:2018sgi}, \end{enumerate} where $\Delta$ and $c$ are positive constants of order one and the reduced Planck mass is $M_\textrm{pl}=1/\sqrt{8\pi G}$. The second Swampland criterion is violated in the $\Lambda$\text{CDM} model, since a positive cosmological constant or being at the minimum of a potential with positive energy density violates the bound \cite{Agrawal:2018own}, thus a rolling scalar field potential, i.e a quintessence model would be required. Hence, if the data supports the second Swampland criterion, it would imply hints for deviations of the $\Lambda$\text{CDM} model. The Swampland criteria aim to find constructions that are compatible with a quantum theory of gravity and it has been found that specific quintessence models can satisfy the Swampland criteria at late times \cite{Agrawal:2018own}. In Ref.~\cite{Elizalde:2018dvw} the authors used Gaussian Processes to reconstruct the form of the potential from the $H(z)$ data, finding hints of invalidating the Swampland criteria, while a similar analysis was performed in \cite{Yang:2020jze}. In Refs.~\cite{Colgain:2019joh,Banerjee:2020xcn} it was found that quintessence models and current data prefer a lower value of $H_0$ than the $\Lambda$\text{CDM} model, thus providing robust test of the Swampland conjectures. Other analyses on the other hand, have found that string-inspired quintessence models with exponential potentials are ruled out by observations and that Swampland conjectures are in tension with viable single-field quintessence models \cite{Akrami:2018ylq,Raveri:2018ddi}. It was also proposed though that this issue might be resolved with multi-field models \cite{Garg:2018reu,Akrami:2020zfz}. See also Ref.~\cite{Heisenberg:2018yae} for the implications of the swampland conjectures on dark energy. On the other hand, our motivation for using cosmography and Machine Learning (ML), both being model independent techniques, is because choosing a specific model can lead to model bias, which in turn would affect the conclusions drawn about fundamental physics. ML algorithms can help to remove biases due to choosing a priori a specific defined model and they are also ideal for events that are not well understood such as dark energy, dark matter or modifications of gravity. Another advantage is that we reconstruct the data without making assumptions on flatness or a dark energy model. Here we use a particular ML method known as the genetic algorithms (GA), which can be defined as a stochastic search approach. However, in our analysis we will use both cosmography and the GA so as to compare the two methods and examine which one provides better constraints given the current data. In particular, we will focus on quintessence as an example of our approach and using the latest compilation of the Hubble parameter $H(z)$ and the growth rate data $f\sigma_8(z)$ we analyze the cosmological implications on two Swampland criteria providing constraints both via Machine Learning and cosmography. For the former approach, we reconstruct the Hubble function $H(z)$ and $f\sigma_8(z)$ using the GA, while with the later method we can express the Swampland conjectures solely via the cosmographic parameters. Finally, in order to test and search for deviations from GR we use our ML reconstructions to analyze two phase diagrams, $H-f\sigma_8$ and $\eta-f\sigma_8$, where $\eta$ is the anisotropic stress parameter. This conjoined diagrams have the asset of helping to break degeneracies between observations that are geometrical against those that come from gravitational effects and makes clearer even visually which redshift ranges should be the target of future surveys to discriminate among the plethora of DE and MG models. This approach has been used for different comparison of models, see for example Refs.~\cite{Sagredo:2018rvc,Moresco:2017hwt,Basilakos:2017rgc,Linder:2016xer,Matsumoto:2020kdu}. Our paper is organized as follows: In Sec.~\ref{sec:theory} we present the theoretical framework including the quintessence reconstruction and the cosmographic expansion. In Sec.~\ref{sec:data} we describe the data used in our analysis and in Sec.~\ref{sec:GA} we outline our ML method, the Genetic Algorithms (GA). Then in Sec.~\ref{sec:results} we set out our results and in Sec.~\ref{sec:diagrams} we provide two phase diagrams derived through our ML reconstructions. Finally in Sec.~\ref{sec:conclusions} we present our conclusions. \section{Theory \label{sec:theory}} Here we present some theoretical aspects of our analysis related to the reconstruction of quintessence and the cosmographic expansion. \subsection{Quintessence reconstruction} At late times, the Friedmann equations including quintessence can be written as \begin{eqnarray} H^{2} &=&\frac{8\pi G}{3} \left(\rho_{m}+\frac{1}{2} \dot{\phi}^{2}+V(\phi)\right), \\ \dot{H} &=&-4 \pi G\left(\rho_{m}+\dot{\phi}^{2}\right), \end{eqnarray} where $H\equiv\frac{\dot{a}}{a}$, for $a=\frac1{1+z}$ and after setting $x\equiv1+z$ they can be solve for the potential and the kinetic terms and be rewritten as \cite{Sahni:2006pa} \begin{eqnarray} \frac{8 \pi G}{3 H_{0}^{2}} V(x) &=&\frac{H(x)^{2}}{H_{0}^{2}}-\frac{x}{6 H_{0}^{2}} \frac{d (H(x)^{2})}{d x}-\frac{1}{2} \Omega_{\mathrm{m},0} x^{3}\label{eq:Vx0},~~ \\ \frac{8 \pi G}{3 H_{0}^{2}}\left(\frac{d \phi}{d x}\right)^{2} &=&\frac{2}{3 H_{0}^{2} x} \frac{d \ln H}{d x}-\frac{\Omega_{\mathrm{m},0} x}{H^{2}}\label{eq:phix0}. \end{eqnarray} It is more convenient to rescale all variables and use dimensionless quantities, which can be done for example by introducing the Planck mass $M_\textrm{pl}\equiv\sqrt{\frac{\hbar c}{8 \pi G}}=\sqrt{\frac{1}{8 \pi G}}$ in natural units ($\hbar=c=1$) and the fact that the critical density is $\rho_c=\frac{3H_0^2}{8\pi G}$. Then we can make the redefinitions \begin{eqnarray} E(z)&\equiv&H(z)/H_0, \nonumber \\ \tilde{\phi}(z)&\equiv&\frac{\phi(z)}{\sqrt{3}M_\textrm{pl}},\nonumber \\ \tilde{V}(z)&\equiv&\frac{V(z)}{\rho_c}, \end{eqnarray} and rewrite the reconstruction equations for the scalar field as \begin{eqnarray} \tilde{V}(x) &=&E(x)^2-\frac{x}{6} \frac{d (E(x)^{2})}{d x}-\frac{1}{2} \Omega_{\mathrm{m},0} x^{3}\label{eq:Vx}, \\ \left(\frac{d \tilde{\phi}}{d x}\right)^{2} &=&\frac{2}{3x} \frac{d \ln E}{d x}-\frac{\Omega_{\mathrm{m},0} x}{E(x)^{2}}\label{eq:phix}. \end{eqnarray} To reconstruct the potential we then integrate Eq.~(\ref{eq:phix}) to determine $\tilde{\phi}(x)$ up to a constant, then we write $x$ as a function of $\tilde{\phi}$ i.e $x(\tilde{\phi})$ and insert it in Eq.~(\ref{eq:Vx}) to find the potential in terms in the scalar field $\tilde{V}(\tilde{\phi})$. For the ML approach the function $H(x)$ and the parameters $H_0$ and $\Omega_{\mathrm{m},0}$ will be given by the GA fits to the data, as described in Section \ref{sec:data}, while in the case of cosmography we will determine the function $H(x)$ from the cosmographic reconstruction and we will assume a Planck 2018 prior on $\Omega_{\mathrm{m},0}$. Note that using the aforementioned equations one may try to reconstruct any DE model, e.g. the constant equation of state $w$=const model or other parameterized $w(z)$ models \cite{Scherrer:2015tra}. \subsection{Cosmography \label{sec:cosmo}} Cosmography is a model independent series expansion in terms of the redshift $z$ that relates the cosmological quantities, such as the Hubble parameter and luminosity distance, to a set of cosmographic coefficients defined as the nth derivative of the scale factor \cite{Visser:2004bf,Capozziello:2008qc,Capozziello:2019cav,Aviles:2012ay}: \begin{eqnarray} H & \equiv & \frac{1}{a} \frac{d a}{d t}, \hspace{12mm} q \equiv -\frac{1}{a H^{2}} \frac{d^{2} a}{d t^{2}}, \\ j & \equiv & \frac{1}{a H^{3}} \frac{d^{3} a}{d t^{3}}, \hspace{5.5mm} s \equiv \frac{1}{a H^{4}} \frac{d^{4} a}{d t^{4}}, \\ l & \equiv & \frac{1}{a H^{5}} \frac{d^{5} a}{d t^{5}}, \hspace{4mm} m \equiv \frac{1}{a H^{6}} \frac{d^{6} a}{d t^{6}}. \end{eqnarray} With simple algebra we can relate these quantities, evaluated today, i.e. at $z=0$, to the series expansions of the Hubble parameter and the luminosity distance. For example, following Ref.~\cite{Visser:2004bf} we find that the luminosity distance in a flat Universe $(\Omega_k=0)$ can be written up to fifth order in redshift as: \begin{eqnarray} d_L(z)&=& \frac{c}{H_0}\Big[z + \frac12 (1-q_0) z^2 + \frac16 (-1 - j_0 + q_0 + 3 q_0^2) z^3 \nonumber \\ &+& \frac{1}{24} (2 + 5 j_0 (1 + 2 q_0) - q_0 (2 + 15 q_0 (1 + q_0)) + s_0) z^4 \nonumber \\ &+& \frac{1}{120}\big((-6+10 j_0^2-l_0-j_0(27+5q_0(22+21q_0))\nonumber\\ &+&3 q_0(2+q_0(27+5q_0(11+7q_0))-5s_0)-11s_0)\big)z^5\nonumber \\ &+& \mathcal{O}(z^6)\Big],\label{eq:cosmodL} \end{eqnarray} while by inverting the equation of the luminosity distance that relates it to the Hubble parameter for a flat Universe, i.e. $d_L(z)=\frac{c}{H_0}(1+z)\int_0^z\frac{1}{H(u)/H_0}du$, and solving for $H(z)$ we find \begin{eqnarray} H(z)/H_0&=& 1 + (1 + q_0) z + \frac12 (j_0 - q_0^2) z^2 \nonumber \\ &+&\frac16 \left(3 q_0^2 (1 + q_0) - j_0 (3 + 4 q_0) - s_0\right) z^3 \nonumber \\ &+& \frac{1}{24} \big(-4j_0^2+l_0-3 q_0^2 (4+q_0 (8+5 q_0))\nonumber \\ &+&j_0 (12+q_0 (32+25 q_0))+(8+7 q_0) s_0\big) z^4\nonumber \\ &+&\mathcal{O}(z^5).\label{eq:cosmoH} \end{eqnarray} Note that going from the luminosity distance given by Eq.~\eqref{eq:cosmodL} to the Hubble parameter given by Eq.~\eqref{eq:cosmoH}, implies the use of differentiation and the presence of a term $1+z$, both of which reduce the order of the polynomial from fifth order to only fourth. This reduction of the polynomial will also be observed later on, when we derive the potential as a function of the scalar field and the cosmographic parameters. Furthermore, it should be noted that there is an issue related to the convergence of the truncation order of the cosmographic series and the redshift range of the data. In Refs.~\cite{Cattoen:2007id,Lazkoz:2013ija,Guimaraes:2010mw} it has been suggested that the variable $y=\frac{z}{1+z}$ avoids the aforementioned convergence issues and is more suitable for parameterizing cosmological distances, where now $y$ lays in the redshift interval $\left[0,1\right]$ which encloses the range of all possible observations. Using Eq.~\eqref{eq:cosmoH} we can now use the quintessence reconstruction set of equations given by Eqs.~\eqref{eq:Vx}-\eqref{eq:phix} to relate the cosmographic parameters to the potential $V(\phi)$, which after some simple algebra can be written in terms of the redshift $z$ as \begin{eqnarray} \tilde{V}(z)&=&\frac{1}{6}\left(4-2q_0-3\Omega_{\mathrm{m}, 0}\right)+\frac{1}{6}\left(8-2j_0+6q_0-9\Omega_{\mathrm{m},0}\right)z\nonumber\\ &+&\frac{1}{6}\left(4+8q_0+j_0(4+q_0)+s_0-9\Omega_{\mathrm{m},0}\right)z^2\nonumber\\ &+&\frac{1}{18}(j_0^2-l_0-9\Omega_{\mathrm{m},0}-j_0q_0(7+3q_0)-7s_0-3q_0s_0)z^3\nonumber \\ &+&\mathcal{O}(z^4), \end{eqnarray} and the derivative of the scalar field $\left(\frac{d\tilde{\phi}'}{dz}\right)^2$ as \begin{eqnarray} \tilde{\phi}'(z)^2&=&\frac{2(1+q_0)}{3}-\Omega_{\mathrm{m},0}+\frac{1}{3}\big(-4+2j_0-4q_0^2 \nonumber\\ &+& 6q_0(\Omega_{\mathrm{m},0}-1)+3\Omega_{\mathrm{m},0}\big)z +\frac{1}{3}\big(6-s_0-3\Omega_{\mathrm{m},0} \nonumber\\ &+& j_0(-8-7q_0+3\Omega_{\mathrm{m},0})+4q_0(3+4q_0+2q_0^2 \nonumber\\ &-& 3(1+q_0)\Omega_{\mathrm{m},0})\big)z^2+\mathcal{O}(z^3). \end{eqnarray} Solving for $\tilde{\phi}$ by integrating the kinetic term over the redshift, will give two branches as \begin{equation} \tilde{\phi}(z)=\tilde{\phi}_0+\epsilon \int_0^z\sqrt{\tilde{\phi}'(u)^2}du, \end{equation} where $\epsilon=\pm1$. We can then express the potential in terms of the cosmographic parameters and the scalar field $\tilde{\phi}$ as \begin{eqnarray} \tilde{V}(\tilde{\phi})&=&\tilde{V}_0+\tilde{V}_1(\tilde{\phi}-\tilde{\phi}_0)+\tilde{V}_2(\tilde{\phi}-\tilde{\phi}_0)^2+\tilde{V}_3(\tilde{\phi}-\tilde{\phi}_0)^3\nonumber\\ &+&\mathcal{O}(\tilde{\phi}^4),\label{eq:Vphi} \end{eqnarray} where we have set {\small \begin{eqnarray} \tilde{V}_0&=& \frac{1}{6}\left(4-2q_0-3\Omega_{\mathrm{m},0}\right),\\ \tilde{V}_1&=& \epsilon^{-1}\frac{8-2j_0+6q_0-9\Omega_{\mathrm{m}, 0}}{2\sqrt{6+6q_0-9\Omega_{\mathrm{m},0}}},\\ \tilde{V}_2&=& \frac{1}{8}\left(15+6q_0+\frac{4(j_0-1)^2}{\left(2+2q_0-3\Omega_{\mathrm{m}, 0}\right)^2}+\frac{4(2j_0+3q_0+s_0)}{2+2q_0-3\Omega_{\mathrm{m},0}}\right),\nonumber \\ \tilde{V}_3&=&\cdots, \end{eqnarray} }where we do not show the term $\tilde{V}_3$ as it is too long and complicated, but can be easily derived from the previous expressions. We can now also calculate the effective mass of the scalar field as: \begin{eqnarray} m_{\phi}^2&=&\frac{d^2 V}{d\phi^2} \nonumber \\ &=&\frac{d^2 \tilde{V}}{d\phi^2} H_0^2 \nonumber \\ &=&2\tilde{V}_2 + 6 \tilde{V}_3 (\tilde{\phi}-\tilde{\phi}_0)+\cdots, \end{eqnarray} where the coefficients $V_2$ and $V_3$ were given earlier. Then, the second Swampland conjecture (SC2) can be written as in terms of the cosmographic parameters as \begin{eqnarray} M_\textrm{pl}\frac{|V'(\phi)|}{V}&=& \frac{\tilde{V}'(\tilde{\phi})}{\sqrt{3} \tilde{V}(\tilde{\phi})}\nonumber \\ &=&S_0 + S_1 (\tilde{\phi}-\tilde{\phi}_0)+S_2 (\tilde{\phi}-\tilde{\phi}_0)^2+\cdots,~~~~ \end{eqnarray} where the coefficients $S_0$, $S_1$ and $S_2$ are given by \begin{eqnarray} S_0 &=& \frac{\tilde{V}_1}{\sqrt{3}\tilde{V}_0}, \\ S_1 &=& -\frac{\tilde{V}_1^2-2 \tilde{V}_0 \tilde{V}_2}{\sqrt{3}\tilde{V}_0^2},\\ S_2 &=& -\frac{\tilde{V}_1^3-3\tilde{V}_0\tilde{V}_1\tilde{V}_2+3 \tilde{V}_0^2\tilde{V}_3}{\sqrt{3}\tilde{V}_0^3}. \end{eqnarray} Note that in the case of the cosmological constant model, we have that $\tilde{\phi}'(z)=0$ and $\tilde{\phi}(z)=\tilde{\phi}_0$, which implies that $\tilde{V}(\tilde{\phi})=\tilde{V}_0=$constant, hence that $m_\phi^2=0$ and $S_0=0$ as expected. Note also that one has to take the limit to $w\rightarrow-1$ before differentiating, as the limit and the derivatives do not commute in this case. We will present the results from the fits to the data and the cosmographic reconstructions of the conjectures in Sec.\ref{sec:results}. \section{Data \label{sec:data}} \subsection{The $H(z)$ data} The Hubble expansion data is obtained by two approaches that are interdependent. First, through the clustering of galaxies or quasars, which represents a direct probe of the Hubble expansion by finding out the BAO peak in the radial direction \cite{Gaztanaga:2008xz}. Second, by the differential age method, which is connected to the redshift drift of distant objects over long time periods, normally more than a decade. This is due to the fact that in metric theories, under the assumption of the Friedmann-Robertson-Walker metric, the Hubble parameter can also be written in terms of the time derivative of the redshift as $H(z)=-\frac{1}{1+z}\frac{dz}{dt}$ \cite{Jimenez:2001gg}. In our analysis we use the $36$ points of the compilation from Ref.~\cite{Arjona:2018jhh}, where the redshift ranges from $0.07\le z \le 2.34$ as can be seen in Table~\ref{tab:Hzdata}. By minimizing the $\chi^2$ analytically over $H_0$ we find \begin{eqnarray} \chi^2_\textrm{H}&=&A-\frac{B^2}{\Gamma},\label{eq:chi2H}\\ H_0&=&\frac{B}{\Gamma},\label{eq:H0bf} \end{eqnarray} where the parameters $A$, $B$ and $\Gamma$ are defined as \begin{eqnarray} A&=&\sum_i^{N_\textrm{H}}\left(\frac{H_i}{\sigma_{H_i}}\right)^2, \\ B&=&\sum_i^{N_\textrm{H}}\frac{H_i~E^\textrm{th}(z_i)}{\sigma_{H_i}^2}, \\ \Gamma&=&\sum_i^{N_\textrm{H}}\left(\frac{E^\textrm{th}(z_i)}{\sigma_{H_i}}\right)^2, \end{eqnarray} and we designate the theoretical value of the Hubble parameter as $E^\textrm{th}(z)=H^\textrm{th}(z)/H_0$ and $N_\textrm{H}=36$. This compilation may be used to measure different cosmological parameters such as the Hubble constant $H_0$, the transition redshift $z_t$, the curvature parameter $\Omega_k$ along with distance redshift data and also constrain the non-relativistic matter and DE parameters, as shown in Ref.~\cite{Yu:2017iju}. \begin{table}[!t] \caption{The $H(z)$ compilation used in our analysis (in units of $\textrm{km}~\textrm{s}^{-1} \textrm{Mpc}^{-1}$). This data, presented in Ref.~\cite{Arjona:2018jhh}, is partly based on those of Refs.~\cite{Moresco:2016mzx} and \cite{Guo:2015gpa}.\label{tab:Hzdata}} \centering \begin{tabular}{cccccccccc} \\ \hline\hline $z$ & $H(z)$ & $\sigma_{H}$ & Ref. \\ \hline $0.07$ & $69.0$ & $19.6$ & \cite{Zhang:2012mp} \\ $0.09$ & $69.0$ & $12.0$ & \cite{STERN:2009EP} \\ $0.12$ & $68.6$ & $26.2$ & \cite{Zhang:2012mp} \\ $0.17$ & $83.0$ & $8.0$ & \cite{STERN:2009EP} \\ $0.179$ & $75.0$ & $4.0$ & \cite{MORESCO:2012JH} \\ $0.199$ & $75.0$ & $5.0$ & \cite{MORESCO:2012JH} \\ $0.2$ & $72.9$ & $29.6$ & \cite{Zhang:2012mp} \\ $0.27$ & $77.0$ & $14.0$ & \cite{STERN:2009EP} \\ $0.28$ & $88.8$ & $36.6$ & \cite{Zhang:2012mp} \\ $0.35$ & $82.7$ & $8.4$ & \cite{Chuang:2012qt} \\ $0.352$ & $83.0$ & $14.0$ & \cite{MORESCO:2012JH} \\ $0.3802$ & $83.0$ & $13.5$ & \cite{Moresco:2016mzx} \\ $0.4$ & $95.0$ & $17.0$ & \cite{STERN:2009EP} \\ $0.4004$ & $77.0$ & $10.2$ & \cite{Moresco:2016mzx} \\ $0.4247$ & $87.1$ & $11.2$ & \cite{Moresco:2016mzx} \\ $0.44$ & $82.6$ & $7.8$ & \cite{Blake:2012pj} \\ $0.44497$ & $92.8$ & $12.9$ & \cite{Moresco:2016mzx} \\ $0.4783$ & $80.9$ & $9.0$ & \cite{Moresco:2016mzx} \\ \hline\hline \end{tabular}~~~~~~~~ \begin{tabular}{cccccccccc} \\ \hline\hline $z$ & $H(z)$ & $\sigma_{H}$ & Ref. \\ \hline $0.48$ & $97.0$ & $62.0$ & \cite{STERN:2009EP} \\ $0.57$ & $96.8$ & $3.4$ & \cite{Anderson:2013zyy} \\ $0.593$ & $104.0$ & $13.0$ & \cite{MORESCO:2012JH} \\ $0.60$ & $87.9$ & $6.1$ & \cite{Blake:2012pj} \\ $0.68$ & $92.0$ & $8.0$ & \cite{MORESCO:2012JH} \\ $0.73$ & $97.3$ & $7.0$ & \cite{Blake:2012pj} \\ $0.781$ & $105.0$ & $12.0$ & \cite{MORESCO:2012JH} \\ $0.875$ & $125.0$ & $17.0$ & \cite{MORESCO:2012JH} \\ $0.88$ & $90.0$ & $40.0$ & \cite{STERN:2009EP} \\ $0.9$ & $117.0$ & $23.0$ & \cite{STERN:2009EP} \\ $1.037$ & $154.0$ & $20.0$ & \cite{MORESCO:2012JH} \\ $1.3$ & $168.0$ & $17.0$ & \cite{STERN:2009EP} \\ $1.363$ & $160.0$ & $33.6$ & \cite{Moresco:2015cya} \\ $1.43$ & $177.0$ & $18.0$ & \cite{STERN:2009EP} \\ $1.53$ & $140.0$ & $14.0$ & \cite{STERN:2009EP} \\ $1.75$ & $202.0$ & $40.0$ & \cite{STERN:2009EP} \\ $1.965$ & $186.5$ & $50.4$ & \cite{Moresco:2015cya} \\ $2.34$ & $222.0$ & $7.0$ & \cite{Delubac:2014aqe} \\ \hline\hline \end{tabular} \end{table} \subsection{The growth-rate data} We now present the growth-rate data $f\sigma_8$ that can be used to constrain the matter density parameter $\Omega_{\mathrm{m},0}$ in a model independent fashion. The data based on the compilation provided in Table I of Ref.~\cite{Sagredo:2018ahx}, where the authors tested the internal robustness of the dataset by analyzing different subsets in the data and using Bayesian model comparison. This compilation is derived through the redshift-space distortions where it is driven the combination $f\sigma_8(a)\equiv f(a)\cdot \sigma(a)$. The value of $f\sigma_8(a)$ can be obtained directly from the ratio of the monopole to the quadrupole of the redshift-space power spectrum, which relies on the parameter $\beta=f/b_0$, where $b_0$ is the bias and $f$ is the growth rate assuming linear theory \cite{Percival:2008sh,Song:2008qt,Nesseris:2006er}. In Ref.~\cite{Sagredo:2018ahx} it was shown that $f\sigma_8(a)$ is independent of the bias, since the latter cancels out from the previous expression. The advantage of using $f\sigma_8(a)$ rather than the growth-rate $f(z)$ is that $f\sigma_8(a)$ is directly connected to the power spectrum of peculiar velocities of galaxies \cite{Nesseris:2014mea} and also it has been proven to be a good discriminator of DE models \cite{Song:2008qt}. For further details on the covariance matrix of the data and how to properly account for the Alcock-Paczynski effect, see for example Refs.~\cite{Sagredo:2018ahx}, \cite{Nesseris:2017vor} and \cite{Kazantzidis:2018rnb}. A publicly available RSD likelihood for MontePython based on the aforementioned growth-rate data compilation was initially presented in Ref.~\cite{Arjona:2020yum} and has been recently used to constrain MG models \cite{Cardona:2020ama}. Furthermore, it was shown in Ref.~\cite{Arjona:2020kco} that the matter density parameter $\Omega_{\mathrm{m},0}$ and $\sigma_8$ can be inferred in a model independent fashion via the reconstruction of $f\sigma_8$ using the following expressions \begin{eqnarray}\label{eq:matter} \Omega_{\mathrm{m},0}&=&\frac{1}{3\int_0^1 dx \frac{f\sigma_8{}(x)}{f\sigma_8{}(1)} \int_0^x dy \frac1{y} \frac{f\sigma_8{}(y)}{f\sigma_8{}(1)}} \\ \sigma_8&=&\int_0^1 \frac{f\sigma_8(x)}{x} dx \label{eq:sigma8}, \end{eqnarray} which only involve integrations over the $f\sigma_8$ reconstructions and are parameter free. Note that Eqs.~(\ref{eq:matter}) and (\ref{eq:sigma8}) can be found by direct manipulations on the definition of $f\sigma_{8}$ which is defined as \begin{eqnarray} f \sigma_{8}(a) & \equiv & f(a) \cdot \sigma(a) \nonumber\\ &=&\frac{\sigma_{8}}{\delta_{m}(1)} a \delta_{m}^{\prime}(a), \end{eqnarray} and through the differential equation that is satisfied by the growth factor $ \delta_{m}(a)$ \begin{equation}\label{eq:fs8} \delta_{m}^{\prime \prime}(a)+\left(\frac{3}{a}+\frac{H^{\prime}(a)}{H(a)}\right) \delta_{m}^{\prime}(a)-\frac{3}{2} \frac{\Omega_{\mathrm{m}, 0}}{a^{5} H(a)^{2} / H_{0}^{2}} \delta_{m}(a)=0, \end{equation} where $f(a)$ is the growth rate and $\sigma(a)=\sigma_{8} \frac{\delta_{m}(a)}{\delta_{m}(1)}$ and Eq.~(\ref{eq:fs8}) equation holds assuming a homogeneous and isotropic universe with no dark energy perturbations and neglecting neutrinos. \subsection{The $E_\textrm{g}$ data} The perturbed flat Friedmann-Lemaitre-Robertson-Walker (FLRW) metric, in the conformal Newtonian gauge, is defined as \begin{equation} ds^2=-(1+2\Psi)dt^2 +a(t)^2(1-2\Phi)dx^2, \end{equation} where the terms $\Psi$ and $\Phi$ represent the two scalar gravitational potentials and $a$ is the scale factor. The ratio of the gravitational potentials define what is known as the gravitational slip $\eta=\frac{\Phi}{\Psi}$ which has to be unity in GR. These potentials must obey the two Poisson equations in Fourier space: \begin{eqnarray} -\frac{k^2}{a^2}(\Phi+\Psi)&=&4 \pi G_\textrm{N} \Sigma(k,a) \rho_m \delta_m, \\ -\frac{k^2}{a^2}\Psi&=&4 \pi G_\textrm{N} \mu(k,a) \rho_m \delta_m, \end{eqnarray} where $G_\textrm{N}$ is the bare Newton's constant and $\Sigma$ and $\mu$ parameterize deviations in GR, which is recovered for $\Sigma=2$ and $\mu=1$. Recently, the $E_\textrm{g}$ statistic was used to test the aforementioned relations, with the hope of being model independent at the linear order \cite{Zhang:2007nk,Reyes:2010tr}. This $E_\textrm{g}$ test can also be written as the expectation value of the ratio of lensing and galaxy clustering observables at a scale $k$ as follows \begin{equation} E_\textrm{g}=\left\langle \frac{a \nabla^2(\Psi+\Phi)}{3 H_0^2 f \delta_m}\right\rangle. \end{equation} To perform a model independent reconstruction on the gravitational slip $\eta$ we reconstruct two quantities through the $E_\textrm{g}$ and $f\sigma_8$ data. First, the quantity $P_2(z)$ defined as $P_2=\frac{\Omega_\textrm{m,0}\Sigma}{f}$ which depends on the lensing potential and the growth rate. In GR this simplifies to $P_2=\frac{2\Omega_\textrm{m,0}}{f}$, which for GR we then have $E_\textrm{g}=\frac{\Omega_\textrm{m,0}}{f}$. In general, $E_\textrm{g}$ can be connected to the $P_2$ statistic of Ref.~\cite{Pinho:2018unz} as $P_2=2 E_\textrm{g}$. Second, the quantity $P_3$ defined as $P_3=\frac{\left(f\sigma_8(z)\right)'}{f\sigma_8(z)}$, where the prime is the derivative with respect to $\ln a$. Then $\eta$ can be defined as \cite{Pinho:2018unz} \begin{equation} \eta(z)=\frac{3P_2(z)(1+z)^3}{2E(z)^2\left(P_3(z)+2+\frac{E'(z)}{E(z)}\right)}-1, \label{eq:etade} \end{equation} where $E(z)\equiv H(z)/H_0$. For completeness the data points used in our analysis are given in Table~\ref{tab:Eg}. \begin{table}[!t] \caption{The $E_\textrm{g}$ data used in this analysis as compiled by Refs.~\cite{Pinho:2018unz} and \cite{Skara:2019usd}. Notice that some of the points in the references mentioned before were duplicates as they come from the same surveys, even though with combinations of different external probes, so we use only one of the measurements to avoid strong correlations. Here we only display the points we used in the analysis.\label{tab:Eg}} \begin{centering} \begin{tabular}{cccc} $z$ & $E_\textrm{g}$ & $\sigma_{E_\textrm{g}}$ & \\ \hline 0.267 & 0.43 & 0.13 & \\ 0.270 & 0.40 & 0.05 & \\ 0.305 & 0.27 & 0.08 & \\ 0.320 & 0.40 & 0.09 & \\ 0.554 & 0.26 & 0.07 & \\ 0.570 & 0.30 & 0.07 & \\ 0.600 & 0.16 & 0.09 & \\ 0.860 & 0.09 & 0.07 & \end{tabular} \par \end{centering} \end{table} \section{Genetic Algorithms \label{sec:GA}} The Genetic Algorithms (GA) are a group of machine learning (ML) techniques that are designed to perform non-parametric reconstruction of data and are constructed on the concept of grammatical evolution, conveyed by the genetic operations of crossover and mutation. For an in-depth discussion on the GA and several applications to cosmology see Refs.~\cite{Bogdanos:2009ib,Nesseris:2010ep, Nesseris:2012tt,Nesseris:2013bia,Sapone:2014nna,Arjona:2020doi,Arjona:2020kco,Arjona:2019fwb,Arjona:2020axn}. The GA emulate the idea of evolution by the application of natural selection; a group of individuals evolves over time under the effect of the stochastic operators of mutation, i.e a random modification in an individual, and crossover, i.e. the merger of two or more different individuals, producing two new individuals (children). Every individual's ``reproductive success'', usually quantified through a $\chi^2$ statistic, is taken to be proportional to its fitness, measuring how precisely each individual of the population fits the data. In what follows, we will outline how to reconstruct the Hubble parameter $H(z)$ from the Hubble expansion history $H(z)$ data, $f\sigma_8(z)$ from the growth-rate data derived via the redshift-space distortions (RSD) and $P_2(z)$ from the Eg data. The reconstruction of $H(z)$, along with that of $H_0$, is needed for Eqs.~(\ref{eq:Vx}) and (\ref{eq:phix}), while our reconstruction for $f\sigma_8(z)$ is used to infer model independently the value of $\Omega_{\mathrm{m},0}$ by means of Eq.~(\ref{eq:matter}). Our reconstruction of $P_2(z)$ from the Eg data is used at a latter stage in Sec.~\ref{sec:diagrams}. The outline to perform the reconstructions proceeds as follows. An initial population of functions is randomly selected so that every member of the population holds initial guesses for $H(z)$, $f\sigma_8(z)$ and $P_2(z)$. We also impose reasonable physical priors, e.g. the Hubble parameter today is given by the Hubble constant $H(z = 0) = H_0$, which then allows us to estimate $H_0$ directly from the $H(z)$ data. For the $f\sigma_8(z)$ reconstruction we assume that the Universe at early times went through a phase of matter domination $(z\simeq100)$, then the linear growth acts as $\delta_m(a)\simeq a$ at high redshifts. However we make no assumption of a DE model or on the curvature of the Universe. Next, the fitness of each member is computed through a $\chi^2$ statistic, using as input the $H(z)$, growth and Eg data. Afterwards, the mutation and crossover operators are applied stochastically to the best-fitting functions in each generation, selected via the tournament selection method, see for more details \cite{Bogdanos:2009ib}. This process is then repeated with different random seeds thousands of times in order to ensure convergence and not to bias the results due to a specific choice of the random seed. After the GA code has converged, the final output is a reconstruction of $H(z)$, $f\sigma_8(z)$ and $P_2(z)$. For the estimation of the errors on the reconstructed functions we implement an analytical approach developed by Refs.~\cite{Nesseris:2012tt,Nesseris:2013bia}, where the errors are derived via a path integral over the whole functional space that can be scanned by the GA. The GA path integral approach has been compared with bootstrap Monte-Carlo error estimates \cite{Nesseris:2012tt} finding excellent agreement with both approaches. To sum up, with the GA we can reconstruct any cosmological function, for example the $H(z)$, $f\sigma_8(z)$ and $P_2(z)$ that we consider here, by applying the algorithm to any dataset of interest. There are no requirements on the specific cosmological model or assumptions on DE, hence our results are model independent. Besides executing a large number of GA runs with different random seed numbers, in order to avoid spurious reconstructions and overfitting we have imposed that all reconstructed functions, as well as their derivatives, are continuous in the range of redshifts we consider. \section{Results \label{sec:results}} \subsection{Genetic Algorithm reconstructions} In this section we will now discuss our ML fits to the data and the corresponding reconstructions of the Swampland conjectures. First, we show in Table~\ref{tab:GAchi2} the best-fit $\chi^2$ for the GA functions for the $\Lambda$CDM model. As can be seen, in all cases the GA out-performs the $\Lambda$CDM model in terms of the best-fit $\chi^2$. Then, in the left panel of Fig.~\ref{fig:pot_errkin} we present our GA reconstruction of the scalar field potential as a function of the scalar field for the redshift range $z \in [0,1.92]$. We can see a parabolic shape of the potential, thus pointing toward some deviations from the $\Lambda$CDM model since for the latter the potential should be flat. The black solid line corresponds with the GA best-fit and the different colours represent the errors for our reconstructions at different redshifts. In the right panel of Fig.~\ref{fig:pot_errkin} we see that the reconstructed kinetic term is positive from $z=0$ to $z~\sim 1.9$, hence our reconstructions of the Swampland conjectures within that redshift range should be free from ghosts and instabilities in this redshift range. The blue solid line corresponds to the GA best-fit and the grey region to the $1\sigma$ errors. Note that the fact that at some redshift the kinetic term of the quintessence scalar field is negative, is actually a common issue of such reconstruction methods. This happens not only for quintessence models, but even for scalar tensor theories, where one naively may expect that due to the extra degree of freedom this would not happen\footnote{See for example Fig.~5 in Ref.~\cite{Nesseris:2006jc}, where a similar phenomenon is also observed.}. This simply means that at some redshift the reconstruction breaks down, as obviously the kinetic term has to be both real and positive. However, due to the larger errors this is not a big problem per se, it just limits our ability to use this particular model at all redshifts. In Fig.~\ref{fig:pot_err} we present our GA reconstructions of the first and second Swampland conjectures on the left and right panels respectively. In both cases the blue solid line and the grey region corresponds to the GA best-fit and the $1\sigma$ error respectively. As can be seen, from $z=0.8$ onward the errors become so large that we can only draw some conclusions at low redshifts with the current available data. In specific we can see how both reconstructions are consistent with the conjectures, being both order unity, although at the same time being consistent with the $\Lambda$CDM model at the $1\sigma$ level. \begin{table}[] \centering \begin{tabular}{cccc} Model & $H(z)$ & $f \sigma_{8}$ & $P_{2}$ \\ \hline$\chi_{\Lambda \mathrm{CDM}}^{2}$ & 19.476 & 12.238 & 10.516 \\ \hline$\chi_{G A}^{2}$ & 17.670 & 12.220 & 5.422 \\ \hline \end{tabular} \caption{The $\chi^2$ for $\Lambda$CDM and GA using the growth $f\sigma_8$, the Hubble rate $H(z)$ and the $E_g$ statistics.} \label{tab:GAchi2} \end{table} \begin{figure*}[!t] \centering \hspace*{-4mm} \includegraphics[width = 0.45\textwidth]{V_phi.pdf} \includegraphics[width = 0.48\textwidth]{kinetic.pdf} \caption{The GA reconstruction of the scalar field potential (left) and the kinetic term (right) for the redshift range $z \in [0,1.92]$. The left panel suggests a parabolic shape of the potential, thus pointing toward some deviations from the $\Lambda$CDM model since for the latter the potential should be constant. The black solid line corresponds with the GA best-fit, the different colors represent the errors for our reconstructions at different redshifts and that $\Delta \tilde{\phi}=\tilde{\phi}(z)-\tilde{\phi}_0$. In the right panel we see that the reconstructed kinetic term is positive from $z=0$ to $z~\sim 1.9$, hence our reconstructions of the Swampland conjectures within that redshift range should be free from ghosts and instabilities. The blue solid line corresponds to the GA best-fit and the grey region to the $1\sigma$ errors. Note that in the right panel we truncate the error in the unphysical region where the kinetic term changes sign. \label{fig:pot_errkin}} \vspace*{-1mm} \end{figure*} \begin{figure*}[!t] \centering \hspace*{-4mm} \includegraphics[width = 0.48\textwidth]{SC1.pdf} \includegraphics[width = 0.48\textwidth]{SC2.pdf} \caption{The GA reconstruction of the first and second Swampland conjectures on the left and right panels respectively. As can be seen, from $z=0.8$ onward the errors become so large that we can only draw some conclusions at low redshifts with the current available data. Both reconstructions are consistent with the conjectures and the $\Lambda$CDM model at the $1\sigma$ level. Notice that $\Delta \phi=\phi(z)-\phi_0$ and that we truncate the error in the unphysical region where the quantities would change sign (as we take the absolute values). \label{fig:pot_err}} \vspace*{-1mm} \end{figure*} \subsection{Cosmographic reconstructions} We now present the cosmographic reconstructions of the potentials and the first and second Swampland conjectures. We find that the standard quintessence reconstruction via cosmography suffers from two issues: first, it is only valid when the square of the kinetic term given via Eq.~\eqref{eq:phix0} is positive. Second, a value of the matter density $\Omega_{\mathrm{m},0}$ from an external source is required, as the $H(z)$ cannot provide it since Eq.~\eqref{eq:cosmoH} only depends on the cosmographic parameters and not at all on $\Omega_{\mathrm{m},0}$. We address the first issue by performing the reconstruction only in the redshift range where the kinetic term is positive, while for the value of the matter density parameter we will assume in what follows the Planck prior $\Omega_m=0.315 \pm0.007$ \cite{Aghanim:2018eyx}. \begin{table*}[] \centering \begin{tabular}{c|cccccccc} Order/param. & $q_0$ & $j_0$ & $s_0$ & $l_0$ & $h$ & $\chi^2$ & $|\Delta \textrm{AIC}_c|$& $|\Delta \textrm{BIC}|$\\\hline 2nd & $0.144\pm0.093$ &$-$ & $-$ & $-$ & $0.586\pm0.023$ & $24.772$ & $5.302$& $5.302$\\ 3rd & $-0.347\pm0.195$ &$0.432\pm0.247$ & $-$ & $-$ & $0.662\pm0.040$ & $19.172$& $2.088$& $3.286$\\ 4th & $-0.851\pm0.373$ &$2.222\pm1.412$ & $2.326\pm4.029$ & $-$ & $0.718\pm0.057$ & $17.301$& $2.758$& $4.998$\\ 5th & $-0.824\pm0.727$ &$2.133\pm3.951$ & $2.091\pm12.525$ & $9.661\pm34.294$ & $0.714\pm0.075$ & $17.304$& $5.470$& $8.585$\\ \end{tabular} \caption{The best fit parameters of the cosmographic expansions. For comparison the $\Lambda$CDM model has best-fit parameters $(\Omega_{\mathrm{m},0},h)=(0.259\pm0.029,0.704\pm0.023)$ with $\chi^2=19.470$, while the GA has $\chi^2=17.670$. The values of the differences for the AIC and BIC, are given with respect to those of the $\Lambda$CDM model.} \label{tab:cosmoparams} \end{table*} \begin{figure*}[!t] \centering \hspace*{-4mm} \includegraphics[width = 0.49\textwidth]{cosmo_phipz.pdf} \includegraphics[width = 0.48\textwidth]{cosmo_V.pdf} \caption{Left: Reconstruction of the kinetic term of the scalar field for various orders of the cosmographic expansion for the Planck prior $\Omega_m=0.315 \pm0.007$ \cite{Aghanim:2018eyx}. As can be seen the third order expansion is positive in the range $z\in[0,0.3]$, while in the other cases the reconstruction breaks down as the field is complex. Right: the third order cosmographic reconstruction on the scalar field potential, where $\Delta \tilde{\phi}=\tilde{\phi}(z)-\tilde{\phi}_0$. We only plot the quantities in the regions where the scalar field kinetic term is positive, which in this case is in the range $z\in[0,0.3]$. As can be seen, in the aforementioned redshift range the cosmographic and the GA reconstructions are consistent.\label{fig:pot_errcosm0}} \vspace*{-1mm} \end{figure*} \begin{figure*}[!t] \centering \hspace*{-4mm} \includegraphics[width = 0.49\textwidth]{cosmo_SC1.pdf} \includegraphics[width = 0.49\textwidth]{cosmo_SC2.pdf} \caption{The third order cosmographic reconstruction on the scalar field potential (left) and the Swampland conjectures SC1 (center) and SC2 (right), where $\Delta \tilde{\phi}=\tilde{\phi}(z)-\tilde{\phi}_0$. We only plot the quantities in the regions where the scalar field kinetic term is positive, which in this case is in the range $z\in[0,0.3]$. As can be seen, in the aforementioned redshift range the cosmographic and the GA reconstructions of the Swampland conjectures are consistent. Note that we truncate the error in the unphysical region where the quantities would change sign (as we take the absolute values).\label{fig:pot_errcosm}} \vspace*{-1mm} \end{figure*} Next, in Table~\ref{tab:cosmoparams} we show the best-fit parameters of the cosmographic parameters for the $H(z)$ data and their $\chi^2$ values for various orders of the series expansion. We also give the values of the corrected AIC and BIC tools, which are defined as \cite{Liddle:2007fy}: \begin{eqnarray} \textrm{AIC}&=& \chi^2_\textrm{min}+2k,\\ \textrm{AIC}_c&=& \textrm{AIC}+\frac{2k(k+1)}{N-k-1},\\ \textrm{BIC} &=& \chi^2_\textrm{min}+k \ln N, \end{eqnarray} where $k$ is the number of free parameters and $N$ the number of data points, which for the Hubble data is equal to $N=36$, while $\chi^2_\textrm{min}$ is the value of the $\chi^2$ at the minimum. In Table~\ref{tab:cosmoparams} the values of the differences of the AIC${}_c$ and BIC are given with respect to those of the $\Lambda$CDM model. In either case, the condition $\Delta \textrm{AIC}_c\le 2$ implies the consistency between the models, while the inequalities $2<\Delta \textrm{AIC}_c\le4$ and $4<\Delta \textrm{AIC}_c\le 7$ indicate a mild and positive evidence against the model with higher value of AIC. On the other hand, when $\Delta \textrm{AIC}_c \ge 10$ suggests strong evidence. The same applies to the BIC test \cite{Nesseris:2012cq}. As can be seen, in all cases the $\Lambda$CDM model seems to be favored by the data, which is also consistent with previous analyses \cite{Basilakos:2018arq}. Note that we cannot apply either of the criteria to the GA as it is a non-parametric approach. Going back to the scalar field, the aforementioned issue with the kinetic term is also shown in Fig.~\ref{fig:pot_errcosm0} where in the left panel we show a reconstruction of the kinetic term of the scalar field for various orders of the cosmographic expansion for the Planck prior $\Omega_m=0.315 \pm0.007$ \cite{Aghanim:2018eyx}. As can be seen the third order expansion is positive in the range $z\in[0,0.3]$, while in the other cases the reconstruction breaks down as the field is complex. Thus, out of all cases only the second and third order expansions have a positive kinetic term at small redshifts and from these two, only the third order expansion has non-trivial phenomenology, hence we focus on this in what follows. In the right panel of Fig.~\ref{fig:pot_errcosm0} we show the third order cosmographic reconstruction on the scalar field potential. We note that we only plot the quantities in the regions where the scalar field kinetic term is positive, which in this case is in the range $z\in[0,0.3]$. As can be seen, in the aforementioned redshift range the cosmographic and the GA reconstructions are consistent. Finally, we show the results for the Swampland conjecture reconstructions in Fig.~\ref{fig:pot_errcosm}. As can be seen both cases are in good agreement in the range $z\in [0,0.3]$ within the errors with the corresponding GA reconstructions. However, given that the reconstructions are limited in range, this also significantly reduces their appeal. \section{Phase Diagrams \label{sec:diagrams}} In MG theories it is frequently assumed that the background level behaves as the $\Lambda$CDM model and its perturbations evolve differently, see for example Ref.~\cite{Troster:2020kai}. This happens in well known DE and MG models such as the so-called designer $f(R)$ models \cite{Multamaki:2005zs,delaCruzDombriz:2006fj,Pogosian:2007sw,Nesseris:2013fca,Arjona:2018jhh} or the designer Horndeski family of models (HDES) \cite{Arjona:2019rfn} whose background is exactly that of the $\Lambda$CDM model. However, in general, MG models can have departures from $\Lambda$CDM both at the background and at the perturbation level, as is for example the case for the Hu-Sawicki $f(R)$ gravity model \cite{Hu:2007nk} or the Kinetic Gravity Braiding (KGB) theory \cite{Kimura:2010di}. In this section we search for potential deviations from $\Lambda$CDM at the background and at the perturbation level by applying our Machine Learning reconstructions to cosmological observations. In specific we use $H(z)$ expansion rate data, the growth-rate $f\sigma_8(z)$ compilation and the Eg statistics to present two phase diagrams, $H-f\sigma_8$ and $\eta-f\sigma_8$, where $\eta$ is a parameter which defines the departure of gravity from GR and can be explored by gravitational lensing. The anisotropic stress $\eta$ is conventionally modeled through the parameter $\eta=\frac{\Phi}{\Psi}$, where $\Phi$ and $\Psi$ are the Newtonian potentials and are considered equal in GR in the absence of anisotropic stresses from other sources like neutrinos. Then, any departure of $\eta$ from unity would hint to modified gravity or if neglected, it could bias the cosmological parameters inferred from the data \cite{Arjona:2019zqg}. With our GA reconstructions we demonstrate how deviations from the $\Lambda$CDM model appears in the $\eta-f\sigma_8$ diagram. In specific, in the left panel of Fig.~\ref{fig:fs8-H}, we plot the $H-f\sigma_8$ diagram for the redshift range of $z\in \left[0,1.9\right]$ using the $H(z)$ and $f\sigma_8$ data. The theoretical prediction of the $\Lambda$CDM model would correspond to a point at $(1,1)$, something which the GA best-fit reconstructions confirm within the errors. The different colors represent our reconstructions for several redshifts with their respective errors. In the right panel of Fig.~\ref{fig:fs8-H} we have the $\eta-f\sigma_8$ diagram for the redshift range of $z\in \left[0.27,0.86\right]$ using the $E_g$ and $f\sigma_8$ data. The theoretical prediction of the $\Lambda$CDM model would correspond to a point at $(1,1)$. The different colors represent our reconstructions for several redshifts with their respective errors. With our GA best-fit reconstructions we find a $\sim 2\sigma$ deviation of the anisotropic stress from unity at $z\sim 0.3$ and a $\sim 4\sigma$ at $z\sim 0.9$, thus hinting toward some deviations from GR, something which was also seen in Ref.~\cite{Arjona:2020kco}. \begin{figure*}[!t] \centering \hspace*{-4mm} \includegraphics[width = 0.48\textwidth]{fs8_Hz.pdf} \includegraphics[width = 0.48\textwidth]{fs8_eta.pdf} \caption{Left: $H-f\sigma_8$ diagram for the redshift range of $z\in \left[0,1.9\right]$ using the $H(z)$ and $f\sigma_8$ data. The theoretical prediction of the $\Lambda$CDM model would correspond to a point at $(1,1)$, something which the GA best-fit reconstructions confirm within the errors. The different colors represent our reconstructions for several redshifts with their respective errors. Right: $\eta-f\sigma_8$ diagram for the redshift range of $z\in \left[0.27,0.86\right]$ using the $E_g$ and $f\sigma_8$ data. The theoretical prediction of the $\Lambda$CDM model would correspond to a point at $(1,1)$. The different colours represent our reconstructions for several redshifts with their respective errors. With our GA best-fit reconstructions we find a $\sim 2\sigma$ deviation of the anisotropic stress from unity at $z\sim 0.3$ and a $\sim 4\sigma$ at $z\sim 0.9$.\label{fig:fs8-H}} \vspace*{-1mm} \end{figure*} \begin{figure}[!t] \centering \hspace*{-4mm} \includegraphics[width = 0.48\textwidth]{s8_om.pdf} \caption{Plot for $\sigma_8$ against $\Omega_m$ from our ML reconstructions on the growth rate $f\sigma_8$ data (gray point), the Planck 2018 TT,TE,EE+lowE+lensing values (blue point) and the DES Y1 2018 results (orange point). \label{fig:S8tension}} \vspace*{-1mm} \end{figure} These deviations could be due to new physics or for the presence of unaccounted systematics and perhaps non-negligible radiative processes, as could be the case for the $E_g$ data. The diagrams presented will be useful for future large-scale structure observations to rule out some of the modified gravity models. Finally, we also report our results from our ML reconstruction through Eqs.~(\ref{eq:matter}) and (\ref{eq:sigma8}) for the quantity $S_8$, quantified as $S_8=\sigma_8 \sqrt{\Omega_m / 0.3}$. This parameter is in tension between Planck and cosmic shear measurements such as Kids-1000 and DES Y1 above the 2$\sigma$ level, where the latter have a preference for a lower value, see for example Fig.~1 of \cite{DiValentino:2020vvd} where the tension can be visualized in the $\sigma_8 - \Omega_{\mathrm{m},0}$ diagram. With the GA we find $S_8=0.733\pm 0.275$ and in Fig.~\ref{fig:S8tension} we also provide a similar diagram, being our results more in agreement with DES Y1. Overall, these results are consistent with Ref.~\cite{DiValentino:2020vvd} and the analysis of \cite{Muir:2020puy} from the DES Y1 release, where the authors split the matter density into a ``geometric" and ``growth" $\Omega_{\mathrm{m},0}$. They are in mild tension with the Planck 2018 results though, something that has been noted repeatedly in the literature, see for example Ref.~\cite{Nesseris:2017vor} and references there in. Note that the errors with the GA approach are considerably larger than those from DES Y1 and the Planck 2018 results being one of the reasons the fact that with our reconstructions we are not making assumptions on flatness or a dark energy model, hence more agnostic. \section{Conclusions \label{sec:conclusions}} Through a plethora of cosmological probes it has been established that the Universe is expanding in an accelerated way, being one of the biggest findings in modern cosmology and implying that $70\%$ of the mass-energy density is made of an unknown content, usually referred to as dark energy (DE), spreading uniformly over the Universe. The cosmological constant $\Lambda$ represents one of the simplest contenders of DE, however some of it caveats is that the observed energy scale of DE is much smaller than the energy scale predicted by a vacuum energy in particle physics \cite{Weinberg:1988cp}. Other viable alternatives come from the inclusion of slowly rolling scalar fields which control the accelerated expansion of the Universe at late times, being Quintessence an example of these type of models. In the framework of string phenomenology, at present we should necessarily be in an epoch of quintessence and in the continuous search for a theory of quantum gravity there has been some Swampland conjectures proposed, which refer to some criteria that must be hold in order to have effective field theories consistent with quantum gravity \cite{Cicoli:2012tz}. Some of these conjectures have important implications in cosmology as is the case of the Swampland criteria that we have analyzed in this paper. In our analysis we have presented two model independent reconstructions of the Swampland conjectures with ML and cosmography placing constraints on the first and second Swampland conjectures, finding that at low redshifts both approaches give consistent results and that, neither conjecture is ruled out by the data at a statistically significant level. However, while the swampland bounds seem to be satisfied by the low redshift data, their reconstructed scalar potential, see Fig.~\ref{fig:pot_errkin}, has a shape similar to a parabola. Thus, while these reconstructions may be superficially consistent with the swampland bounds, it should be noted that exponential potentials are what naturally emerges in UV complete settings. Comparing both of our approaches, the ML algorithm seems a more robust approach for the following reasons: First, in cosmography we are assuming flatness and a Planck 2018 prior on $\Omega_{\mathrm{m},0}$, while with the GA we do not make assumptions on flatness, $\Omega_{\mathrm{m},0}$ or on a dark energy model. Second, our ML reconstructions allow us to place constraints at higher redshifts where the kinetic term reconstructed goes up to $z\sim 1.9$. while for cosmography only extends up to $z\sim 0.3$, as the scalar field kinetic term becomes negative beyond that. Finally, using the growth rate data $f\sigma_8(z)$ we perform a model independent probe of modified gravity cosmologies through two phase diagrams, $H-f\sigma_8$ and $\eta-f\sigma_8$ where the anisotropic stress parameter $\eta$ is reconstructed through the $E_g$ statistics, related to gravitational lensing data. We see that the first diagram $H-f\sigma_8$ is consistent within the errors with the $\Lambda$CDM model, while the second diagram $\eta-f\sigma_8$ has a $\sim 2\sigma$ deviation of the anisotropic stress from unity at $z\sim 0.3$ and a $\sim 4\sigma$ at $z\sim 0.9$, thus pointing toward mild deviations from GR, which could be the future target of upcoming large-scale structure surveys. \section*{Acknowledgements} The authors acknowledge support from the Research Project PGC2018-094773-B-C32 and the Centro de Excelencia Severo Ochoa Program SEV-2016-0597. S.~N. also acknowledges support from the Ram\'{o}n y Cajal program through Grant No. RYC-2014-15843.\\ \textbf{Numerical Analysis Files}: The Genetic Algorithm codes used by the authors in the analysis of the paper can be found at \href{https://github.com/snesseris}{https://github.com/snesseris} and \href{https://github.com/RubenArjona}{https://github.com/RubenArjona}.\\ \begin{appendix} \section{Error analysis \label{sec:errors}} In Eqs.~(\ref{eq:Vx}) and (\ref{eq:phix}) we have rewritten the reconstruction equations for the scalar field in terms of dimensionless variables. Then, by standard error propagation we find that the error for the potential and the kinetic term can be expressed as \begin{eqnarray} \delta \tilde{V}&=&2E\delta E-\frac{x}{6}\frac{d}{dx}\left(2E\delta E\right)-\frac{1}{2}\delta \Omega_{\mathrm{m},0}x^3,\\ \delta \left(\frac{d \tilde{\phi}}{d x}\right)^{2} &=&\frac{2}{3x}\frac{d}{dx}\Big(\frac{\delta E}{E}\Big)+\frac{2\Omega_{\mathrm{m},0}x\delta E}{E^3}-\frac{x\delta \Omega_{\mathrm{m},0}}{E^2}.~~~~ \end{eqnarray} Since the Swampland conjecture 2 (SC2) is defined as \begin{equation} \textrm{SC2}=\frac{\left|V_{,\phi}\right|}{V}\equiv \frac{\left|(dV/dz)/(d\phi/dz)\right|}{V}=\frac{\left|V'/\phi'\right|}{V}, \end{equation} then the error will be given by \begin{equation} \delta \textrm{SC2}=\frac{\left|\left(\frac{\delta V'\phi'-\delta \phi'V'}{\phi'^2}\right)\right| V-\delta V \left|V'/\phi'\right| }{V^2}, \end{equation} where the prime is the derivative with respect to the redshift $z$. \end{appendix} \bibliographystyle{apsrev4-2}

2d38bd2f782d69872204eacb7a3c803b5a50f794

\section{Proof of the First Zonklar Equation} \section*{Acknowledgments} We gratefully acknowledge the support of NVIDIA Corporation with the donation of the Titan V GPU used for this research. This work was partially supported by WAC@Lucca funded by Fondazione Cassa di Risparmio di Lucca, AI4EU - an EC H2020 project (Contract n. 825619), and upon work from COST Action 16101 ``Action MULTI-modal Imaging of FOREnsic SciEnce Evidence (MULTI-FORESEE)'', supported by COST (European Cooperation in Science and Technology). \ifCLASSOPTIONcaptionsoff \newpage \fi {\small \bibliographystyle{IEEEtran} \section{Introduction} \label{introduction} \IEEEPARstart{A}{nomalies} represent a controversial phenomenon in the scientific world. Although they can lead to fascinating discoveries, sometimes they are a symptom of something unexpected that just happened. Even though they can manifest in different ways, all kinds of anomalies origin from a common basic principle: an unexpected prediction from a given theory from what is believed to be a proper answer. \begin{figure}[t] \includegraphics[width=\linewidth]{images/mocca_sketch.pdf} \caption{Schematic representation of the {MOCCA}{} approach. Each feature space is represented as an $x-y$ plane. The cyan dots represent the centroids of the anomaly-free images while the green (red) dots represent the normal (anomalous) samples, respectively. $\tau_i$ is the distance between the deep representation of a given input image from the centroid at layer $i$.} \label{fig:arch_simpl} \end{figure} Concerning the Deep Learning (DL) field, an anomaly might be thought of as an out-of-distribution sample presented as input to a Deep Neural Network (DNN). More specifically, from a statistical point of view~\cite{hawkins1980identification, edgeworth1887xli}, we can discern among outliers and novelties that are described by the same probability distribution of the normal data and anomalies that are instead characterized by completely different statistics. Being able to detect such events is an attractive feature, especially concerning applications such as surveillance systems~\cite{vinayakumar2019deep, zavrak2020anomaly, alhakami2019network, hasan2016learning}, medical diagnosis~\cite{stafford1994application,fernando2020neural,ouardini2019towards,schlegl2017unsupervised}, fraud detection~\cite{roy2018deep,pumsirirat2018credit,lebichot2019deep,fiore2019using}, and defect detection~\cite{tout2017automatic,kumar2008computer}. Indeed the task of Anomaly Detection (AD)~\cite{aggarwal2015outlier,chalapathy2019deep} is among the most active research fields in the machine learning community. Since the cost to collect large amounts of anomalous samples is prohibitive, the AD is usually considered as an unsupervised problem with the training databases containing non-anomalous class instances only. Thus, to detect anomalies, deep models are typically trained on in-manifold samples only to learn an effective boundary that captures the concept of normality from the distribution of one kind of data only. In recent years, One-Class (OC) approaches to AD have drawn the scientific community's interest. Especially, autoencoders~\cite{zhou2017anomaly,gong2019memorizing,wang2020advae} and GANs~\cite{schlegl2019f,akcay2018ganomaly,li2019mad} based approaches reached the highest performance available in the literature. In the Machine Learning (ML) field, commonly adopted approaches leverage the models' final output only, thus interpreting a neural network as a single computational block that performs an input-to-output mapping. Concerning such a point of view, throughout this manuscript, we refer to such an approach as ``holistic" interpretation. Specifically, what we mean by ``holistic" is that both the training and test phases rely on the output of the last layer only, i.e., there is no information extracted from the intermediate levels of the architecture. In such a context, our contribution stems from a different interpretation of the mapping represented by a DNN. We show that by leveraging the deep representations extracted at various depths in both the training and inference phases of a learning model, a neural network reaches higher performance on the AD task than when only the last layer's output is considered. We propose a novel framework, named {Multi-layer One-Class ClassificAtion}{} ({MOCCA}{}), to train and test deep learning models on the AD task. The innovation in our work is the explicit optimization of the intermediate representations and their use in the test phase for the task at hand. {MOCCA}{} leverages the multi-layer structure of deep architectures, differently from commonly used approaches that consider a neural network as a single computational block, i.e., using the output of the last layer only. During training, each layer's feature space is optimized for AD, whereas in the test phase, the deep representations extracted from the trained layers are combined to detect anomalies. To prove the effectiveness of our strategy, we apply it to autoencoders. Specifically, with {MOCCA}{}, we split the training process into two steps. First, the autoencoder is trained on the reconstruction task only. Then, we only retain the encoder tasked with minimizing the $L_2$ distance between the output representation and a reference point, the anomaly-free training data centroid, at each considered layer. Subsequently, we combine the deep features extracted at the various trained layers of the encoder model to detect anomalies at inference time. We show a schematic view of our approach in \autoref{fig:arch_simpl}. Our contributions can be summarized as follows: \begin{itemize} \item we formulate a ``multi-layer'' based approach to AD, named {MOCCA}{}, that explicitly optimizes the representations extracted at different layers of a deep learning model during training, and then combines them in the test phase to detect anomalies; \item we perform extensive experiments on publicly available single-image AD datasets, namely, {CIFAR10}}%~\cite{krizhevskycifar10}}{} and {MVTec AD~\cite{bergmann2019mvtec}}{}, and empirically show that models trained with the {MOCCA}{} approach reach higher performance compared to the state-of-the-art; \item we perform experiments on the {ShanghaiTech~\cite{luo2017revisit}}{} dataset, and show that, even though our method is not tailored for video-based AD, it delivers models with performance comparable to state-of-the-art approaches specially designed for such a task. Thus, showing the high generalization capability of our technique; \item we perform a model analysis to give insights into how our approach works and empirically analyze the benefits of exploiting the representations generated at different layers of a learning model. \end{itemize} The remainder of the paper is organized as follows. In \autoref{rel_works}, we briefly review the related works, while in \autoref{approach}, we describe our approach to the anomaly detection task. In \autoref{datasets} and \autoref{exp_results}, we present the datasets we used and report the obtained results on them, respectively. In \autoref{model_analysis} we perform an analysis of the models, and, finally, in \autoref{conclusions}, we conclude the paper. \section{Related works} \label{rel_works} \begin{figure*}[t] \includegraphics[width=\linewidth]{images/figure_anomaly_new_arch_2.pdf} \caption{Schematic representation of the {MOCCA}{} training. Left: two-stage training for single image input. Right: end-to-end training for video-based AD. To exploit the time correlation among frames, LSTMs are used instead of ``selector" modules. The superscript $s$ ($h$) refers to the \emph{soft} (\emph{hard}) boundary settings.} \label{fig:arch} \end{figure*} The latest approaches to the AD task are mainly based on reconstruction and discrimination techniques. Autoencoders~\cite{borghesi2019semisupervised,pawlowski2018unsupervised,li2019specae,zhou2017anomaly} and GANs~\cite{zhou2020sparse,ngo2019fence,sabokrou2018adversarially} belong to the former class while the latter approach gathers techniques such as the one-class classification~\cite{wang2019gods,hong2020latent,razzak2020one}. Concerning GAN-based approaches, in~\cite{sabokrou2018adversarially}, the authors exploit a reconstruction technique that leverages an autoencoder and a CNN that are adversarially trained. In AnoGAN~\cite{schlegl2017unsupervised} the generator learns to reconstruct the input sample through latent space optimization, and the discriminator generates deep representations for both the original and the reconstructed samples, while in~\cite{zenati2018efficient}, the authors propose to learn an encoder network that maps the input samples directly to the generator's latent space. A slightly different approach is proposed in~\cite{akcay2018ganomaly}, where an explicit latent space minimization is obtained by learning an encoder model. The OC-GAN approach is introduced in~\cite{perera2019ocgan}, where authors use a denoising autoencoder network and a classifier in order to learn the latent representations of the normal samples in an adversarial manner. In~\cite{an2015vaead} variational autoencoders are used to detect anomalies by exploiting the reconstruction probability as the objective. In~\cite{abati2019latentspace} the authors combine a reconstruction approach based on autoencoders with an autoregressive model that learns a factorization of the latent space distribution. In~\cite{bergmann2018improving}, the authors use the structural similarity index metric (SSIM) to train autoencoders while~\cite{huang2019inverse} propose the Inverse-Transform AutoEncoder (ITAE) based on the use of autoencoders that reconstruct images after the application of a set of specific transformations. The One-Class (OC) approach has a long history starting from the study of shallow models. Indeed, first attempts in such a direction date back to the 2000s with the proposal of the One-Class SVM~\cite{scholkopf2000ocsvm,tax2004svdd}. In~\cite{erfani2016high}, a hybrid approach is proposed based on deep autoencoders and OC-SVM, while in~\cite{chalapathy2018ocnn} the authors trained their models with an OC-SVM equivalent loss function. One of the first proposals concerning an end-to-end training approach to OC-AD is proposed in~\cite{ruff2018deep}, where the code generated by an encoder is mapped to a point within a hypersphere so that the normal samples remained inside of it while anomalous ones lay outside. Lastly, in~\cite{oza2019occnn}, the authors use an encoder for getting the latent representations of the normal samples, and a pseudo-negative class is created using zero-centered Gaussian noise in the same latent space. Most recently, Venkataramanan et al.~\cite{venkataramanan2020attention} exploit a variational autoencoder combined with a specialized attention mechanism with the final goal of performing anomaly localization. In~\cite{zaheer2020old}, the authors tackle the problem of the stability training of GANs when there are not lots of data available. A semi-supervised approach is proposed in~\cite{ruff2019deep}, and in~\cite{bergmann2020uninformed}, the authors exploit a student–teacher framework to perform anomaly detection and pixel-precise anomaly segmentation at the same time. In \cite{sultani2018real}, they leverage a multiple instance learning approach while in \cite{gong2019memorizing} a technique named MemAE is introduced where authors have added a memory module to deep autoencoders. Compared to all the works mentioned above, {MOCCA}{} differs from them on two key aspects. On the one side, it exploits the deep representation extracted at various layers of the learning model, both at training and inference time, which contrasts to classical methodology in which only the final output is considered to fulfill the task. On the other hand, it does not make any assumption on the deep features' statistical distributions. Combining these two properties allows the model to adjust each single feature space at its best to accomplish the AD task. \section{Proposed approach} \label{approach} As a general conception, DNNs are a sequence of transformations that approximate a function $f_{\theta}:\ \mathcal{X} \rightarrow \mathcal{Y}$ where $\mathcal{X} \subseteq \mathbb{R}^d$ and $\mathcal{Y} \subseteq \mathbb{R}^m$ are the input and output space, respectively, and $\theta$ are the parameters to be learned at training time. We refer to such an approach as ``holistic" (see \autoref{introduction} for more details) in the sense that the entire net is considered as a single computational block that given an input, returns an output. As opposed to such a point of view, with {MOCCA}{} we adopt a ``multi-layer" interpretation of the learning models where we consider a DNN as a sequence of single transformations each mapping its input to a more representative space: \begin{flalign} \label{eq:layerwise} \resizebox{1.\hsize}{!}{$f_{\theta}(\mathbf{x}) = \phi_m(\mathbf{\theta}_m; \mathbf{o}_{m-1}) \circ \phi_{m-1}(\mathbf{\theta}_{m-1}; \mathbf{o}_{m-2}) \circ .... \circ \phi_1(\mathbf{\theta}_1; \mathbf{x})$} \end{flalign} where each $\phi_i$ term represents the operation performed by a specific layer, and the matrices $\mathbf{\theta}_i$ represent their weights and biases. The output of each operation is reported as $\mathbf{o}_i$, while $\mathbf{x}$ is the network input. Our intuition is that the outputs $\mathbf{o}_i$ of the various layers, i.e., the representations generated at different depths of a DNN, can be exploited to enhance the performance of a learning model on the AD task compared to when the entire decision process leverages the last layer output only. Indeed, it has been already shown in literature~\cite{massoli2020detection,papernot2018deep,carrara2019adversarial} that deep features extracted at various layers of a model can help a DNN to fulfill its task. However, it is not enough to combine the representations at test time only. Instead, all the layers must be trained to a common aim. As mentioned in \autoref{introduction}, our base network is an autoencoder where both the encoder and the decoder are Deep Convolutional Neural Network (DCNN). With {MOCCA}{} we formulate the training process as a two-stage procedure in which we first train the full autoencoder on the reconstruction task only, and then we specialize only the encoder to detect anomalies by exploiting an OC-like objective~\cite{ruff2018deep} applied to different layers of the network. However, we empirically observe that a single-step end-to-end training, in which we optimize the reconstruction and the OC objectives simultaneously, is more effective than the two-step one for video-based AD. A schematic representation of the {MOCCA}{} training procedures is presented in \autoref{fig:arch}. As one can see from the figure, we process the model inner layers' output using ``selector'' and ``LSTM'' modules concerning single-image and video-based data type, respectively. Concerning the ``selector'' blocks, they are made of an average pooling operation or a two-layer neural network concerning the {CIFAR10}}%~\cite{krizhevskycifar10}}{} and {MVTec AD~\cite{bergmann2019mvtec}}{}, respectively. Specifically, concerning the {CIFAR10}}%~\cite{krizhevskycifar10}}{} dataset, we use only the pooling operation to fully assess the real advantages brought by {MOCCA}{}. As mentioned above, we exploit the OC objective and we evaluate it by using the deep features extracted at different depths of the encoder model. Specifically, we considered two variants for such an objective function termed \emph{soft-} and \emph{hard-}boundary. The first one is expressed as follows: \begin{align} \label{eq:obj_layer_s} \mathcal{L}_{j}^{s} = R_j^2 + \frac{1}{|B|\cdot\nu}\sum^{|B|}_{i} \mathrm{max}\{0, \parallel \phi_j(\mathbf{x}_i; \theta) - \mathbf{c}_j \parallel^2 - R_j^2\} \end{align} The goal of such a loss is to minimize the volume of the hypersphere at each layer $j$, centered at $\bf c_j$ and with radius $R_j$, that is interpreted as the boundary region for normal data~\cite{tax2004svdd}. Then, the goal of~\autoref{eq:obj_layer_s} is to minimize the radius, $R_j$, of such spheres (one for each trained layer). In other words, we expect the ``normal'' data to lie within a sphere, at each layer, while the anomalous samples are expected to remain outside of it. The second addend in the equation penalizes ``normal'' data points that lie outside the sphere after being passed through the network. The radius $R_j$ is a scalar quantity evaluated as the $1-\nu$ quantile of the features' distance distribution, in a mini-batch, from the centroid $\bf c_j$. We re-evaluate the radius at each layer at regular intervals while training. A decreasing value of the radius at each layer is an indicator of converging training. The other terms in the equation have the following interpretation: $|B|$ is the mini-batch size, $\nu$ is a hyperparameter that allows controlling the fraction of allowed outliers, $\phi_j$ represents the function that the layer $j$ carries out, and $\mathbf{x}_i$ is the model input. Concerning the \emph{hard-}boundary loss, it is expressed as follows: \begin{align} \label{eq:obj_layer_h} \mathcal{L}_{j}^{h} &= \frac{1}{|B|\cdot\nu}\sum^{|B|}_{i} \parallel \phi_j(\mathbf{x}_i; \theta) - \mathbf{c}_j \parallel^2 \end{align} Differently from~\autoref{eq:obj_layer_s}, ~\autoref{eq:obj_layer_h} simply tries to reduce as much as possible the distance of each sample from the layer's centroid by employing a quadratic loss. After the first training step in which we tasked the full autoencoder with the reconstruction objective, we retain only the encoder and perform an initial forward step on the whole training dataset (that contains non-anomalous samples only) to extract deep features at different depths. Subsequently, we evaluate the centroids, at each layer, as the average of those features. We performed experiments in which we tested the hypothesis of using medoids instead of centroids, but we did not observe any improvement. Once we evaluate the centroids, they are kept fix while training the encoder. We also experimented with several strategies to re-evaluate them after a specific number of training iterations, but we did not observe tangible improvements. Regarding the video-based AD, we initialize the centroids at the beginning of the training, i.e., with the model not trained. Considering a set of layers $\mathcal{J} = \{ j\ |\ j=\ 0,\ 1,\ ... J\}$, we formalize the {MOCCA}{} objective, during the second-step of the training, as: \begin{flalign} \label{eq:obj_full} \mathcal{L}^{s/h} &= \frac{1}{|\mathcal{J}|}\sum_{j}^{|J|}\mathcal{L}_{j}^{s/h} + \frac{\lambda}{2}\sum_{p}^{|P|}\parallel \theta_p \parallel^2 \end{flalign} where $|J|$ is the number of layers we consider, and the sum runs over the layer indexes $j$. The last term of the objective is the $L_2$ regularization for the model parameters. \section{Experiments} \label{exp_results} In this section, we report our experimental results. However, before that, we describe the various metrics we use to assess the models' performance. \input{tables/cifar10_auroc} \subsection{Metrics} To assess the performance of the models trained with {MOCCA}{} and compare them to the other approaches in the literature, we exploit two metrics: the Area Under the Curve (AUC) and the maximum Balanced Accuracy (maxBA). The former metric is the area under the Receiver Operating Characteristics curve. Instead, concerning the latter, the Balanced Accuracy (BA) represents the arithmetic mean between the sensitivity, i.e., percentage of anomalous samples correctly detected, and the specificity, i.e., same as the sensitivity but for non-anomalous samples: \begin{flalign} \label{eq:mba} \mathrm{BA} = \frac{TP}{2\cdot (TP\ +\ FN)} + \frac{TN}{2\cdot (TN\ +\ FP)} \end{flalign} where $TP$ and $FN$ are the true positives and the false negatives, respectively, and $TN$ and $FP$ are the true negatives and the false positives, respectively. In the AD context, it is useful to quote both the AUC and the maxBA metrics. The former one provides an aggregate measure of the performance of a model across all possible classification thresholds. Instead, maxBA is a measure of performance at a specific threshold that could be used in production. It selects the threshold for which the balanced accuracy measure, i.e., the average among the correctly classified images for anomalous (true positives) and anomaly-free test images (true negatives), is maximum and reports the obtained BA. We evaluate both metrics only on the {MVTec AD~\cite{bergmann2019mvtec}}{} dataset since for the {CIFAR10}}%~\cite{krizhevskycifar10}}{} and {ShanghaiTech~\cite{luo2017revisit}}{} datasets we only found the AUC values reported in the literature. Concerning the anomaly score for a given input image, we evaluate its value as: \begin{flalign} \label{eq:anscore} \tau_j(\mathbf{x}) &= \parallel \phi_j(\mathbf{x}, \theta) - \mathbf{c}_j \parallel^2 \\ \nonumber \gamma(\mathbf{x}) &= \frac{1}{|\mathcal{J}|}\sum_j^{|\mathcal{J}|} \begin{cases} \tau_j(\mathbf{x}) & \text{hard boundary}\\ \\ \tau_j(\mathbf{x}) - R_j^2 & \text{soft boundary}\\ \end{cases} \end{flalign} where $\mathbf{x}$ is the input image, $\mathcal{J} = \{ j\ |\ j=\ 0,\ 1,\ ... J\}$ is the set of layers we consider, $\phi_j(\mathbf{x}, \theta)$ is the feature vector extracted at layer $j$, and $\mathbf{c}_j$ and $R_j$ are the center of the hypersphere and its radius at the layer $j$, respectively and $\gamma$ is the anomaly score. We refer the reader to \autoref{approach} for further details on the meaning of the boundaries. Concerning the textures-type classes from the {MVTec AD~\cite{bergmann2019mvtec}}{}, we evaluate the anomaly score as the maximum among the scores relative to each of the 64x64 patches of the given image: \begin{flalign} \label{eq:mvtec_pathces_score} \gamma^{h/s}(\mathbf{x}) = \mathrm{max}\big\{\gamma^{h/s}(\mathrm{patch}_i))\ \ |\ \ i\ =\ 1,\ 2,\ ...,\ 64 \big\} \end{flalign} where the superscripts $s$ and $h$ correspond to when we apply a ``soft" or ``hard" boundary while training the model, respectively. More details on how we extract patches from a single image can be found in \autoref{mvtecpreprocess}. Lastly, considering video-based input we consider a single input clip as made of 16 frames. We then apply a sliding window technique to move through all the frames of a given video and construct the input clips. Since each frame can appear multiple times across different clips, we evaluate its score as the mean value among all of its scores. Moreover, a single frame can have different scores in different clips having a different time correlation, captured by the LSTMs (see \autoref{fig:arch}), with all the other frames. For such a reason, we normalize the score of each frame to the maximum and minimum values of the scores within the clips in which the frame under analysis is present: \begin{flalign} \label{eq:videoan} \gamma^{h/s}(\mathbf{x}_i) = \frac{\langle \gamma^{h/s}(\mathbf{x}_i) \rangle - \mathrm{max_{clips}}\langle \gamma^{h/s}(\mathbf{x}_i) \rangle}{\mathrm{max_{clips}}\langle \gamma^{h/s}(\mathbf{x}_i) \rangle - \mathrm{min_{clips}}\langle \gamma^{h/s}(\mathbf{x}_i) \rangle} \end{flalign} Finally, we add a reconstruction term to the score. \subsection{Experimental results - CIFAR10} Concerning the {CIFAR10}}%~\cite{krizhevskycifar10}}{} dataset, we instantiate each class as a single AD problem, and we train ten different seeded models on each of them. Such a procedure allows us to quote a mean AUC value and the corresponding variance. We report the results in \autoref{tab:cfr10}. As we can see from \autoref{tab:cfr10}, our approach reaches the highest performance on six out of ten classes. Moreover, on class-1, class-5, class-7, class-8, and class-9, the {MOCCA}{} method performs better than the state-of-the-art (SotA) results concerning both the ``soft" and the ``hard" boundaries. As reported in \autoref{train_details}, on the {CIFAR10}}%~\cite{krizhevskycifar10}}{} dataset we use a LeNet-like architecture as in~\cite{ruff2018deep}. Moreover, to better emphasize that our approach's higher performance is not due to a mere addition of more models to the baseline, we use averaging pooling layers as ``selectors" blocks. Thus, since we use the same architecture as in~\cite{ruff2018deep}, we can conclude that the higher performance of our models are only due to the use of {MOCCA}{} and not because we use deeper models or because we add more branches to the base architecture. To summarize the previous results, we report in \autoref{tab:cfr10_means} the AUC values, for each model in \autoref{tab:cfr10}, averaged among all the ten classes of the dataset. \input{tables/cifar10_auroc_means} From \autoref{tab:cfr10_means}, it is clear that our approach reaches the highest performance concerning both types of boundary settings. Moreover, we can appreciate that we obtain higher performance, also considering larger models such as LSA~\cite{abati2019latentspace}. \subsection{Experimental results - MVTec AD} \input{tables/mvtec_tables_no_means} Regarding the {MVTec AD~\cite{bergmann2019mvtec}}{} dataset, also, in this case, we consider each class as an independent AD problem. As reported in \autoref{mvtecpreprocess}, the dataset classes are divided into texture- and object-like sets. For each class, we report the maxBA and the AUC in \autoref{tab:mvtec_balacc} and \autoref{tab:mvtec_auroc}, respectively. Regarding the texture-type of classes, we see from the tables that the {MOCCA}{} approach allows our models to reach the highest performance on three out of five classes concerning both the \emph{hard} and \emph{soft} boundaries. Similar reasonings hold in the case of object-type classes, too. Concerning the results from~\cite{venkataramanan2020attention}, it is important to highlight that, even though we report their results, they should not directly compared with others. The reason for that is because in~\cite{venkataramanan2020attention}, the models are trained on more data rather than on MVTec AD only. Thus, those results are not directly comparable with the other methods. Due to the very low number of test images available in the dataset, typically large variations in the performance of the model are observed among the different classes. Thus, to better compare the performance of the various approaches, we report in \autoref{tab:mvtec_means} the overall mean values for the maxBA and the AUC evaluated among all classes of the dataset. \input{tables/mvtec_means} From the table, we conclude that the {MOCCA}{} approach allows us to reach the highest performance on both types of metrics considering both the \emph{soft} and \emph{hard} type of boundary. \subsection{Experimental results - ShanghaiTech} Differently from the {CIFAR10}}%~\cite{krizhevskycifar10}}{} and {MVTec AD~\cite{bergmann2019mvtec}}{} datasets, the {ShanghaiTech~\cite{luo2017revisit}}{} concerns the video-based AD task. Although we test models trained with {MOCCA}{} against such a protocol, it is essential to stress that our approach is not specially designed for the video-based scenario. We report our results in \autoref{tab:shanghai} and others available in the literature. \input{tables/shanghaitech_auroc} From \autoref{tab:shanghai}, we can see that our approach's performance is utterly comparable to the current SotA models, specifically designed to handle video-based input. Thus, showing that our method is applicable to both the image- and video-based anomaly detection tasks. Indeed, the only modification we apply to {MOCCA}{} for video-based contexts is to move to a single-step training, based on the same objectives, and to substistute the ``selector" modules with LSTMs. \section{Conclusions} \label{conclusions} The anomaly detection task is still an open challenge in many scientific fields. Several approaches have been proposed to tackle this problem in the context of deep learning, typically based on an unsupervised training paradigm. Indeed, being rare events, collecting anomalous samples to construct a supervised training dataset might be extremely expensive. Thus, approaches in which neural networks automatically learn the concept of ``normality" from non-anomalous data only represent a promising solution. We propose to adopt a multi-layer approach, named {MOCCA}{}, to exploit the output of a deep model at different depths to detect anomalous input in the one-class setting. Differently from the usual ``holistic" interpretation of a learning model in which a neural network is considered a single computational block, {MOCCA}{} explicitly leverages the networks' multi-layer composition. Specifically, we show that such an approach enhances a neural network's discrimination capability. We conduct extensive experiments on three different datasets and perform an analysis of the models to support our intuitions. We test our method against the single-image AD task showing that it improves the state-of-the-art both on the CIFAR10 and MVTec AD datasets. Specifically, concerning the performance averaged among all the classes, {MOCCA}{} improves upon the literature results with both the \emph{soft} and \emph{hard} type of boundary. We acknowledge the best improvement concerning the overall maxBA on the MVTec AD dataset that overcomes the state-of-the-art results by 6\%. Moreover, even though our approach is not tailored for the video-based AD task, we test it also using such a protocol by employing the ShanghaiTech dataset. From the experimental results, we see that with {MOCCA}{}, the models' performance is utterly comparable to what was obtained by approaches specially designed for such a task. Thus, showing the high generalization capability of our method. Finally, we report insights about the behavior of models trained with {MOCCA}{} by performing an ablation study and reporting the different CDFs of the distance of the deep representations from the centroids of a given class across different layers. Such an analysis, pointed out that the benefits from using {MOCCA}{} are two-fold: on the one side, we obtain discriminative deep features from more layers, and on the other hand, we are able to set more discriminative thresholds. \section{Datasets and Training} \label{datasets} This section reports the used datasets and provides details about the training procedure that we adopt. \subsection{CIFAR10} The {CIFAR10}}%~\cite{krizhevskycifar10}}{} dataset contains 50K training images and 10K test ones shared among ten different classes. We preprocess the images by applying a global contrast normalization procedure using the $L_1$ norm, and then we normalize them to be in the range $[0, +1]$. Given each class, which we refer to as the ``normal class", we have 5000 images to train the model, and we evaluate each model's performance on the whole test set. With such a training approach, the model only sees instances from the ``normal class" and never sees any anomaly while learning. \subsection{MVTec}\label{mvtecpreprocess} The {MVTec AD~\cite{bergmann2019mvtec}}{} dataset comprises $\sim$3.6K and $\sim$1.7K high-resolution images to train and test DNNs, respectively, shared among 15 classes which are divided into two categories: textures (5 classes) and objects (10 classes). The dataset is split into two sets: one for training purposes containing ``normal" images only and one specifically designed to test the models' performance. Specifically, the latter one contains anomalous images, with defects of different types and non-anomalous ones. We apply two different preprocessing operations to objects- and texture-type classes. Concerning the formers, we first resize the image to 128x128 pixels and then apply a random rotation in the range $[-\pi/4, +\pi/4]$ when the anomaly of the object is not related to its orientation. Instead, we first resize images to 512x512 pixels in the latter type of classes, and then we crop 64x64 non-overlapping patches used as input to the network. Moreover, we augment the data by exploiting a random rotation in the range $[0, +\pi/4]$. Finally, we normalize all the objects- and texture-type images to be in the range $[-1, +1]$. In \autoref{fig:mvtec_example} we show an example of textures- and objects-type images from the dataset. \begin{figure}[!ht] \includegraphics[width=\linewidth]{images/mvtec_smaller.pdf} \caption{Samples from different classes of the {MVTec AD~\cite{bergmann2019mvtec}}{} dataset. Top: texture classes. Bottom: object classes. We highlight in red the anomalies.} \label{fig:mvtec_example} \end{figure} \subsection{ShanghaiTech} The {ShanghaiTech~\cite{luo2017revisit}}{} dataset is one of the largest video anomaly datasets. It comprises over 270,000 training frames from 13 scenes with complex light conditions and camera angles, accounting for 130 abnormal events. We follow the same preprocessing strategy as in~\cite{abati2019latentspace}, i.e., we use a MOG-based approach to estimate the background and remove it from the frames. By employing such a procedure, we eliminate the necessity of background estimation and let the model focus on foreground objects only. Given a video, we construct clips made by 16 frames to be used as input to the learning models. To exploit the temporal correlation among frames, we employ LSTM cells (we refer the reader to \autoref{approach} for more details about our models' architecture). Finally, we resize each frame to 256x512 pixels to feed models. In \autoref{fig:shanghaitect_example} we report an example of ``normal" and anomalous frames from two different videos. \begin{figure}[!ht] \includegraphics[width=\linewidth]{images/shanghaitech_example_smaller.pdf} \caption{Samples of ``normal" (left) and anomalous (right) frames from the {ShanghaiTech~\cite{luo2017revisit}}{} dataset. We highlight in red the anomalies.} \label{fig:shanghaitect_example} \end{figure} \subsection{Training details}\label{train_details} Concerning the {CIFAR10}}%~\cite{krizhevskycifar10}}{} dataset, we use a LeNet-like architecture as in~\cite{ruff2018deep}, made of three convolutional layers and one fully connected layer after them. We use the Adam~\cite{kingma2014adam} optimizer for both pre-train the full architecture and train the encoder with learning rates of $1.e^{-3}$ and $1.e^{-4}$, respectively. We set the encoder code's size equals to 128 and the value of the parameter $\nu$ in the range [0, 0.1]. Finally, we use a batch size of 256. As we mentioned in \autoref{approach}, concerning the {CIFAR10}}%~\cite{krizhevskycifar10}}{} dataset, we use an average pooling operation as the ``selector" module. Thus, we emphasize that the higher performance reached by using {MOCCA}{} is not due to larger, deeper, or more models. Instead, the benefits of using {MOCCA}{} stand from its ability to exploit the representations generated at different depths of a learning model. To our aim, we train ten different seeded models on each class, considering the other nine as anomalies. Such a procedure allows us to estimate the mean response of our approach and its standard deviation. Regarding the {MVTec AD~\cite{bergmann2019mvtec}}{} and the {ShanghaiTech~\cite{luo2017revisit}}{} datasets, we use a residual-like structure that comprises four and five residual blocks, respectively, followed by two fully connected layers. For this dataset, the ``selector" blocks consist of a convolutional layer followed by a pooling operation, a batch norm layer, and a final fully connected layer. In video-based AD, we substitute the ``selector" networks with LSTM cells to exploit the time correlation among the frames within a given input clip. To train models on those two datasets, we use again the Adam~\cite{kingma2014adam} optimizer and a learning rate in the set $\{10^{-2}, 10^{-3}\}$ that we drop by a factor of ten at specific epochs depending on the class under study. Being each class of each dataset an independent AD problem, we use different hyperparameters to train the models on each of them. Moreover, we do not always use the same set of layers to evaluate the objective in \autoref{eq:obj_full}. Indeed, we train the models by using different layer combinations and finally select the best performing one in each class. To allow the researchers to reproduce our work, we made the code publicly available on GitHub\footnote{\url{https://github.com/fvmassoli/mocca-anomaly-detection.git}}. \section{Model Analysis}\label{model_analysis} In this section, we look in more detail at the behavior of our models. First, we focus on an ablation study to show the impact of using a different number of layers to evaluate a specific image's anomaly score. Specifically, we prove that with {MOCCA}{}, we effectively succeed in exploiting the deep representations extracted at different depths of a DNN. To our aim, we perform the ablation study considering the ``Leather" class of the {MVTec AD~\cite{bergmann2019mvtec}}{} dataset. We report the results in \autoref{tab:ablation_mvtec}. \input{tables/ablation_study_tile_mvtec} As described in \autoref{train_details}, the encoder's architecture consists of four residual blocks followed by two fully connected layers. The indexes in the first column of \autoref{tab:ablation_mvtec} correspond to the layers' ordering where the $0$-th layer is the closest to the input. The results in \autoref{tab:ablation_mvtec} should be interpreted as follows. Each row in the table represents a different model that we trained with {MOCCA}{} by considering the output from the layers listed in the first column. For example, the first row represents the results we obtained by considering the output (in the training and test phases) from the last layer only, while in the second row we consider the layer 5 and 6 together. We aim at showing that by exploiting the output at different layers while training a learning model, we can use the output from those same layers at inference time to enhance the network's discrimination power. On the contrary, we experimentally observed that training the model using the last layer's output only and then using more layers at inference time always gave worse results. Such an observation is one of the key points on which we base our approach. As it is clear from the table, independently from the type of boundary we apply, we obtain higher results by utilizing more layers. This result supports our intuition that the features extracted at different depths help to detect anomalies in the input images. By carefully looking at \autoref{tab:ablation_mvtec} we notice that the maxBA improves until we add layers 4 and 5 to the last one. Moreover, we can notice that, in the case of the \emph{hard} boundary setting, we can obtain a slight improvement by adding layer 3. Finally, we notice that by adding more layers, we do not see any further improvement. We can interpret such behavior by considering that since the first layers are closer to the input data, they specialize on simple patterns. On the contrary, higher layers generate representations that amplify aspects of the input that are important for discrimination~\cite{lecun2015deep}, thus more useful to fulfill the final task. Hence, adding layers that are too close to the input data does not improve the learning model's overall performance. We then focus our attention on the distribution of the distances among the features and the centroids, of a given class, at different layers. Specifically, we compare our training approach against the ``holistic" approach, i.e., when the learning model is considered a single computational block. As specified in \autoref{introduction}, ``holistic" refers to the approach similar to~\cite{ruff2018deep} where the last layer's output only is used to train and test the encoder on the AD task. To our aim, we train two identical models once with the {MOCCA}{} approach and then by evaluating the OC loss on the last layer only (``holistic"). We report the resulting Cumulative Density Function (CDF) in \autoref{fig:mvtec_cdf}. \begin{figure}[!h] \includegraphics[width=\linewidth]{images/mvtec_dist_cdf.pdf} \caption{CDF of the test images distance from the ``normal" class centroid for the {MOCCA}{} approach (top row) and for the ``holistic" one (bottom row). The blue (red) line represents the CDF of ``normal" (anomalous) images.} \label{fig:mvtec_cdf} \end{figure} Concerning the model trained with {MOCCA}{}, we see that the distributions for ``normal" images always lies at the left of the corresponding for anomalous ones (as one would expect). Moreover, we see that the CDFs of ``normal" images rise faster than the ones of anomalous samples. Thus, allowing one to set a more discriminative threshold on the anomaly score. On the contrary, we see that by considering the last layer only while training on the AD task, the distributions of distances for anomalous and ``normal" images are highly overlapped even in the last layer. Thus, by training with {MOCCA}{}, we have a double gain: on the one side, we obtain discriminative deep features from more layers, and on the other hand, we are able to set more discriminative thresholds.

dbf4bb53110df7647812ad77da7451167835f5fe

\section{Introduction} \subsection{Funk and Hilbert metrics} The Hilbert metric in the interior of a convex set $K\subset\mathbb R^n$ is given by $$d^H_K(x,y)=\frac12 \log \frac{|ay||xb|}{|ax||yb|}, $$ where $a,x,y,b$ appear in that order on the line through $x,y$ with $a,b\in\partial K$. It is evidently invariant under invertible projective maps. Hilbert discovered this metric while attempting to generalize the Beltrami-Klein model of hyperbolic geometry. It is moreover a \emph{projective} metric, meaning that straight segments are geodesic. Hilbert's fourth problem \cite{hilbert_problems} asks for a construction of all projective metrics. A different example of a projective metric, which we now recall, was discovered by Funk \cite{funk}. It is given by the non-symmetric distance function $$d^F_K(x,y)=\log\frac{|xb|}{|yb|},$$ where $b$ is the intersection of the ray $\vv{xy}$ with $\partial K$. Various problems in Hilbert geometry have been raised and studied, notably those of volume growth \cite{colbois_verovic, vernicos_polytopes, berck_bernig_vernicos, tholozan} and Gromov hyperbolicity \cite{benoist}. The Funk metric, on the other hand, has been for the most part denied the spotlight. In this paper we attempt to argue that at least from the perspectives of convex geometry and billiard dynamics, it is the Funk metric that is most natural to study. Let $K\subset \mathbb{R}^n$ be a convex compact set with non-empty interior, placed in the affine space $\mathbb{R}^n$. The \emph{Funk metric} on $K$ is arguably the simplest (non-symmetric) metric one can define naturally in its interior. Indeed, it is the non-reversible Finsler metric whose tangent unit ball at $x\in \inter(K)$ is $B_xK=K\subset T_x\mathbb{R}^n=\mathbb{R}^n$, namely $K$ itself with the origin fixed at $x$. The Funk and Hilbert metrics are closely related: $$d^H_K(x,y)=\frac12(d^F_K(x,y)+d^{F}_K(y,x)).$$ It is a remarkable fact that the symmetrization of the Funk metric is already projectively invariant. This hints that the Funk metric comes close to being projectively invariant itself. We shall pursue this idea along several paths. \subsection{Summary of results} If $K\subset\mathbb{R}\mathbb{P}^n$ is a convex body with interior, we may fix a hyperplane at infinity $\eta$ that is disjoint from $K$, and consider the corresponding $\eta$-Funk metric on $K$, denoted $\Funk_K^\eta$. Changing $\eta$ results in a change to the Funk metric, but of a very particular type: the Finsler norm is changed by an exact 1-form. This is the core reason behind the projective invariance of many metric invariants associated to the Funk metric, as follows. For a submanifold $X\subset \inter(K)$, let $\Funk^\eta_K(X)$ stand for the Finsler metric that is induced on $X$ from $\Funk_K^\eta$, that is it is the intrinsic metric induced on $X$ from the Funk metric. When $\eta$ is clear from context, e.g. when $K$ belongs to an affine space, we simply write $\Funk_K(X)$. \begin{MainTheorem}\label{main:projective_invariants} Assume $K\subset \mathbb{R}^n$ is $C^2$-smooth and strictly convex. Let $\Omega\subset \inter(K)$ be a smooth $k$-dimensional submanifold with boundary. The following are invariant under the action of any invertible projective map $g$ keeping $K$ compact. \begin{enumerate} { \item The $k$-dimensional Holmes-Thompson volume of $\Funk_K(\Omega)$, denoted $\vol^{F}_K(\Omega)$. \item The geodesics of $\Funk_K(\Omega)$, as well as the length of closed geodesics. \item Assume $k=n$, that is $\Omega$ is a smooth domain. Then the Finsler billiard orbits in $\Funk_K(\Omega)$, as well as the length of the periodic orbits, are invariant.} \end{enumerate} \end{MainTheorem} Therefore, whenever we discuss any of those invariants, we may assume $K\subset\mathbb{R}\mathbb{P}^n$ is a convex set, without specifying an affine chart. Here and in the following, smoothness and strict convexity assumptions in statements concerning volumes can be relaxed through a continuity argument. The \emph{reverse Funk metric} is defined similarly to the Funk metric by setting $B_xK=a_x(K)$, where $a_x$ is the antipodal map with respect to $x$, or equivalently $d_K^{RF}(x,y)=d_K^F(y,x)$. For $K, \eta$ as before we denote the corresponding reverse Funk metric on $\inter(K)$ by $\RFunk_K^\eta$. Theorem \ref{main:projective_invariants} applies equally to $\RFunk$. In a sense, those invariants are more natural than the corresponding ones for the Hilbert metric, which is most clearly manifested by their invariance under projective duality. Recall that the \emph{polar} convex body $K^\vee\subset \mathbb{P}^\vee=(\mathbb{R}\mathbb P^n)^\vee$ is the closure of the set of all hyperplanes disjoint from $K$. \begin{MainTheorem}\label{main:duality} Assume $K\subset \inter(L)$ are $C^2$-smooth, strictly convex bodies in $\mathbb{R}\mathbb{P}^n$. Then \begin{enumerate} \item $\vol^F_L(K)=\vol^{F}_{K^\vee}(L^\vee)$. \item The billiard orbits in $\Funk_L(K)$ and $\RFunk_{K^\vee}(L^\vee)$ are in a natural bijective correspondence. Moreover, a periodic orbit corresponds to a periodic orbit, and they have equal length. \item The Holmes-Thompson volumes of $\Funk_L(\partial K)$ and $\Funk_{K^\vee}(\partial L^\vee)$ coincide. \item The geodesics of $\Funk_L(\partial K)$ and $\RFunk_{K^\vee}(\partial L^\vee)$ are in a natural bijective correspondence, and the respective closed geodesics have equal length. \end{enumerate} \end{MainTheorem} In the limit where $K$ is shrinking to a fixed point inside $L$ through homotheties, we get corresponding statements for linear spaces with Minkowski norms, as follows. Let $K, L\subset V=\mathbb{R}^n$ be fixed convex sets with $0$ in their interior, and consider $K\subset (V, \|\bullet\|_L)$ and $L^o\subset (V^*, \|\bullet\|_{K^o})$. The first statement becomes a tautology, as both volumes, when properly rescaled, approach the symplectic volume of $K\times L^o\subset V\oplus V^*$. The second statement is the duality of the Minkowski billiards $(K, L)$ and $(L^o,-K^o)$, first established in \cite{gutkin_tabachnikov}, and which permeates the work of Artstein-Avidan and Ostrover on the relationship between Minkowski billiards, symplectic geometry and the Mahler conjecture \cite{artstein_ostrover, artstein_karasev_ostrover}, see also \cite{ostrover} for a survey. It is the approach of \cite{artstein_ostrover} that we take to establish the duality. The third statement was established in \cite{holmes_thompson} in the particular case of $K=L$. The fourth statement is the (non-symmetric) Sch\"affer dual girth conjecture \cite{schaeffer}, which can also be understood as a duality of gliding billiard orbits. It was established in \cite{alvarez-paiva} together with the general case of the third statement. The first statement implies an isomorphic form of duality in Hilbert geometry. Let $\vol_K^H$ be the Holmes-Thompson volume of the Hilbert metric in $K$. \begin{MainCorollary}\label{cor:hilbert_volume} Given $K\subset \inter(L)$ in $\mathbb{R}\mathbb{P}^n$, $\vol^H_L(K)\leq \frac{1}{2^n}{2n\choose n} \vol_{K^\vee}^H(L^\vee)$. Equality is attained if and only if $K$ is an ellipsoid, and $L$ is a simplex. \end{MainCorollary} { The third statement implies a duality in Hilbert geometry in the projective plane. \begin{MainCorollary}\label{cor:hilbert_plane} Let $K\subset\inter(L)$ be two convex sets in $\mathbb{R}\mathbb{P}^2$. Then the Hilbert lengths $\vol^H_L(\partial K)$ and $\vol^H_{K^\vee}(\partial L^\vee)$ coincide. \end{MainCorollary} } It is not hard to verify that for an ellipsoid $B\subset\mathbb{R}\mathbb{P}^n$, the Funk billiard (though not the metric itself) in a domain $\Omega\subset \inter(B)$ coincides with that of the Beltrami-Klein hyperbolic metric. It follows in particular that for an ellipsoid $K\subset B\subset \mathbb{R}\mathbb P^n$, the Funk billiard in $K$ is completely integrable, see \cite{veselov}. We provide a Funk perspective and an alternative proof of this result in the projective plane, as follows. \begin{MainTheorem}\label{main:conics} Let $K\subset \inter(B)$ be nested ellipses in $\mathbb{R}\mathbb{P}^2$. Then every orbit of the $B$-Funk billiard in $K$ has a caustic, which is a conic in the dual pencil defined by $\partial K, \partial B$. Furthermore, if $Q$ is a caustic for $\Funk_B(K)$, and $\widehat Q\subset \mathbb P^\vee$ is the caustic of the dual orbit in $\RFunk_{K^\vee}(B^\vee)$, then $\widehat Q^*, \partial B, \partial K, Q$ form a harmonic quadruplet of conics. \end{MainTheorem} Here $Q^*\subset\mathbb P^\vee$ denotes the dual quadric, consisting of all tangent hyperplanes to $Q\subset\mathbb{P}$. For a harmonic quadruplet of quadrics, see Definition \ref{def:harmonic_quadrics}. In particular, we can interpret the Poncelet porism of any pair of nested conics in terms of closed Funk (equivalently, hyperbolic) billiard orbits. In the second part of this note, we consider the Holmes-Thompson volume of balls in the Funk metric. Assume $K\subset\mathbb{R}^n$, and let $B_r^F(q, K)$ denote the outward $r$-ball in the Funk metric on $K$ centered at $q$. Denoting by $K^z\subset (\mathbb{R}^n)^*$ the dual body with respect to $z$, and by $\omega_n$ the volume of the Euclidean unit ball in $\mathbb{R}^n$, one has $\vol^{F}_K(B_r^F(0,K))=\frac{1}{\omega_n}\int_{(1-e^{-r})K}|K^z|dz$. We show that for $K$ unconditional and any $r$, the volume of $B_r^F(0, K)$ is maximized when $K$ is an ellipsoid, which translates into the following equivalent systems of affine inequalities. \begin{MainTheorem}\label{main:bodies_inequality} For an unconditional convex body $K\subset \mathbb{R}^n$, one has \begin{itemize} \item For all $0< \rho<1$ $$ \int_{\rho K} |K^z|dz\leq n\omega_n^2\int_0^\rho\frac{t^{n-1}dt}{(1-t^2)^{\frac{n+1}{2}}}.$$ \item For all $j\geq 0$, $$\int_{K\times K^o}\langle x, \xi\rangle^{2j}dxd\xi\leq \frac{(2\pi)^n}{n+2j}\frac{1}{ (n+2j)!!(n-2)!!(2j-1)!!} \left(\frac2\pi\right)^{\frac{1-(-1)^n}{2} }.$$ \end{itemize} Equality in each case is attained uniquely by ellipsoids. \end{MainTheorem} \noindent For $j=1$, this is a theorem of Ball \cite{ball_remarks}. In the limit $\rho\to0$, which corresponds to Funk balls of infinitesimal radius, the Blaschke-Santal\'o inequality is recovered: $$|K\times K^o|\leq |B^n\times (B^n)^o|=\omega_n^2.$$ On the other end of the scale, one has the following result of \cite{berck_bernig_vernicos}. \begin{MainProposition}\label{prop:centro_affine_main} Assume $K\subset\mathbb{R}^n$ is $C^2$-smooth and strictly convex, and $0\in\inter(K)$. Then as $\rho\to 1^{-}$, $$\int_{\rho K}|K^z|dz\sim 2^{-\frac{n-1}{2}}\frac{\omega_n}{n-1}\frac{1}{(1-\rho)^{\frac{n-1}{2}}}\Omega_n(K),$$ where $\Omega_n(K)$ is the centro-affine surface area of $K$. \end{MainProposition} For the definition of centro-affine surface area, see eq. \eqref{eq:centro_affine}. Therefore in the limit of large Funk balls, Theorem \ref{main:bodies_inequality} yields the centro-affine isoperimetric inequality: $$\Omega_n(K)\leq \Omega_n(B^n)=n\omega_n.$$ Thus Funk geometry interpolates between those two extremes through a continuum of affine inequalities. As both the Blaschke-Santal\'o and the centro-affine inequalities are valid for arbitrary convex bodies with centroid at the origin, we naturally conjecture that Theorem \ref{main:bodies_inequality} remains true in this generality. The Colbois-Verovic volume entropy conjecture \cite{colbois_verovic} asserts that the volume growth entropy of metric balls in Hilbert geometry is maximized by ellipsoids, or equivalently by smooth strictly convex bodies. This was recently proved by Tholozan \cite{tholozan} and Vernicos-Walsh \cite{vernicos_walsh} using very different methods, and previously established up to dimension $3$ in \cite{vernicos3d}. Theorem \ref{main:bodies_inequality} yields yet another proof of the volume entropy conjecture, albeit only for unconditional bodies. We write $B^H_r(q,K)$ for the metric ball in the Hilbert metric. \begin{MainCorollary}\label{cor:entropy} For $K\subset \mathbb{R}^n$ a convex unconditional body, $$\limsup_{R\to \infty}\frac{1}{R}\log \vol_K^H (B^H_R(q,K))\leq n-1.$$ \end{MainCorollary} Theorem \ref{main:bodies_inequality} follows from a more general functional form of the same inequalities. Below $\mathcal L$ is the Legendre transform, defined in eq. \eqref{eq:legendre}. \begin{MainTheorem}\label{main:inequality_functional} For an unconditional Borel function $f=e^{-\phi}:\mathbb{R}^n\to[0,\infty)$, one has the equivalent systems of inequalities \begin{itemize} \item For all $j\geq 0$, $$\int_{\mathbb{R}^n\times\mathbb{R}^n}\langle x,\xi\rangle ^{2j}e^{-\phi(x)-\mathcal L\phi(\xi)}dx d\xi\leq (2\pi)^n \frac{(n-2+2j)!!}{(n-2)!!(2j-1)!!}.$$ \item For all $0\leq \rho<1$, $$\int_{\mathbb{R}^n\times\mathbb{R}^n}e^{-\phi(x)-\mathcal L \phi (\xi)+\rho\langle x,\xi\rangle}dxd\xi\leq \frac{(2\pi)^n}{(1-\rho^2)^{\frac n 2}}.$$ \end{itemize} Equality in each case is attained uniquely, up to equality a.e., by $f$ that is a multiple of a gaussian. \end{MainTheorem} In the limit $\rho\to0$ we recover the functional Blaschke-Santal\'o inequality for unconditional functions. For $j=1$, this is a theorem of Huang and Li \cite{huang_li}. The proof we give is inspired by, and largely the same as the proof of the functional Blaschke-Santal\'o inequality of Lehec \cite{lehec}. By Proposition \ref{prop:centro_affine_main}, the centro-affine surface area of $K$ can be viewed as a regularization of the total Holmes-Thompson volume of $K$ equipped with its Funk metric. It excels at capturing the volume growth of the metric, but has the drawback of being dependent on the precise way in which the volume is exhausted, namely through metric balls centered at a point. As a consequence, it depends on the choice of a center point; worse yet, it is not a projective invariant, defying the expectations one might entertain in light of Theorem \ref{main:projective_invariants} i). In an attempt to remedy this situation, in section \ref{sec:projective_total_volume} we propose a different regularization of the total volume, which turns out to be a projective invariant of the body alone. For simplicity we focus on the projective plane, although it is likely that a similar procedure can be carried out in greater generality. To state precisely, we use the standard Euclidean structure on $\mathbb{R}^{3}$, and locally identify $\mathbb{R}\mathbb{P}^2$ with $S^2$. Given $\Omega\subset \inter(K)$ in $S^2$, it then holds that $$\vol^F_K(\Omega)=\frac{1}{\omega_n}\int_{\Omega\times K^\vee}\frac{dxd\xi}{\langle x,\xi\rangle^3},$$ where $K^\vee\subset S^2$ is the polar set given by $K^\vee=\{\xi\in S^2:\forall x\in K, \langle x,\xi\rangle\geq 0\}$. For a convex set $K\subset S^2$ we define, borrowing a term from \cite{brylinski}, the \emph{Beta function of $K$} by the integral $$B_K(z)=\int_{K\times K^\vee}\langle x,\xi\rangle^zdxd\xi, \quad \Re z>-1.$$ \begin{MainTheorem}\label{main:beta} Let $K\subset S^2$ be $C^2$-smooth and strictly convex. Then $B_K(z)$ extends as a meromorphic function with simple poles contained in $\{-\frac52,-\frac72,-\frac 92,\dots\}$. Moreover, the value $B_K(-3)$ is a projective invariant of $K\subset\mathbb{R}\mathbb{P}^2$. \end{MainTheorem} While the affinity to Funk volume is transparent, an explicit description of $B_K(-3)$ in terms of the Funk metric on $K$ remains to be found. The regularized total volume was inspired by the theory of M\"obius energy of knots \cite{ohara_knots}, and particularly its extensions to linking and Riesz energy \cite{ohara_solanes_linking, ohara_solanes_riesz}. \subsection{Plan of the paper} In section \ref{sec:preliminaries} we recall the basics of the various geometries that we use, mostly to fix notation. The rest of the paper consists of three main parts, namely sections \ref{sec:outlook}-\ref{sec:girth}, \ref{sec:every_scale}-\ref{sec:inequalities} and \ref{sec:projective_total_volume}, that are largely independent of each other. In Section \ref{sec:outlook} we dwell on the projective nature of the Funk metric, and its transformation under projective maps. Definition \ref{def:conomadic} is key. We establish the duality property of the Funk volume element, proving part i) of Theorems \ref{main:projective_invariants} and \ref{main:duality}, and Corollary \ref{cor:hilbert_volume}. In section \ref{sec:billiards} we discuss general Funk billiards and prove their projective invariance and duality properties, namely part ii) of Theorems \ref{main:projective_invariants} and \ref{main:duality}. Then in section \ref{sec:conics} we focus on Funk billiards inside an ellipse, which is just the standard hyperbolic billiard, and prove Theorem \ref{main:conics}. In section \ref{sec:girth}, we establish the Funk analogue of Sch\"affer's dual girth conjecture, completing the proof of Theorems \ref{main:projective_invariants} and \ref{main:duality}. In section \ref{sec:every_scale} we begin the study of the volume of balls in Funk geometry, establishing Proposition \ref{prop:centro_affine_main}. Then in section \ref{sec:inequalities} we prove the various inequalities, namely Theorems \ref{main:bodies_inequality},\ref{main:inequality_functional} and Corollary \ref{cor:entropy}. Section \ref{sec:projective_total_volume} is dedicated to the proof of Theorem \ref{main:beta}. \subsection{Acknowledgements} This project was born out of many fruitful discussions with Alina Stancu, whose interest and encouragement made this work possible. I greatly benefited from discussions with, and ideas contributed by Shiri Artstein-Avidan, Yaron Ostrover, Bo'az Klartag and Gil Solanes, to whom much gratitude and appreciation are extended. Thanks are also due to Constantin Vernicos for several useful comments on a first draft of the paper. The project got started during the author's stay in Montreal as a CRM-ISM postdoctoral fellow; the support provided by those institutes, as well as the excellent working atmospheres of UdeM, Concordia and McGill Universities, are gratefully acknowledged. \section{Preliminaries}\label{sec:preliminaries} \subsection{Convexity} A convex body $K\subset V=\mathbb{R}^n$ is a compact convex set, which we will henceforth assume to have non-empty interior. We write $|K|$ for the Euclidean volume of $K$, and $\mathcal H^{n-1}$ for the Hausdorff measure on $\partial K$. The Euclidean unit ball is $B^n$, and $\omega_n:=|B^n|=\frac{\pi^{\frac n 2}}{\Gamma(\frac n 2 +1)}$. The support function of $K\subset V$ is $h_K(\xi)=\sup_{x\in K}\langle x,\xi\rangle:V^*\to \mathbb{R}$, which is a convex function. We say that $K$ is smooth and strictly convex if $\partial K$ is $C^2$-smooth of strictly positive gaussian curvature. By $\mathcal K_0(V)$ we denote the class of convex bodies with $0\in\inter(K)$. For $K\in\mathcal K_0(V)$, the \emph{Minkowski functional} of $K$ is $\|x\|_K=\inf\{t>0: \frac{x}{t}\in K\}$. If $K=-K$, it is a norm for which $K$ is the unit ball. The cone measure on $\partial K$ is denoted $\mu_K$, and has $\int_K d\mu_K=n|K|$. If $\partial K$ is $C^1$ smooth, it is given by $d\mu_K(x)=\langle x, \nu_x\rangle d\mathcal H^{n-1}(x)$, where $\nu_x$ is the outward unit normal. The \emph{polar} (or \emph{dual}) convex body is $K^o=\{\xi\in V^*: \langle \xi,x\rangle\leq 1\}\in\mathcal K_0(V^*)$. It holds that $\|\xi\|_{K^o}=h_K(\xi)$. For $z\in \inter(K)$, we write $K^z=(K-z)^o$. For a Borel function $\phi:V\to \mathbb{R}\cup\{+\infty\}$, its \emph{Legendre transform} is \begin{equation}\label{eq:legendre}\mathcal L\phi(\xi)=\sup_{x\in\mathbb{R}^n} (\langle x,\xi\rangle -\phi(x)): V^*\to \mathbb R\cup\{+\infty\}. \end{equation} If $\phi$ is defined on a subset of $V$, we first extend it by $+\infty$. The function $\mathcal L\phi$ is always convex, and if $\phi$ is convex then $\mathcal L^2\phi=\phi$. An easy computation shows that $\mathcal L(\frac1p\|x\|_K^p)=\frac1q\|\xi\|_{K^o}^q$ for $p^{-1}+q^{-1}=1$. This includes the case $p=\infty, q=1$ which reads $\mathcal L(-\log \mathbbm 1_K)(\xi)=\|\xi\|_{K^o}=h_K(\xi)$. Writing $\phi_z=\phi(z+\bullet)$, one immediately finds that $$\mathcal L\phi_z(\xi)=\sup_{x} (\langle x,\xi\rangle-\phi(x+z))=\mathcal L\phi(\xi)-\langle z,\xi\rangle.$$ For much more information on convexity, we refer to \cite{schneider_book}. \subsection{Projective geometry} Denote $V=\mathbb{R}^{n+1}$. We often write $\mathbb{P}=\mathbb{R}\mathbb{P}^n=\mathbb{P}(V)$ for the $n$-dimensional real projective space when the dimension is clear from context, and $\mathbb{P}^\vee=\mathbb{P}(V^*)$ for the dual projective space. The group of isomorphisms of the projective space is $\PGL(n+1)=\GL(n+1)/\mathbb{R}^*$, and its elements are the invertible projective maps, also known as homographies. Projective duality establishes a bijection between the points of $\mathbb P$ and the hyperplanes of $\mathbb P^\vee$, by assigning to a point $x\in\mathbb P$ the hyperplane $\mathbb P(x^\perp)\subset\mathbb{P}^\vee$, also denoted $x^\perp$. This allows to identify $\mathbb{P}^\vee$ with the set of hyperplanes of $\mathbb{P}$, and vice versa. A \emph{convex body} $K\subset\mathbb{P}$ is the image of a closed proper convex cone in $\mathbb{R}^{n+1}$, namely one that does not contain a line. Equivalently, for any hyperplane $H\subset\mathbb{P}\setminus K$, $K\subset\mathbb{P}\setminus H=\mathbb{R}^n$ is a convex body. We will often use the same notation for both the body and the cone it defines in $V$. The \emph{polar} (or \emph{dual}) convex body is $K^\vee\subset\mathbb{P}^\vee$, which is the closure of the set of all hyperplanes disjoint from $K$. $K^\vee$ is a convex body, and if $K$ is smooth and strictly convex, then so is $K^\vee$. The \emph{normal map}, also called the \emph{Legendre transform}, is the bijection $\mathcal L_K:\partial K\to\partial K^\vee$, given by $x\mapsto T_x\partial K$. The linear and projective polar bodies are closely related, see Lemma \ref{lem:two_dualities}. Let $a, b, c, d$ lie on a line in $\mathbb{P}$, in that order. Their \emph{cross ratio} is $[a, b, c, d]:=\frac{|ac||bd|}{|ab||cd|}$. It is invariant under $\PGL(n+1)$. A \emph{pencil of hyperplanes} in $\mathbb P$ is the image under projective duality of a line in $\mathbb{P}^\vee$. The cross ratio of four hyperplanes on a pencil is defined as their cross ratio in $\mathbb{P}^\vee$. It can be computed by intersecting the hyperplanes with a generic line, and taking the cross ratio of the respective intersection points. A \emph{quadric} $E=[A]\subset \mathbb{P}$ is given by the homogeneous quadratic equation $\{x: \langle A x, x\rangle=0\}$, for some symmetric matrix $A=A_E$. A \emph{linear pencil} of quadrics is any family of quadrics of the form $[tA_1+sA_2]$, $t, s\in\mathbb{R}$. A \emph{dual pencil} is a family of the form $[(tA_1^{-1}+sA_2^{-1})^{-1}]$, $t, s\in\mathbb{R}$. Given two non-degenerate quadrics $E_1, E_2\in\mathbb R\mathbb{P}^n$, the projective map $A=E_1^{-1}E_2\in\PGL(n+1)$ is defined as follows. Let $Q_1, Q_2$ be quadratic forms on $\mathbb{R}^{n+1}$ such that $E_j=\{Q_j=0\}$, which are uniquely defined up to a scalar multiple. Then $A$ is represented by the endomorphism $\overline A$ given by setting $Q_2(x,y)=Q_1(\overline Ax,y)$ for all $x,y\in\mathbb{R}^{n+1}$. \begin{Definition}\label{def:harmonic_quadrics} Four quadrics $Q_1,Q_2,Q_3,Q_4\subset\mathbb{P}$ form a \emph{ nic quadruplet} if $$Q_2^{-1}Q_1=Q_4^{-1}Q_3.$$ \end{Definition} We refer to e.g. \cite{projective_book} for an overview to projective geometry, and to \cite{complex_convexity} for an introduction to convexity in projective geometry. \subsection{Finsler geometry} A \emph{non-reversible Finsler manifold} $(M,\phi)$ is a smooth manifold equipped with a function $\phi :TM\setminus \underline 0\to [0,\infty)$, which restricts to a non-symmetric norm on each tangent space. The assumed smoothness of $\phi$ depends on the problem at hand. The tangent unit ball at $x$ is $B_xM=\{v\in T_xM: \phi(x,v)\leq 1\}$, and the cotangent ball is $B_x^*M=(B_xM)^o\subset T_x^*M$. The co-ball bundle is $B^*M=\cup_{x\in M}B_x^*M$. The \emph{Holmes-Thompson} volume is the measure $\mu^{HT}:=\frac{1}{\omega_n}\pi_*(\vol_{2n}|_{B^*M})$, where $\vol_{2n}$ is the symplectic volume on $T^*M$, and $\pi:T^*M\to M$ the natural projection. By $\Dens(V)$ we denote the one-dimensional real line of Lebesgue measures, or \emph{densities}, on $V$. A measure on a manifold can be identified with a section of the line bundle $\Dens(TM)$. The Holmes-Thompson measure has $\mu^{HT}_x(B_xM)=\frac{1}{\omega_n}|B_xM||B^*_xM|$. For an illuminating discussion of the Holmes-Thompson volume in Finsler geometry, which plays a central role in this work, we refer to \cite{alvarez_thompson, alvarez_berck}. Both Hilbert and Funk geometries are Finsler, and we refer to \cite{hilbert_handbook} for a comprehensive account. Let us quote a few facts we will use. The Funk metric of $K\subset\mathbb{R}^n$ is given by the Finsler norm $\phi(x,v)=\|v\|_{K-x}$. The \emph{outward ball} in the Funk geometry of $K\subset\mathbb{R}^n$, of radius $r$ and centered at $q$, is the set $$B_r^F(q, K)=\{x\in \inter(K): d^F_K(q,x)\leq r.\}$$ Extrinsically, $B^F_r(q, K)=q+(1-e^{-r})(K-q)$. Given a compact domain $\Omega\subset \inter(K)$ with piecewise smooth boundary, one may, following \cite{gutkin_tabachnikov}, consider the Funk billiard map inside $\Omega$ by requiring that whenever $x,y,z\in\partial \Omega$ are consecutive points in a billiard orbit, then $$\frac{\partial}{\partial y}(d^F(x,y)+d^F(y,z))=0.$$ \section{A projective outlook on the Funk metric} \label{sec:outlook} \subsection{The projective co-nomadic Finsler structure} We start with some terminology. \begin{Definition}\label{def:conomadic} Two Finsler structures $F_1$, $F_2$ on $M$ are \emph{co-translate} if $F_2-F_1$ is a $1$-form on $M$. A \emph{co-nomadic Finsler structure} on a manifold is an equivalence class of co-translate Finsler structures. { Similarly, we define an \emph{exact co-nomadic Finsler structure} to be the equivalence class of Finsler norms up to an exact $1$-form on $M$. } \end{Definition} Geometrically, a co-nomadic Finsler structure is the data of all cotangent unit balls, fixed up to translation in each cotangent space. Given a co-nomadic Finsler structure $[F]$, represented by a Finsler norm $F$, its symmetrization $F^S(v):=\frac{1}2(F(v)+F(-v))$ is clearly a reversible Finsler metric that is independent of $F$. The Holmes-Thompson volume element of a co-nomadic Finsler structure is similarly well-defined. { Furthermore, if $N\subset M$ is a $k$-dimensional submanifold, it inherits a co-nomadic Finsler structure, in particular its $k$-dimensional Holmes-Thompson volume is well-defined.} { \begin{Lemma}\label{lem:exact_property} Given an \emph{exact} co-nomadic Finsler structure $[F]$ on a manifold $M$, represented by the Finsler norm $F$, and an oriented closed curve, its length only depends on $[F]$. Moreover, the geodesics of $F$ only depend on $[F]$. \end{Lemma} \proof The first statement is clear. For the second, recall that a geodesic is any curve that locally extremizes the length functional of a curve with fixed endpoints. Modifying the norm by an exact $1$-form simply adds a constant to the length functional. \endproof } Take $x\in \inter(K)$, and consider the cone $K^\vee\subset V^*$. For any non-zero $\xi\in V^*$, set $B_\xi:=K^\vee\cap (x^\perp +\xi)-\xi\subset x^\perp$. It is a convex body (projectively equivalent to $K^\vee$). Furthermore, $B_{t\xi}=tB_{\xi}$ for all $t\neq 0$. Finally if $\xi=\xi_1-\xi_2\in x^\perp$, we get $B_{\xi_2}=B_{\xi_1}+\xi$. Consequently, $B_\bullet$ defines a convex set in $(V^*/x^\perp)^*\otimes x^\perp=T^*_x\mathbb{P}$, up to translation. Let us call it the \emph{projective co-nomadic Finsler structure on $\inter(K)$}. To obtain a true Finsler structure, one can proceed in several ways. \begin{enumerate} \item Fix a Euclidean structure on $V$. One can then take $\xi=x^\perp$ and obtain the fixed cotangent ball $K^\vee\cap (x^\perp + x)-x\subset T^*_x\mathbb{P}=T_x\mathbb{P}=x^\perp$. We call it the orthogonal Finsler structure. \item Fix $\eta\in \inter(K^\vee)$, and take $B_x^*K:=K^\vee\cap (x^\perp +\eta_0)-\eta_0$ for any $\eta_0\in\eta\subset V^*$. As we will see, this is simply the reverse Funk metric, see Proposition \ref{prop:projective_funk}. \item Fix an affine-equivariant point selector $S$ in the interior of convex bodies (which need only be defined on projective images of $K^\vee$), such as the center of mass or Santal\'o point. We then get a projectively invariant construction of a Finsler metric on $\inter(K)$ which is co-translate to the reverse Funk metric. \end{enumerate} Given an extra input $\alpha$ from the above list, we will say that the projective co-nomadic Finsler structure is \emph{anchored} by it. The following lemma relates the notions of linear and projective polarity. It assumes $V=\mathbb{R}^{n+1}$ is equipped with the standard Euclidean structure, and so all linear spaces are identified with their duals. \begin{Lemma}\label{lem:two_dualities} Identify a convex body $K\subset \mathbb{R}\mathbb{P}^n$ with the cone $\tilde K\subset V=\mathbb{R}^{n+1}$, and with $K_t=\tilde K\cap \{x_{n+1}=t\}$. For $x\in V\setminus \{0\}$, write $[x]=\mathbb{R} x\in \mathbb{R}\mathbb{P}^n$. Then for $\hat p=(p,1)\in \inter(K_1)\subset\mathbb{R}^n$, the set $-(K_1-\hat p)^o\subset \mathbb{R}^n=\{x_{n+1}=0\}$ coincides with $\pi_n(K^\vee\cap ([\hat p]^\perp+e_{n+1}))$, where $\pi_n$ projects orthogonally to $\mathbb{R}^n$. \end{Lemma} \proof Consider $(-y,z)\in \mathbb{R}^n\oplus \mathbb{R}$ which lies in $K^\vee\cap ([\hat p]^\perp+e_{n+1})$. Then for all $(\kappa,1)\in K_1$ we have $\langle (-y, z-1), (p,1)\rangle=\langle -y,p\rangle +(z-1)=0$, and $\langle (-y,z),(\kappa,1)\rangle =\langle -y,\kappa\rangle +z\geq 0$. Thus $z=1+\langle y,p\rangle$, and $\langle y,\kappa-p\rangle \leq 1$. That is, $\pi_n(K^\vee\cap ([\hat p]^\perp+e_{n+1}))\subset \{(-y,1): y\in (K_1-\hat p)^o\}$. For the opposite inclusion, start with $(y,1)\in (K_1-\hat p)^o$, and verify that $(-y,1+\langle y, p\rangle )\in K^\vee\cap ([\hat p]^\perp+e_{n+1})$. \endproof In particular if $K\subset \mathbb{R}\mathbb{P}^n$ is given in the affine chart $\{x_{n+1}=1\}=\mathbb{R}^n+e_{n+1}$ by $K_1\subset \mathbb{R}^n$, then $K^\vee\cap \{x_{n+1}=1\}=-K_1^o+e_{n+1}$. \begin{Proposition}\label{prop:projective_funk} Let $K\subset \mathbb{R}\mathbb P^n$ be a convex set, and fix $\eta\in \inter(K^\vee)$. The $\eta$-anchored projective Finsler metric then coincides with $\RFunk_K^{\eta}$. \end{Proposition} \proof Using a Euclidean structure to identify $T_x\mathbb P(V)=T_x^*\mathbb P(V)=x^\perp$, the cotangent ball of the Euclidean-anchored Finsler metric is $K^\vee\cap (x^\perp +x)$, where $x\in S^n$ is identified with $x\in \mathbb P(V)$. Assume $\eta=e_{n+1}$. The affine chart is $\{x_{n+1}=1\}$, where the Funk cotangent unit ball at $\hat p =(p,1)=\frac{x}{x_{n+1}}=:A(x)$ is $(K_1-p)^o\subset e_{n+1}^\perp=T_p^*K_1]$. The identification between the two tangent spaces is the differential at $x$ of the map $A:S^{n}\to \{x_{n+1}=1\}$, namely $d_xA(v)=\frac{v}{x_{n+1}}-\frac{x}{x_{n+1}^2}v_{n+1}: x^\perp \to e_{n+1}^\perp$. The dual map is $$d_xA^*:e_{n+1}^\perp \to x^\perp,\quad d_xA^*(u)=\frac{u}{x_{n+1}}-\frac{1}{x_{n+1}^2}\langle x,u\rangle e_{n+1}. $$ That is, $d_xA^*$ is a projection to $x^\perp$ parallel to $e_{n+1}$, followed by a $\frac{1}{x_{n+1}}$- homothety. Thus the inverse map $(d_xA^*)^{-1}$ is just the orthogonal projection on $e_{n+1}^\perp$, followed by an $x_{n+1}$-homothety: $$(d_xA^*)^{-1}=x_{n+1}\pi_n.$$ Observe that $K^\vee\cap (x^\perp +e_{n+1})=x_{n+1}(K^\vee\cap (x^\perp +x))$. The image of the $\eta$-anchored cotangent ball at $x$, viewed inside $x^\perp=T_x^*\mathbb{R}\mathbb P^n$, is therefore $$(x^\perp +x)\cap K^\vee -\frac{e_{n+1}}{x_{n+1}} = (x^\perp +\frac{e_{n+1}}{x_{n+1}})\cap K^\vee -\frac{e_{n+1}}{x_{n+1}}, $$ which is mapped by $(d_xA^*)^{-1}$ to $$ x_{n+1}\pi_n ((x^\perp +\frac{e_{n+1}}{x_{n+1}})\cap K^\vee ) =\pi _n((x^\perp+e_{n+1})\cap K^\vee)= -(K_1-\hat p)^o$$ using Lemma \ref{lem:two_dualities}. \endproof In light of the discussion following Definition \ref{def:conomadic}, Theorem \ref{main:projective_invariants} part i) follows. \begin{Corollary} The symmetrization of the projective co-nomadic Finsler metric is the Hilbert metric. \end{Corollary} \proof Since the Hilbert metric is the symmetrization of the Funk metric, this follows immediately from Proposition \ref{prop:projective_funk} \endproof \begin{Remark} Thus we obtain a construction of the Hilbert metric which is transparently projectively invariant and Finslerian at the same time. An equivalent description can be found in \cite[section 3]{vernicos_yang}. \end{Remark} A careful examination reveals that the Funk metric itself is projectively invariant, up to the addition of an \emph{exact} 1-form. \begin{Proposition}\label{prop:funk_exact} Fix $\eta, \theta\in \inter(K^\vee)$, and let $F_\eta, F_\theta$ be the corresponding Funk metrics on $\inter(K)$. Then $F_\theta-F_\eta$ is an exact 1-form on $\inter(K)$. The same holds for the reverse Funk metric. \end{Proposition} \proof The statements for the Funk and reverse Funk metrics are trivially equivalent. Assume $\eta=e_{n+1}$, and use a Euclidean structure to identify $T_x^*K$ with $x^\perp$. Examining the proof of Proposition \ref{prop:projective_funk} and using the notation therein, the corresponding cotangent balls differ by a shift of $w=\frac{\theta}{\langle x,\theta\rangle}-\frac{e_{n+1}}{x_{n+1}}$. To represent this translation in the fixed affine hyperplane $\{y_{n+1}=1\}$, we put $\hat y=(y,1)$ , $x=\frac{\hat y}{|\hat y|}$, $\theta=(\alpha, \beta)\in \mathbb{R}^n\oplus \mathbb{R}$. Now apply $(d_xA^*)^{-1}=x_{n+1}\pi_n$ to $w$ to get $$s(y):=(d_xA^*)^{-1}w=\frac{x_{n+1}} {\langle x,\theta\rangle}\pi_n(\theta)=\frac{1}{\langle y, \alpha\rangle +\beta}\alpha.$$ Considered as a $1$-form, $s(y)$ is exact: \[ s(y)=d\log(\langle y,\alpha\rangle+\beta),\] concluding the proof. \endproof { Thus the exact co-nomadic class of the Funk metric is projectively invariant.} \subsection{The volume in Funk geometry}\label{sec:volume_funk} For a convex set $K\subset \mathbb{P}$, let us construct a smooth measure $\widetilde \mu$ on $\inter(K)$. Given $x\in \inter(K)$, choose a density $d_x=\alpha_x\otimes\nu_x\in \Dens(T_x\mathbb{P})=\Dens(x^*\otimes V/x)=\Dens(V^*/x^\perp)\otimes \Dens((x^\perp)^*)$. Then $\nu_x^*\in \Dens(x^\perp)$ is defined by having $\nu_x^*\otimes\nu_x$ the symplectic (Liouville) volume on $x^\perp\oplus (x^\perp)^*$. Choose $\xi\in V^*/x^\perp$ with $\alpha_x(\xi)=1$, and consider the intersection in $V^*$ of the cone $K^\vee$ with the affine subspace $x^\perp+\xi$. We evaluate $c(x)=\nu_x^*(K^\vee\cap (x^\perp +\xi))$ and set $\widetilde\mu_x:=c(x)d_x\in\Dens(T_x\mathbb{P})$. It is easily verified that the choice of $\alpha_x,\nu_x$ does not matter. Thus $\widetilde\mu$ is well-defined in $\inter(K)$. \begin{Lemma}\label{lem:holmes_thompson} $\frac{1}{\omega_n}\widetilde\mu$ is the Holmes-Thompson volume of the projective co-nomadic Finsler structure inside $K$. \end{Lemma} \proof Immediate from the construction. \endproof Next we construct a measure on $\mathbb P\times \mathbb P^\vee$, which is analogous to the symplectic volume on $V\oplus V^*$. The group of projective automorphisms $\PGL(V)$ acts on $\mathbb{P}^\vee$ by $g(\xi):=g^{-*}\xi$. It holds that \begin{align*} \Dens(T_{x,\xi}(\mathbb{P}\times \mathbb{P}^\vee))&=\Dens(x^*\otimes V/x\oplus \xi^*\otimes V^*/\xi)\\&=\Dens^*(x)^{n+1}\otimes\Dens^*(\xi)^{n+1}\otimes \Dens(V\oplus V^*)\\&=\Dens^*(x)^{n+1}\otimes\Dens^*(\xi)^{n+1}=\Dens^*(x\oplus \xi)^{n+1}, \end{align*} where all equalities are equviariant for the stablizer of $(x,\xi)$ in $\PGL(V)$. Denote the incidence manifold $\mathcal Z=\{(x,\xi): x\perp \xi \}\subset \mathbb P\times \mathbb P^\vee$. We thus arrive at \begin{Proposition}\label{prop:invariant_measure} There is a one-dimensional space of projective-invariant measures on $\mathbb P\times \mathbb P^\vee\setminus \mathcal Z$, with canonic normalization. Using the standard Euclidean structure on $\mathbb{R}^{n+1}$ to identify $\mathbb{P}$ and $\mathbb{P}^\vee$ locally with the unit sphere, it is given by \[d\mu(x,\xi)= |\langle x,\xi \rangle| ^{-(n+1)}d\sigma_x d\sigma_\xi\] where $\sigma_x,\sigma_\xi$ are the standard rotationally-invariant measures on each sphere. \end{Proposition} \begin{Lemma}\label{lem:two_measures} For a measurable set $A\subset \inter(K)$, it holds that $\widetilde\mu(A)=\mu(A\times K^\vee)$. \end{Lemma} \proof We have $$\mu(A\times K^\vee)=\int_{A}dx\int_{K^\vee}\frac{d\xi}{|\langle x,\xi\rangle|^{n+1}}.$$ Next we use the Euclidean structure for all choices in the definition of $\widetilde\mu$. Assume $x=e_{n+1}$, so $T_x\mathbb P=T_x^*\mathbb P=e_{n+1}^\perp$. Choose $d_x=dx$, $\alpha_x=x$, so that $\xi=x$, and $\nu_x$, $\nu_x^*$ are both the Euclidean volume. So it remains to check that $$\int_{K^\vee\cap S^n}\frac{d\xi}{|\xi_{n+1}|^{n+1}}=|K^\vee\cap (e_{n+1}+e_{n+1}^\perp)|.$$ As the jacobian of the map \begin{equation}\label{eq:projection} S^n\to e_{n+1}+e_{n+1}^\perp,\quad \xi\mapsto \xi/\xi_{n+1}\end{equation} is $1/|\xi_{n+1}|^{n+1}$, the claim follows. \endproof \begin{Corollary}\label{cor:dual_volumes} The Holmes-Thompson volume of the Funk metric in the interior of $L$, denoted $\vol^F_L$, is a projective invariant of $L$, that is independent of a choice of a hyperplane at infinity. Furthermore, if $K\subset\inter(L)$ is convex, there is a duality of Funk volumes: $\vol_L^F(K)=\vol_{K^\vee}^F(L^\vee)$. \end{Corollary} \proof The first part follows from Proposition \ref{prop:projective_funk} and the paragraph after Definition \ref{def:conomadic}. The second is immediate from Lemmas \ref{lem:holmes_thompson} and \ref{lem:two_measures}. \endproof Let $\vol^H_K$ denote the Holmes-Thompson volume of the Hilbert metric in $\inter(K)$. \begin{Corollary}\label{cor:hilbert_volume_sec3} Given $K\subset \inter(L)$ in $\mathbb{R}\mathbb{P}^n$, $\vol^H_L(K)\leq\frac{1}{2^n}{2n\choose n} \vol_{K^\vee}^H(L^\vee)$. Equality is uniquely attained when $L$ is a simplex, and $K$ is an ellipsoid. \end{Corollary} \proof Choose an affine chart and assume $0\in\inter(K)$. By the Rogers-Shephard inequality and the duality of volumes, \begin{align*}\vol^H_L(K)=\frac{1}{\omega_n}\int_K|\frac12 (L^z-L^z)|dz\leq \frac1{2^n\omega_n}{2n\choose n}\int_K |L^z|dz&=\frac1{2^n}{2n\choose n}\vol^F_{L}(K)\\&=\frac1{2^n}{2n\choose n}\vol^F_{K^o}(L^o).\end{align*} On the other hand by the Brunn-Minkowski inequality, $$\vol^F_{K^o}(L^o)=\frac{1}{\omega_n}\int_{L^o}|(K^o)^z|dz\leq \frac{1}{\omega_n}\int_{L^o}|\frac12 ((K^o)^z-(K^o)^z)|dz=\vol^H_{K^o}(L^o).$$ The Rogers-Shephard inequality becomes an equality when $L$ is a simplex, while the Brunn-Minkowski inequality above gives equality for symmetric bodies, that is $(K^o)^z$ must have antipodal symmetry (with respect to some point) for every $z\in L^o$. A \emph{projective center} of a convex body in $\mathbb{R}\mathbb{P}^n $ is a center of antipodal symmetry for the body in some affine chart containing $K$. The sets $(K^o)^z$ are projectively equivalent to $K$, and we find that the set $L^o\subset \inter(K^\vee)$ parametrizes different choices of hyperplanes $\eta^\perp$ at infinity, for which $K\subset\mathbb{R}\mathbb{P}^n\setminus \eta^\perp$ has antipodal symmetry with some center point. But that implies that every $\eta\in L^o$ is a projective center for $K^\vee\subset\mathbb{P}^\vee$. By \cite[Theorem 9-3.]{kelly_straus}, we conclude that $K^\vee$, and therefore $K$, must be a convex quadric, that is an ellipsoid. \endproof \section{Funk billiards and duality}\label{sec:billiards} Let $K\subset \inter(L)$ in $\mathbb{P}^n$ be a pair of convex bodies. Using the $L$-Funk metric on $K$, we can consider the corresponding billiard dynamics inside $K$ following \cite{gutkin_tabachnikov}, which we call the $L$-Funk billiard in $K$. As it is a non-reversible Finsler structure, some care should be taken, however the results we need go through unaltered. In particular, using \cite[Lemma 3.3]{gutkin_tabachnikov} we get the following description of the reflection law, illustrated in figure \ref{fig:reflectionla}. Assume $q_1\in\partial K$, and $v\in S_{q_1}K$ is an incoming ray. Extend the ray until its first intersection $p_0$ with $\partial L$. Let $Q_1$ be a hyperplane tangent to $\partial K$ at $q_1$, and $P_0$ the hyperplane tangent to $L$ at $p_0$. Let $z_1=\mathbb{R}\mathbb{P}^{n-2}$ be the intersection $Q_1\cap P_0$, and note that $z_1\subset\mathbb \mathbb{R}\mathbb{P}^n\setminus L$. Let $P_1$ be the other tangent hyperplane to $\partial L$ through $z_1$, and $p_1\in\partial L$ the tangency point. The vector $v'\in S_{q_1}K$ pointing towards $p_1$ is then the outgoing ray. \begin{figure}[h] \centering \includegraphics[width=0.4\linewidth]{reflectionlaw2.pdf} \caption{Funk billiard reflection law} \label{fig:reflectionla} \end{figure} \begin{Remark}\label{rem:proper} It is clear from this description that a proper Funk billiard trajectory when $K, L$ are smooth and strictly convex, will extend indefinitely, in terms of bounces, as a proper billiard trajectory. \end{Remark} One can also consider the billiard defined by the reverse Funk metric $\RFunk_L(K)$. Its billiard trajectories coincide with the time-reversed Funk billiard trajectories of $\Funk_L(K)$. The same billiard can also be considered from an outer perspective: the phase space then consists of all $(n-2)$-dimensional planes $z\subset\mathbb{P}^n$ that do not intersect $L$, and the billiard map mapping $z_0$ to $z_1$ is defined by the same diagram, which is reproduced in figure \ref{fig:outer_billiard} from the outer billiard perspective. By analogy with the affine setting, we call this the \emph{outer Funk billiard} on $L$ with geometry set by $K$. \begin{figure}[h] \centering \includegraphics[width=0.4\linewidth]{outer_billiard} \caption{Funk outer billiard reflection law} \label{fig:outer_billiard} \end{figure} \begin{exmp} Assume $K\subset B\subset\mathbb{R}^2$ are two discs centered at the origin. Then the resulting Funk billiard inside $K$ coincides with the standard Euclidean one, as illustrated in figure \ref{fig:circles}. Indeed, take $p_0,p_1\in\partial B$, and let $z$ be the intersection of the tangents to $B$ at those points. Since $\measuredangle op_0z=\measuredangle op_1z=\measuredangle oq_1z=\frac\pi 2$, the points $p_0, p_1,q_1, z,o$ all lie on the circle having $zo$ as its diameter. It follows that the incidence angle is $\measuredangle p_0q_1z=\measuredangle p_0p_1z$, while the reflected angle is $\measuredangle p_1q_1z=\measuredangle p_1p_0z$. It remains to note that $|zp_0|=|zp_1|$. \end{exmp} \begin{figure}[h] \centering \includegraphics[width=0.4\linewidth]{circles} \caption{Funk billiard in concentric discs} \label{fig:circles} \end{figure} Examining the Funk billiard law description, we see it is projectively invariant. The following provides an alternative proof of this fact, and extends the projective invariance to include the length spectrum. \begin{Proposition}\label{prop:billiard_invariant} Choose a hyperplane at infinity $\eta\in\inter(L^\vee)$. The corresponding $L$-Funk billiard reflection law in the interior of $K$ is then independent of $\eta$, and thus projectively invariant. Furthermore, the $L$-Funk length of a periodic orbit is also projectively invariant. \end{Proposition} \proof Let $d_{\eta}(x,y)$ be the $(L,\eta)$-Funk distance from $x$ to $y$. Replacing $\eta$ by $\eta'$, we have $d_{\eta'}(x,y)=d_{\eta}(x,y)+h(y)-h(x)$ for some $h=h_{\eta,\eta'}$ by Proposition \ref{prop:funk_exact}. Now let $p,r\in\partial K$ be fixed points. Then $p,q,r$ are consecutive reflection points of the billiard if $q$ is an extremal point for $d_\eta(p,q)+d_\eta(q,r)$. Since $d_{\eta'}(p,q)+d_{\eta'}(q,r)=d_\eta(p,q)+d_\eta(q,r)+h(r)-h(p)$, this condition is independent of $\eta$. Finally, if $p_0,\dots, p_{N-1}, p_N=p_0$ is a periodic orbit, then evidently $$\sum_{j=0}^{N-1} d_{\eta'}(p_j,p_{j+1})= \sum_{j=0}^{N-1} d_{\eta}(p_j,p_{j+1})+\sum_{j=0}^{N-1}(h(p_{j+1})-h(p_j))= \sum_{j=0}^{N-1} d_{\eta}(p_j,p_{j+1}),$$ concluding the proof. \endproof \begin{exmp}\label{exm:hyperbolic} Consider the Euclidean ball $B^n\subset\mathbb{R}^n$. The Funk Finsler norm in $\inter(B^n)$ is given by $$ \phi_x(v)=\frac{\sqrt{(1-|x|^2)|v|^2+\langle x,v\rangle^2}}{1-|x|^2}+\frac{\langle x,v\rangle}{1-|x|^2}.$$ The first summand is the Beltrami-Klein model of hyperbolic geometry . The second summand is the $1$-form $\frac{1}{1-|x|^2}\sum_{j=1}^nx_j dx_j$, which is exact. Consequently, for any domain $K\subset\inter(B^n)$, the $B^n$-Funk billiard in $K$ coincides with the hyperbolic billiard, and the length of periodic orbits is the same in both geometries. \end{exmp} A careful examination of the reflection law shows that it is anti-symmetric under the duality $(K,L)\leftrightarrow (L^\vee, K^\vee)$, namely a Funk billiard trajectory in one naturally defines a reverse Funk billiard trajectory in the other, which we call the dual trajectory, as follows. Let us write $Q_j=\mathcal L_K(q_j)=T_{q_j}\partial K\in\partial K^\vee$ for $q_j\in\partial K$, and similarly $P_j\in\partial L^\vee$ for $p_j\in \partial L$. We identify the phase space with a subset of $\{(q,p): q\in \partial K, p\in \partial L^\vee\}$, where $(q,p)$ represents the outgoing ray from $q$ in the direction of $p$. The billiard tranformation $(q_0,p_0)\mapsto (q_1,p_1)$ can be understood as two dual billiards unraveling simultaneously, taking alternating steps: the Funk billiard $\Funk_L(K)$, and the reverse Funk billiard $\RFunk _{K^\vee}(L^\vee)$. In $\mathbb P$, the ray arrives at $q_1\in\partial K$, and $Q_1=\mathcal L_K(q_1)$ is marked on $\partial K^\vee$. In $\mathbb P^\vee$, the interval from $Q_1$ to $P_0=\mathcal L_L(p_0)$ (which avoids $L^\vee$) is extended until its second intersection $P_1$ with $\partial L^\vee$. Now $p_1=\mathcal L_{L^\vee}(P_1)\in \partial L$ is marked, and the outgoing ray from $q_1$ is $(q_1, p_1)$. This dual dynamic is illustrated in figure \ref{fig:dual_billiards}, where the same letter is used for both a point and its dual hyperplane. \begin{figure}[h] \centering \includegraphics[width=1\linewidth]{dual_billiards2} \caption{Dual Funk - reverse Funk billiard dynamics} \label{fig:dual_billiards} \end{figure} We thus established the following. \begin{Proposition}\label{prop:duality_billiards} Let $q_0,q_1,\dots \in\partial K$ be an $L$-Funk billiard trajectory in $K$. Let $p_j\in\partial L$ be the first intersection of the extension of the oriented segment $[q_j, q_{j+1}]$ with $\partial L$, and $P_j$ the hyperplane tangent to $L$ at $p_j$. Then $P_0,P_1,\dots$ is a $K^\vee$-reverse Funk billiard trajectory in $L^\vee$. In the plane, the two orbits have equal rotation number. \end{Proposition} \begin{Remark} The stark similarity to the duality in Minkowski billiards is not without reason. In the limit when $L$ is much bigger than $K$, the Funk billiard becomes Minkowski, much like the hyperbolic plane is almost Euclidean in small scale. More precisely, take $K\subset L\subset\mathbb{R}^n$, fix $a\in L$ and set $L_R=a+R(L-a)$. Then as $R\to\infty$, the Funk billiard orbits in $K$ converge to the Minkowski billiard orbits with geometry given by $L$ centered at $a$. In particular, if $L$ is the Euclidean ball and $a$ its center, we recover the Euclidean billiard inside $K$. \end{Remark} Consider the case of the projective plane. Since the Funk billiard trajectory consists of straight segments and preserves the symplectic form on the phase space \cite[Theorem 4.3]{gutkin_tabachnikov}, Birkhoff's theorem on the existence of two $m$-periodic orbits, for any $m\geq 2$ and any rotation number, remains valid. As an example, the two $2$-periodic orbits when $K\subset L$ are two non-concentric circles appear in figure \ref{fig:2_periodic_circles}. \begin{figure}[h] \centering \includegraphics[width=0.4\linewidth]{2_periodic} \caption{2-periodic orbits in a pair of discs} \label{fig:2_periodic_circles} \end{figure} To motivate the next result, let us consider 2-periodic orbits in general. Assume that $q_0, q_1\in\partial K$ are the bounce points of a 2-periodic orbit $\gamma$ of the $L$-Funk billiard in $K$, and $p_0, p_1\in \partial L$ the corresponding intersection points, appearing on the line in the order $p_1, q_0, q_1, p_0$ as in figure \ref{fig:2_periodic_general}. Note that $\gamma$ is also a reverse Funk billiard orbit. Its Funk length is twice the Hilbert distance between the points: $$\textrm{Length}^F_L(\gamma)=\textrm{Length}^{RF}_L(\gamma)=\log \frac {|p_1q_1|}{|p_1q_0|}+\log \frac {|q_0p_0|}{|q_1p_0|}=2d^H_L(q_0, q_1).$$ \begin{figure}[h] \centering \includegraphics[width=0.48\linewidth]{2_periodic_general} \caption{2-periodic orbit} \label{fig:2_periodic_general} \end{figure} Since, $$d^H_L(q_0,q_1)=\frac12 \log [p_1, q_0, q_1, p_0]=\frac12 \log [P_1, Q_0, Q_1, P_0]=d^H_{K^\vee}(P_0, P_1),$$ it holds that dual 2-periodic orbits have equal length. In fact, the duality of billiard orbits extends to the metric realm in full generality. We now assume $K\subset \inter(L)$ are smooth, strictly convex bodies in $\mathbb{R}\mathbb{P}^n$. \begin{Theorem}\label{thm:dual_lengths} Dual periodic orbits in $\Funk_L(K)$ and $\RFunk_{K^\vee}(L^\vee)$ have equal lengths. \end{Theorem} We will need the following auxiliary fact of independent interest. \begin{Proposition}\label{prop:unique_symplecic} There is a $\GL(V)$-invariant non-degenerate 2-form $\omega$ on $\mathbb P\times\mathbb P^\vee\setminus \mathcal Z$. Such a form is unique up to constant, and moreover it is canonically normalized. Furthermore, it is a symplectic form. \end{Proposition} \proof Let $H=\Stab_{x,\xi}\subset \GL(V)$ be the stabilizer of a fixed point $(x,\xi)$, which acts on $T_x\mathbb P$ by the full linear group, and similarly on $T_\xi\mathbb P^\vee$. Observe that $$ \wedge^2 T^*_{x,\xi}(\mathbb P\times\mathbb P^\vee)=\wedge^2(x\otimes x^\perp\oplus \xi\otimes \xi^\perp)=x^{\otimes 2}\otimes \wedge^2 x^\perp\oplus \xi^{\otimes 2}\otimes \wedge^2 \xi^\perp\oplus x\otimes x^\perp \otimes \xi\otimes\xi^\perp. $$ The first and second summand evidently have no $H$-invariant elements. As for the third summand, we may write $$x\otimes x^\perp \otimes \xi\otimes\xi^\perp= (x\otimes \xi)\otimes (x^\perp\otimes\xi^\perp).$$ There is a natural identification $\theta:x\otimes\xi\to \mathbb{R}$, and a natural pairing $\eta:x^\perp\otimes\xi^\perp \to \mathbb{R}$, which is moreover non-degenerate as $(x,\xi)\notin \mathcal Z$. We then let $\omega_{x,\xi}$ correspond to $\theta\otimes\eta^{-1}$. It is immediate from the construction that $\omega$ is non-degenerate, and it is evidently canonically normalized. For uniqueness, it remains to show that any $H$-invariant non-trivial pairing $\psi:x^\perp\otimes \xi^\perp\to \mathbb{R}$ must be a multiple of $\eta^{-1}$. Writing $\xi^\perp=W$ and identifying $x^\perp=W^*$, $H$ acts on $x^\perp\otimes \xi^\perp=W^*\otimes W$ by $\GL(W)$, admitting a unique invariant line. Finally, let us verify that $\omega$ is symplectic, that is $d\omega=0$. As $d\omega\in\Omega^3(\mathbb P\times\mathbb P^\vee\setminus \mathcal Z)$ is $\GL(V)$-invariant, it suffices to show that there are no non-trivial invariant $3$-forms. We have an invariant decomposition \begin{align*} \wedge^3 T^*_{x,\xi}(\mathbb P\times\mathbb P^\vee)&=x^{\otimes 3}\otimes \wedge^3 x^\perp\oplus \xi^{\otimes 3}\otimes \wedge^3 \xi^\perp\\&\oplus x^{\otimes 2}\otimes \wedge^2 x^\perp \otimes \xi\otimes\xi^\perp\oplus x\otimes x^\perp \otimes \xi^{\otimes 2}\otimes \wedge^2\xi^\perp. \end{align*} As before, the first and second summands do not have invariant lines. Considering the third summand, we have $$ x^{\otimes 2}\otimes \wedge^2 x^\perp \otimes \xi\otimes\xi^\perp=x\otimes \wedge^2 x^\perp\otimes\xi^\perp .$$ One can choose an element $g\in H$ acting by the identity on $x^\perp$ and $\xi^\perp$, and rescaling $x$ non-trivially. Thus there are no invariant elements in the third summand, and similiarly neither in the fourth.\endproof \textit{Proof of Theorem \ref{thm:dual_lengths}.} Recall that by Proposition \ref{prop:billiard_invariant}, the Funk length of a periodic orbit is a projective invariant. Now fix $\eta\in\mathbb P^\vee$, and define $H_\eta:(\mathbb P\setminus \eta^\perp)\times \mathbb P^\vee\setminus \mathcal Z\to T^*\mathbb P$ by identifying, for any $x\notin\eta^\perp$, $x^*=\eta$, and setting \begin{equation}\label{eq:identification}H_\eta(x,\xi)=-\left(\hat\eta \mapsto (x^\perp+\hat \eta)\cap \xi -\hat\eta\right) \in \eta^*\otimes x^\perp=x\otimes x^\perp=T_x^*\mathbb P.\end{equation} For $\eta\neq \eta'$, we have in the common domain of definition the equality $$H_{\eta'}(x,\xi)=H_\eta(x,\xi)+\beta_{\eta,\eta'}(x),$$ where $\beta_{\eta,\eta'}\in\Omega^1(\mathbb P\setminus(\eta^\perp\cup\eta'^\perp))$ is a closed $1$-form by the proof of Proposition \ref{prop:funk_exact}. Let $\omega_0$ be the canonic symplectic form on $T^*\mathbb{P}$. We claim that for distinct $\eta,\eta'\in\mathbb P^\vee$, the $2$-forms $H_\eta^*\omega_0$ and $H_{\eta'}^*\omega_0$ coincide on their common domain of definition. In light of the previous paragraph, it suffices to show that $\overline\beta^*\omega_0=0$, where $\overline\beta(x,\xi)=(x,\beta(x))$, and $\beta\in \Omega^1(U)$ is a closed form, defined in a neighborhood $U\subset\mathbb P$. We may choose coordinates locally such that $\omega_0=\sum dx_i\wedge d\xi_i$, while $\beta=\sum \frac{\partial f}{\partial x_i}dx_i$, and denote $B:=D_x\beta=(\frac{\partial^2f}{\partial x_i\partial x_j})$. Choose $u,v\in T_{x,\xi}(\mathbb P\times \mathbb P^\vee)$, and write $\hat u=dx(u), \hat v=dx(v)\in T_x\mathbb P$. We have \begin{align*}\overline\beta^*\omega_0(u,v)&=\omega_0((\hat u, B\hat u), (\hat v, B\hat v))\\&=\sum_i( \hat u_i (B\hat v)_i- \hat v_i (B\hat u)_i)=\sum_{i,j}B_{ij}\hat u_i\hat v_j-\sum_{i,j}B_{ji}\hat u_i\hat v_j=0.\end{align*} It follows that as $\eta\in\mathbb P^\vee$ varies, $H_\eta^*\omega_0$ patch to a globally defined form $\omega'\in \Omega^2(\mathbb P\times\mathbb P^\vee\setminus \mathcal Z)$. Let us check that $\omega'$ is projectively invariant. First we compute that for $g\in \GL(V)$, \begin{align*} H_\eta\circ g(x,\xi)=H_\eta(gx,g^{-*}\xi)=(gx, \hat\eta\mapsto (g^{-*}x^\perp+\hat \eta)\cap g^{-*}\xi-\hat\eta)=gH_{g^*\eta}(x,\xi).\end{align*} Consequently, $$g^*\omega'=g^*H_\eta^*\omega_0= H_{g^*\eta}^* g^*\omega_0= H_{g^*\eta}^*\omega_0=\omega'.$$ It follows by Proposition \ref{prop:unique_symplecic} that $\omega'=c\omega$, for some $c\in\mathbb{R}$. { To find the value of $c$, we note that $\omega^n=\mu$ defines the Holmes-Thompson volume by Lemma \ref{lem:two_measures}, which by definition corresponds to $\omega_0^n$, and so $c^n=1$. It is also clear that $c>0$, e.g. by writing both forms explicitly using Euclidean coordinates at a point where $\xi=x$. Thus $c=1$.} Replacing $T^*\mathbb P$ with $T^*\mathbb P^\vee$, one similarly has the map $$\overline{H_\theta}:\mathbb{P}\times (\mathbb P^\vee\setminus \theta^\perp)\setminus \mathcal Z\to T^*\mathbb P^\vee, (x,\xi)\mapsto \left(\hat\theta \mapsto (\xi^\perp+\hat \theta)\cap x -\hat\theta\right) \in \theta^*\otimes \xi^\perp=\xi\otimes\xi^\perp, $$ and the various forms $\overline{H_\theta}^*\omega_0^\vee$ as $\theta$ varies patch to give $\omega''=\omega'=c'\omega$. The constant is $c'=1$, as the sign is flipped twice compared to $H_\eta$: once due to the change of order of $\mathbb{P}, \mathbb{P}^\vee$, and a second time due to the sign change in the definition of $\overline H_\theta$. Fix $\eta\in \inter(L^\vee)$, $\theta\in \inter(K)$. It holds by Proposition \ref{prop:projective_funk} that $$S^*\Funk^\eta_L(K)=H_\eta(K\times\partial L^\vee),\quad S^*\RFunk^\theta_{K^\vee}(L^\vee)=\overline{H_\theta}(\partial K\times L^\vee).$$ Consider now $\Sigma:=\partial (K\times L^\vee)\subset\mathbb{P}\times\mathbb{P}^\vee\setminus\mathcal Z$. Denote by $\alpha_0, \overline\alpha_0$ the Liouville 1-forms on $T^*\mathbb P$, resp. $T^*\mathbb{P}^\vee$. Define $\alpha_\eta=H_\eta^*\alpha_0$, $\overline\alpha_\theta=\overline{H_\theta}^*\overline\alpha_0$. It holds that $d\alpha_\eta-d\overline \alpha_\theta=\omega-\omega=0$, and so $\alpha_\eta-\overline \alpha_\theta=df$ for some function $f$ defined near $\Sigma$. A dual pair of billiard trajectories corresponds to a trajectory on $\Sigma$, as we now describe. It first follows an integral curve of $\Ker \omega|_{K\times \partial L^\vee}$ while on $\Sigma_1:=\inter(K)\times \partial L^\vee$, and of $\Ker\omega|_{\partial K\times L^\vee}$ while on $\Sigma_2 = \partial K\times \inter(L^\vee)$, corresponding to the straight segments of $\Funk_L(K)$ and $\RFunk_{K^\vee}(L^\vee)$, respectively. By Remark \ref{rem:proper}, it can only arrive to $\partial K\times\partial L^\vee$ transversally, and so the reflection law translates to simply switching from one integral curve to another, which similarly must also be transversal to $\partial K\times\partial L^\vee$. Such a trajectory on $\Sigma$ is called a generalized characteristic. Let a generalized characteristic $\Gamma$ correspond to a dual pair of orbits $\gamma\subset \Funk_L(K)$ and $\overline\gamma\subset\RFunk_{K^\vee}(L^\vee)$. Noting that $\alpha_\eta|_{\Gamma\cap \partial K\times L^\vee}=0$, $\overline \alpha_\theta|_{\Gamma\cap K\times\partial L^\vee}=0$, we conclude that \[\textrm{Length}(\gamma)=\int_\Gamma \alpha_\eta=\int_\Gamma \overline\alpha_\theta=\textrm{Length}(\overline\gamma).\]\qedhere \section{Integrability of Funk billiard in conics}\label{sec:conics} Recall that by Example \ref{exm:hyperbolic}, the reflection laws of the $B^n$-Funk and Beltrami-Klein hyperbolic metrics coincide. Similarly, the Funk geodesics on a manifold $M\subset \inter(B^n)$ are also those of the metric induced from hyperbolic space. In particular, if $K$ is an ellipsoid nested in $B^n$, the Funk billiard in $K$ is completely integrable \cite{veselov, tabachnikov_hyperbolic}. We now recover and refine this observation in the case of the projective plane by analyzing the Funk billiard directly. Let $K\subset \inter(B)$ be nested conics in $\mathbb{P}^2$. \begin{Theorem} Any given orbit in $\Funk_B(K)$ remains tangent to a conic $Q_{i}$, while the points of the outer orbit all lie on a conic $Q_o$. Furthermore, $Q_o$ belongs to the linear pencil of conics defined by $Q_K=\partial K,Q_B=\partial B$, while $Q_i$ belongs to the dual pencil. Moreover, the quadruplet $Q_i, Q_K, Q_B, Q_o$ is harmonic. \end{Theorem} \proof We first observe that by Proposition \ref{prop:duality_billiards}, the existence of an inner caustic implies the existence of an outer caustic and vice versa. Moreover, the outer caustic belongs to the linear pencil of $Q_K, Q_B$ if and only if the inner caustic lies on the dual pencil. Let us establish the existence of an outer caustic which belongs to the linear pencil through $Q_K$ and $Q_B$. We may assume $B$ is the unit disc. By applying a projective transformation preserving $B$, we may further assume that $Q_K$ is a centered ellipse, that is $Q_K=\{q:\langle Aq,q\rangle=1\}$. Consider a point $z_0\in B^c$, and let $z_1$ be its image under the outer Funk billiard map. Each $z_j$ belongs to a unique conic in the linear pencil, which is parametrized by $t\in \mathbb{R}$, $Q_t:=\{\langle (A+tI)v, v\rangle =1+t\}$. It remains to show that $t_1=t_2$, or equivalently \begin{equation}\label{eq:t1t2} \frac{\langle Az_0,z_0\rangle -1}{1-|z_0|^2}= \frac{\langle Az_1,z_1\rangle -1}{1-|z_1|^2}.\end{equation} Let $p_0\in Q_ B$ be the tangency point of $[z_0,z_1]$. Denoting $\|v\|_A:=\langle Av, v\rangle$, we have $$ \frac{|z_0|^2-1}{|z_1|^2-1}=\frac{|z_0-p_0|^2}{|z_1-p|^2}=\frac{\|z_0-p_0\|_A^2}{\|z_1-p_0\|_A^2}, $$ the last inequality due to the fact that $z_0,pp_0,z_1$ lie on one line. Similarly, $$ \frac{\langle Az_0,z_0\rangle -1}{\langle Az_1,z_1\rangle -1}=\frac{\|z_0-q_0\|_A^2}{\|z_1-q_1\|_A^2}.$$ Thus \eqref{eq:t1t2} becomes $$ \frac{\|z_0-q_0\|_A}{\|z_0-p_0\|_A}=\frac{\|z_1-q_1\|_A}{\|z_1-p_0\|_A}\iff \frac{\sin\measuredangle_Az_0q_0p_0}{\sin\measuredangle_Az_0p_0q_0}=\frac{\sin\measuredangle_Az_1q_1p_0}{\sin\measuredangle_Az_1p_0q_1},$$ where $\measuredangle_A$ is the angle with respect to $\|\bullet\|_A$. As $\measuredangle_Az_0p_0q_0+\measuredangle_Az_1p_0q_1=\pi$, it remains to notice that $\measuredangle_Az_0q_0p_0=\frac12 \measuredangle_A q_0oq_1=\measuredangle_Az_1q_1p_0$. \begin{figure}[h] \centering \includegraphics[width=0.48\linewidth]{conics1} \caption{Outer caustic} \label{fig:conics1} \end{figure} For the last statement, we consider a point $z\in Q_t$, tangent to $Q_B$ at $p$ and to $Q_K$ at $q$ as in figure \ref{fig:conics2}. \begin{figure}[h] \centering \includegraphics[width=0.48\linewidth]{conics2_new} \caption{Relating the inner and outer caustics} \label{fig:conics2} \end{figure} Define the conic $Q$ by $Q^{-1}:=(Q_B^{-1}Q_o)Q_K^{-1}$. To show that $Q_i=Q$, we ought to prove that $p-q$ is tangent to $Q$. Now the line through $p,q$ is represented in $\mathbb{R}^3$ by $w:=(p,1)\times (p-q,0)=(J(p-q), \det(p,p-q))$, where $J$ is the counterclockwise rotation by $\frac\pi 2$. It remains to check that $$\langle Q^{-1}w, w\rangle=0\iff \langle (I+tA^{-1})J(p-q), J(p-q)\rangle -(1+t)\det(p,p-q)^2=0.$$ Putting $t=-\frac{\langle Az, z\rangle-1}{|z|^2-1}$ and $ z=p+sJp=q+rJAq$ for some $s, r\in \mathbb{R}$, we find $$ |z|^2-1=s^2,\quad \langle Az,z\rangle =1+r^2\langle AJAq, JAq\rangle.$$ Since $\det(p,p-q)^2+\langle p,p-q\rangle ^2=|p-q|^2$, the claimed equality becomes $$ s^2\langle p,p-q\rangle^2=r^2\|JAq\|_A^2(\|J(p-q)\|_{A^{-1}}^2-\det(p,p-q)^2).$$ The left hand side can be rewrriten as \begin{align*}\langle sp, p-q\rangle^2&=\langle sJp, J(p-q)\rangle^2=\langle z-p, J(p-q)\rangle^2=\\&\langle z-q, J(p-q)\rangle^2=r^2\langle JAq, J(p-q)\rangle^2=r^2\langle Aq, p-q\rangle^2.\end{align*} It thus remains to check that whenever $|p|=1$ and $\langle Aq,q\rangle=1$, one has $$ \langle Aq, p-q\rangle^2=\|JAq\|_A^2(\|J(p-q)\|_{A^{-1}}^2-\det(q,p-q)^2).$$ We can write \begin{align*}\|J(p-q)\|_{A^{-1}}^2-\det(q,p-q)^2&=\|J(p-q)\|_{A^{-1}}^2-\langle J(p-q), q\rangle^2\\&=\|J(p-q)\|_{A^{-1}}^2-\langle Aq, J(p-q)\rangle_{A^{-1}}^2\\&=\langle J(p-q), \nu \rangle_{A^{-1}}^2\end{align*} where $\nu=\frac{Jq}{\|Jq\|_{A^{-1}}}$, since $\|Aq\|_{A^{-1}}^2=\langle A^{-1}Aq, Aq\rangle=1$ and $\langle \nu, Aq\rangle_{A^{-1}}=0$. It remains to check that $$ \|Jq\|_{A^{-1}}\langle JAq, J(p-q)\rangle=\|JAq\|_A\langle A^{-1}Jq, J(p-q)\rangle,$$ or equivalently that $$\|Jq\|_{A^{-1}}JAq=\|JAq\|_A A^{-1}Jq.$$ If $JAq=\lambda A^{-1}Jq$ then $$\lambda =\frac{\|JAq\|_A}{\|A^{-1}Jq\|_A}=\frac{\|JAq\|_A}{\|Jq\|_{A^{-1}}},$$ so it remains to check that $JA$ and $A^{-1}J$, or equivalently $AJA$ and $J$, are proportional. Both are anti-symmetric $2\times 2$ matrices, and so that must be the case. \endproof \begin{Remark} When $K$ approaches $B$, $Q_i$ and $Q_o$ become polars of one another with respect to $\partial B$. \end{Remark} \section{Sch\"affer's dual girth conjecture}\label{sec:girth} The Gutkin-Tabachnikov duality of Minkowski billiards has a continuous counterpart, namely Sch\"affer's dual girth conjecture \cite{schaeffer}, which is a theorem of \'Alvarez Paiva \cite{alvarez-paiva}. The setting of Funk geometry is no different, providing another collection of projective invariants enjoying duality. { \begin{Theorem}\label{thm:girth} Let $L$ be a smooth, strictly convex body in $\mathbb{P}^n$. \begin{enumerate} \item Let $M\subset\inter(L)$ be a compact submanifold with boundary, and consider $\Funk^\eta_L(M)$. Then its Holmes-Thompson volume, its geodesics and the length of closed geodesics are all independent of $\eta$, and thus constitute projective invariants of $(M,L)$. \item If $K\subset\inter(L)$ is strictly convex, then $\Funk_L(\partial K)$ and $\RFunk_{K^\vee}(\partial L^\vee)$ have the same Holmes-Thompson volumes and length spectra. \end{enumerate} \end{Theorem} \proof The first claim follows immediately by Proposition \ref{prop:funk_exact} and Lemma \ref{lem:exact_property}, as the Funk metric induces on $M$ a projectively-invariant exact co-nomadic Finsler structure. For the second claim, we follow closely the proof in \cite{alvarez-paiva}. Write for simplicity $\partial K$ for $\Funk^\eta_L(\partial K)$, and similarly $\partial L^\vee$ for $\RFunk^\theta_{K^\vee}(\partial L^\vee)$. Define a smooth embedding $\Phi_\eta:S^*\partial K\to \partial K\times\partial L^\vee$ as follows. Let $\pi_x: T_x^*K\to T_x^*\partial K$ be the natural projection. Recall that by Proposition \ref{prop:projective_funk}, $S^*_xK=H_\eta(x\times \partial L^\vee)$, where $H_\eta: K\times L^\vee\to T^*K$ is given by eq. \eqref{eq:identification}. Now $\pi_x(S^*_xK)=B_x^*\partial K$, and $S^*_x\partial K$ is diffeomorphic to its preimage under $\pi_x$, namely the shadow boundary $Z_x\subset S_x^* K$. Denote $\tau_x=\pi_x^{-1}:S^*_x\partial K\to Z_x$, and define $\tau_\eta:S^*\partial K\to S^*K, (x,\xi')\mapsto (x,\tau_x(\xi'))$. Set $$\Phi_\eta(x,\xi')=H_\eta^{-1}\circ\tau_\eta(x, \xi').$$ Let us describe the image of $\Phi_\eta$. It holds that $(x,\xi)\in \Image(\Phi_\eta)$ if and only if the restricted projection $\pi_x:T_\xi S^*_xK\to T_x^*\partial K$ is not an isomorphism, or equivalently if the natural pairing $T_x\partial K\otimes T_\xi\partial L^\vee\to \mathbb{R}$ is degenerate. The latter pairing is induced from the natural non-degenerate pairing $T_x\mathbb P\otimes T_\xi \mathbb P^\vee\to \mathbb{R}$, which is given by identifying $\xi=x^*$, $(x^\perp)^*=\xi^\perp$, so that $T_x \mathbb{P}=x^*\otimes (x^\perp)^*=\xi\otimes \xi^\perp=T_\xi^*\mathbb{P}^\vee$. In particular, $\Image(\Phi_\eta)$ does not depend on $\eta$. Turning to the dual manifold, define $\overline{\Phi_\theta}=\overline{H_\theta}^{-1}\circ\tau_\theta:S^*\partial L^\vee\to \partial K\times\partial L^\vee$ in full analogy with $\Phi_\eta$. Observe that the images of the two maps $\Phi_\eta$, $\overline \Phi_\theta$ coincide by their common description as the degeneracy locus of the pairing $T_x\partial K\otimes T_\xi\partial L^\vee\to \mathbb{R}$. Let $\alpha_1$ be the canonical contact form on $S^*\partial K$, and $\overline\alpha_1$ that of $S^*\partial L^\vee$. Define $\alpha_\eta =(\Phi_\eta^{-1})^*\alpha_1\in\Omega^1(\Image(\Phi_\eta))$ and $\overline\alpha_\theta =(\overline\Phi_\theta^{-1})^*\alpha_1\in\Omega^1(\Image(\Phi_\eta))$, which coincide with the restriction to $\Image(\Phi_\eta)$ of the corresponding $1$-forms on $\partial K\times\partial L^\vee$ in the proof of Theorem \ref{thm:dual_lengths}, in particular by that proof $\alpha_\eta-\overline\alpha_\theta$ is an exact form. We now have \begin{align*}\vol^{HT}(\Funk_L(\partial K))&=\int_{S^*\partial K}(d\alpha_1)^{n-2}\wedge\alpha_1=\int_{\Image(\Phi_\eta)}d\alpha_\eta^{n-2}\wedge\alpha_\eta\\&=\int_{\Image(\overline\Phi_\theta)}d\overline\alpha_\theta^{n-2}\wedge\overline\alpha_\theta=\vol^{HT}(\Funk_{K^\vee}(\partial L^\vee)).\end{align*} Similarly, the length spectrum of $\Funk^\eta_L(\partial K)$ coincides with the action spectrum of $\alpha_\eta$ on $\Image(\Phi_\eta)$, which is also the action spectrum of $\overline \alpha_\theta$. \endproof } We remark that the Holmes-Thompson volumes of $\Funk_L(\partial K)$ and $\RFunk_L(\partial K)$ trivially coincide. \textit{Proof of Corollary \ref{cor:hilbert_plane}.} This follows at once from Theorem \ref{thm:girth}, since the Hilbert length of a curve coincides with its 1-dimensional Funk Holmes-Thompson volume. \qed. \section{Mahler volumes of every Funk scale}\label{sec:every_scale} Define the \emph{Funk-Mahler volume of radius $0<r<\infty$ (around $q\in \inter(K)$)} by $$M_r(K,q)=\omega_n r^{-n}\vol_K^F(B_r^F(q, K)), \quad M_r(K)=\inf_{q\in \inter(K)}M_r(K,q).$$ In Euclidean terms, we have $$M_r(K,q)=r^{-n}\int_{z\in q+(1-e^{-r})(K-q)}|K^z|dz,$$ where $K^z$ denotes the dual body with respect to $z$. Evidently, $M_r(K,q)$ is invariant under the full linear group with $q$ at the origin. \begin{Lemma}\label{lem:mahler_convex} For any convex $K$, $M_r(K, q)$ is a strictly convex function of $q$. In particular, if $K=-K$ then $M_r(K)=M_r(K, 0)$. \end{Lemma} \proof The function $z\mapsto |K^z|$ is well-known to be strictly convex in $z\in \inter(K)$. This follows at once by writing $$|K^z|=\omega_n\int_{S^{n-1}}\max_{x\in K}\langle \theta, x-z\rangle ^{-n}d\sigma(\theta),$$ where $\sigma$ is the rotation-invariant probability measure on $S^{n-1}$, and noticing that for any fixed $x,\theta$, the function $z\mapsto (\langle x,\theta\rangle -\langle z, \theta\rangle)^{-n}$ is strictly convex. It follows that $$M_r(K,q)=r^{-n}\int_{(1-e^{-r})K}|K^{e^{-r}q+w}|dw$$ is convex as well. If $K=-K$ then $M_r(K, q)=M_r(K, -q)$, and so $q=0$ must be a minimum. \endproof The Funk-Mahler volume enjoys invariance under duality, as follows. \begin{Proposition} It holds for all $K\subset \mathbb{R}^n$ that $M_r(K, q)=M_r(K^q, q)$. In particular, if $K=-K$ then $M_r(K)=M_r(K^o)$. \end{Proposition} \proof We may assume $q=0$. Put $\rho=1-e^{-r}$, so that $B_r^F=\rho K$ is the ball of radius $r$ in the Funk metric, centered at the origin. Then $$M_r(K,0)=\frac{\omega_n}{r^n}\vol_K^F(\rho K)=\frac{\omega_n}{r^n}\vol_{\rho^{-1} K^o}^F(K^o)=M_r(\rho^{-1}K^o,0)=M_r(K^o, 0),$$ where the second equality is due to Corollary \ref{cor:dual_volumes}, the penultimate equality holds since $K^o$ is the ball of radius $r$ around the origin in the Funk geometry of $\rho^{-1}K^o$, and the last equality is due to $\GL$-invariance. The second statement now follows by Lemma \ref{lem:mahler_convex}. \endproof \begin{Remark} In projective-invariant terms, this duality assumes the following form. For $K\subset\mathbb{P}$, $q\in \inter(K)$ and $\eta\in\inter(K^\vee)$, let $B^\eta_r(K, q)$ be the outward ball of radius $r$ centered at $q$ in $\Funk_\eta(K)$. Then $B^\eta_r(K, q)\subset \Funk_\eta(K)$ and $B_r^q(K^\vee, \eta)\subset \Funk_q(K^\vee)$ have equal volume. \end{Remark} As $r\to0$, $M_r(K,q)\to M_0(K,q):=|K||K^q|$. Thus the Mahler volume $M_0(K):=\lim_{r\to 0} M_r(K)$ is recovered at infinitesimal scale. On the opposite end of the scale, the asymptotics of $M_r(K,q)$ is governed by the regularity of the boundary of $K$. When $K$ is smooth and strictly convex, we recover the centro-affine surface area as $r\to\infty$, as established in \cite{berck_bernig_vernicos}. Recall that the \emph{centro-affine surface area relative to $q\in \inter(K)$} is given by \begin{equation}\label{eq:centro_affine}\Omega_n(K, q)=\int_{\partial K}\frac{k_p^{\frac 12}}{\langle p-q, \nu_p\rangle^{\frac{n+1}{2}}}d\mu_K(p;q)=\int_{\partial K}\frac{k_p^{\frac 12}}{\langle p-q, \nu_p\rangle^{\frac{n-1}{2}}}d\mathcal H^{n-1}(p),\end{equation} where $k_p$ is the Gaussian curvature of $\partial K$ at $p$, $\nu_p$ the unit outward normal, and $\mu_K(p;q)$ the cone measure with the origin at $q$. As \cite{berck_bernig_vernicos} focuses on the Hilbert metric with the Busemann volume, we provide for the sake of completenss the details of the computation in our case. \begin{Lemma}\label{lem:volume_ellipsoid} Let $A$ be a positive-definite $n\times n$ matrix, and $Q_A=\{x\in\mathbb{R}^n:\langle Ax, x\rangle \leq 1\}$. Then for $p\in\partial Q_A$ and $0<\rho <1$, we have $$|Q_A^{\rho p}|=\omega_n\frac{(\det A)^{1/2}}{(1-\rho^2)^{\frac{n+1}{2}}}=\omega_n\frac{1}{(1-\rho^2)^{\frac{n+1}{2}}}\frac{k_p^{1/2}}{\langle p,\nu_p\rangle^{\frac{n+1}{2}}}.$$ \end{Lemma} \proof For the first equality, we compute that $Q_A^{\rho p}=\{ \xi: \langle B(\xi-c),\xi-c\rangle\leq 1\}$, where, denoting $B_0=A^{-1}-\rho^2pp^T$, we have \begin{align*}c&=\rho B_0^{-1}p, \qquad B=\frac{B_0}{1+\rho^2\langle B_0^{-1}p,p\rangle}. \end{align*} Now by the Sherman-Morrison formula, $$B_0^{-1}=A+\frac{\rho^2}{1-\rho^2}App^TA\Rightarrow \langle B_0^{-1}p,p\rangle =\frac{1}{1-\rho^2},$$ while by the matrix determinant lemma we h ave $$\det B_0 =(1-\rho^2)\det A^{-1}.$$ It follows that $$|Q_A^{\rho p}|=\omega_n\sqrt{\det B}^{-1}=\omega_n\sqrt{\frac{\det A}{1-\rho^2}}(1-\rho^2)^{-\frac{n}{2}}=\omega_n\frac{\sqrt{\det A}}{(1-\rho^2)^{\frac{n+1}{2}}}$$ For the second equality, recall that $\nu_x=\frac{Ax}{|Ax|}$. It follows that for $v\in T_pQ_A$, $$D_p\nu_x (v)=\frac{Av}{|Ap|}-\langle \frac{Av}{|Ap|}, \nu_p\rangle \nu_p=\frac{1}{|Ap|}\pi_{T_pQ_A}\circ A.$$ Take an orthonormal basis $v_1,\dots, v_{n-1}$ of $T_pQ_A$, and evaluate the volume of $A(v_1\wedge\dots\wedge v_{n-1}\wedge p)$ in two ways. We get $$|A(v_1\wedge\dots\wedge v_{n-1}\wedge p)|=\det A\cdot\langle p, \nu_p\rangle=\det(\pi_{T_pQ_A}\circ A)\cdot|Ap|.$$ Noting also that $\langle p, \nu_p\rangle=\frac{1}{|Ap|}$ since $\langle Ap, p\rangle =1$, we find $\det A=\frac{k_p}{\langle p, \nu_p\rangle^{n+1}}.$ \endproof \begin{Lemma}\label{lem:dual_volume} For $K\in\mathcal K_0(\mathbb{R}^n)$ smooth and strictly convex, and $p\in \partial K$, we have $$|K^{\rho p}|= (1-\rho)^{-\frac{n+1}{2}}\left(2^{-\frac{n+1}{2}} \omega_n \frac{k_p^{1/2}}{\langle p,\nu_p\rangle^{\frac{n+1}{2}}}+err(p)\right),$$ where $err(p)\to 0$ as $\rho\to 1$, uniformly in $p\in\partial K$. \end{Lemma} \proof Throughout the proof, $c,c'$ will be positive constants that are bounded away from $0$ and $\infty$, uniformly in $p\in\partial K$, and may change between occurrences. As both $|K^{\rho p}|$ and the main term on the right hand side are $\SL(n)$-invariant, we may assume without loss of generality that $p=-e_n$ and $T_p\partial K=\Span(e_1,\dots,e_{n-1})$. We may further assume that $y_{n}=-1+\frac {a_p} 2|y'|^2$ is the quadratic approximation of $\partial K$ near $p$, where we write $y'=(y_1,\dots,y_{n-1})$. Let $Q$ be the ellipsoid $a_p|y'|^2 +y_n^2=1$, osculating $\partial K$ at $p$. Write $1-\rho =\epsilon^2$, and note that $c\leq a_p\leq c'$. Let $\nu_x$ be the Euclidean normal at $x\in\partial K$. One has \[\|\nu_x\|_{K^{\rho p}}=h_K(\nu_x)-\langle \nu_x,\rho p\rangle=\langle x-\rho p, \nu_x\rangle. \] We may therefore parametrize the points of $K^{\rho p}$ by $$\zeta_K:[0,1]\times\partial K\to K^{\rho p}, \quad \zeta_K(\alpha, x)=\alpha\frac{\nu_x}{\langle x-\rho p, \nu_x\rangle}.$$ Then by Cauchy-Schwartz, $$\langle x-\rho p, \nu_x\rangle\geq \langle x-p, \nu_x\rangle -(1-\rho)|\langle p, \nu_x\rangle| \geq \langle x-p, \nu_x\rangle- \epsilon^2.$$ Let $H$ be the hyperplane parallel to $T_p\partial K$ at distance $\epsilon$ from it, and $H^+, H^-$ the half-spaces it defines, with $p\in H^+$. We claim that $|\zeta_K([0,1]\times \partial (K\cap H^-))|\leq c\epsilon^{-n} $. Consider $x\in\partial K\cap H^-$, so $|x-p|\geq \epsilon$. If $\nu_x$ is proportional to $x-p$ then $\langle x-p, \nu_x\rangle |x-p|\geq \epsilon$, and $|\zeta_K(\alpha, x)| \leq \frac2 {\epsilon}$. Otherwise, consider the affine plane $P=p+\Span(x-p, \nu_x)$, and consider the planar convex body $K_P=K\cap P$. Using Euclidean cooridnates $(x_1,x_2)$ in which $T_pK_P=\mathbb{R} e_1$, $Q_P=Q\cap P$ is osculated at $p$ by the parabola $x_2=-1+\frac12 a_p x_1^2$. \begin{figure}[h] \centering \includegraphics[width=0.4\linewidth]{centro_affine_2d} \caption{} \label{fig:reflectionlaw} \end{figure} We orient $\partial K$ in such a way that the tangent at $x$ makes an acute angle with $x-p$. Let $y\in H\cap \partial K_P$ belong to the oriented arc $[p,x]\subset\partial K_P$. It holds by the convexity of $K_P$ that $$\langle x-p, \nu_x\rangle\geq |x-p|\sin \measuredangle pxy=|p-y|\sin\measuredangle pyx\geq c\sqrt{\epsilon}\sin\measuredangle pyx.$$ Denoting by $v$ the tangent ray to $\partial K_P$ at $y$, we have $$c\sqrt \epsilon=\measuredangle pyz\leq \measuredangle pyx\leq \pi-\measuredangle(y-p, v)= \pi-c\sqrt \epsilon.$$ Therefore, $$\langle x-p, \nu_x\rangle\geq c\sqrt\epsilon \sin\measuredangle pyx\geq c\epsilon.$$ Thus again we find that $|\zeta_K(\alpha, x)|\leq \frac c\epsilon$. That is, $\zeta_K([0,1]\times \partial K\cap H^-)\subset \frac{c}{\epsilon}B^n$, and so \begin{equation}\label{eq:volume_bound}|\zeta_K([0,1]\times(\partial K\cap H^-))|\leq c\epsilon^{-n}.\end{equation} Define $z_K(y'), z_Q(y'):\mathbb{R}^{n-1}\to \mathbb{R}$ for $|y'|<c\sqrt{\epsilon}$ by having $x_K(y'):=(y', z_K(y'))\in \partial K, x_Q(y'):=(y', z_Q(y'))\in \partial Q$, and let $\nu_K(y'), \nu_Q(y')$ be the correspondning unit normals. Then $z_K(y')=z_Q(y')+O(|y'|^3)=-1+\frac{a_p}{2}|y'|^2+O(|y'|^3)$. \begin{figure}[h] \centering \includegraphics[width=0.4\linewidth]{centro_affine} \caption{} \label{fig:reflectionlaw} \end{figure} Define the ellipsoids $Q_{\pm} = \{y: \frac12 (a_p\pm 2\epsilon^{1/3})|y'|^2+y_n^2=1\}.$ It holds that $z_{Q_\pm}(y')=-1+\frac{1}{2}(a_p\pm 2\epsilon^{1/3})|y'|^2+O(|y'|^3),$ and so for $\epsilon\ll1$, $$Q_+\cap H^+\subset K\cap H^+\subset Q_-\cap H^+.$$ Moreover, as $|\nu_K(y')-\nu_Q(y')|=O(|y'|)$ and $|y'|\leq c \epsilon^{1/2}\ll\epsilon^{1/3}$, it is evident that $$\zeta_{Q_-}([0,1]\times Q_-\cap H^+)\subset \zeta_{K}([0,1]\times K\cap H^+)\subset \zeta_{Q_+}([0,1]\times Q_+\cap H^+).$$ By estimate \eqref{eq:volume_bound} applied to $Q_\pm$ and combined with Lemma \ref{lem:volume_ellipsoid}, we have $$ \zeta_{Q_\pm}([0,1]\times Q_\pm\cap H^+)=\epsilon^{-(n+1)}\left( 2^{-\frac{n+1}{2}} \omega_n \frac{k_p^{1/2}}{\langle p,\nu_p\rangle^{\frac{n+1}{2}}}+err_\pm(p)\right), $$ and applying \eqref{eq:volume_bound} for $K$, the claim follows. \endproof \begin{Proposition}\label{prop:centro_affine} For $q\in \inter(K)$, $$\lim_{R\to \infty}\frac{\vol^F_K(B_R^F(q, K))}{e^{\frac{n-1}{2}R}}=\frac{1}{n-1}2^{-\frac{n-1}{2}} \Omega_n(K,q).$$ \end{Proposition} \proof We may put $q=0$. Parametrize the outward Funk ball of radius $R$ around $0$ by $p\in\partial K$, $0\leq r\leq R$ using $\phi(p,r)=(1-e^{-r})p$. Then $$ \vol^F_K(B_R^F(0,K))=\int_0^R\int_{\partial K} |K^{\phi(p,r)}|\mathrm{Jac}(\phi(p,r))d\mathcal H^{n-1}(p)dr.$$ Now $\mathrm{Jac}(\phi(p,r))=(1-e^{-r})^{n-1}e^{-r}\langle p, \nu_p\rangle$, while by Lemma \ref{lem:dual_volume} $$ |K^{\phi(p,r)}|= 2^{-\frac{n+1}{2}} \omega_n e^{\frac{n+1}{2}r} \frac{k_p^{1/2}}{\langle p,\nu_p\rangle^{\frac{n+1}{2}}}+o(e^{\frac{n+1}{2}r}),$$ where the error term is uniform in $p$. Applying l'Hospital's rule we find \begin{align*}\lim_{r\to \infty}\frac{\vol_K^F(B_r^F(0,K))}{e^{\frac{n-1}{2}r}}&=\lim_{r\to\infty}\frac{2}{n-1} \frac {(1-e^{-r})^{n-1}e^{-r}\int_{\partial K} |K^{\phi(p,r)}|\langle p, \nu_p\rangle d\mathcal H^{n-1}(p)}{e^{\frac{n-1}{2}r}} \\&=\frac{1}{n-1}2^{-\frac{n-1}{2}} \int_{\partial K}\frac{k_p^{1/2}}{\langle p,\nu_p\rangle^{\frac{n-1}{2}}}d\mathcal H^{n-1}(p).\end{align*} \endproof \section{Maximizing the Funk-Mahler volume}\label{sec:inequalities} Let us introduce for convenience $\rho=1-e^{-r}$, and the modified Funk-Mahler volumes $$\widetilde M_\rho(K,q):=r^nM_{r}(K,q)=\int_{z\in q+\rho (K-q)}|K^z|dz,\quad \widetilde M_\rho(K)=\inf_{q\in \inter(K)}\widetilde M_\rho(K, q).$$ \begin{Lemma} Let $B\subset\mathbb{R}^n$ be an ellipsoid. Then $$\widetilde M_\rho(B)=n\omega_n^2\int_0^\rho\frac{t^{n-1}dt}{(1-t^2)^{\frac{n+1}{2}}}.$$ \end{Lemma} \proof By Lemmas \ref{lem:mahler_convex} and \ref{lem:volume_ellipsoid}, \[\widetilde M_\rho(B)=\widetilde M_\rho(B^n,0)=\int_{\rho B}|B^z|dz=\omega_n\int_{\rho B}\frac{dz}{(1-|z|^2)^{\frac{n+1}{2}}}=n\omega_n^2 \int_0^\rho\frac{t^{n-1}dt}{(1-t^2)^{\frac{n+1}{2}}}.\] \endproof In light of the discussion in section \ref{sec:every_scale}, and inspired by the Blaschke-Santal\'o and the centro-affine isoperimetric inequalities, we propose a unified inequality containing both of the above as limiting cases. \begin{Conjecture}\label{conj:bodies} Fix any $0\leq \rho<1$, and let $K\in\mathcal K(\mathbb{R}^n)$ have $0$ as its centroid. \begin{enumerate} \item $\widetilde M_\rho(K,0)\leq \widetilde M_\rho(B^n)$. \item If $\widetilde M_\rho(K,0)= \widetilde M_\rho(B^n)$ then $K$ is an ellipsoid. \end{enumerate} \end{Conjecture} This is not to say that $\widetilde M_\rho(K)=\widetilde M_\rho(K,0)$: as $\rho\to 0$, $\widetilde M_\rho(K,q)$ attains its minimum close to the Santal\'o point of $K$, which generally differs from the centroid. We will prove the conjecture in the restricted class of unconditional convex bodies. To this end, we must expand our discussion to include log-concave functions. We write $f_q(\bullet)=f(q+\bullet)$. \begin{Definition} For $f:\mathbb{R}^n\to \mathbb{R}_+$ a Borel function, put $\phi(x)=-\log f(x):\mathbb{R}^n\to \mathbb{R}\cup\{+\infty\}$. Define for $\rho\in\mathbb C$ the Funk-Mahler volumes around $0$ \begin{align*}\widetilde M_\rho(f,0)&:=\frac{1}{n!}\displaystyle\int_{z\in\mathbb{R}^n} e^{-\phi(\frac{z}{\rho})}\displaystyle\int_{\xi\in \mathbb{R}^n} e^{-\mathcal L\phi_z(\xi)} d\xi dz\\&=\frac{1}{n!}\rho^n\displaystyle\int_{\mathbb{R}^n\times \mathbb{R}^n} e^{-\phi(x)-\mathcal L \phi (\xi)+\rho\langle x,\xi\rangle}dxd\xi,\end{align*} whenever the integral converges. Define also $$\widetilde M_\rho(f, q):=\widetilde M_\rho(f_q,0), \quad \widetilde M_\rho(f):=\inf_{q\in \mathbb{R}^n} \widetilde M_\rho(f, q).$$ \end{Definition} As $\rho\to 0$, $$\widetilde M_\rho(f,0)\sim \frac 1{n!} \rho^n \int_{\mathbb{R}^n\times \mathbb{R}^n} e^{-\phi(z)-\mathcal L \phi(\xi)}dxd\xi,$$ which is the functional Mahler product. \begin{Lemma}\label{lem:functional_invariance} It holds that $\widetilde M_\rho(f,0)=\widetilde M_\rho(\lambda f\circ A,0)$ for any $\lambda> 0$ and $A\in\GL_n(\mathbb{R})$. Furthermore, $\widetilde M_\rho(f,q)$ is a convex function of $q$, and strictly convex where finite. \end{Lemma} \proof The first statement is immediate from definition by the $\GL$-equivariance of the Legendre transform, and since $\mathcal L(\phi+c)=\mathcal L\phi -c$. For the second, observe that $$ \widetilde M_\rho(f,q)=\frac{1}{n!}\int_{\mathbb{R}^n\times\mathbb{R}^n} e^{-\phi(\frac{w}{\rho})+\langle \xi,w\rangle-\mathcal L\phi(\xi)}e^{(1-\rho)\langle \xi,q\rangle}dwd\xi,$$ and note that $t\mapsto e^{(1-\rho)t}$ is strictly convex. \endproof The functional Funk-Mahler volumes generalize those previously introduced for convex bodies, as follows. \begin{Lemma}\label{lem:body_vs_function} $\widetilde M_\rho(\mathbbm 1_K, 0)=\widetilde M_\rho(K ,0)$. \end{Lemma} \proof Recall that $\mathcal L(-\log \mathbbm 1_K)(\xi)=h_K(\xi)=\|\xi\|_{K^o}$. Thus $$\widetilde M_\rho(\mathbbm 1_K, 0)=\frac{1}{n!}\int_{\rho K\times\mathbb{R}^n}e^{-\|\xi\|_{K^z}}dzd\xi=\frac{1}{n!}\int_{\rho K}dz \int_0^\infty r^{n-1}e^{-r}dr\int_{\partial K^z}d\mu_{K^z}, $$ where $d\xi=r^{n-1}drd\mu_{K^z}$. Thus $$\widetilde M_\rho(\mathbbm 1_K, 0)= \frac{1}{n!}(n-1)!n\int_{\rho K} |K^z|dz=\widetilde M_\rho(K, 0). \qedhere$$ Using this representation, let us exhibit a curious property of the Funk-Mahler volume of a convex body, which at present we only formally establish for zonoids. \begin{Proposition}\label{prop:to_the_infinity_and_beyond} Assume $0\in \inter(K)$. There is then a continuous extension of $\widetilde M_\rho(K, 0)$ to $\rho\in[0,1)\times i\mathbb{R}$, which is analytic in $(0,1)\times i\mathbb{R}$. \\Furthermore if $K$ is a zonoid, then $$\lim_{t\to \pm\infty} \widetilde M_{it}(K,0)=i^n\frac{(2\pi)^n}{n!}.$$ \end{Proposition} \proof For the first statement, put $\rho=s+it$, write $$\widetilde M_{\rho}(K,0)= \widetilde M_{\rho}(\mathbbm 1_K,0)=\frac{(s+it)^n}{n!}\int_{K\times\mathbb{R}^n} e^{-h_K(\xi)+s\langle x,\xi\rangle +it \langle x,\xi\rangle}dxd\xi, $$ and note that $$\frac{1}{n!}\int_{K\times\mathbb{R}^n} e^{-h_K(\xi)+s\langle x,\xi\rangle }dxd\xi =\frac{1}{s^n}\widetilde M_s(K,0)<\infty$$ for $0\leq s<1$. For the second statement, let $\mathcal F(f)(y)=\int_{\mathbb{R}^n}f(x)e^{-i\langle x, y\rangle }dx$ be the Fourier transform. Write \begin{align*}\widetilde M_{it}(K,0)&=\frac{i^n}{n!} t^n\int_{\mathbb{R}^n\times\mathbb{R}^n}\mathbbm 1_K(x) e^{-h_K(\xi)}e^{it\langle x, \xi\rangle}dxd\xi\\&=\frac{i^n}{n!}\int_{\mathbb{R}^n} \mathbbm 1_K(\frac{x}{t})\mathcal F(e^{-h_K})(-x)dx\\&=\frac{i^n}{n!}\int_{\mathbb{R}^n}e^{-h_K(\xi)}\int_{tK}e^{i\langle x,\xi\rangle} dxd\xi ,\end{align*} and note that $\mathcal F(e^{-h_K})\in L^1(\mathbb{R}^n)$ when $K$ is a zonoid. Thus as $t\to\pm \infty$, $$ \widetilde M_{it}(K,0)\to \frac{i^n}{n!}\int_{\mathbb{R}^n}\mathcal F(e^{-h_K})=\frac{i^n}{n!}(2\pi)^ne^{-h_K(0)}=i^n\frac{(2\pi)^n}{n!}.$$ \endproof Motivated by Conjecture \ref{conj:bodies}, and inspired by the results of \cite{lehec}, we boldly put forward the following \begin{Conjecture}\label{conj:functional_general} Consider two Borel functions $\phi,\psi:\mathbb{R}^n\to \mathbb{R}\cup\{+\infty\}$ such that for all $x,\xi\in\mathbb{R}^n$ one has $\phi(x)+\psi(\xi)\geq \langle x, \xi\rangle $. Assume also $\int xe^{-\phi(x)}dx=0$. Then for all $0\leq \rho<1$, \begin{equation}\label{eq:general_inequality} \int_{\mathbb{R}^n\times\mathbb{R}^n}e^{-\phi(x)-\psi(\xi)+\rho\langle x,\xi\rangle}dxd\xi\leq (2\pi)^n(1-\rho^2)^{-\frac n 2}.\end{equation} Equality is attained uniquely, up to equality a.e., by the pairs of functions \begin{equation}\label{eq:pairs}\phi(x)=\frac 12 \langle Ax, x\rangle +c, \quad \psi(\xi)=\mathcal L\phi(\xi)=\frac 12 \langle A^{-1}x, x\rangle -c,\end{equation} for some $A>0$ and $c\in\mathbb{R}$. \end{Conjecture} We remark that here and in the following, one may restrict attention to dual pairs of convex functions $\phi, \psi=\mathcal L\phi$, without loss of generality. In all that follows, we take \emph{a.e.-uniqueness} to mean uniqueness up to equality a.e. Let us check the trivial statement in the conjecture. \begin{Lemma} Pairs $\phi,\psi$ as in \eqref{eq:pairs} give equality in \eqref{eq:general_inequality}. \end{Lemma} \proof Taking $q(x)=\frac12 |x|^2$, $$\mathcal Lq_z(\xi)=\frac{|\xi|^2}{2}-\langle \xi,z\rangle,$$ and so \begin{align*} &\int_{\mathbb{R}^n\times\mathbb{R}^n} e^{-\frac{|z|^2}{2\rho^2}-\frac{|\xi|^2}{2}+\langle \xi, z\rangle}d\xi dz=\rho^n\int e^{-\frac12 (|w|^2+|\xi|^2-2\rho\langle w,\xi\rangle )}d\xi dw \\&=\rho^n\int e^{-\frac12 (|w-\rho\xi|^2+(1-\rho^2)|\xi|^2)}d\xi dw =\int e^{-\frac12 |w|^2}dw\int e^{-\frac12 (1-\rho^2)|\xi|^2} d\xi \\&=(2\pi)^n\rho^n(1-\rho^2)^{-\frac{n}{2}}. \end{align*} By Lemma \ref{lem:functional_invariance}, we get equality for all pairs as claimed. \endproof An immediate corollary is a functional version of Conjecture \ref{conj:bodies}, as follows. \begin{Conjecture}\label{conj:functional_santalo} Let $f:\mathbb{R}^n\to \mathbb{R}_+$ be an integrable Borel function, and assume also that $\int x f(x)dx=0$. Then for all $0\leq \rho<1$, \[ \widetilde M_\rho(f,0)\leq \frac{1}{n!}(2\pi)^n\rho^n(1-\rho^2)^{-\frac n 2}. \] Equality is attained uniquely by multiples of gaussians: $f(x)=\lambda e^{-\frac 12 \langle Ax, x\rangle}$. \end{Conjecture} \begin{Remark} Note that the implication Conjecture \ref{conj:functional_general} $\Rightarrow$ Conjecture \ref{conj:functional_santalo} is valid separately for each $n$, and also separately for any family of functions with certain symmetry, such as even or unconditional functions. \end{Remark} We now prove Conjecture \ref{conj:functional_general} under the extra assumption of unconditionality. Recall that $\phi$ is unconditional if $\phi(x_1,\dots, x_n)=\phi(|x_1|,\dots, |x_n|)$. \begin{Theorem}\label{thm:functional_unconditional} Consider two unconditional Borel functions $\phi,\psi:\mathbb{R}^n\to \mathbb{R}\cup\{+\infty\}$ such that for all $x,\xi\in\mathbb{R}^n$ one has $\phi(x)+\psi(\xi)\geq \langle x, \xi\rangle $. Then for all $0\leq \rho<1$, $$ \int_{\mathbb{R}^n\times\mathbb{R}^n}e^{-\phi(x)-\psi(\xi)+\rho\langle x,\xi\rangle}dxd\xi\leq (2\pi)^n(1-\rho^2)^{-\frac n 2}.$$ Equality is attained a.e.-uniquely by Legendre-dual pairs $$\phi(x)=\frac 12 \langle Ax, x\rangle +c, \quad \psi(\xi)=\mathcal L\phi(\xi)=\frac 12 \langle A^{-1}x, x\rangle -c.$$ \end{Theorem} \proof Since $\phi(x)=\phi(-x)$, the claimed inequality is equivalent to \begin{equation}\label{eq:ineq1} I:=\int_{\mathbb{R}^n\times\mathbb{R}^n} e^{-\phi(x)-\psi(\xi)}\cosh(\rho \langle x, \xi\rangle)d\xi dx\leq (2\pi)^n(1-\rho^2)^{-n/2}.\end{equation} The left hand side can be rewritten as \begin{align*}J&=\sum_{j=0}^\infty \frac{\rho^{2j}}{(2j)!}\int_{\mathbb{R}^n\times\mathbb{R}^n}e^{-\phi(x)-\psi(\xi)}\langle x, \xi\rangle^{2j}dxd\xi\\&=\sum_{j=0}^\infty\frac{\rho^{2j}}{(2j)!}\sum_{|I|=2j}{2j\choose I}\int_{\mathbb{R}^n}x^{I}e^{-\phi(x)}dx\int_{\mathbb{R}^n}\xi^{I}e^{-\psi(\xi)}d\xi,\nonumber\end{align*} where the last sum is over all $n$-tuples $I=(i_1,\dots, i_n)$ of non-negative integers such that $|I|:=\sum_{\nu=1}^n i_\nu =2j$, and we use the notation $x^I:=\prod_{\nu=1}^n x_\nu^{i_\nu}$, and $${2j\choose I}:=\frac{(2j)!}{i_1!\cdots i_n!}$$ for the multinomial coefficient. Observe that by unconditionality, if $I$ contains an odd index then the corresponding integrals vanish. Therefore, \begin{align}\label{eq:eq1} J&= \sum_{j=0}^\infty\frac{\rho^{2j}}{(2j)!}\sum_{|I|=j}{2j\choose 2I}\int_{\mathbb{R}^n}x^{2I}e^{-\phi(x)}dx\int_{\mathbb{R}^n}\xi^{2I}e^{-\psi(\xi)}d\xi\\&= 4^{n}\sum_{j=0}^\infty\frac{\rho^{2j}}{(2j)!}\sum_{|I|=j}{2j\choose 2I}\int_{\mathbb{R}^n_+}x^{2I}e^{-\phi(x)}dx\int_{\mathbb{R}^n_+}\xi^{2I}e^{-\psi(\xi)}d\xi. \nonumber\end{align} Making the change of coordinates $x=e^u:=(e^{u_i})_{i=1}^n$, $\xi=e^v=(e^{v_i})$, we may write \begin{equation}\label{eq:eq2}\int_{\mathbb{R}_+^n}e^{-\phi(x)}x^{2I}dx\int_{\mathbb{R}_+^n}e^{-\psi(\xi)}\xi^{2I}d\xi=\int_{\mathbb{R}^n} f(u)du\int_{\mathbb{R}^n} g(v)dv,\end{equation} where $f(u)=e^{-\phi(e^u)}e^{\langle 2I+1, u\rangle}$, $g(v)=e^{-\psi(e^v)}e^{\langle 2I+1, v\rangle}$, and $I+c:=(i_\nu+c)_{\nu=1}^n$. Denote $$h(w)=e^{-\frac12 |e^w|^2}e^{\langle 2I+1, w\rangle},\quad w\in \mathbb{R}^n,$$ where $|e^w|=\sqrt{\sum_{i=1}^n e^{2w_i}}$. Since $\phi(x)+\psi(\xi)\geq \langle x, \xi\rangle$, we immediately find that $$f(u)g(v)\leq h\left(\frac{u+v}{2}\right)^2,$$ and by the Prekopa-Leindler inequality we have \begin{align*} \int_{\mathbb{R}^n} f\int_{\mathbb{R}^n} g&\leq \left(\int_{\mathbb{R}^n} h\right)^2=\left(\int_0^\infty e^{-\frac{|p|^2}{2}}p^{2I}dp\right)^2=\prod_{\nu=1}^n\left(\int_0^\infty e^{-\frac{t^2}{2}} t^{2i_\nu}dt\right)^2\\&=\left( \frac{\sqrt{2\pi}}{2}\right)^{2n} \prod_{\nu=1}^n ((2i_\nu-1)!!)^2=\frac{(2\pi)^n}{4^n} \prod_{\nu=1}^n ((2i_\nu-1)!!)^2.\end{align*} Combined with \eqref{eq:ineq1}, \eqref{eq:eq1}, \eqref{eq:eq2}, it remains to verify that $$ \sum_{j=0}^\infty \frac{\rho^{2j}}{(2j)!}\sum_{|I|=j}{2j\choose 2I}\prod_{\nu=1}^n ((2i_\nu-1)!!)^2 =(1-\rho^2)^{-n/2},$$ or equivalently $$ \sum_{j=0}^\infty \rho^{2j}\sum_{|I|=j}{2j\choose 2I}\prod_{\nu=1}^n \frac{(2i_\nu-1)!!}{(2i_\nu)!!} =(1-\rho^2)^{-n/2}.$$ By Newton's binomial formula, \[ (1-\rho^2)^{-\frac12}=\sum_{i=0}^\infty{-1/2 \choose i}(-1)^i\rho^{2i}=\sum_{i=0}^\infty\frac{(2i-1)!!}{2^i i!}\rho^{2i}=\sum_{i=0}^\infty\frac{(2i-1)!!}{(2i)!!}\rho^{2i}, \] and it remains to raise both sides to power $n$. For equality, we must have an equality at each application of Prekopa-Leindler, which is characterized in \cite{dubuc}. Namely for each $I$, one has a.e. equalities $f=\lambda h(\bullet + C)$, $g=\lambda^{-1}h(\bullet - C)$. Already for a single multi-index $I$ this forces the existence of a diagonal matrix $A$ and $c\in\mathbb{R}$ such that a.e. one has $$\phi(x)=\frac 12 \langle Ax, x\rangle +c,\quad \psi(\xi)=\frac 12 \langle A^{-1}x, x\rangle -c.$$ \endproof By the previous discussion, we immediately get \begin{Corollary} For all unconditional Borel functions $f:\mathbb{R}^n\to \mathbb{R}_+$ it holds that \[ \widetilde M_\rho(f,0)\leq \frac{1}{n!}(2\pi)^n\rho^n(1-\rho^2)^{-\frac n 2}, \] with equality attained a.e.-uniquely by $f(x)=\lambda e^{-\frac12 \langle Ax, x\rangle}$ with some diagonal matrix $A>0$ and scalar $\lambda>0$. \end{Corollary} \proof Follows directly from Theorem \ref{thm:functional_unconditional}, since $\widetilde M_\rho(f,0)=\widetilde M_\rho(f)$ by unconditionality and the second part of Lemma \ref{lem:functional_invariance}. \endproof In particular, we proved a functional inequality of independent interest. \begin{Corollary}\label{cor:even_moments} Consider an unconditional Borel function $f=e^{-\phi}:\mathbb{R}^n\to \mathbb{R}_+$. Then $$ I_{2j}(f):=\int_{\mathbb{R}^n\times\mathbb{R}^n}\langle x,\xi\rangle ^{2j}e^{-\phi(x)-\mathcal L\phi(\xi)}dx d\xi\leq (2\pi)^n \frac{(n-2+2j)!!}{(n-2)!!(2j-1)!!}.$$ Equality is attained a.e.-uniquely by multiples of gaussians $f(x)=\lambda e^{-\frac 12 \langle Ax, x\rangle}$, for some diagonal $A>0$ and scalar $\lambda>0$. \end{Corollary} The case of $j=0$ is the Blaschke-Santal\'o inequality for unconditional functions. The case of $j=1$ was established previously in \cite{huang_li}. \proof Denote $\psi=\mathcal L\phi$. In the proof of Theorem \ref{thm:functional_unconditional}, we have seen that $I_{2j}(e^{-\phi})$ is maximized a.e.-uniquely by multiples of gaussians. The maximal value $I_{2j}(G)$ can be computed directly, or alternatively it can be deduced from Theorem \ref{thm:functional_unconditional} as follows. We have $$\sum_{j=0}^\infty\frac{\rho^{2j}}{(2j)!}I_{2j}(G)= (2\pi)^n(1-\rho^2)^{-n/2},$$ and by Newton's binomial \[ (1-\rho^2)^{-\frac n2}=\sum_{j=0}^\infty{-n/2 \choose j}(-1)^j\rho^{2j}=\sum_{i=0}^\infty\frac{(n+2j-2)!!}{(n-2)!!(2j)!! }\rho^{2j}.\] Consequently, $$I_{2j}(G)=(2\pi)^n(2j)!\frac{(n+2j-2)!!}{(n-2)!!(2j)!! }=(2\pi)^n\frac{(n-2+2j)!!}{(n-2)!!(2j-1)!!}.$$ \endproof We can now deduce the corresponding inequalities for unconditional convex bodies. Define $$I_{2j}(K):=\int_{K\times K^o} \langle x,\xi\rangle^{2j}dxd\xi.$$ First, we prove two simple relations. \begin{Lemma} For any convex body $K\in\mathcal K_0(\mathbb{R}^n)$, \begin{equation}\label{eq:gaussian_body} I_{2j}(K)=\frac{1}{((n+2j)!!)^2} \left(\frac 2\pi \right)^{\frac12(1-(-1)^n)}I_{2j}(e^{-\frac12 \|x\|_K^2}).\end{equation} If $K=-K$ then \begin{equation}\label{eq:mahler_to_moments}\widetilde M_\rho(K)= \sum_{j=0}^\infty {n+2j \choose n}I_{2j}(K)\rho^{n+2j}.\end{equation} \end{Lemma} \proof We will use the change of variables $x=tp$, $\xi=s\eta$, with $p\in\partial K, \eta\in \partial K^o$. We have $dx=r^{n-1}drd\mu_K(p), d\xi =s^{n-1}dsd\mu_{K^o}(\eta)$. We may write \begin{align*}I_{2j}(e^{-\frac12 \|x\|_K^2})&=\int_{\mathbb{R}^n\times \mathbb{R}^n}\langle x, \xi\rangle ^{2j}e^{-\frac{\|x\|^2_K}{2}}e^{-\frac{\|\xi\|^2_{K^o}}{2}}dxd\xi\\&=\int_{[0,\infty)^2}e^{-\frac {r^2}{2}-\frac{s^2}{2}}r^{n+2j-1}s^{n+2j-1}drds\int_{\partial K\times\partial K^o}\langle p, \eta\rangle ^{2j}d\mu_K(p)d\mu_{K^o}(\eta) \\&=\left(\int_0^\infty r^{n+2j-1} e^{-r^2/2}dr\right)^2\int_{\partial K\times\partial K^o}\langle p, \eta\rangle ^{2j}d\mu_K(p)d\mu_{K^o}(\eta)\\ &=2^{n+2j-2}\Gamma(\frac{n}{2}+j)^2\int_{\partial K\times\partial K^o}\langle p, \eta\rangle ^{2j}d\mu_K(p)d\mu_{K^o}(\eta).\end{align*} On the other hand, \begin{align*} I_{2j}(K)&=\int_{K\times K^o}\langle x, \xi\rangle^{2j}dxd\xi\\&=\int_{\partial K\times \partial K^o} \langle p, \eta\rangle ^{2j}d\mu_K(p)d\mu_{K^o}(\eta)\int_{[0,1]^2}r^{n+2j-1}s^{n+2j-1}drds\\&=\frac{1}{(n+2j)^2}\int_{\partial K\times \partial K^o} \langle p, \eta\rangle ^{2j}d\mu_K(p)d\mu_{K^o}(\eta), \end{align*} and it remains to combine the two equalities to find \begin{align*} I_{2j}(K)&=\frac{1}{2^{n+2j-2}(n+2j)^2\Gamma(\frac n 2+j )^2}I_{2j}(e^{-\frac12\|x\|^2_K})\\&=\frac{1}{((n+2j)!!)^2} \left(\frac 2\pi \right)^{\frac12(1-(-1)^n)}I_{2j}(e^{-\frac12 \|x\|_K^2}).\end{align*} Now assume $K=-K$. By Lemmas \ref{lem:body_vs_function} and \ref{lem:mahler_convex}, \begin{align*} \widetilde M_\rho(K)=\widetilde M_\rho(K,0)=\widetilde M_\rho(\mathbbm 1_K, 0)=&\frac{\rho^n}{n!}\int_{K\times\mathbb{R}^n}e^{-\|\xi\|_{K^o}+\rho\langle x,\xi\rangle}dxd\xi\\&=\frac{\rho^n}{n!}\int_{K\times\mathbb{R}^n}e^{-\|\xi\|_{K^o}}\cosh(\rho\langle x,\xi\rangle)dxd\xi \\&=\sum_{j=0}^\infty \frac{1}{n!(2j)!}\rho^{n+2j}\int_{K\times\mathbb{R}^n}\langle x, \xi\rangle^{2j}e^{-\|\xi\|_{K^o}}dxd\xi. \end{align*} We may write \begin{align*}I_{2j}(K)=\int_{K\times K^o}\langle x,\xi\rangle^{2j}dxd\xi&=\int_{K\times\partial K^o}\langle x,\eta\rangle^{2j}dxd\mu_{K^o}(\eta)\int_0^1s^{2j+n-1}ds\\&=\frac{1}{n+2j}\int_{K\times\partial K^o}\langle x,\eta\rangle^{2j}dxd\mu_{K^o}(\eta).\end{align*} Therefore, \begin{align*}\int_{K\times\mathbb{R}^n}\langle x, \xi\rangle^{2j}e^{-\|\xi\|_{K^o}}dxd\xi &=\int_{K\times [0,\infty)\times \partial K^o} s^{2j+n-1}e^{-s}\langle x,\eta\rangle ^{2j}dxds d\mu_{K^o}(\eta)\\&=(2j+n-1)!\int_{K\times\partial K^o}\langle x, \eta\rangle^{2j}dxd\mu_{K^o}(\eta)\\&=(2j+n)!I_{2j}(K),\end{align*} and so \[ \widetilde M_\rho(K) = \sum_{j=0}^\infty \frac{(n+2j)!}{n!(2j)!}\rho^{n+2j}I_{2j}(K). \] \endproof \begin{Corollary}\label{cor:bodies_moments} For $K\subset \mathbb{R}^n $ convex and unconditional, $$ I_{2j}(K):=\int_{K\times K^o} \langle x,\xi\rangle^{2j}dxd\xi\leq \frac{(2\pi)^n}{n+2j}\frac{1}{ (n+2j)!!(n-2)!!(2j-1)!!} \left(\frac2\pi\right)^{\frac{1-(-1)^n}{2} }.$$ Equality is attained uniquely by ellipsoids. \end{Corollary} \proof We apply Corollary \ref{cor:even_moments} to $f(x)=e^{-\phi(x)}=e^{-\frac12 \|x\|_K^2}$. By eq. \eqref{eq:gaussian_body}, $I_{2j}(K)$ is uniquely maximized by ellipsoids, and the maximal value is readily computed. \endproof \begin{Corollary}\label{cor:bodies_mahler} For any $0\leq r<\infty$, and $K\subset \mathbb{R}^n$ convex and unconditional, $M_r(K)\leq M_r(B^n)$. Equality is attained uniquely by ellipsoids. \end{Corollary} \proof We only need to note that by Corollary \ref{cor:bodies_moments}, each summand of \eqref{eq:mahler_to_moments} is uniquely maximized by ellipsoids. \endproof \textit{Proof of Corollary \ref{cor:entropy}}. Since $d^H=\frac12(d^F+d^{RF})\geq \frac12d^F$, it follows that $B^H_r(q, K)\subset B^F_{2r}(q, K)$. Write $c_n$ for a constant depending only on $n$ which may change between occurrences. By Proposition \ref{prop:centro_affine} we have $\vol^F(B^F_R(q,B^n))\leq c_ne^{\frac{n-1}{2}R }$. Using the Rogers-Shephard inequality as in the proof of Corollary \ref{cor:hilbert_volume_sec3}, and applying Corollary \ref{cor:bodies_mahler}, we find \begin{align*} \vol_K^H(B_r^H(q,K))&\leq c_n \vol_K^F(B_r^H(q,K))\leq c_n\vol_K^F(B_{2r}^F(q,K))\leq c_n\vol_K^F(B_{2r}^F(q,B^n))\\&\leq c_ne^{\frac{n-1}{2}2r}=c_ne^{(n-1)r},\end{align*} completing the proof. \qed Using the same methods as before, we can also show the following. \begin{Theorem} Let $f=e^{-\phi}:\mathbb{R}^n_+\to \mathbb{R}_+$ be a Borel function. Then \begin{itemize} \item For all $j\geq 0$, $$I^+_j(f):=\int_{\mathbb{R}_+^n\times\mathbb{R}_+^n}\langle x,\xi\rangle^{j} e^{-\phi(x)-\mathcal L\phi(\xi)}dx d\xi\leq I^+_j(e^{-\frac{|x|^2}{2}}).$$ \item If $n=1$, $$I_{2j+1}^+(f)=\int_0^\infty x^{2j+1} e^{-\phi(x)}dx \int_0^\infty \xi^{2j+1}e^{-\mathcal L\phi(\xi)} d\xi\leq (2j)!!^2.$$ \item For all $0\leq \rho<1$, $$\int_{\mathbb{R}_+\times\mathbb{R}_+}e^{-\phi(x)-\mathcal L\phi (\xi)}\sinh (\rho x\xi)dxd\xi\leq\frac{1}{\sqrt{1-\rho^2}}\arctan\frac{\rho}{\sqrt{1-\rho^2}}.$$ \end{itemize} Equality in all cases is a.e.-uniquely attained by $\lambda e^{-\frac12 \langle Ax, x\rangle}$, where $A>0$ is diagonal and $\lambda>0$. \end{Theorem} \proof The first two inequalities can be obtained by trivially adjusting the proof of Theorem \ref{thm:functional_unconditional}. For the last inequality, we should show that $$\sum_{j=0}^\infty\frac{(2j)!!^2}{(2j+1)!}\rho^{2j+1}=\sum_{j=0}^\infty\frac{(2j)!!}{(2j+1)!!}\rho^{2j+1}=\frac{1}{\sqrt{1-\rho^2}}\arctan\frac{\rho}{\sqrt{1-\rho^2}}.$$ This can be done by observing that $$ \frac{(2j)!!}{(2j+1)!!}=\int_0^{\pi/2}\sin^{2j}tdt,$$ and so \begin{align*}\sum_{j=0}^\infty\frac{(2j)!!}{(2j+1)!!}\rho^{2j+1}&=\int_{0}^{\frac \pi 2}(\rho\sin t)^{2j+1}dt=\int_0^{\frac \pi 2}\frac{\rho \sin t}{1-\rho^2\sin^2t}dt\\&=\int_0^1 \frac{2\rho ds}{s^2+2(1-2\rho^2)s +1}\\&=\frac{1}{\sqrt{1-\rho^2}}\left(\arctan \frac{\sqrt{1-\rho^2}}{\rho}-\arctan \frac{1-2\rho^2}{2\rho \sqrt{1-\rho^2}}\right)\\&=\frac{1}{\sqrt{1-\rho^2}}\arctan \frac{\rho}{\sqrt{1-\rho^2}}.\end{align*} Equality cases are settled using \cite{dubuc} as in Theorem \ref{thm:functional_unconditional}. \endproof As a sidenote, the discussion above suggests an extension of the centro-affine surface area to functions. \begin{Definition} Let $f:\mathbb{R}^n\to[0,\infty)$ be a Borel function. Its centro-affine surface area is \begin{align*}\Omega_n(f)&:=\frac{1}{n!}\limsup_{\rho\to1^-} (1-\rho^2)^{n/2}\displaystyle\int_{\mathbb{R}^n\times \mathbb{R}^n} e^{-\phi(x)-\mathcal L \phi (\xi)+\rho\langle x,\xi\rangle}dxd\xi.\end{align*} \end{Definition} Evidently, $\Omega_n(f)=\Omega_n(\lambda f\circ A)$ for any $\lambda> 0$ and $A\in\GL_n(\mathbb{R})$. \section{Regularizing the total Funk volume} \label{sec:projective_total_volume} We use the notation of subsection \ref{sec:volume_funk}. Recall that we use the standard Euclidean structure on $V=\mathbb{R}^{n+1}$, and identify the double cover of $\mathbb{P}$ and $\mathbb{P}^\vee$ with the unit sphere. Let $x_0\in x$ be a unit vector, well-defined up to sign. We write $\sigma_x, \sigma_\xi$ for the standard Euclidean measure on the corresponding copy of $S^2$. \begin{Lemma}\label{lem:lemma91}For $g\in \PGL(V)$, set $\chi_g(x,\xi)=|gx_0||g^{-*}\xi_0|$. Then \[g_*(\sigma_x\sigma_\xi)=\chi_{g^{-1}}^{-n-1}\sigma_x\sigma_\xi\] \end{Lemma} \proof Denote by $\Stab^+(x,\xi)\subset\Stab(x,\xi)$ the subgroup of the stabilzer in $\PGL(V)$ that has $\sign \det g|_x=\sign \det g^*|_\xi$. Then $\Stab^+(x,\xi)$-equivariantly we have $$\Dens(T_{x,\xi}(\mathbb{P}\times \mathbb{P}^\vee))=\Dens^*(x)^{n+1}\otimes\Dens^*(\xi)^{n+1}=x^{n+1}\otimes \xi^{n+1},$$ and so \[g_*(\sigma_x\sigma_\xi)(gx,g^{-*}\xi)=(|gx_0||g^{-*}\xi_0|)^{n+1}\sigma_x\xi=\chi_g(x,\xi)^{n+1}\sigma_x\sigma_\xi.\] Hence \[g_*(\sigma_x\sigma_\xi)(x,\xi)=\chi_g(g^{-1}x,g^*\xi)^{n+1}\sigma_x\sigma_\xi=\chi_{g^{-1}}^{-n-1}(x,\xi)\sigma_x\sigma_\xi,\] as claimed. \endproof Define $f_0\in C(\mathbb{P}\times\mathbb{P}^\vee)$ by $f_0(x,\xi)=|\langle x_0,\xi_0\rangle|$. It is easy to check that \[ g^*f_0=\chi_g^{-1}f_0.\] Define $\mu_z(x,\xi)=f_0(x,\xi)^z\sigma_x\sigma_\xi\in \mathcal M(\mathbb{P}\times\mathbb{P}^\vee\setminus \mathcal Z)$. For $\textrm{Re}z>-1$, it extends as an absolutely continuous measure on $\mathbb{P}\times\mathbb{P}^\vee$. \begin{Lemma}\label{lem:measure_push} $g_*\mu_z=\chi_{g^{-1}}^{-(n+1)-z}\mu_z$. \end{Lemma} \proof It is easy to check that $g_*\mu_z=(g^{-1})^*f_0^z\cdot g_*(\sigma_x\sigma_\xi)$. It remains to put the previous computations together. \endproof As expected from Proposition \ref{prop:invariant_measure}, we see that $z=-(n+1)$ corresponds to the invariant measure $\mu$. Henceforth we use the same notation for a point on $S^n$ and the corresponding line in $\mathbb{R}\mathbb{P}^n$. The action of $\GL(V)$ on $S^n$ is $g(x)=\frac{gx}{|gx|}$. Let $K\subset S^n\subset V=\mathbb{R}^{n+1}$ be a proper convex compact set, and $K^\vee\subset S^n$ its polar body, namely $K^\vee=\{\xi\in S^{n}: \forall x\in K, \langle x,\xi\rangle\geq 0\}$. They correspond to a dual pair of convex sets in $\mathbb{P}$ and $\mathbb{P}^\vee$, also denoted $K, K^\vee$. Define the \emph{Beta function} of $K\subset S^n$ by \[\mathbb{C}\ni z\mapsto B_K(z)=\int_{K\times K^\vee}\langle x,\xi\rangle ^z\sigma_x\sigma_\xi =\langle \mathbbm1_{K\times K^\vee}, \mu_z\rangle. \] Clearly, $B_K(z)=B_{K^\vee}(z)$. Assume henceforth that $K$ is smooth and strictly convex. \begin{Lemma}\label{lem:convergence_domain} The integral $B_K(z)$ is convergent if and only if $\mathrm{Re} z>-\frac{n+3}{2}$. \end{Lemma} \proof The zeros of $\langle x,\xi\rangle $ on $K\times K^\vee$ are $\mathcal Z\cap K\times K^\vee$, which is an $(n-1)$-dimensional submanifold in $\partial K\times\partial K^\vee$. Let $x_1,\dots,x_{n-1}$ be local coordinates on $\partial K$, and $\xi_1,\dots, \xi_{n-1}$ the Legendre-dual coordinates on $\partial K^\vee$, that is $\mathcal Z\cap (K\times K^\vee)=\{\xi_j=x_j\}$. Let $x_n$ be the remaining coordinate near $\partial K$, so that locally $ K=\{x_n\geq 0\}$, and $K^\vee=\{\xi_n\geq 0\}$. We put $\xi=x+y$, and use the notation $v^j=(v_i)_{i=1}^{j}$. When restricted to $\{x_n=\xi_n=0\}$, $f_0( (x^{n-1},0), (\xi^{n-1},0)) =\langle A_x y^{n-1}, y^{n-1}\rangle+O(|y^{n-1}|^3)$, for some $A_x>0$. It follows that in general, $$f_0(x,\xi)=\langle A_{x}y^{n-1}, y^{n-1}\rangle+b_xx_n+c_x\xi_n +O(|y^{n-1}|^3+|y^{n-1}|(x_n+\xi_n)+x_n^2+\xi_n^2),$$ where $A_x, b_x, c_x>0$ are uniformly bounded from above and below by strict convexity. Thus $\int_{K\times K^\vee} f_0(x,\xi)^z$ converges if and only if \[\int_{y_1^2+\dots+y_{n-1}^2<1}\int_{x_n, \xi_{n}=0}^1 ( y_1^2+\dots+y_{n-1}^2+x_n+\xi_{n})^z<\infty, \] which immediately reduces to $2(\mathrm{Re}z+2)+(n-2)>-1\iff \Re z>-\frac{n+3}{2}$. \endproof \emph{In all the following we assume $n=2$}, as already this case appears to be interesting and involved. We proceed to prove Theorem \ref{main:beta}. Denote $D_+=\{(x,\xi); \langle x,\xi\rangle> 0\}\subset S^2\times S^2$. The group $\SO(3)$ acts diagonally on $S^2\times S^2$, leaving $D_+$ invariant. By $\Omega^{i,j}(S^2\times S^2)$ we denote the differential forms of bidegree $(i,j)$. We will use the function $R_z(t)=\frac{1-t^{z}}{1-t^2}$, which is smooth in $t>0$ and analytic in $z\in\mathbb C$. \begin{Proposition}\label{prop:primitive} For $z\in \mathbb{C}$, $z\neq -1$, define $\omega_z\in{ \Omega^{1,1}(D_+)^{\SO(3)} }$ by \[\omega_z:=\frac{1}{z+1}R_{z+1}(\langle x,\xi\rangle)\langle dx,\xi\rangle\wedge \langle x, d\xi\rangle.\] Then $d_xd_\xi\omega_z=\langle x,\xi\rangle^z \sigma_x\sigma_\xi$. \end{Proposition} \proof Rather than verify the claimed equality, we will construct from scratch the form $\omega_z$ with the required derivative. We write $S^2_x$, $S^2_\xi$ for the corresponding copies of the sphere. Fix $\xi$ and use it as the North Pole with associated polar coordinates $\theta, \phi$ of $x\in S^2$, where $\cos\theta=\langle x,\xi\rangle$, and $\phi$ is the azimuth. We have $\sigma_x=\sin\theta d\theta d\phi$. It holds that $d\phi(u)=-\frac{1}{\sin\theta}d\theta(x\times u)$, and differentiating $\cos\theta=\langle x,\xi\rangle$ we find \begin{equation}\label{eq:dphi} -\sin \theta d\theta=\langle dx, \xi\rangle\Rightarrow d\phi=\frac{1}{\sin^2\theta}\langle x\times dx, \xi\rangle=\frac{\langle x\times dx, \xi\rangle}{1-\langle x,\xi\rangle^2} \end{equation} The forms $d\theta$, $\langle x\times dx,\xi\rangle=\sin^2\theta d\phi$ are linearly independent in $S^2\setminus \{\pm \xi\}$, and vanish at $\pm\xi$. Assume a 1-form $\zeta_z$ on $S^2_x$ of the form $\zeta_z=b_z(\theta)d\phi$ is given, such that $d_\xi\omega_z= \zeta_z \sigma_\xi.$ Then \[ d\zeta_z=b_z'(\theta) d\theta d\phi+ c_0 b_z(0)\delta_\xi, \] where $c_0=2\pi$ and $\delta_\xi$ is the delta measure at $\xi$. Thus if $b_z(0)=0$, and $b_z'(\theta)=\sin\theta\cos^z\theta$, the desired equation would be satisfied. For this we take \[b_z(\theta)=\frac{1-\cos^{z+1}\theta}{z+1 } \] Now fix $x$. We should find an $\mathbb{R}^3$-valued $\Stab(x)=\SO(2)$-equivariant $1$-form $\eta_z$ on $S^2_\xi$ with \begin{equation}\label{eqn:eta_differential}d_\xi\eta_z=\frac{1-\langle x,\xi\rangle^{z+1}}{1-\langle x,\xi\rangle^2} \sigma_\xi \xi, \end{equation} whence we may take $\omega_z= \frac{1}{z+1} \langle x\times dx, \eta_z\rangle $. It is easy to verify that $$\eta_z=a_z(\theta)d\phi \cdot x+c_z(\theta)d\theta \cdot x\times\xi,$$ where $c_z(\theta)=\frac{1-\cos^{z+1}\theta}{\sin\theta}$ and $a_z(\theta)=\int_0^\theta c_z(\tau)\cos\tau d\tau$ satisfies \eqref{eqn:eta_differential}. The verification boils down to the equality $$\frac{1-\cos^{z+1}\theta}{\sin\theta}\cos\theta d\theta d\phi\cdot x-\frac{1-\cos^{z+1}\theta}{\sin\theta}d\theta \cdot x\times d\xi=\frac{1-\cos^{z+1}\theta}{\sin\theta}d\theta d\phi\cdot \xi,$$ which in turn follows from \eqref{eq:dphi} applied to $S^2_\xi$, by taking scalar products with $x,\xi$ and $x\times\xi$. We find that \[\omega_z=-\frac{1}{z+1}\langle x\times dx, \eta_z\rangle=\frac{1}{z+1}\frac{1-\langle x,\xi\rangle^{z+1}}{1-\langle x,\xi\rangle^2}\langle dx,\xi\rangle\wedge \langle x, d\xi\rangle.\] \endproof It follows by the Stokes theorem and using continuity to pass to the boundary of $D_+$, that for $\Re z\geq 0$, \begin{equation}\label{eq:stokes_beta}B_K(z) = \int_{ K\times K^\vee} \langle x,\xi\rangle^z = \int_{\partial K\times \partial K^\vee} \omega_z. \end{equation} \begin{exmp} Let us compute $B_K(z)$ for the spherical disc $K=\{x_3\geq \frac{1}{\sqrt 2}\}$, satisfying $K=K^\vee$. Parametrize $\partial K$ by $x(\alpha)=\frac{1}{\sqrt 2}(\cos \alpha, \sin\alpha, 1)$, and $\partial K^o$ by $\xi(\beta)=\mathcal L(x(\beta))=\frac{1}{\sqrt 2}(-\cos \beta, -\sin\beta, 1)$, $0\leq \alpha,\beta\leq 2\pi$. Then: \[\langle x,\xi\rangle =\frac12(1-\cos(\alpha-\beta))=\sin^2\frac{\alpha-\beta}{2}\] \[ dx=\frac{1}{\sqrt 2}(-\sin\alpha,\cos\alpha,0)d\alpha,\quad d\xi=\frac{1}{\sqrt 2}(\sin\beta,-\cos\beta,0)d\beta \] \[\langle dx,\xi\rangle = \frac12\sin(\alpha-\beta)d\alpha,\quad \langle d\xi,x\rangle =\frac12\sin(\beta-\alpha)d\beta \] Thus \begin{align*} B_K(z)&=-\frac{1}{4(z+1)} \int_0^{2\pi}\int_0^{2\pi}\frac{1-(\sin^2\frac{\alpha-\beta}{2})^{z+1}}{1-\sin^4\frac{\alpha-\beta}{2}}\sin^2(\alpha-\beta)d\alpha d\beta \\& =-\frac{1}{2(z+1)}\int_{-\pi}^{\pi}(2\pi-2|\gamma|)\frac{1-(\sin^2\gamma)^{z+1}}{1-\sin^4\gamma}\sin^22\gamma d\gamma\\&= -\frac{8}{z+1}\int_0^{\pi}(\pi-\gamma)\frac{1-(\sin^2\gamma)^{z+1}}{1+\sin^2\gamma}\sin^2\gamma d\gamma\\&= -\frac{8}{z+1}\int_0^{\pi}\frac \pi 2\frac{1-(\sin^2\gamma)^{z+1}}{1+\sin^2\gamma}\sin^2\gamma d\gamma \\&= -\frac{8\pi}{z+1}\int_0^{\frac\pi2}\frac{1-(\sin^2\gamma)^{z+1}}{1+\sin^2\gamma}\sin^2\gamma d\gamma, \end{align*} where the penultimate equality is by symmetrizing the integrand around $\frac\pi 2$. Putting $t=\sin^2\gamma$ we get \begin{align*} B_K(z)&=-\frac{4\pi}{z+1}\int_0^1\frac{1-t^{z+1}}{1+t}\frac{\sqrt t}{\sqrt{1-t}}dt=-\frac{4\pi}{z+1}(\pi(1-\frac{1}{\sqrt 2})-\int_0^1 \frac{t^{z+\frac 32}dt}{(1+t)\sqrt{1-t}} ) \\&=-\frac{4\pi}{z+1}\left(\pi(1-\frac{1}{\sqrt 2})-\sqrt\pi \frac{\Gamma(z+\frac 52)}{\Gamma(z+3)}{}_2F_1(1,z+\frac52,z+3;-1)\right) .\end{align*} In particular, $B_K(-3)=2\pi^2$. \end{exmp} \begin{Proposition} For $K\subset S^2$ smooth and strictly convex, $B_K(z)$ extends as a meromorphic function with simple poles that are contained in $\{-\frac52, -\frac72,-\frac 92,\dots\}$. \end{Proposition} \proof Let $\mathcal L:\partial K\to \partial K^\vee$ be the Legendre transform, defined by $\mathcal L x=\partial K^\vee\cap x^\perp$. It points in the unique direction perpendicular to both $x$ and $T_x\partial K$: since $\langle x,\xi\rangle:\partial K\times\partial K^\vee\to \mathbb{R}$ is smooth and non-negative, its zeros are all minima and hence critical points. That is, for fixed $\xi=\mathcal L x$, we have $\langle dx,\xi\rangle =0$. Parametrize $\partial K$ by arc-length, denoted $x(\alpha)$, and $\partial K^\vee$ by $\xi(\beta)=\mathcal L(x(\beta))$. It then holds that \[ \langle x(\alpha), \xi(\beta)\rangle =(\beta-\alpha)^2 u(\alpha,\beta)\] with $u$ smooth, and by strict convexity $u>0$. Similarly we have $$\langle x'(\alpha),\xi(\beta)\rangle =(\beta-\alpha) v_1(\alpha,\beta),\quad \langle \xi'(\beta),x(\alpha)\rangle =(\beta-\alpha) v_2(\alpha,\beta),$$ with $v_1,v_2$ smooth and non-zero when $|\alpha-\beta|\leq \epsilon$. Here $\epsilon>0$ is fixed, and depends on the lower bound on the curvatures of $K, K^\vee$. Recall eq. \eqref{eq:stokes_beta}: \[B_K(z) =-\frac{1}{z+1}\int_{\partial K\times \partial K^\vee} \frac{1-\langle x,\xi\rangle^{z+1}}{1-\langle x,\xi\rangle^2}\langle dx,\xi\rangle\wedge \langle d\xi,x\rangle .\] Writing $\beta=\alpha+\phi$, it follows that up to an addition of an entire function of $z$ and the factor $\frac{1}{z+1}$, $B_K(z)$ equals \begin{align*}B^\epsilon_K(z)&:=\int_{-\epsilon}^\epsilon d\phi\int_{\alpha=0}^{2\pi}\frac{1-\langle x(\alpha),\xi(\alpha+\phi)\rangle^{z+1}}{1-\langle x(\alpha),\xi(\alpha+\phi)\rangle^{2}} \langle x'(\alpha), \xi(\alpha+\phi)\rangle \langle \xi'(\alpha+\phi), x(\alpha)\rangle d\alpha \\ &=h_\epsilon(K)-\int_{-\epsilon}^\epsilon|\phi|^{2z+4} d\phi\int_{\alpha=0}^{2\pi} u(\alpha,\alpha+\phi)^{z+1}\tilde v(\alpha,\phi)d\alpha , \end{align*} where $h_\epsilon(K)$ is a constant depending only on $K$, and $ \tilde v$ is a smooth non-zero function. The inner integral, denoted $f_z(\phi)$, is an analytic family of smooth functions as $z\in\mathbb C$. The function $\int_{-\epsilon}^\epsilon|\phi|^{2z+4} f_z(\phi)d\phi$ is then meromorphic in $\mathbb{C}$, with simple poles that are contained in a subset of $\{-\frac52,-\frac 72,\dots\}$, corresponding to the simple poles of the meromorphic family $\{|\phi|^{2z+4}\}$ of even homogeneous distributions on $\mathbb{R}$. Also by Lemma \ref{lem:convergence_domain}, $z=-1$ is an analytic point of $B_K(z)$. This completes the proof of the meromorphic extendibility of $B_K(z)$ with poles as stated. \endproof \begin{Proposition} The value $B_K(-3)$ is a projective invariant of $K\subset\mathbb{R}\mathbb{P}^2$. \end{Proposition} \proof Recall the symbol $\chi_g(x,\xi)$ defined in Lemma \ref{lem:lemma91}. For any fixed $g\in \PGL(V)$, we have by Lemma \ref{lem:measure_push} the following equality for $\textrm{Re}z>-1$: \begin{align}\label{eq:invariance}&\langle \mathbbm 1_{g^{-1}K\times (g^{-1}K)^\vee}, \mu_z\rangle=\langle g^*\mathbbm1_{K\times K^\vee},\mu_z\rangle=\langle \mathbbm1_{K\times K^\vee}, g_*\mu_z\rangle\\=&\langle \chi_{g^{-1}}^{-3-z}\mathbbm 1_{K\times K^\vee}, \mu_z\rangle.\nonumber\end{align} We already established that one end (and hence both) of this equation admits a meromorphic extension in $z\in \mathbb C$, which is analytic at $z=-3$. Let us verify that the value of the right hand side at $z=-3$ equals $B_K(-3)$, implying projective invariance. Fix $g\in \PGL(3)$, and consider $h(z)=\langle \chi_{g^{-1}}^{-3-z} \mathbbm 1_{K\times K^\vee}, \mu_z\rangle=\int_{K\times K^\vee} \chi_{g^{-1}}^{-3-z} d\mu_z$. Write $\psi(x,\xi)=\chi_{g^{-1}}(x,\xi)$, and note $\psi$ is a smooth positive function. We now use Proposition \ref{prop:primitive} and the Stokes theorem to write \begin{align*}h(z)&= \int_{K\times K^\vee} \psi(x,\xi)^{-z-3}d_xd_\xi\omega_z\\&=\int\limits_{\partial K\times K^\vee}\psi^{-z-3}d_\xi\omega_z+(z+3)\int\limits_{K\times K^\vee}\psi^{-z-4} d_x \psi\wedge d_\xi \omega_z \\ &=\int\limits_{\partial K\times \partial K^\vee}\psi^{-z-3}\omega_z- (z+3)\int\limits_{K\times K^\vee} d_\xi(\psi^{-z-4}d_x \psi)\wedge\omega_z \\ &+(z+3)\int\limits_{\partial K\times K^\vee}\psi^{-z-4}d_\xi \psi\wedge\omega_z +(z+3)\int\limits_{ K\times \partial K^\vee}\psi^{-z-4} d_x \psi\wedge \omega_z\end{align*} The first summand admits a meromorphic extension by the same proof as that for $B_K$, and its value at $z=-3$ is $\int_{\partial K\times \partial K^\vee}\omega_{-3}=B_K(-3)$. Consider next the second summand; by Lemma \ref{lem:convergence_domain}, the form $\omega_z$ is $L^1$ in $K\times K^\vee$ for $\Re z>-\frac72$, implying the second summand is analytic in this range, and vanishes at $z=-3$. The last two summands are treated similarly, so let us focus on the first one. Written explicitly, it is $$\frac{z+3}{z+1}\int_{\partial K\times K^\vee} \frac{1-\langle x,\xi\rangle^{z+1}}{1-\langle x,\xi\rangle^2}\langle dx,\xi\rangle\wedge\langle x,d\xi\rangle\wedge \psi^{-z-4}d_\xi\psi.$$ It is easy to see that up to a summand that is analytic near, and vanishes at, $z=-3$, the integral equals $$I=\frac{z+3}{z+1}\int_{\partial K\times K^\vee} \langle x,\xi\rangle^{z+1}\langle x,d\xi\rangle \wedge\langle dx,\xi\rangle\wedge \psi^{-z-4}d_\xi\psi.$$ It holds that $$\frac{1}{z+2}d(\langle x,\xi\rangle^{z+2})\wedge\langle dx,\xi\rangle\wedge \psi^{-z-4}d_\xi\psi=\langle x,\xi\rangle^{z+1}\langle x, d\xi\rangle \wedge \langle dx, \xi\rangle\wedge \psi^{-z-4}d_\xi\psi.$$ Consequently we can integrate by parts to obtain \begin{align*}I=I_1+I_2&=\frac{z+3}{(z+1)(z+2)}\int\limits_{\partial K\times\partial K^\vee}\langle x,\xi\rangle ^{z+2} \langle dx,\xi\rangle\wedge\psi^{-z-4}d_\xi\psi\\&-\frac{z+3}{(z+1)(z+2)}\int_{\partial K\times K^\vee} \langle x,\xi\rangle ^{z+2} d(\langle dx,\xi\rangle \wedge\psi^{-z-4}d_\xi\psi). \end{align*} Let us rewrite the first summand as \begin{align}\nonumber I_1&=\frac{1}{(z+1)(z+2)}\int\limits_{\partial K\times \partial K^\vee}d_x(\langle x,\xi\rangle^{z+3})\wedge \psi^{-z-4}d_\xi\psi\\&=- \frac{1}{(z+1)(z+2)}\int\limits_{\partial K\times\partial K^\vee} \langle x,\xi\rangle^{z+3}d_x(\psi^{-z-4}d_\xi\psi)\label{eq:I1_2} \end{align} We use the arc-length parametrization $x(\alpha)$, $0\leq \alpha\leq L$ for $\partial K$, and $\xi (\beta)=\mathcal L(x(\beta))$, $0\leq \beta\leq L$ for $\partial K^\vee$. Recall that by strict convexity, $\langle x(\alpha), \xi(\beta)\rangle=|\beta-\alpha|^2u(\alpha, \beta)$ with $u>0$ smooth. We will use $C$ to denote a positive constant that only depends on $K$ and $g$, which may differ between occurrences. Consider the integral in \eqref{eq:I1_2}. As $z\to-3$, the integrand is majorized by $C|\langle x, \xi\rangle|^{-1/4}$, which is integrable on $\partial K\times\partial K^\vee$. Thus by Lebesgue's dominated convergence theorem, $$\lim _{z\to -3} I_1=-\frac12\int_{\partial K\times\partial K^\vee}d_x(\psi^{-1}d_\xi\psi) =0.$$ Next we consider $I_2$. Let $P: S^2\to \partial K^\vee$ be the nearest-point projection, well-defined in some open neighborhood $U$ of $\partial K^\vee$. Define $K^\vee_\epsilon=\{\xi\in U: \langle \xi, \mathcal L(P\xi)\rangle \leq \epsilon\}$, which is a one-sided tube around $\partial K^\vee$. We parametrize $\partial K\times K^\vee_\epsilon$ by $(x(\alpha), \xi(\beta, s))$, $0\leq \alpha,\beta\leq L$, $0\leq s\leq \epsilon$, where $\xi=\xi(\beta, s)$ is the point that has $P(\xi)=\xi(\beta)$, $\langle \xi, \mathcal L (P(\xi))\rangle=s$. Write $\eta=d(\langle dx,\xi\rangle \wedge\psi^{-z-4}d_\xi\psi)$. We claim that the integral \begin{equation}\label{eq:integral}\int_{\partial K\times K^\vee}\langle x,\xi\rangle^{z+2}\eta\end{equation} converges for $\Re z>-\frac 72$. It is enough to consider the integral \eqref{eq:integral} in the smaller domain $\{|\alpha-\beta|\leq \epsilon, 0\leq s\leq \epsilon\}$. There $\langle x(\alpha), \xi(\beta, s)\rangle \geq C\max((\beta-\alpha)^2, s)\geq C((\beta-\alpha)^2+s) $. It remains to check that for $\Re z>-\frac72$, \[ \int_0^\epsilon\int_0^\epsilon\frac{dsd\phi}{(s+\phi^2)^{-z-2}}<\infty, \] which is clear. Consequently, $\lim_{z\to -3} I_2=0$ due to the $(z+3)$ factor. This concludes the proof of the statement. \endproof Let us derive some equivalent expressions for $B_K(-3)$, working towards an explicit geometric formula. One has \begin{align*}B_K(-3)&=-\frac12\left. \int_{\partial K\times\partial K^\vee}\frac{1-\langle x,\xi\rangle^{z+1}}{1-\langle x,\xi\rangle ^2}\langle dx,\xi\rangle\wedge\langle x,d\xi\rangle\right|_{z=-3} \\&=\frac12 \left.\int_{\partial K\times\partial K^\vee}\langle x,\xi\rangle^{z+1} \frac{1-\langle x,\xi\rangle^{-z-1}}{1-\langle x,\xi\rangle ^2}\langle dx,\xi\rangle\wedge\langle x,d\xi\rangle\right|_{z=-3}\\&=\frac12 \left.\int_{\partial K\times\partial K^\vee}\langle x,\xi\rangle^{z+1}\langle dx,\xi\rangle\wedge\langle x,d\xi\rangle\right|_{z=-3} \\&=\frac1{2(z+2)}\left.\int_{\partial K\times\partial K^\vee}\langle dx,\xi\rangle \wedge d(\langle x, \xi\rangle^{z+2})\right|_{z=-3} \\&=\frac1{2}\left.\int_{\partial K\times\partial K^\vee}\langle x,\xi\rangle^{z+2}\langle dx,d\xi\rangle\right|_{z=-3}. \end{align*} \begin{align}B_K(-3)&=\frac1{2}\left.\int_{\partial K\times\partial K^\vee}\frac{\langle dx,d\xi\rangle}{\langle x,\xi\rangle^z}\right|_{z=1} \label{eq:mahler_volume} \\&=\frac12\left.\int_{\partial K\times\partial K^\vee} \frac{\langle dx,\xi\rangle\wedge \langle x,d\xi\rangle}{\langle x,\xi\rangle^z}\right|_{z=2}\label{eq:mahler_volume2} \end{align} \begin{Proposition} Let $\partial K$ be parametrized by arc-length as $x(\alpha)$. The geodesic curvature of $\partial K$ is $\kappa(\alpha)=\det (x(\alpha), x'(\alpha), x''(\alpha))$. Then $$B_K(-3)= \frac12 \lim_{\epsilon\to 0}\int_{x\in\partial K, \xi\in \partial K^\vee, \langle x,\xi\rangle \geq\epsilon^2}\frac{\langle dx,\xi\rangle\wedge\langle x,d\xi\rangle}{\langle x,\xi\rangle^2}+\frac{2\sqrt 2}{\epsilon}\int_{\partial K}\sqrt {\kappa(\alpha )}d\alpha.$$ \end{Proposition} \proof We carry out the regularization explicitly, using formula \eqref{eq:mahler_volume2}. Put $L=\textrm{Length}(x(\alpha))$, $\xi(\beta)=\mathcal L(x(\beta))=x(\beta)\times x'(\beta)$, and let $\alpha,\beta\in\mathbb{R}$ be $L$-periodic parameters. We will write $x(\alpha)>\xi(\beta)$ if $\beta<\alpha<\beta+\frac{L}{2}$. Define for small values $|t|\leq \epsilon$ \[g_K(t)=\left\{\begin{array}{cc}\int_{\langle x,\xi\rangle =t^2, x>\xi}\langle dx,\xi\rangle&,\qquad t\geq0 \\ \int_{\langle x,\xi\rangle =t^2, x<\xi}\langle dx,\xi\rangle&,\qquad t<0 \end{array}\right.\] Clearly $g_K$ is smooth in $[-\epsilon,\epsilon]$, and $g_K(0)=0$. Let $S(x,\xi)=\sqrt{\langle x,\xi\rangle}$ be a signed square root defined on $\partial K\times \partial K^\vee$ , which is non-negative for $x\geq \mathcal L(\xi)$, and non-positive for $x\leq L(\xi)$. Put $t=S(x,\xi)$, so $t^2=f_0(x,\xi)=\langle x,\xi\rangle$ and $2tdt=\langle dx,\xi\rangle +\langle x,d\xi\rangle$, and $\langle dx,\xi\rangle \wedge\langle x,d\xi\rangle=-2tdt\langle dx,\xi\rangle$. It then holds that \begin{align*}B_K(-3)&= \frac12 \int_{\langle x,\xi\rangle \geq\epsilon^2}\frac{\langle dx,\xi\rangle\wedge\langle x,d\xi\rangle}{\langle x,\xi\rangle^2} - \int_0^{\epsilon}\frac{g_K(t)-g_K(-t)-2g_K'(0)t}{t^3}dt +\frac{2}{\epsilon}g_K'(0) , \end{align*} and the middle summand approaches $0$ as $\epsilon\to 0$. It remains to compute $g'_K(0)$. Note that \[g_K(-t)= \int_{S(x,\xi)=-t}\langle dx,\xi \rangle =- \int_{S(\xi,x)=t}\langle d\xi,x\rangle =-g_{K^\vee}(t).\] Parametrizing $\alpha=\beta+s_t(\beta)$, we have \[t^2=\langle x(\alpha), \xi(\beta)\rangle =\frac12 s_t(\beta)^2\kappa(\beta)+O(s_t^3).\] Thus \[s_t(\beta)=\frac{\sqrt2}{\sqrt{\kappa(\beta)}}t+ O(t^2),\] and writing \[g_K(t)= \int_0^L \langle x'(\beta +s_t(\beta)),x(\beta)\times x'(\beta)\rangle d\beta,\] \[\langle x'(\beta +s_t(\beta)),x(\beta)\times x'(\beta)\rangle= \sqrt{2\kappa(\beta)}t+O(t^2), \] we find \[g_K'(0)=g_{K^\vee}'(0)= \sqrt 2\int_0^L\sqrt{\kappa(\beta)}d\beta. \] \endproof \bibliographystyle{abbrv}

a07277f53aa549bdca2fc168455ff116be0a9670

\section{Introduction} A key objective of mobile robotic systems is to execute safe, reliable motion while avoiding obstacles in the shortest amount of time possible. While mobile robots have demonstrated considerable success in recent years, they still fail to maneuver in environments at the speed and agility of humans. Traditional navigation algorithms require explicit perception, mapping, localization and control for collision free motion. Often, high-speed navigation using these traditional approaches is severely limited by the sensor's field of view (FOV) as well as the computational requirements needed for explicit mapping. This requires a robot to frequently reduce its speed to rescan the environment, construct a map and replan a new trajectory. This is in contrast to human navigation where cognitive psychologists have hypothesized that humans (i) actively differentiate between occupied and free spaces based on observations, (ii) make predictions of occupied spaces beyond line of sight and (iii) use these predictions for navigation to help improve robustness and agility~\cite{doi:10.1177/105971230000800301, buckner2010role,Schwartenbeck411272}. \begin{figure}[] \centering \includegraphics[width=0.70\columnwidth]{images/first_image_option1.png} \caption{A motivating example of our approach. We leverage occupancy map prediction to allow the robot to plan more robust trajectories at higher speeds compared to those limited by the sensor horizon.} \label{fig:first_figure} \end{figure} In this paper, we lay the foundation for developing algorithms that provide these capabilities to robotic systems. Our intuition is that as humans navigate, they leverage spatial cues within the environment to generate predictions of future spaces and use that as part of the planning process. Our objective is to mimic this predictive capability in robotic systems (see Fig.~\ref{fig:first_figure}). Our approach similarly learns to predict spaces beyond the line of sight and further uses the predicted areas as part of the robot controller for planning. Our specific contributions include: \begin{itemize} \item Novel perception algorithms that predict future occupancy maps using generative neural networks. \item A controller that leverages the predicted occupancy map during planning. \item Real-world hardware experiments using a robotic car to demonstrate higher speed navigation with improved reliability compared to a controller not operating on predicted regions of the map. \end{itemize} \begin{figure*}[] \centering \includegraphics[width=1.75\columnwidth]{images/hs_sys_diagram.png} \caption{This diagram describes the overall perception and control pipeline. The perception module receives on-board sensor data from the car and produces a predicted occupancy map using a U-Net style generative neural network. The control algorithm receives the robot state, predicted occupancy map and goal point and generates collision-free trajectories. } \label{fig:hs_sys_diagram} \end{figure*} \section{Related Work} Previous work has presented different strategies to predict unknown parts of the ma . A variety of methods incorporate map predictions to speed exploration of an environment \cite{chang2007p,shrestha2019learned,Katyal2019,katyal2018,ramakrishnan2020occupancy}. Predominantly, these methods are applied to exploration and therefore do not stress the capability of the map prediction module with respect to the control pipeline. In particular, this paper extends our prior work, \cite{Katyal2019}, by using a balanced classification loss instead of a regression loss in addition to data augmentation and noise suppression techniques required to generate accurate predictions using real data. An alternative to map prediction is presented in \cite{strom2015predictive}; this method predicts the next best viewpoint for exploration by utilizing experience from previously mapped environments. Our method shows that map prediction effectively increases the amount of information available from the map and improves point-to-point high speed navigation. Adjacent to map prediction is the problem of planning the shortest path to a goal in an unknown environment. These methods perform some level of inference over the environment to inform motion planning. For example, in \cite{stein2018learning}, they learn how to plan waypoints to a goal using a partial map of the environment. Unlike our method, this approach does not consider robot dynamics. Both \cite{richter2014high} and \cite{omri2020} use prior experience to reduce the size of viable environment hypotheses. Specifically, \cite{richter2014high} learns the probability of collision of motion primitives whereas \cite{omri2020} utilizes experience-based map predictions in a belief space planner. Similar to our approach, Elfhafsi \textit{et al.} \cite{Elhafsi2020} incorporate map prediction with global path planning that respects the system dynamics, but all of their verification was done in simulation. Existing work has also considered applying reinforcement learning (RL) in tandem with occupancy maps or depth information to navigate an unknown environment. In \cite{DBLP:journals/corr/KahnVPAL17}, the method uses RGB images and predicts the probability and variance of collision while navigating towards a goal, but is incapable of global planning. In \cite{karkus2017qmdp}, the approach models the world as a POMDP and take an end-to-end approach to navigate to a goal. Additionally the method presented in \cite{li2019deep} uses a partial map as inputs to a deep RL policy to navigate in unknown environments. While these methods have shown success in application, the data requirements of deep RL policies are large and often impractical in real world scenarios. In addition, our method offers insight in the process by intermediately predicting a map. A key challenge in high-speed navigation in unknown environments is balancing the speed of navigation against the information gained from its sensors. Our method alleviates this tension by learning to infer beyond the FOV of our sensor, but many other methods alter the planning and control pipelines to be reactive to changes in the environment. For instance \cite{kousik2018bridging, lopez2017aggressive} focus on quickly planning trajectories in order to react to new obstacles. In \cite{tordesillas2020faster}, the method simultaneously plans two trajectories, a safe trajectory in known space and an aggressive trajectory into unknown spac . For instance, \cite{gao2019flying} and \cite{Florence2020} plan trajectories and verify safety using only point clouds, similar to \cite{lopez2017aggressive}. As mentioned earlier, our strategy differs from these methods as they all focus on shortening the time between sensor reading and control generation, whereas our method infers beyond the sensor FOV to amplify the information available. \section{Preliminaries} \subsection{Problem Formulation} Our objective is to enable an unmanned ground vehicle (UGV) to navigate, at a high speed to a goal position in an unknown environment using RGBD and tracking cameras for perception. Our approach makes use of RGBD mapping, map prediction, and a receding-horizon controller to achieve this objective. For map prediction, the goal of our network architecture is to learn a function that maps an input occupancy map to an expanded occupancy map that extends beyond the FOV of the sensor. More formally, we are learning the function \[ f : M_{in} \mapsto M_{out} \] \noindent where \begin{math}M_{in}\end{math} represents the input occupancy map, and \begin{math}M_{out}\end{math} represents the predicted, expanded output occupancy map. Components of the function $f$ include an encoder $f_{enc}(M_{in})\mapsto h\in \mathcal{H}$ which maps the input occupancy map to a hidden state and $f_{dec}(h) \mapsto (M_{out})$, which is a decoding function mapping the hidden state to an expanded, predicted occupancy map. The controller accepts robot odometry state, $\mathbf{x_{robot}}$, the desired goal, $\mathbf{G_{robot}}$, a dynamics model of the robot, $\mathbf{VDM_{robot}}$, and the predicted occupancy map, $M_{out}$ to produce the desired robot velocity, $v$ and turn angle, $\theta$. \subsection{Platform}\label{sec:platform} The platform we use for our evaluation is the MIT Race car \cite{cain2017high} built on the 1/10-scale Traxxas Rally Car platform, as shown in Fig.~\ref{fig:rc_car}. This RC car has a reported maximum speed of 40 m/s and contains Intel Realsense D435 and T265 cameras. The onboard Nvidia\textsuperscript{\tiny\textregistered} Jetson TX2 computer runs our perception, mapping and planning software integrated with the Robot Operating System (ROS)~\cite{ros}. Our main interface to the RC car is the variable electronic speed controller (VESC) interface that provides vehicle state information and receives commands including desired velocity and turn angle. \section{Approach} \subsection{Perception} The objective of the perception algorithm is to observe co-registered RGB and depth data to produce an occupancy map for planning. As demonstrated in Fig.~\ref{fig:hs_sys_diagram}, RTAB-Map~\cite{labbe2019rtab} is used to create 2D occupancy maps using the RGB and depth images from a Realsense D435 and visual inertial odometry from a Realsense T265. To improve our mapping performance, we limit the D435's depth sensor range to 3 meters, and we apply gradient filtering on the raw depth images. The depth image gradients are calculated using a Sobel filter with a 5 $\times$ 5 kernel. All pixels with a gradient magnitude larger than twice the median are discarded. This removes "ghost noise" near sharp edges in the image. We use RTAB-Map to generate an updated map at approximately 3 Hz while running on the Nvidia\textsuperscript{\tiny\textregistered} Jetson TX2 hardware on the car. \subsubsection{Neural Network Architecture} \begin{figure}[] \centering \includegraphics[width=0.6\columnwidth]{images/updated_rccar2.jpg} \caption{MIT Racecar with Intel Realsense Cameras} \label{fig:rc_car} \end{figure} We use a U-Net style neural network architecture~\cite{DBLP:journals/corr/RonnebergerFB15,pix2pix2017} to receive the occupancy map provided by RTAB-Map and predict an expanded occupancy map. The U-Net neural network is a generative architecture used in several image completion algorithms including~\cite{pix2pix2017,DBLP:conf/iccv/XieLLCZLWD19,DBLP:conf/eccv/YanLLZS18}. The U-Net network consists of skip connections allowing a direct connection between the layers \begin{math}i\end{math} and \begin{math}n-i\end{math}. These skip connections enable the option to bypass the bottleneck associated with the downsampling layers and significantly increases the accuracy of predicted occupancy regions~\cite{DBLP:journals/corr/abs-1803-02007}. Our implementation of the encoder network consists of 7 convolution, batch normalization and ReLU layers where each convolution consists of a $4 \times 4$ filter and stride length of 2. The number of filters for the 7 layers in the encoder network are: (64, 128, 256, 512, 512, 512, 512). Similarly, the decoder network consists of 7 upsampling layers with the following number of filters: (512, 1024, 1024, 1024, 512, 256, 128). \subsubsection{Loss Function} To train our network, we use a class-balanced cross-entropy loss function as described in~\cite{cui2019classbalancedloss}. The discrete classes used for labeling each pixel of the occupancy map include occupied, unoccupied and unknown spaces. Because there are significantly more pixels associated with unoccupied and unknown spaces versus obstacles, we apply class balancing techniques on the cross entropy loss with additional 5$\times$ weight added to the occupied space loss. This results in predictions where the edges representing obstacles are far more pronounced as seen in Fig.~\ref{fig:weight_balance}. \subsubsection{Post Processing} During our testing on the robotic car, we frequently observed small noise artifacts being generated by the neural network, a condition commonly found in generative neural networks~\cite{kaneko2020NR-GAN}. While seemingly minor and transient, these artifacts caused significant instability issues during control as the trajectory planner would often abruptly change the planned path in response to these artificial obstacles or would fail to find a valid trajectory. To alleviate this, we apply a traditional morphological closing operation with a 5 $\times$ 5 kernel to suppress the noise generated by the neural network with results shown in Fig.~\ref{fig:morphological}. The observed map and the filtered predictive map are combined to create the planning map; any unknown space from the observed map is filled in with data from the predictive map. \subsubsection{Training Details} We generate the datasets in an unsupervised manner. As the robot navigates a new environment, the robot collects data that consists of a submap that corresponds to the current occupancy map based on the sensor's horizon as well as the expanded ground truth map after the environment has been explored. We explored various sizes of the submap and found a map corresponding to 6$\,$m $\times$ 6$\,$m provided by best geometric size of the submap given the characteristics of the sensor. We used a map resolution of 0.05 meters per pixel so the input occupancy map image resolution was 120 $\times$ 120 pixels. Further, we experimented with various predicted region sizes and found predicting a region of 7.5$\,$m $\times$ 7.5$\,$m corresponding to an image size of 150 $\times$ 150 provided the optimal accuracy and performance characteristics for the controller. Further, due to limited amounts of real world data available, we perform data augmentation techniques to apply random rotations to the occupancy map training data. This allows us to be more robust to various hallway configurations as shown in Fig.~\ref{fig:data_augmentation}. \begin{figure}[] \centering \includegraphics[width=0.45\columnwidth]{images/100_no_weight.png} \includegraphics[width=0.45\columnwidth]{images/100_with_weight.png} \caption{Predicted occupancy map generated without class balancing weight (Left) and with class balancing weight (Right) where white represents unoccupied space, grey is occupied and black is unknown. The class balancing weight produces stronger edges for obstacles in the predicted occupancy map.} \label{fig:weight_balance} \end{figure} \begin{figure}[] \centering \includegraphics[width=0.45\columnwidth]{images/prediction49_output_before_morph.png} \includegraphics[width=0.45\columnwidth]{images/prediction49_output.png} \caption{(Left) Generated occupancy map with noise artifacts. (Right) Predicted occupancy map after morphological close operation (white is unoccupied space, grey is occupied and black is unknown).} \label{fig:morphological} \end{figure} \begin{figure*}[] \centering \includegraphics[width=1.6\columnwidth]{images/data_augmentation.png} \caption{Three examples from our training set of occupancy maps and their resulting expanded predictive map along with the ground truth (white is unoccupied space, light grey is occupied and dark grey unknown). Augmenting our training data with random rotations, allows the network prediction to be more robust to different environment configurations encountered by the robot.} \label{fig:data_augmentation} \end{figure*} \begin{figure}[] \centering \includegraphics[width=0.7\columnwidth]{images/prediction_example_cleaned.png} \caption{Visualization of system during hardware experiment. Known map is enclosed by the red boundary. The brown path is the smoothed RRT path to goal, and the purple path is the local optimized direct transcription trajectory.} \label{fig:hardware_experiment} \end{figure} \subsection{Control Algorithm} The objective of the control algorithm as described in Fig.~\ref{fig:hs_sys_diagram} is to compute a collision free path to the goal, generate a series of feasible trajectories to waypoints, and send control commands to the mobile robot to follow the computed trajectory. We adapt the controller proposed in our prior work,~\cite{basescu2020direct}. While initially developed for fixed-wing flight, we believe it is particularly well suited for high-speed navigation. The receding horizon allows for rapid replanning while using a dynamically built map, and the trajectory generation and tracking allows for a high-rate, dynamically feasible control output. \subsubsection{Dynamics Model} We use a simple bicycle acceleration model to describe the robot's dynamics. The equations of motion are as follows: \begin{align} \dot{x} &= v*cos(\theta) \nonumber \\ \dot{y} &= v*sin(\theta) \nonumber \\ \dot{v} &= u_0 \\ \dot{\theta} &= v * tan(\delta) / L \nonumber \\ \dot{\delta} &= u_1 \nonumber \\ \nonumber \end{align} The state is written as $\vect{x}=\begin{bmatrix} x, y, v, \theta, \delta \end{bmatrix}$ where $x$ and $y$ are 2D position, $v$ is forward velocity, $\theta$ is orientation, and $\delta$ is turn angle. The input, $\vect{u}=\begin{bmatrix} u_0, u_1 \end{bmatrix}$, represents acceleration and turn angle velocity, respectively, and $L$ is the wheel base length. \subsubsection{Control Strategy} Here, we review the receding horizon controller proposed in \cite{basescu2020direct}, which can be decomposed into three main stages. In the first stage, a path to goal is generated using a standard rapidly-exploring random tree (RRT) \cite{lavalle1998rapidly}. The resulting path is pruned and then smoothed using G2 Continuous Cubic B\'ezier Spiral Path Smoothing (G2CBS) \cite{yang2010analytical}. The curvature along the smoothed path, $\begin{bmatrix} x(s), y(s) \end{bmatrix}$, is calculated as: \begin{align} \kappa(s)=\frac{(y''(s)x'(s)-x''(s)y'(s))}{(x'(s)^2+y'(s)^2)^{\frac{3}{2}}} \nonumber \end{align} This curvature is then mapped to velocity based upon $v_{max}$ and $v_{min}$, the maximum and minimum velocity. \nonumber \begin{align} v(s)=\frac{d\vect{x}}{dt}(s)=v_{max}-\kappa(s)*\frac{v_{max}-v_{min}}{2} \end{align} The path's velocity parameterization is used to reparametrize the path by time. \begin{align} t=\int_{0}^{s} \frac{1}{v(s)}ds \end{align} In our implementation, we modified the RRT to improve performance in a dynamically built map by initializing the RRT tree with the raw RRT path to goal from the previous control iteration. Before initialization, we check the path for collisions and truncate it if a collision is detected. This initialization results in faster RRT computation and increased path consistency between iterations. In the second stage, a dynamically feasible trajectory from the current state to a horizon point is generated. The horizon point is selected as a time horizon selected along the parameterized RRT path. We utilize the same direct transcription feasibility problem as formulated in \cite{basescu2020direct} and refer the reader to their formulation. This approach discretizes the trajectory into $N$ knot points using a variable time interval $dt$. Let $\mathbf{x}_0(t_k)$ be the position of the robot and $\mathbf{u}_0(t_k)$, $k < N$, be the input at the $k^{th}$ knot point where $t_{k+1} = t_k + dt$. We modify this feasibility problem by introducing a cost function to penalize large $dt$ (therefore encouraging high speeds) as well as slightly penalizing the input to encourage smoother trajectories. The objective function is shown below. \begin{align} J(\mathbf{x}_0(t_k), \mathbf{u}_0(t_k), dt) = \sum_{k=0}^{N-1} \mathbf{u}_0(t_k)^T \enspace \mathbf{R_c} \enspace \mathbf{u}_0(t_k) + dt \end{align} In order to track the dynamically feasible trajectory, time-varying LQR (TVLQR) is performed in the third stage. The control signal is generated as: \begin{align} \mathbf{u}(t_k,\mathbf{x}) &= \vect{K}(t_k)(\mathbf{x}-\mathbf{x}_0(t_k))+\mathbf{u}_{0}(t_k). \end{align} where $\vect{K}(t_k)$ is the optimal gain matrix. \subsubsection{Control Parameters} The control pipeline executes at a rate of 5$\,$Hz, or a control interval of $T = 0.2 \,$seconds. The control signal is calculated from the odometry and TVLQR gains at a rate of 50$\,$Hz. The max time and max iterations for the RRT search were set to 0.05$\,$seconds and 20000 iterations. The maximum velocity for RRT path parameterization was set to the maximum allowed velocity specified in each hardware trial. For RRT sampling, we use an an obstacle avoidance radius of 0.4$\,$m. For direct transcription, we used $N = 10$ knot points and a time horizon of $H = 2\,$seconds. We set $\vect{\delta}_f = \begin{bmatrix} 0.1, 0.1, 0.1, 0.25, 100.0 \end{bmatrix}$, $\mathbf{R_c} = diag(0.1, 0.1)$ and use an obstacle radius of 0.35$\,$m. The costs for TVLQR are as follows: \begin{align} Q &= diag(10, 10, 10, 10, 10) \nonumber \\ Q_f &= diag(1, 1, 5, 1, 1) \nonumber \\ R &= diag(1, 1) \nonumber \\ \nonumber \end{align} \begin{figure*} \centering \includegraphics[width=0.48\columnwidth]{images/1603893080305236947.jpg} \includegraphics[width=0.48\columnwidth]{images/1603893082841778333.jpg} \includegraphics[width=0.48\columnwidth]{images/1603893085907992712.jpg} \includegraphics[width=0.48\columnwidth]{images/1603893088442458503.jpg} \caption{This sequence of images represents a trajectory taken by the robot during our experimental evaluation.} \label{fig:car_isc_pics} \end{figure*} The acceleration and turn angle control bounds were set to [-2.5, 2.5] m/s and [-1.5, 1.5] rad/s respectively. The velocity minimum bound was set to 0.5 m/s and turn angle state bounds were set to [-0.3, 0.3] rad. \section{Experimental Evaluation} We conduct preliminary hardware experiments to validate our approach using a robotic car based on MIT's open-source race car~\cite{cain2017high} as described in~\ref{sec:platform}. Our map prediction network was trained on indoor scenes consisting primarily of straight and turning corridors. Similar to~\cite{DBLP:journals/corr/abs-2002-05700}, we conduct both zero-shot and continual learning scenarios. In our zero-shot experiments, we evaluate the performance on new environments (Fig.~\ref{fig:car_isc_pics}) not seen by the robot. In the continual learning evaluation, we allow the robot to collect data in a semi-supervised manner from the new environment, fine tune the network offline and reevaluate performance. We assessed the maximum speed allowed by the robot and the number of successful trials, which we defined as reaching the target goal without collision. A visualization of our system during the hardware experiment is shown in Fig. \ref{fig:hardware_experiment}. The results are summarized in Table~\ref{table:results}. Without map prediction, the robot's maximum speed, $v_{max}$, was 3 m/s without collision. When evaluating without map prediction with the maximum speed of 4 m/s, the robot was only able to successfully reach the goal 1/5 attempts. With map prediction, we were able to achieve success 3/5 trials. After allowing the network to fine tune on the new environment, we were able to increase the ratio to 4/5 successful trials showing the ability to continually learn as the robot explores new environments. For comparison, the trajectories with and without prediction for one of the trials where $v_{max}=4\,$m/s is shown in Fig.~\ref{fig:hardware_trajectories}. Without map prediction, the robot limited by the sensor's field of view, plans a waypoint in unknown space and is not able to react in time once the map has been updated to reflect the true occupied space. In contrast, with map prediction, we are able to plan with longer horizons resulting in smoother trajectories that allow the robot to reach the desired goal. \begin{table}[] \centering \setlength{\tabcolsep}{1.2mm} \renewcommand{\arraystretch}{1.1} \begin{tabular}{|c|c|c|} \hline \textbf{Algorithm} & \textbf{Max Speed} & \textbf{Success Rate} \\ \hline \hline Without Map Prediction & 3 m/s & 5/5 \\ \hline Without Map Prediction & 4 m/s & 1/5 \\ \hline With Map Prediction (Zero-shot) & 4 m/s & 3/5 \\ \hline With Map Prediction (Fine Tuned) & 4 m/s & 4/5 \\ \hline \end{tabular} \caption{This table captures the results of our preliminary hardware experiments on our modified MIT race car. } \label{table:results} \end{table} \begin{figure} \centering \includegraphics[width=1\columnwidth]{images/trajectories.png} \caption{Example trajectories of car with max velocity of 4 m/s with and without map prediction.} \label{fig:hardware_trajectories} \end{figure} \section{Discussion and Conclusion} In this paper, we describe an approach that enables high speed navigation based on predicted occupancy maps. Specifically, we present a generative neural network architecture approach to collect self-supervised training data and training methodology that allows the robot to predict occupied spaces beyond the line of sight of the camera. Further, we present several engineering solutions in the perception algorithm that were required for prediction to work on real data on the physical hardware including data augmentation, using a class-balanced loss function and traditional computer vision methods for noise suppression. In addition, we also present a real-time controller that leverages the predicted occupancy map as part of the planning algorithm. At a max velocity of 3 m/s, 3 Hz mapping rate, 3 m sensor range, and 1 sec time horizon, planned trajectories will almost always be within the known region of the map. It isn't surprising, thus, that the system is successful without map prediction with these parameters. When max velocity is increased to 4 m/s, planned trajectories will often be within unknown space of the map (outside of the sensor range). This can cause trajectories to plan through unseen walls, causing failure. The predicted occupancy map is able to help address this limitation by providing a longer horizon for planning which accounts for the improved performance at 4 m/s in our preliminary hardware evaluation. While the results are promising, there are many opportunities for future work. One area to explore is to extend our preliminary hardware experiments to more thorough training and testing in various indoor and outdoor environments to assess the impact of map prediction for high speed navigation. Another area of future work is to attempt to bypass explicit mapping completely and develop prediction techniques on raw sensor data. We believe this would improve the performance significantly as map prediction is currently the bottleneck from a computational perspective. Another area to explore is to improve continual learning techniques as the robot enters new environments. It is unrealistic to expect the training data to capture the full distribution of the environments that the robot will expect to see. We believe further research is needed to improve the data efficiency of continual learning techniques so that the robot can improve performance in real time as it explores new environments. In spite of these limitations, we believe we have shown the promise of developing predictive capabilities that improve navigation performance of mobile robots and are continually developing techniques to further extend these capabilities as described above. \clearpage \bibliographystyle{IEEEtran}

6896e5c856770f0bf1e32dab8b9fb4b4c1a27de1

\section{Introduction: Quantum chaos conjecture and many-body systems} \label{intro} The ubiquitousness of {\em random matrix theory} (RMT) \cite{Mehtabook,sarnak,forrester} descriptions for a diverse range of phenomena in Nature and Society is the example par excellence of effectiveness of mathematics. In quantum dynamical systems, the fact that the fluctuations in the spectra of unitary evolution operators can be described in terms of structureless ensembles of RMT, characterised solely by unitary and anti-unitary symmetries, has been identified as a defining property of {\em quantum chaos}. The presence of RMT spectral correlations has been related to Hamiltonian chaos of the corresponding limiting classical dynamical system via the so-called {\em quantum chaos conjecture}\footnote{Sometimes referred to also as Bohigas-Giannoni-Schmit conjecture.} (QCC) \cite{casati,berry,bohigas}, while the absence thereof is linked to integrability or regularity of the corresponding classical motion via the Berry-Tabor conjecture \cite{berrytabor,marklof}. A heuristic proof of QCC in terms of semiclassical periodic orbit theory has been a decades long tour de force \cite{berry85,sieberrichter,mueller}, while a rigorous proof has so far been possible only in a rather restricted setting of completely connected quantum graphs \cite{weidenmueller}. In systems which lack a small parameter (e.g. an effective Planck's constant), such as extended (many-body) systems of locally interacting quantum spins or fermions, the mechanisms for the validity of QCC have remained obscure despite a plethora of empirical evidence, see e.g. Refs.~\cite{mila,montambaux,hsu,P99,lea}. For such systems, lacking any meaningful limiting classical chaotic dynamics, the agreement of spectral fluctuations with RMT can be considered as the most versatile definition of {\em quantum chaos} and as a robust empirical method for detection of quantum (non)integrability. Recently we proposed a rigorous methodology which lead to the first proof of the emergence of RMT spectral 2-point correlation functions in the thermodynamic limit, for a particular locally interacting chain of quantum spins ${1/2}$~\cite{BKP18}. In this paper, we present a generalisation of such methodology and show that it can be extended to a much broader class of systems. In particular, we reformulate our approach in the general language of local quantum circuits --- which are the standard minimal model of quantum many-body (extended) systems with local interactions~\cite{nahum,chalker,chalker2} --- and show that it leads to exact results (in the thermodynamic limit) whenever the ``local gates" (the unitary matrices encoding the nearest-neighbour interactions of local quantum circuits) are \emph{dual-unitary}, i.e. they generate unitary evolution in both time and space. The method described here applies to quantum circuits of qudits (or arbitrary spins) and can easily account for spatially inhomogeneous interactions. Moreover, in contrast to Ref.~\cite{BKP18}, here we focus on the generic case of systems without anti-unitary symmetries (like time-reversal) while only sketch the extension of the results to generic time-reversal invariant case. In summary, in this paper we identify the key mathematical steps for approaching the problem of characterizing quantum ergodicity \cite{zelditch,alicki} and quantum chaos \cite{tolya,chalker,ljubotina} through spectral correlations in extended quantum spin lattice systems. Although our results hold in a specific setting (and in the thermodynamic limit only), they represent the first rigorous proof of the emergence of RMT behaviour in a class of extended quantum spin systems to the best of our knowledge. They pertain to both the case of quenched disorder and the clean limit, and provide the first proof of the QCC in the many-body realm. The rest of the paper is laid out as follows. In Sec.~\ref{sec:basic} we introduce the basic concepts and provide their definitions. In Sec.~\ref{sec:main} we state and interpret our main results, while in Sec.~\ref{sec:proofs} we elaborate the proofs. In Sec.~\ref{sec:discussion} we discuss some straightforward extensions and generalisations of our results. While the treatment of spatially inhomogeneous dual-unitary interactions in Sec.~\ref{Sec:inhom} is rigorous, the extensions of the techniques to study fluctuations (\ref{sec:fluctuations}) and singular disorder distributions (\ref{sec:nonisotropic}) are speculative at this point. \section{Basic concepts} \label{sec:basic} \subsection{Floquet quantum circuits} In this work we consider a class of quantum many-body systems known as \emph{Floquet local quantum circuits}. They consist of a set of $2L$, $L\in\mathbb N$, quantum variables with $d$ internal states (``qudits''), that can be thought of as arranged on a 1-dimensional periodic lattice $\Lambda_L=\frac{1}{2}\mathbb Z_{2L}$. The Hilbert space of the system is \begin{equation} \mathcal H_{2L} = (\mathcal H_1)^{\otimes 2 L} = \mathbb C^{\mathcal N}, \end{equation} where the ``local Hilbert space'' $\mathcal H_1 = \mathbb C^d$ is the Hilbert space of a single qudit and $\mathcal N = d^{2L}$ is the dimension of $\mathcal H_{2L}$. In these systems the time evolution is discrete and generated by integer powers of the unitary operator \begin{equation} \mathbb U_{L} :=\prod_{x \in \mathbb Z_{L}} \eta_{x,L}(U_{x,1}) \prod_{x \in \mathbb Z_{L}+\frac{1}{2}}\!\!\eta_{x,L}(U_{x,\frac{1}{2}}) \,, \label{eq:Floquet} \end{equation} conventionally called the ``Floquet operator''. In writing Eq.~\eqref{eq:Floquet} we introduced the following definitions. \begin{itemize} \item[(i)] We indicated by $\eta_{x,n} \!:\, {\rm End}(\mathcal H_2)\,\rightarrow\,{\rm End}(\mathcal H_{2n})$, with $n\in \mathbb N$ and $x\in \Lambda_{n}$, the linear map defined by \begin{equation} \eta_{x,n}(O):=\Pi_{2n}^{2x-1} (O \otimes\mathbbm{1}_{2(n-1)}) \Pi_{2n}^{-(2x-1)}. \end{equation} Here $\mathcal H_n=(\mathcal H_1)^{\otimes n}$ denotes the Hilbert space of a periodic qudit chain of $n$ sites, while $\Pi_n$ and $\mathbbm{1}_n$ designate respectively the periodic shift operator and the identity operator over $\mathcal H_n$. Explicitly we have \begin{equation} \Pi_n \ket{j_1}\otimes \ket{j_2}\otimes\cdots\ket{j_n} = \ket{j_n} \otimes \ket{j_1}\otimes \cdots\ket{j_{n-1}}, \end{equation} with $\Pi^n_n = \mathbbm{1}_n$, where \begin{equation} \mathcal R = \{\ket{j};\; j=0,\ldots, d-1 \}, \label{eq:realbasis} \end{equation} is the canonical orthonormal basis of $\mathcal H_1=\mathbb C^d$. \item[(ii)] We introduced the function $U_{\cdot, \cdot} : \Lambda_L \times \frac{1}{2}\mathbb Z_2\,\rightarrow\,{\rm U}(d^2)$, where ${\rm U}(N)$ denotes the group of $N\times N$ unitary matrices. The operators $U_{x,\frac{1}{2}}, U_{x,1}\equiv U_{x,0}\in{\rm U}(d^2)$ define the interaction among neighbouring qudits at sites $x-\frac{1}{2},x$ for half-odd integer and integer times respectively. These operators encode all physical information about the dynamics and will be referred to as the ``local gates''. \end{itemize} This setting can be regarded as the simplest possible modelling for extended quantum many-body systems~\cite{chalker,chalker2}. Indeed, it captures what can be considered the primary and most essential feature of an extended system, i.e., the locality of the interactions. It is precisely this feature that distinguishes an extended system from a single quantum variable with arbitrary many internal states, for example, a single (arbitrary high) spin. Note that even though local quantum circuits describe time-depended dynamics, they can also be thought of as approximations of time-independent local interactions obtained through the use of the Suzuki-Trotter decomposition \cite{suzuki,osborne}. Finally, we stress that this precise setting emerges naturally in the context of quantum simulation, for instance, through the use of the recently developed Google's Sycamore processor~\cite{google}. The time evolution generated by \eqref{eq:Floquet} admits the following convenient graphical representation \begin{align} &\mathbb U_{L}^t= \begin{tikzpicture}[baseline=(current bounding box.center), scale=0.55] \foreach \i in {0,1,2}{ \draw[thick, dotted] (-9.5,2*\i-1.7) -- (0.4,2*\i-1.7); \draw[thick, dotted] (-9.5,2*\i-1.3) -- (0.4,2*\i-1.3); } \foreach \i in {3,...,13}{ \draw[gray, dashed] (-12.5+\i,-2.1) -- (-12.5+\i,4.3); } \foreach \i in {-1,...,5}{ \draw[gray, dashed] (-9.75,3-\i) -- (.75,3-\i); } \foreach \i in{1.5,2.5,3.5}{ \draw[thick] (0.5,2*\i-0.5-3.5) arc (45:-90:0.17); \draw[thick] (-10+0.5+0,2*\i-0.5-3.5) arc (90:270:0.15); } \foreach \i in{0.5,1.5,2.5} { \draw[ thick] (0.5,1+2*\i-0.5-3.5) arc (-45:90:0.17); \draw[ thick] (-10+0.5,1+2*\i-0.5-3.5) arc (270:90:0.15); } \foreach \i in {1,2}{ \Text[x=1.25,y=-2+2*\i]{\scriptsize$\i$} } \foreach \i in {1,3}{ \Text[x=1.25,y=-2+\i]{\small$\frac{\i}{2}$} } \foreach \i in {1,3,5}{ \Text[x=-7.5+\i-2,y=-2.6]{\small$\frac{\i}{2}$} } \foreach \i in {1,2,3}{ \Text[x=-7.5+2*\i-2,y=-2.6]{\scriptsize${\i}$ } } \foreach \jj[evaluate=\jj as \j using -2*(ceil(\jj/2)-\jj/2)] in {-1,-3,-5}{ \foreach \i in {1} { \draw[thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=myred5, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[thick] (-2*\i+2,-1.35-\jj) -- (-2*\i+2.15,-1.35-\jj) -- (-2*\i+2.15,-1.5-\jj);% } \foreach \i in {2} { \draw[thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=myred4, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[thick] (-2*\i+2,-1.35-\jj) -- (-2*\i+2.15,-1.35-\jj) -- (-2*\i+2.15,-1.5-\jj);% } \foreach \i in {3} { \draw[thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=myred3, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[thick] (-2*\i+2,-1.35-\jj) -- (-2*\i+2.15,-1.35-\jj) -- (-2*\i+2.15,-1.5-\jj);% } \foreach \i in {4} { \draw[thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=myred2, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[thick] (-2*\i+2,-1.35-\jj) -- (-2*\i+2.15,-1.35-\jj) -- (-2*\i+2.15,-1.5-\jj);% } \foreach \i in {5} { \draw[thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=myred1, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[thick] (-2*\i+2,-1.35-\jj) -- (-2*\i+2.15,-1.35-\jj) -- (-2*\i+2.15,-1.5-\jj);% } } \foreach \jj[evaluate=\jj as \j using -2*(ceil(\jj/2)-\jj/2)] in {-4,-2,0}{ \foreach \i in {1} { \draw[thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=myred8, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[thick] (-2*\i+1,-1.35-\jj) -- (-2*\i+1.15,-1.35-\jj) -- (-2*\i+1.15,-1.5-\jj);% } \foreach \i in {2} { \draw[thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=myred10, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[thick] (-2*\i+1,-1.35-\jj) -- (-2*\i+1.15,-1.35-\jj) -- (-2*\i+1.15,-1.5-\jj);% } \foreach \i in {3} { \draw[thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=myred6, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[thick] (-2*\i+1,-1.35-\jj) -- (-2*\i+1.15,-1.35-\jj) -- (-2*\i+1.15,-1.5-\jj);% } \foreach \i in {4} { \draw[thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=myred9, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[thick] (-2*\i+1,-1.35-\jj) -- (-2*\i+1.15,-1.35-\jj) -- (-2*\i+1.15,-1.5-\jj);% } \foreach \i in {5} { \draw[thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=myred1, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[thick] (-2*\i+1,-1.35-\jj) -- (-2*\i+1.15,-1.35-\jj) -- (-2*\i+1.15,-1.5-\jj);% } } \Text[x=-2,y=-2.6]{$\cdots$} \Text[x=0.47,y=-2.6]{\small $L\equiv 0$} \Text[x=-4,y=-3.35]{\small $x$} \Text[x=1.25,y=4]{\small $t$} \Text[x=2,y=1]{\small$\tau$} \Text[x=1.25,y=3.2]{$\vdots$} \end{tikzpicture}, \end{align} where each local gate is represented by \begin{equation} \label{eq:Ugate} U_{x,\tau}\equiv U_{x,{\rm mod}(\tau,1)}=\begin{tikzpicture}[baseline=(current bounding box.center), scale=.7] \draw[ thick] (-4.25,0.5) -- (-3.25,-0.5); \draw[ thick] (-4.25,-0.5) -- (-3.25,0.5); \draw[ thick, fill=myred, rounded corners=2pt] (-4,0.25) rectangle (-3.5,-0.25); \draw[thick] (-3.75,0.15) -- (-3.75+0.15,0.15) -- (-3.75+0.15,0); \Text[x=-4.25,y=-0.75]{} \end{tikzpicture},\qquad x\in \Lambda_L,\quad\tau\in\frac{1}{2}\mathbb Z, \end{equation} and different shades illustrate distinct matrices. The function ${\rm mod}(x,n)$ indicates the remainder upon division by $n$. Note that leftmost and rightmost gates are connected because of periodic boundary conditions. Finally, we point out that the dynamics generated by \eqref{eq:Floquet} are \emph{time-reversal invariant} if there exist a unitary operator $\mathbb K\in {\rm U}(d^{2L})$ such that~\cite{Mehtabook} \begin{equation} \mathbb K \mathbb U_{L}^{\phantom{T}} \mathbb K^\dag = \mathbb U_{L}^T \qquad\text{and}\qquad \mathbb K^{{T}}=\pm \mathbb K\,, \label{eq:Tsym} \end{equation} where $(\cdot)^T$ denotes transposition in the canonical basis \eqref{eq:realbasis} and $(\cdot)^\dagger$ Hermitian conjugation. Symmetric and antisymmetric matrices correspond respectively to cases where the anti-unitary operator implementing time reversal on the Hilbert space squares to plus or minus one~\cite{Mehtabook}. The first, ``regular'', kind of time-reversal symmetry emerges in physical systems with integer total angular momentum and is associated with orthogonal ensembles of RMT, while the second characterises systems with half-odd integer spin and is associated with symplectic ensembles~\cite{Mehtabook}. \subsection{Our Setting} Here we consider local gates of the form \begin{subequations} \begin{align} &U_{x+\frac{1}{2},\frac{1}{2}} = (u_{x} \otimes u_{x+\frac{1}{2}})\, U = \begin{tikzpicture}[baseline=(current bounding box.center), scale=.7] \draw[ thick] (-4.5,0.75) -- (-3.25,-0.5); \draw[ thick] (-4.25,-0.5) -- (-3,0.75); \draw[ thick, fill=myred, rounded corners=2pt] (-4,0.25) rectangle (-3.5,-0.25); \draw[thick] (-3.75,0.15) -- (-3.75+0.15,0.15) -- (-3.75+0.15,0); \draw[ thick, fill=myblue3, rounded corners=2pt] (-3.25,.5) circle (.15); \draw[ thick, fill=myblue4, rounded corners=2pt] (-4.25,.5) circle (.15); \draw[thick] (-2*2.875+2.4,0.05+.5) -- (-2*2.875+2.55,0.05+.5) -- (-2*2.875+2.55,-0.1+.5); \draw[thick] (-2*2.875+1.45,-0.1+.5) -- (-2*2.875+1.45,0.05+.5) -- (-2*2.875+1.6,0.05+.5); \Text[x=-4.25,y=-0.75]{} \end{tikzpicture}\,,\label{eq:Floquetgates1}\\ &U_{x,1} =( w_{{\rm mod}(x-\frac{1}{2},L)} \otimes w_{x})\, W=\begin{tikzpicture}[baseline=(current bounding box.center), scale=.7] \draw[ thick] (-4.5,0.75) -- (-3.25,-0.5); \draw[ thick] (-4.25,-0.5) -- (-3,0.75); \draw[ thick, fill=myorange, rounded corners=2pt] (-4,0.25) rectangle (-3.5,-0.25); \draw[thick] (-3.75,0.15) -- (-3.75+0.15,0.15) -- (-3.75+0.15,0); \draw[ thick, fill=myblue1, rounded corners=2pt] (-3.25,.5) circle (.15); \draw[ thick, fill=myblue2, rounded corners=2pt] (-4.25,.5) circle (.15); \draw[thick] (-2*2.875+2.4,0.05+.5) -- (-2*2.875+2.55,0.05+.5) -- (-2*2.875+2.55,-0.1+.5); \draw[thick] (-2*2.875+1.45,-0.1+.5) -- (-2*2.875+1.45,0.05+.5) -- (-2*2.875+1.6,0.05+.5); \Text[x=-4.25,y=-0.75]{} \end{tikzpicture}\,,\qquad x\in \mathbb Z_{L}, \label{eq:Floquetgates2} \end{align} \end{subequations} where $U,W\in {\rm U}(d^2)$ act non-trivially on a pair of neighbouring qudits and $u_{x},w_{x}\in {\rm U}(d)$ on a single one (we hence represented them graphically as balls acting on a single wire). Therefore we have \begin{align} &\mathbb U_{L} = \begin{tikzpicture}[baseline=(current bounding box.center), scale=0.55] \draw[thick, dotted] (-9.5,-1.7) -- (0.4,-1.7); \draw[thick, dotted] (-9.5,-1.3) -- (0.4,-1.3); \foreach \i in {3,...,13}{ \draw[gray, dashed] (-12.5+\i,-2.1) -- (-12.5+\i,0.3); } \foreach \i in {3,...,5}{ \draw[gray, dashed] (-9.75,3-\i) -- (.75,3-\i); } \foreach \i in{0.5} { \draw[ thick] (0.5,1+2*\i-0.5-3.5) arc (-45:90:0.17); \draw[ thick] (-10+0.5,1+2*\i-0.5-3.5) arc (270:90:0.15); } \foreach \i in{1.5}{ \draw[thick] (0.5,2*\i-0.5-3.5) arc (45:-90:0.17); \draw[thick] (-10+0.5+0,2*\i-0.5-3.5) arc (90:270:0.15); } \foreach \i in {0,1}{ \Text[x=1.25,y=-2+2*\i]{\scriptsize$\i$} } \foreach \i in {1}{ \Text[x=1.25,y=-2+\i]{\small$\frac{\i}{2}$} } \foreach \i in {1,3,5}{ \Text[x=-7.5+\i+1-3,y=-2.6]{\small$\frac{\i}{2}$} } \foreach \i in {1,2,3}{ \Text[x=-7.5+2*\i-2,y=-2.6]{\scriptsize${\i}$} } \foreach \i in {0,...,4}{ \draw[thick] (-.5-2*\i,-1) -- (0.525-2*\i,0.025); \draw[thick] (-0.525-2*\i,0.025) -- (0.5-2*\i,-1); \draw[thick, fill=myorange, rounded corners=2pt] (-0.25-2*\i,-0.25) rectangle (.25-2*\i,-0.75); \draw[thick] (-2*\i,-0.35) -- (-2*\i+0.15,-0.35) -- (-2*\i+0.15,-0.5);% } \foreach \jj[evaluate=\jj as \j using -2*(ceil(\jj/2)-\jj/2)] in {0} \foreach \i in {1,...,5} { \draw[thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=myred, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[thick] (-2*\i+1,-1.35-\jj) -- (-2*\i+1.15,-1.35-\jj) -- (-2*\i+1.15,-1.5-\jj);% } \draw[ thick, fill=myblue1, rounded corners=2pt] (0.5,-1) circle (.15); \draw[ thick, fill=myblue2, rounded corners=2pt] (-0.5,-1) circle (.15); \draw[ thick, fill=myblue3, rounded corners=2pt] (-1.5,-1) circle (.15); \draw[ thick, fill=myblue4, rounded corners=2pt] (-2.5,-1) circle (.15); \draw[ thick, fill=myblue5, rounded corners=2pt] (-3.5,-1) circle (.15); \draw[ thick, fill=myblue6, rounded corners=2pt] (-4.5,-1) circle (.15); \draw[ thick, fill=myblue7, rounded corners=2pt] (-5.5,-1) circle (.15); \draw[ thick, fill=myblue8, rounded corners=2pt] (-6.5,-1) circle (.15); \draw[ thick, fill=myblue9, rounded corners=2pt] (-7.5,-1) circle (.15); \draw[ thick, fill=myblue10, rounded corners=2pt] (-8.5,-1) circle (.15); \draw[ thick, fill=myblue8, rounded corners=2pt] (0.5,0) circle (.15); \draw[ thick, fill=myblue9, rounded corners=2pt] (-0.5,0) circle (.15); \draw[ thick, fill=myblue1, rounded corners=2pt] (-1.5,0) circle (.15); \draw[ thick, fill=myblue3, rounded corners=2pt] (-2.5,0) circle (.15); \draw[ thick, fill=myblue2, rounded corners=2pt] (-3.5,0) circle (.15); \draw[ thick, fill=myblue4, rounded corners=2pt] (-4.5,0) circle (.15); \draw[ thick, fill=myblue10, rounded corners=2pt] (-5.5,0) circle (.15); \draw[ thick, fill=myblue9, rounded corners=2pt] (-6.5,0) circle (.15); \draw[ thick, fill=myblue8, rounded corners=2pt] (-7.5,0) circle (.15); \draw[ thick, fill=myblue7, rounded corners=2pt] (-8.5,0) circle (.15); \foreach \jj[evaluate=\jj as \j using -2*(ceil(\jj/2)-\jj/2)] in {0} \foreach \i in {1,...,5}{ \draw[thick] (-2*\i+1.4,-.95-\jj) -- (-2*\i+1.55,-.95-\jj) -- (-2*\i+1.55,-1.1-\jj); } \foreach \jj[evaluate=\jj as \j using -2*(ceil(\jj/2)-\jj/2)] in {0} \foreach \i in {1,...,5}{ \draw[thick] (-2*\i+2.4,0.05-\jj) -- (-2*\i+2.55,0.05-\jj) -- (-2*\i+2.55,-0.1-\jj); } \foreach \jj[evaluate=\jj as \j using -2*(ceil(\jj/2)-\jj/2)] in {0} \foreach \i in {1,...,5}{ \draw[thick] (-2*\i+1.45,-0.1-\jj) -- (-2*\i+1.45,0.05-\jj) -- (-2*\i+1.6,0.05-\jj); } \foreach \jj[evaluate=\jj as \j using -2*(ceil(\jj/2)-\jj/2)] in {0} \foreach \i in {0,...,4}{ \draw[thick] (-2*\i+.45,-1.1-\jj) -- (-2*\i+.45,-0.95-\jj) -- (-2*\i+.6,-0.95-\jj); } \Text[x=-2,y=-2.6]{$\cdots$} \end{tikzpicture}. \label{eq:diagramFloquet} \end{align} In particular, it is immediate to see that, choosing local gates \eqref{eq:Floquetgates1}--\eqref{eq:Floquetgates2} with \begin{equation} U=U^T, \qquad W=W^T, \qquad w_{x}^{\phantom{T}}=u^{T}_{x},\qquad \forall\,\, x\in \Lambda_L, \label{eq:timereversal} \end{equation} the condition \eqref{eq:Tsym} is fulfilled with \begin{equation} \mathbb K=\prod_{x \in \mathbb Z_{L}+\frac{1}{2}} \eta_{x,L}(W)= \mathbb K^T \,. \end{equation} Namely, the dynamics generated by \eqref{eq:diagramFloquet} are time-reversal invariant. Here we consider both the time-reversal-invariant and the non-time-reversal-invariant cases. We remark that in \eqref{eq:Floquetgates1}--\eqref{eq:Floquetgates2} we assumed the 2-site gates $U,W$ to be the same for all $x$. In physical terms this means that we consider interactions that are \emph{homogeneous} in space, while we allow for some position-dependent `external fields' (encoded in the single-site gates $u_{x}, w_x$). The extension of our results to fully inhomogeneous systems is discussed in Sec.~\ref{Sec:inhom}. \subsection{Spectral form factor} \label{sec:SFF} The objective of this paper is the study of spectral statistics of the Floquet operator. Namely, we consider the distribution of the elements of the spectrum of the unitary Floquet matrix \begin{equation} {\rm spect}[\mathbb U_{L}]= \{e^{i \varphi_j};\,j=1,2\ldots,{\cal N}\}\,, \end{equation} where $\varphi_j$ --- conventionally referred to as \emph{quasienergies} --- can be taken to be in $[0,2\pi)$. Considering ${\rm spect}[\mathbb U_{L}]$ as a one-dimensional gas on the circle ${\cal S}^1$, we analyse its 2-point correlation functions, specifically, the \emph{spectral form factor} (SFF) defined as \begin{equation} K(t,L) := \mathbb E\!\left[\,|{\rm tr}\, \mathbb U_{L}^t|^2\,\right] = \mathbb E\left[\sum_{j,j'=1}^{\mathcal N} e^{i (\varphi_j-\varphi_{j'}) t}\right], \qquad t,L\in\mathbb N\,. \label{eq:SFF} \end{equation} Here $\mathbb E[\cdot]$ is an average over an ensemble of similar systems. The average is necessary to smear out the fluctuations of $|{\rm tr}\, \mathbb U_{L}^t|^2$, which do not die out even in the limit of large $L$, and extract the universal behaviour. We shall see later that very mild forms of averaging are sufficient (we remind the reader that the results are most interesting in the limit of clean systems), specifically we will consider cases where $u_x$ and $w_x$ are {\em i.i.d. } for $x\in\Lambda_L$ densely covering an arbitrary small ball around the identity in ${\rm SU}(d)$. The SFF is directly connected to the Fourier transform of the quasi-energy 2-point function. Indeed, introducing the $n$-point function as \begin{equation} \rho_{n}(\vartheta_1,\ldots,\vartheta_n) := \mathbb E\left[ \sum_{j_1\neq\ldots \neq j_n=1}^{\mathcal N} \prod_{k=1}^n \!\delta(\vartheta_k-\varphi_{j_k})\!\right], \end{equation} the SFF can be expressed as \begin{equation} K(t,L) = \int\limits_{[0,2\pi]^{2}} \!\!{\rm d} \vartheta_1 {\rm d} \vartheta_2\,\, e^{i (\vartheta_1- \vartheta_2) t} \rho_{2}(\vartheta_1, \vartheta_2)+ \mathcal N. \label{eq:SFF2pointf} \end{equation} Since this quantity measures correlations between quasienergy levels at arbitrary distance, it is very convenient to analyse extended systems for large volume $L$ where neighbouring quasienergy levels become exponentially close in $L$ and one has to look at correlations on larger scales. \subsection{Spectral form factor for random unitary matrices} \label{sec:RMT} Before moving to the analysis of \eqref{eq:SFF} for local quantum circuits let us briefly recall our point of reference: the SFF of random unitary matrices. In this case the average $\mathbb E\left[\cdot\right]$ in \eqref{eq:SFF} is replaced by the integration over an ensemble of random unitary matrices of dimension ${\cal N}$, i.e. \begin{equation} K_{\rm ens}(t,{\cal N}) := \int |{\rm tr}\, \mathbb U^t|^2 {\rm d}\mu_{\rm ens}(\mathbb U), \end{equation} and the result depends on the precise form of the measure ${\rm d}\mu_{\rm ens}(\mathbb U)$ of the ensemble considered. Specifically, in this paper we are interested in the two most common cases: (i) systems without anti-unitary symmetries, and (ii) systems with regular time-reversal symmetry (squaring to the identity). In these two cases the relevant ensembles of random matrices are two of Dyson's circular ensembles: the {\em Circular Unitary Ensemble} (CUE) and the {\em Circular Orthogonal Ensemble} (COE). The CUE measure is the invariant Haar measure over ${\rm U}(\mathcal N)$, while the COE measure is defined for \emph{symmetric unitary matrices} and is uniquely specified by the property of being invariant under orthogonal transformations~\cite{Mehtabook}. In these two cases the result reads as~\cite{Haakebook} \begin{align} K_{\rm CUE}(t,{\cal N}) &= {\rm min}(t, \mathcal N)\,,\\ K_{\rm COE}(t,{\cal N}) &= 2{\rm min}(t, \mathcal N)\left(1-\sum_{m=1}^{{\rm min}(t, \mathcal N)} \frac{1}{2m+2{\rm max}(t,\mathcal N)-\mathcal N-1}\right)\!. \end{align} In particular, in the thermodynamic limit $L\to\infty$ (${\cal N}\to\infty$) they simplify to \begin{equation} \lim_{{\cal N}\to\infty} K_{\rm CUE}(t,{\cal N}) = t,\qquad\qquad \lim_{{\cal N}\to\infty} K_{\rm COE}(t,{\cal N}) = 2 t\,. \label{eq:SFFRMTTL} \end{equation} The main result of our paper is the proof that one recovers the r.h.s.\ of \eqref{eq:SFFRMTTL} by computing exactly the expression \eqref{eq:SFF} for a broad class of Floquet quantum circuits. \subsection{Spectral form factor of Floquet quantum circuits} For local quantum circuits the SFF \eqref{eq:SFF} can be represented diagrammatically as follows \begin{equation} K(t,L) = \mathbb E\left[ {\rm tr}(\mathbb U_{L})^t {\rm tr}(\mathbb U_{L}^\dag)^t\right]= \mathbb E\Biggl[\begin{tikzpicture}[baseline=(current bounding box.center), scale=0.49] \foreach \i in {1,...,5}{ \draw[thick, dotted] (2*\i+2-12.5+0.255,-1.75-0.1) -- (2*\i+2-12.5+0.255,4.25-0.1); \draw[thick, dotted] (2*\i+2-11.5-0.255,-1.75-0.1) -- (2*\i+2-11.5-0.255,4.25-0.1);} \foreach \i in {1,...,5}{ \draw[thick] (2*\i+2-11.5,4) arc (-45:175:0.15); \draw[thick] (2*\i+2-11.5,-2) arc (315:180:0.15); \draw[thick] (2*\i+2-0.5-12,-2) arc (-135:0:0.15); } \foreach \i in {2,...,6}{ \draw[thick] (2*\i+2-2.5-12,4) arc (225:0:0.15); } \foreach \i in {0,1,2}{ \draw[thick, dotted] (-9.5,2*\i-1.745) -- (0.4,2*\i-1.745); \draw[thick, dotted] (-9.5,2*\i-1.255) -- (0.4,2*\i-1.255); } \foreach \i in{1.5,2.5,3.5}{ \draw[thick] (0.5,2*\i-0.5-3.5) arc (45:-90:0.15); \draw[thick] (-10+0.5+0,2*\i-0.5-3.5) arc (45:270:0.15); } \foreach \i in{0.5,1.5,2.5} { \draw[ thick] (0.5,1+2*\i-0.5-3.5) arc (-45:90:0.15); \draw[ thick] (-10+0.5,1+2*\i-0.5-3.5) arc (315:90:0.15); } \foreach \jj[evaluate=\jj as \j using -2*(ceil(\jj/2)-\jj/2)] in {-1,-3,-5}{ \foreach \i in {1,...,5} { \draw[thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=myorange, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[thick] (-2*\i+2,-1.35-\jj) -- (-2*\i+2.15,-1.35-\jj) -- (-2*\i+2.15,-1.5-\jj);% } } \foreach \jj[evaluate=\jj as \j using -2*(ceil(\jj/2)-\jj/2)] in {-4,-2,0}{ \foreach \i in {1,...,5} { \draw[thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=myred, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[thick] (-2*\i+1,-1.35-\jj) -- (-2*\i+1.15,-1.35-\jj) -- (-2*\i+1.15,-1.5-\jj);% } } \foreach \jj in {0,2,4}{ \draw[ thick, fill=myblue1, rounded corners=2pt] (0.5,-1+\jj) circle (.15); \draw[ thick, fill=myblue2, rounded corners=2pt] (-0.5,-1+\jj) circle (.15); \draw[ thick, fill=myblue3, rounded corners=2pt] (-1.5,-1+\jj) circle (.15); \draw[ thick, fill=myblue4, rounded corners=2pt] (-2.5,-1+\jj) circle (.15); \draw[ thick, fill=myblue5, rounded corners=2pt] (-3.5,-1+\jj) circle (.15); \draw[ thick, fill=myblue6, rounded corners=2pt] (-4.5,-1+\jj) circle (.15); \draw[ thick, fill=myblue7, rounded corners=2pt] (-5.5,-1+\jj) circle (.15); \draw[ thick, fill=myblue8, rounded corners=2pt] (-6.5,-1+\jj) circle (.15); \draw[ thick, fill=myblue9, rounded corners=2pt] (-7.5,-1+\jj) circle (.15); \draw[ thick, fill=myblue10, rounded corners=2pt] (-8.5,-1+\jj) circle (.15); \draw[ thick, fill=myblue8, rounded corners=2pt] (0.5,\jj) circle (.15); \draw[ thick, fill=myblue9, rounded corners=2pt] (-0.5,\jj) circle (.15); \draw[ thick, fill=myblue1, rounded corners=2pt] (-1.5,\jj) circle (.15); \draw[ thick, fill=myblue3, rounded corners=2pt] (-2.5,\jj) circle (.15); \draw[ thick, fill=myblue2, rounded corners=2pt] (-3.5,\jj) circle (.15); \draw[ thick, fill=myblue4, rounded corners=2pt] (-4.5,\jj) circle (.15); \draw[ thick, fill=myblue10, rounded corners=2pt] (-5.5,\jj) circle (.15); \draw[ thick, fill=myblue9, rounded corners=2pt] (-6.5,\jj) circle (.15); \draw[ thick, fill=myblue8, rounded corners=2pt] (-7.5,\jj) circle (.15); \draw[ thick, fill=myblue7, rounded corners=2pt] (-8.5,\jj) circle (.15); } \foreach \jj in {0,-2,-4}{ \foreach \i in {1,...,5}{ \draw[thick] (-2*\i+1.4,-.95-\jj) -- (-2*\i+1.55,-.95-\jj) -- (-2*\i+1.55,-1.1-\jj);} \foreach \i in {1,...,5}{ \draw[thick] (-2*\i+2.4,0.05-\jj) -- (-2*\i+2.55,0.05-\jj) -- (-2*\i+2.55,-0.1-\jj);} \foreach \i in {1,...,5}{ \draw[thick] (-2*\i+1.45,-0.1-\jj) -- (-2*\i+1.45,0.05-\jj) -- (-2*\i+1.6,0.05-\jj);} \foreach \i in {0,...,4}{ \draw[thick] (-2*\i+.45,-1.1-\jj) -- (-2*\i+.45,-0.95-\jj) -- (-2*\i+.6,-0.95-\jj);} } \def0{0} \def-9{-9} \foreach \i in {1,...,5}{ \draw[ thick, dotted] (2*\i+2-1.485+0.25+0-11,-2.5-0.1+-9+2.5) -- (2*\i+2-1.485+0.25+0-11,3.5-0.1+-9+2.5); \draw[ thick, dotted] (2*\i+2-0.525-0.25+0-11,-2.5-0.1+-9+2.5) -- (2*\i+2-0.525-0.25+0-11,3.5-0.1+-9+2.5); } \foreach \i in {0,1,2}{ \draw[ thick, dotted] (1.75+0-11,2*\i-1.25+-9+2.5) -- (11.5+0-11,2*\i-1.25+-9+2.5); \draw[ thick, dotted] (1.5+0-11,2*\i-.76+-9+2.5) -- (11.5+0-11,2*\i-.76+-9+2.5); } \foreach \i in {1,...,5} { \draw[ thick] (2*\i+2-1.5-11+0,3.5+-9+2.5) arc (135:-0:0.15); \draw[ thick] (2*\i+2-.5-11+0,3.5+-9+2.5) arc (-325:-180:0.15); \draw[ thick] (2*\i+2-1.5-11+0,-2.5+-9+2.5) arc (-45:180:-0.15); \draw[ thick] (2*\i+2-0.5-11+0,-2.5+-9+2.5) arc (45:-180:0.15); } \foreach \i in {3,...,5} { \draw[ thick] (0+.5,2*\i-0.5-3.5+-9) arc (45:-90:0.15); \draw[ thick] (0-10+0.5+0,2*\i-0.5-3.5+-9) arc (45:280:0.15); } \foreach \i in{2,...,4} { \draw[ thick] (0+.5,1+2*\i-0.5-3.5+-9) arc (-45:90:0.15); \draw[ thick] (0-10+0.5,1+2*\i-0.5-3.5+-9) arc (-45:-280:0.15); } \foreach \i in {1,...,5} { \draw[ thick] (0+.5-2*\i,6+-9) -- (0+1-2*\i,5.5+-9); \draw[ thick] (0+1.5-2*\i,6+-9) -- (0+1-2*\i,5.5+-9); } \foreach \jj[evaluate=\jj as \j using -2*(ceil(\jj/2)-\jj/2)] in {0,...,3} \foreach \i in {1,...,5} { \draw[ thick] (0+.5-2*\i-1*\j,2+1*\jj+-9) -- (0+1-2*\i-1*\j,1.5+\jj+-9); \draw[ thick] (0+1-2*\i-1*\j,1.5+1*\jj+-9) -- (0+1.5-2*\i-1*\j,2+\jj+-9); } \foreach \i in {0,...,4} { \draw[ thick] (0-.5-2*\i,1+-9) -- (0+0.5-2*\i,0+-9); \draw[ thick] (0-0.5-2*\i,0+-9) -- (0+0.5-2*\i,1+-9); \draw[ thick, fill=mygreen, rounded corners=2pt] (0-0.25-2*\i,0.25+-9) rectangle (0+.25-2*\i,0.75+-9); \draw[thick] (0-2*\i,0.65+-9) -- (0+.15-2*\i,.65+-9) -- (0+.15-2*\i,0.5+-9); } \foreach \jj[evaluate=\jj as \j using -2*(ceil(\jj/2)-\jj/2)] in {-1,1,3} \foreach \i in {1,...,5} { \draw[ thick] (0+.5-2*\i-1*\j,1+1*\jj+-9) -- (0+1-2*\i-1*\j,1.5+\jj+-9); \draw[ thick] (0+1-2*\i-1*\j,1.5+1*\jj+-9) -- (0+1.5-2*\i-1*\j,1+\jj+-9); \draw[ thick, fill=myblue4, rounded corners=2pt] (0+0.75-2*\i-1*\j,1.75+\jj+-9) rectangle (0+1.25-2*\i-1*\j,1.25+\jj+-9); \draw[thick] (0+1-2*\i-1*\j,1.65+1*\jj+-9) -- (0+1.15-2*\i-1*\j,1.65+1*\jj+-9) -- (0+1.15-2*\i-1*\j,1.5+1*\jj+-9); } \foreach \jj[evaluate=\jj as \j using -2*(ceil(\jj/2)-\jj/2)] in {0,2,4}{ \foreach \i in {1,...,5} { \draw[ thick] (0+.5-2*\i-1*\j,1+1*\jj+-9) -- (0+1-2*\i-1*\j,1.5+\jj+-9); \draw[ thick] (0+1-2*\i-1*\j,1.5+1*\jj+-9) -- (0+1.5-2*\i-1*\j,1+\jj+-9); \draw[ thick, fill=myblue, rounded corners=2pt] (0+0.75-2*\i-1*\j,1.75+\jj+-9) rectangle (0+1.25-2*\i-1*\j,1.25+\jj+-9); \draw[thick] (0+1-2*\i-1*\j,1.65+1*\jj+-9) -- (0+1.15-2*\i-1*\j,1.65+1*\jj+-9) -- (0+1.15-2*\i-1*\j,1.5+1*\jj+-9); }} \foreach \jj in {0.5,2.5,4.5}{ \draw[ thick, fill=mygray1, rounded corners=2pt] (.5+0,-.5+\jj+-9) circle (.15); \draw[ thick, fill=mygray4, rounded corners=2pt] (-.5+0,-.5+\jj+-9) circle (.15); \draw[ thick, fill=mygray3, rounded corners=2pt] (-1.5+0,-.5+\jj+-9) circle (.15); \draw[ thick, fill=mygray2, rounded corners=2pt] (-2.5+0,-.5+\jj+-9) circle (.15); \draw[ thick, fill=mygray1, rounded corners=2pt] (-3.5+0,-.5+\jj+-9) circle (.15); \draw[ thick, fill=mygray5, rounded corners=2pt] (-4.5+0,-.5+\jj+-9) circle (.15); \draw[ thick, fill=mygray4, rounded corners=2pt] (-5.5+0,-.5+\jj+-9) circle (.15); \draw[ thick, fill=mygray8, rounded corners=2pt] (-6.5+0,-.5+\jj+-9) circle (.15); \draw[ thick, fill=mygray7, rounded corners=2pt] (-7.5+0,-.5+\jj+-9) circle (.15); \draw[ thick, fill=mygray6, rounded corners=2pt] (-8.5+0,-.5+\jj+-9) circle (.15); } \foreach \jj in {1.5,3.5,5.5} { \draw[ thick, fill=mygray3, rounded corners=2pt] (.5+0,-.5+\jj+-9) circle (.15); \draw[ thick, fill=mygray6, rounded corners=2pt] (-.5+0,-.5+\jj+-9) circle (.15); \draw[ thick, fill=mygray1, rounded corners=2pt] (-1.5+0,-.5+\jj+-9) circle (.15); \draw[ thick, fill=mygray2, rounded corners=2pt] (-2.5+0,-.5+\jj+-9) circle (.15); \draw[ thick, fill=mygray3, rounded corners=2pt] (-3.5+0,-.5+\jj+-9) circle (.15); \draw[ thick, fill=mygray4, rounded corners=2pt] (-4.5+0,-.5+\jj+-9) circle (.15); \draw[ thick, fill=mygray5, rounded corners=2pt] (-5.5+0,-.5+\jj+-9) circle (.15); \draw[ thick, fill=mygray6, rounded corners=2pt] (-6.5+0,-.5+\jj+-9) circle (.15); \draw[ thick, fill=mygray7, rounded corners=2pt] (-7.5+0,-.5+\jj+-9) circle (.15); \draw[ thick, fill=mygray8, rounded corners=2pt] (-8.5+0,-.5+\jj+-9) circle (.15); } \foreach \jj in {0,-2,-4}{ \foreach \i in {1,...,5}{ \draw[thick] (-2*\i+1.4+0,-.95-\jj-0.5+-9+1.5) -- (-2*\i+1.55+0,-.95-\jj-0.5+-9+1.5) -- (-2*\i+1.55+0,-1.1-\jj-0.5+-9+1.5);} \foreach \i in {1,...,5}{ \draw[thick] (-2*\i+2.4+0,0.05-\jj-0.5+-9+1.5) -- (-2*\i+2.55+0,0.05-\jj-0.5+-9+1.5) -- (-2*\i+2.55+0,-0.1-\jj-0.5+-9+1.5);} \foreach \i in {1,...,5}{ \draw[thick] (-2*\i+1.45+0,-0.1-\jj-0.5+-9+1.5) -- (-2*\i+1.45+0,0.05-\jj-0.5+-9+1.5) -- (-2*\i+1.6+0,0.05-\jj-0.5+-9+1.5);} \foreach \i in {0,...,4}{ \draw[thick] (-2*\i+.45+0,-1.1-\jj-0.5+-9+1.5) -- (-2*\i+.45+0,-0.95-\jj-0.5+-9+1.5) -- (-2*\i+.6+0,-0.95-\jj-0.5+-9+1.5);} } \end{tikzpicture}\Biggr], \end{equation} where we represented the trace in the {\em forward time sheet} (${\rm tr}\, \mathbb U_{L}^t$) using the diagram \eqref{eq:diagramFloquet} and that in the {\em backward time sheet} (${\rm tr}\, (\mathbb U_{L}^\dag)^t$) by introducing \begin{align} U^\dag & = \begin{tikzpicture}[baseline=(current bounding box.center), scale=.7] \draw[ thick] (-4.25,0.5) -- (-3.25,-0.5); \draw[ thick] (-4.25,-0.5) -- (-3.25,0.5); \draw[ thick, fill=myblue, rounded corners=2pt] (-4,0.25) rectangle (-3.5,-0.25); \draw[thick] (-3.75,0.15) -- (-3.75+0.15,0.15) -- (-3.75+0.15,0); \Text[x=-4.25,y=-0.75]{} \end{tikzpicture}\,, & W^\dag &=\begin{tikzpicture}[baseline=(current bounding box.center), scale=.7] \draw[ thick] (-4.25,0.5) -- (-3.25,-0.5); \draw[ thick] (-4.25,-0.5) -- (-3.25,0.5); \draw[ thick, fill=myblue4, rounded corners=2pt] (-4,0.25) rectangle (-3.5,-0.25); \draw[thick] (-3.75,0.15) -- (-3.75+0.15,0.15) -- (-3.75+0.15,0); \Text[x=-4.25,y=-0.75]{} \end{tikzpicture}\,, & u_{x}^\dag, w_{x}^\dag= \begin{tikzpicture}[baseline=(current bounding box.center), scale=.7] \draw[ thick] (-4.25,0.5) -- (-4.25,-0.5); \draw[ thick, fill=mygray4, rounded corners=2pt] (-4.25,0) circle (.15); \draw[thick, rotate around = {-45:(0.525-4.77,0.375-0.4)}] (.45-4.77,0.3-0.4) -- (.45-4.77,0.45-0.4) -- (.6-4.77,0.45-0.4); \Text[x=-4.25,y=-0.75]{} \end{tikzpicture}\,. \end{align} Once again shades of the same colour denote different matrices. Note that top and bottom lines at the same positions within both sheets are connected because of the traces. Folding the backward sheet (blue) underneath the forward one (red) we write the folded circuit representation of the SFF \begin{equation} K(t,L)= \mathbb{E}\Biggl[ \begin{tikzpicture}[baseline=(current bounding box.center), scale=0.55] \foreach \i in {1,...,5}{ \draw[very thick, dotted] (2*\i+2-12.5+0.255,-1.75-0.1) -- (2*\i+2-12.5+0.255,4.25-0.1); \draw[very thick, dotted] (2*\i+2-11.5-0.255,-1.75-0.1) -- (2*\i+2-11.5-0.255,4.25-0.1);} \foreach \i in {1,...,5}{ \draw[very thick] (2*\i+2-11.5,4) arc (-45:175:0.15); \draw[very thick] (2*\i+2-11.5,-2) arc (315:180:0.15); \draw[very thick] (2*\i+2-0.5-12,-2) arc (-135:0:0.15); } \foreach \i in {2,...,6}{ \draw[very thick] (2*\i+2-2.5-12,4) arc (225:0:0.15); } \foreach \i in {0,1,2}{ \draw[very thick, dotted] (-9.5,2*\i-1.745) -- (0.4,2*\i-1.745); \draw[very thick, dotted] (-9.5,2*\i-1.255) -- (0.4,2*\i-1.255); } \foreach \i in{1.5,2.5,3.5}{ \draw[very thick] (0.5,2*\i-0.5-3.5) arc (45:-90:0.15); \draw[very thick] (-10+0.5+0,2*\i-0.5-3.5) arc (45:270:0.15); } \foreach \i in{0.5,1.5,2.5} { \draw[very thick] (0.5,1+2*\i-0.5-3.5) arc (-45:90:0.15); \draw[very thick] (-10+0.5,1+2*\i-0.5-3.5) arc (315:90:0.15); } \foreach \jj[evaluate=\jj as \j using -2*(ceil(\jj/2)-\jj/2)] in {-1,-3,-5}{ \foreach \i in {1,...,5} { \draw[very thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[very thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[very thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[very thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=OliveGreen, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[ thick] (-2*\i+2,-1.35-\jj) -- (-2*\i+2.15,-1.35-\jj) -- (-2*\i+2.15,-1.5-\jj);% } } \foreach \jj[evaluate=\jj as \j using -2*(ceil(\jj/2)-\jj/2)] in {-4,-2,0}{ \foreach \i in {1,...,5} { \draw[very thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[very thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[very thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[very thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=mygreen, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[ thick] (-2*\i+1,-1.35-\jj) -- (-2*\i+1.15,-1.35-\jj) -- (-2*\i+1.15,-1.5-\jj);% } } \foreach \jj in {0,2,4}{ \draw[ thick, fill=myyellow1, rounded corners=2pt] (0.5,-1+\jj) circle (.15); \draw[ thick, fill=myyellow2, rounded corners=2pt] (-0.5,-1+\jj) circle (.15); \draw[ thick, fill=myyellow3, rounded corners=2pt] (-1.5,-1+\jj) circle (.15); \draw[ thick, fill=myyellow4, rounded corners=2pt] (-2.5,-1+\jj) circle (.15); \draw[ thick, fill=myyellow5, rounded corners=2pt] (-3.5,-1+\jj) circle (.15); \draw[ thick, fill=myyellow6, rounded corners=2pt] (-4.5,-1+\jj) circle (.15); \draw[ thick, fill=myyellow7, rounded corners=2pt] (-5.5,-1+\jj) circle (.15); \draw[ thick, fill=myyellow8, rounded corners=2pt] (-6.5,-1+\jj) circle (.15); \draw[ thick, fill=myyellow9, rounded corners=2pt] (-7.5,-1+\jj) circle (.15); \draw[ thick, fill=myyellow10, rounded corners=2pt] (-8.5,-1+\jj) circle (.15); \draw[ thick, fill=myyellow8, rounded corners=2pt] (0.5,\jj) circle (.15); \draw[ thick, fill=myyellow9, rounded corners=2pt] (-0.5,\jj) circle (.15); \draw[ thick, fill=myyellow1, rounded corners=2pt] (-1.5,\jj) circle (.15); \draw[ thick, fill=myyellow3, rounded corners=2pt] (-2.5,\jj) circle (.15); \draw[ thick, fill=myyellow2, rounded corners=2pt] (-3.5,\jj) circle (.15); \draw[ thick, fill=myyellow4, rounded corners=2pt] (-4.5,\jj) circle (.15); \draw[ thick, fill=myyellow10, rounded corners=2pt] (-5.5,\jj) circle (.15); \draw[ thick, fill=myyellow9, rounded corners=2pt] (-6.5,\jj) circle (.15); \draw[ thick, fill=myyellow8, rounded corners=2pt] (-7.5,\jj) circle (.15); \draw[ thick, fill=myyellow7, rounded corners=2pt] (-8.5,\jj) circle (.15); } \foreach \jj in {0,-2,-4}{ \foreach \i in {1,...,5}{ \draw[ thick] (-2*\i+1.4,-.95-\jj) -- (-2*\i+1.55,-.95-\jj) -- (-2*\i+1.55,-1.1-\jj);} \foreach \i in {1,...,5}{ \draw[ thick] (-2*\i+2.4,0.05-\jj) -- (-2*\i+2.55,0.05-\jj) -- (-2*\i+2.55,-0.1-\jj);} \foreach \i in {1,...,5}{ \draw[ thick] (-2*\i+1.45,-0.1-\jj) -- (-2*\i+1.45,0.05-\jj) -- (-2*\i+1.6,0.05-\jj);} \foreach \i in {0,...,4}{ \draw[ thick] (-2*\i+.45,-1.1-\jj) -- (-2*\i+.45,-0.95-\jj) -- (-2*\i+.6,-0.95-\jj);} } \end{tikzpicture}\Biggr]\,, \label{eq:SFFfolded} \end{equation} where we introduced ``doubled'' or thickened wires \begin{equation} \label{eq:thickwire} \begin{tikzpicture}[baseline=(current bounding box.center), scale=0.8] \def\eps{0.5}; \draw[very thick] (-3.25,0.5) -- (-3.25,-0.5); \draw[ thick] (-2.0,0.5) -- (-2.0,-0.5); \draw[ thick] (-1.75,0.7) -- (-1.75,-0.3); \Text[x=-2.65,y=0.05]{$=$} \end{tikzpicture}\,, \end{equation} and ``doubled'' gates \begin{eqnarray} \label{eq:doublegate} &\begin{tikzpicture}[baseline=(current bounding box.center), scale=.7] \def\eps{0.5}; \Wgategreen{-3.75}{0}; \Text[x=-3.75,y=-0.6]{} \end{tikzpicture} = \begin{tikzpicture}[baseline=(current bounding box.center), scale=.7] \draw[thick] (-1.65,0.65) -- (-0.65,-0.35); \draw[thick] (-1.65,-0.35) -- (-0.65,0.65); \draw[ thick, fill=myblue, rounded corners=2pt] (-1.4,0.4) rectangle (-.9,-0.1); \draw[thick] (-1.15,0) -- (-1,0) -- (-1,0.15); \draw[thick] (-2.25,0.5) -- (-1.25,-0.5); \draw[thick] (-2.25,-0.5) -- (-1.25,0.5); \draw[ thick, fill=myred, rounded corners=2pt] (-2,0.25) rectangle (-1.5,-0.25); \draw[thick] (-1.75,0.15) -- (-1.6,0.15) -- (-1.6,0); \Text[x=-2.25,y=-0.6]{} \end{tikzpicture} = U\otimes U^{*}\,, \qquad \begin{tikzpicture}[baseline=(current bounding box.center), scale=.7] \def\eps{0.5}; \Wgateolivegreen{-3.75}{0}; \Text[x=-3.75,y=-0.6]{} \end{tikzpicture} = \begin{tikzpicture}[baseline=(current bounding box.center), scale=.7] \draw[thick] (-1.65,0.65) -- (-0.65,-0.35); \draw[thick] (-1.65,-0.35) -- (-0.65,0.65); \draw[ thick, fill=myblue4, rounded corners=2pt] (-1.4,0.4) rectangle (-.9,-0.1); \draw[thick] (-1.15,0) -- (-1,0) -- (-1,0.15); \draw[thick] (-2.25,0.5) -- (-1.25,-0.5); \draw[thick] (-2.25,-0.5) -- (-1.25,0.5); \draw[ thick, fill=myorange, rounded corners=2pt] (-2,0.25) rectangle (-1.5,-0.25); \draw[thick] (-1.75,0.15) -- (-1.6,0.15) -- (-1.6,0); \Text[x=-2.25,y=-0.6]{} \end{tikzpicture} = W\otimes W^{*}\,, \\ \nonumber & \begin{tikzpicture}[baseline=(current bounding box.center), scale=.7] \draw[very thick] (-4.25,0.5) -- (-4.25,-0.5); \draw[ thick, fill=myYO, rounded corners=2pt] (-4.25,0) circle (.15); \draw[thick, rotate around = {-45:(0.525-4.77,0.375-0.4)}] (.45-4.77,0.3-0.4) -- (.45-4.77,0.45-0.4) -- (.6-4.77,0.45-0.4); \Text[x=-4.25,y=-0.75]{} \end{tikzpicture} = \begin{tikzpicture}[baseline=(current bounding box.center), scale=.7] \draw[ thick] (-4,0.5) -- (-4,-0.5); \draw[ thick, fill=mygray4, rounded corners=2pt] (-4,0) circle (.15); \draw[thick, rotate around = {135:(0.525-4.27-0.25,0.375-0.35)}] (.45-4.27-.25,0.3-0.35) -- (.45-4.27-.25,0.45-0.35) -- (.6-4.27-.25,0.45-0.35); \draw[ thick] (-4.25,0.5) -- (-4.25,-0.5); \draw[ thick, fill=myblue10, rounded corners=2pt] (-4.25,0) circle (.15); \draw[thick, rotate around = {-45:(0.525-4.77,0.375-0.4)}] (.45-4.77,0.3-0.4) -- (.45-4.77,0.45-0.4) -- (.6-4.77,0.45-0.4); \Text[x=-4.25,y=-0.75]{} \end{tikzpicture} = u_{x}\otimes u_{x}^{*},\,\, w_{x}\otimes w_{x}^{*}\,. \end{eqnarray} Here and in the following $(\cdot)^*$ denotes complex conjugation in the canonical basis \eqref{eq:realbasis}. \subsection{Local disorder averaging} \label{sec:Ave} As mentioned in Sec.~\ref{sec:SFF} the definition of SFF requires an average. Since our interest is mainly on clean systems, we consider averages over \emph{onsite disorder} that can be made arbitrary weak. This kind of disorder is arguably the most harmless form of disorder that one can introduce in the system because it does not couple different sites. In particular, we focus on the following generic model of on-site disorder where the local gates \eqref{eq:Floquetgates1}--\eqref{eq:Floquetgates2} are specified by fixed unitary interactions $U, W \in {\rm U}(d^2)$, and site-dependent local gates $u_{x}, w_{x} \in {\rm SU}(d)$ of the general form \begin{equation} u_{x} = e^{i \boldsymbol{\theta}_{0,x}\cdot\boldsymbol{\sigma}}, \qquad w_{x} = e^{i \boldsymbol{\theta}_{1,x}\cdot\boldsymbol{\sigma}^T}, \qquad x\in\Lambda_L,\;\;\boldsymbol{\theta}_{\iota,x}\in\mathbb R^{d^2-1}\,. \label{eq:urx} \end{equation} The vector $\boldsymbol{\sigma}=(\sigma_1,\sigma_2,\ldots, \sigma_{d^2-1})$ is formed by Generalised Gell-Mann matrices $\sigma_a$~\cite{GenGellMann} (Pauli matrices for $d=2$, Gell-Mann matrices for $d=3$, etc.), the Hermitian generators of $\mathfrak{su}(d)$, and $\boldsymbol{\sigma}^T=(\sigma_1^T,\sigma_2^T,\ldots, \sigma_{d^2-1}^T)$ is the vector of the corresponding transposed generators. The expectation can be explicitly written in terms of a factorised measure as: \begin{equation} \mathbb E[f] = \int f(\boldsymbol\theta) \prod_{x=0}^{L-1}\prod_{\iota,\iota'=0}^1 g_{\iota\iota'}(\boldsymbol{\theta}_{\iota,x+\frac{\iota'}{2}}) {\rm d}^{d^2-1}\boldsymbol{\theta}_{\iota, x+\frac{\iota'}{2}} \,, \quad \boldsymbol \theta \equiv (\boldsymbol{\theta}_{\iota,x})^{\iota=0,1}_{x\in\Lambda_L}\,. \label{eq:measure} \end{equation} where $g_{\iota\iota'}\in L^1[\mathbb R^{d^2-1}]$ are arbitrary probability densities of i.i.d. random variables $\boldsymbol{\theta}_{\iota,x}$. Note that distributions on integer ($\iota'=0$) and half-odd-integer ($\iota'=1$) sublattices are generally different. \subsection{Space-time duality} \label{sec:duality} The key property of $\mathbb E[\cdot]$ (\ref{eq:measure}) is the factorization with respect to a spatial coordinate $x$. This means that, even though the diagram \eqref{eq:SFFfolded} cannot be thought of as the trace of the product of $t$ transfer matrices in the time direction (because the average couples different time layers), it can be thought of as the trace of the product of $L$ transfer matrices in the space direction. Specifically, \begin{equation} K(t,L)=\begin{tikzpicture}[baseline=(current bounding box.center), scale=0.5] \Text[x=-10,y=4.15]{} \foreach \i in{1.5,2.5,3.5}{ \draw[very thick] (-10+0.5+0,2*\i-0.5-3.5) arc (45:270:0.15); } \foreach \i in{0.5,1.5,2.5} { \draw[very thick] (-10+0.5,1+2*\i-0.5-3.5) arc (315:90:0.15); } \end{tikzpicture}\,\mathbb{E}\Bigl[\begin{tikzpicture}[baseline=(current bounding box.center), scale=0.5] \foreach \i in {5}{ \draw[very thick, dotted] (2*\i+2-12.5+0.255,-1.75-0.1) -- (2*\i+2-12.5+0.255,4.25-0.1); \draw[very thick, dotted] (2*\i+2-11.5-0.255,-1.75-0.1) -- (2*\i+2-11.5-0.255,4.25-0.1);} \foreach \i in {5}{ \draw[very thick] (2*\i+2-11.5,4) arc (-45:175:0.15); \draw[very thick] (2*\i+2-11.5,-2) arc (315:180:0.15); \draw[very thick] (2*\i+2-0.5-12,-2) arc (-135:0:0.15); } \foreach \i in {6}{ \draw[very thick] (2*\i+2-2.5-12,4) arc (225:0:0.15); } \foreach \jj[evaluate=\jj as \j using -2*(ceil(\jj/2)-\jj/2)] in {-1,-3,-5}{ \foreach \i in {1} { \draw[very thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[very thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[very thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[very thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=OliveGreen, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[ thick] (-2*\i+2,-1.35-\jj) -- (-2*\i+2.15,-1.35-\jj) -- (-2*\i+2.15,-1.5-\jj);% } } \foreach \jj[evaluate=\jj as \j using -2*(ceil(\jj/2)-\jj/2)] in {-4,-2,0}{ \foreach \i in {1} { \draw[very thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[very thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[very thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[very thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=mygreen, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[ thick] (-2*\i+1,-1.35-\jj) -- (-2*\i+1.15,-1.35-\jj) -- (-2*\i+1.15,-1.5-\jj);% } } \foreach \jj in {0,2,4}{ \draw[ thick, fill=myyellow1, rounded corners=2pt] (0.5,-1+\jj) circle (.15); \draw[ thick, fill=myyellow2, rounded corners=2pt] (-0.5,-1+\jj) circle (.15); \draw[ thick, fill=myyellow8, rounded corners=2pt] (0.5,\jj) circle (.15); \draw[ thick, fill=myyellow9, rounded corners=2pt] (-0.5,\jj) circle (.15); } \foreach \jj in {0,-2,-4}{ \foreach \i in {1}{ \draw[ thick] (-2*\i+1.4,-.95-\jj) -- (-2*\i+1.55,-.95-\jj) -- (-2*\i+1.55,-1.1-\jj);} \foreach \i in {1}{ \draw[ thick] (-2*\i+2.4,0.05-\jj) -- (-2*\i+2.55,0.05-\jj) -- (-2*\i+2.55,-0.1-\jj);} \foreach \i in {1}{ \draw[ thick] (-2*\i+1.45,-0.1-\jj) -- (-2*\i+1.45,0.05-\jj) -- (-2*\i+1.6,0.05-\jj);} \foreach \i in {0}{ \draw[ thick] (-2*\i+.45,-1.1-\jj) -- (-2*\i+.45,-0.95-\jj) -- (-2*\i+.6,-0.95-\jj);} } \end{tikzpicture}\Bigr]\mathbb{E}\Bigl[\begin{tikzpicture}[baseline=(current bounding box.center), scale=0.5] \foreach \i in {5}{ \draw[very thick, dotted] (2*\i+2-12.5+0.255,-1.75-0.1) -- (2*\i+2-12.5+0.255,4.25-0.1); \draw[very thick, dotted] (2*\i+2-11.5-0.255,-1.75-0.1) -- (2*\i+2-11.5-0.255,4.25-0.1);} \foreach \i in {5}{ \draw[very thick] (2*\i+2-11.5,4) arc (-45:175:0.15); \draw[very thick] (2*\i+2-11.5,-2) arc (315:180:0.15); \draw[very thick] (2*\i+2-0.5-12,-2) arc (-135:0:0.15); } \foreach \i in {6}{ \draw[very thick] (2*\i+2-2.5-12,4) arc (225:0:0.15); } \foreach \jj[evaluate=\jj as \j using -2*(ceil(\jj/2)-\jj/2)] in {-1,-3,-5}{ \foreach \i in {1} { \draw[very thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[very thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[very thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[very thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=OliveGreen, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[ thick] (-2*\i+2,-1.35-\jj) -- (-2*\i+2.15,-1.35-\jj) -- (-2*\i+2.15,-1.5-\jj);% } } \foreach \jj[evaluate=\jj as \j using -2*(ceil(\jj/2)-\jj/2)] in {-4,-2,0}{ \foreach \i in {1} { \draw[very thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[very thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[very thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[very thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=mygreen, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[ thick] (-2*\i+1,-1.35-\jj) -- (-2*\i+1.15,-1.35-\jj) -- (-2*\i+1.15,-1.5-\jj);% } } \foreach \jj in {0,2,4}{ \draw[ thick, fill=myyellow1, rounded corners=2pt] (0.5,-1+\jj) circle (.15); \draw[ thick, fill=myyellow2, rounded corners=2pt] (-0.5,-1+\jj) circle (.15); \draw[ thick, fill=myyellow8, rounded corners=2pt] (0.5,\jj) circle (.15); \draw[ thick, fill=myyellow9, rounded corners=2pt] (-0.5,\jj) circle (.15); } \foreach \jj in {0,-2,-4}{ \foreach \i in {1}{ \draw[ thick] (-2*\i+1.4,-.95-\jj) -- (-2*\i+1.55,-.95-\jj) -- (-2*\i+1.55,-1.1-\jj);} \foreach \i in {1}{ \draw[ thick] (-2*\i+2.4,0.05-\jj) -- (-2*\i+2.55,0.05-\jj) -- (-2*\i+2.55,-0.1-\jj);} \foreach \i in {1}{ \draw[ thick] (-2*\i+1.45,-0.1-\jj) -- (-2*\i+1.45,0.05-\jj) -- (-2*\i+1.6,0.05-\jj);} \foreach \i in {0}{ \draw[ thick] (-2*\i+.45,-1.1-\jj) -- (-2*\i+.45,-0.95-\jj) -- (-2*\i+.6,-0.95-\jj);} } \end{tikzpicture}\Bigr]\mathbb{E}\Bigl[\begin{tikzpicture}[baseline=(current bounding box.center), scale=0.5] \foreach \i in {5}{ \draw[very thick, dotted] (2*\i+2-12.5+0.255,-1.75-0.1) -- (2*\i+2-12.5+0.255,4.25-0.1); \draw[very thick, dotted] (2*\i+2-11.5-0.255,-1.75-0.1) -- (2*\i+2-11.5-0.255,4.25-0.1);} \foreach \i in {5}{ \draw[very thick] (2*\i+2-11.5,4) arc (-45:175:0.15); \draw[very thick] (2*\i+2-11.5,-2) arc (315:180:0.15); \draw[very thick] (2*\i+2-0.5-12,-2) arc (-135:0:0.15); } \foreach \i in {6}{ \draw[very thick] (2*\i+2-2.5-12,4) arc (225:0:0.15); } \foreach \jj[evaluate=\jj as \j using -2*(ceil(\jj/2)-\jj/2)] in {-1,-3,-5}{ \foreach \i in {1} { \draw[very thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[very thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[very thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[very thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=OliveGreen, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[ thick] (-2*\i+2,-1.35-\jj) -- (-2*\i+2.15,-1.35-\jj) -- (-2*\i+2.15,-1.5-\jj);% } } \foreach \jj[evaluate=\jj as \j using -2*(ceil(\jj/2)-\jj/2)] in {-4,-2,0}{ \foreach \i in {1} { \draw[very thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[very thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[very thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[very thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=mygreen, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[ thick] (-2*\i+1,-1.35-\jj) -- (-2*\i+1.15,-1.35-\jj) -- (-2*\i+1.15,-1.5-\jj);% } } \foreach \jj in {0,2,4}{ \draw[ thick, fill=myyellow1, rounded corners=2pt] (0.5,-1+\jj) circle (.15); \draw[ thick, fill=myyellow2, rounded corners=2pt] (-0.5,-1+\jj) circle (.15); \draw[ thick, fill=myyellow8, rounded corners=2pt] (0.5,\jj) circle (.15); \draw[ thick, fill=myyellow9, rounded corners=2pt] (-0.5,\jj) circle (.15); } \foreach \jj in {0,-2,-4}{ \foreach \i in {1}{ \draw[ thick] (-2*\i+1.4,-.95-\jj) -- (-2*\i+1.55,-.95-\jj) -- (-2*\i+1.55,-1.1-\jj);} \foreach \i in {1}{ \draw[ thick] (-2*\i+2.4,0.05-\jj) -- (-2*\i+2.55,0.05-\jj) -- (-2*\i+2.55,-0.1-\jj);} \foreach \i in {1}{ \draw[ thick] (-2*\i+1.45,-0.1-\jj) -- (-2*\i+1.45,0.05-\jj) -- (-2*\i+1.6,0.05-\jj);} \foreach \i in {0}{ \draw[ thick] (-2*\i+.45,-1.1-\jj) -- (-2*\i+.45,-0.95-\jj) -- (-2*\i+.6,-0.95-\jj);} } \end{tikzpicture}\Bigr]\mathbb{E}\Bigl[\begin{tikzpicture}[baseline=(current bounding box.center), scale=0.5] \foreach \i in {5}{ \draw[very thick, dotted] (2*\i+2-12.5+0.255,-1.75-0.1) -- (2*\i+2-12.5+0.255,4.25-0.1); \draw[very thick, dotted] (2*\i+2-11.5-0.255,-1.75-0.1) -- (2*\i+2-11.5-0.255,4.25-0.1);} \foreach \i in {5}{ \draw[very thick] (2*\i+2-11.5,4) arc (-45:175:0.15); \draw[very thick] (2*\i+2-11.5,-2) arc (315:180:0.15); \draw[very thick] (2*\i+2-0.5-12,-2) arc (-135:0:0.15); } \foreach \i in {6}{ \draw[very thick] (2*\i+2-2.5-12,4) arc (225:0:0.15); } \foreach \jj[evaluate=\jj as \j using -2*(ceil(\jj/2)-\jj/2)] in {-1,-3,-5}{ \foreach \i in {1} { \draw[very thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[very thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[very thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[very thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=OliveGreen, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[ thick] (-2*\i+2,-1.35-\jj) -- (-2*\i+2.15,-1.35-\jj) -- (-2*\i+2.15,-1.5-\jj);% } } \foreach \jj[evaluate=\jj as \j using -2*(ceil(\jj/2)-\jj/2)] in {-4,-2,0}{ \foreach \i in {1} { \draw[very thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[very thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[very thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[very thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=mygreen, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[ thick] (-2*\i+1,-1.35-\jj) -- (-2*\i+1.15,-1.35-\jj) -- (-2*\i+1.15,-1.5-\jj);% } } \foreach \jj in {0,2,4}{ \draw[ thick, fill=myyellow1, rounded corners=2pt] (0.5,-1+\jj) circle (.15); \draw[ thick, fill=myyellow2, rounded corners=2pt] (-0.5,-1+\jj) circle (.15); \draw[ thick, fill=myyellow8, rounded corners=2pt] (0.5,\jj) circle (.15); \draw[ thick, fill=myyellow9, rounded corners=2pt] (-0.5,\jj) circle (.15); } \foreach \jj in {0,-2,-4}{ \foreach \i in {1}{ \draw[ thick] (-2*\i+1.4,-.95-\jj) -- (-2*\i+1.55,-.95-\jj) -- (-2*\i+1.55,-1.1-\jj);} \foreach \i in {1}{ \draw[ thick] (-2*\i+2.4,0.05-\jj) -- (-2*\i+2.55,0.05-\jj) -- (-2*\i+2.55,-0.1-\jj);} \foreach \i in {1}{ \draw[ thick] (-2*\i+1.45,-0.1-\jj) -- (-2*\i+1.45,0.05-\jj) -- (-2*\i+1.6,0.05-\jj);} \foreach \i in {0}{ \draw[ thick] (-2*\i+.45,-1.1-\jj) -- (-2*\i+.45,-0.95-\jj) -- (-2*\i+.6,-0.95-\jj);} } \end{tikzpicture}\Bigr]\mathbb{E}\Bigl[\begin{tikzpicture}[baseline=(current bounding box.center), scale=0.5] \foreach \i in {5}{ \draw[very thick, dotted] (2*\i+2-12.5+0.255,-1.75-0.1) -- (2*\i+2-12.5+0.255,4.25-0.1); \draw[very thick, dotted] (2*\i+2-11.5-0.255,-1.75-0.1) -- (2*\i+2-11.5-0.255,4.25-0.1);} \foreach \i in {5}{ \draw[very thick] (2*\i+2-11.5,4) arc (-45:175:0.15); \draw[very thick] (2*\i+2-11.5,-2) arc (315:180:0.15); \draw[very thick] (2*\i+2-0.5-12,-2) arc (-135:0:0.15); } \foreach \i in {6}{ \draw[very thick] (2*\i+2-2.5-12,4) arc (225:0:0.15); } \foreach \jj[evaluate=\jj as \j using -2*(ceil(\jj/2)-\jj/2)] in {-1,-3,-5}{ \foreach \i in {1} { \draw[very thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[very thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[very thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[very thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=OliveGreen, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[ thick] (-2*\i+2,-1.35-\jj) -- (-2*\i+2.15,-1.35-\jj) -- (-2*\i+2.15,-1.5-\jj);% } } \foreach \jj[evaluate=\jj as \j using -2*(ceil(\jj/2)-\jj/2)] in {-4,-2,0}{ \foreach \i in {1} { \draw[very thick] (.5-2*\i-1*\j,-2-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[very thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-2-\jj); \draw[very thick] (.5-2*\i-1*\j,-1-1*\jj) -- (1-2*\i-1*\j,-1.5-\jj); \draw[very thick] (1-2*\i-1*\j,-1.5-1*\jj) -- (1.5-2*\i-1*\j,-1-\jj); \draw[thick, fill=mygreen, rounded corners=2pt] (0.75-2*\i-1*\j,-1.75-\jj) rectangle (1.25-2*\i-1*\j,-1.25-\jj); \draw[ thick] (-2*\i+1,-1.35-\jj) -- (-2*\i+1.15,-1.35-\jj) -- (-2*\i+1.15,-1.5-\jj);% } } \foreach \jj in {0,2,4}{ \draw[ thick, fill=myyellow1, rounded corners=2pt] (0.5,-1+\jj) circle (.15); \draw[ thick, fill=myyellow2, rounded corners=2pt] (-0.5,-1+\jj) circle (.15); \draw[ thick, fill=myyellow8, rounded corners=2pt] (0.5,\jj) circle (.15); \draw[ thick, fill=myyellow9, rounded corners=2pt] (-0.5,\jj) circle (.15); } \foreach \jj in {0,-2,-4}{ \foreach \i in {1}{ \draw[ thick] (-2*\i+1.4,-.95-\jj) -- (-2*\i+1.55,-.95-\jj) -- (-2*\i+1.55,-1.1-\jj);} \foreach \i in {1}{ \draw[ thick] (-2*\i+2.4,0.05-\jj) -- (-2*\i+2.55,0.05-\jj) -- (-2*\i+2.55,-0.1-\jj);} \foreach \i in {1}{ \draw[ thick] (-2*\i+1.45,-0.1-\jj) -- (-2*\i+1.45,0.05-\jj) -- (-2*\i+1.6,0.05-\jj);} \foreach \i in {0}{ \draw[ thick] (-2*\i+.45,-1.1-\jj) -- (-2*\i+.45,-0.95-\jj) -- (-2*\i+.6,-0.95-\jj);} } \end{tikzpicture}\Bigr] \begin{tikzpicture}[baseline=(current bounding box.center), scale=0.5] \Text[x=0.5,y=4.15]{} \foreach \i in{1.5,2.5,3.5}{ \draw[very thick] (0.5,2*\i-0.5-3.5) arc (45:-90:0.15); } \foreach \i in{0.5,1.5,2.5} { \draw[very thick] (0.5,1+2*\i-0.5-3.5) arc (-45:90:0.15); } \end{tikzpicture}\,. \label{eq:SFFTMspace} \end{equation} In equations this is expressed as \begin{eqnarray} K(t,L) &=& \mathbb E\left[\left({\rm tr}\,\mathbb U_{L}^t\right) \left({\rm tr}\,\mathbb U_{L}^t\right)^*\right] = \mathbb E\left[ {\rm tr}\left(\mathbb U^{\phantom{*}}_{L}\otimes \mathbb U^*_{L}\right)^t\right] \nonumber\\ &=& \mathbb E\left[ {\rm tr}\left(\prod_{x=1}^L \tilde{\mathbb U}_{t}^{\phantom{*}}(x) \otimes \tilde{\mathbb U}_{t}(x)^*\right)\right] = {\rm tr}\left(\mathbb E\left[\tilde{\mathbb U}_{t}^{\phantom{*}} \otimes \tilde{\mathbb U}_{t}^*\right]\right)^L \nonumber\\ &=& {\rm tr}\,\mathbb T^L, \label{eq:SFFduality} \end{eqnarray} where the tensor product operates between the two different time sheets, and we introduced the following definitions: \begin{itemize} \item[(i)] ``Dual'' Floquet operator propagating in the space-direction over the Hilbert space ${\cal H}_{2t}$ of $2t$ qudits, explicitly depending on the position $x\in \mathbb Z_L$: \begin{equation} \tilde{\mathbb U}_{t}(x) := \prod_{\tau \in \mathbb Z_{t}+\frac{1}{2}}\!\!\eta_{\tau,t}(\tilde U\,( u_{x-\frac{1}{2}} \otimes w_{x-\frac{1}{2}}^T)) \prod_{\tau \in \mathbb Z_{t}} \eta_{\tau,t}(\tilde W \, (w_{x} \otimes u_x^T))\,. \end{equation} Here $\tilde{U}$, $\tilde{W} \in {\rm End}(\mathcal H_2)$ are the ``dual" 2-body interaction gates defined via the space-time duality mapping $\,\tilde{} : {\rm End}(\mathcal H_2) \to {\rm End}(\mathcal H_2)$. Specifically, for any $O\in{\rm End}(\mathcal H_2)$ with matrix elements \begin{equation} O_{i_1i_2,j_1j_2} = \bra{i_1}\otimes \bra{i_2} O \ket{j_1}\otimes \ket{j_2}\,, \end{equation} we define \begin{equation} \tilde{O}_{jl,ik} := O_{ij,kl},\quad i,j,k,l\in\{0,1, \ldots, d-1\}\,. \label{eq:duallocalgate} \end{equation} We see that $\tilde{U}_{ij,kl}$ and $\tilde{W}_{ij,kl}$ correspond to a particular reshuffling of the indices of ${U}_{ij,kl}$ and ${W}_{ij,kl}$. \item[(ii)] SFF--transfer matrix: \begin{equation} \mathbb T:= \mathbb E\bigl[\tilde{\mathbb U}_t^{\phantom{*}} \otimes \tilde{\mathbb U}_t^*\bigr]\in {\rm End}({\cal H}_{2t}\otimes{\cal H}_{2t}). \label{eq:SFFTM} \end{equation} Note that $\mathbb T$ \emph{does not} depend on position $x$ due to the identical distribution of $(\boldsymbol{\theta}_{\iota,x-\frac{1}{2}},\boldsymbol{\theta}_{\iota,x})$ for all $x\in\mathbb Z_L$. \end{itemize} More specifically, performing explicitly the average via \eqref{eq:measure}, we find \begin{equation} \mathbb T = (\tilde{\mathbb U} \otimes \tilde{\mathbb U}^*) \mathbb O^\dag_1 (\tilde{\mathbb W} \otimes \tilde{\mathbb W}^*) \mathbb O^{\phantom{\dag}}_0, \label{eq:SFFTM2} \end{equation} where we introduced \begin{align} \tilde{\mathbb U} &:= \!\!\prod_{\tau \in \mathbb Z_{t}+\frac{1}{2}}\!\!\eta_{\tau,t}(\tilde U)\,, \label{eq:Utildee}\\ \tilde{\mathbb W} &:= \prod_{\tau \in \mathbb Z_{t}} \eta_{\tau,t}(\tilde W)\,, \label{eq:Utildeo}\\ \mathbb O_{\iota'} &:= \mathbb O_{0\iota'} \mathbb O_{1\iota'}= \mathbb O_{1\iota'} \mathbb O_{0\iota'}\,,\\ \mathbb O_{\iota\iota'} &:= \int\!{\rm d}^{d^2-1}\boldsymbol{\theta}\,g_{\iota\iota'}(\boldsymbol{\theta}) \exp\left(i\boldsymbol\theta\cdot (\boldsymbol{M}_{\iota}\otimes \mathbbm{1}_{2t}-\mathbbm{1}_{2t}\otimes \boldsymbol{M}^*_{\iota})\right)\,. \label{eq:defOa} \end{align} Here $\boldsymbol{M}_\iota = (M_{1,\iota},M_{2,\iota},\ldots, M_{d^2-1,\iota})$ with $M_{a, \iota}$ denoting the representation of the Hermitian generators of ${\mathfrak su}(d)$ in the lattice made of integer and half-odd-integer time indices \begin{equation} M_{a, \iota} :=\sum_{\tau \in \mathbb Z_t + \frac{1}{2}\iota} \!\!\!\!\!\!\sigma_{a,\tau}\,,\quad \sigma_{a,\tau} : = \Pi^{2\tau}_{2t}(\sigma_a\otimes \mathbbm{1}_{2t-1})\Pi^{-2\tau}_{2t}\,,\quad \iota\in\{0,1\}\,. \label{eq:sublatticeM} \end{equation} Normalisability and non-negativity of the probability densities $g_{\iota\iota'}$ imply the following important properties of the operators $\mathbb O_{\iota\iota'}$: \begin{enumerate} \item[(a)] $\mathbb O_{\iota\iota'}$ is a \emph{non-expansive mapping}: \begin{equation} \!\!\!\| \mathbb O_{\iota\iota'} \| \le \int\!{\rm d}^{d^2-1}\boldsymbol{\theta}\, |g_{\iota\iota'}(\boldsymbol{\theta})| \left\|\exp\left(i\boldsymbol\theta\cdot (\boldsymbol{M}_{\iota}\otimes \mathbbm{1}_{2t}-\mathbbm{1}_{2t}\otimes \boldsymbol{M}^*_{\iota})\right)\right\| = 1\,. \label{eq:propi} \end{equation} \item[(b)] Let $\mathbb O_{\iota\iota'} \ket{B} = e^{i\phi} \ket{B}$ for some $\ket{B}\in \mathcal H_{2t}\otimes \mathcal H_{2t}$ and $\phi\in\mathbb R$. Then: \begin{equation} \phi = 0,\,\, \textrm{and}\,\, (M_{a,\iota}\otimes \mathbbm{1}_{2t}-\mathbbm{1}_{2t}\otimes M^*_{a,\iota})\ket{B} = 0\,,\,\, \forall a\in\{1,2,\ldots,d^2-1\}\,. \label{eq:propii} \end{equation} Indeed, the assumptions imply \begin{equation} e^{i\boldsymbol\theta\cdot (\boldsymbol{M}_{\iota}\otimes \mathbbm{1}_{2t}-\mathbbm{1}_{2t}\otimes \boldsymbol{M}^*_{\iota})}\ket{B} = e^{i\phi} \ket{B}, \label{eq:propsup} \end{equation} for a dense set $\mathcal S \ni \boldsymbol{\theta}$ with positive measure, e.g. the support of $g_{\iota\iota'}$ or its dense subset. Without loss of generality we can assume that $\mathcal S$ contains the origin $\boldsymbol 0\in\mathcal S$. Taking a partial derivative $\frac{\partial}{\partial\theta_a}|_{\boldsymbol\theta=\boldsymbol 0}$ of (\ref{eq:propsup}) we obtain (\ref{eq:propii}). \end{enumerate} As convenient examples we can consider a Gaussian measure \begin{equation} g_{\iota\iota'}(\boldsymbol \theta)= \prod_{a=1}^{d^2-1}\frac{1}{\sqrt{2\pi}\nu_{a\iota\iota'}} \exp\left(-{\frac{1}{2}} \frac{\theta_a^2}{\nu^2_{a\iota\iota'}}\right)\,, \end{equation} or a box-measure \begin{equation} g_{\iota\iota'}(\boldsymbol \theta)= \prod_{a=1}^{d^2-1} \frac{1}{2\nu_{a\iota\iota'}} \Theta(\nu_{a\iota\iota'}-|\theta_a|)\,, \end{equation} where choosing sufficiently small nonvanishing variabilities $\nu_{a\iota\iota'} > 0$ allows for arbitrary concentration of measure around the identity in ${\rm SU}(d)$, and hence description of an ``almost clean system''. \subsection{Relevant limits} For the class of dual-unitary circuits, which is the main focus of this paper and will be elaborated in the next section, the expression (\ref{eq:SFFduality}) in terms of the map $\mathbb T$ [which is, in fact, a vectorised form of a completely positive trace preserving and unital mapping over ${\rm End}(\mathcal{H}_{2t})$] allows us to explicitly compute the SFF at any fixed $t$ in the thermodynamic limit ${L\to\infty}$. In particular, as we show below, we find that \begin{equation} \lim_{L\to\infty}K(t,L) = \lim_{\mathcal{ N}\to\infty}K_{\rm RMT}(t,\mathcal{N})=t,\qquad \forall t. \end{equation} This fact is quite remarkable and signals a special property of the dual-unitary systems considered here. Indeed, in generic quantum chaotic systems one typically observes that there exists a timescale $t^*(L)$ such that the universal behaviour described by RMT emerges only for times ${t>t^*(L)}$. This timescale, usually referred to as the Thouless (or Ehrenfest) time, is typically observed to grow monotonically with $L$~\cite{ljubotina,chalker,largeq}. The fact that for us, instead, $t^*(L)$ is strictly equal to zero can be interpreted as a sort of ``critical chaotic" (scale-free) property of dual-unitary systems. It would certainly be desirable to address spectral correlations on other scales. For instance, those on the scale of mean level spacing (which becomes exponentially small in $L$ for a many-body system) and correspond to times $t$ of the order of Heisenberg time $\mathcal N = 2^L$. This would require studying the scaling limit \begin{equation} \mathcal{K}(\tau) = \lim_{L\to\infty} 2^{-L} K\left(\lfloor 2^L \tau\rfloor,L\right). \end{equation} At the moment, however, we do not foresee any method to rigorously attack this challenging issue. \section{Statement of the main results} \label{sec:main} \subsection{Exact SFF at large $L$ for dual-unitary circuits} \label{sec:leadingeigs} Using the representation \eqref{eq:SFFduality} we see that to compute the SFF one has to determine the spectrum of $\mathbb T$. In particular, to obtain $K(t,L)$ in the large size limit $L\to\infty$, it is sufficient to find all the eigenvalues with maximal magnitude. To achieve this goal we consider a special class of local quantum circuits called \emph{dual-unitary} circuits~\cite{BKP19}. These systems are characterised by the property that their dual local gates (cf. \eqref{eq:duallocalgate}) are unitary. Specifically, we consider local gates $U$ and $W$ (cf.~(\ref{eq:Floquetgates1}, \ref{eq:Floquetgates2})) that simultaneously fulfil \begin{align} &U^\dag U = U U^{\dag}=\mathbbm{1}, & &\tilde U^\dag \tilde U = \tilde U \tilde U^{\dag}=\mathbbm{1}\,,\label{eq:dualunitarityU}\\ &W^\dag W= W W^{\dag}=\mathbbm{1}, & &\tilde W^\dag \tilde W = \tilde W \tilde W^{\dag}=\mathbbm{1}\,. \label{eq:dualunitarityW} \end{align} The above conditions admit non-trivial solutions for any local dimension $d\geq2$~\cite{Gutkin20,ClLa20b}. A complete classification of solutions, however, has been achieved only for $d=2$~\cite{BKP19}. An immediate consequence of \eqref{eq:dualunitarityU} and \eqref{eq:dualunitarityW} is that both $\tilde{\mathbb U}$ \eqref{eq:Utildee} and $\tilde{\mathbb W}$ \eqref{eq:Utildeo} are unitary. This allows us to prove the following Lemma: \begin{lemma} \label{lemma1} For dual-unitary circuits the matrix $\mathbb T$ (\ref{eq:SFFTM}) fulfils the properties: \begin{itemize} \item[(i)] $|\lambda|\leq 1$ for all $\lambda\in {\rm spect}(\mathbb T)$. \item[(ii)] If $\mathbb T\!\ket{A} =e^{i \phi} \!\ket{A}$, $\phi\in\mathbb R$, then \begin{eqnarray} &&(\tilde{\mathbb U} \otimes \tilde{\mathbb U}^*)\cdot (\tilde{\mathbb W} \otimes \tilde{\mathbb W}^*)\ket{A}=e^{i \phi}\!\ket{A}\,,\nonumber \\ && ( M_{a, \iota}\otimes \mathbbm{1}_{2t}-\mathbbm{1}_{2t}\otimes M_{a, \iota}^*)(\tilde{\mathbb W} \otimes \tilde{\mathbb W}^*) \ket{A} = 0\,,\label{eq:conditionsstateA}\\ && ( M_{a, \iota}\otimes \mathbbm{1}_{2t}-\mathbbm{1}_{2t}\otimes M_{a, \iota}^*)\ket{A} = 0, \quad \iota\in\{0,1\},\;a\in\{1,2,\ldots,d^2-1\}\,. \nonumber \end{eqnarray} \item[(iii)] For any unimodular eigenvalue $\lambda$, $|\lambda| =1$, its algebraic and geometric multiplicities coincide (i.e. its Jordan blocks are trivial). \end{itemize} \end{lemma} In essence, (i) and (iii) mean that $\mathbb T$ is a linear non-expansive mapping over $\mathcal H_{2t}\otimes \mathcal H_{2t}$, while (ii) suggests that computation of invariant subspaces can be reduced to simpler algebraic problems. Indeed, it ensures that for dual-unitary circuits \begin{equation} \mathcal N(\phi):=\frac{1}{L} \sum_{\ell = 1}^L e^{-i\phi \ell} K(t,\ell) \end{equation} approaches a finite value when $L\to\infty$ which is given by the number of linearly independent solutions $\ket{A}$ of the system of equations \eqref{eq:conditionsstateA}. The number of solutions for $\phi=0$ give the SFF averaged over the system size, while showing that $\phi=0$ is the only phase for which there are nontrivial solutions gives the thermodynamic limit $\lim_{L\to\infty} K(t,L)$. In order to further simplify the conditions (\ref{eq:conditionsstateA}) we introduce a vector-operator isomorphism $\mathcal H \otimes \mathcal H\overset{vo}{\longleftrightarrow} {\rm End}(\mathcal H)$. Specifically, we define in the canonical basis \eqref{eq:realbasis}: \begin{equation} \ket{j}\otimes \ket{j'} \overset{vo}{\longleftrightarrow} \ket{j}\bra{j'}\,. \end{equation} This implies that the problem of finding all linearly independent states $\ket{A}$ solving \eqref{eq:conditionsstateA} is mapped to the one of finding all linearly independent operators $A$ satisfying, for all $a\in\{1,2,\ldots, d^2-1\}$ and $\iota \in\{0,1\}$: \begin{equation} \tilde{\mathbb U} \tilde{\mathbb W} A \tilde{\mathbb W}^\dag \tilde{\mathbb U}^\dag =e^{i \phi} A,\quad [M_{a, \iota},A] = 0, \quad [\tilde{\mathbb W}^\dag M_{a, \iota} \tilde{\mathbb W},A]=0\,. \label{eq:CondSU2} \end{equation} The above conditions can be simplified further by making use of an explicit parametrisation of a dual-unitary matrices. Specifically, we parametrise the matrix $D \in {\rm End}(\mathbb C^d\otimes \mathbb C^d)$ fulfilling $D D^\dag= D^\dag D=\mathbbm{1}$ and $\tilde D \tilde D^\dag=\tilde D^\dag \tilde D=\mathbbm{1}$ as follows \begin{equation} D = (u_1\otimes u_2) S e^{i J s_3 \otimes s_3} (u_3\otimes u_4), \qquad\qquad J\in[0,\pi], \label{eq:parametrisationgen} \end{equation} where $u_j\in {\rm U}(d)$ are arbitrary unitary matrices (`local gates') and $S\in {\rm End}(\mathbb C^d\otimes \mathbb C^d)$ is the SWAP operator defined as \begin{equation} S \ket{j_1}\otimes \ket{j_2} = \ket{j_2}\otimes \ket{j_1}\qquad \forall j_1,j_2\in\{0,1,\ldots d-1\}\,. \end{equation} Finally, here and in the following $s_1,s_2,s_3$ designate the `spin matrices' carrying $d-$dimensional irreducible representation of $SU(2)$ over $\mathcal H_1$, satisfying \begin{equation} [s_a,s_b]= i\sum_{c=1}^3 \epsilon_{abc} s_c, \end{equation} where $\epsilon_{abc}$ is the three-dimensional Levi-Civita tensor, and we choose $s_3$ to be diagonal in the canonical basis \eqref{eq:realbasis} \begin{equation} s_3 = {\rm diag}\left( -\frac{d-1}{2},-\frac{d-3}{2},\ldots,\frac{d-1}{2}\right)\,. \label{eq:diag} \end{equation} Local embeddings into ${\rm End}(\mathcal H_{2t})$ are, like in \eqref{eq:sublatticeM}, defined as \begin{equation} s_{a,\tau} : = \Pi^{2\tau}_{2t}(s_a\otimes \mathbbm{1}_{2t-1})\Pi^{-2\tau}_{2t}\,. \end{equation} Note that for $d=2$ the parametrisation (\ref{eq:parametrisationgen}) exhausts all dual-unitary circuits~\cite{BKP19} while for $d>2$ it characterises a physically interesting sub-class~\cite{ClLa20b}. Plugging \eqref{eq:parametrisationgen} in the definitions \eqref{eq:Utildee}--\eqref{eq:Utildeo} we find \begin{align} &\tilde{\mathbb U} = e^{i \theta} e^{i \boldsymbol\alpha_0\cdot \boldsymbol M_0} e^{i \boldsymbol\alpha_1\cdot \boldsymbol M_1} \,\tilde{\mathbb V} \, e^{i \boldsymbol\beta_0\cdot \boldsymbol M_0} e^{i \boldsymbol\beta_1\cdot \boldsymbol M_1},\nonumber \\ &\tilde{\mathbb W} = e^{i \theta'} e^{i \boldsymbol\gamma_0\cdot \boldsymbol M_0} e^{i \boldsymbol\gamma_1\cdot \boldsymbol M_1} \,\tilde{\mathbb V}' \, e^{i \boldsymbol\delta_0\cdot \boldsymbol M_0} e^{i \boldsymbol\delta_1\cdot \boldsymbol M_1}, \label{eq:UGE} \end{align} where $\boldsymbol{\alpha}_\iota,\boldsymbol{\beta}_\iota,\boldsymbol{\gamma}_\iota,\boldsymbol{\delta}_\iota\in\mathbb R^{d^2-1}$, $\theta,\theta'\in\mathbb R$, and we introduced \begin{align} &\tilde{\mathbb V} := (S e^{i J s_3 \otimes s_3})^{\otimes t}\,, \label{eq:genmag}\\ &\tilde{\mathbb V}' := \Pi_{2t} (S e^{i J' s_3 \otimes s_3})^{\otimes t} \Pi^\dag_{2t}\,.\nonumber \end{align} We are then able to simplify the conditions for the existence of unimodular eigenvalues and write their invariant eigenoperator spaces in terms of a simple algebraic commutant: \begin{lemma} \label{lemma2} For $J, J'\neq 0$ the conditions \eqref{eq:CondSU2} cannot be met unless $\phi=0$. In this case, they are equivalent to \begin{equation} [A, M_{a,\iota}]=0, \quad [A, M_{ab,\iota}]=0\,, \qquad a,b\in\{1,2,\ldots,d^2-1\}, \; \iota\in\{0, 1\}\,. \label{eq:finalconditions} \end{equation} Here we introduced the 2-site magnetization operators of the even and odd spin sub-lattices \begin{equation} M_{a b,\iota}:= \sum_{\tau\in\mathbb Z_t+\frac{1}{2}\iota} \!\!\!\sigma_{a, \tau} \sigma_{b,\tau+\frac{1}{2}}\,, \qquad\qquad \iota\in\{0,1\}\,. \label{eq:doubleM} \end{equation} \end{lemma} As a corollary of Lemma~\ref{lemma1} and Lemma~\ref{lemma2}, we can express the SFF in the limit $L\to\infty$ in terms of the dimension of the eigenspace of eigenvalue 1, which in turn (Lemma~\ref{lemma2}) equals the dimension (in ${\rm End}({\cal H}_{2t})$) of the commutant $\mathcal M'$ of the set \begin{equation} \mathcal M:= \{ M_{a,\iota} \}_{a,\iota} \cup \{ M_{ab,\iota} \}_{a,b,\iota}\,. \label{eq:finalset} \end{equation} Namely, \begin{equation} \lim_{L\to\infty} K(t,L) = \dim \mathcal M'. \label{eq:limitcommutant} \end{equation} In fact, $\mathcal M'$ can be completely characterised: \begin{theorem} \label{theorem1} The commutant $\mathcal M'$ is the span of the representation of the cyclic group $C_t$ of even-site translations on a periodic chain of $2t$ spins: \begin{equation} \mathcal M' = {\rm span}\{ \Pi_{2t}^{2\tau};\,\, \tau = 0,1,\ldots t-1\}\,. \label{eq:commutant} \end{equation} \end{theorem} Hence, we arrive at the following corollary, which summarises our first main result \begin{corollary} For local quantum circuits \eqref{eq:Floquet} with local gates of the form (\ref{eq:Floquetgates1},\ref{eq:Floquetgates2}) and \begin{equation} U = (u_1\otimes u_2) S e^{i J s_3 \otimes s_3} (u_3\otimes u_4),\qquad W = (u'_1\otimes u'_2) S e^{i J' s_3 \otimes s_3} (u'_3\otimes u'_4) \end{equation} where $u_j, u'_j\in {\rm U}(d)$ and $J, J'\neq 0$, the SFF \eqref{eq:SFF} averaged according to the measure \eqref{eq:measure} fulfils \begin{equation} \lim_{L\to\infty} K(t,L) = t\,. \label{eq:mainresult} \end{equation} \end{corollary} This is precisely the CUE result for \emph{all} times. It is remarkable that 2-point spectral correlations of dual-unitary circuits agree with RMT at all scales. Note that the restriction $J, J'\neq 0$ for the validity of the statement is not surprising. Indeed, for ${J=0}$ the gate $U$ does not encode interactions among the qudits, they are evolved in an entirely independent fashion (an analogous conclusion holds concerning $W$ for ${J'=0}$). This means that if one of $J$ and $J'$ is equal to zero not all the qudits are coupled by the dynamics and one cannot expect $\mathbb U_{L}$ in Eq.~\eqref{eq:Floquet} to behave like a random matrix on the whole Hilbert space. \subsection{Results on SFF at large $L$ for \emph{T-symmetric} dual-unitary circuits} To obtain circuits with time-reversal symmetry one has to choose local gates $U, W \in {\rm U}(d^2)$ and on-site disorder $u_{x}, w_{x} \in {\rm U}(d)$ which are compatible with the conditions \eqref{eq:timereversal}. A generic choice is \begin{equation} U=U^T,\qquad\qquad W=W^T, \label{eq:TIEc1} \end{equation} and gates $u_{x}, w_{x}$ of the form \begin{equation} u_{x} = e^{i \boldsymbol{\theta}_{x}\cdot\boldsymbol{\sigma}}, \qquad w_{x} = e^{i \boldsymbol{\theta}_{x}\cdot\boldsymbol{\sigma}^T} = u_x^T, \qquad x\in\Lambda_L,\;\;\boldsymbol{\theta}_{x}\in\mathbb R^{d^2-1}. \label{eq:urxT} \end{equation} where now $\boldsymbol{\theta}_{x}$ is the same in both $u_{x}$ and $w_{x}$. The expectation can again be explicitly written in terms of a factorised measure as: \begin{equation} \mathbb E_T[f] = \int f(\boldsymbol\theta) \prod_{x=0}^{L-1}\prod_{\iota'=0}^1 g_{\iota'}(\boldsymbol{\theta}_{x+\frac{\iota'}{2}}) {\rm d}^{d^2-1}\boldsymbol{\theta}_{x+\frac{\iota'}{2}} \,, \quad \boldsymbol \theta \equiv (\boldsymbol{\theta}_{x})_{x\in\Lambda_L}\,. \label{eq:measureT} \end{equation} where $g_{\iota'}\in L^1[\mathbb R^{d^2-1}]$ is a pair of arbitrary probability densities of i.i.d. random variables $\boldsymbol{\theta}_{x}$ on integer ($\iota'=0$) and half-odd-integer ($\iota'=1$) sublattices. Considering local gates fulfilling the conditions (\ref{eq:TIEc1}, \ref{eq:urxT}) and averaging according to the measure $\mathbb E_T[\cdot] $, one can repeat the reasoning of Sec.~\ref{sec:duality} and conclude \begin{eqnarray} K_T(t,L) = \mathbb E_T\left[\left({\rm tr}\,\mathbb U_{L}^t\right) \left({\rm tr}\,\mathbb U_{L}^t\right)^*\right] = {\rm tr}\,\mathbb T_T^L, \label{eq:SFFdualityT} \end{eqnarray} with \begin{equation} \mathbb T_T : = (\tilde{\mathbb U} \otimes \tilde{\mathbb U}^*) {\mathbb O}_{T,1}^\dag (\tilde{\mathbb W} \otimes \tilde{\mathbb W}^*) {\mathbb O}^{\phantom{\dag}}_{T, 0}\,. \end{equation} Here $\tilde{\mathbb U}$, and $\tilde{\mathbb W}$ are defined as in Eqs.~(\ref{eq:Utildee},\,\ref{eq:Utildeo}) while time-reversal symmetric averaging operator ${\mathbb O}_{T, \iota'}$ reads as \begin{equation} {\mathbb O}_{T, \iota'} := \int\!{\rm d}^{d^2-1}\boldsymbol{\theta}\,g_{\iota'}(\boldsymbol{\theta}) \exp\left(i\boldsymbol\theta\cdot (\boldsymbol{M} \otimes \mathbbm{1}_{2t}-\mathbbm{1}_{2t}\otimes \boldsymbol{M}^*)\right)\,. \label{eq:OopT} \end{equation} Here $\boldsymbol{M} = (M_{1},M_{2},\ldots, M_{d^2-1})$ with $M_{a}$ denoting the representation of the $a$-th Hermitian generator of ${\mathfrak su}(d)$ in the \emph{full} time lattice \begin{equation} M_{a} :=\sum_{\tau \in \Lambda_t} \sigma_{a,\tau} = M_{a,0}+M_{a,1}. \label{eq:fulllatticeM} \end{equation} Assuming that, together with the conditions (\ref{eq:TIEc1}, \ref{eq:urxT}), the local gates also fulfil (\ref{eq:dualunitarityU}, \ref{eq:dualunitarityW}) (i.e. they are \emph{dual-unitary}), and noting that the averaging operator also satisifies the properties (\ref{eq:propi},\ref{eq:propii}), one can immediately write an analogue of the Lemma~\ref{lemma1} for the transfer matrix $\mathbb T_T$ (its proof \ref{sec:prooflemma1} carries over). Namely, one can prove that $\mathbb T_T$ is again a linear non-expansive mapping over $\mathcal H_{2t}\otimes \mathcal H_{2t}$, and its eigenvectors corresponding to unimodular eigenvalues are determined by the conditions \eqref{eq:conditionsstateA} with $M_{a,\iota}$ replaced by $M_{a}$. Following the logic of Sec.~\ref{sec:leadingeigs}, one can map the problem of finding all linear independent eigenvectors of $\mathbb T_T$ corresponding to unimodular eigenvalues to a simpler algebraic problem. In this case the equivalent algebraic problem is to find all linearly independent operators $A$ fulfilling \begin{equation} \tilde{\mathbb U} \tilde{\mathbb W} A \tilde{\mathbb W}^\dag\,\tilde{\mathbb U}^\dag =e^{i \phi} A,\quad [M_{a},A] = 0, \quad [\tilde{\mathbb W}^\dag M_{a} \tilde{\mathbb W},A]=0\,. \label{eq:CondSU2T} \end{equation} for some $\phi\in\mathbb R$ and all $a\in\{1,2,\ldots,d^2-1\}$. Using the explicit from \eqref{eq:parametrisationgen} of the 2-site dual-unitary gates we can again simplify the conditions~\eqref{eq:CondSU2T} and write their invariant eigenoperator spaces in terms of simple algebraic commutants. In this case, to lift some technical complications, we specialise the treatment to the case of qubits ($d=2)$: \begin{lemma} \label{lemma3} For $J, J'\neq 0$ and $d=2$ the conditions \eqref{eq:CondSU2T} cannot be met unless $\phi=0$. In this case, they are equivalent to \begin{equation} [A, M_{a}]=0,\; [A, M_{ab, \iota}+R_{2t} M_{ab, \iota}R_{2t}]=0\,, \quad a,b\in\{1,2,3\}\,, \; \iota\in\{0, 1\}\,, \label{eq:finalconditionsT} \end{equation} where the 2-site magnetization operators of the even and odd spin sub-lattices are defined in \eqref{eq:doubleM} and $R_{2t}$ is the reflection of the time lattice $\Lambda_t$ around the centre, i.e. \begin{equation} R_{2t} \ket{j_1}\otimes \ket{j_2}\otimes\cdots\ket{j_{2t}} = \ket{j_{2t}} \otimes \cdots\ket{j_2} \otimes \ket{j_{1}}. \end{equation} \end{lemma} As before, this lemma allows us to express the SFF in the limit $L\to\infty$ in terms of the dimension (in ${\rm End}({\cal H}_{2t})$) of the commutant $\mathcal M'_T$ of the set \begin{equation} \mathcal M_T := \{ M_{a} \}_{a} \cup \{ M_{ab,\iota} +R_{2t} M_{ab, \iota}R_{2t} \}_{a,b,\iota}\,. \end{equation} Namely, \begin{equation} \lim_{L\to\infty} K_T(t,L) = \dim \mathcal M_T'. \end{equation} The commutant $\mathcal M_T'$ can again be completely characterised by proving a statement analogous to Theorem~\ref{theorem1}, which we present here without proof. \begin{conjecture} \label{conjecture1} The commutant $\mathcal M'_T$ is the linear span of the representation of the dihedral group $D_t$, i.e. the symmetry group of a polygon of $t$ vertices, on a periodic chain of $2t$ spins: \begin{equation} \mathcal M'_T = {\rm span}\{R_{2t}^n\Pi_{2t}^{2\tau};\,\, \tau = 0,1,\ldots t-1,\, n= 0,1\}\,. \label{eq:commutantT} \end{equation} \end{conjecture} The number of independent elements of the dihedral group is established by: \begin{lemma} \label{lemma4} The number of linearly independent elements in the representation of $D_t$ in $\mathcal H_t$ formed by $\{R_{2t}^n\Pi_{2t}^{2\tau}\}^{n=0,1}_{\tau=0,1,\ldots t-1}$ is $2t$. \end{lemma} Hence, we arrive at the following corollary: \begin{corollary} For local quantum circuits \eqref{eq:Floquet} with $d=2$, local gates of the form (\ref{eq:Floquetgates1},\ref{eq:Floquetgates2}), and \begin{equation} U = (u_1\otimes u_2) S e^{i J s_3 \otimes s_3} (u_1^{T}\otimes u_2^{T}),\quad W = (u'_1\otimes u'_2) S e^{i J' s_3 \otimes s_3} (u^{\prime\, T}_1\otimes u^{\prime\, T}_2), \end{equation} where $u_j, u'_j\in {\rm U}(2)$ and $J, J'\neq 0$, the SFF \eqref{eq:SFF} averaged according to the measure \eqref{eq:measureT} fulfils \begin{equation} \lim_{L\to\infty} K_T(t,L) = 2t\,. \end{equation} \end{corollary} This is precisely the COE result for \emph{all} times. Once again we see that 2-point spectral correlations of dual-unitary circuits agree with RMT at all scales. The proof of Conjecture~\ref{conjecture1} follows the same ideas as that of Theorem~\ref{theorem1} but is technically more involved and will be presented elsewhere~\cite{COE}. A similar proof, for the special case of the time-reversal symmetric self-dual kicked Ising model, has been presented in the supplemental material of Ref.~\cite{BKP18}. \section{Proofs} \label{sec:proofs} \subsection{Proof of Lemma \ref{lemma1}} \label{sec:prooflemma1} As a consequence of properties (\ref{eq:propi},\ref{eq:propii}) the spectra of $\mathbb O_{\iota\iota'}$ belong to the open unit disk with 1 attached to it, ${\mathcal D}_1 = \{z\in\mathbb C; |z|<1\} \cup \{1\}\supset {\rm spect}(\mathbb O_{\iota\iota'})$. Since $\mathbb O_{0\iota'}$ and $\mathbb O_{1\iota'}$ commute, ${\rm spect}(\mathbb O_{\iota'}=\mathbb O_{0\iota'}\mathbb O_{1\iota'}) \subset {\mathcal D}_1$, and all eigenvectors $\ket{R}$ of $\mathbb O_{\iota'}$ with unique unimodular eigenvalue $1$ are characterised by \begin{equation} (M_{a, \iota}\otimes \mathbbm{1}_{2t}-\mathbbm{1}_{2t}\otimes M_{a, \iota}^*)\ket{R} = 0, \quad \iota\in\{0,1\},\;a\in\{1,2,\ldots,d^2-1\}\,. \label{eq:unique} \end{equation} Since $M_{a, \iota}\otimes \mathbbm{1}_{2t}-\mathbbm{1}_{2t}\otimes M_{a, \iota}^*$ are Hermitian, exactly the same conditions uniquely characterise the eigenvalue $1$ eigenvectors of $\mathbb O^\dagger_{\iota'}$. We now turn to SFF transfer matrix \eqref{eq:SFFTM2} and write \begin{equation} \mathbb T^\dag \mathbb T= \mathbb O_0^{{\dag}} (\tilde{\mathbb W} \otimes \tilde{\mathbb W}^*)^\dag \mathbb O_1^{\phantom{\dag}} \mathbb O_1^{{\dag}} (\tilde{\mathbb W} \otimes \tilde{\mathbb W}^*) \mathbb O_0^{\phantom{\dag}}\,. \label{eq:TT} \end{equation} Let $\ket{A}$ be a normalised eigenvector of $\mathbb T$ associated to the eigenvalue $\lambda$. Considering the expectation value of \eqref{eq:TT} we have, since $\bra{B}\mathbb O^\dagger_{\iota'} \mathbb O_{\iota'}\ket{B} \le \langle B|B\rangle$, $\bra{B}\mathbb O_{\iota'} \mathbb O^\dagger_{\iota'}\ket{B} \le \langle B|B\rangle$, for any $\ket{B}$: \begin{eqnarray} |\lambda|^2 &=& \braket{A|\mathbb T^\dag \mathbb T|A} = \braket{A|\mathbb O_0^{{\dag}} (\tilde{\mathbb W} \otimes \tilde{\mathbb W}^*)^\dag \mathbb O_1 \mathbb O_1^{{\dag}} (\tilde{\mathbb W} \otimes \tilde{\mathbb W}^*) \mathbb O_0 |A} \nonumber \\ &\le & \braket{A|\mathbb O_0^{{\dag}} (\tilde{\mathbb W} \otimes \tilde{\mathbb W}^*)^\dag (\tilde{\mathbb W} \otimes \tilde{\mathbb W}^*) \mathbb O_0 |A} = \braket{A|\mathbb O_0^{{\dag}} \mathbb O_0 |A} \leq 1\,, \label{eq:boundx} \end{eqnarray} which proves point (i). The eigenvalue $\lambda$ is unimodular only if both inequalities in (\ref{eq:boundx}) are saturated. The second one implies (\ref{eq:unique}) for $\ket{R}=\ket{A}$, i.e. the second line of \eqref{eq:conditionsstateA}, while the first one implies (\ref{eq:unique}) for $\ket{R}= (\tilde{\mathbb W} \otimes \tilde{\mathbb W}^*) \ket{A}$, i.e. the third line of \eqref{eq:conditionsstateA}. Since $\mathbb O_0 \ket{A}=\ket{A}$, $\mathbb O^\dagger_1 (\tilde{\mathbb W} \otimes \tilde{\mathbb W}^*) \ket{A} = (\tilde{\mathbb W} \otimes \tilde{\mathbb W}^*) \ket{A}$ we have the first line of \eqref{eq:conditionsstateA}. This proves point (ii). Finally, we prove the last point by contradiction: assuming that the eigenvalue $\lambda$ corresponds to a non-trivial Jordan block, there must exist a normalised vector $\ket{B}$ such that \begin{equation} \mathbb T \ket{B}=\lambda \ket{B} + \alpha \ket{A}\,,\qquad \alpha\neq0\,, \end{equation} where $\ket{A}$ is the eigenvector corresponding to the eigenvalue $\lambda$ (where we can choose $\braket{A|B}=0$). Reasoning as above we have \begin{equation} \braket{B|\mathbb T^\dag \mathbb T|B} = 1+ |\alpha|^2 \ge 1\,, \end{equation} which is a contradiction. \subsection{Proof of Lemma \ref{lemma2}} Plugging \eqref{eq:UGE} into the first condition \eqref{eq:CondSU2} and using the second condition to commute $\boldsymbol\alpha\cdot \boldsymbol M_{\iota}$ around $A$ we bring the conditions \eqref{eq:CondSU2} to the following form \begin{align} \tilde{\mathbb V}\,\tilde{\mathbb V}^{\prime} A \tilde{\mathbb V}^{\prime\dag} \tilde{\mathbb V}^\dag &\!\!=e^{i \phi} A, & [M_{a,\iota},A] &\!=\!0, & [\tilde{\mathbb V}^{\prime \dag} M_{a,\iota} \tilde{\mathbb V}^{\prime},A]&=0, \label{eq:intermediatecond} \end{align} where $a\in\{1,2,3\}$, $\iota\in\{0,1\}$. Let us now consider more closely the operator in the last relation. Considering for example $\iota=0$ sublattice and a combination of generators which yields the first spin matrix $s_1 = \boldsymbol \alpha \cdot \boldsymbol \sigma$, \begin{align} \tilde{\mathbb V}^{\prime \dag} (\boldsymbol\alpha\!\cdot\!\boldsymbol M_{0})\tilde{\mathbb V}^{\prime} &= \sum_{\tau \in\mathbb Z_t} \exp\left[- i J' s_{3, \tau-\frac{1}{2}} s_{3,\tau} \right]s_{1,\tau-\frac{1}{2}} \exp\left[i J' s_{3, \tau-\frac{1}{2}} s_{3,\tau} \right]\,, \end{align} where we used $S^\dag (\mathbbm{1}\otimes s_a) S = s_a\otimes \mathbbm{1}$. Resolving the identity in an eigenbasis of $s_{3}$ we find \begin{align} \tilde{\mathbb V}^{\prime \dag} (\boldsymbol\alpha\!\cdot\!\boldsymbol M_{0})\tilde{\mathbb V}^{\prime} &= \sum_{\tau\in\mathbb Z_t} s_{1,\tau-\frac{1}{2}} \cos(J's_{3,\tau}) + s_{2,\tau\!-\!\frac{1}{2}} \sin(J' s_{3, \tau}). \label{eq:VMVe} \end{align} Next, we consider \begin{align} &\tilde{\mathbb V}^{\prime \dag} (\boldsymbol\alpha\!\cdot\!\boldsymbol M_{0}) \tilde{\mathbb V}^{\prime} - e^{i\frac{\pi}{2} \boldsymbol\alpha\cdot\boldsymbol M_{0}} \tilde{\mathbb V}^{\prime \dag} (\boldsymbol\alpha\!\cdot\!\boldsymbol M_{0}) \tilde{\mathbb V}^{\prime} e^{-i\frac{\pi}{2}\boldsymbol\alpha\cdot\boldsymbol M_{0}}= \notag\\ & = 2\!\sum_{\tau\in\mathbb Z_t}\! s_{2,\tau-\frac{1}{2}} \sin(J' s_{3, \tau}).\label{eq:VM1Vod} \end{align} Since $\sin(J' s_{3})$ is Hermitian and traceless it can be expanded in terms of the generators $\{\sigma_a\}$, i.e. \begin{equation} \sin(J' s_{3}) = \boldsymbol{c}(J')\cdot \boldsymbol\sigma\,, \quad\textrm{where}\quad \boldsymbol{c}(J') \neq \boldsymbol 0 \;\;\textrm{for}\;\; J'\neq 0\,. \end{equation} Furthermore, since the adjoint representation of ${\rm SU}(d)$ is irreducible, we can for any non-vanishing vector $\boldsymbol{\beta}\in\mathbb R^{d^2-1}$, and $b \in\{1,\ldots,d^2-1\}$ find a vector $\boldsymbol\gamma\in\mathbb R^{d^2-1}$, such that \begin{equation} e^{i \boldsymbol\gamma\cdot\boldsymbol{M}_\iota} (\boldsymbol{\beta}\cdot\boldsymbol{\sigma}_ {\tau+\frac{1}{2}\iota}) e^{-i \boldsymbol\gamma\cdot\boldsymbol{M}_\iota}= \sigma_{b,\tau+\frac{1}{2}\iota},\quad \tau\in\mathbb Z_t\,. \end{equation} This means that, conjugating the operators on r.h.s. of \eqref{eq:VM1Vod} with appropriate $\boldsymbol\gamma\cdot\boldsymbol{M}_\iota$ on integer ($\iota=0$) and half-odd integer ($\iota=1$) spin sub-lattices independently, we can produce any operator of the form $M_{a b,1}$ (cf.~\eqref{eq:doubleM}). Since $A$ commutes with $M_{a,\iota}$, we have for $J'\neq 0$: \begin{equation} [A,M_{ab,1}]=0,\qquad a,b\in\{1,2,\ldots,d^2-1\}. \label{eq:AdoubleModd} \end{equation} To obtain an analogous statement for $M_{ab,0}$ we first note that combining the first and last relation of \eqref{eq:intermediatecond} yields \begin{equation} [\tilde{\mathbb V} M_{a,0} \tilde{\mathbb V}^\dag,A] =0\,, \qquad a\in\{1,2,\ldots, d^2-1\}\,. \end{equation} Proceeding as before we find \begin{equation} \tilde{\mathbb V} (\boldsymbol\alpha\!\cdot\!\boldsymbol M_{1}) \tilde{\mathbb V}^{\dag} = \sum_{\tau\in\mathbb Z_t} \cos(J s_{3,\tau}) s_{1,\tau+\frac{1}{2}}-\sin(J s_{3, \tau})s_{2,\tau+\frac{1}{2}}, \label{eq:VMVo} \end{equation} and \begin{align} \!\!\!\!\!\! e^{i\frac{\pi}{2} \boldsymbol\alpha\cdot\boldsymbol M_{1}} \tilde{\mathbb V} (\boldsymbol\alpha\!\cdot\!\boldsymbol M_{1}) \tilde{\mathbb V}^{\dag} e^{-i\frac{\pi}{2}\boldsymbol\alpha\cdot\boldsymbol M_{1}}- \tilde{\mathbb V} (\boldsymbol\alpha\!\cdot\!\boldsymbol M_{1}) \tilde{\mathbb V}^{\dag} & = 2\!\sum_{\tau\in\mathbb Z_t}\!\sin(J s_{3, \tau}) s_{2,\tau+\frac{1}{2}}. \end{align} Assuming $J\neq0$, we can repeat the reasoning after \eqref{eq:VM1Vod} and find: \begin{equation} [A,M_{ab,0}]=0,\qquad a,b\in\{1,2,\ldots, d^2-1\}. \label{eq:AdoubleMeven} \end{equation} Now we note that $\mathbb V$ and $\mathbb V'$ can be written in terms of double magnetisations~\eqref{eq:doubleM}. This can be seen by observing that \begin{equation} \sum_{a=1}^{d^2-1} \sigma_a\otimes\sigma_a \end{equation} is the quadratic Casimir operator of the representation of ${\rm SU}(d)$ over $\mathbb C^d \otimes \mathbb C^d$. Therefore, we must have \begin{equation} \sum_{a=1}^{d^2-1} \sigma_a\otimes\sigma_a = c_+ \frac{\mathbbm{1}+S}{2} + c_- \frac{\mathbbm{1}-S}{2} , \end{equation} for some $c_\pm\in\mathbb R$. Indeed, the symmetric and antisymmetric subspaces of $\mathbb C^d \otimes \mathbb C^d$ contain irreducible representations. Fixing the constants using the explicit form of $\{\sigma_a\}$ (see, e.g., Ref.~\cite{GenGellMann}) we find $c_\pm=\pm 2-2/d$. Using \begin{equation} S = e^{- i \frac{\pi}{2}}e^{i \frac{\pi}{2} S}, \end{equation} and writing $s_3=\sum_a \gamma_a \sigma_a$ we finally find \begin{align} \mathbb V &= e^{- i \frac{\pi}{2} t}\exp\Bigl[i \frac{\pi}{2} \sum_{\tau\in\mathbb Z_t} S_\tau\Bigr] \exp\Bigl[{i J\sum_{ab}\gamma_a\gamma_b M_{ab,0}}\Bigr] \notag\\ &= e^{- i \frac{\pi}{2d} t (1-d)}\exp\Bigl[{i \frac{\pi}{4} \sum_{a=1}^{d^2-1} M_{aa,0}}\Bigr] \exp\Bigl[{i J \sum_{ab} \gamma_a\gamma_b M_{ab,0}}\Bigr]\,, \label{eq:VMMMd}\\ \mathbb V' &= e^{- i \frac{\pi}{2} t} \Pi_{2t}\exp\Bigl[i \frac{\pi}{2} \sum_{\tau\in\mathbb Z_t} S_\tau\Bigr] \exp\Bigl[{i J' \sum_{a,b} \gamma_a\gamma_b M_{ab,0}}\Bigr] \Pi^\dag_{2t} \notag\\ &= e^{- i \frac{\pi}{2d} t (1-d)}\exp\Bigl[{i \frac{\pi}{4} \sum_{a=1}^{d^2-1} M_{aa,1}}\Bigr] \exp\Bigl[{i J'\sum_{ab}\gamma_a\gamma_b M_{ab,1}}\Bigr]\,, \label{eq:V'MMMd} \end{align} where we introduced \begin{equation} S_\tau := \mathbbm{1}_{2\tau-1}\otimes S \otimes\mathbbm{1}_{2t-2\tau-1}. \end{equation} Equations \eqref{eq:AdoubleModd}, \eqref{eq:AdoubleMeven}, \eqref{eq:VMMMd} and \eqref{eq:V'MMMd} imply that if $J,J'\neq0$ \begin{align} \tilde{\mathbb V}\,\tilde{\mathbb V}^{\prime} A \tilde{\mathbb V}^{\prime \dag} \tilde{\mathbb V}^\dag = A\,, \end{align} so that the first of conditions \eqref{eq:intermediatecond} can be fulfilled only for $\phi=0$ and in that case it follows from \eqref{eq:AdoubleModd}, \eqref{eq:AdoubleMeven}, and the second of \eqref{eq:intermediatecond}. This proves the Lemma. \subsection{Proof of Theorem \ref{theorem1}} To prove the Theorem we use of the following Lemma. \begin{lemma} \label{lemma5} Let $\mathcal K \subseteq {\rm End}(\mathcal H_{2t})$ be a multiplicative algebra of operators over $\mathcal H_{2t}$, generated by elements of $\mathcal M = \{ M_{a,\iota} \}_{a,\iota} \cup \{ M_{ab,\iota} \}_{a,b,\iota}$. The representations of $\mathcal K$ over the eigenspaces $\{\mathcal W_k \subset \mathcal H_{2t} \}_{k=0}^{t-1}$ of the 2-site shift operator $(\Pi_{2t})^2\mathcal W_k = e^{-2\pi i k/t}\mathcal W_k$ are all irreducible and inequivalent. \end{lemma} Our statement (Theorem~\ref{theorem1}) follows from a simple combination of Lemma~\ref{lemma5} (which is proven later below) and the Schur's Lemma. If some $A$ fulfils the conditions \eqref{eq:finalconditions}, it commutes with all elements of the algebra $\mathcal K$ generated by $\{M_{a, \iota}\}$ and $\{M_{ab, \iota}\}$. Since the representations of the algebra $\mathcal K$ in the eigenspaces $\mathcal W_k$ are irreducible and inequivalent, Schur's Lemma implies \begin{equation} A = \sum_{k=0}^{t-1} c_k Q_k, \label{eq:sol} \end{equation} where $c_k\in\mathbb C$ are arbitrary coefficients and \begin{equation} Q_k := \frac{1}{t}\sum_{\tau=0}^{t-1} e^{2\pi i \tau k/t} \Pi_{2t}^{2\tau}\,,\qquad k\in\{0,\ldots,t-1\}, \label{eq:Zk} \end{equation} are orthogonal projectors on $\{\mathcal W_k\}_{k=0}^{t-1}$, i.e. $Q_k Q_{k'} = \delta_{k,k'} Q_k$. This proves that $A$ is a linear combination of $t$ cyclic translations $\Pi_{2t}^{2\tau}$, i.e. $\mathcal K' = \mathcal M' = {\rm span}\{\Pi^{2\tau}_{2t};\,\tau=0,1\ldots,t-1\}$, concluding the proof of Theorem \ref{theorem1}. \subsection{Proof of Lemma \ref{lemma5}} We begin by introducing the following shorthand notation \begin{align} S_{\pm}& := \sum_{\tau\in \Lambda_t} s_{\pm,\tau} \left(s_{3,\tau}+\frac{d-1}{2}\right)\,,\label{eq:Spm} \\ R_{\pm, n}& := \sum_{\tau\in \Lambda_t} \left(s_{3,\tau}+\frac{d-1}{2}\right) s^n_{\pm,\tau+\frac{1}{2}}\,,\label{eq:Rpm} \\ T_{\pm, n}& := \sum_{\tau\in \Lambda_t} \left(s_{3,\tau}-\frac{d-1}{2}\right) s^n_{\pm,\tau-\frac{1}{2}}\,,\label{eq:Tpm} \\ M_{\pm\mp}& := \sum_{\tau\in \Lambda_t} s_{\pm,\tau}s_{\mp,\tau+\frac{1}{2}},\label{eq:Mpmmp} \\ Z_\iota & := \sum_{t\in\mathbb Z_t} s_{3,\tau+\frac{1}{2}\iota},\label{eq:Zdef} \end{align} where $n\in\{1,\ldots,d-1\}$ and \begin{equation} s_{\pm,\tau} = \frac{s_{1,\tau} \pm i s_{2,\tau}}{\sqrt 2},\qquad \tau\in\Lambda_t, \end{equation} are local spin raising/lowering operators. Using the commutation relations among $\{\sigma_a\}$ it is straightforward to show that all operators \eqref{eq:Spm}--\eqref{eq:Zdef} can be expressed as linear combinations of $M_{a,\iota}$, $M_{ab,\iota}$ (cf.~\eqref{eq:sublatticeM} and \eqref{eq:doubleM}). In addition, we also introduce the set of vectors (states) in $\mathcal H_{t}$ \begin{equation} \!\!\!\!\!\mathcal S_k : =\!\! \begin{cases} \begin{aligned} & \ket{0}^{\otimes 2t} \cup \{ \ket{n}_0;\, n\in \mathcal I\}\,\cup \\ &\{\ket{n, {\boldsymbol\nu}, {m}}_{\ell, 0};\, n, m \in \mathcal I, \boldsymbol\nu\in \mathcal J^{\ell-2}, 2\le\ell\le 2t\} \end{aligned} & k=0\\ \\ \begin{aligned} &\{ \ket{n}_k;\, n\in \mathcal I\}\,\cup \\ &\{\ket{n, {\boldsymbol\nu}, {m}}_{\ell, k};\, n, m \in \mathcal I, \boldsymbol\nu\in \mathcal J^{\ell-2}, 2\le\ell\le 2t\} \end{aligned} & k\!\in\!\{1,\ldots,\!2t\!-\!1\!\} \end{cases} \label{eq:states} \end{equation} where we defined the sets \begin{align} \mathcal I &:= \{1,\ldots,d-1\},\label{eq:setI}\\ \mathcal J &:= \{0, 1,\ldots,d-1\}, \end{align} and the vectors \begin{align} &\ket{n}_{k}:=\frac{1}{\sqrt{2t}}\sum_{j=0}^{2t-1} e^{i\frac{\pi k}{t} j}\Pi^j_{2t} s^{n}_{+,0} \ket{0}^{\otimes 2t},\label{eq:statesm1}\\ &{\ket{n_1, \underbrace{0\cdots0}_{\ell_1-1}, n_{2}, \cdots, n_a, \underbrace{0\cdots0}_{\ell_a-1}, n_{a+1}}\!{}_{\ell, k}}\notag\\ &\qquad\qquad := \frac{1}{\sqrt{2t}}\sum_{j=0}^{2t-1} e^{i\frac{\pi k}{t} j}\Pi^j_{2t} s^{n_1}_{+,0}s_{+,\frac{\ell_1}{2}}^{n_{2}}\!\!\cdots s_{+,\frac{\ell-\ell_a}{2}}^{n_{a}}s_{+,\frac{\ell-1}{2}}^{n_{a+1}} \ket{0}^{\otimes 2t}\!\!,\label{eq:statesm>1} \end{align} with $\{\ell_j\}_{j=1}^a\subset\{1,\ldots, 2t-1\}$ fulfilling \begin{equation} \mathbb N \ni \ell :=1+\sum_{j=1}^a\ell_{j}\leq 2t\,. \end{equation} We note that the `empty' state $\ket{0}\in \mathcal R$ (cf. \eqref{eq:realbasis}) satisfies \begin{equation} s_3\ket{0}= -\frac{d-1}{2}\ket{0}, \end{equation} (cf. \eqref{eq:diag}). The integer $\ell$ shall be referred to as the \emph{length} of the states~\eqref{eq:statesm>1} (one can verify that $\ell\geq 2$), while the states \eqref{eq:statesm1} have conventionally a unit length. For each value of $k$ the set \eqref{eq:states} contains $(d-1)^2d^{\ell-2}$ states for every length $\ell\geq2$ and $d-1$ states for $\ell=1$. Note that for each $k$ the states in \eqref{eq:states} have \emph{momentum} $\pi k/t$, i.e. \begin{equation} \Pi_{2t} \ket{n, {\boldsymbol \nu}, m}_{\ell, k} = e^{-i \frac{\pi k}{t}} \ket{n, {\boldsymbol \nu}, m}_{\ell, k}\,. \end{equation} The set $\mathcal S_k$ is complete in $\mathcal V_k$ --- the eigenspace of single-site shift $\Pi_{2t}$ associated with momentum $\pi k/t$ --- but are not all linearly independent: While for $\ell < t$ the states are clearly orthonormal, some of the states with $\ell \geq t$ can be represented by a string with a shorter length or they have multiple representations with the same length. One can then construct a basis $\mathcal B_k$ of $\mathcal V_k$ by extracting from $\mathcal S_k$ the maximal subset of linearly independent vectors. Here we want to prove the following (Lemma \ref{lemma5}): The representation of the algebra $\mathcal K$ generated by $\{M_{a,\iota}\}$ and $\{M_{ab,\iota}\}$ is irreducible in \begin{equation} \mathcal W_k\equiv{\rm span}(\mathcal B_k \cup \mathcal B_{k+t}),\qquad\qquad k\in\{0,1,\ldots,t-1\}\,, \end{equation} specifically in the eigenspace of $(\Pi_{2t})^2$ corresponding to the eigenvalue $e^{-{2\pi i k}/{t}}$. Moreover, the irreducible representations in different $\mathcal W_k$ are inequivalent. Noting that $\mathcal W_k$ are closed under the action of $\mathcal K$ (all generators commute with $(\Pi_{2t})^2$) we have that the following three requirements imply the statement of the Lemma: \begin{itemize} \item[(1)] All vectors in $\mathcal B_k$ are mapped into one another by elements of the algebra $\mathcal K$. \item[(2)] There is an element of $\mathcal K$ mapping $\ket{1}_k$ and $\ket{1}_{k+t}$ one into another. \item[(3)] There is no unitary matrix $C$ such that \begin{align} (Z_1)_k &= C (Z_1)_p C^{\dag},\label{eq:condC1}\\ (Z_0)_k &= C (Z_0)_p C^{\dag}, \label{eq:condC2}\\ (M_{+-}^2)_k &= C (M_{+-}^2)_p C^{\dag},\label{eq:condC3} \end{align} where $(\cdot)_p$ denotes the projection to $\mathcal W_p$, if $p\neq k$. \end{itemize} \noindent\emph{Proof of (1)}. We prove the statement by showing the validity of a sufficient condition: all states \eqref{eq:states} are mapped into one-another by elements of the algebra $\mathcal K$. This condition is sufficient because the elements of $\mathcal B_k$ are a subset of the states \eqref{eq:states}. We begin by proving that one can map $\ket{1}_k$ into every state \eqref{eq:states}. First, we note that using \begin{equation} S_{\pm} \ket{n}_{k} = n \ket{n\pm1}_{k} \label{eq:Spmrel} \end{equation} we can map $\ket{1}_{k}$ to all states $\ket{n}_{k}$ with $n\in\{2,\ldots,d-1\}$ and, for $k=0$, also to $\ket{0}^{\otimes 2 t}$. Next, we observe that using \begin{equation} R_{+,1} \ket{n}_{k} = n \ket{n, 1}_{2,k}\label{eq:Rprel} \end{equation} and then repeatedly applying \begin{equation} S_{+} \ket{n,m}_{k} = n \ket{n+1, m}_{2,k}+ m \ket{n, m+1}_{2,k} \end{equation} we can map $\ket{1}_{k}$ into every state $\ket{n, m}_{2,k}$ with $n,m\in\{1,\ldots,d-1\}$. We proceed using an inductive argument. Assuming that we can access every state $\ket{m, {\boldsymbol \nu}, m}_{\ell', k}$ of length $\ell'<\ell$ we shall prove that we can access every state of length $\ell$, for $\ell \geq 3$. This follows straightforwardly from the relations \begin{align} R_{+,1} \ket{n, {\boldsymbol\nu}, m}_{\ell-1, k} &\simeq m \ket{n, {\boldsymbol\nu}, m, 1}_{\ell, k}\,,\label{eq:indstep1}\\ M_{-+} \ket{{n, {\boldsymbol\nu}, 1}}_{\ell-1,k} &\simeq \ket{n,{\boldsymbol\nu},0,1}_{\ell, k} \,,\label{eq:indstep2} \end{align} where $\simeq$ denotes equality up to states of length $<\ell$ which can be accessed by assumption, and the repeated application of \begin{align} S_{+} \ket{n, {\boldsymbol\nu}, m, m_2}_{\ell, k} &\simeq m_2 \ket{n, {\boldsymbol\nu}, m, m_2+1}_{\ell, k}\,,\label{eq:fe}\\ S_{+} \ket{n, {\boldsymbol\nu}, 0, m_2}_{\ell, k} &\simeq m_2 \ket{n, {\boldsymbol\nu}, 0, m_2+1}_{\ell, k}\,. \label{eq:se} \end{align} In Eq.~\eqref{eq:fe}, $\simeq$ denotes equality up to states of the form $\ket{n', {\boldsymbol\nu}', m', m_2}_{\ell, k}$ that are accessed at the previous step ($n$, $\boldsymbol \nu$, and $m$ are arbitrary). Analogously in Eq.~\eqref{eq:se}, $\simeq$ denotes equality up to states of the form $\ket{n', {\boldsymbol\nu}', 0, m_2}_{\ell, k}$. This means that for every state $\ket{n, {\boldsymbol\nu}, m}_{\ell, k}$ there exist an operator $B_{n, \boldsymbol \nu, m}\in\mathcal K$ such that \begin{equation} B_{n, \boldsymbol \nu, m}\ket{1}_{k}= \ket{n, {\boldsymbol\nu}, m}_{\ell, k}\,. \end{equation} Then we can construct an operator mapping the arbitrary vector $\ket{n', {\boldsymbol \nu}', m'}_{\ell', k}$ into the arbitrary vector $\ket{n, {\boldsymbol \nu}, m}_{\ell, k}$ \begin{equation} A_{n, \boldsymbol \nu, m; n', \boldsymbol \nu', m'}=B^{\phantom{\dag}}_{n, \boldsymbol \nu, m} \ket{1}_{\!k\,\,k}\!\!\bra{1} B^\dag_{n', \boldsymbol \nu', m'}\,. \end{equation} To prove that this operator is in $\mathcal K$ we fist note that, since the generators are Hermitian, we have that if ${B^{\phantom{\dag}}_{n, \boldsymbol \nu, m}\in\mathcal K}$ also ${B^\dag_{n, \boldsymbol \nu, m}\in\mathcal K}$. We then just need to prove that $\ket{1}_{k}\prescript{}{k}{\bra{1}}\in\mathcal K$. This is explicitly done by observing \begin{align} &\ket{1}_{k} \prescript{}{k}{\bra{1}}= \frac{(S_{-})^{d-2} (S_{+})^{d-2} T_{-,d-1}^{2t-1} R_{+,d-1}^{2t-1}}{\prescript{}{k}{\braket{1|(S_{-})^{d-2} (S_{+})^{d-2} T_{-,d-1}^{2t-1} R_{+,d-1}^{2t-1}|1}_k}}\,, & &k=0 \lor d\neq 2\,,\\ &\ket{1}_{k} \prescript{}{k}{\bra{1}}= \frac{T_{-,1}^{2t-2} R_{+,1}^{2t-2}}{\prescript{}{k}{\braket{1|T_{-,1}^{2t-2} R_{+,1}^{2t-2}|1}_k}}\,, & &k\neq 0 \wedge d= 2\,. \end{align} \noindent\emph{Proof of (2)}. This point is immediate. Indeed, one can directly verify that \begin{equation} (Z_{0}-Z_{1}) \ket{1}_{k} = \ket{1}_{k+t}\,. \end{equation} \noindent\emph{Proof of (3)}. The statement is trivial whenever $k$ or $p$ are 0. Indeed, $\ket{0}^{\otimes 2t}$ is the only eigenstate of $Z_0+Z_1$ corresponding to eigenvalue $2t$ and does not appear for $p\neq0$. This means that the sum of \eqref{eq:condC1} and \eqref{eq:condC2} can never be satisfied. To prove (3) for $k,p\neq0$ we note that \begin{equation} \ket{1}^{(0)}_{k} : = \frac{\ket{1}_{k}+\ket{1}_{k+t}}{\sqrt 2}\,, \end{equation} is the only eigenstate of $Z_{1}$ and $Z_{0}$ with eigenvalues $t$ and $t-2$ respectively. This means that \eqref{eq:condC1} and \eqref{eq:condC2} can be fulfilled only if $\ket{1}^{(0)}_{k}$ is an eigenstate of $C$. In turn, this implies that \begin{equation} \tensor*[^{(0)}_{k}]{\braket{1 |(M_{+-})^2| 1}}{_k^{(0)}} = e^{-2i \pi k/t}\,, \end{equation} is invariant under the mapping implemented by $C$. Since $e^{-2i \pi k/t}\neq e^{-2i \pi p/t}$ for $k\neq p \in \{1,\ldots, t-1\}$ we conclude that there can be no transformation $C$ fulfilling \eqref{eq:condC2}. \subsection{Proof of Lemma \ref{lemma3}} We consider $d=2$, where $s_a = \frac{1}{2}\sigma_a$, $a\in\{1,2,3\}$, and begin by writing the analogue of \eqref{eq:intermediatecond}. To this aim we plug the form \eqref{eq:parametrisationgen} in the definitions (\ref{eq:Utildee}, \ref{eq:Utildeo}) and use the constraints (\ref{eq:TIEc1}, \ref{eq:urxT}) to find \begin{align} &\tilde{\mathbb U} = e^{i \theta} e^{i \boldsymbol{\alpha} \cdot \boldsymbol M}\,\tilde{\mathbb V}\, e^{i \boldsymbol\beta\cdot\boldsymbol M}, & &\tilde{\mathbb W} = e^{i \theta'} e^{i \boldsymbol\gamma\cdot\boldsymbol M}\, \tilde{\mathbb V}'\, e^{i \boldsymbol\delta\cdot \boldsymbol M}\!, \label{eq:UTIE} \end{align} where $\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\gamma},\boldsymbol{\delta}\in\mathbb R^3$, $\theta,\theta'\in\mathbb R$, $\boldsymbol M = (M_1,M_2,M_3)$, while $\tilde{\mathbb V}$ and $\tilde{\mathbb V}'$ are defined in \eqref{eq:genmag}. Substituting now \eqref{eq:UTIE} into the first condition \eqref{eq:CondSU2T} and using the second condition to commute $\boldsymbol\alpha\cdot\boldsymbol M$ around $A$ we find the desired analogue of \eqref{eq:intermediatecond} \begin{align} \tilde{\mathbb V}\,\tilde{\mathbb V}^{\prime} A \tilde{\mathbb V}^{\prime \dag} \tilde{\mathbb V}^\dag &\!\!=e^{i \phi} A, & [M_a,A] &\!=\!0, & [\tilde{\mathbb V}^{\prime \dag} M_a \tilde{\mathbb V}^{\prime},A]&=0, & a&\in\{1,2,3\}\,. \label{eq:intermediatecondTIE} \end{align} Next we consider \begin{align} \tilde{\mathbb V}^{\prime \dag} M_1 \tilde{\mathbb V}' &=\tilde{\mathbb V}^{\prime \dag} M_{1,1} \tilde{\mathbb V}' +\tilde{\mathbb V}^{\prime \dag} M_{1,0} \tilde{\mathbb V}' \notag\\ &= \sin(2J') (M_{32,1}+M_{23,1}) + \cos(2J') M_{1},\\ \tilde{\mathbb V} M_1 \tilde{\mathbb V}^\dag &= \tilde{\mathbb V} M_{1,1} \tilde{\mathbb V}^\dag+\tilde{\mathbb V} M_{1,0} \tilde{\mathbb V}^\dag \notag\\ &= - \sin(2J) (M_{32,0}+M_{23,0}) + \cos(2J) M_{1}\,. \end{align} where we used \eqref{eq:VMVe} and \eqref{eq:VMVo} specialised to the case $d=2$. From these relations we see that, for $J, J'\neq 0$, $A$ commutes with $\{M_{32, \iota}+M_{23, \iota}\}_{\iota=0,1}$. Using the second of \eqref{eq:intermediatecondTIE} to make generic ${\rm SU}(2)$ rotations we find that actually $A$ commutes with the following 10 operators \begin{align} \{M_{12, \iota}+M_{21, \iota},&M_{13, \iota}+M_{31, \iota},\notag\\ &M_{23, \iota}+M_{32, \iota}, M_{11, \iota}-M_{22, \iota},M_{11, \iota}-M_{33, \iota}\}_{\iota=0,1}. \label{eq:MsymTIE} \end{align} In fact, $A$ commutes also with $\{M_{aa,\iota}\}$. To show that we consider the following objects \begin{align} &P_{\bar \iota} (P_{\iota}P_{\bar \iota})^{\frac{t}{2}-1} (M_{11,\iota}-M_{33,\iota}) ({P_{ \iota}P_{\bar \iota}})^{1-\frac{t}{2}} P^{-1}_{\bar \iota}\notag\\ &= - S_3 M_{11,\iota} -M_{33,\iota}, \qquad\qquad\qquad\quad t\,\,\text{even}\label{eq:stringteven}\\ \notag\\ &(P_{\iota}P_{\bar \iota})^{\frac{t-1}{2}} (M_{11,\iota}-M_{33,\iota}) (P_{\iota}P_{\bar \iota})^{-\frac{t-1}{2}} \notag\\ &= - S_3 M_{11,\bar \iota} + M_{33,\bar \iota}, \qquad\qquad\qquad \quad t\,\,\text{odd}\label{eq:stringtodd} \end{align} for $\iota\in\{0,1\}$, $\bar{\iota}:=1-\iota$, and we defined \begin{equation} S_a = \sigma^{\otimes 2t}_a= i^{2t} e^{i \frac{\pi}{2} M_a}, \qquad\qquad P_{\iota} = e^{i \frac{\pi}{4} (M_{11,\iota}-M_{22,\iota})}\,. \end{equation} The left hand sides of \eqref{eq:stringteven} and \eqref{eq:stringtodd} commute with $A$ by construction, therefore we find \begin{align} &[A, S_3 M_{11, \iota} +M_{33, \iota}]=0, & & t\,\,\text{even}\label{eq:stringteven1}\\ & [A, S_3 M_{11, \iota} - M_{33, \iota}]=0, & & t\,\,\text{odd}\label{eq:stringtodd1}. \end{align} Using that $A$ commutes with $\{S_a\}_{a=1,2,3}$ and with the operators \eqref{eq:MsymTIE} we then have \begin{align} &[A, (S_a+\mathbbm{1}_t) M_{bb, \iota}]=(S_a+\mathbbm{1}_t) [A, M_{bb, \iota}]=0, & & t\,\,\text{even}\label{eq:stringteven2}\\ & [A, (S_a-\mathbbm{1}_t) M_{bb, \iota} ]=(S_a-\mathbbm{1}_t) [A, M_{bb, \iota}]=0, & & t\,\,\text{odd}\,.\label{eq:stringtodd2} \end{align} At this point we observe \begin{equation} {\rm spect}(S_1+S_2+S_3)= \{(-1)^{t} 3,(-1)^{t+1}\}\,, \end{equation} where the multiplicities (both algebraic and geometrical) of the two eigenvalues are $2^{2t-2}$ and $3 \cdot 2^{2t-2}$ respectively. Thus, by summing (\ref{eq:stringteven2}, \ref{eq:stringtodd2}) over $a$ we obtain that the commutators are multiplied by invertible operators. Therefore, we finally arrive at \begin{equation} [A, M_{bb, \iota}]=0\,,\qquad b\in\{1,2,3\}\,. \label{eq:AdoubleMevenTsym} \end{equation} Putting it all together we find \begin{equation} [A, M_{ab, \iota} +R_{2t} M_{ab, \iota} R_{2t} ]=0\,. \end{equation} At this point, considering $\tilde{\mathbb V}\,\tilde{\mathbb V}'$ and using (\eqref{eq:VMMMd}, \eqref{eq:V'MMMd}) we have \begin{align} \tilde{\mathbb V}\,\tilde{\mathbb V}^{\prime} A \tilde{\mathbb V}^{\prime \dag} \tilde{\mathbb V}^\dag = A\, \end{align} so that the first of conditions \eqref{eq:intermediatecondTIE} can be fulfilled only for $\phi=0$, and in that case it follows from \eqref{eq:AdoubleMevenTsym}, and the second of \eqref{eq:intermediatecondTIE}. This proves the Lemma. \subsection{Proof of Lemma \ref{lemma4}} To prove the statement we note that the set of operators $\{R_{2t}^n\Pi_{2t}^{2\tau}\}^{n=0,1}_{\tau=0,1,\ldots t-1}$ can be written as \begin{equation} \Pi_{2t}^{2\tau} = \sum_{k=0}^{t-1} e^{-2\pi i \tau k/t} Q_k, \qquad\qquad R_{2t} \Pi_{2t}^{2\tau}= \sum_{k=0}^{t-1} e^{-2\pi i \tau k/t} Q'_k, \end{equation} where $Q_k$ are the orthogonal projectors defined in Eq.~\eqref{eq:Zk} and we introduced \begin{equation} Q'_k := \frac{1}{t}\sum_{\tau=0}^{t-1} e^{2\pi i \tau k/t} R_{2t} \Pi_{2t}^{2\tau}\,, \qquad k\in\{0,\ldots,t-1\}. \label{eq:Z'k} \end{equation} Since the mapping between $\{R_{2t}^n\Pi_{2t}^{2\tau}\}^{n=0,1}_{\tau=0,1,\ldots t-1}$ and $\{Q_k, Q'_k\}_{k=0,1,\ldots t-1}$ is invertible it is sufficient to prove that the latter operators are linearly independent. To this aim we note that $\{Q_k, Q'_k\}_{k=1,\ldots t-1}$ are obviously linearly independent. This can be seen by writing them in a basis of eigenstates of $\Pi_{2t}$ and noting distinct operators are non zero on distinct, non-overlapping, blocks. Moreover, noting that all $\{Q_k, Q'_k\}_{k=1,\ldots t-1}$ are zero when reduced to the block of zero two-momentum (i.e. the one composed by eigenstates of $\Pi_{2t}$ with eigenvalues $1$ and $-1$), we have that the only two operators which can be linearly dependent are $Q_0$ and $Q'_0$. To prove that such operators are independent we note that \begin{equation} Q_0 \ket{n}_0 = Q'_0 \ket{n}_0 = \ket{n}_0\,,\quad Q_0 \ket{n}_t = -Q'_0 \ket{n}_t = \ket{n}_t\,,\quad n\in\mathcal I\,, \end{equation} where $\ket{n}_0, \ket{n}_t$ are defined in Eq.~\eqref{eq:states} and the set $\mathcal I$ in \eqref{eq:setI}. This implies that \begin{equation} \alpha Q_0+\beta Q'_0 =0 \end{equation} only if $\alpha=\beta=0$ and concludes the proof. \section{Discussion of the results and their possible extensions} \label{sec:discussion} In this section we discuss some generalisations and extensions of our results. While the extension to inhomogeneous interactions in Sec.~\ref{Sec:inhom} is rigorous, the other two subsections discussing fluctuations of SFF (\ref{sec:fluctuations}) and singular disorder distributions (\ref{sec:nonisotropic}) are currently of speculative nature. \subsection{Spatially inhomogeneous interactions} \label{Sec:inhom} The space-time duality approach adopted in this paper treats separately each point in space and it is therefore convenient to study general inhomogeneous interactions. However, our results as elaborated in Sec.~\ref{sec:main} are not directly applicable to this case. Here we explain how to extend them focussing for simplicity on the case of no time reversal symmetry. We begin by observing that for position-dependent local gates, Eq.~\eqref{eq:SFFduality} is substituted with \begin{equation} K(t,L) = {\rm tr}\left( \mathbb T_1 \mathbb T_2 \cdots \mathbb T_L\right)\,, \label{eq:inhomoK} \end{equation} where $\mathbb T_x$ is defined as in Eq.~\eqref{eq:SFFTM2} with $\mathbb O$ given in Eq.~\eqref{eq:defOa} while $\tilde{\mathbb U}$ and $\tilde{\mathbb W}$ are replaced by \begin{align} \tilde{\mathbb W}_x &:= \prod_{\tau \in \mathbb Z_{t}} \eta_{\tau,t}(\tilde W_{x+\frac{1}{2}}),\\ \tilde{\mathbb U}_x &:= \!\prod_{\tau \in \mathbb Z_{t}+\frac{1}{2}}\!\eta_{\tau,t}(\tilde U_x). \end{align} Then we choose dual-unitary gates $U_x,W_x$ of the form \eqref{eq:parametrisationgen}, assuming $J_x, J_x' \neq 0$ for all $x$. We use Lemma~\ref{lemma1}, Lemma~\ref{lemma2}, and Theorem~\ref{theorem1} to find \begin{equation} K(t,L) = t + {\rm tr}\left( \mathbb R_1 \mathbb R_2 \cdots \mathbb R_L\right)\,, \label{eq:inhomoK2} \end{equation} where $\mathbb R_{x}= (\mathbbm{1}-\mathbb P) \mathbb T_x (\mathbbm{1}-\mathbb P)$ and $\mathbb P$ is the projector onto the eigenspace of $\mathbb T_x$ associated to the eigenvalue 1. In writing \eqref{eq:inhomoK2} we used that, due to Lemma~\ref{lemma1}, Lemma~\ref{lemma2}, and Theorem~\ref{theorem1}, such eigenspace is the same for all $x$. From \eqref{eq:inhomoK2} we see that to recover the result \eqref{eq:mainresult} we need to show that \begin{equation} \lim_{L\to\infty} {\rm tr}\left( \mathbb R_1 \mathbb R_2 \cdots \mathbb R_L\right) = 0\,. \label{eq:dream} \end{equation} For example this would hold if $\| \mathbb R_{x}\| <1$, where $\|\cdot\|$ denotes the operator norm. The results of Sec.~\ref{sec:main}, however, are not sufficient to infer \eqref{eq:dream}. To overcome this problem, we focus on the case of $L$ even and make the following different replacement \begin{equation} \mathbb T_{x-1} \mathbb T_x = \mathbb P + \tilde{\mathbb R}_{x-1, x} \quad \Rightarrow \quad K(t,L) = t + {\rm tr}(\tilde{\mathbb R}_{1,2} \tilde{\mathbb R}_{3,4} \cdots \tilde{\mathbb R}_{L-1,L})\,. \label{eq:replacement} \end{equation} where \begin{equation} \tilde{\mathbb R}_{x-1, x}= (\mathbbm{1}-\mathbb P) \mathbb T_{x-1} \mathbb T_x (\mathbbm{1}-\mathbb P). \end{equation} For the new operator $\tilde{\mathbb R}_{x-1, x}$ we are able to prove the following theorem (the proof is provided at the end of the subsection): \begin{theorem} \label{theorem2} For $J_x,J'_x \neq0$, \begin{equation} \|\tilde{\mathbb R}_{x-1, x}\| < 1. \end{equation} \end{theorem} This means that if there is a {\em finite density} of points $x$ with nonzero couplings ($J_x,J'_x\neq 0$), i.e. finite density of non-SWAP gates, then \begin{equation} \lim_{L\to\infty} {\rm tr}\left( \tilde{\mathbb R}_{1,2} \cdots \tilde{\mathbb R}_{L-1,L}\right) = 0\,, \label{eq:dream2} \end{equation} and \begin{equation} \lim_{L\to\infty} K(t,L) = {\rm tr}\,\mathbb P = t\,. \end{equation} A completely analogous treatment holds for $L$ odd, e.g., considering \begin{equation} K(t,L) = t + {\rm tr}(\tilde{\mathbb R}_{12} \cdots \tilde{\mathbb R}_{L-2,L-1}{\mathbb R}_{L})\,. \end{equation} \subsubsection{Proof of Theorem~\ref{theorem2}} To prove the statement it is sufficient to show that if \begin{equation} \braket{A| (\mathbb T_{x-1} \mathbb T_x)^\dag \mathbb T_{x-1} \mathbb T_x|A}=1 \end{equation} for some state $\ket{A}$, then \begin{equation} \mathbb P \ket{A}= \ket{A}\,. \end{equation} This follows immediately from Theorem~\ref{theorem1} and the following two lemmas \setcounter{lemma}{5} \begin{lemma} \label{lemma6} For dual-unitary circuits, if a state $\ket{A}$ fulfils \begin{equation} \braket{A|(\mathbb T_0 \mathbb T)^\dag\mathbb T_0 \mathbb T|A}=1 \label{eq:EV} \end{equation} where $\mathbb T_0$ and $\mathbb T$ are transfer matrices of the form \eqref{eq:SFFTM2} (with unitary matrices $\tilde{\mathbb U}_0$ and $\tilde{\mathbb W}_0$ and $\tilde{\mathbb U}$ and $\tilde{\mathbb W}$ respectively), then \begin{align} & ( M_{a, \iota}\otimes \mathbbm{1}_{2t}-\mathbbm{1}_{2t}\otimes M_{a, \iota}^*) (\tilde{\mathbb W}_0 \otimes \tilde{\mathbb W}^*_0) (\tilde{\mathbb U} \otimes \tilde{\mathbb U}^*) (\tilde{\mathbb W} \otimes \tilde{\mathbb W}^*)\ket{A}=0\,,\label{eq:conditionsstateA1} \\ & ( M_{a, \iota}\otimes \mathbbm{1}_{2t}-\mathbbm{1}_{2t}\otimes M_{a, \iota}^*) (\tilde{\mathbb U} \otimes \tilde{\mathbb U}^*) (\tilde{\mathbb W} \otimes \tilde{\mathbb W}^*)\ket{A}=0\,,\label{eq:conditionsstateA2}\\ & ( M_{a, \iota}\otimes \mathbbm{1}_{2t}-\mathbbm{1}_{2t}\otimes M_{a, \iota}^*) (\tilde{\mathbb W} \otimes \tilde{\mathbb W}^*) \ket{A} = 0\,,\label{eq:conditionsstateA3}\\ &( M_{a, \iota}\otimes \mathbbm{1}_{2t}-\mathbbm{1}_{2t}\otimes M_{a, \iota}^*)\ket{A} = 0\,, \label{eq:conditionsstateA4} \end{align} where $\iota\in\{0,1\}$, $a\in\{1,2,\ldots,d^2-1\}$. \end{lemma} \begin{lemma} \label{lemma7} For $\tilde{\mathbb U}$ and $\tilde{\mathbb W}$ of the form \eqref{eq:UGE} with $J\neq 0$ and $J'\neq 0$ the conditions \eqref{eq:conditionsstateA2}, \eqref{eq:conditionsstateA3}, and \eqref{eq:conditionsstateA4} are equivalent to \begin{equation} [A, M_{a,\iota}]=0\,, \quad [A, M_{ab,\iota}]=0\,, \quad a,b\in\{1,2,\ldots,d^2-1\},\quad \iota\in\{0, 1\}\,. \end{equation} \end{lemma} \subsubsection{Proof of Lemma \ref{lemma6}} \label{sec:prooflemma6} Considering the expectation value \eqref{eq:EV}, and using that \begin{equation} \bra{B}\mathbb O^\dagger_{\iota'} \mathbb O^{\phantom{\dag}}_{\iota'}\ket{B} \le \langle B|B\rangle\,,\qquad \bra{B}\mathbb O^{\phantom{\dag}}_{\iota'} \mathbb O^\dagger_{\iota'}\ket{B} \le \langle B|B\rangle\,, \end{equation} for any $\ket{B}$, we have \begin{eqnarray} 1 &=& \braket{A|(\mathbb T_0 \mathbb T)^\dag\mathbb T_0 \mathbb T |A}\nonumber\\ &=& \| \mathbb O_1^{{\dag}} (\tilde{\mathbb W}_0 \otimes \tilde{\mathbb W}^*_0) \mathbb O_0 (\tilde{\mathbb U} \otimes \tilde{\mathbb U}^*) \mathbb O_1^{{\dag}} (\tilde{\mathbb W} \otimes \tilde{\mathbb W}^*) \mathbb O_0 |A \rangle\|^2 \label{eq:boundxinho} \\ &\le & \| \mathbb O_1^{{\dag}} (\tilde{\mathbb W}_0 \otimes \tilde{\mathbb W}^*_0) \mathbb O_0 (\tilde{\mathbb U} \otimes \tilde{\mathbb U}^*) \mathbb O_1^{{\dag}} (\tilde{\mathbb W} \otimes \tilde{\mathbb W}^*) |A \rangle\|^2 \nonumber \\ &\le & \| \mathbb O_1^{{\dag}} (\tilde{\mathbb W}_0 \otimes \tilde{\mathbb W}^*_0) \mathbb O_0 (\tilde{\mathbb U} \otimes \tilde{\mathbb U}^*) (\tilde{\mathbb W} \otimes \tilde{\mathbb W}^*) |A \rangle\|^2 \nonumber \\ &\le & \| \mathbb O_1^{{\dag}} (\tilde{\mathbb W}_0 \otimes \tilde{\mathbb W}^*_0) (\tilde{\mathbb U} \otimes \tilde{\mathbb U}^*) (\tilde{\mathbb W} \otimes \tilde{\mathbb W}^*) |A \rangle\|^2 \nonumber \\ &\le & \| (\tilde{\mathbb W}_0 \otimes \tilde{\mathbb W}^*_0) (\tilde{\mathbb U} \otimes \tilde{\mathbb U}^*) (\tilde{\mathbb W} \otimes \tilde{\mathbb W}^*) |A \rangle\|^2 = \| | A \rangle\|^2 =1\,. \nonumber \end{eqnarray} This can hold only if all four inequalities in (\ref{eq:boundxinho}) are saturated. Using \eqref{eq:unique} we see that this happens only if the conditions (\ref{eq:conditionsstateA1}\,--\, \ref{eq:conditionsstateA4}) are satisfied. \subsubsection{Proof of Lemma \ref{lemma7}} Plugging the forms \eqref{eq:UGE} we bring (\ref{eq:conditionsstateA2}\,--\,\ref{eq:conditionsstateA4}) in the form \begin{align} [M_{a,\iota},A] &\!=\!0, & [\tilde{\mathbb V}^{\prime \dag} M_{a,\iota} \tilde{\mathbb V}^{\prime},A]&=0, & [\tilde{\mathbb V}^{\prime \dag} \tilde{\mathbb V}^\dag M_{a,\iota} \tilde{\mathbb V} \tilde{\mathbb V}^{\prime},A]&=0, \label{eq:intermediatecondinho} \end{align} where $a\in\{1,2,\ldots,d^2-1\}$, $\iota\in\{0,1\}$. As shown in the proof of Lemma \ref{lemma2} we have that the first two relations \eqref{eq:intermediatecondinho} imply that for $J'\neq 0$: \begin{equation} [A,M_{ab,1}]=0,\qquad a,b\in\{1,2,\ldots,d^2-1\}. \label{eq:AdoubleModdinho} \end{equation} Using the above relation and Eq.~\eqref{eq:V'MMMd} we have that the third of \eqref{eq:intermediatecondinho} is equivalent to \begin{align} [\tilde{\mathbb V}^\dag M_{a,\iota} \tilde{\mathbb V},A]=0. \end{align} Proceeding as in~(\ref{eq:VMVo}, \ref{eq:AdoubleMeven}) we then find that for $J\neq0$: \begin{equation} [A,M_{ab,0}]=0,\qquad a,b\in\{1,2,\ldots, d^2-1\}. \label{eq:AdoubleMeveninho} \end{equation} \subsection{Fluctuations of the spectral form factor} \label{sec:fluctuations} The approach presented in the current manuscript can also be applied to study the \emph{fluctuations} of the SFF~\eqref{eq:SFF} (this idea has recently been exploited in Ref.~\cite{Fluctuations} in the special case of the self-dual kicked Ising model and in Ref.~\cite{largeq} for the large $d$ asymptotics of Floquet chains with Haar random interactions). Specifically, it can be employed to compute the higher moments of $|{\rm tr}\, \mathbb U_{L}^t|^{2}$ with respect to the {\em i.i.d.} on-site disorder distribution \begin{equation} K_n(t,L) := \mathbb E\left[ |{\rm tr}\, \mathbb U_{L}^t|^{2n}\right] = \mathbb E\left[\left(\sum_{i,j=1}^{\mathcal N} e^{i (\varphi_{i}-\varphi_{j}) t}\right)^{\!\!\!n\;}\right]\!\!,\,\,\, t,L\in\mathbb N,\,\,\, n\geq1. \label{eq:nSFF} \end{equation} Indeed, exploiting the space-time duality described in Sec.~\ref{sec:duality} we can express the above quantities as \begin{equation} K_n(t,L) = {\rm tr}\,\mathbb T^L_n, \end{equation} with \begin{equation} \mathbb T_n = (\tilde{\mathbb U}^{\otimes n} \otimes (\tilde{\mathbb U}^*)^{\otimes n}) \mathbb O_{1;n}^\dag (\tilde{\mathbb W}^{\otimes n} \otimes (\tilde{\mathbb W}^*)^{\otimes n}) \mathbb O^{\phantom{\dag}}_{0;n}, \label{eq:SFFTM2n} \end{equation} where $\tilde{\mathbb U}$ and $\tilde{\mathbb W}$ are defined in (\ref{eq:Utildee}, \ref{eq:Utildeo}) while we introduced \begin{align} \mathbb O_{\iota';n} &:= \mathbb O_{0\iota';n} \mathbb O_{1\iota';n}= \mathbb O_{1\iota';n} \mathbb O_{0\iota';n}\,,\\ \mathbb O_{\iota\iota';n} &:= \int\!{\rm d}^{d^2-1}\boldsymbol{\theta}\,g_{\iota\iota'}(\boldsymbol{\theta}) \exp\left(i\boldsymbol\theta\cdot (\boldsymbol{M}_{\iota;n}^{\phantom{*}}\otimes \mathbbm{1}_{2t}-\mathbbm{1}_{2t}\otimes \boldsymbol{M}^*_{\iota;n})\right)\,. \end{align} and finally $\boldsymbol{M}_{\iota;n} = (M_{1,\iota;n},M_{2,\iota;n},\ldots, M_{d^2-1,\iota;n})$ with \begin{equation} M_{a, \iota;n} := \sum_{j=0}^{n-1} \mathbbm{1}_t^{\otimes j}\otimes\! M_{a, \iota}\!\otimes \mathbbm{1}_t^{\otimes (n-1-j)},\quad a\in\{1,\ldots,d^2-1\},\,\, \iota\in\{0,1\}, \end{equation} where the generalised sublattice magnetisation $M_{a, \iota}$ is defined in Eq.~\eqref{eq:sublatticeM}. Note that, for definiteness, here we considered systems without time-reversal symmetry. Using that the matrix $\mathbb T_n$ has the same structure as $\mathbb T$ in \eqref{eq:SFFTM2} we can directly repeat the treatment described in Sec.~\ref{sec:leadingeigs}. In particular, for local gates of the form \eqref{eq:parametrisationgen} we find \begin{equation} \lim_{L\to\infty} K_n(t,L) = {\rm dim}\, \mathcal M_n', \end{equation} where we introduced the set \begin{equation} \mathcal M_n:= \{ M_{a,\iota; n} \}_{a,\iota} \cup \{ M_{ab,\iota;n} \}_{a,b,\iota}\,, \label{eq:finalsetn} \end{equation} and the coproduct of $n$ double magnetizations (cf. \eqref{eq:doubleM}) \begin{equation} M_{a b,\iota; n}:= \!\sum_{j=1}^n\! \mathbbm{1}_t^{\otimes (j-1)}\otimes\! M_{ab, \iota}\!\otimes \mathbbm{1}_t^{\otimes (n-j)},\,\, a,b\in\{1,\ldots,d^2-1\},\,\iota\in\{0,1\}\,. \label{eq:doubleMn} \end{equation} All elements of \eqref{eq:finalsetn} are invariant under permutations of the $n$ copies of $\mathcal H_t$ in $\mathcal H_t^{\otimes n}$ and under 2-site translations within each copy, i.e., they commute with \begin{align} A_{p; \tau_1, \ldots \tau_n} &= \Gamma(p) (\Pi_{2t}^{2 \tau_1} \otimes\cdots\otimes \Pi_{2t}^{2 \tau_n}), \quad \tau_1,\ldots, \tau_n\in \{0,\ldots,t-1\}\,, \end{align} where $\Gamma(\cdot)$ is a representation of $S_n$, the symmetric group of $n$ letters, on $\mathcal H_{t n}\equiv \mathcal H_t^{\otimes n}$ and $p\in S_n$. Specifically, \begin{equation} \Gamma(p) \ket{A_1}\otimes\ket{A_2}\otimes\cdots\otimes \ket{A_n} = \ket{A_{p(1)}}\otimes\ket{A_{p(2)}}\otimes\cdots\otimes \ket{A_{p(n)}}. \end{equation} This means that \begin{equation} \mathcal A_n = {\rm span}\{ A_{p; \tau_1,\tau_2 \ldots \tau_n};\,\, \tau_1,\tau_2 \ldots, \tau_n\in \{0,\ldots,t-1\},\, p\in S_n\} \end{equation} is a vector subspace of $\mathcal M_n'$ and hence \begin{equation} \lim_{L\to\infty} K_n(t,L)={\rm dim}\, \mathcal M_n'\geq n!\cdot t^n = \lim_{\mathcal N \to\infty} \int |{\rm tr}\,\mathbb{U}^t|^{2n} {\rm d}\mu_{\rm CUE}(\mathbb{U}), \label{eq:fluctuations} \end{equation} where ${\rm d}\mu_{\rm CUE}(\mathbb{U})$ is the CUE measure and $\mathcal N$ is the dimension of the matrix $\mathbb{U}$. In analogy with what happens for $n=1$ (c.f. Theorem~\ref{theorem1}) we expect $\mathcal A_n$ and $\mathcal M_n'$ to actually coincide, leading to an equality sign in \eqref{eq:fluctuations}. However, we leave the formal proof of this statement to future work. Similar conclusions (with CUE replaced by COE) hold in the time-reversal invariant case. \subsection{Singular on-site disorder distributions} \label{sec:nonisotropic} As discussed in Sec.~\ref{sec:SFF} we expect our results to be stable under modifications of the averaging procedure as long as such modifications do not introduce spatial correlations. In our setting this can be verified explicitly by considering \emph{singular} disorder distributions of local gates $u_x,w_x$, Eq.~\eqref{eq:urx}, supported on lower-dimensional submanifolds of ${\rm SU}(d)$ that include the identity. For instance, one can imagine having some of the components of $\boldsymbol{\theta}_{\iota,x}$ in \eqref{eq:urx} (or $\boldsymbol{\theta}_{x}$ in~\eqref{eq:urxT}) set to zero for all $\iota$ and $x$. Physically, this choice describes a weaker external noise where, for example, the random magnetic fields are imposed only along certain specific directions rather than isotropically. In this case we expect that away from certain ``resonances", namely for almost all 2-site dual-unitary gates $U, W$, the treatment described in the previous sections is still applicable. In particular we anticipate that the SFF will still be characterised by the commutants of \eqref{eq:commutant} or \eqref{eq:commutantT} depending on whether or not the problem is time-reversal symmetric. Let us illustrate the main steps that can be used to prove this idea. We consider for simplicity the case of qubits ($d=2$) and absence of time-reversal symmetry. Denoting by $\mathcal I$ the subset of indices $\mathcal I\subset\{1,2,3\}$ such that $\{\theta_{a, \iota, x}\}_{a\in \mathcal I}$ are \emph{not} set to zero, and repeating the steps of Sec.~\ref{sec:duality} and \ref{sec:leadingeigs}, one readily finds the following analogue of the conditions \eqref{eq:CondSU2} \begin{equation} \begin{split} &\tilde{\mathbb U} \tilde{\mathbb W} A \tilde{\mathbb W}^\dag \tilde{\mathbb U}^\dag =e^{i \phi} A\,,\qquad [M_{a, \iota},A] = 0\,,\\ & [\tilde{\mathbb W}^\dag M_{a, \iota} \tilde{\mathbb W},A]=0\,,\qquad\qquad\qquad\qquad\qquad \iota\in\{0,1\}\,,\,\, a\in\mathcal I\,. \end{split} \label{eq:CondSU1} \end{equation} At this point we note that if $\mathcal I $ has at least two elements these conditions are equivalent to \eqref{eq:CondSU2}. This follows by observing that if $A$ commutes with $M_{a, \iota}$ and $M_{b, \iota}$ it also commutes with their commutator \begin{equation} [M_{a, \iota},M_{b, \iota}] = i \sum_{c=1}^3\epsilon_{abc} M_{c, \iota}\,. \end{equation} Here we used that $\{M_{a, \iota}\}$ generate the $\mathfrak{su}(2)$ algebra. A completely analogous reasoning applies for the set of equivalent matrices $\{\tilde{\mathbb W}^\dag M_{a, \iota} \tilde{\mathbb W}\}$. The only possible non-trivial choice is then to take the set $\mathcal I$ composed by single element, which, without loss of generality, can be set to $3$. This corresponds to $u_x,w_x$ being restricted to some ${\rm U}(1)$ subgroup of ${\rm SU}(2)$. To show that the our treatment applies also in this case we need to prove the analogue of Lemma~\ref{lemma2}. Namely, we need to show that if $A$ fulfils \eqref{eq:CondSU1} then it commutes with $\{M_{a,\iota}\}_{a=1,2,3; \iota=0,1}$, $\{M_{ab,\iota}\}_{a,b=1,2,3; \iota=0,1}$. This can be done by considering the following set of $15$ operators \begin{equation} \begin{split} \mathcal S_1 =&\{ \{M_{3, \iota}\}_{\iota=0,1}, \{[N_{\iota}, M_{3, \iota'}]\}_{\iota,\iota'=0,1}, \{[[N_\iota, M_{3, \iota'}], M_{3, \iota'}]\}_{\iota,\iota'=0,1}, \\ &\{N_{\iota}\}_{\iota=0,1}, [[N_0, M_{3, 0}], N_0], [[N_0, M_{3,1}], N_0], [[N_1, M_{3,0}], N_1 \} \, , \end{split} \label{eq:U1ops} \end{equation} where we introduced the short-hand notation \begin{equation} N_\iota:=\tilde{\mathbb W}^\dag M_{3, \iota} \tilde{\mathbb W}^{\phantom{\dag}}\!\!, \quad \iota=0,1. \end{equation} Since the operators \eqref{eq:U1ops} are constructed by taking commutators of $M_{3, \iota}$, and $\tilde{\mathbb W}^\dag M_{3, \iota} \tilde{\mathbb W}$, they commute with $A$. Moreover, they can be written as linear combinations of $\{M_{a,\iota}\}_{a=1,2,3; \iota=0,1}$ and $\{M_{ab,1}\}_{a,b=1,2,3}$. This means that if we can prove that the operators in $\mathcal S_1$ are \emph{linearly independent} we immediately have \begin{equation} [A, M_{a, \iota}]=0,\qquad [A, M_{ab, 1}]=0\,,\qquad a,b\in\{1,2,3\},\qquad \iota\in\{0,1\}\,. \end{equation} We could not prove explicitly the linear independence of the set \eqref{eq:U1ops} but we verified numerically that it holds almost always (away from special, measure-zero set of $U$, $W$ --- the so-called ``resonances''). An analogous reasoning considering the set of $15$ operators \begin{equation} \begin{split} \mathcal S_2 =&\{ \{[[\tilde N_\iota, M_{3, \iota'}], M_{3, \iota'}]\}_{\iota,\iota'=0,1}, \{\tilde N_{\iota}\}_{\iota=0,1}, [[\tilde N_0, M_{3, 0}], \tilde N_0],\\ & [[\tilde N_0, M_{3,1}], \tilde N_0], \{[\tilde N_{\iota}, M_{3, \iota'}]\}_{\iota,\iota'=0,1}, [[\tilde N_1, M_{3,0}], \tilde N_1] \} \, , \end{split} \label{eq:U1ops2} \end{equation} with \begin{equation} \tilde{N}_\iota:=\tilde{\mathbb U} \tilde{\mathbb W} \tilde{\mathbb W}^\dag M_{3, \iota} \tilde{\mathbb W} \tilde{\mathbb W}^\dag \tilde{\mathbb U}^\dag=\tilde{\mathbb U} M_{3, \iota} {\tilde{\mathbb U}}^\dag\,, \end{equation} leads to \begin{equation} [A, M_{a, \iota}]=0,\qquad [A, M_{ab, 0}]=0\,,\qquad a,b\in\{1,2,3\},\qquad \iota\in\{0,1\}\,. \end{equation} This means that \begin{equation} \lim_{L\to\infty} K(t,L) = \dim \mathcal M' = t, \end{equation} where in the second step we applied Theorem~\ref{theorem1}. \begin{acknowledgements} This work has been supported by the EU Horizon 2020 program through the European Research Council (ERC) Advanced Grant OMNES No. 694544, by the Slovenian Research Agency (ARRS) under the Programme P1-0402, and by the Royal Society through the University Research Fellowship No. 201101 (BB). BB acknowledges useful discussions with Andrea De Luca. \end{acknowledgements}

aafaaf97f109df6817c46fda1c41db7999f723eb

\section{Introduction} Online learning is a fundamental optimization paradigm employed in settings where one needs to make decisions in an uncertain environment. Such methods are essential for a range of practical applications: ad-serving~\citep{mcmahan2013ad}, dynamic pricing~\citep{LobelLV17,mao2018contextual}, or recommender systems~\citep{abernethy2007online} are only a few examples. These techniques are highly dependent on access to certain user data, such as search history, list of contacts, etc. which may expose sensitive information about a particular person. As these tools become ubiquitous on the internet, one can witness a surge in the collection of user data at massive scales. This is a tremendous problem, since by obtaining information about the behavior of algorithms run on these data, adversarial entities may learn potentially sensitive information; this could then be traced to a particular user, even if the users were anonymized to begin with~\citep{dwork2014algorithmic}. To mitigate the threat of diminishing user privacy, one can leverage the power of differential privacy~\citep{dwork2006calibrating}, a notion of privacy which ensures that the output of an algorithm is not sensitive to the presence of a particular user's data. Therefore, based on this output, one can not determine whether a user presents one or more given attributes. Differentially private learning algorithms have been studied in several settings, and a large number of recent works addressed the challenge of designing general optimization primitives with privacy guarantees~\citep{jain2012differentially, agarwal2017price, bassily2014private, bassily2014private2, abadi2016deep, wang2017differentially, iyengar2019towards}. In this paper, we further advance this line of research by offering differentially private algorithms for a very general task -- the bandit convex optimization problem in the case where the space of decisions that the learning algorithm can make exhibits complex geometry. Bandit convex optimization is an extremely general framework for online learning, which is motivated by the natural setting where, after making a decision the algorithm only learns the loss associated with its action, and nothing about other possible decisions it could have made (as opposed to the weaker \textit{full information} model where losses associated to all the possible decisions are revealed). Algorithms for this problem are highly dependent on the geometric properties of the space of decisions -- and their performance usually depends on the ability to perform certain projections onto this space~\citep{ben2001lectures,jaggi2013revisiting}. For large scale problems, this requirement may be prohibitive, as decisions may have to satisfy certain constraints (the set of recommendations must be diverse enough, or the set of ads to be displayed satisfy a given budget). Projection-free methods overcome this issue by exploiting the fact that some canonical decision sets often encountered in applications (matroid polytope, submodular base polytope, flow polytope, convex relaxations of low-rank matrices) have efficient linear optimization oracles. One can therefore use these efficient oracles in conjunction with the projection-free method to obtain algorithms that can be deployed for real-world applications. In this work we bridge the requirements of privacy and efficiency for online learning, building on the works of~\citep{garber2019improved, garber2013playing} to obtain the first differentially private algorithm for projection-free bandit optimization. To do so we leverage a generic framework for online convex optimization in the presence of noise, which we then adapt to our specific setting in a modular fashion. \medskip {\bf Our Contributions.} We give the first differentially private algorithm for the bandit convex optimization problem in the projection-free setting (we defer the definition of $(\epsilon,\delta)$-privacy to Definition~\ref{def:dp} and the problem statement to Section~\ref{sec:prelim}). We summarize the regret guarantees of our algorithm in the following theorem and compare it with the state of the art guarantees in the private and non-private settings. Our main focus is on the dependency on the dimension $n$ of the ambient space, the number $T$ of iterations, and the privacy budget $\epsilon$. For ease of comparison, we use the $\widetilde{O}$ notation to hide poly-logarithmic factors in $n$ and $T$, as well as parameters such as the Lipschitz constant of the loss functions. The precise guarantees can be found in Lemma~\ref{lem:reglap} (for $(\epsilon,0)$-privacy) and Lemma~\ref{lem:reggauss} (for $(\epsilon,\delta)$-privacy). \begin{thm}\label{thm:maininfo} Let $\D\subseteq \R^n$ be a convex domain for which we have access to a linear optimization oracle. Assume that for every $t \geq 1$, $f_t$ is convex and $L$-Lipschitz. Furthermore suppose that $\max_{\x,\y\in\D} \|x-y\| \leq D$. Then there exists an algorithm $\textsc{PrivateBandit}$ (Algorithm~\ref{alg:bandit}) which performs projection-free convex optimization in the bandit setting such that one of the following two properties holds: \begin{itemize}[leftmargin=*,noitemsep,topsep=0pt] \item the algorithm is $(\epsilon, 0)$-differentially private and, assuming $L=O(1)$ and $D=O(1)$, achieves an expected regret of $$\Reg_T = \widetilde{O}\bigg(\frac{T^{3/4}n^{3/2}} {\epsilon} \bigg) \,.$$ \item the algorithm is $(\epsilon, \delta)$-differentially private and, assuming $L=O(1)$ and $D=O(1)$, achieves an expected regret of $$\Reg_T = \widetilde{O}\bigg(\frac{(T^{3/4}n^{1/2}+T^{1/2}n) \log^{O(1)}(1/\delta) }{\epsilon} \bigg)\,,$$ whenever $\delta = 1/(n+T)^{O(1)}$. \end{itemize} \end{thm} In the non-private setting, the state of the art regret guarantee for projection-free bandit optimization is $\widetilde{O}(n^{1/2} T^{3/4})$ due to Garber and Kretzu~\citep{garber2019improved}\footnote{Their paper allows for a trade-off among parameters. This bound is optimized for the case when $T\gg n$.}. The regret guarantee of our algorithm matches the guarantee of ~\citep{garber2019improved} up to a $n/\epsilon$ factor in the $(\epsilon, 0)$ regime, and a $1/\epsilon$ factor in the $(\epsilon, \delta)$-regime, whenever $T \geq n^2$. Prior works in the private setting require projections to be available. The state of the art guarantees for private bandit optimization with projections are achieved by the work of Smith and Thakurta~\citep{thakurta2013nearly}. Smith and Thakurta focus on $(\epsilon,0)$-privacy and obtain a regret bound of $\widetilde{O}(n T^{3/4}/\epsilon)$. A variant of their algorithm can be used for $(\epsilon,\delta)$-privacy and obtains a regret bound of $\widetilde{O}(\sqrt{n} T^{3/4}/\epsilon)$. Our algorithm's guarantee matches the best guarantee with projections for $(\epsilon,\delta)$-privacy and is worse by a $\sqrt{n}$ factor for $(\epsilon,0)$-privacy. We leave it as an interesting open problem to improve the bound for $(\epsilon, 0)$-privacy to match the one using projections. \medskip {\bf Our Techniques.} In the process of obtaining our main result, we develop the common abstraction of noisy mirror descent to capture both online bandit optimization and private optimization (the \textsc{NoisyOCO} framework). This allows us to analyze the impact of the noise introduced to protect privacy on the regret of the online optimization. Once the framework is set up, one only needs to analyze the noise level to ensure the appropriate privacy guarantee and one immediately obtains the corresponding regret bound. However, analyzing the noise is in itself a non-trivial challenge. In the case of $(\epsilon,\delta)$-privacy, we give a strong concentration bound allowing us to match the privacy-regret trade-off achieved with projections (see Lemmas~\ref{lem:randvec} and \ref{lem:privgauss}). In this case, the straightforward approach leads to worse bounds and one of our main contributions is to improve the bound under $(\epsilon,\delta)$-differential privacy by using concentration bounds and ignoring the tail. By contrast, in $(\epsilon,0)$-differential privacy, one cannot ignore what happens in the tail of the distribution and understanding the algorithm in that regime seems difficult. We believe our framework is general and it facilitates further progress in differentially private optimization. We demonstrate our framework by instantiating it with the Laplace mechanism (to obtain an $(\epsilon, 0)$-private algorithm) and with the Gaussian mechanism (to obtain an $(\epsilon, \delta)$-private algorithm). It would be interesting to apply our framework to other notions of differential privacy, such as concentrated differential privacy~\citep{dwork2016concentrated,bun2016concentrated} and Renyi differential privacy~\citep{mironov2017renyi}. While we resort to established techniques from differential privacy (Gaussian and Laplacian mechanisms, tree based aggregation), properly integrating them with optimization methods does require some degree of care. For example, our $(\epsilon, \delta)$-privacy bound is derived using a matrix concentration inequality which crucially relies on a randomized smoothing technique used for obtaining gradient estimates. This is a key ingredient to obtaining the correct $\widetilde{O}(n^{1/2})$ dependence in dimension in the $(\epsilon, \delta)$ regime. The final algorithm is simple but effective, matching the best known bound with projection in $(\epsilon, \delta)$-DP. We see it as a proof of concept for a general approach to deriving differentially private optimization methods. Previous results in this area can be recovered by following our approach: inject the maximum amount of noise as to not change the convergence rate asymptotically, then analyze the privacy loss. This is very simple, but it paves the way for further development of practical differentially private algorithms, without requiring major changes in their implementation -- simply replace certain components of the algorithm with a black box implementation of the required differentially private mechanism. \medskip {\bf Other Related Work.} The theory of online convex optimization is truly extensive, and has seen a lot of developments in the recent years. Here, we will refer to the relevant works that involve projection-free and/or differentially private online learning algorithms. The class of projection-free online learning algorithms was initiated by the work of~\citep{hazan2012projection} in the context of online convex optimization, where full gradients are revealed after making a decision. This was further extended to multiple regimes~\citep{garber2013playing, garber2013linearly, garber2015faster} including the bandit setting~\citep{chen2018projection,garber2019improved}. As discussed above, Smith and Thakurta~\citep{thakurta2013nearly} achieve the state of the art regret guarantee for $(\epsilon,0)$-private online bandit optimization when projections are available. For general Lipschitz functions, their regret is $\widetilde{O}(n T^{3/4}/\epsilon)$. In the specific case where the adversary is oblivious and the loss functions are strongly-convex, they improve this to $\widetilde{O}(n T^{2/3}/\epsilon)$. In a different line of work,~\citep{agarwal2017price} obtained improved bounds for the case where losses are linear functions and for the multi-armed bandit problem (a generalization of the learning with experts framework), with regret $\widetilde{O}(T^{2/3}/\epsilon)$ and $\widetilde{O}(nT^{2/3}/\epsilon)$ respectively. These results, however, concern only a restricted class of bandit optimization problems, so for the general bandit convex optimization problem the result of~\citep{thakurta2013nearly} still stands as the best one. In fact, even in the non-private setting improving the regret of the bandit convex optimization problem from $\widetilde{O}(T^{3/4})$ to $\widetilde{O}(T^{2/3})$~\citep{awerbuch2004adaptive,dani2006robbing} or below~\citep{dani2008price,abernethy2009competing,bubeck2017kernel} requires stronger access to the set of actions than just projections (such as via a self-concordant barrier), and involves performing expensive computations. Indeed,~\citep{bubeck2017kernel} is the first to achieve both optimal regret $\widetilde{O}(T^{1/2})$ and polynomial running time per iteration. \section{Preliminaries}\label{sec:prelim} {\bf Bandit Convex Optimization.} In the bandit convex optimization problem, an algorithm iteratively selects actions $\x_t$ (using a possibly randomized strategy) from a convex set $\D \subseteq \R^n$. After selecting an action, the loss caused by this choice $f_t(\x_t)$ is revealed, where $f_t : \D \rightarrow \R$ is a convex function unknown to the algorithm. After performing this for $T$ iterations, the algorithm compares its total loss $\sum_{t=1}^T f_t(\x_t)$ to the smallest loss it could have incurred by choosing a fixed strategy throughout all the iterations $\min_{\x \in \D} \sum_{t=1}^T f_t(\x)$. The difference between these two losses is called \textit{regret}: \[ \Reg_T = \sum_{t=1}^T f_t(\x_t) - \min_{\x \in \D} \sum_{t=1}^T f_t(\x) \] and the goal is to minimize its expectation over the randomized choices of the algorithm. \medskip {\bf Differential Privacy.} Differential privacy~\citep{dwork2006calibrating} is a rigorous framework used to control the amount of information leaked when performing computation on a private data set. In our framework, we seek algorithms which ensure that the amount of information an adversary can learn about a particular loss function $f_t$ is minimal, i.e. it is almost independent on whether $f_t$ appears or not in the sequence of loss functions occurring throughout the execution of the algorithm. For completeness, we define differential privacy in the context of loss functions encountered in the bandit convex optimization problem. \begin{definition}[$(\epsilon,\delta)$-differential privacy]\label{def:dp} A randomized online learning algorithm $\calA$ on the action set $\D$ is $(\epsilon,\delta)$-differentially private if for any two sequences of loss functions $F = (f_1, \dots, f_T)$ and $F' = (f_1', \dots, f_T')$ differing in at most one element, for all $S \subseteq \D^T$ it holds that \[ \Pr[\calA(F) \in S] \leq e^\eps \Pr[ \calA(F') \in S ] + \delta\,. \] \end{definition} One obstacle that may occur in the context of bandit optimization is that changing a single loss function may alter the set of actions returned in the future by the algorithm. \medskip {\bf The Projection-Free Setting.} While classical online optimization methods have a long history of developments, these rely in general on the ability to perform projections onto the feasible set $\D$ of actions. For example, one may want to choose actions that correspond to points inside a matroid polytope, or other complicated domains. In such situations, it is computationally infeasible to perform projections onto $\D$, and designing algorithms where all the actions lie inside this domain becomes a challenging task. In the case of online optimization, this issue is mitigated by \textit{projection-free methods}~\citep{jaggi2013revisiting,garber2015faster,dudik2012lifted,shalev2011large}, where the complexity of the high complexity of the description of $\D$ is balanced by the existence of a linear optimization oracle over this domain. Among these, the conditional gradient method (also known as Frank-Wolfe)~\citep{bubeck2015convex} is the best known one. In our setting, we treat the case where, although $\D$ may be very complicated, we have access to such an oracle which given any direction $\vv\in\R^n$ returns $\arg\min_{\x \in \D} \langle \vv, \x \rangle$. Such oracles are easily available for interesting domains such as the matroid polytope, or the submodular base polytope. \medskip {\bf Parameters and Assumptions.} We write vectors and matrices in bold. We use $\langle \x, \y \rangle$ to represent inner products, and $\|\x\|$ to represent the $\ell_2$ norm of a vector $\|\x\| = \left(\sum_i x_i\right)^{1/2}$. When considering other norms than $\ell_2$ we explicitly specify them $\|\x\|_p = \left(\sum_i x_i^p \right)^{1/p}$. We let $B_p^n$ be the $n$ dimensional unit $\ell_p$ ball and $S_p^n$ the boundary of $B_p^n$ i.e. the $n$ dimensional unit $\ell_p$ sphere. We consider optimizing over a convex domain $\D \subseteq \R^n$, for which we have access to a linear optimization oracle. We define the diameter of the domain as $D = \max_{\x,\y\in\D} \|x-y\|$. We say that a function $f:\D\rightarrow \R$ is $L$-Lipschitz iff $\vert f(\x) - f(\y) \vert \leq L \|\x-\y\|$ for all $\x,\y\in\D$ and that a differentiable function $f$ is $\beta$-smooth iff $\|\nabla f(\x) - \nabla f(\y)\| \leq \beta\|\x-\y\|$ for all $\x,\y \in \D$. We say that $f$ is $\alpha$-strongly convex iff $\|\nabla f(\x) - \nabla f(\y)\| \geq \alpha\|\x-\y\|$. In our setting, all functions $f_t$ satisfy the standard assumption of being $L$-Lipschitz. Just like in prior works~\citep{thakurta2013nearly,agarwal2017price}, we further assume that the number of iterations $T$ we run the algorithm for is known ahead of time. This assumption can be eliminated via a standard reduction using the doubling trick (see~\citep{auer1995gambling} and~\citep{chen2018projection}), which invokes the base algorithm repeatedly by doubling the horizon $T$ at each invocation, at the expense of adding an extra $O(\log T)$ factor in the privacy loss. For simplicity further assume that all $f_t$'s are defined within a region that slightly exceeds the boundary of $\D$. This assumption is required, since one of the techniques employed here requires having $f_t$ defined over $\D \oplus \ndel B_2^n$ for a small scalar $\ndel$. This assumption can be removed via a simple scaling trick, whenever $\D$ contains an $\ell_2$ ball centered at the origin (similarly to~\citep{garber2019improved}); we explain how to do so in Appendix~\ref{sec:domain}. Finally, in order to be able to appropriately privatize the losses $f_t(\x_t)$ we require bounding their magnitude. To do so we assume that each $f_t$ achieves $0$ loss at some point within $\D$, which via the Lipschitz condition and the diameter bound automatically implies that $\vert f_t(\x_t) \vert \leq LD$ for all $t$. Other related works~\citep{flaxman2004online, agarwal2010optimal, thakurta2013nearly} simply use a fixed upper bound $\vert f_t(\x_t) \vert \leq B$ for some fixed parameter $B$, but we prefer this new convention to reduce the number of parameters to control, since we are focused mainly in the regret dependence in $T$, $n$ and $\epsilon$. \medskip {\bf Mirror Maps and the Fenchel Conjugate.} In general, convex optimization implicitly relies on the existence of a \textit{mirror map} $\omega : \D \rightarrow \R$ with desirable properties (see~\citep{ben2001lectures} for an extensive treatment of these objects). This is used in order to properly interface iterates and gradient updates, since in Banach spaces these are of different types. In our specific case we use $\omega(\x) = \frac{1}{2} \|\x\|_2^2$, although other choices can be used depending on the geometry of $\D$. We define the Fenchel conjugate of $\omega$ as $\omega^* : \R^n \rightarrow \R$ such that \begin{equation}\label{eq:fenchelyoung} \omega^*(\y) = \max_{\x \in \D} \langle \y, \x \rangle - \omega(\x)\,. \end{equation} Furthermore, one has that whenever $\omega$ is strongly convex, $\omega^*$ is smooth and differentiable~\citep{nesterov2005smooth}, and its gradient satisfies \begin{equation} \nabla \omega^*(\y) = \arg\max_{\x \in \D} \langle \y, \x \rangle - \omega(\x)\,. \end{equation} \medskip {\bf Smoothing.} We use the randomized smoothing technique from~\citep{flaxman2004online} in order to smoothen the loss functions $f_t$. This technique is crucial to obtain gradient estimators despite having only value access. \begin{lem}[\citep{flaxman2004online}]\label{lem:smooth} Let $f:\mathbb{R}^{n}\rightarrow\mathbb{R}$ be a convex and $L$-Lipschitz function. Then the smoothing $\hatf(\boldsymbol{x})=\mathbb{E}_{\uu\sim B_{2}^n}f\left(\boldsymbol{x}+\ndel \uu\right)$ satisfies the following properties: (1) $\left|f(\x)-\hatf(\x)\right|\leq\ndel L$, (2) $\widehat{f}$ is convex and $L$-Lipschitz, (3) $\nabla\hatf(\x)=\frac{n}{\ndel}\cdot\mathbb{E}_{\uu\sim S_{2}^n}f(\x+\ndel \uu)\cdot \uu$. \end{lem} \medskip {\bf Tree Based Aggregation.} An essential ingredient of the algorithm is maintaining partial sums of the gradient estimators witnessed so far. We use a variant of the algorithm from~\citep{dwork2010differential, jain2012differentially}, as implemented in~\citep{agarwal2017price}. We use the algorithm as a black box and only rely on its properties that are stated in Theorem~\ref{thm:treebthm} below. We include a description of the \textsc{TreeBasedAgg} algorithm in Appendix~\ref{sec:tbagg_app} for completeness. \begin{thm}[\citep{jain2012differentially,agarwal2017price}] \label{thm:treebthm} Let $\{\ell_t\}_{t=1}^T$ be a sequence of vectors in $\R^n$, and let $Y_1$ and $Y_2$ be promises such that $\| \ell_t \|_1 \leq Y_1$ and $\| \ell_t \|_2 \leq Y_2$ for all $t$. Let $\epsilon, \delta > 0$, and $\lambda \geq \frac{Y_1 \log T}{\eps}$ and $\sigma \geq \frac{Y_2}{\eps} \log T \log \frac{\log T}{\delta}$. There is an algorithm, \textsc{TreeBasedAgg}, that first outputs $\widehat{L}_0$ and then iteratively takes $\ell_t$ as input and returns an approximate partial sum $\widehat{L}_t$ for $1\leq t \leq T$. The algorithm can be specified with a noise distribution $\mathcal{P}$ over $\R^n$ so that the output sequence $\{\widehat{L}_t\}_{t=1}^T$ satisfies $\widehat{L}_t = \sum_{s=1}^t \ell_s + \sum_{r=1}^{\lceil \log T \rceil}Z_r$, where $Z_r \sim \mathcal{P}$, and furthermore: \begin{itemize}[leftmargin=*,noitemsep,topsep=0pt] \item when $\mathcal{P}$ is coordinate-wise $Lap(0, \lambda)$, the sequence $\{\widehat{L}_t\}_{t=1}^T$ is $(\epsilon, 0)$-differentially private. \item when $\mathcal{P}$ is coordinate-wise $\mathcal{N}(0,\sigma^2)$, the sequence $\{\widehat{L}_t\}_{t=1}^T$ is $(\epsilon, \delta)$-differentially private. \end{itemize} \end{thm} \section{Algorithm}\label{sec:mainalg} The algorithm is described in Algorithm~\ref{alg:bandit}. It builds on the work of Garber and Kretzu \citep{garber2019improved} and uses the smoothing (Lemma~\ref{lem:smooth}) and tree aggregation (Theorem~\ref{thm:treebthm}) routines designed in previous work (see Section~\ref{sec:prelim}). The algorithm follows the structure of an online mirror descent algorithm. It performs a sequence of iterations, and in each iteration it makes a guess $x_t$ based on the previous outcomes. The iterations are divided into $\Tr$ batches, each of size $\Tb$ (thus, $T=\Tr\cdot \Tb$). Each batch $R$ is treated as a round for online mirror descent with a twist: in parallel, we compute the best regularized response $\xtil_R$ for the revealed outcomes in the first $R-1$ batches (line~$14$) and use the previously computed $\xtil_{R-1}$ for all iterations in batch $R$ (lines $6$ to $12$). Three notices are in order: \begin{itemize}[leftmargin=*,noitemsep,topsep=0pt] \item Computing the best regularized response to previous outcomes requires maintaining the sum of gradients in previous rounds. The tree-based aggregation method (Theorem~\ref{thm:treebthm}) is used to maintain these sums accurately while preserving privacy (line~$13$). \item In each iteration of a batch, the algorithm only has access to the function value, not the gradient, so we use the smoothing technique (Lemma~\ref{lem:smooth}): the function value at a perturbation of $\xtil_{R-1}$ is used to obtain a stochastic estimate of the gradient of a smoothed proxy of the objective function. Thus, each iteration in the same batch uses a different perturbation of the same response $\xtil_{R-1}$. \item We only compute an approximation of the best regularized response, using the conditional gradient method in line 14. The precision to which this is computed is chosen in such a way that the number of iterations required by conditional gradient matches the number of iterations in a batch, so that we can charge each call to the linear optimization oracle over $\mathcal{D}$ to one iteration of the bandit algorithm. \end{itemize} \setlength{\columnsep}{2.5cm} \setlength{\multicolsep}{0.0pt} \begin{algorithm*}[h] \caption{\textsc{PrivateBandit}$(T, \Dist, D)$} \label{alg:bandit} \begin{algorithmic}[1] \INPUT time horizon $T$, symmetric noise distribution $\Dist$, diameter of domain $D$. \STATE $\Tr = T^{1/2}, \Tb = \frac{T}{\Tr}, \eta = \frac{D}{T^{3/4} n^{1/2} L}$, $\ndel = \frac{Dn^{1/2}}{T^{1/4}}$. \STATE Initialize \textsc{TreeBasedAgg} for a sequence of length $\Tr$ and noise $\Dist$. \FOR{$R=1$ {\bfseries to} $\Tr$} \STATE \textbf{execute in parallel:} \begin{multicols}{2} \STATE $\gtil_R = 0$\label{line:batch-start} \FOR{$r=1$ {\bfseries to} $\Tb$} \STATE $t = (R-1)\Tb+r$ \STATE Sample $\uu_t \sim S_2(1)$ \STATE $\x_{t}=\xtil_{R-1} \textcolor{blue}{ + \ndel \uu_t}$\label{line:answer} \STATE Query $F_t = \frac{n}{\ndel}f_t(\xtil_{R-1} +\ndel \uu_t)$ \STATE $\gtil_R = \gtil_R + F_t \cdot \uu_t$ \ENDFOR\label{line:batch-end} \STATE \COMMENT{Update the partial sum of noisy gradients.} $\stil_{R} = \textsc{TreeBasedAgg}\left(\gtil_R, R\right)$.\label{line:tree}\vfill\columnbreak \STATE Solve via conditional gradient \[ \min_{\x\in\mathcal{D}}\frac{1}{2}\left\Vert \x\right\Vert _{2}^{2}-\left\langle \eta \stil_{R-1},\x\right\rangle \] to precision $\epscg = D/\Tb^{1/2}$. Let $\xtil_R$ be the output.\label{line:cg} \end{multicols} \ENDFOR \end{algorithmic} \end{algorithm*} The algorithm needs to be specified with a noise distribution $\mathcal{P}$ over $\R^n$, which we use for privatizing the partial sums in order to strike the right tradeoff between privacy and regret. To obtain an $(\epsilon,0)$-private algorithm, we set $\mathcal{P}$ to be coordinate-wise Laplace noise $Lap(0,\lambda)$. To obtain an $(\epsilon,\delta)$-private algorithm, we set $\mathcal{P}$ to be coordinate-wise Gaussian noise $\mathcal{N}(0,\sigma^2)$. The precise choice for the parameters $\lambda$ and $\sigma^2$ are established in Lemmas~\ref{lem:reglap} and \ref{lem:reggauss}. We analyze the regret and privacy guarantees of the algorithm in Sections~\ref{sec:regret} and \ref{sec:privacy}, respectively. While our algorithm roughly follows the line of the one from~\citep{garber2019improved}, the version employed here is a slight simplification and generalization of it, since in particular we do not require any specific stopping conditions and case analysis for solving the inner conditional gradient routine, and we can extend it to more general geometries defined by the mirror map. Also, the $\textsc{NoisyOCO}$ framework allows us to handle the noise introduced by the differentially private mechanisms without making any further changes to the algorithm or its analysis. This framework may be of further interest for designing differentially private optimization methods. In the following section we analyze the regret of Algorithm~\ref{alg:bandit}, for which we prove the following regret bound. \begin{lem}\label{lem:finalregret} Let $\mu = \E_{X\sim\Dist}\|X\|$, and let $D$ be the diameter of the domain $\mathcal{D}$, $n$ the ambient dimension, and $L$ an upper bound on the Lipschitz constant of the loss functions $f_t$. Then the algorithm \textsc{PrivateBandit} obtains a regret of $$O\bigg( T^{3/4} n^{1/2} LD + T^{1/4} D \mu \log T \bigg) \,.$$ \end{lem} \section{Noisy Mirror Descent Framework and Regret Analysis} In this section, we sketch the regret analysis for Algorithm~\ref{alg:bandit}. We derive the algorithm's regret guarantee via the \textsc{NoisyOCO} framework --- a meta-algorithm for online convex optimization with noise --- that we describe and analyze in Section~\ref{sec:noisyoco}. In Section~\ref{sec:regret}, we show that Algorithm~\ref{alg:bandit} is an instantiation of this meta-algorithm and we derive its regret guarantees from the guarantee for \textsc{NoisyOCO}. \subsection{Noisy Mirror Descent Framework} \label{sec:noisyoco} Here we describe and analyze the \textsc{NoisyOCO} algorithm (Algorithm~\ref{alg:noisyoco}) for online convex optimization with noise. We assume that we perform online convex optimization over a convex domain $\D$ endowed with a strongly convex mirror map $\omega : \D \rightarrow \R$ such that $\max_{\x\in\D} \omega(\x) \leq D_{\omega}^2$. We also assume $(\kappa, \gamma)$-noisy gradient access, defined as follows: \begin{itemize}[leftmargin=*,noitemsep,topsep=0pt] \item a noisy gradient oracle for $f_t$; given $\x$, it returns a randomized $\gtil = \textsc{NoisyGrad}(f_t, \x)$ such that $\E \gtil = \nabla f_t(\x)$, and $\E \| \gtil \|^2 \leq \kappa^2$, \item a noisy gradient oracle for $\omega^*$; given $\g$, it returns a randomized $\xtil = \textsc{NoisyMap}(\g)$ such that $\E \|\nabla \omega^*(\g) - \xtil \| \leq \gamma$. \end{itemize} \begin{algorithm}[h] \caption{\textsc{NoisyOCO}$(T)$} \label{alg:noisyoco} \begin{algorithmic}[1] \INPUT time horizon $T$. \STATE $\eta = D_{\omega}^{1/2} / (\kappa T^{1/2}), \s_0 = 0$ \FOR{$t=1$ {\bfseries to} $T$} \STATE $\xtil_t = \textsc{NoisyMap}(-\eta \s_{t-1})$ \STATE {\bfseries output} $\xtil_t$ and {\bfseries query} $\gtil_t = \textsc{NoisyGrad}(f_t, \xtil_t)$ \STATE $\s_t = \s_t + \gtil_t$ \ENDFOR \end{algorithmic} \end{algorithm} Under these conditions we can derive the following regret guarantee. \begin{lem}\label{lem:noisyoco} Given an instance of online convex optimization with $(\kappa, \gamma)$-noisy gradient access, the algorithm \textsc{NoisyOCO} obtains an expected regret of $$\Reg_T = O\left(T^{1/2} \kappa D_{\omega} + T \kappa \gamma \right)\,.$$ \end{lem} We give the full proof in Appendix~\ref{sec:pf_noisy_oco}. \subsection{Regret Analysis} \label{sec:regret} The regret analysis is based on the guarantee for $\textsc{NoisyOCO}$ from Lemma~\ref{lem:noisyoco}. It follows from mapping the steps in Algorithm~\ref{alg:bandit} to the framework from \textsc{NoisyOCO}, and bounding the parameters involved. In order to do so, we explain how the \textsc{NoisyGrad} and \textsc{NoisyMap} routines are implemented by Algorithm~\ref{alg:bandit}. We then proceed to bound the $\kappa$ and $\gamma$ parameters corresponding to this specific instantiation, which will yield the desired result. Here we describe the steps required for analysis. We offer detailed proofs in the appendix. The first step is to reduce the problem to minimizing regret on a family of functions $\{\tilf_R\}_{R=1}^{\Tr}$, where $\tilf_R = \sum_{t=1}^{\Tr} \hatf_{(\Tb-1)\cdot R + t}$. This introduces two sources of error: one from using the smoothed $\hatf$ instead of $f$, and another from using different iterates $x_t = \xtil_{R-1}+\uu_t$ in the same round, even though batching iterations effectively constrains all the iterates in a fixed batch to be equal. These errors are easy to control, and add at most $O(T\ndel L)$ in regret. For the family of functions $\{\tilf_R\}_{R=1}^{\Tr}$ we implement $\textsc{NoisyGrad}$ as: \begin{align*} \gtil_R &:= \textsc{NoisyGrad}(\tilf_R, \x_t) \\ &= \sum_{t = (R-1)\Tb+1}^{R \Tb} \frac{n}{\ndel} f_t (\x_t + \ndel \uu_t) \uu_t\,, \end{align*} which is an unbiased estimator for $\nabla \hatf_R(\x_t)$, per Lemma~\ref{lem:smooth}. Thus we bound $\kappa^2$ by showing in Lemma 10.2 that $\E \|\gtil_R\|^2 \leq \Tb \cdot \left(LD n/\ndel \right)^2 + \Tb^2 L^2$. Furthermore, the output of \textsc{NoisyMap} is implemented in line 14 by running conditional gradient to approximately minimize a quadratic over the feasible domain $\mathcal{D}$. The error in the noisy map implementation comes from (1) only approximately minimizing the quadratic, (2) using a noisy partial sum of gradient estimators rather than an exact one, and (3) using a stale partial sum approximation $\stil_{R-1}$ for round $R+1$, instead of $\stil_{R}$. We show in Corollary 10.5 that the error parameter corresponding to this \textsc{NoisyMap} implementation can be bounded as $\gamma \leq \eta\left(\lceil \log T \rceil \cdot \mu + \kappa \right) + \sqrt{20} \frac{D}{\sqrt{ \Tb}}$. In Appendix~\ref{sec:regretanalysis}, we bound the specific parameters corresponding to these implementations. Plugging these with bounds inside Lemma~\ref{lem:noisyoco} yields the proof of Lemma~\ref{lem:finalregret}, after appropriately balancing the parameters $\Tb, \Tr, \ndel$. \section{Privacy Analysis}\label{sec:privacy} In this section, we instantiate Algorithm~\ref{alg:bandit} with appropriate noise distribution $\mathcal{P}$ in order to derive our $(\epsilon,0)$-private algorithm and our $(\epsilon,\delta)$-private algorithm. As mentioned earlier, we use Laplace noise for $(\epsilon,0)$-privacy and obtain the guarantee in Lemma~\ref{lem:reglap}, and we use Gaussian noise for $(\epsilon,\delta)$-privacy and obtain the guarantee in Lemma~\ref{lem:reggauss}. First, we describe the proofs for the $(\epsilon, 0)$-privacy regime, where we employ Laplace noise, since they show how this framework allows us to trade regret and privacy. \begin{lem}[Privacy with Laplace noise]\label{lem:lappriv} Let $\mathcal{P}$ be coordinate-wise $Lap(0, T^{1/2} nL \log T /\epsilon)$. The algorithm \textsc{PrivateBandit}$(T, \mathcal{P})$ is $(\epsilon, 0)$-differentially private. \end{lem} \begin{proof} First we bound the $\ell_1$ norm of the vectors whose partial sums are maintained by the tree based aggregation method in Algorithm~\ref{alg:bandit}. Since each vector contributing to that partial sum is obtained by adding up $\Tb = T^{1/2}$ vectors, each of which is a unit $\ell_2$ vector scaled by a constant that is absolutely bounded by $M = LD \frac{n}{\ndel} = T^{1/4}n^{1/2} L$, we naively bound the $\ell_1$ norm of each of them by $\Tb \cdot n^{1/2} \cdot M \leq T^{1/2} n L$. Therefore, by Theorem~\ref{thm:treebthm}, releasing $\Tr = T^{1/2}$ such partial sums causes a loss of privacy of at most $\epsilon$ whenever $$\lambda \geq \frac{ T^{1/2} n L \log T }{\epsilon}\,.$$ \end{proof} Using Lemma~\ref{lem:lappriv} we can now bound the regret of the $(\epsilon,0)$-differentially private algorithm. \begin{lem}[Regret with Laplace noise] \label{lem:reglap} Let $\mathcal{P}$ be coordinate-wise $Lap(0, T^{1/2} nL \log T /\epsilon)$. The algorithm \textsc{PrivateBandit}$(T, \mathcal{P})$ has regret $$\reg_T = O\bigg(T^{3/4}n^{1/2}LD + \frac{T^{3/4}n^{3/2} L D \log^2 T}{\epsilon} \bigg)\,.$$ \end{lem} \begin{proof} Per Lemma~\ref{lem:finalregret} we only need to upper bound the expected $\ell_2$ norm of an $n$-dimensional vector where each coordinate is independently sampled from $Lap(0, T^{1/2} nL \log T /\epsilon)$. Indeed, we have $$\mu = O\bigg(n^{1/2} \cdot T^{1/2} nL \log T /\epsilon\bigg)\,.$$ Plugging this into the regret guarantee from Lemma~\ref{lem:finalregret} we obtain the desired result. \end{proof} We notice that the regret guarantee we achieved has an undesirable dependence in dimension. In the remainder of this section, we show that we can obtain a improved guarantees if we settle for $(\epsilon, \delta)$-differential privacy instead, which we achieve by using Gaussian noise. This is also more properly suited to our setting, since the regret bound we proved depends on $\ell_2$ norms of the injected noise vectors, which is exactly what governs the privacy loss in this case. A novel and precise error analysis in Lemmas~\ref{lem:randvec} and \ref{lem:privgauss} enables us to obtain the same regret bound as when projection is available for $(\epsilon, \delta)$-privacy. We additionally use the fact that the randomized smoothing technique further constrains the norm of the vectors $\gtil_R$ we employ to update the partial sums, with high probability. In order to do this, we resort to a concentration inequality for random vectors (Lemma~\ref{lem:randvec}) which allows us to obtain an improved guarantee on privacy, at the expense of a slight increase in the $\delta$ parameter. Compared to the $(\epsilon, 0)$ case, allowing this low probability failure event enables us to save roughly a factor of $\widetilde{O}(T^{1/4}n^{-1/2})$ in the norm of the vectors we use to update the partial sums via tree based aggregation. In turn, these allow us to use less noise to ensure privacy, and therefore we obtain an improved regret. In order to do so, we first require a high probability bound on the $\ell_2$ norm of a sum of random vectors, multiplied by an adversarially chosen set of scalars. \begin{lem}\label{lem:randvec} Let $\uu_1, \dotsm \uu_k \sim B_2(1)$ be a set of independent random vectors in $\R^n$. Then, with probability at least $1-\delta$, one has that for any vector $\cc \in \R^k$ such that $\|\cc\| \leq \Delta$: \begin{align*} \left\Vert \sum_{i=1}^k \uu_i c_i \right\Vert \leq 10 \Delta \bigg( \log\frac{ n+k}{\delta} + \sqrt{\left(1+\frac{k}{n}\right)\log \frac{n+k}{\delta}}\bigg) \end{align*} \end{lem} \begin{proof} Consider the family of matrices $\{\Z_i\}_{i=1}^k \in \R^{n\times k}$ where $\Z_i$ has its $i^{th}$ column equal to $\uu_i$ and all the other entries are $0$. Therefore $\E \Z_k = 0$ and $\|\Z_k\| \leq 1$. Furthermore, by definition $\Z_i \Z_i^\top = \uu_i \uu_i^{\top}$ and $\Z_i^\top \Z_i = \|\uu_i\|^2 \cdot \ones_i \ones_i^\top$. Therefore $\E \Z_i \Z_i^\top = \mI / n$ and $\E \Z_i \Z_i^{\top} = \ones_i \ones_i^\top$. So \begin{align*} \sigma^2 &= \max\left\{\left\|\E \sum_{i=1}^k \Z_i \Z_i^\top\right\|, \left\|\E \sum_{i=1}^k \Z_i^\top \Z_i\right\|\right\} \\ &= \max \left\{ \left\| \frac{k}{n} \cdot \mI \right\|, \left\| \mI \right\| \right\} \leq 1+k/n\,. \end{align*} Letting $\Z = \sum_{i=1}^k \Z_i$, and using matrix Bernstein~\citep{tropp2015introduction}, we see that \[ \Pr\left[\left\|\Z \right\| \geq t \right] \leq (n+k)\exp( -t^2 / (2 (1+k/n) + 2t/3 ) )\,. \] Hence for $t =10\left(\log \frac{n+k}{\delta} + \sqrt{\left(1+\frac{k}{n}\right) \log\frac{ n+k}{\delta}}\right)$ one has that $\|\Z\| \leq t$ with probability at least $1-\delta$. Therefore with probability at least $1-\delta$ we have $\left\|\Z\cc \right\| \leq \|\Z\| \|\cc\| \leq 10\Delta \left(\log\frac{ n+k}{\delta}+ \sqrt{\left(1+\frac{k}{n}\right)\log \frac{n+k}{\delta}}\right)$ which implies what we needed. \end{proof} Now we can obtain a tighter bound the privacy loss when using Gaussian noise. \begin{lem}[Privacy with Gaussian noise]\label{lem:privgauss} Let $\mathcal{P}$ be coordinate-wise $\mathcal{N}(0, \sigma^2)$, where \begin{align*} \sigma &= \frac{ T^{1/4}n^{1/2}L \log T \log(T/\delta) }{\epsilon} \\ &\cdot \left( \log\frac{n+T}{\delta}+\sqrt{\left(1+\frac{T^{1/2}}{n}\right)\log \frac{n+T}{\delta}} \right) \end{align*} The algorithm \textsc{PrivateBandit}$(T, \mathcal{P})$ is $(\epsilon, \delta)$-differentially private. \end{lem} \begin{proof} Using Lemma~\ref{lem:randvec} and union bound we have that with probability at least $1-\delta_{0} T^{1/2}$ the $\ell_2$ norm of each of the $\Tb$ vectors contributing to the partial sums maintained in the tree based aggregation routine is $M = O\bigg(T^{1/4}n^{1/2}L \left( \log\frac{n+T}{\delta_0}+\sqrt{(1+\frac{\Tb}{n})\log \frac{n+T}{\delta_0}} \right)\bigg)$. By Theorem~\ref{thm:treebthm} maintaining these partial sums is thus $(\epsilon, \delta_0 T^{1/2} + \delta_1)$-differentially private, where $\epsilon = \frac{M \log T \log (T/\delta_1) }{\sigma}$. Hence setting $\delta_1 = \delta/2$, $\delta_0 = \delta/(2 T^{1/2})$ and \begin{align*} \sigma &= \frac{M \log T \log(T/\delta_1)}{\eps} =\frac{ T^{1/4}n^{1/2}L \log T \log(T/\delta) }{\epsilon} \\ &\cdot \left( \log\frac{n+T}{\delta}+\sqrt{\left(1+\frac{T^{1/2}}{n}\right)\log \frac{n+T}{\delta}} \right) \end{align*} yields an $(\epsilon, \delta)$-differentially private algorithm. \end{proof} \begin{lem}[Regret with Gaussian noise]\label{lem:reggauss} Let $\delta =1/(n+T)^{O(1)}$. The algorithm \textsc{PrivateBandit}$(T, \mathcal{N}(0,\sigma^2))$ where $\sigma$ is chosen according to Lemma~\ref{lem:privgauss} such that the algorithm is $(\epsilon,\delta)$-private has regret \begin{align*} \Reg_T &= O\bigg( T^{3/4} n^{1/2} LD \\ &+ \frac{T^{1/2}n LD \log^2 T \log(T/\delta)\log((n+T)/\delta)}{\eps} \\ &+\frac{ T^{3/4} n^{1/2} LD \log^2 T \log(T/\delta) \sqrt{\log((n+T)/\delta)} }{\eps} \bigg)\,. \end{align*} \end{lem} \begin{proof} Per Lemma~\ref{lem:finalregret} we only need to upper bound the expected norm of an $n$-dimensional vector where each coordinate is sampled from $\mathcal{N}(0, \sigma^2)$. In this case we have $\mu = O(n^{1/2} \sigma)$, so plugging it into the regret guarantee from Lemma~\ref{lem:finalregret} we obtain regret \begin{align*} \Reg_T &= O\bigg( T^{3/4} n^{1/2} LD + T^{1/4} n^{1/2} D \sigma \log T \bigg) \end{align*} which implies the result after substituting $\sigma$. \end{proof} The proof of Theorem~\ref{thm:maininfo} now follows from combining Lemmas~\ref{lem:lappriv}, ~\ref{lem:reglap}, ~\ref{lem:privgauss} and~\ref{lem:reggauss}. \section{Discussion and Open Problems} We saw how one can a derive differentially private algorithm starting from a very basic framework for noisy online convex optimization. Our analysis builds on advances in both differential privacy and online convex optimization, combines their techniques in non-trivial ways and introduces new ideas. Among others, a novel and precise error analysis in Lemmas~\ref{lem:randvec} and \ref{lem:privgauss} enables us to obtain the same regret bound as when projection is available for $(\epsilon, \delta)$-privacy, in contrast with $(\epsilon, 0)$-privacy. To the best of our knowledge, this is a rare case where such a difference between the two privacy settings arise. We think it is an interesting direction for future work to obtain an analogous improvement even in the $(\epsilon, 0)$-privacy setting. It would be interesting to see if this generic method in conjunction with tools from differential privacy can be used to obtain more private learning algorithms. A few outstanding questions remain. Since $\widetilde{O}(T^{3/4})$ is also the best known regret bound in the non-private setting, it would be interesting to improve this result, which may lead to an improved differentially private algorithm. Furthermore, in the non-private setting with projections, obtaining algorithms with lower regret requires a stronger control of the geometry of the domain -- this makes differential privacy more difficult to achieve since even the simplest algorithms with improved regret require solving a linear system of equations, which is much more sensitive to noise than vanilla gradient steps. Developing a more fine-grained understanding of these problems via differential privacy poses an exciting challenge. \section*{Acknowledgments} AE was supported in part by NSF CAREER grant CCF-1750333, NSF grant CCF-1718342, and NSF grant III-1908510. HLN was supported in part by NSF CAREER grant CCF-1750716 and NSF grant CCF-1909314. AV was supported in part by NSF grant CCF-1718342. \bibliographystyle{plainnat}

ae1db48accdef388d52676a8770c84bd05e8cc6a

\section{Introduction} Classical molecular dynamics (MD) is a compute-intensive technique that enables quantitative studies of molecular processes. Of the possible modeling approaches, classical all-atom MD represents all of the atoms of a chosen system explicitly (including solvent) and accounts for inter-atomic forces through classical bonded and non-bonded potentials. It has seen remarkable developments due to its fidelity and it has been applied with success to problems such as conformational changes, folding, binding, permeation, and many others. \cite{Schulten_discovery_2009} It has, however, two important limitations: first, it needs extensive and carefully-optimized tables of inter-atomic potentials (known as force fields)\cite{ponder_force_2003}; second, it is compute-intensive, and despite heroic efforts and progress in accelerating MD codes \cite{martinez-rosell_drug_2017}, it still struggles to reach the temporal scales of several important physiological processes. Machine learning (ML) potentials have become especially attractive with the advent of deep neural network (DNN) architectures, which enable the example-driven definition of arbitrarily complex functions and their derivatives. As such, DNNs offer a very promising avenue to embed fast-yet-accurate potential energy functions in MD simulations, after training on large-scale databases obtained from more expensive approaches. One particularly interesting feature of neural network potentials is that they can learn many-body interactions. The SchNet architecture \cite{schutt2017schnet, schutt2018schnet}, for instance, learns a set of features using continuous filter convolutions on a graph neural network and predicts the forces and energy of the system. SchNet was originally used in quantum chemistry to predict energies of small molecules form their atomistic representations. A key feature of using SchNet is that the model is inherently transferable across molecular systems. More recently, this has been extended to learn a potential of mean force which involves averaging of a potential over some coarse-grained degrees of freedom \cite{ruza2020temperature, duvenaud2015convolutional,husic2020coarse, wang2019machine, nuske2019coarse, wang2020ensemble,ZhangHan2018_CG}, which however pose challenges in their parameterization \cite{wang2019coarse,D0CC03512B}. Indeed, molecular modeling on a more granular scale has been tackled by so-called coarse-graining (CG) approaches before \cite{marrink_perspective_2013,machado_sirah_2019,Saunders2013,Izvekov2005,Noid2013,ClementiCOSB}, but it is particularly interesting in combination with DNNs. Here, we introduce TorchMD, a molecular dynamics code built from scratch to leverage the primitives of the ML library PyTorch \cite{NEURIPS2019_pytorch}. TorchMD enables the rapid prototyping and integration of machine-learned potentials by extending the bonded and non-bonded force terms commonly used in MD with DNN-based ones of arbitrary complexity. The two key points of TorchMD are that, being written in PyTorch, it is very easy to integrate other ML PyTorch models, like ab-initio neural network potentials (NNPs) \cite{gao_torchani_2020, schutt2018schnet} and machine learning coarse-grained potentials \cite{wang2019machine, husic2020coarse}. Secondly, TorchMD provides the capability to perform end-to-end differentiable simulations \cite{D0CC03512B,wang2020endtoend}, being differentiable on all of its parameters. Similarly, Jax\cite{jax2018github} was used to perform end-to-end molecular simulations on Lennard-Jones systems\cite{schoenholz2019jax} and for biomolecular systems as well\cite{timemachine}. Other efforts have tackled the integration of MD codes with DNN libraries, although in different contexts. For all-atom models, \citet{wang2020endtoend} demonstrated the use of graph networks to recover empirical atom types. Ab-initio QM-based training of potentials is being tackled by several groups, including \citet{yao_tensormol-01_2018, schnetpack, gao_torchani_2020} but not using a differentiable PyTorch environment. This paper provides an account of the capabilities of TorchMD (Section \ref{sec:features}), highlighting the functional forms supported and an effective fitting strategy for data-driven DNN potentials. All of the TorchMD code, including the CLN tutorial and the corresponding training data, is open-source and available at \url{github.com/torchmd}. \section{Methods} \label{sec:features} \subsection{TorchMD simulations} TorchMD is, at first glance, a standard molecular dynamics code. It offers NVT ensemble simulations including a Langevin thermostat. Starting atomic velocities are derived from a Maxwell-Boltzmann distribution. Integration is done using the velocity Verlet algorithm. Long-range electrostatics are approximated using the reaction field method \cite{rfe}. TorchMD also supports simulations of periodic systems. Minimization is done using the L-BFGS algorithm. Because it is written in Python using PyTorch arrays, it is also very simple to modify and simulations can be run on any devices supported by PyTorch (CPU, GPU, TPU). However, unlike specialized MD codes \cite{harvey2009acemd} it is not designed for speed. TorchMD uses chemical units consistent with classical MD codes such as ACEMD\cite{harvey2009acemd}, namely kcal/mol for energies, K for temperatures, g/mol for masses and \r{A} for distances. \subsection{Analytical potentials} TorchMD supports reading AMBER force field parameters through parmed \cite{parmed}. In addition to that, to allow for faster prototyping and development, it implements its own easy to read YAML-based force field format. An example force field YAML file for the simulation of a water box is given in Figure \ref{fig:yaml}. Currently, TorchMD missing features include hydrogen bond constraints and neighbour lists. \begin{figure}[h!] \centering \inputminted[linenos]{yaml}{Figures/water_forcefield.yaml} \caption{An example YAML forcef ield for water molecules} \label{fig:yaml} \end{figure} TorchMD implements the functional form of the AMBER potential \cite{maier_ff14sb:_2015}. It offers all basic AMBER terms: harmonic bonds, angles, torsions and non-bonded Van der Waals and electrostatic energies. The above potentials are implemented as follows. The bonded potential terms are calculated as $$V_\mathit{bonded} = k_0 (r - r_{eq})^2$$ where $k_0$ is the force constant, $r$ the distance between the bonded atoms and $r_{eq}$ the equilibrium distance between them. The angle terms are calculated as $$V_\mathit{angle} = k_\theta (\theta - \theta_{eq})^2$$ where $\theta$ the angle between the three bonded atoms, $k_\theta$ is the angular force constant, and $\theta_{eq}$ the equilibrium angle. The torsion terms are calculated as $$V_\mathit{torsion} = \sum_{n=1}^{n_{max}} k_n (1 + \cos(n \phi - \gamma))$$ where $\phi$ the dihedral angle between the four atoms, $\gamma$ is the phase offset and $k_n$ the amplitude of the harmonic component of periodicity~$n$. The non-bonded Van der Waals (VdW) terms are calculated as $$V_\mathit{VdW} = \frac{A}{r^{12}}-\frac{B}{r^6}$$ where $A = 4 \epsilon \sigma^{12}$ and $B = 4 \epsilon \sigma^6$ with $\epsilon$ being the well depth of the interaction of two atoms and $\sigma$ the distance at which the energy is zero. The VdW potential also supports a cutoff by using a switching distance. Its energy is then obtained by multiplying the $V_\mathit{VdW}$ term with the scaling factor $$S = 1 - 6 x^5 + 15 x^4 -10 x^3$$ $$\mbox{with} \quad x = (r - r_s)/(r_c - r_s)$$ where $r_{s}$ is the switching distance and $r_{c}$ the cutoff distance. Electrostatics without cutoff are implemented using the following potential. $$V_\mathit{electrostatic} = k_{e} \frac{q_i q_j}{r}$$ where $k_{e} = \frac{1}{4 \pi \epsilon_0}$ is Coulomb's constant, $q_i$ and $q_j$ the charges of the two atoms and $r$ the distance between them. Electrostatics with cutoff are modified to use the reaction field method \cite{rfe} as follows $$V_\mathit{electrostatic} = k_{e} q_i q_j\left(\frac{1}{r}+{k}_{\mathit{rf}}{r}^{2}-{c}_{\mathit{rf}}\right)$$ $${k}_{\mathit{rf}}=\left(\frac{1}{{r_\mathit{c}}^3}\right)\left(\frac{{\epsilon}_{\mathit{sol}}-1}{2{\epsilon}_{\mathit{sol}}+1}\right)$$ $${c}_{\mathit{rf}}=\left(\frac{1}{{r}_{\mathit{c}}}\right)\left(\frac{3{\epsilon}_{\mathit{sol}}}{2{\epsilon}_{\mathit{sol}}+1}\right)$$ where $r_c$ corresponds to the cutoff distance and $\epsilon_{sol}$ to the solvent dielectric constant. In addition to the above, TorchMD also trivially allows the use of any other external potential $V_{ext}$ written in PyTorch which takes atomic coordinates as input and outputs energy and forces. Thus the total potential is calculated as \begin{equation} \begin{split} V_\mathit{total} & = \sum^{n_{bonds}} V_\mathit{bonded} + \sum^{n_{angles}} V_\mathit{angle} + \sum^{n_{torsions}} \!\!\! V_\mathit{torsion} \\ & + \sum_{i}^{n_{atoms}}\sum_{j<i}^{n_{atoms}} (V_\mathit{VdW} + V_\mathit{electrostatic}) \\ & + \sum^{n_{ext}} V_\mathit{ext} \end{split} \end{equation} Since PyTorch offers automatic differentiation, there is no need to calculate analytical gradients from the forces. Forces can be obtained with a single autograd PyTorch call on the total energy of the system. Analytical gradients have been nevertheless implemented for all analytical AMBER potential terms for performance reasons. \subsection{Training machine learning potentials} TorchMD provides a fully usable code for training neural network potentials in PyTorch called TorchMD-Net(\url{github.com/torchmd/torchmd-net}). Currently we are using a SchNet-based\cite{schutt2017schnet} model. However, it would be straightforward to derive the forces from non-parametric kernel methods like FCHL,\cite{christensen2020fchl} by providing a simple force calculator class, or other ML potentials. This object just takes as input the positions and box every timestep and returns the external energies and forces computed with the method of choice. For the present work, we take the featurization and atom-wise layer from SchNetPack\cite{schnetpack}, but rewrote entirely the training and inference parts. In particular, to allow training on multiple GPUs, the network is trained using the PyTorch lightning framework \cite{falcon2019pytorch}. TorchMD can also run concurrently a set of identical simulations by just changing the random number generator seed, arranging the neural network potential into a batch for speed, thus recovering, at least partially, the efficiency of optimized molecular dynamics codes. For the QM9 dataset, we trained the model using a standard loss function using mean square error over the energies. For the coarse-grain model, training is performed using the bottom-up ``force matching'' approach, focused on reproducing thermodynamics of the system from atomistic simulations, as described in previous work\cite{wang2019machine, husic2020coarse}. \section{Results} \label{sec:cln} To demonstrate functionalities of TorchMD, here we present some application examples. Firstly, a set of typical MD use cases (water box, small peptide, protein and ligand) mainly to assess speed and energy conservation. Secondly, we validate the code on QM9, a dataset of 134k small molecule conformations with energies \cite{ramakrishnan2014quantum}. In this case, however we cannot run any dynamical simulations as the dataset only presents ground state conformations of the molecules, so we are mainly validating the training. Then, we demonstrate end-to-end differentiable capabilities of TorchMD by recovering force field parameters from a short MD trajectory. Finally, we present a coarse-grained simulation of a miniprotein, chignolin,\cite{honda2008crystal} using NNP trained on all-atom MD simulation data. Here, we also describe how to produce a neural network-based coarse-grained model of chignolin, although the methods exposed are general to any other protein. A step-by-step example of training and simulating CG model is presented in the tutorial available in the \url{github.com/torchmd/torchmd-cg} repository. \subsection{Simulations of all-atom systems and performance} \begin{center} \begin{table} \begin{tabular}{ lccc } \hline System & Atoms & TorchMD & ACEMD \\ \hline Water & 291 & 6 min 56 s & 7 s \\ Di-alanine & 688 & 8 min 44 s & 8 s \\ Trypsin & 3,248 & 13 min 2 s & 16 s \\ \hline \end{tabular} \caption{Performance comparison for 50,000 steps at 1 fs/timestep on different systems.} \label{tab:performance} \end{table} \end{center} The performance of TorchMD is compared against ACEMD3\cite{harvey2009acemd}, a high-performance molecular dynamics code. In Table \ref{tab:performance} we can see the three different test systems comprised of a simple periodic water box of 97 water molecules, alanine dipeptide and trypsin with the ligand benzamidine bound to it. As it can be seen, TorchMD is around 60 times slower on the test systems than ACEMD3 running on a TITAN V NVIDIA GPU. Most of the performance discrepancy can be attributed to the lack of neighbour lists for non-bonded interactions in TorchMD and is currently prohibitive for much larger systems as the pair distances cannot fit into GPU memory. This is not a strongly limiting factor for the CG simulations conducted in this paper as the number of beads remains relatively low for the test case. However, we believe that, with an upcoming implementation of neighbour lists, TorchMD can reach much better performance, albeit still slower than highly specialized codes as ACEMD3 due to the generic nature of PyTorch operations in addition with the PyTorch library overhead. To evaluate the correctness of the TorchMD implementation of the AMBER force-field we compared it against OpenMM for 14 different systems ranging from ions, water boxes, small molecules to whole proteins, thus testing all the different force-field terms. In all 14 test cases the potential energy difference between OpenMM and TorchMD was lower than $10^{-3}$ kcal/mol when computed with the same parameters. Energy conservation was validated with TorchMD by running a NVE simulation of a periodic water box for 1 ns with 1 fs timestep. Energy conservation normalized per degree of freedom was calculated as $E_{total} / n_{dof} R $ where $n_{dof}=870$ is the number of degrees of freedom of the system and $R$ the ideal gas constant. We obtained a mean value of $1.1~ 10^{-5}$ K per degree of freedom showing a good energy conservation. \subsection{Training validation on the QM9 dataset} The network was trained using Adam optimizer\cite{ kingma2014adam} with a learning rate scheduler on multiple GPUs by using PyTorch Lightning \cite{falcon2019pytorch}. An example of the configuration file for QM9 training is presented in Figure \ref{fig:train_input}. We performed multiple trainings using TorchMD-Net with different amounts of training data (Fig.\ref{fig:training}). The learning rate scheduler was determined with a patience of $10$ on a validation subset of $5\%$ of all data chosen at random. The performance reported is for the randomly chosen test set. The linear shape of the test performance in the log-log scale demonstrate the correctness of the training \cite{vapnik2013nature}. With the current set of hyperparameters (Fig.\ref{fig:train_input}) we report a best performance of $10$ meV for 100,000 training points, marginally better than the reported best performance of SchNet for QM9 \cite{schnetpack}. \begin{figure}[htpb!] \inputminted[linenos]{yaml}{Figures/train.yaml} \caption{An example of a training input file for training QM9} \label{fig:train_input} \end{figure} \begin{figure}[htb!] \centering \includegraphics[width=8cm]{Figures/learning_curve.png} \caption{Learning curve for the QM9 dataset.} \label{fig:training} \end{figure} \subsection{Demonstration of end-to-end differentiable simulations} The availability of automatic differentiation (AD) within a molecular dynamics package is beneficial beyond ML applications. Being able to compute gradients for all numerical operations opens up new avenues for sensitivity analysis, force field optimization and steered MD simulations, as well as simulations under highly complex constraints and restraints. To demonstrate these capabilities, the present example infers force field parameters from a short MD trajectory. First, a small water box containing 97 water molecules and one Na$^+$/Cl$^-$ ion pair was simulated using the TIP3P water model with flexible bonds and angles. After energy minimization and NVT equilibration at 300 K, the simulation was run for 10 ps in the microcanonical ensemble. The simulation used a 1 fs time step, a 9 \AA\ cutoff with 7.5 \AA\ switch distance, and reaction field electrostatics. Coordinates and velocities were saved every 10 steps. Next, all partial atomic charges $q$ in the system were annihilated (in practice, they were scaled by 0.01 to ensure non-vanishing gradients of the electrostatic potential). In order to infer $q$ from the MD trajectory, the integrator was initialized with snapshots $r(t_i)$, $v(t_i)$ from the trajectory. Then, 10 steps of simulation were run with the modified charges and the final positions from this short simulation were compared with the respective subsequent trajectory snapshot $r(t_{i+1})$. In other words, the simulation served as a parameterized propagator $Q: (r(t), v(t); q) \mapsto r(t + \delta_t)$ with $\delta_t = 10$ fs. Due to the AD capabilities within TorchMD, this propagator is end-to-end differentiable. \begin{figure}[htbp] \centering \includegraphics[width=\columnwidth]{Figures/loss_water.png} \includegraphics[width=\columnwidth]{Figures/charges_water.png} \caption{Inference of partial atomic charges $q$ from a short trajectory. Training loss (top) and charges (bottom) during training.} \label{fig:water} \end{figure} To recover the charges, we minimized the loss function $$ L\big(r(t_i), v(t_i); q\big) = \| Q(r(t_i), v(t_i); q) - r(t_{i+1}) \|_2^2, $$ i.e.\ the mean-squared distance between the ground-truth trajectory and the propagated coordinates (taking into account periodic boundary conditions). This loss function is differentiable with respect to the charges $q$ so that gradients can be obtained via backpropagation. Training was performed using Adam with a learning rate of $10^{-3}$ over one snapshot at a time. To enforce net neutrality, the positive charges ($q_H$ and $q_{\mathrm{Na}^+}$) were implicitly obtained from the oxygen and chlorine charges and only $q_O$ and $q_{\mathrm{Cl}^-}$ were explicitly optimized. Figure \ref{fig:water} shows the evolution of the training loss and the partial atomic charges during training. After just one epoch (1000 iterations), the original charges were recovered up to 3\% accuracy. \subsection{Coarse-graining all-atom systems} For our last application example, we built two coarse-grained models of chignolin: one solely based on $\alpha$-carbon atoms (CA) and another one based on $\alpha$-carbon and $\beta$-carbon atoms (CACB) (Fig.\ref{fig:cln-cg}). \begin{figure}[h!] \centering \includegraphics[width=\columnwidth]{Figures/beads_cln.png} \caption{Miniprotein chignolin: heavy-atom representation (left) and coarse-grained representations: CA beads connected by bonds (middle), CA and CB beads connected by bonds (right). The beads in coarse grained representations were coloured by bead type.} \label{fig:cln-cg} \end{figure} \subsubsection{Training data} We selected the CLN025 variant of chignolin (sequence YYDPETGTWY), which forms a $\beta$-hairpin turn while folded (Figure~\ref{fig:cln-cg}). Due to its small size (10 amino-acids) and fast folding, it has been extensively studied with MD \cite{lindorff2011fast, beauchamp2012simple, husic2016optimized, mckiernan2017modeling, sultan2018automated, scherer2019variational}. Training data was obtained from all-atom simulation of the protein in explicit solvent with ACEMD\cite{harvey2009acemd} on the GPUGRID.net distributed computing network.\cite{buch2010high} The system containing one chignolin chain was solvated in a cubic box of 40 \AA, containing 1881 water molecules and two Na$^+$ ions. The system was simulated at 350 K with CHARMM22* force field \cite{Piana2011} and TIP3P model of water.\cite{Jorgensen1983} A Langevin integrator was used with a damping constant of \SI{0.1}{\per\pico\second}. Integration time step was set to 4 fs, with heavy hydrogen atoms (scaled up to four times the hydrogen mass) and holonomic constrains on all hydrogen-heavy atom bond terms \cite{Feenstra1999}. Electrostatics were computed using Particle Mesh Ewald with a cutoff distance of 9 \AA\ and a grid spacing of 1 \AA. We used an adaptive sampling approach \cite{Doerr2014} where new simulations were started from the least explored states. As a result we obtained a total simulation time of \SI{180}{\us} with forces and coordinates saved every 100 ps giving a total of $1.8 \times 10^{6}$ frames. To obtain the training data for the CA model, the initial training set of coordinates and forces was filtered to retain only CA atoms positions and forces. In this example, a coarse grained system contains 10 beads, built out of seven unique types of beads, one for each amino acid type. The training set for CACB model was prepared in a similar fashion, filtering both CA and CB atoms, and achieving 19 beads and 8 unique types of beads, as all CA atoms were classified as one bead type with the exception of glycine and each CB was assigned an amino acid-specific bead type. Details of bead selection for both models are described in Supporting Methods. \subsubsection{Neural network potential training} \label{Pot_fit} For coarse-grained simulations it is important to provide some prior potentials in order to limit the space that the dynamics can visit to the space sampled in the training data\cite{wang2019machine}. All the terms of the force-field could be applied, but for simplicity we limit them to bonds and repulsions. Bonds prevent the protein polymer chain from breaking and repulsions stop computing NNP on very close atom distances where there is no data. For pairwise bonded terms, we used the all-atom training data to construct distance histograms for each pair of bonded bead types. Specifically, for each bonded pair, we counted the fraction of time that the respective distance spent in an equally-spaced bin in a distance range appropriate to the bead selection, 3.55~\AA\ and 4.2~\AA\ for all bonds between $\alpha$ carbon beads, and 1.3~\AA\ and~1.8 \AA\ for all bonds between $\alpha$ carbon and $\beta$ carbon. The distance distributions were Boltzmann-inverted to obtain free-energy profiles, and these were used to fit the equilibrium distance $r_0$ and the spring constant $k$ of the respective harmonic potential $$ V^\mathrm{(prior)}_\mathrm{harmonic}(r) = k (r-r_0)^2 + V_0, $$ where $r$ is the distance between the beads involved in the bond. Priors for non-bonded repulsive terms were derived analogously. Distance histograms were constructed with 30 equally-spaced distance bins between 3\,\AA\ and 6\,\AA\ and were used to fit the parameter $\epsilon$ of the repulsive potential $$V^\mathrm{(prior)}_\mathrm{repulsive}(r) = 4 \epsilon r^{-6} + V_0,$$ where $r$, as above, is the distance between the non-bonded beads. In fitting the potential curves we corrected for the reference state by normalizing counts of each bin by the volume of the corresponding spherical shell. Non-linear curve fits were performed with the Levenberg-Marquardt method of the SciPy package \cite{virtanen_scipy_2020}. The parameters of the prior forces are stored in a YAML force field file. Plots presenting the quality of fits are included in the Supporting Information (Fig. S1-S4) as well as YAML files describing prior force field. Based on resulting prior force field and input coordinates, we calculated a set of prior forces acting on the beads and then deducted them form true forces, resulting in a set of forces that we refer to as delta-forces. Along with coordinates, delta-forces were used as the input for training. In the case of CA model embeddings correspond to integers unique for each amino acid type. For CACB model all $\alpha$ carbons have the same embedding with the exception of glycine and each $\beta$ carbon has an embedding unique for each amino acid type. The network was trained using force matching approach, where a predicted force is compared to a true force from the training set \cite{wang2019machine, husic2020coarse}. In the example presented here, the network consisted of 3 interaction layers, 128 filters used in continuous-filter convolution, 128 features to describe atomic environments, 9 \AA\ cutoff radius and 150 Gaussian functions for CA model and $300$ Gaussians for CACB model. Increasing the number of Gaussian functions for CACB model was found to provide a higher stability of the model and prevent forming collapsed non-physical structures during the simulation. Models for simulation were selected when the validation loss reached a plateau. The training and validation loss as well as learning rates are presented in Supporting Figure 5. \subsubsection{Simulation of the NNP} \begin{figure}[h!] \inputminted[linenos, breaklines]{yaml}{Figures/input.yaml} \caption{An example of a simulation input file} \label{fig:input} \end{figure} \begin{figure*}[htpb!] \centering \includegraphics[width=\textwidth]{Figures/Fig_TICA.png} \caption{Two-dimensional free energy surfaces for the reference all-atom MD simulations (left) and the two coarse-grained models, CA (center) and CACB (right). The free energy surface for each simulation set was obtained by binning over the first two TICA dimensions, dividing them into a 120 $\times$ 120 grid, and averaging the weights of the equilibrium probability in each bin computed by the Markov state model. The reference MD simulations plot displays the locations of the three energy minima on the surface, corresponding to folded state (red dot), unfolded conformations (blue dot) and a misfolded state (orange dot). Both reference MD and coarse-grained simulations were performed at 350K. } \label{fig:Fig_TICA_CA} \end{figure*} The combinations of the force fields covering prior forces and the trained networks are used to simulate both CA and CACB systems with TorchMD. We introduce the parameters of the simulation as a YAML-formatted configuration file (Figure \ref{fig:input}), although the simulation can be also started from command line. The network is introduced to TorchMD as an external force, with specified network's location, embeddings and a calculator. An external force calculator class must have a "calculate" method that returns a tuple with energy and forces tensors. In our case, for both models, we run the simulation at 350K for 10 ns with 1 fs timestep, saving the output every 1000 fs. Note that while the simulations use a small timestep, the effective dynamics of the coarse-grained systems is much faster than the all-atom MD system as the coarse-grained model is supposed to reproduce the energetics but with much faster kinetics. Since TorchMD can easily handle parallel dynamics, we concurrently run ten simulations, of which five start from the folded state and five from unfolded conformations. The free energy surfaces obtained with a time-lagged independent component analysis (TICA)\cite{perez2013identification} for the all-atom baseline simulations and the coarse-grained simulations obtained with TorchMD are presented in Figure \ref{fig:Fig_TICA_CA}. The energy landscapes are obtained from binning the configurations over the first two TICA dimensions and computing the average of the equilibrium probability on each bin, obtained by Markov state model analysis of the microstate of each configuration. To support TICA plots, we included plots with RMSD values for the first 2 ns of representative trajectories for both models with different starting points (Fig.~\ref{fig:Fig_RMSD_all}). Plots presenting full trajectories are included in Supporting Information (Fig.S6-S9). Neither SchNet nor prior energy terms can enforce chirality in the system, because they both work purely on the distances between the beads. Therefore, the RMSD plots were supplemented with RMSD values of the trajectory's mirror image. Results show that the coarse-grained simulations for both models were able to obtain several folding and unfolding events for chignolin. The energy landscapes for the CA model show that it captured the folded state as a global minimum of energy. The simulations also covers other minima representing unfolded and misfolded states. However, they do not recreate the energy barriers connecting these basins (as expected), which is better seen on the one-dimensional free energy surfaces (Fig.S10). The CACB model also detects the global minimum correctly, but fails at guessing the free energy of the unfolded region. Overall the simulation is less stable than for CA model and the misfolded state minimum is incorrectly located. \begin{figure*}[h!] \centering \includegraphics[width=\textwidth]{Figures/RMSD_paper.png} \caption{The RMSD values across the first 2 ns of the unmodified trajectory (\emph{True}, red) and a mirror image of the original trajectory (\emph{Mirror}, gray) for CA model (on the left) and CACB model (on the right). Trajectory 4 (top left panel) and Trajectory 1 (top right panel) are examples of a trajectories started from the folded state for CA model and CACB model, respectively. Trajectory 8 (bottom left panel) and Trajectory 7 (bottom right panel) are examples of trajectories started from elongated chain for CA model and CACB model, respectively. A moving average of 100 frames is represented as darker lines. The full 10 ns of each simulation is included in Supporting Figures S6-S9.} \label{fig:Fig_RMSD_all} \end{figure*} \section{Conclusion} In this paper we demonstrated TorchMD, a PyTorch-based molecular dynamics engine for biomolecular simulations with machine learning capabilities. We have shown several possible applications ranging from Amber all-atom simulations, to end-to-end learning of parameters and finally a coarse-grained neural network potential for protein folding. In particular, building a NNP for protein folding requires supplementing it with asymptotic, analytical potentials for bonds and repulsions to prevent exploring conformations not visited in the training data in which the predictions of NNP are unreliable. We have shown how to coarse-grain a protein into either $\alpha$-carbon atoms or $\alpha$-carbon and $\beta$-carbon atoms. Currently, CA model seems to work the best but future research will indicate which models are better suited for a more diverse set of targets. TorchMD end-to-end differentiability of its parameters is a feature that projects such as the Open Force Field Initiative \cite{mobley_openforcefield_2018} can potentially exploit. Furthermore, for additional speed we plan to facilitate the integration of machine learning potentials in OpenMM\cite{eastman_openmm_2017} and ACEMD \cite{harvey2009acemd}. Meanwhile, we believe that TorchMD can play an important role by facilitating experimentation between ML and MD fields, speeding up the model-train-evaluate prototyping cycle, and promoting the adoption of data-based approaches in molecular simulations. All the code machinery to produce the models is made available for practitioners at \url{github.com/torchmd}. \begin{acknowledgement}. SD thanks the Chan Zuckerberg initiative for funding. We thank the volunteers of GPUGRID.net for donating computing time. This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 823712. \end{acknowledgement}

80baa517b0579560d7adaa035f016582beaa433c

\section{Introduction} As described in \cite{LiangPathConstruction12}, algorithm STPC (Spanning Tree-based Path Construction) builds three independent spanning trees \cite{GraphTheory2005} ($\mathcal{T}_1$, $\mathcal{T}_2$ and $\mathcal{T}_3$) wrt node $r$ in the 3-vertex-connected graph $\mathcal{G}^*_{ex}$ ($\mathcal{G}^*_{ex}$ is obtained by adding virtual nodes and virtual links to the original graph $\mathcal{G}$) for constructing a sufficient number of simple paths such that all links in $\mathcal{G}$ can be identified. Based on these three independent spanning trees, STPC constructs 3 internally vertex disjoint paths wrt each node in $\mathcal{G}$, thus resulting in $3|V|$ paths totally, where $V$ is the set of nodes in $\mathcal{G}$. However, there exists redundant information in these $3|V|$ paths since some of them are exactly the same. The main objective of this report is to prove that the number of distinct paths constructed by STPC in fact equals $|L|$, where $L$ is the set of links in $\mathcal{G}$. By linear algebra, we know that the minimum number of the required path measurements for full link identifications is $|L|$ and all these paths must be linearly independent. Therefore, STPC is characterized by high efficiency in that all constructed distinct paths are linearly independent, i.e., all link metrics in $\mathcal{G}$ can be identified without conducting any unnecessary path measurements. In the end, we show some additional simulation results to verify the superior efficiency of STPC and STLI. The terms and notations used in this report are defined in Table \ref{t notion}. \begin{table}[b] \renewcommand{\arraystretch}{1.3} \caption{Main Notations} \label{t notion} \vspace{-.5em} \centering \begin{tabular}{r|m{6.5cm}} \hline \textbf{Symbol} & \textbf{Meaning} \\ \hline $V(\mathcal{G})$, $L(\mathcal{G})$ & set of nodes/links in graph $\mathcal{G}$\\ \hline $m$, $n$ & number of nodes/links in $\mathcal{G}$ \\ \hline $\mathcal{G}+l$ & add a link: $\mathcal{G}+l=(V(\mathcal{G}),L(\mathcal{G})\cup \{l\})$, where the end-points of link $l$ are in $V(\mathcal{G})$\\ \hline $\mathcal{G} \setminus \mathcal{G}^{'}$ & From $\mathcal{G}$, delete all nodes in common with $\mathcal{G}'$ and their incident links \\ \hline $\mathcal{G} \cup \mathcal{G}^{'} $ & graph union: $ \mathcal{G} \cup \mathcal{G}^{'}=(V(\mathcal{G}) \cup V(\mathcal{G}'), L(\mathcal{G}) \cup L(\mathcal{G}'))$ \looseness=-1\\ \hline $\mathcal{P}$ & simple path, defined as a graph with $V(\mathcal{P})=\{v_0,\ldots,v_k\}$ and $L(\mathcal{P})=\{v_0v_1, v_1v_2,\ldots,v_{k-1}v_k\}$, where $v_0,\ldots,v_k$ are distinct nodes\\ \hline ${\mu}_i$ & ${\mu}_i\in V(\mathcal{G})$ is the $i$-th monitor in $\mathcal{G}$ \\ \hline $w_l$, $c_{\mathcal{P}}$ & metric of link $l$, sum metric of path $\mathcal{P}$ \looseness=-1\\ \hline \end{tabular} \vspace{-0mm} \end{table} \section{Independent Spanning Trees} With respect to $\mathcal{G}^*_{ex}$, let $\mathcal{P}_{STPC}$ denote the path set obtained by running STPC algorithm. As discussed before, the goal is to prove that the number of distinct paths in $\mathcal{P}_{STPC}$ is the number of links in $\mathcal{G}$, i.e., $|\mathcal{P}_{STPC}|=|L|$. For this purpose, we first briefly discuss how to construct the 3 independent spanning trees, some unique features of which will be used in later proofs. \subsection{Three Independent Spanning Trees} In this section, we consider how to construct three independent spanning trees wrt node $r$ in $\mathcal{G}^*_{ex}$. \begin{definition} (\cite{Cheriyan88}) \emph{Nonseparating ear decomposition:} Nonseparating ear decomposition of $\mathcal{G}$ is a decomposition $V(\mathcal{G})=V(\mathcal{E}_1\cup \mathcal{E}_2 \cup \cdots \cup \mathcal{E}_{n_e})$ (each $\mathcal{E}_i$ is called an \emph{ear}, $n_e$ ears in total) such that \begin{enumerate} \item $\mathcal{E}_1$ is an induced cycle, \item $\mathcal{E}_i$ ($2\leq i \leq n_e$) is an induced simple path with only its end-points in common with $\mathcal{E}_1 \cup \cdots \cup \mathcal{E}_{i-1}$, \item Let $\overline{\mathcal{G}_i}:=\mathcal{G}\setminus (\mathcal{E}_1\cup \mathcal{E}_2 \cup \cdots \cup \mathcal{E}_{i})$. For each $\mathcal{E}_i$, $\overline{\mathcal{G}_i}$ is connected and each internal node of $\mathcal{E}_i$ has a neighbor in $\overline{\mathcal{G}_i}$, and \item $|\mathcal{E}_{n_e}|=3$ and the internal node of $\mathcal{E}_{n_e}$ does not appear in other ears. \end{enumerate} \end{definition} \begin{definition} (\cite{Even76}) \emph{s-t numbering:} s-t numbering of $\mathcal{G}$ is a one-one function $f:\ V\rightarrow [1\cdots n]$ ($n=|V|$) such that \begin{enumerate} \item $f(s)=1$, $f(t)=n$, and \item For each vertex $v$ in $\mathcal{G}\setminus \{s,t\}$, it has a neighbor $u$ with $f(u)<f(v)$ and a neighbor $w$ with $f(w)>f(v)$. \end{enumerate} \end{definition} Suppose $\mathcal{E}_j$ is the first ear with node $v$ on it, then $j$ is called the \emph{ear level} of $v$, denoted by $g(v)=j$. It is proved in \cite{Cheriyan88} that a 3-vertex-connected graph $\mathcal{G}^*_{ex}$ has a nonseparation ear decomposition, by which each node in $\mathcal{G}^*_{ex}$ can be assigned an ear level and an s-t number. With the knowledge of the ear level and the s-t number of each node, three independent spanning trees can be constructed accordingly. Without loss of generality, in the sequel, we assume ${\mu}_1$ is a non-cutvertex monitor\footnote{A non-cutvetex monitor can be found since it it impossible that all monitors are cutvertices in an identifiable network according to the minimum monitor placement algorithm MMP \cite{MaInfocomOnline}.} in $\mathcal{G}$ and virtual node $r$ connects to ${\mu}_2$, i.e., in the definition \cite{LiangPathConstruction12} of $\mathcal{G}^*_{ex}$, ${\mu}_i$ equals ${\mu}_2$. Due to the special structure of $\mathcal{G}^*_{ex}$, the complicated algorithm \cite{Cheriyan88} of finding $\mathcal{E}_1$ can be avoided, i.e., simply choose $r{\mu'}_1{\mu}_1{\mu}'_2r$ as $\mathcal{E}_1$. Then the rest ears can be found by the ear decomposition algorithm described in the proof of Theorem 1 \cite{Cheriyan88}. To construct 3 independent spanning trees, one rule for selecting the last ear $\mathcal{E}_{n_e}$ is that its internal node must be ${\mu}_2$. One preferable property of nonseparation ear decomposition is that the s-t numbers can be naturally computed in the process of selecting each ear, following Algorithm \ref{alg_stNum} \begin{algorithm}[tb] \small \SetKwInOut{Input}{input}\SetKwInOut{Output}{output} \Input{Ear decomposition of graph $\mathcal{G}$} \Output{s-t number of each node in $\mathcal{G}$} \ForEach {ear $\mathcal{E}_i$} { \eIf {i=1} { $f(r)=1$, $f({\mu}'_2)=2$, $f({\mu}_1)=3$, $f({\mu}'_1)=4$\; } { Suppose the end-points of $\mathcal{E}_i$ are $v_1$ and $v_2$ with $f(v_1)<f(v_2)$ ($\eta = f(v_2)$) and $\psi = |\mathcal{E}_i|-2$, then\\ (i) for each existing node $w$ with $f(w)\geq\eta$, $f(w)\leftarrow f(w)+\psi$, and\\ (ii) following the direction from $v_1$ to $v_2$ on $\mathcal{E}_i$, all internal nodes of $\mathcal{E}_i$ are sequentially numbered from $\eta$ to $\eta+\psi-1$. } } \caption{s-t numbering of $\mathcal{G}$}\label{alg_stNum} \vspace{-.25em} \end{algorithm} \normalsize By running Algorithm \ref{alg_stNum}, the final s-t numbers for $r$ and ${\mu}'_1$ are $1$ and $n$ (the number of nodes in $\mathcal{G}^*_{ex}$), respectively. With the computed ear level and s-t numbers of a given 3-vertex-connected graph $\mathcal{G}^*_{ex}$, it is easy to show that each node $v$ in $\mathcal{G}^*_{ex}\setminus \{r,{\mu}'_1,{\mu}_2\}$ has three neighbors $w_1$, $w_2$ and $w_3$ such that \begin{enumerate} \item ear level $g(w_1)>g(v)$; \item $f(w_2)<f(v)$ and $g(w_2)\leq g(v)$, and \item $f(w_3)>f(v)$ and $g(w_3)\leq g(v)$. \end{enumerate} The above three properties are also the three rules to construct the corresponding independent spanning trees: \begin{itemize} \item $\mathcal{T}_1$: First, add link $r\mu_2$ to $\mathcal{T}_1$. Next, for each node $v$ in $\mathcal{G}_m\setminus \{r, {\mu}_2\}$, $\mathcal{T}_1$ involves link $vw$, where $w$ is a neighbor node of $v$ with $g(w)>g(v)$; \item $\mathcal{T}_2$: First, add link $r{\mu'}_2$ to $\mathcal{T}_2$. Next, for each node $v$ in $\mathcal{G}_m\setminus \{r, {\mu}'_2\}$, $\mathcal{T}_2$ involves link $vw$, where $w$ is a neighbor node of $v$ with $f(w)<f(v)$ and $g(w)\leq g(v)$; \item $\mathcal{T}_3$: First, add link $r{\mu'}_1$ to $\mathcal{T}_3$. Next, for each node $v$ in $\mathcal{G}_m\setminus \{r, {\mu}'_1\}$, $\mathcal{T}_3$ involves link $vw$, where $w$ is a neighbor node of $v$ with $f(w)>f(v)$ and $g(w)\leq g(v)$. \end{itemize} Accordingly, three independent spanning trees $\mathcal{T}_1$, $\mathcal{T}_2$ and $\mathcal{T}_3$ are constructed in $\mathcal{G}^*_{ex}$. \section{Number of Distinct Paths Constructed by STPC} With the constructed three independent spanning trees, let $\mathcal{G}_m:=\mathcal{T}_1\cup \mathcal{T}_2\cup \mathcal{T}_3$. Then each node in $\mathcal{G}_{m}\setminus r$ (virtual nodes ${\mu}'_1$ and ${\mu}'_2$ are also included) has three internally vertex disjoint paths (denoted by $\mathcal{X}^{(1)}$, $\mathcal{X}^{(2)}$ and $\mathcal{X}^{(3)}$) to $r$, each along the corresponding spanning tree. Combining any two of these three paths, we get three cycles, i.e., $\mathcal{C}^{(1)}=\mathcal{X}^{(1)}\cup \mathcal{X}^{(2)}$, $\mathcal{C}^{(2)}=\mathcal{X}^{(2)}\cup \mathcal{X}^{(3)}$ and $\mathcal{C}^{(3)}=\mathcal{X}^{(1)}\cup \mathcal{X}^{(3)}$. Let $C$ be the set of all these cycles. To count the number of distinct paths constructed by STPC, we first count the number of distinct cycles in $C$. Next, removing all virtual links and the resulting isolated nodes in each cycle $\mathcal{C}_i$ of $C$, one monitor-to-monitor simple path $\mathcal{Y}_i$ (if $||\mathcal{Y}_i||\geq 1$) is formed, thus generating a path set ${Y}$. We then study the number of linearly independent paths in ${Y}$ and yield path set ${Y}'$ by removing the linearly dependent paths in ${Y}$. Finally, we show that for each path $\mathcal{Y}'_i$ with more than 2 monitors in ${Y}'$, it can be shortened to a path, denoted by $\mathcal{Z}_i$, with only 2 monitors while retaining the property of full tree link identification in $\mathcal{G}^*_{ex}$. We can prove path set $\{\mathcal{Z}_i\}$ has $||\mathcal{G}_m||-|\mathbb{V}|$ paths ($\mathbb{V}$ is the set of virtual links in $\mathcal{G}_m$), all of which are linearly independent and can cover all paths constructed by STPC for tree link identification in $\mathcal{G}$. Therefore, for tree link identification, the number of distinct paths constructed by STPC is exactly the number of tree links. In the end, we show that the auxiliary algorithm of STPC constructs one path for each non-tree link in $\mathcal{G}$, thus completing the proof of Theorem IV.2 in \cite{LiangPathConstruction12} as a link in $\mathcal{G}$ is either a tree or non-tree link. \subsection{Number of Distinct Cycles} \label{sec_numDisCycl} Let $\mathcal{X}^{(1)}_v$, $\mathcal{X}^{(2)}_v$ and $\mathcal{X}^{(3)}_v$ denote three internally vertex disjoint paths from $v$ ($v\in(\mathcal{G}_m\setminus r)$) to $r$ along the corresponding independent spanning trees. Combining any two of these paths, node $v$ is associated with three cycles, i.e., $\mathcal{C}^{(1)}_v=\mathcal{X}^{(1)}_v\cup \mathcal{X}^{(2)}_v$, $\mathcal{C}^{(2)}_v=\mathcal{X}^{(2)}_v\cup \mathcal{X}^{(3)}_v$ and $\mathcal{C}^{(3)}_v=\mathcal{X}^{(1)}_v\cup \mathcal{X}^{(3)}_v$. Therefore, there are $3(|\mathcal{G}_m|-1)$ cycles, forming cycle set $C$. In this section, we investigate on counting the distinct cycles in $C$. For illustrative purpose, we color the links on $\mathcal{T}_1$ as blue, $\mathcal{T}_2$ as green, and $\mathcal{T}_3$ as red in the sequel. Since the independent trees are constructed wrt $r$, each link in $\mathcal{T}_1$, $\mathcal{T}_2$ and $\mathcal{T}_3$ has a natural direction toward $r$ when selecting $\mathcal{X}^{(i)}_v$ ($i=1,2,3$). Therefore, for the simplicity of the following proofs, each link in $\mathcal{T}_1$, $\mathcal{T}_2$ and $\mathcal{T}_3$ is assigned a direction toward $r$. Note this direction is for and only for the theorem proof and only exists on $\mathcal{T}_1$, $\mathcal{T}_2$ and $\mathcal{T}_3$, meaning the assumption that the original graph is undirected still holds. With this notation, links $l_1$ on $\mathcal{T}_i$ and $l_2$ on $\mathcal{T}_j$ ($i\neq j$) with different colors and different directions might correspond to the same link in $\mathcal{G}_m$. \begin{lemma} The number of distinct cycles in $C$ is the number of links (both real and virtual) in $\mathcal{G}_m$, i.e., $|C|=||\mathcal{G}_m||$. Furthermore, all these distinct cycles are linearly independent. \end{lemma} \begin{proof} In $\mathcal{G}_m$, let $b_1$ be the number of links appearing only in $\mathcal{T}_1$ (colored as blue), and $b_2$ be the number of links appearing in two trees: one is $\mathcal{T}_1$ (colored as blue), the other one is either $\mathcal{T}_2$ (colored as green) or $\mathcal{T}_3$ (colored as red). Then we have: \begin{equation}\label{eq_red_num} b_1 + b_2 = |\mathcal{G}_m|-1, \end{equation} since $\mathcal{T}_1$ is a spanning graph of $\mathcal{G}_m$. Consider ear $\mathcal{E}_i$. Let $\delta_i$ be the number of newly added nodes to $\mathcal{E}_1\cup\cdots\cup \mathcal{E}_{i-1}$ by $\mathcal{E}_i$. Let $v_{0}\cdots v_{\delta_i+1}$ denote the nodes in $\mathcal{E}_i$, where $v_0$ and $v_{\delta_i+1}$ are end-points already existed in previous ears. Then we can derive the relationship among the number of links, the number of nodes and the number of ears in $\mathcal{G}_m$. Suppose there are $n_e$ ears in total, then $\mathcal{E}_1$ with $\delta_1$ new nodes has $\delta_1$ links since $\mathcal{E}_1$ is a cycle, whereas $\mathcal{E}_i$ ($2\leq i \leq {n_e}$) which is a path with $\delta_i$ new nodes has $\delta_i+1$ new links. In addition to these links, there exist $b_1$ links only appearing in the blue tree and not involving in any ears. Hence, we have \begin{equation}\label{eq_linkNodeRelation} \begin{aligned} ||\mathcal{G}_m|| & = b_1 +\delta_1 + \sum^{n_e}_{i=2}(\delta_i+1)\\ &=b_1+\delta_1+\sum^{n_e}_{i=2}\delta_i+\sum^{n_e}_{i=2}1\\ &=b_1+|\mathcal{G}_m|+{n_e}-1. \end{aligned} \end{equation} To count distinct cycles in $C$, we consider merging the cycles obtained from each pair of independent spanning trees. \begin{figure}[tb] \centering \includegraphics[width=1.8in]{CombineRedGreen} \caption{Paths generated by merging red and green path segments.} \label{Fig:CombineRedGreen} \end{figure} (i) We first consider the cycles obtained by combining the red and green paths. For ear $\mathcal{E}_i$, each newly added node in $\mathcal{E}_i$ corresponds to a cycle obtained by combining the red and green paths. However, as Fig.~\ref{Fig:CombineRedGreen} shows, each link between $v_1$ and $v_{\delta_i}$ corresponds to two colors, red and green, with opposite directions. Therefore, cycles formed by combining the red and green paths wrt $v_1\cdots v_{\delta_i+1}$ are identical. Thus, ear $\mathcal{E}_i$ only contributes one distinct cycle when combining the red and green paths for any of the newly added nodes $v_1\cdots v_{\delta_i}$. \begin{figure}[tb] \centering \includegraphics[width=1.8in]{CombineBlueWithOthers} \caption{Paths generated by merging blue and red/green path segments.} \label{Fig:CombineBlueWithOthers} \end{figure} (ii) Next, we consider merging the cycles generated by combining the red and the blue paths. For the $\delta_i$ new nodes in ear $\mathcal{E}_i$, each path formed by combining the red and the blue paths are distinct from each other. However, as Fig.~\ref{Fig:CombineBlueWithOthers} displays, the $red+blue$ cycle wrt $v_1$ might be identical with the $red+blue$ cycle wrt $v_0$ (when $v_0$ was first added in the previous operations). Thus, the same cycle might be counted twice. Let $\varepsilon_i$ indicate if link $v_0v_1$ is colored as blue, i.e., $\varepsilon_i=1$ if true, and $\varepsilon_i=0$ otherwise. With this notation, the number of distinct cycles formed by combining the red and the blue paths is $\delta_i-\varepsilon_i$ for ear $\mathcal{E}_i$ ($2\leq i \leq n_e $), and $\delta_1-1-\varepsilon_1$ for ear $\mathcal{E}_1$. (iii) Finally, we consider merging the cycles formed by combining the green and the blue paths. Following the same argument in (ii), each path formed by combining the green and the blue paths are distinct from the $\delta_i$ new nodes in ear $\mathcal{E}_i$. However, the cycles wrt to $v_{\delta_i}$ and $v_{\delta_i+1}$ might be identical (see Fig.~\ref{Fig:CombineBlueWithOthers}). Let $\varepsilon'_i$ indicate if link $v_{\delta_i}v_{\delta_i+1}$ is colored blue, i.e., $\varepsilon'_i=1$ if true, and $\varepsilon'_i=0$ otherwise. Accordingly, the number of distinct cycles formed by combining the green and the blue paths is $\delta_i-\varepsilon'_i$ for ear $\mathcal{E}_i$ ($2\leq i \leq n_e$), and $\delta_1-1-\varepsilon'_1$ for ear $\mathcal{E}_1$. Therefore, the number of distinct cycles, denoted by $Q_i$, contributed by ear $\mathcal{E}_i$ is: \begin{equation}\label{eq_earContribution} \begin{aligned} Q_1&=1+\delta_1-1-\varepsilon_1+\delta_1-1-\varepsilon'_1,\\ Q_i&=1+\delta_i-\varepsilon_i+\delta_i-\varepsilon'_i\ \ (2\leq i \leq n_e). \end{aligned} \end{equation} Thus, the number of distinct cycles in $C$ is \begin{equation}\label{eq_totalCycles} \begin{aligned} |C|&=Q_1+\sum^{n_e}_{i=2}Q_i\\ &=2\delta_1-\varepsilon_1-\varepsilon'_1-1+2\sum^{n_e}_{i=2}\delta_i+\sum^{n_e}_{i=2}1-\sum^{n_e}_{i=2}(\varepsilon_i+\varepsilon'_i)\\ &=2|\mathcal{G}_m|+n_e-2-\sum^{n_e}_{i=1}(\varepsilon_i+\varepsilon'_i)\\ &=2|\mathcal{G}_m|+n_e-2-b_2 \end{aligned}. \end{equation} Subtracting (\ref{eq_linkNodeRelation}) from (\ref{eq_totalCycles}) and utilizing (\ref{eq_red_num}), we get \begin{equation}\label{eq_concCyclNum} |C|=||\mathcal{G}_m||. \end{equation} Let $\mathcal{G}'_{m}:=\mathcal{G}_m$ except all virtual links/nodes in $\mathcal{G}_m$ are real links/nodes in $\mathcal{G}'_{m}$. Suppose cycle measurement is allowed and $r$ is the only monitor in $\mathcal{G}'_{m}$, then all cycle measurements associated with $C$ is sufficient to identify all links in $\mathcal{G}'_{m}$ (following the same method in STLI, see \cite{LiangPathConstruction12}). Moreover, we have $||\mathcal{G}_m||=||\mathcal{G}'_{m}||$; therefore, all cycles in $C$ are linearly independent. \end{proof} \subsection{Number of Linearly Independent Paths After Removing All Virtual Links} \label{sec_NumLinearIndePaths} In section \ref{sec_numDisCycl}, we get a cycle set $C$ with $||\mathcal{G}_m||$ linearly independent cycles. For any cycle in $C$, removing all involved virtual links and the resulting isolated nodes, we can obtain a monitor-to-monitor simple path since each cycle in $C$ is obtained by combining two internally vertex disjoint paths ($\mathcal{X}^{(i)}$ and $\mathcal{X}^{(j)}$) ($i,j\in\{1,2,3\}$, $i\neq j$) along two independent spanning trees and virtual links only appear at the end of $\mathcal{X}^{(i)}$ and $\mathcal{X}^{(j)}$. Therefore, following the same operation to each cycle in $C$ iteratively, we can generate a new set ${Y}$ with each element $\mathcal{Y}_i$ representing a monitor-to-monitor simple paths. In this section, we investigate on the number of linearly independent paths in ${Y}$. Note that some paths in ${Y}$ might contain more than 2 monitors. We will show how to shorten these paths in the next section. Let $\mathbb{V}$ denote the set of virtual links\footnote{$\mathcal{G}_m$ is a spanning graph of $\mathcal{G}^*_{ex}$. Thus, the number of virtual links in $\mathcal{G}_m$ is less than that in $\mathcal{G}^*_{ex}$.} in $\mathcal{G}_m$. The goal is to prove the number of linearly independent paths in ${Y}$ is $||\mathcal{G}_m||-|\mathbb{V}|$, and these linearly independent paths are sufficient to identify all tree links in the original graph $\mathcal{G}$. For all redundant paths generated by mapping from $C$ to ${Y}$, they can be divided into 3 categories, i.e., 1) trivial topology, 2) linearly dependent paths, and 3) duplicate paths. Note in Fig.~\ref{Fig:TrivialTopo}--Fig.~\ref{Fig:OtherVirtualLinks}, for all paths (or path segments) colored as blue/green/red, they represent virtual links if they are outside $\mathcal{G}$; and real simple paths otherwise. \begin{figure}[tb] \centering \includegraphics[width=1.5in]{TrivialTopo} \caption{Combining red and green paths wrt ${{\mu}_1}$.} \label{Fig:TrivialTopo} \end{figure} 1) Trivial topology (a graph with no nodes or links). As shown in Fig.~\ref{Fig:TrivialTopo}, combining the red and green paths wrt ${\mu}_1$ ($\mathcal{G}_m$ involves $r{\mu'}_1{\mu}_1{\mu}'_2r$), cycle $r{\mu'}_1{\mu}_1{\mu}'_2r$ is formed. However, the corresponding path of $r{\mu'}_1{\mu}_1{\mu}'_2r$ after removing the virtual links and the resulting isolated nodes is empty. Therefore, this trivial path has no contribution for identifying real links in $\mathcal{G}$. 2) Linearly dependent paths. The following 3 cases can generate redundant linearly dependent paths. \begin{figure}[tb] \centering \includegraphics[width=1.5in]{LinearDepPaths1} \caption{Path construction wrt ${{\mu}_2}$.} \label{Fig:LinearDepPaths1} \end{figure} (i) Consider the three cycles constructed wrt ${\mu}_2$ (shown in Fig.~\ref{Fig:LinearDepPaths1}). The paths associated with $blue+red$ cycle and $blue+green$ cycle are ${\mu}_ie_1{\mu}_2$ and ${\mu}_2e_2{\mu}_j$, respectively. However, the path associated $red+green$ cycle, ${\mu}_ie_1{\mu}_2e_2{\mu}_j$, is the sum of previously constructed paths ${\mu}_ie_1{\mu}_2$ and ${\mu}_2e_2{\mu}_j$. Therefore, redundant path ${\mu}_ie_1{\mu}_2e_2{\mu}_j$ is linearly dependent with the other two. \begin{figure}[tb] \centering \includegraphics[width=1.8in]{LinearDepPaths2} \caption{Path construction wrt ${{\mu}_a}$.} \label{Fig:LinearDepPaths2} \end{figure} (ii) Now consider the case shown in Fig.~\ref{Fig:LinearDepPaths2}. For $\mathcal{G}_m$, there exists one and only one monitor ${\mu}_a$ with link ${\mu}'_2{\mu}_a$ colored as blue in the ${\mu}'_2\rightarrow {\mu}_a$ direction, according to the rules for blue tree construction. Now consider ${\mu}_a$. According to the s-t numbering rule and the processing of ear decomposition, we have $f({\mu}'_2)<f({\mu}_a)$ and $g({\mu}'_2)<g({\mu}_a)$; therefore, ${\mu}'_2{\mu}_a$ is colored\footnote{Note ${\mu}_a$ might have more than one neighbor $w$ with $f(w)<f({\mu}_a)$ and $g(w)<g({\mu}_a)$, in which case we can still choose ${\mu}'_2{\mu}_a$ to color it as green.} as green in the other direction. Then ${\mu}_2e_1{\mu}_a$ and ${\mu}_ie_2{\mu}_a$ are two paths obtained from $green+blue$ cycle and $green+red$ cycle wrt ${\mu}_a$, respectively. However, the path generated by the $blue+red$ cycle is the sum of ${\mu}_2e_1{\mu}_a$ and ${\mu}_ie_2{\mu}_a$, thus redundant in ${Y}$. \begin{figure}[t] \centering \includegraphics[width=1.8in]{LinearDepPaths3} \caption{Path construction wrt ${{\mu}_b}$.} \label{Fig:LinearDepPaths3} \end{figure} (iii) Analogously, there exists one and only one monitor ${\mu}_b$ with link ${\mu}'_1{\mu}_b$ colored as blue in the ${\mu}'_1\rightarrow {\mu}_b$ direction (see Fig.~\ref{Fig:LinearDepPaths3}). Based on the nonseparating ear decomposition, we have $f({\mu}'_1)>f({\mu}_b)$ and $g({\mu}'_1)<g({\mu}_b)$; therefore, ${\mu}'_1{\mu}_b$ is colored\footnote{Note ${\mu}_b$ might have more than one neighbor $u$ with $f(u)>f({\mu}_a)$ and $g(u)<g({\mu}_a)$, in which case we can still choose ${\mu}'_1{\mu}_b$ to color it as red.} as red in the other direction. Then ${\mu}_be_1{\mu}_2$ and ${\mu}_be_2{\mu}_j$ are two paths obtained from $red+blue$ cycle and $red+green$ cycle wrt ${\mu}_b$, respectively. However, the path generated by the $blue+green$ cycle is the sum of ${\mu}_be_1{\mu}_2$ and ${\mu}_be_2{\mu}_j$, thus not providing new information for link identifications in $\mathcal{G}$. 3) Duplicate paths. Each of the following 3 cases can generate the same path twice. \begin{figure}[tb] \centering \includegraphics[width=1.8in]{DupPath12} \caption{Duplicate paths wrt ${{\mu}_1}$ and ${\mu}'_2$.} \label{Fig:DupPath12} \end{figure} (i) We have considered combining the red and green paths wrt ${\mu}_1$. Now consider the paths associated with the $blue+red$ cycle and $blue+green$ cycle. As Fig.~\ref{Fig:DupPath12} displays, the paths associated with these two cycles are exactly the same, i.e., path ${\mu}_1e_1{\mu}_2$ is generated twice. (ii) Following the similar argument, the paths associated with the $blue+red$ cycle and $blue+green$ cycle wrt ${\mu}'_2$ are also the same, i.e., path ${\mu}_2e_2{\mu}_a$ (shown in Fig.~\ref{Fig:DupPath12}) is generated twice when mapping from $C$ to ${Y}$. Note ${\mu}_a$ cannot be the same as ${\mu}_2$ since ${\mu}_2$ only appears in the last ear. Therefore, path ${\mu}_2e_2{\mu}_a$ is non-trivial. \begin{figure}[tb] \centering \includegraphics[width=1.8in]{DupPath3} \caption{Duplicate paths wrt ${\mu}'_1$.} \label{Fig:DupPath3} \end{figure} (iii) Similar to (i) and (ii), the paths associated with the $blue+red$ cycle and $blue+green$ cycle wrt ${\mu}'_1$ are also the same, i.e., ${\mu}_be_1{\mu}_2$ as shown in Fig.~\ref{Fig:DupPath3} appears twice when mapping from $C$ to ${Y}$. In addition, ${\mu}_be_1{\mu}_2$ is non-trivial since ${\mu}_2$ has the highest ear level which means ${\mu}_b\neq {\mu}_2$. \begin{figure}[tb] \centering \includegraphics[width=3.3in]{OtherVirtualLinks} \caption{Other virtual links connecting to ${\mu}'_1$ or ${\mu}'_2$.} \label{Fig:OtherVirtualLinks} \end{figure} 4) We have discussed 7 cases in 1)--3), each case providing a redundant path when mapping from $C$ to ${Y}$. For the spanning graph $\mathcal{G}_m$, the minimum number of virtual links in $\mathcal{G}_m$ is also 7, i.e., $r{\mu'}_1$, $r{\mu'}_2$, ${\mu}'_1{\mu}_1$, ${\mu}'_2{\mu}_1$, ${\mu}'_2{\mu}_a$ (Fig.~\ref{Fig:LinearDepPaths2}), and ${\mu}'_1{\mu}_b$ (Fig.~\ref{Fig:LinearDepPaths3}). However, in addition to these 7 virtual links, there might exist other virtual links in $\mathcal{G}_m$. As Fig.~\ref{Fig:OtherVirtualLinks} shows, suppose there are other $N_1$ virtual links (colored as green and no color on the other direction as ${\mu}'_1{\mu}_b$ has been colored as blue) connecting ${\mu}'_2$ and ${\mu}_{gi}$ ($i=1,\cdots,N_1$) and $N_2$ virtual links (colored as red and no color on the other direction as ${\mu}'_1{\mu}_b$ has been colored as blue) connecting ${\mu}'_1$ and ${\mu}_{ri}$ ($i=1,\cdots,N_2$). For the 3 paths constructed wrt each node in $\{{\mu}_{g1},\cdots,{\mu}_{gN_1},{\mu}_{r1},\cdots,{\mu}_{rN_2}\}$, there exists one path which is linearly dependent with the other two. To prove this claim, consider ${\mu}_{g1}$ as an example. Two paths\footnote{Note ${\mu}_c$ can be any node in $\{{\mu}_{r1},\cdots,{\mu}_{rN_2},{\mu}_b\}$ in Fig.~\ref{Fig:OtherVirtualLinks}.} ${\mu}_2e_1{\mu}_{g1}$ and ${\mu}_ce_2{\mu}_{g1}$ (see Fig.~\ref{Fig:OtherVirtualLinks}) can be obtained from the $green+blue$ cycle and $green+red$ cycle wrt ${\mu}_{g1}$. However, the path associated with the $blue+red$ cycle wrt ${\mu}_{g1}$ is the sum of ${\mu}_2e_1{\mu}_{g1}$ and ${\mu}_ce_2{\mu}_{g1}$, thus linearly dependent with the other two. The same argument can be applied to other nodes in $\{{\mu}_{g1},\cdots,{\mu}_{gN_1},{\mu}_{r1},\cdots,{\mu}_{rN_2}\}$. Therefore, we can identify another $N_1+N_2$ linearly dependent paths in ${Y}$ regarding these $N_1+N_2$ virtual links. In sum, 1)--4) cover all the possible cases of redundant paths when mapping from $C$ to ${Y}$ and the number of these redundant paths is $7+N_1+N_2$, which is also the number of virtual links (denoted by $|\mathbb{V}|$) in $\mathcal{G}_m$. Therefore, removing these redundant paths, the cardinality of the resulting path set ${Y}'$ is $||\mathcal{G}_m||-(7+N_1+N_2)=||\mathcal{G}_m||-|\mathbb{V}|$. Now we can explore if ${Y}'$ is sufficient to identify all tree links in $\mathcal{G}^*_{ex}$. Recall that each path (for identifying tree links) constructed by STPC consists of one or two path segments, each path segment terminating at the first monitor it encounters. If we let each segment terminate at the last monitor before traversing a virtual link and combine any two segments wrt real node $v$ ($v$ can be either monitor or non-monitor), then a new measurement path set $J$ is formed. It is easy to show (using STLI) that this new path set $J$ is sufficient to identify all tree links in $\mathcal{G}^*_{ex}$. Recall the mapping process from cycle set $C$ to path set ${Y}$. Each path in ${Y}$ is obtained by removing the virtual links and all resulting isolated nodes of the corresponding cycle in $C$. Then each path $\mathcal{P}_J$ in $J$ can be obtained by applying the same operation to the corresponding cycle $C_J$ ($C_J\supset \mathcal{P}_J$) in $C$; therefore, each path in $J$ must be involved as one element in ${Y}$. Accordingly, ${Y}$ is sufficient to identify all tree links in $\mathcal{G}^*_{ex}$. Since ${Y}'$ is obtained by removing redundant paths in the procedure of mapping $C$ to ${Y}$, then ${Y}'$ is also sufficient to identify all tree links in $\mathcal{G}^*_{ex}$, i.e., the measurement matrices associated with $J$, ${Y}$ and ${Y}'$ have the same column rank. In sum, for graph $\mathcal{G}_m$ with $||\mathcal{G}_m||-|\mathbb{V}|$ real links, the property of ${Y}'$ is that it contains exactly $||\mathcal{G}_m||-|\mathbb{V}|$ paths, with each path containing only tree links in $\mathcal{G}^*_{ex}$. Therefore, the measurement matrix associated with ${Y}'$ is a square matrix with full rank. \subsection{Number of Distinct Paths by STPC} \begin{theorem}\label{thm:number of distinct path} The number of distinct paths constructed by STPC equals $n$, the number of links in $\mathcal{G}$. \end{theorem} \begin{proof} We have proved that the corresponding measurement matrix of ${Y}'$ is square and sufficient to identify all tree links in $\mathcal{G}^*_{ex}$. In this section, we prove that all paths with more than 2 monitors in ${Y}'$ can be shortened to form a new path set ${Z}$ which can cover all the paths obtained by STPC for tree link identification, i.e., paths obtained by line 1--11 in STPC algorithm. 1) First, we consider the number of generated paths wrt to node $v$ (line 5--9 in STPC). (i) If $v$ is a monitor in $\{{\mu}_2,{\mu}_a,{\mu}_b,{\mu}_{g1},\cdots,{\mu}_{gN_1},{\mu}_{r1},\cdots,{\mu}_{rN_2}\}$ (see Fig.~\ref{Fig:OtherVirtualLinks}), then one path in $\{\mathcal{P}_{v1},\mathcal{P}_{v2},\mathcal{P}_{v3}\}$ contains no links, resulting to be an invalid path and discarded without appending to $\mathbf{R}$ in line 10. In this case, only two paths are generated in line 6. In 2) and 4) of Section \ref{sec_NumLinearIndePaths}, we have shown that one linearly dependent path, which is the combination of the other two paths wrt the same node, has been removed, thus not existed in ${Y}'$. Therefore, in ${Y}'$, there are also two paths wrt to nodes belonging to $\{{\mu}_2,{\mu}_a,{\mu}_b,{\mu}_{g1},\cdots,{\mu}_{gN_1},{\mu}_{r1},\cdots,{\mu}_{rN_2}\}$. (ii) If $v$ is ${\mu}_1$, then line 6 of STPC only generates one path, since the other two paths contain only one node, i.e., ${\mu}_1$ itself. In this case, we have also shown that only path (${\mu}_1e_1{\mu}_2$ colored as blue in Fig.~\ref{Fig:DupPath12}) wrt ${\mu}_1$ is retained in ${Y}'$ after removing one invalid (Section \ref{sec_NumLinearIndePaths}-1)) and one duplicate (Section \ref{sec_NumLinearIndePaths}-3)-(i)) path. (iii) Apart from the above two cases, each node in $\mathcal{G}$ corresponds to 3 measurement paths both in ${Y}'$ and $\mathcal{P}_{STPC}$, where $\mathcal{P}_{STPC}$ is the path set by STPC after executing line 1--11. \begin{figure}[tb] \centering \includegraphics[width=2.3in]{LongShortPaths} \caption{Shortening a long path in ${Y}'$.} \label{Fig:LongShortPaths} \end{figure} 2) Second, we compare the differences between the two path segments obtained by the two methods wrt to node $v$ on the same tree $\mathcal{T}_i$. Let $\mathcal{S}_a$ and $\mathcal{S}_b$ denote the corresponding path segments associated with ${Y}'$ and $\mathcal{P}_{STPC}$, respectively. Both $\mathcal{S}_a$ and $\mathcal{S}_b$ start at $v$ and go toward $r$ along the same tree $\mathcal{T}_i$. The only difference is that $\mathcal{S}_a$ terminates at the last monitor before traversing a virtual link and $\mathcal{S}_b$ terminates at the first monitor it encounters (as shown in Fig.~\ref{Fig:LongShortPaths}). In Fig.~\ref{Fig:LongShortPaths}, observe that wrt ${\mu}_{first}$, $\mathcal{S}_c$ is constructed and $W_{\mathcal{S}_c}$ can be computed through measuring the paths constructed wrt ${\mu}_{first}$ in ${Y}'$. Therefore, when constructing path segment $\mathcal{S}_a$ wrt $v$, there is no need to terminate at ${\mu}_{last}$. $\mathcal{S}_a$ can also terminate at ${\mu}_{first}$ since $W_{\mathcal{S}_c}$ can be known from other path measurements directly (there are only two monitors on $\mathcal{S}_c$) or indirectly (in Fig.~\ref{Fig:LongShortPaths}, $W_{{{\mu}_i\rightarrow {\mu}_{last}}}$ and $W_{{{\mu}_{first}\rightarrow {\mu}_{i}}}$ are obtained from paths constructed wrt ${\mu}_i$ and ${\mu}_{first}$, respectively). Using this operation, each path with more than 2 monitors in ${Y}'$ can be shortened. 3) Let ${Y}''$ denote the path set obtained from ${Y}'$ after the above shortening process. We can show in set ${Y}''$, for a monitor ${\mu}^*$ which is not in $\{{\mu}_2,{\mu}_a,{\mu}_b,{\mu}_{g1},\cdots,{\mu}_{gN_1},{\mu}_{r1},\cdots,{\mu}_{rN_2}\}$, the corresponding measurement paths can be further shortened. Let $\mathcal{S}^*_1$, $\mathcal{S}^*_2$ and $\mathcal{S}^*_3$ denote the corresponding shortened path segments wrt ${\mu}^*$, then the associated measurement paths in ${Y}''$ are $\mathcal{S}^*_1\cup \mathcal{S}^*_2$, $\mathcal{S}^*_2\cup \mathcal{S}^*_3$ and $\mathcal{S}^*_3\cup \mathcal{S}^*_1$, which can be further shortened to $\mathcal{S}^*_1$, $\mathcal{S}^*_2$ and $\mathcal{S}^*_3$, respectively, forming the same 3 paths as those obtained by STPC. Suppose the newly formed set from ${Y}''$ by path shortening is ${Z}$, then ${Z}$ is also sufficient to identify all tree links in $\mathcal{G}_m$. This is because, from the perspective of linear algebra, this shortening operation means that one part of the original linear equation can be eliminated by subtracting the linear combinations of some other rows; therefore, the resulting matrix rank is the same as the original matrix. Since we have proved that the measurement matrix associated with ${Y}'$ has a full rank, the measurement matrix associated with ${Z}$ also has a full rank. To identify tree links in $\mathcal{G}^*_{ex}$, based on the above three arguments, we know that the number of paths constructed wrt $v$ are the same in both ${Y}'$ and $\mathcal{P}_{STPC}$. Moreover, wrt the same node and along the same independent spanning tree(s), the paths obtained by these two methods are exactly the same after the path shortening operation. Thus, the newly generated path set ${Z}$ can cover all the paths selected by STPC for tree link identification. The number of paths in ${Z}$ is the number of tree links in $\mathcal{G}^*_{ex}$; therefore, the number of distinct paths constructed by line 1--11 in STPC is exactly the number of tree links in $\mathcal{G}^*_{ex}$. Finally, we consider the non-tree links of the original graph $\mathcal{G}$. As $\mathcal{G}_m=\mathcal{T}_1 \cup \mathcal{T}_2 \cup \mathcal{T}_3$, all non-tree links are involved in set $L(\mathcal{G}\setminus \mathcal{G}_m)$. With the knowledge of tree link metrics, based on the auxiliary algorithm of STPC for non-tree link identification, in line 12--15, each non-tree link $l$ (with link metric $W_l$) corresponds to one new path measurement involving $W_l$ as the only unknown link metric. Therefore, the number of newly added paths by line 12--15 equals the number of non-tree links in $\mathcal{G}$ and they are linearly independent with all other selected paths. A link in $\mathcal{G}$ is either a tree or non-tree link; therefore, the number of distinct paths constructed by STPC equals the number of links in $\mathcal{G}$. \end{proof} \section{Performance Evaluations on Random Graphs} Besides the main simulation results in \cite{LiangPathConstruction12}, we also simulate random graphs with a different number of links and observe similar results in Table \ref{t sparse}. \begin{table*}[tb] \renewcommand{\arraystretch}{1.3} \caption{Sparsely-Connected Random Graphs (ER: $p=0.0390$, RG: $d_c=0.11943$, BA: $\varrho=3$, $I_{\mbox{\tiny max}}=3\times n$)} \label{t sparse} \vspace{-.5em} \centering \begin{tabular}{c|c|c|c|c|c|c|c|c|c|c|c} \hline graph & $\overline{n}$ & ${m}$ & $\overline{\kappa}$ & $r_{\mbox{\small succ}}$ & ${\Upsilon}$ & $\tSTPC$ (s)& $\tRWPC$ (s)& $\tSTLI$ (ms)& $\tMILI$ (ms)& $\hSTPC$& $\hRWPC$ \\ \hline ER & 438.48 & 150 & 9.39 & 99.00\% & 99.76\% & 10.21 & 38.73 & 5.27 & 17.8 & 17.94 & 14.2 \\ \hline RG & 449.02 & 150 & 14.96 & 4.00\% & 93.91\% & 10.62 & 109.21 & 5.10 & 23.34 & 22.48 & 14.43 \\ \hline BA & 441 & 150 & 3 & 72.00\% & 99.70\% & 8.41 & 90.48 & 5.39 & 20.45 & 15.08 & 8.85 \\ \hline \end{tabular} \vspace{-0mm} \end{table*} \bibliographystyle{IEEEtran}

6c00d7b2ae4dc6aad93a342456653a7ec2f29ff0

\section{Experimental Methods} The experiments of freshwater solidification in a convective cell with different inclination angles (here we use the coordinate system attached to the cell) are conducted in a classical Rayleigh-B\'enard (RB) convection system (see Fig.~\ref{SM_expsetup}). Fig.~\ref{SM_expsetup}(a) shows the sketch of the experimental cell. The experimental cell consists of a top cooling plate (with temperature $T_t$), a heating bottom plate (with temperature $T_b$), and plexiglas sidewalls with height H = 240 mm (length $L_x =240$ mm and width $L_z =60$ mm, i.e., aspect ratio $\Gamma = L_x/H = $1.0). The working fluid, which is confined between the top and bottom plates, is deionized and ultrapure water, and the water density and thermal expansion coefficient around the density peak temperature, $T_c=$4$^\circ$C, are shown in Fig.~\ref{SM_expsetup}(d) and (e). Before conducting any experiments, water is boiled twice to be degassed. The temperature of both the top and bottom plates are well controlled at constant temperatures ($T_t < 0^\circ$C, and $T_b > 0^\circ$C) by the circulating bath (PolyScience PP15R-40), with the temperature fluctuations less than $\pm0.2$K. Silicone O-rings are attached in between the top plate and the sidewall, and also in between the sidewall and the bottom plate, to seal the cell. In order to compensate the volume change during the solidification of freshwater, an expansion vessel is connected to the experimental cell through a rubber tube, which is open to the atmosphere so that the pressure of the experimental cell remains unchanged. Six resistance thermistors (44000 series thermistor element, see the sketch in Fig.~\ref{SM_expsetup}(c)) are embedded into the top and bottom plates, respectively (see the sketch in Fig.~\ref{SM_expsetup}(a), the black dots on the top and bottom plates show the placement of the thermistors). There are two steps to avoid the heat exchange between the experimental cell and the surrounding environment, 1) the experimental cell is wrapped in a sandwich structure: insulation foam, aluminum plate (helping to support and acting as the temperature measuring spot of the temperature sensor, PT 100, of the Proportional-Integral-Derivative (PID) system), and insulation foam; 2) a PID (Proportional-Integral-Derivative) controller is used to control the temperature of the surrounding environment outside the experimental cell (see Fig.~\ref{SM_expsetup}(b)). \begin{figure*}[!htb] \centering \includegraphics[width=0.8\linewidth]{SM_expsetup.pdf} \caption{(a) Sketch of the experimental system for Rayleigh-Bénard convection coupled with solidification of fresh water under different inclination angles. (b)The PID (Proportional-Integral-Derivative) controller and the temperature sensor used in the PID system. (c) The sketch of resistance thermistor, 44000 series thermistor element, which is used to measure the top and bottom plates temperature. (d) The nonmonotonic relationship of the water density with temperature for cold water near $T_c$ from Ref. \cite{Gebhart1977A}. (e) the thermal expansion coefficient of water as a function of the temperature.} \label{SM_expsetup} \end{figure*} During the experiments and numerical simulations, in order to focus on the effect of inclination, we have limited the cold plate to a constant undercooling temperature of $T_t = -10^\circ$C, which is a typical value in winter. The effects of changing the undercooling temperature, $T_t$, are qualitatively predictable and are not expected to change the occurrence of different forms of the ice front morphology upon increasing the inclination angle, $\beta$, as well as its physical explanation (by the boundary layer theory based model). The only effects that $T_t$ will bring are that, 1) the ice thickness at the equilibrium state of the system is thicker (thinner) with decreasing (increasing) $T_t$; 2) the corresponding ice front morphology is similar in the shape compared with that of the current study but will be different in the extension and the local curvature. We have conducted several typical sets of experiments (we only perform the typical cases of the experiments and did not conduct every experiment corresponding to the simulations because the experiments are time consuming): 1) for investigation of the effects of different heating condition, we have performed $T_b = 6^\circ$C, $10^\circ$C and $12^\circ$C, at the system inclination angle of $\beta=90$; 2) for the investigation of the influence of the cell inclination, we have done several different inclination angles, $\beta$ (unit: deg), i.e., $\beta=0$, 50, 90, 180, at the hot plate temperature $T_b =10^\circ$C; In the experiments, the ice forms on the cold plate and grows in thickness until the system reaches a statistical equilibrium state. The statistical equilibrium state is reached when the standard deviation of the ice thickness time series is less than 0.5\%. Changing the heating condition, $T_b$ can adjust the intensity of the two counterrotating convective rolls. Recall that the water density reaches the maximum, $\rho_c$, at $T_c$ (which is approximately $4^\circ$). Then depending on the value of the hot plate temperature, $T_b$, the flow configuration can be classified into two regimes: 1) When $4^\circ \text{C}< T_b \leq 8^\circ \text{C}$, the anticlockwise convective roll (with the temperature ranging from $T_c$ to $T_b$ and the corresponding thermal driving $T_b-T_c \leq 4$K) has weaker intensity than that of the clockwise (with the temperature ranging from $T_0$ to $T_c$ and the corresponding thermal driving $T_c-T_0 \approx 4$K). In this regime, the whole ice front is shielded by the colder clockwise roll and is away from the influence of the hot plumes from the hot plate. So the ice front is flat with a tilting angle. 2) When $T_b > 8^\circ \text{C}$, the anticlockwise convective roll (with the temperature ranging from $T_c$ to $T_b$ and the corresponding thermal driving $T_b-T_c > 4$K) has stronger intensity than that of the clockwise (with the temperature ranging from $T_0$ to $T_c$ and the corresponding thermal driving $T_c-T_0 \approx 4$K). So the anticlockwise roll penetrates the clockwise roll and finally affects the ice front morphology, which represents thinner in thickness with higher $T_b$ but similar shape. With the knowledge of these, we later limited our investigation to the case with $T_b =10^\circ$C and performed more experimental investigation under different inclination angles (see Fig.~\ref{SM_moreexppic}). \begin{figure*}[!htb] \centering \includegraphics[width=0.6\linewidth]{SM_moreexppic_v2.pdf} \caption{More results of the experiments ((a), (c), (e)) compared with the simulations (with properly considering the water density anomaly) ((b), (d), (f)). Parameter chosen: (a) and (b): $T_b = 10^\circ$C and $\beta=0$; (c) and (d): $T_b = 10^\circ$C and $\beta=50$; (e) and (f): $T_b = 10^\circ$C and $\beta=180$. The unit for $\beta$ is deg. The blue dashed line in (a), (c), and (e) shows the ice front position which is guide to the eye. (b), (d), and (f) are the temperature field overlapped with isotherms as well as the velocity vectors. The ice front is represented by the black thick line and the $T_c$-isotherm is represented by the red thick line. The blue shaded area is the ice layer.} \label{SM_moreexppic} \end{figure*} Fig.~\ref{SM_moreexppic} reports more experimental results with the parameter $T_b=10^\circ$C and $T_t=-10^\circ$C, with inclination angle, $\beta=0$(panels (a) and (b)), $\beta=50$(panels (c) and (d)), and $\beta=180$(panels (e) and (f)). Fig.~\ref{SM_moreexppic}(a), (c), and (e) are from experiments, and Fig.~\ref{SM_moreexppic}(b), (d), and (f) are from the corresponding simulations. From Fig.~\ref{SM_moreexppic}, we can conclude that the results from the experiments and numerical simulations under different inclination angles agree well with each other. {\color{black} \section{Numerical Methods} The simulations are performed by means of the \textsc{CH4-PROJECT} code \cite{calzavarini2019eulerian}, which adopts a Lattice-Boltzmann algorithm for the description of fluid and temperature dynamics, and an enthalpy method for the ice evolution. The method has been validated against experiments \cite{wang2020growth}, while the code has been intensively tested in \cite{Esfahani2018Basal}. In the simulation, we use the Boussinesq approximation, which means that the density is regarded as a constant value except for that in the buoyancy term in the momentum equation. The non-monotonic relationship of the water density as a function of the temperature is $\rho_\text{w} = \rho_c(1-\alpha^*|T_b-4|^q )$ \cite{Gebhart1977A}, with $\rho_c = 999.972~kg/m^3$ the maximum density corresponding to $T_c \approx 4^\circ$C, $\alpha^* = 9.30\times10^{-6} (K^{-q})$, and $q=1.895$. All the physical properties of water and ice phase, except for $\rho_w$ in the buoyancy term are evaluated at the mean temperatures in each phase which are $(T_\text{b}+0)/2$ and $(T_\text{t}+0)/2$, respectively. In the simulations we neglect the microscopic physics leading to kinetic undercooling, Gibbs-Thomson effect and the anisotropic growth/melting \cite{dash2006physics}. Furthermore, we assume the ice and water density remain the same, i.e., $\rho_I = \rho_w$, to satisfy the incompressibility of the flow in the simulation. The relevant equations that govern the fluid flow and the boundary conditions are reported here below (the energy equation will be presented later), \begin{equation} \begin{split} &\vec{\nabla} \cdot \vec{u} = 0,\\ & \frac{\partial \vec{u}}{\partial t} + \vec{u}\cdot \vec{\nabla} \vec{u} = -\frac{\vec{\nabla} p}{\rho_\text{c}} +\nu_w \nabla^2 \vec{u} +\alpha^*g|T-4|^q \hat e_z, \end{split} \label{governing_eqn} \end{equation} where $\vec{u}(x,y,t)$, $p(x,y,t)$ are fluid velocity, pressure, respectively and we denote with $x$ as the horizontal direction and with $y$ the vertical direction; $\nu_w$, $\rho$, $g$ are the kinematic viscosity of water, the density, and the acceleration of gravity, respectively. $\hat e_z$ is the unit vector pointing in the direction opposite to that of gravity. \begin{figure*}[t!hb] \centering \includegraphics[width=0.6\linewidth]{sketch.pdf} \caption{Schematic description of the problem considered. The blue shaded region corresponds to the solid phase (ice) and the red shaded region to the liquid phase (water). $T_0$ is the melting temperature. The top plate is cooled at constant temperature $T_t$, and the bottom plate is heated at the constant temperature $T_b$. The ice-water interface is fixed at $T_0$ and is of no slip condition.The two side walls are of adiabatic conditions.} \label{sketch} \end{figure*} The boundary conditions corresponding to the governing equations above are isothermal at the top and bottom plates, no-slip at the bottom plate and at the ice-water interface, adiabatic at the lateral boundaries, and no-slip and freezing (namely, Stefan condition \cite{ bodenschatz2000recent}) at the phase-changing interface (see also figure \ref{sketch}). The boundary conditions read: \begin{equation} \begin{split} & T(x,0,t) = T_\text{b},\\ & T(x,1,t) = T_\text{t},\\ & \vec{u}(x,0,t) = 0,\\ & \vec{u}(x,1-h_\text{i}(x,t),t) = 0,\\ &\frac{\partial {T(x,y,t)}}{\partial x}|_{x=0} = 0,\\ &\frac{\partial {T(x,y,t)}}{\partial x}|_{x=L_x/H} = 0,\\ &L \rho_\text{i} V_n = \vec{n} \cdot \vec{q}_\text{w} - \vec{n}\cdot \vec{q}_{\text{i}}, \end{split} \label{governing_eqn_bcs} \end{equation} where $L$ is the latent heat, $V_n$ is the normal speed of the ice-water interface, and $h_i(x,t)$ the dimensionless ice layer thickness at the position $x$, $\vec{q}$ is the heat flux vector, $\vec{n}$ is a unit normal at the ice-water interface pointing into the liquid. The subscripts I and w refer to the ice and the water, respectively. The heat flux reads $q_\text{i} = -k_\text{i} \vec{\nabla} T_\text{i}$ and $q_\text{w} = -k_\text{w} \vec{\nabla} T_\text{w}$. The boundary condition at the ice-water interface requires particular care due to its time and space dependent character. So an useful method is to separate the total enthalpy $h$ into sensible heat and latent heat \cite{Moritz2019An}: \begin{equation} \text{h} = \left\{ \begin{split} &L\phi_w + \text{C}_\text{pi}T, ~~~~~~~~~~~~~~~~~~~~~~~~~~~when~ T<T_0,\\ &L\phi_w + \text{C}_\text{pi}T_0, ~~~~~~~~~~~~~~~~~~~~~~~~~~when ~T=T_0,\\ &L\phi_w + \text{C}_\text{pi}T_0 + \text{C}_\text{pw}(T-T_0), ~~~~~when ~T>T_0. \end{split} \right. \label{enthalpy} \end{equation} where $T_0$ is the phase change temperature ($T_0=0$), and $\phi_w(x,y,t)$ is the liquid fraction in the system and the relation between $h_i(x,t)$ and $\phi_w(x,y,t)$ is $h_i(x,t) = 1 - \int_{0}^{1} \phi_w(x,y,t)\, dy$. In the ice phase, $\phi_w=0$, and in the water phase, $\phi_w=1$, which leads to an additional source term, $S_1$ (the formulation is shown later), from the latent heat contribution at the ice-water interface in the energy conservation equation. We use the Lattice Boltzmann method (LBM) which is able to capture the turbulent convective dynamics in the water phase and also describe the phase change process at the ice-water interface. The basic principle and formulation of the method are described in \cite{succi2001lattice}, while the specific applications to phase-change have been presented in \cite{huber2008lattice,Esfahani2018Basal}. It is noteworthy that the key to accurately solve such problems is to recover the diffusion term in the energy conservation equation exactly and, similarly to \cite{chen2017a}, we implement the correction when the investigated domain consists of heterogeneous media which lead to another additional source term, $S_2$, in the energy conservation equation. So the energy equation with consideration of two source terms $S_1$ and $S_2$ reads \begin{equation} \sigma (\rho \text{C}_\text{p})_0 \frac{\partial T}{\partial t} + \vec{ \nabla} \cdot (\sigma (\rho \text{C}_\text{p})_0 T\vec{u}) = \vec{\nabla} \cdot (k \vec{\nabla} T) +S_1+S_2, \label{source_term} \end{equation} where $T(x,y,t)$ is the temperature fields (all temperatures are measured in Celsius), and $k$ is the thermal conductivity. When it is in water phase $k = k_{w}$, $\text{C}_\text{p}= \text{C}_\text{pw}$, and when it is in ice phase $k = k_{i}$, $\text{C}_\text{p}= \text{C}_\text{pi}$. The first source term is $S_1=- L\rho\frac{\partial \phi_w}{\partial t}$ and the second source term $S_2=-\sigma k \vec{\nabla} T \vec{\nabla} \frac{1}{\sigma} - \frac{\rho \text{C}_\text{p}}{\sigma} T \vec{u} \vec{\nabla} \sigma$. Here $\sigma = \frac{\rho \text{C}_\text{p}}{(\rho \text{C}_\text{p})_0}$ is the ratio of heat capacitance (which is variable and depends on the type of phase, i.e., ice or water) and $(\rho \text{C}_\text{p})_0$ is reference heat capacitance which is taken as constant \cite{chen2017a}. In this study we focus on the morphology of the ice front at the statistical equilibrium state. In order to speed up the process of reaching an equilibrium state, all the simulations are started with a thin layer of flat ice. \section{Systematic investigation of various inclination angles} We have conducted a systematical exploration of how $\beta$ (unit:deg) affects the ice morphology, ranging from $\beta = 0$ to $\beta = 180$ (here we use the coordinate system attached to the cell). Fig.~\ref{SM_systematic_angle} reports the time sequences of the temperature field under different inclination angles, $\beta$ with the parameters $T_t = -10^\circ$C. We focus on the morphology of the ice front at the statistical equilibrium state. In order to speed up the process of reaching an equilibrium state, all the simulations are started with a thin layer of flat ice. The boundary conditions are no-slip for the velocity, adiabatic at the sidewalls, and constant temperatures at the top and bottom plates. The initial condition is still fluid at linear temperature profile from $T_0$ to $T_b$. We assume thermophysical properties to be constant except for the density in the buoyancy term. The real water density property near to the density peak temperature, $T_c$, is well described with the equation $\rho(T)=\rho_c(1-\alpha^*|T-T_\text{c}|^q )$, with $\rho_c = 999.972~kg/m^3$ the maximum density corresponding to $T_c \approx 4^\circ$C, $\alpha^* = 9.30\times10^{-6} (K^{-q})$, and $q=1.895$ This form of equation gives the maximum density, $\rho_{c} = 999.972 kg/m^3$, at the density peak temperature $T_c$ \cite{Gebhart1977A}. \begin{figure*}[!htb] \centering \includegraphics[width=0.8\linewidth]{SM_systematic_angle.pdf} \caption{The systematic investigation of the influence of the inclination angle on the ice front morphology. The snapshots at the internal, middle, and outside circles are the temperature field 10min, 2.5hr, and 2.5d (when the system has reached the statistical equilibrium state) after the start of the simulations, respectively. The black thin lines are the isotherms. The red dashed line shows the ice front, and the red thick line shows the $T_c$-isotherm. The sketch in the middle shows the coordinate system (attached to the system) and how the inclination angle, $\beta$ (unit: deg), is defined.} \label{SM_systematic_angle} \end{figure*} Fig.~\ref{SM_systematic_angle} shows the temperature field (overlapped with isotherms) with varying the system inclination angle, $\beta$ (unit: deg), from 0 (Rayleigh-B\'enard convection with heating from below and freshwater solidification from above) to $180$ (flipped Rayleigh-B\'enard convection with heating from above and freshwater solidification from below), and here we use the coordinate system attached to the cell. The snapshots at the internal, middle, and outside circles are the temperature field 10min, 2.5hr, and 2.5d (when the system has reached the statistical equilibrium state) after the start of the simulations, respectively. In the RB (small $\beta$) and the flipped-RB (large $\beta$) regimes, the ice front is heavily affected by the intensive convective flow, and thus the ice front represents complex forms of morphology different from those near the vertical convection cases. % % \section{The boundary layer model and its input parameter} It has been mentioned that to locate the ice front profile, we need an input from the simulation results, namely the ice thickness at the starting point of the thermal boundary layer, $h_i(0)$. On top of this, the ice morphology can be predicted. Fig.~\ref{hinput} report the $h_i(0)$ as a function of the inclination angle, $\beta$. It can been observed that $h_i(0)$ is roughly constant and around the level of the averaged ice thickness $h_i$, which is shown in Fig.~3 of the main paper. \begin{figure*}[t] \centering \includegraphics[width=0.5\linewidth]{SM_inputh.pdf} \caption{$h_i(0)$ as a function of the inclination angle, $\beta$. $h_i(0)$ is the ice thickness at $x=0$, which is the input parameter of the boundary layer model.} \label{hinput} \end{figure*} {\color{black} \section{The buoyancy-intensity model} From the ice front morphology of the numerical simulation shown in Figure.4(c)-(m) of the main paper, we can observe that along x-axis, the ice thickness firstly increases and later decreases. We indicate with $\Delta x_0$, the position where the ice thickness reaches the maximum (which is here the a dimensionless position normalized by $H$) The prescribed system inclination results in different levels of thermal stratification, which induces modifications of the ice front morphology. When $\beta = 0$, there is a stably-stratified layer (from $T_0$ to $T_c$) on top of the unstably-stratified layer (from $T_c$ to $T_b$). As the cell is tilted with a small $\beta$, the convection is strong enough to squeeze the stably-stratified layer to be closely attached to the ice front, so the ice morphology is influenced by the one-roll convective flow. As $\beta$ increases towards $90$, the inclination of the temperature gradient with respect to the gravity is strong enough to break down the stratification and thus the stably-stratified layer is set into motion in the form of a clockwise convective roll (an upward cold water current) which competes with the initially anticlockwise roll (downward warm water current). The shielding effect is prominent because the ice thickness reaches a local maximum, and the hot plumes impacting on the top part of the ice induces a local minimum of the ice thickness. The flow motion of the original stably-stratified layer (from $T_0$ to $T_c$) intensifies as $\beta$ increases but still when $\beta<90$. After $\beta=90$, the stratification configuration flips over. The temperature difference of the unstably-stratified layer is now $(T_c-T_0)\approx 4$K. The unstably stratified layer is directly in contact with the ice front. From the above described flow structure dependence on the inclination, we can deduce that there are two aspects of convection motion: 1) The intrinsic aspect, which means that even without inclination the convection exists, and this aspect of convective motion occurs in the initially unstably-stratified layer, i.e., from $T_c$ to $T_b$ when $\beta<90$ and from $T_0$ to $T_c$ when $\beta>90$. 2) The inclination-induced aspect, which means the initially stably-stratified layer can be set into convective motion when the inclination is introduced into the system. This aspect of convective motion occurs in the initially stably-stratified layer, i.e., from $T_0$ to $T_c$ when $\beta<90$ and from $T_c$ to $T_b$ when $\beta>90$. That's why in the definition of the buoyancy intensity, $I_1(\beta)$ and $I_2(\beta)$, there are two parts: the first part that includes $g_x$ results from the intrinsic aspect, and the second part that includes $g_y$ results from the inclination-induced aspect. In order to display $I_1(\beta)$ and $I_2(\beta)$ in a more compact form, we introduce the Heaviside step function, $\mathcal{H}$, to turn on or off the intrinsic aspect. So the final definition of $I_1(\beta)$ and $I_2(\beta)$ is of this kind of form (see Eqn (\ref{buoyancyintensity})) \begin{equation} \begin{split} &I_1(\beta) =g_x (1- \tfrac{\rho(T_m)}{\rho_c}) - g_y \mathcal{H}[\beta-90^{\circ}] (1-\tfrac{\rho(T_0)}{\rho_c});\\ & I_2(\beta) =g_x (1-\tfrac{\rho(T_{m2})}{\rho_c}) + g_y \mathcal{H}[90^{\circ}-\beta] (1-\tfrac{\rho(T_b)}{\rho_c}).\\ \end{split} \label{buoyancyintensity} \end{equation} The competition of the two convective rolls result in a local minimum heat transfer rate where the ice thickness reaches the maximum, and we use the buoyancy-intensity ratio to quantify this competition effect which give good agreement with the trend we observed in the numerical simulations (see Figure.4(b) of the main paper). This provides further evidence that the competition of two convective rolls accounts for the form of ice front morphology.\\ % % % % % % % } % % % % % % % % % %

b49735ac3088aedc1637609a2bf48c7c613155cd

\section{Introduction} \label{section: introduction} The Shannon entropy and other classical information measures serve as a powerful tool in various combinatorial and graph-theoretic applications (e.g., \cite{BabuR14}--\cite{Friedgut04}, \cite{Galvin14}--\cite{Radhakrishnan01} and \cite{Spencer85}) such as the method of types, applications of Shearer's lemma, sub and supermodularity properties of information measures and their applications, entropy-based proofs of Moore bound for irregular graphs, Bregman's theorem on the permanent of square matrices with binary entries, Spencer's theorem in discrepancy theory, etc. The enumeration of discrete structures that satisfy certain local constraints, and particularly the enumeration of independent sets in graphs is of interest in discrete mathematics. Many important structures can be modeled by independent sets in a graph, i.e., subsets of vertices in a graph where none of them are connected by an edge. For example, if a graph models some kind of incompatibility, then an independent set in this graph represents a mutually compatible collection. Upper bounding the number of independent sets in a regular graph was motivated in \cite{Alon91} by a conjecture which has several applications in combinatorial group theory. A survey paper on upper bounding the number of independent sets in graphs, along with some of their applications, is provided in \cite{Samotij15}. The problem of counting independent sets in graphs received, in general, significant attention in the literature of discrete mathematics over the last three decades, and also in the information theory literature (\cite{MadimanT_ISIT07, MadimanT_IT10}). A tight upper bound on the number of independent sets in finite and undirected general graphs was proved in the special setting of bipartite regular graphs in \cite{Kahn01}, and it was conjectured there to hold for general (irregular) graphs (2001, see Conjecture~4.2 in \cite{Kahn01}). A decade later (2010), it was extended in \cite{Zhao10} to regular graphs (that are not necessarily bipartite); a year later (2011), it was proved in \cite{GalvinZ11} for graphs with small degrees (up to~5). Finally, this conjecture was recently (2019) proved in general \cite{SahSaStZhao19}, by utilizing a new approach. The reader is referred to \cite{SahSaStZhao19b} for an announcement on the solution of this conjecture as a frustrating combinatorial problem for two decades, along with the history and ramifications of this problem, and some reflections of the authors on their work in \cite{SahSaStZhao19}. The recently-introduced proof of the conjecture for {\em general} undirected graphs \cite{SahSaStZhao19} uses an induction on the number of vertices in a graph, and it obtains a recurrence inequality whose derivation involves some judicious applications of H\"{o}lder's inequality (see Sections~2 and 4 in \cite{SahSaStZhao19}). It settled for the first time Conjecture~4.2 in \cite{Kahn01} by an interesting approach, which is unrelated to information theory. The possibility of generalizing the information-theoretic proof in \cite{Kahn01} to irregular bipartite graphs was left in \cite{SahSaStZhao19} as an open issue. It should be noted that by proving Kahn's conjecture for irregular bipartite graphs, this readily enables to extend the proof to general undirected graphs (by invoking Zhao's inequality, see Lemma~3 in \cite{GalvinZ11}). The main contribution of this work is the extension of Kahn’s information-theoretic proof technique for bipartite regular graphs \cite{Kahn01} to handle irregular bipartite graphs. In particular, when the bipartite graph is regular on one side, but it may be irregular in the other, the extended entropy-based proof technique yields the same bound that was conjectured by Kahn \cite{Kahn01} and proved by Sah et al. \cite{Zhao10}. The structure of the paper is as follows: Section~\ref{section: preliminaries} provides preliminaries and notation that are essential for the analysis in this paper. Section~\ref{section: scientific merit} explains (in more details) the scientific merit and contributions of the present work; for the sake of causal presentation, we provide these explanations after Section~\ref{section: preliminaries}. Sections~\ref{section: An IT proof of Theorem 3} and~\ref{section: An IT proof of Zhao's inequality} are the core of this work. \section{Preliminaries and Notation} \label{section: preliminaries} We provide in this section the notation and preliminary material which is essential for the presentation in this paper. \subsection{Notation and Basic Properties of the Entropy} \label{subsection: notation} The following notation is used in the present paper: \begin{itemize} \item $\naturals \eqdef \{1, 2, \ldots \}$ denotes the set of natural numbers. \item $X^n \eqdef (X_1, \ldots X_n)$ denotes an $n$-dimensional random vector of discrete random variables, having a joint probability mass function (PMF) that is denoted by $\pmfOf{X^n}$. \item For every $n \in \naturals$, $\OneTo{n} \eqdef \{1, \ldots, n\}$; \item $X_{\set{S}} \eqdef (X_i)_{i \in \set{S}}$ is a random vector for an arbitrary nonempty subset $\set{S} \subseteq \OneTo{n}$; if $\set{S} = \emptyset$, then conditioning on $X_{\set{S}}$ is void. \item $\I{E}$ denotes the indicator of an event $E$; i.e., it is equal to~1 if this event is satisfied, and it is zero otherwise. \item Let $X$ be a discrete random variable that takes its values on a set $\set{X}$, and let $\pmfOf{X}$ be the PMF of $X$. The {\em Shannon entropy} of $X$ is given by \begin{IEEEeqnarray}{rCl} \label{eq: entropy} \Ent{X} \eqdef -\sum_{x \in \set{X}} \pmfOf{X}(x) \, \log \pmfOf{X}(x), \end{IEEEeqnarray} where throughout this paper, we take all logarithms to base~2. \item For $p \in [0,1]$, \begin{IEEEeqnarray}{rCl} \label{eq2: EntBin} \EntBin{p} \eqdef -p \log p -(1-p) \log(1-p), \end{IEEEeqnarray} where $\EntBin{\cdot}$ is the {\em binary entropy function}. By continuous extension, the convention $0 \log 0 = 0$ is used. \item Let $X$ and $Y$ be discrete random variables with a joint PMF $\pmfOf{XY}$, and having a marginal PMF of $X$ given $Y$ which is denoted by $\CondpmfOf{X}{Y}$. Let $X$ and $Y$ take their values in the sets $\set{X}$ and $\set{Y}$, respectively. The {\em conditional entropy} of $X$ given $Y$ is defined as \begin{IEEEeqnarray}{rCl} \label{eq: conditional entropy 1} \EntCond{X}{Y} & \eqdef & -\sum_{(x,y) \in \set{X} \times \set{Y}} \pmfOf{XY}(x,y) \log \CondpmfOf{X}{Y}(x|y) \\[0.1cm] \label{eq: conditional entropy 1.5} &=& \sum_{y \in \set{Y}} \pmfOf{Y}(y) \, \EntCond{X}{Y=y}, \end{IEEEeqnarray} and \begin{IEEEeqnarray}{rCl} \EntCond{X}{Y} &=& \Ent{X,Y} - \Ent{Y}. \label{eq: conditional entropy 2} \end{IEEEeqnarray} \end{itemize} This paper relies on the following basic properties of the Shannon entropy: \begin{itemize} \item Entropies and conditional entropies of discrete random vectors are nonnegative. \item If $\set{X}$ is a finite set, then \begin{IEEEeqnarray}{rCl} \label{eq: UB Ent} \Ent{X} \leq \log \card{\set{X}}, \end{IEEEeqnarray} with equality in \eqref{eq: UB Ent} if and only if $X$ is equiprobable over the set $\set{X}$. \item Conditioning cannot increase the entropy, i.e., \begin{IEEEeqnarray}{rCl} \label{eq: conditioning reduces entropy} \EntCond{X}{Y} \leq \Ent{X}, \end{IEEEeqnarray} with equality in \eqref{eq: conditioning reduces entropy} if and only if $X$ and $Y$ are independent. \item Generalizing \eqref{eq: conditional entropy 2} to $n$-dimensional random vectors gives the {\em chain rule for the entropy}: \begin{IEEEeqnarray}{rCl} \label{eq1: chain rule} \Ent{X^n} &=& \Ent{X_1} + \EntCond{X_2}{X_1} + \ldots + \EntCond{X_n}{X_1, \ldots, X_{n-1}} \\[0.1cm] \label{eq2: chain rule} &=& \sum_{i=1}^n \EntCond{X_i}{X^{i-1}}. \end{IEEEeqnarray} \item The following {\em subadditivity property of the entropy} is implied by \eqref{eq: conditioning reduces entropy} and \eqref{eq2: chain rule}: \begin{IEEEeqnarray}{rCl} \label{eq: subadditivity Ent} \Ent{X^n} \leq \sum_{i=1}^n \Ent{X_i}, \end{IEEEeqnarray} with equality in \eqref{eq: subadditivity Ent} if and only if $X_1, \ldots, X_n$ are independent random variables. \end{itemize} \subsection{Shearer's Lemma} \label{subsection: Shearer's Lemma} Shearer's lemma extends the subadditivity property \eqref{eq: subadditivity Ent} of the entropy. Due to its simplicity and usefulness in this paper (and elsewhere), we state and prove it here. \begin{proposition}[Shearer's Lemma, \em{\cite{ChungGFS86}}] \label{proposition: Shearer's Lemma} Let $X_1, \ldots, X_n$ be discrete random variables, and let the sets $\set{S}_1, \ldots, \set{S}_m \subseteq \OneTo{n}$ include every element $i \in \OneTo{n}$ in {\em at least} $k \geq 1$ of these subsets. Then, \begin{IEEEeqnarray}{rCl} \label{eq: Shearer's Lemma} k \Ent{X^n} \leq \sum_{j=1}^m \Ent{X_{\set{S}_j}}. \end{IEEEeqnarray} \end{proposition} As a special case of \eqref{eq: Shearer's Lemma}, setting $\set{S}_i \eqdef \{i\}$ as singletons for all $i \in \OneTo{n}$ gives \eqref{eq: subadditivity Ent} by having $k=1$ and $m=n$. \begin{IEEEproof} Let $\set{S} = \{i_1, \ldots, i_\ell\}$ with $1 \leq i_1 < \ldots < i_\ell \leq n$. By invoking the chain rule in this order, \begin{IEEEeqnarray}{rCl} \Ent{X_{\set{S}}} &=& \Ent{X_{i_1}} + \EntCond{X_{i_2}}{X_{i_1}} + \ldots + \EntCond{X_{i_\ell}}{X_{i_1}, \ldots, X_{i_{\ell-1}}} \nonumber \\[0.1cm] &\geq& \sum_{i \in \set{S}} \EntCond{X_i}{X^{i-1}} \nonumber \\ &=& \sum_{i=1}^n \I{i \in \set{S}} \, \EntCond{X_i}{X^{i-1}}, \label{eq: conditioning reduces entropy 2} \end{IEEEeqnarray} where the last inequality holds since an additional conditioning cannot increase the entropy (i.e., $\EntCond{X}{Y,Z} \leq \EntCond{X}{Y}$ for all $X,Y$ and $Z$). By assumption, for $i \in \OneTo{n}$, \begin{IEEEeqnarray}{rCl} \label{eq: degree} \sum_{j=1}^m \I{i \in \set{S}_j} \geq k, \end{IEEEeqnarray} since the number of subsets $\{\set{S}_j\}_{j=1}^m$ that include $i$ as an element is at least $k$. Consequently, it follows that \begin{IEEEeqnarray}{rCl} \sum_{j=1}^m \Ent{X_{\set{S}_j}} & \geq & \sum_{j=1}^m \sum_{i=1}^n \bigl\{ \I{i \in \set{S}_j} \, \EntCond{X_i}{X^{i-1}} \bigr\} \label{151220a1} \\ &=& \sum_{i=1}^n \Biggl\{ \sum_{j=1}^m \I{i \in \set{S}_j} \, \EntCond{X_i}{X^{i-1}} \Biggr\} \label{151220a2} \\ &\geq& k \sum_{i=1}^n \EntCond{X_i}{X^{i-1}} \label{151220a3} \\[0.1cm] &=& k \Ent{X^n}, \label{151220a4} \end{IEEEeqnarray} where \eqref{151220a1} follows from \eqref{eq: conditioning reduces entropy 2}; \eqref{151220a2} holds by interchanging the order of summation; \eqref{151220a3} holds by \eqref{eq: degree}; and \eqref{151220a4} holds by the chain rule for the Shannon entropy. \end{IEEEproof} \begin{remark} \label{remark: Sheaerer's lemma} Inequality \eqref{eq: Shearer's Lemma} holds even if the sets $\set{S}_1, \ldots, \set{S}_m$ are not necessarily included in $\OneTo{n}$. To verify it, define the subsets $\set{S}'_j \eqdef \set{S}_j \cap [1:n]$ for all $j \in \OneTo{m}$. The subsets $\set{S}'_1, \ldots, \set{S}'_m$ are all included in $\OneTo{n}$, and every element $i \in \OneTo{n}$ continues to be included in at least $k \geq 1$ of these subsets. Hence, Proposition~\ref{proposition: Shearer's Lemma} can be applied to the subsets $\set{S}'_1, \ldots, \set{S}'_m$. By the monotonicity property of the entropy, the inclusion $\set{S}'_j \subseteq \set{S}_j$ implies that $\Ent{X_{\set{S}'_j}} \leq \Ent{X_{\set{S}_j}}$ for all $j \in \OneTo{m}$, which then yields the statisfiability of \eqref{eq: Shearer's Lemma}. \end{remark} \begin{remark} \label{remark2: Shearer's lemma} A generalized inequality which extends both Shearer's lemma and Han's inequality is provided in \cite[Proposition~1]{MadimanT_IT10}. \end{remark} \vspace*{0.1cm} Shearer's lemma and some of its variants (see \cite{Friedgut04} and \cite{GavinkyLSS14}) have been successfully applied in various occasions (see, e.g., \cite{Friedgut04, Galvin14, GavinkyLSS14, Kahn01, Kahn02, Radhakrishnan01}). Shearer's lemma is also instrumental in this paper. \subsection{Graphs, Independent Sets, and Tensor Products} \label{subsection: Graphs} Let $G$ be an undirected graph, and let $\V{G}$ and $\E{G}$ denote respectively the sets of vertices and edges in $G$. A graph $G$ is called {\em $d$-regular} if the degree of all vertices in $\V{G}$ is equal to $d$. Otherwise, if the graph $G$ is not $d$-regular for some $d \in \naturals$, then $G$ is an {\em irregular} graph. A graph is called {\em bipartite} if it has two types of vertices, and an edge cannot connect vertices of the same type; we refer to the vertices of a bipartite graph $G$ as left and right vertices. A graph $G$ is called {\em complete} if every vertex $v \in \V{G}$ is connected to all the other vertices in $\V{G} \setminus \{v\}$ (and not to itself); similarly, a bipartite graph is called complete if every vertex is connected to all the vertices of the other type in the graph. A complete $(d-1)$-regular graph is denoted by $K_d$, having a number of vertices $\bigcard{\V{K_d}} = d$, and a number of edges $\bigcard{\E{K_d}} = \tfrac12 \, d(d-1)$. Likewise, a complete $d$-regular bipartite graph is denoted by $K_{d,d}$, having a number of vertices $\bigcard{\V{K_{d,d}}} = 2d$ (i.e., $d$ vertices of each of the two types), and a number of edges $\bigcard{\E{K_{d,d}}} = d^2$. An {\em independent set} of an undirected graph $G$ is a subset of its vertices such that none of the vertices in this subset are adjacent (i.e., none of them are joined by an edge). Let $\indset{G}$ denote the set of all the independent sets in $G$, and let $\bigcard{\indset{G}}$ denote the number of independent sets in $G$. Similarly to \cite{Alon91, GalvinZ11, Galvin14, Kahn01, MadimanT_ISIT07, MadimanT_IT10, Radhakrishnan01, SahSaStZhao19, Samotij15, Zhao10} (and references therein), our work considers the question of how many independent sets can $G$ have. The {\em tensor product} $G \times H$ of two graphs $G$ and $H$ is a graph such that the following holds: \begin{itemize} \item The vertex set of $G \times H$ is the Cartesian product $\V{G} \times \V{H}$; \item Two vertices $(g,h), (g', h') \in \V{G \times H}$ are adjacent if and only if $g$ is adjacent to $g'$, and $h$ is adjacent to $h'$, i.e., $(g,g') \in \E{G}$ and $(h,h') \in \E{H}$. This is denoted by $\bigl( (g,h), (g', h') \bigr) \in \E{G \times H}$. \end{itemize} In general, the following identities hold: \begin{IEEEeqnarray}{rCl} & \bigcard{\V{G \times H}} = \bigcard{\V{G}} \, \bigcard{\V{H}}, \label{eq: vertices of tensor product} \\[0.1cm] & \bigcard{\E{G \times H}} = 2 \, \bigcard{\E{G}} \, \bigcard{\E{H}}. \label{eq: edges of tensor product} \end{IEEEeqnarray} By the definition of a complete $d$-regular graph $K_d$, the graph $K_2$ is specialized to two vertices that are connected by an edge. Let us label the two vertices in $K_2$ by 0 and~1. For a graph $G$, the tensor product $G \times K_2$ is a bipartite graph, called the {\em bipartite double cover} of $G$, where the set of vertices in $G \times K_2$ is given by \begin{IEEEeqnarray}{rcl} \label{eq: vertices of bipartite double cover} \V{G \times K_2} = \bigl\{ (v,i) : v \in \V{G}, \, i \in \{0,1\} \bigr\}, \end{IEEEeqnarray} and its set of edges is given by \begin{IEEEeqnarray}{rCl} \label{eq: edges of bipartite double cover} \E{G \times K_2} = \bigl\{ \bigl( (u,0), (v,1) \bigr) : (u,v) \in \E{G} \bigr\}. \end{IEEEeqnarray} Hence, every edge $e = (u,v) \in \E{G}$ is mapped into the two edges $\bigl( (u,0), (v,1) \bigr) \in \E{G \times K_2}$ and $\bigl( (v,0), (u,1) \bigr) \in \E{G \times K_2}$ (since the graph $G$ is undirected). This implies that the numbers of vertices and edges in $G \times K_2$ are doubled in comparison to their respective numbers in $G$; moreover, every edge in $G$, which connects a pair of vertices of specified degrees, is mapped into two edges in $G \times K_2$ where each of these two edges connects a pair of vertices of the same specified degrees. \subsection{Upper Bounds on the Number of Independent Sets} \label{subsection: UB - Independent sets} The present subsection introduces the relevant results to this paper. The following theorem provides a tight upper bound on the number of independent sets in bipartite regular graphs, and its derivation in \cite{Kahn01} makes a clever use of Shearer's lemma (Proposition~\ref{proposition: Shearer's Lemma}). \begin{theorem}[Kahn 2001, \cite{Kahn01}] \label{theorem: Kahn 2001} If $G$ is a bipartite $d$-regular graph with $n$ vertices, then \begin{IEEEeqnarray}{rCl} \label{eq: Kahn 2001} \bigcard{\indset{G}} \leq \bigl( 2^{d+1} - 1 \bigr)^{\frac{n}{2d}}. \end{IEEEeqnarray} Furthermore, if $n$ is an even multiple of $d$, then the upper bound in the right side of \eqref{eq: Kahn 2001} is tight, and it is obtained by a disjoint union of $\frac{n}{2d}$ complete $d$-regular bipartite graphs $(K_{d,d})$. \end{theorem} Kahn's result was later extended by Zhao \cite{Zhao10} for a general $d$-regular graph via a brilliant combinatorial reduction to the setting of $d$-regular bipartite graphs, which proved \cite[Conjecture~4.1]{Kahn01}. \begin{theorem}[Zhao 2010, \cite{Zhao10}] \label{theorem: Zhao10} The upper bound on the number of independent sets in \eqref{eq: Kahn 2001} continues to hold for all $d$-regular graphs with $n$ vertices. \end{theorem} Recently, Sah {\em et al.} \cite{SahSaStZhao19} proved Kahn's conjecture in \cite[Conjecture~4.2]{Kahn01} (made eighteen years earlier) for an upper bound on the number of independent sets in a general undirected graph with no isolated vertices. The proof in \cite{SahSaStZhao19} is combinatorial, and it extends the result in Theorem~\ref{theorem: Zhao10} as follows. \begin{theorem}[Sah et al. 2019, \cite{SahSaStZhao19}] \label{theorem: Sah et al., 2019} Let $G$ be an undirected graph without isolated vertices or multiple edges connecting any pair of vertices. Let $\dgr{v}$ denote the degree of a vertex $v \in \V{G}$. Then, \begin{IEEEeqnarray}{rCl} \label{eq: Sah et al.} \bigcard{\indset{G}} \leq \prod_{(u,v) \in \E{G}} (2^{\dgr{u}} + 2^{\dgr{v}} - 1)^{\frac1{\dgr{u} \dgr{v}}} \end{IEEEeqnarray} with an equality if $G$ is a disjoint union of complete bipartite graphs. \end{theorem} Let $K_{\dgr{u}, \dgr{v}}$ be a complete bipartite graph where the degrees of its left and right vertices are equal to $\dgr{u}$ and $\dgr{v}$, respectively. Then, the number of independent sets in such a complete bipartite graph is equal to \begin{IEEEeqnarray}{rcl} \label{eq: complete bipartite graph} \bigcard{\indset{K_{\dgr{u}, \dgr{v}}}} = 2^{\dgr{u}} + 2^{\dgr{v}} - 1 \end{IEEEeqnarray} since every subset of the left vertices, as well as every subset of the right vertices, forms an independent set of $K_{\dgr{u}, \dgr{v}}$; on the other hand, any subset which contains both left and right vertices is not an independent set (since the bipartite graph $K_{\dgr{u}, \dgr{v}}$ is complete). Note that the substraction by~1 in the right side of \eqref{eq: complete bipartite graph} is because, in the counting of the number of subsets of left vertices ($2^{\dgr{v}}$) or right vertices ($2^{\dgr{u}}$), the empty set is counted twice. Hence, \eqref{eq: Sah et al.} can be rewritten in an equivalent form as \begin{IEEEeqnarray}{rCl} \label{eq2: Sah et al.} \bigcard{\indset{G}} \leq \prod_{(u,v) \in \E{G}} \bigcard{\indset{K_{\dgr{u}, \dgr{v}}}}^{\frac1{\dgr{u} \dgr{v}}}. \end{IEEEeqnarray} Since $\E{K_{\dgr{u}, \dgr{v}}} = \dgr{u} \dgr{v}$, it follows that the bound in \eqref{eq: Sah et al.} (or \eqref{eq2: Sah et al.}) is achieved by the complete bipartite graph $K_{\dgr{u}, \dgr{v}}$. More generally, the bound is achieved by a disjoint finite union of such complete bipartite graphs since the number of independent sets in a disjoint union of graphs is equal to the product of the number of independent sets in each of these component graphs. For the extension of the validity of Theorem~\ref{theorem: Kahn 2001} to Theorem~\ref{theorem: Zhao10}, obtained by relaxing the requirement that the graph is bipartite, the following inequality was introduced by Zhao for every finite graph $G$ \cite[Lemma~2.1]{Zhao10}: \begin{IEEEeqnarray}{rCl} \label{eq: Zhao's inequality} \bigcard{\indset{G}}^2 \leq \bigcard{\indset{G \times K_2}}, \end{IEEEeqnarray} which relates the number of independent sets in a graph to the number of independent sets in the bipartite double cover of this graph. The transition from Theorem~\ref{theorem: Kahn 2001} to Theorem~\ref{theorem: Zhao10}, as introduced in \cite{Zhao10}, is a one-line proof. Let $G$ be a $d$-regular graph with $n$ vertices, then $G \times K_2$ is $d$-regular bipartite graph with $2n$ vertices. Hence, \eqref{eq: Kahn 2001} and \eqref{eq: Zhao's inequality} give that \begin{IEEEeqnarray}{rCl} \label{eq: Zhao's proof} \bigcard{\indset{G}}^2 \leq \bigcard{\indset{G \times K_2}} \leq (2^{d+1}-1)^{\frac{2n}{2d}}, \end{IEEEeqnarray} and taking the square root of the leftmost and rightmost sides of \eqref{eq: Zhao's proof} implies that \eqref{eq: Kahn 2001} continues to hold even when the regular graph $G$ is not necessarily bipartite. \section{Scientific Merit and Contributions of the Present Work} \label{section: scientific merit} After introducing Shearer's lemma (Proposition~\ref{proposition: Shearer's Lemma}) and Theorems~\ref{theorem: Kahn 2001}--\ref{theorem: Sah et al., 2019}, we address the scientific merit and contributions of the present work in more details (in comparison to the introduction in Section~\ref{section: introduction}). Theorem~\ref{theorem: Sah et al., 2019} was recently proved in \cite{SahSaStZhao19} (see also \cite{SahSaStZhao19b}) for general graphs, without relying on information theory. The motivation of our work is rooted in the following sentences from \cite[p.~174]{SahSaStZhao19}: \par {\em Kahn's proof \cite{Kahn01} of the bipartite case of Theorem~\ref{theorem: Kahn 2001} made clever use of Shearer's entropy inequality \cite{ChungGFS86}. It remains unclear how to apply Shearer's inequality in a lossless way in the irregular case, despite previous attempts to do so, e.g., \cite[Section~3]{MadimanT_ISIT07} and \cite[Section~5.C]{MadimanT_IT10}}. The present paper gives an information-theoretic proof of Theorem~\ref{theorem: Sah et al., 2019} in a setting where the bipartite graph is regular on one side (i.e., the vertices on the other side of the bipartite graph can be irregular, and have arbitrary degrees). Its contributions are as follows: \begin{itemize} \item Section~\ref{section: An IT proof of Theorem 3} provides a (non-trivial) extension of the proof of \eqref{eq: Kahn 2001}, from regular bipartite graphs \cite{Kahn01} to general bipartite graphs. It leads to an upper bound on the number of independent sets which is in general looser than the bound in \eqref{eq: Sah et al.} (or its equivalent form in \eqref{eq2: Sah et al.}). However, for the family of bipartite graphs that are regular on one side of the graph, our bound in \eqref{eq1: UB cardinality} coincides with the bound in \eqref{eq: Sah et al.}. The main deviation from Kahn's information-theoretic proof in \cite{Kahn01} is that we allow here the bipartite graph to be irregular. This generalization is not trivial in the sense that it requires a more careful analysis and a slightly more complicated version of Shearer’s Lemma (see Remark~\ref{remark: Sheaerer's lemma}). It should be noted, however, that the suggested proof does follow the same recipe of Kahn's proof in \cite{Kahn01} with some further complications that arise from the non-regularity of the bipartite graphs. \item A variant of the proof of Zhao's inequality \eqref{eq: Zhao's inequality} (\cite[Section~2]{Zhao10}) is provided in Section~\ref{section: An IT proof of Zhao's inequality}. \end{itemize} It is interesting to note that the observation that \eqref{eq: Sah et al.} can be extended from (undirected) bipartite graphs to general graphs, by utilizing \eqref{eq: Zhao's inequality}, was made in \cite[Lemma~3]{GalvinZ11}. However, a computer-assisted proof of \eqref{eq: Sah et al.} was restricted there to graphs whose maximal degrees are at most~5 (see \cite[Theorem~2]{GalvinZ11}). \section{An Information-Theoretic Proof of Theorem~\ref{theorem: Sah et al., 2019} for a Family of Bipartite Graphs} \label{section: An IT proof of Theorem 3} The core of the proof of Theorem~\ref{theorem: Sah et al., 2019} is proving \eqref{eq: Sah et al.} for an undirected bipartite graph. We provide here an extension of the entropy-based proof by Kahn \cite{Kahn01} from bipartite $d$-regular graphs to general bipartite graphs, and then we prove \eqref{eq: Sah et al.} for the family of bipartite graphs that are regular on one side. As it is explained in Section~\ref{section: scientific merit}, the proof in the present section follows the same recipe of Kahn's proof in \cite{Kahn01} with some complications that arise from the non-regularity of the bipartite graphs. The following proof deviates from the proof in \cite{Kahn01} already at its starting point, by a proper adaptation of the proof technique to the general setting of irregular bipartite graphs, followed by a bit more complicated usage of Shearer's lemma and a more involved analysis. Consider first a general bipartite graph $G$ with a number of vertices $\card{\V{G}} = n$, and where none of its vertices is isolated. Label them by the elements of $\OneTo{n}$. Let $\set{L}$ and $\set{R}$ be the vertices of the two types in $\V{G}$ (called, respectively, the left and right vertices in $G$), so $\V{G} = \set{L} \cup \set{R}$ is a disjoint union. Let $\set{D}_{\mathrm{L}}$ and $\set{D}_{\mathrm{R}}$ be, respectively, the sets of all possible degrees of vertices in $\set{L}$ and $\set{R}$. For all $d \in \set{D}_{\mathrm{L}}$, let $\set{L}_d$ be the set of vertices in $\set{L}$ having degree $d$, and let $\set{R}_d$ be the set of vertices in $\set{R}$ that are adjacent to vertices in $\set{L}_d$ (note that the vertices in $\set{R}_d$ are not necessarily those vertices in $\set{R}$ having degree $d$, so the definitions of $\set{L}_d$ and $\set{R}_d$ differ, i.e., they are not similar up to a replacement of left vertices of degree $d$ with right vertices of the same degree). Then, \begin{IEEEeqnarray}{rCl} \label{eq: unions} \set{L} = \bigcup_{d \in \set{D}_{\mathrm{L}}} \, \set{L}_d, \quad \set{R} = \bigcup_{d \in \set{D}_{\mathrm{L}}} \, \set{R}_d, \end{IEEEeqnarray} where the first equality in \eqref{eq: unions} is (by definition) a union of pairwise disjoint sets. Let $\set{S} \in \indset{G}$ be an independent set in $G$, which is selected uniformly at random from $\indset{G}$, and let $X^n \eqdef (X_1, \ldots, X_n)$ be given by \begin{IEEEeqnarray}{rCl} \label{eq: indicators} X_i \eqdef \I{i \in \set{S}}, \quad i \in \OneTo{n}, \end{IEEEeqnarray} so the binary random vector $X^n$ indicates which of the $n$ vertices in $\V{G}$ belongs to the randomly selected independent set $\set{S}$. Since $\set{S}$ is equiprobable in $\indset{G}$, we have \begin{IEEEeqnarray}{rCl} \label{eq: Ent01} \Ent{X^n} = \log \, \bigcard{\indset{G}}. \end{IEEEeqnarray} Let $X_{\set{L}} = (X_i)_{i \in \set{L}}$ and $X_{\set{R}} = (X_i)_{i \in \set{R}}$ be used as a shorthand. Then, \begin{IEEEeqnarray}{rCl} \label{eq: Ent02} \Ent{X^n} &=& \Ent{X_{\set{L}}, X_{\set{R}}} \\[0.1cm] \label{eq: Ent03} &=& \Ent{X_{\set{L}}} + \EntCond{X_{\set{R}}}{X_{\set{L}}} \\[0.1cm] \label{eq: Ent04} & \leq & \sum_{d \in \set{D}_{\mathrm{L}}} \Ent{X_{\set{L}_d}} + \EntCond{X_{\set{R}}}{X_{\set{L}}} \\ \label{eq: Ent05} & \leq & \sum_{d \in \set{D}_{\mathrm{L}}} \Ent{X_{\set{L}_d}} + \sum_{d \in \set{D}_{\mathrm{L}}} \EntCond{X_{\set{R}_d}}{X_{\set{L}}} \\ \label{eq: Ent06} &=& \sum_{d \in \set{D}_{\mathrm{L}}} \bigl\{ \Ent{X_{\set{L}_d}} + \EntCond{X_{\set{R}_d}}{X_{\set{L}}} \bigr\}, \end{IEEEeqnarray} where inequalities \eqref{eq: Ent04} and \eqref{eq: Ent05} hold by the subadditivity of the entropy, and due to \eqref{eq: unions}. It should be noted that although the first summand in the right side of \eqref{eq: Ent06} is an entropy of $X_{\set{L}_d}$, the conditioning on $X_{\set{L}}$ (rather than just on $X_{\set{L}_d}$) in the second term leads to a stronger upper bound on $\Ent{X^n}$ (since $\set{L}_d \subseteq \set{L}$, and conditioning reduces the entropy). This is essential for the continuation of the proof (see \eqref{eq: Ent08}). We next upper bound the two summands in the right side of \eqref{eq: Ent06}, starting with the conditional entropy. By invoking the subadditivity property of the entropy, for every $d \in \set{D}_{\mathrm{L}}$, \begin{IEEEeqnarray}{rCl} \label{eq: Ent07} \EntCond{X_{\set{R}_d}}{X_{\set{L}}} \leq \sum_{r \in \set{R}_d} \EntCond{X_r}{X_{\set{L}}}. \end{IEEEeqnarray} For every $r \in \set{R}_d$, let $\Ngb{r}$ be the set of all the vertices that are adjacent to the vertex $r$. Since the graph $G$ is bipartite, we have $\Ngb{r} \subseteq \set{L}$ (but, in general, $\Ngb{r} \not\subseteq \set{L}_d$), and consequently \begin{IEEEeqnarray}{rCl} \label{eq: Ent08} \EntCond{X_r}{X_{\set{L}}} \leq \EntCond{X_r}{X_{\Ngb{r}}}. \end{IEEEeqnarray} Combining \eqref{eq: Ent07} and \eqref{eq: Ent08} gives that \begin{IEEEeqnarray}{rCl} \label{eq: Ent09} \EntCond{X_{\set{R}_d}}{X_{\set{L}}} \leq \sum_{r \in \set{R}_d} \EntCond{X_r}{X_{\Ngb{r}}}. \end{IEEEeqnarray} For $r \in \set{R}_d$, let \begin{IEEEeqnarray}{rCl} \label{eq: def RV Q} Q_r \eqdef \I{ \set{S} \cap \Ngb{r} = \emptyset} \end{IEEEeqnarray} be the indicator function of the event where none of the vertices that are adjacent (in $G$) to the vertex $r$ are included in the (randomly selected) independent set $\set{S}$. Then, \begin{IEEEeqnarray}{rCl} \label{eq: Ent10} \EntCond{X_r}{X_{\Ngb{r}}} \leq \EntCond{X_r}{Q_r} \end{IEEEeqnarray} since the random vector $X_{\Ngb{r}}$ indicates which of the indices $i \in \Ngb{r}$ are included in $\set{S}$, whereas the binary random variable $Q_r$ only indicates if there is such an index. Consequently, \eqref{eq: Ent09} and \eqref{eq: Ent10} imply that \begin{IEEEeqnarray}{rCl} \label{eq: Ent11} \EntCond{X_{\set{R}_d}}{X_{\set{L}}} \leq \sum_{r \in \set{R}_d} \EntCond{X_r}{Q_r}. \end{IEEEeqnarray} For the binary random variable $Q_r$, let \begin{IEEEeqnarray}{rCl} \label{eq: q_r} q_r \eqdef \prob[Q_r = 1]. \end{IEEEeqnarray} By \eqref{eq: def RV Q}, $Q_r = 0$ if and only if $\set{S} \cap \Ngb{r} \neq \emptyset$, which implies that $r \not\in \set{S}$ since there is a vertex in $\Ngb{r}$ that belongs to the independent set $\set{S}$. Therefore, if $Q_r = 0$, then $X_r = 0$ (see \eqref{eq: indicators}), so \begin{IEEEeqnarray}{rCl} \label{eq: Ent12} \EntCond{X_r}{Q_r = 0} = 0. \end{IEEEeqnarray} If $Q_r = 1$, then $X_r \in \{0,1\}$ and it is also equiprobable (the latter holds since given $Q_r = 1$, the independent set $\set{S}$ is uniformly distributed over all the independent sets in $\indset{G}$ that do not include any neighbor of the vertex $r$, so the vertex $r$ can be either removed from or added to such an independent set, while still giving an independent set that does not include any neighbor of $r$). Hence, \begin{IEEEeqnarray}{rCl} \label{eq: Ent13} \EntCond{X_r}{Q_r = 1} = 1. \end{IEEEeqnarray} Hence, from \eqref{eq: q_r}--\eqref{eq: Ent13}, \begin{IEEEeqnarray}{rCl} \EntCond{X_r}{Q_r} &=& q_r \, \EntCond{X_r}{Q_r = 1} + (1-q_r) \, \EntCond{X_r}{Q_r = 0} \nonumber \\ &=& q_r, \label{eq: Ent14} \end{IEEEeqnarray} and the combination of \eqref{eq: Ent11} and \eqref{eq: Ent14} yields \begin{IEEEeqnarray}{rCl} \label{eq: Ent15} \EntCond{X_{\set{R}_d}}{X_{\set{L}}} \leq \sum_{r \in \set{R}_d} q_r. \end{IEEEeqnarray} We next upper bound $\Ent{X_{\set{L}_d}}$, which is the first summand in the right side of \eqref{eq: Ent06}, and here Shearer's lemma (see Proposition~\ref{proposition: Shearer's Lemma}) comes into the picture. Since, by definition, $\set{R}_d$ is the set of the vertices that are connected to the subset $\set{L}_d$ of the degree-$d$ vertices in $\set{L}$, and $\Ngb{r}$ is the set of vertices in $\set{L}$ that are connected to a vertex $r \in \set{R}_d$ in the bipartite graph $G$, then it follows that every vertex in $\set{L}_d$ belongs to at least $d$ of the subsets $\{\Ngb{r}\}_{r \in \set{R}_d}$. Hence, by Shearer's lemma (which, in light of Remark~\ref{remark: Sheaerer's lemma}, it also holds regardless of the fact that, for $r \in \set{R}_d$, the set $\Ngb{r}$ is not necessarily a subset of $\set{L}_d$), \begin{IEEEeqnarray}{rCl} \label{eq: Ent16} \Ent{X_{\set{L}_d}} \leq \frac1d \sum_{r \in \set{R}_d} \Ent{X_{\Ngb{r}}}. \end{IEEEeqnarray} The binary random variable $Q_r$ is a deterministic function of the random vector $X_{\Ngb{r}}$ since, from \eqref{eq: indicators} and \eqref{eq: def RV Q}, $Q_r=1$ if and only if all the entries of $X_{\Ngb{r}}$ are equal to~0. Consequently, for all $r \in \set{R}_d$, \begin{IEEEeqnarray}{rCl} \Ent{X_{\Ngb{r}}} &=& \Ent{X_{\Ngb{r}}, Q_r} \label{eq: Ent17} \\[0.1cm] &=& \Ent{Q_r} + \EntCond{X_{\Ngb{r}}}{Q_r} \label{eq: Ent18} \\ &=& \EntBin{q_r} + \EntCond{X_{\Ngb{r}}}{Q_r}, \label{eq: Ent19} \end{IEEEeqnarray} where the equality in \eqref{eq: Ent19} follows from \eqref{eq2: EntBin} and \eqref{eq: q_r}. Next, from \eqref{eq: q_r}, \begin{IEEEeqnarray}{rCl} \label{eq: Ent20} \EntCond{X_{\Ngb{r}}}{Q_r} = q_r \EntCond{X_{\Ngb{r}}}{Q_r=1} + (1-q_r) \EntCond{X_{\Ngb{r}}}{Q_r=0}. \end{IEEEeqnarray} If $Q_r = 1$, then $X_{\Ngb{r}}$ is a vector of zeros, so \begin{IEEEeqnarray}{rCl} \label{eq: Ent21} \EntCond{X_{\Ngb{r}}}{Q_r=1} = 0. \end{IEEEeqnarray} Otherwise, if $Q_r=0$, then $X_i=1$ for at least one element $i \in \Ngb{r}$; since $\card{\Ngb{r}}= \dgr{r}$ is the degree of the vertex $r$ (by assumption, there are no multiple edges connecting any pair of vertices), it follows that the vector $X_{\Ngb{r}} \in \{0,1\}^{\dgr{r}}$ cannot be the zero vector, so \begin{IEEEeqnarray}{rCl} \label{eq: Ent22} \EntCond{X_{\Ngb{r}}}{Q_r=0} \leq \log(2^{\dgr{r}}-1). \end{IEEEeqnarray} Combining \eqref{eq: Ent17}--\eqref{eq: Ent22} gives \begin{IEEEeqnarray}{rCl} \label{eq: Ent23} \Ent{X_{\Ngb{r}}} \leq \EntBin{q_r} + (1-q_r) \log(2^{\dgr{r}}-1), \end{IEEEeqnarray} and, from \eqref{eq: Ent16} and \eqref{eq: Ent23}, we get the following upper bound on the first summand in the right side of \eqref{eq: Ent06}: \begin{IEEEeqnarray}{rCl} \label{eq: Ent24} \Ent{X_{\set{L}_d}} \leq \frac1d \sum_{r \in \set{R}_d} \bigl\{ \EntBin{q_r} + (1-q_r) \log(2^{\dgr{r}}-1) \bigr\}. \end{IEEEeqnarray} Consequently, combining \eqref{eq: Ent02}--\eqref{eq: Ent06}, \eqref{eq: Ent15} and \eqref{eq: Ent24} implies that \begin{IEEEeqnarray}{rCl} \label{eq: Ent25} \Ent{X^n} &\leq& \sum_{d \in \set{D}_{\mathrm{L}}} \bigl\{ \Ent{X_{\set{L}_d}} + \EntCond{X_{\set{R}_d}}{X_{\set{L}}} \bigr\} \\ \label{eq: Ent26} &\leq& \sum_{d \in \set{D}_{\mathrm{L}}} \Biggl\{ \, \sum_{r \in \set{R}_d} q_r + \frac1d \sum_{r \in \set{R}_d} \bigl\{ \EntBin{q_r} + (1-q_r) \log(2^{\dgr{r}}-1) \bigr\} \Biggr\} \\[0.1cm] \label{eq: Ent27} &=& \sum_{d \in \set{D}_{\mathrm{L}}} \Biggl\{ \frac1d \sum_{r \in \set{R}_d} \bigl\{ \EntBin{q_r} + (1-q_r) \log(2^{\dgr{r}}-1) + q_r \log(2^d) \bigr\} \Biggr\} \\[0.1cm] \label{eq: Ent28} &=& \sum_{d \in \set{D}_{\mathrm{L}}} \Biggl\{ \frac1d \sum_{r \in \set{R}_d} \biggl\{ \EntBin{q_r} + q_r \log \biggl(\frac{2^d}{2^{\dgr{r}}-1} \biggr) + \log(2^{\dgr{r}}-1) \biggr\} \Biggr\}. \end{IEEEeqnarray} Since $q_r \in [0,1]$ for $r \in \set{R}_d$, we next maximize an auxiliary function $f_r \colon [0,1] \to \Reals$, defined as \begin{IEEEeqnarray}{rCl} \label{eq: f_r} f_r(x) \eqdef \EntBin{x} + x \log \biggl(\frac{2^d}{2^{\dgr{r}}-1} \biggr), \quad x \in [0,1], \end{IEEEeqnarray} in order to obtain an upper bound on the right side of \eqref{eq: Ent28} which is independent of $\{q_r\}$. By \eqref{eq2: EntBin}, setting the first derivative of the concave function $f_r(\cdot)$ to zero gives the equation \begin{IEEEeqnarray}{rCl} \label{derivative f_r is 0} \log \biggl(\frac{1-x}{x}\biggr) + \log \biggl(\frac{2^d}{2^{\dgr{r}}-1} \biggr) = 0, \end{IEEEeqnarray} whose solution is given by \begin{IEEEeqnarray}{rCl} \label{eq: solution x} x = \frac{2^d}{2^d + 2^{\dgr{r}} - 1}. \end{IEEEeqnarray} Consequently, it follows from \eqref{eq: Ent25}--\eqref{eq: f_r} and \eqref{eq: solution x} that \begin{IEEEeqnarray}{rCl} \label{eq: Ent29} \Ent{X^n} &\leq& \sum_{d \in \set{D}_{\mathrm{L}}} \Biggl\{ \frac1d \sum_{r \in \set{R}_d} \Bigl\{ f_r(q_r) + \log(2^{\dgr{r}}-1) \Bigr\} \Biggr\} \\ \label{eq: Ent30} &\leq& \sum_{d \in \set{D}_{\mathrm{L}}} \Biggl\{ \frac1d \sum_{r \in \set{R}_d} \Bigl\{ f_r\biggl( \frac{2^d}{2^d + 2^{\dgr{r}} - 1} \biggr) + \log(2^{\dgr{r}}-1) \Bigr\} \Biggr\} \end{IEEEeqnarray} and the calculation of the term in the inner sum in the right side of \eqref{eq: Ent30} gives \begin{IEEEeqnarray}{rCl} && \hspace*{-0.5cm} f_r\biggl( \frac{2^d}{2^d + 2^{\dgr{r}} - 1} \biggr) + \log(2^{\dgr{r}}-1) \nonumber \\[0.1cm] && = \biggEntBin{\frac{2^d}{2^d + 2^{\dgr{r}} - 1}} + \biggl(\frac{2^d}{2^d + 2^{\dgr{r}} - 1}\biggr) \, \log \biggl(\frac{2^d}{2^{\dgr{r}}-1} \biggr) + \log(2^{\dgr{r}}-1) \label{eq: Ent31} \\[0.1cm] && = - \biggl(\frac{2^d}{2^d + 2^{\dgr{r}} - 1}\biggr) \, \log\biggl( \frac{2^d}{2^d + 2^{\dgr{r}} - 1} \biggr) -\biggl(\frac{2^{\dgr{r}} - 1}{2^d + 2^{\dgr{r}} - 1}\biggr) \, \log\biggl( \frac{2^{\dgr{r}} - 1}{2^d + 2^{\dgr{r}} - 1} \biggr) \nonumber \\[0.1cm] && \hspace*{0.4cm} + \biggl(\frac{2^d}{2^d + 2^{\dgr{r}} - 1} \biggr) \, \log \biggl(\frac{2^d}{2^{\dgr{r}}-1} \biggr) + \log(2^{\dgr{r}}-1) \label{eq: Ent32} \\[0.1cm] && = - \biggl(\frac{2^d}{2^d + 2^{\dgr{r}} - 1}\biggr) \, \biggl[ \log\biggl( \frac{2^d}{2^d + 2^{\dgr{r}} - 1} \biggr) - \log\biggl( \frac{2^d}{2^{\dgr{r}} - 1} \biggr) \biggr] \nonumber \\[0.1cm] && \hspace*{0.4cm} -\biggl(\frac{2^{\dgr{r}} - 1}{2^d + 2^{\dgr{r}} - 1}\biggr) \, \log\biggl( \frac{2^{\dgr{r}} - 1}{2^d + 2^{\dgr{r}} - 1} \biggr) + \log(2^{\dgr{r}}-1) \label{eq: Ent33} \\[0.1cm] && = -\biggl(\frac{2^d}{2^d + 2^{\dgr{r}} - 1}\biggr) \, \log \biggl( \frac{2^{\dgr{r}} - 1}{2^d + 2^{\dgr{r}} - 1} \biggr) -\biggl(\frac{2^{\dgr{r}} - 1}{2^d + 2^{\dgr{r}} - 1}\biggr) \, \log\biggl( \frac{2^{\dgr{r}} - 1}{2^d + 2^{\dgr{r}} - 1} \biggr) \nonumber \\[0.1cm] && \hspace*{0.4cm} + \log(2^{\dgr{r}}-1) \label{eq: Ent34} \\[0.1cm] &&= -\log \biggl( \frac{2^{\dgr{r}} - 1}{2^d + 2^{\dgr{r}} - 1} \biggr) + \log(2^{\dgr{r}}-1) \label{eq: Ent35} \\[0.1cm] &&= \log\bigl(2^d + 2^{\dgr{r}} - 1\bigr), \label{eq: Ent36} \end{IEEEeqnarray} where \eqref{eq: Ent31} and \eqref{eq: Ent32} hold, respectively, by \eqref{eq: f_r} and \eqref{eq2: EntBin}. Substituting the equality in \eqref{eq: Ent36} into the upper bound on the entropy in the right side of \eqref{eq: Ent30}, together with \eqref{eq: Ent01}, gives \begin{IEEEeqnarray}{rCl} \label{eq: Ent37} \log \, \bigcard{\indset{G}} &\leq& \sum_{d \in \set{D}_{\mathrm{L}}} \Biggl\{ \frac1d \sum_{r \in \set{R}_d} \log\bigl(2^d + 2^{\dgr{r}} - 1\bigr) \Biggr\}, \end{IEEEeqnarray} which, by exponentiation of both sides of \eqref{eq: Ent37}, gives \begin{IEEEeqnarray}{rCl} \label{eq1: UB cardinality} \bigcard{\indset{G}} \leq \prod_{d \in \set{D}_{\mathrm{L}}} \prod_{r \in \set{R}_d} \bigl(2^d + 2^{\dgr{r}} - 1\bigr)^{\frac1d}. \end{IEEEeqnarray} The upper bound in the right side of \eqref{eq1: UB cardinality} is in general looser than the bound in Theorem~\ref{theorem: Sah et al., 2019}. Indeed, to clarify this point, let $\Gamma_{d,d'}$ denote the fraction of vertices in $\set{R}_d$ having degree $d' \in \set{D}_\mathrm{R}$. Then, \begin{IEEEeqnarray}{rCl} \prod_{d \in \set{D_{\mathrm{L}}}} \prod_{r \in \set{R}_d} (2^d + 2^{\dgr{r}} - 1)^{\frac1d} &=& \prod_{d \in \set{D}_{\mathrm{L}}} \prod_{d' \in \set{D}_{\mathrm{R}}} (2^d + 2^{d'} - 1)^{\frac{\card{\set{R}_d} \, \Gamma_{d,d'}}{d}} \label{eq2: UB cardinality} \\ &=& \prod_{d \in \set{D}_{\mathrm{L}}} \prod_{d' \in \set{D}_{\mathrm{R}}} \Bigl\{ (2^d + 2^{d'} - 1)^{\frac1{d \, d'}} \Bigr\}^{d' \, \card{\set{R}_d} \, \Gamma_{d,d'}} \label{eq3: UB cardinality} \\ &\geq& \prod_{(u,v) \in \E{G}} \bigl( 2^{\dgr{u}} + 2^{\dgr{v}} - 1 \bigr)^{\frac1{\dgr{u} \, \dgr{v}}}, \label{eq4: UB cardinality} \end{IEEEeqnarray} where \eqref{eq2: UB cardinality} holds since, for all $d \in \set{D}_{\mathrm{L}}$, the number of vertices in $\set{R}_d$ with degree $d' \in \set{D}_{\mathrm{R}}$ is equal to $\card{\set{R}_d} \, \Gamma_{d,d'}$; finally, \eqref{eq4: UB cardinality} holds since the number of edges $e = (u,v) \in \E{G}$ that connect left vertices of degree $d$ and right vertices of degree $d'$ is less than or equal to $d' \, \card{\set{R}_d} \, \Gamma_{d,d'}$ (since $d'$ edges emanate from each such right vertex, but these edges are not necessarily connected to left vertices of degree $d$). In view of this explanation, there is however an interesting case where the upper bound in the right side of \eqref{eq1: UB cardinality} and the bound in Theorem~\ref{theorem: Sah et al., 2019} coincide. Let $G$ be a bipartite graph that is $d$-regular on one side (i.e, one type of its vertices have a fixed degree $d$, and the other type of vertices can be irregular with arbitrary degrees). Without any loss of generality, one can assume that the left vertices are $d$-regular (as otherwise, the graph can be flipped without affecting its independent sets, and also the bound in Theorem~\ref{theorem: Sah et al., 2019} is symmetric in the degrees $\dgr{u}$ and $\dgr{v}$). In this setting, $\set{L}_d = \set{L}$ and $\set{R}_d = \set{R}$ (recall that, by assumption, there are no isolated vertices). Consequently, the right side of \eqref{eq1: UB cardinality} is specialized to \begin{IEEEeqnarray}{rCl} \label{eq5: UB cardinality} \bigcard{\indset{G}} \leq \prod_{r \in \set{R}} \bigl(2^d + 2^{\dgr{r}} - 1\bigr)^{\frac1d}. \end{IEEEeqnarray} Since there are exactly $\dgr{r}$ edges connecting each vertex $r \in \set{R}$ with vertices in $\set{L}$, and (by the latter assumption) all of the left vertices in $\set{L}$ are of a {\em fixed degree} $d$, it follows that in this setting, the right side of \eqref{eq5: UB cardinality} can be rewritten in the form \begin{IEEEeqnarray}{rCl} \prod_{r \in \set{R}} \bigl(2^d + 2^{\dgr{r}} - 1\bigr)^{\frac1d} &=& \prod_{r \in \set{R}} \left( \bigl(2^d + 2^{\dgr{r}} - 1\bigr)^{\frac1{d \dgr{r}}} \right)^{\dgr{r}} \\ \label{eq6: UB cardinality} &=& \prod_{(u,v) \in \E{G}} \bigl(2^{\dgr{u}} + 2^{\dgr{v}} - 1\bigr)^{\frac1{\dgr{u} \dgr{v}}}, \end{IEEEeqnarray} which, indeed, shows that the right side of \eqref{eq1: UB cardinality} and the bound in Theorem~\ref{theorem: Sah et al., 2019} coincide for bipartite graphs that are regular on one side of the graph (without restricting the other side). \section{A Variant of the Proof of Zhao's Inequality} \label{section: An IT proof of Zhao's inequality} This section suggests a variant of the proof of Zhao's Inequality in \eqref{eq: Zhao's inequality} (see \cite[Lemma~2.1]{Zhao10}). Although it is somewhat different from the one in \cite{Zhao10}, this forms in essence a reformulation of Zhao’s proof. Let $G$ be a finite graph, and let $\bigcard{\V{G}} = n$. Label the vertices in the left and right sides of the bipartite graph $G \times K_2$ (i.e., the bipartite double cover of $G$) by $\{(i,0)\}_{i=1}^n$ and $\{(i,1)\}_{i=1}^n$, respectively. Choose independently and uniformly at random two independent sets $\set{S}_0, \set{S}_1 \in \indset{G}$. For $i \in \OneTo{n}$, let $X_i, Y_i \in \{0,1\}$ be random variables defined as $X_i = 1$ if and only if $i \in \set{S}_0$, and $Y_i = 1$ if and only if $i \in \set{S}_1$. Then, by the statistical independence and equiprobable selection of the two independent sets from $\indset{G}$, we have \begin{IEEEeqnarray}{rCl} \Ent{X^n, Y^n} &=& \Ent{X^n} + \Ent{Y^n} \label{eq: Ent38} \\ &=& 2 \log \, \bigcard{\indset{G}}, \label{eq: Ent39} \end{IEEEeqnarray} where \eqref{eq: Ent38} holds since $X^n = (X_1, \ldots, X_n)$ and $Y^n = (Y_1, \ldots, Y_n)$ are statistically independent (by construction), and \eqref{eq: Ent39} holds since they both have an equiprobable distribution over a set whose cardinality is $\bigcard{\indset{G}}$. Consider the following set of vertices in $G \times K_2$: \begin{IEEEeqnarray}{rCl} \label{eq: set S} \set{S} &\eqdef& \bigl\{ \set{S}_0 \times \{0\} \bigr\} \bigcup \bigl\{ \set{S}_1 \times \{1\} \bigr\} \\[0.1cm] &=& \bigcup_{i \in \set{S}_0, \, j \in \set{S}_1} \bigl\{ (i,0), (j,1) \bigr\}. \end{IEEEeqnarray} The set $\set{S}$ is not necessarily an independent set in $G \times K_2$ since $\bigl( (i,0), (j,1) \bigr) \in \E{G \times K_2}$ for all $i \in \set{S}_0$ and $j \in \set{S}_1$ for which $(i,j) \in \E{G}$ (see \eqref{eq: edges of bipartite double cover}). We next consider all $(i,j) \in \E{G}$ such that $X_i = Y_j = 1$. To that end, fix an ordering of all the $2^n$ subsets of $\V{G}$, and let $\set{T} \in \V{G}$ be the first subset in this particular ordering that includes exactly one endpoint of each edge $(i,j) \in \E{G}$ for which $X_i = Y_j = 1$. Consider the following replacements: \begin{itemize} \item If $(i,0) \in \set{S}$ and $i \in \set{T}$, then $(i,0)$ is replaced by $(i,1)$; \item Likewise, if $(j,1) \in \set{S}$ and $j \in \set{T}$, then $(j,1)$ is replaced by $(j,0)$. \end{itemize} Let $\widetilde{\set{S}}$ be the set of new vertices after these possible replacements. Then, $\widetilde{S} \in \indset{G \times K_2}$ since all adjacent vertices in $\set{S}$ are no longer connected in $\widetilde{\set{S}}$. Indeed, there is no way that after (say) a vertex $(i,0)$ is replaced by $(i,1)$, there is another replacement of a vertex $(j,1)$ by $(j,0)$, for some $j$ such that $(i,j) \in \E{G}$; otherwise, that would mean that $\set{T}$ contains both $i$ and $j$, which is impossible by construction. Similarly to the way $X^n, Y^n \in \{0,1\}^n$ were defined, let $\widetilde{X}^n, \widetilde{Y}^n \in \{0,1\}^n$ be defined such that, for all $i \in \OneTo{n}$, $\widetilde{X}_i = 1$ if and only if $(i,0) \in \widetilde{\set{S}}$, and $\widetilde{Y}_i = 1$ if and only if $(i,1) \in \widetilde{\set{S}}$. The mapping from $(X^n,Y^n)$ to $(\widetilde{X}^n, \widetilde{Y}^n)$ is injective. Indeed, it is shown to be injective by finding all indices $(i,j) \in \E{G}$ such that $\widetilde{X}_i = \widetilde{X}_j = 1$ or $\widetilde{Y}_i = \widetilde{Y}_j = 1$, finding the first subset $\set{T} \in \V{G}$ according to our previous fixed ordering of the $2^n$ subsets of $\V{G}$ that includes exactly one endpoint of each such edge $(i,j) \in \E{G}$, and performing the reverse operation to return back to $X^n$ and $Y^n$ (e.g., if $(i,j) \in \E{G}$, $\widetilde{X}_i = \widetilde{X}_j = 1$ and $i \in \set{T}$ while $j \not\in \set{T}$, then $\widetilde{X}_i = 1$ is transformed back to $Y_i = 1$, and $\widetilde{X}_j = 1$ is transformed back to $X_j = 1$). Consequently, we get \begin{IEEEeqnarray}{rCl} \Ent{X^n, Y^n} &=& \Ent{\widetilde{X}^n, \widetilde{Y}^n} \label{eq: Ent40} \\ & \leq & \log \, \bigcard{\indset{G \times K_2}}, \label{eq: Ent41} \end{IEEEeqnarray} where \eqref{eq: Ent40} holds by the injectivity of the mapping from $(X^n, Y^n)$ to $(\widetilde{X}^n, \widetilde{Y}^n)$, and \eqref{eq: Ent41} holds since $\widetilde{S}$ is an independent set in $G \times K_2$, which implies that $(\widetilde{X}^n, \widetilde{Y}^n)$ can get at most $\bigcard{\indset{G \times K_2}}$ possible values (by definition, there is a one-to-one correspondence between $\widetilde{\set{S}}$ and $(\widetilde{X}^n, \widetilde{Y}^n)$). Combining \eqref{eq: Ent38}, \eqref{eq: Ent39}, \eqref{eq: Ent40} and \eqref{eq: Ent41} gives \begin{IEEEeqnarray}{rCl} \label{eq: final} 2 \log \, \bigcard{\indset{G}} \leq \log \, \bigcard{\indset{G \times K_2}}, \end{IEEEeqnarray} which gives \eqref{eq: Zhao's inequality} by exponentiation of both sides of \eqref{eq: final}. \subsection*{Acknowledgement} Correspondence with Ashwin Sah and Mehtaab Sawhney (both are currently mathematics graduate students at MIT), and the constructive comments in the review process are gratefully acknowledged. \vspace*{0.2cm}

0531ca21c6cfbd649fa9303010712fc4ac5cc4a6

\section{Introduction} Propositional dynamic logic, PDL, is a well-known modal logic formalizing reasoning about structured actions, e.g.\ computer programs or actions performed by physical agents, and their correctness properties \cite{Fischer1979,Harel2000}. PDL is subject to two limiting design features. First, being based on classical propositional logic, it formalizes actions that modify values of Boolean variables. A more general setting, one where variables take values from an arbitrary set (integers, characters, trees etc.), is offered by variants of first-order Dynamic Logic, DL \cite{Harel1979,Harel2000}; these variants, however, are mostly undecidable. Second, PDL can express the fact that one action is guaranteed to attain a certain goal while another action is not, but it is not able to express that one action is a more efficient way of attaining the goal than another action. In other words, accessibility between states mediated by actions is modelled as a crisp rather than a graded relation; the former approach is a convenient idealization, but the latter one is more realistic and often also practically required. Both of these limitations of ``classical'' PDL are avoided in a \emph{many-valued} setting. In such a setting, values of formulas in states of a Kripke model are taken from an algebra that is typically distinct from the two-element Boolean algebra used in classical PDL. In a many-valued setting, accessibility between states can also be evaluated in such an algebra, naturally leading to a representation of ``costs'' or other ``weights'' associated with performing actions under specific circumstances. Research into many-valued modal logics dates back to the 1960s, see the pioneering \cite{Segerberg1967} and the later \cite{Ostermann1988}. Fitting \cite{Fitting1991,Fitting1992} was the first to study modal logics where both formulas in states and accessibility relations between states in the Kripke model take values from a non-Boolean algebra. Fitting considers finite Heyting algebras; generalizations studied for example in \cite{Bou2011,Caicedo2015,Caicedo2010,Hansoul2013,Vidal2017} focus on various kinds of finite or infinite \emph{residuated lattices} \cite{Galatos2007}. Residuated lattices are algebraic structures related to \emph{substructural logics}, with many important special cases such as Boolean and Heyting algebras, relation algebras, lattice-ordered groups, powersets of monoids, various algebras on the $[0,1]$-interval and so on. Investigations of PDL based on residuated lattices are relatively scarce. The work in \cite{Behounek2008,Hughes2006,Liau1999} focuses on expressivity of PDL with many-valued accessibility, but technical results such as decidability or completeness are not provided. Teheux \cite{Teheux2014} establishes decidability and completeness of PDLs based on finite {\L}ukasiewicz chains and the present author \cite{Sedlar2016} establishes decidability and completeness of PDL extending the paraconsistent modal logic of \cite{Odintsov2010}; both papers, however, deal with crisp acessibility relations. As an attempt to sytematize the work in many-valued PDL, Madeira et al.\ \cite{Madeira2016,Madeira2015} put forward a general method of producing many-valued versions of PDL, based on the matrix representation of Kleene algebras; their method, however, applies only if models are defined to be finite. In this paper we add to this literature by studying PDLs based on finite Full Lambek algebras, that is, residuated lattices with a distinguished, though arbitrary, $0$ element. We assume that both evaluations of formulas in states and accessibility between states are many-valued. Our main technical results are general completeness and decidability proofs for logics in the family. To the best of our knowledge, our results are the first decidability and completeness results concerning non-crisp many-valued PDL. To be more specific, we work with versions of test-free PDL based on finite Full Lambek algebras with canonical constants; we prove that any PDL based on a finite FL-algebra with canonical constants is decidable; we also establish a completeness result for PDLs based on finite commutative integral FL-algebras with canonical constants. The paper is structured as follows. Section \ref{sec--PRE} introduces the general framework of PDL based on finite FL-algebras. We note that, for technical reasons discussed in \S\ref{sec--ISS}, our version of PDL uses the transitive closure operator, or Kleene plus, as primitive instead of the more standard reflexive transitive closure operator, the Kleene star. An informal interpretation of the framework is discussed in \S\ref{sec--MOT}. Section \ref{sec--DEC} establishes our decidability result using a generalization of the smallest filtration technique. Section \ref{sec--COM} establishes the completeness result for PDLs based on finite integral commutative FL-algebras with canonical constants. Our work there builds on the results of \cite{Bou2011}, but the canonical model construction used in our proof is novel to this paper (it is a suitable generalization of the greatest filtration construction, though the model itself is infinite). \section{Preliminaries}\label{sec--PRE} In this section we briefly recall two-valued PDL (\S\ref{sec--PRE1}), and we define FL-algebras and many-valued models for the language of PDL based on them (\S\ref{sec--PRE2}). We point out some basic facts that we will use later on. \subsection{Two-valued PDL}\label{sec--PRE1} We begin by recalling some well-known facts about two-valued test-free PDL; see \cite{Harel2000}. Fix $Ac = \{ \mathrm{a}_{i} \mid i \in \omega \}$, a countable set of atomic action expressions. The set of \emph{standard action expressions}, $\mathit{STA}$, is the closure of $Ac$ under applying binary operators $;$ (``composition''), $\cup$ (``choice'') and unary $*$ (``Kleene star''). That is, $\mathit{STA}$ are regular expressions over $Ac$ without the empty expression. For example, $(\mathrm{a}_0 ; \mathrm{a}_1)^{*} \cup \mathrm{a}_0 $ is in $ \mathit{STA}$. Let $Pr = \{ \mathrm{p}_i \mid i \in \omega \}$ be a countable set of propositional variables. Take $\bm{2}$, the two-element Boolean algebra on the set $\{ 0, 1 \}$ with meet $\sqcap$, join $\sqcup$ and complement $\mathord{-}$; the binary operation $\Rightarrow$ is defined as usual: $a \Rightarrow b := \mathord{-} a \sqcup b$. Formulas of the \emph{standard language for $\bm{2}$}, $Fm(\mathcal{L}^{\mathit{STA}}_{\bm{2}})$, are defined by \[ \varphi \: := \: \mathrm{p} \mid \bar{c} \mid \varphi \land \varphi \mid \varphi \lor \varphi \mid \varphi \to \varphi \mid [\alpha]\varphi \] where $\mathrm{p} \in Pr$, $c \in \bm{2}$ and $\alpha \in \mathit{STA} $. For example, $\mathrm{p}_0 \to [\mathrm{a}_0 ; (\mathrm{a}_1)^{*}] (\mathrm{p}_1 \to \bar{0})$ is a formula of $\mathcal{L}^{\mathit{STA}}_{\bm{2}}$. A \emph{$\bm{2}$-valued frame for $\mathit{STA}$} is $\mathfrak{F} = (S, \{ R_{\alpha} \}_{\alpha \in \mathit{STA}})$ where $S$ is a non-empty set and, for each $\alpha \in \mathit{STA}$, $R_{\alpha}$ is a function from $S \times S$ to $\bm{2}$. We denote $R(\alpha) := \{ (s,t) \mid R_{\alpha}(s,t) = 1 \}$; and the functions in $\{ R_{\alpha} \}_{\alpha \in \mathit{STA}}$ are required to satisfy the following: \begin{enumerate*} \item $R(\alpha \cup \beta) = R(\alpha) \cup R(\beta)$; \item $R(\alpha ; \beta) = R(\alpha) \circ R(\beta)$, the composition of $R(\alpha)$ and $R(\beta)$; \item $R(\alpha^{*}) = R(\alpha)^{*}$, the reflexive transitive closure of $R(\alpha)$. \end{enumerate*} Let $\mathfrak{F} = (S, \{ R_{\alpha} \}_{\alpha \in \mathit{STA}})$ be a $\bm{2}$-valued frame. A \emph{$\bm{2}$-valued model based on $\mathfrak{F}$} is $\mathfrak{M} = (S, \{ R_{\alpha} \}_{\alpha \in \mathit{STA}}, V)$ where $V : Fm(\mathcal{L}^{\mathit{STA}}_{\bm{2}}) \times S \to \bm{2}$ such that \begin{itemize} \item $V(\bar{c}, s) = c$; \item $V(\varphi \land \psi, s) = V(\varphi, s) \sqcap V(\psi, s)$, $V(\varphi \lor \psi, s) = V(\varphi, s) \sqcup V(\psi, s)$, and $V(\varphi \to \psi, s) = V(\varphi, s) \Rightarrow V(\psi, s)$; \item $V([\alpha]\varphi, s) = \bigsqcap_{t \in S} \big (R_{\alpha} (s, t) \Rightarrow V(\varphi, t) \big )$. \end{itemize} \noindent Note that $V([\alpha]\varphi, s) = \bigsqcap_{R_{\alpha}(s, t) = 1} V(\varphi,t)$. A formula $\varphi$ is \emph{valid in $\mathfrak{M}$} iff $V(\varphi,s) = 1$ for all $s$; validity in frames and classes of frames is defined as expected. This is the standard presentation of test-free PDL, phrased in a way that invites generalizations obtained by replacing $\bm{2}$ by another algebra. We will study some such generalizations in this paper but, as we discuss in more detail below, the story is somewhat more complicated. For reasons discussed in \S\ref{sec--ISS}, our generalizations will use a different primitive iteration operator instead of the Kleene star. The operator we will use, however, is conveniently related to the Kleene star. The set of \emph{action expressions over $Ac$}, $\mathit{ACT}$, is the closure of $Ac$ under composition, choice and the unary operator $+$ (``Kleene plus''). Formulas of the \emph{language $\mathcal{L}_{\bm{2}}$} are defined as expected (we omit reference to $\mathit{ACT}$), with $\alpha \in \mathit{ACT}$; for example, $\mathrm{p}_0 \to [\mathrm{a}_0 ; (\mathrm{a}_1)^{+}] (\mathrm{p}_1 \to \bar{0})$ is a formula of $\mathcal{L}_{\bm{2}}$. The definition of \emph{$\bm{2}$-valued frames for $\mathit{ACT}$} is the same as the definition of $\bm{2}$-valued frames for $\mathit{STA}$, with an obvious exception, namely, the requirement that $R(\alpha^{+})$ be the \emph{transitive closure} of $R(\alpha)$, i.e.\ $R(\alpha^{+}) = \bigcup_{n > 0} R^{n}(\alpha)$, where $R^{1}(\alpha) = R(\alpha)$ and $R^{n+1}(\alpha) = R^{n}(\alpha) \circ R(\alpha)$. Compare this with the \emph{reflexive} transitive closure $R(\alpha)^{*} = \bigcup_{n \geq 0} R^{n}(\alpha)$, where $R^{0}(\alpha) = \{ (s,s) \mid s \in S \}$. Models based on frames for $\mathit{ACT}$ are defined as before. \begin{proposition}\label{prop--PRE-Avoiding Kleene star} Let For each $\alpha \in \mathit{ACT}$ and $\varphi \in Fm(\mathcal{L}_{\bm{2}})$, \[V(\varphi\land [\alpha^{+}]\varphi, s) = 1 \quad \text{iff} \quad \forall t ( (s,t) \in R(\alpha)^{*} \implies V(\varphi, t) = 1)\, . \] \end{proposition} Proposition \ref{prop--PRE-Avoiding Kleene star} implies that $\varphi \land [\alpha^{+}]\varphi$ ``simulates'' $[\alpha^{*}]\varphi$ in $\mathcal{L}_{\bm{2}}$. This provides a justification for our using languages based on $\mathit{ACT}$ rather than on $\mathit{STA}$ in what follows. However, we admit that this choice is related to the technical issues discussed in \S\ref{sec--ISS}. \subsection{FL-algebras and finitely-valued PDL}\label{sec--PRE2} In this section we generalize two-valued PDL by replacing the two-element Boolean algebra $\bm{2}$ by a more general structure, namely, a finite FL-algebra. FL-algebras provide semantics for a wide class of \emph{substructural logics} \cite{Galatos2007}. \begin{definition} An \emph{FL-algebra} (``full Lambek algebra'', \cite{Galatos2007}) is a set $X$ with binary operations $\sqcap, \sqcup, \backslash, \cdot, \slash$ and two distinguished elements $1, 0$ such that \begin{itemize} \item $(X, \sqcap, \sqcup)$ is a lattice (let $a \sqsubseteq b$ iff $a \sqcup b = b$); \item $(X, \cdot, 1)$ is a monoid; \item $(\backslash, \cdot, \slash)$ are residuated over $(X, \sqsubseteq)$, i.e.\ \[ a \cdot b \sqsubseteq c \quad\text{ iff }\quad b \sqsubseteq a \backslash c \quad\text{ iff }\quad a \sqsubseteq c \slash b \, ; \] \item $0$ is an arbitrary element of $X$. \end{itemize} \emph{Residuated lattices} are $0$-free reducts of FL-algebras. Each finite FL-algebra $\bm{X}$ contains a least element $\bot^{\bm{X}}$ (for all $a \in X$, $\bot^{\bm{X}} \sqsubseteq a$) and a greatest element $\top^{\bm{X}}$ (for all $a \in X$, $a \sqsubseteq \top^{\bm{X}}$). \end{definition} We usually write $ab$ instead of $a \cdot b$ and $a \Rightarrow b$ instead of $b \slash a$. Two varieties of FL-algebras will be important in this paper: \begin{itemize} \item \emph{commutative} FL-algebras satisfy $ab = ba$ for all $a,b \in X$; \item \emph{integral} FL-algebras satisfy $a \sqsubseteq 1$ for all $a \in X$. \end{itemize} \noindent Note that in commutative FL-algebras $a \backslash b = b \slash a$. \begin{example} The two-element Boolean algebra $\bm{2}$ is a commutative integral FL-algebra, where $\cdot$ is $\sqcap$ and $\backslash$ (identical to $\slash$) is $\Rightarrow$. \end{example} \begin{example}\label{exam:Luk} Let $N > 0$ and define $\bm{N} = (N, max, min, +_N, \to_N )$ where \[ a +_N b = min (a + b, N - 1) \quad \text{and} \quad a \to_N b = max (b - a, 0) \, . \] $\bm{N}$ is a finite commutative integral FL-algebra, with $0$ as the monoid identity with respect to $+_{N}$ and the greatest element under the $\geq$-ordering induced by taking $min$ as join. We note that $\bm{N}$ is isomorphic to the $N$-element {\L}ukasiewicz lattice $\textit{\bm{{\L}}}_{N}$ over $\{ \frac{k}{N-1} \mid k \in N \}$. \end{example} \begin{example} As an example of a non-commutative, non-integral infinite FL-algebra, take the power set of the free monoid over some set $\Sigma$, i.e.\ the set of languages over $\Sigma$, with intersection as meet, union as join, $L \cdot L' := \{ xx' \mid x \in L \And x' \in L' \}$, $\{ \varepsilon \}$ as the monoid identity ($\varepsilon$ is the empty word) and $L \backslash L' := \{ x \in \Sigma \mid L \cdot \{ x \} \subseteq L' \}$, $L' \slash L := \{ x \in \Sigma \mid \{ x \} \cdot L \subseteq L' \}$. \end{example} The following lemma summarizes some of the properties of FL-algebras we will rely on in this paper (we will often say that something holds ``by the properties FL-algebras'' in our proofs). \begin{lemma}\label{lem--APP-Properties of L} Let $\bm{X}$ be an arbitrary FL-algebra. Then \begin{enumerate* \item $a \sqsubseteq b$ iff $1 \sqsubseteq a \Rightarrow b$; \item If $a \sqsubseteq b$ and $c \sqsubseteq d$, then $b \Rightarrow c \sqsubseteq a \Rightarrow d$, $b \backslash c \sqsubseteq a \backslash d$ and $ac \sqsubseteq bd$; \item $(a \sqcup b)c = ac \sqcup ab$ and $c(a \sqcup b) = ca \sqcup cb$; \item $a \Rightarrow (b \sqcap c) = (a \Rightarrow b) \sqcap (a \Rightarrow c)$; \item $a \sqcup b \Rightarrow c = (a \Rightarrow c) \sqcap (b \Rightarrow c)$; \item $a \Rightarrow (b \Rightarrow c) = ab \Rightarrow c$; \item $(a \Rightarrow b)(b \Rightarrow c) \sqsubseteq a \Rightarrow c$; \item $(1 \Rightarrow a) = a$ \end{enumerate*} \end{lemma} If $S$ is a non-empty set, then $\Pi(S)$ is the set of all finite sequences of elements of $S$; that is, $\pi \in \Pi(S)$ iff $\pi$ is a function from some $n \in \omega$, called the length of $\pi$, to $S$. The unique sequence of length $0$ is $\emptyset$. If $\pi$ is a sequence of length $n$ and $s \in S$, then $\pi^{\frown}s$ is the unique sequence of length $n+1$ such that $(\pi^{\frown}s)(k) = \pi(k)$ for all $k < n$ and $(\pi^{\frown}s)(n) = s$. Note that each sequence $\pi$ of length $n > 0$ can be expressed as $( \ldots (\emptyset^{\frown}\pi(0))^{\frown} \ldots)^{\frown}\pi(n-1)$. \begin{definition} Let $\bm{X}$ be a finite FL-algebra and $S$ a non-empty set. A \emph{binary $\bm{X}$-valued relation on $S$} is any function from $S \times S$ to $\bm{X}$. Let $R, Q$ be binary $\bm{X}$-valued relations on a set $S$; then \begin{itemize} \item the \emph{union} of $R$ and $Q$ is the function $R \cup Q$ defined by $(R \cup Q)(s,t) : = R(s,t) \sqcup Q(s,t)$; \item the \emph{composition} of $R$ and $Q$ is the function $R \circ Q$ defined by $(R \circ Q)(s,t) = \bigsqcup_{x \in S} \big ( R(s,x) \cdot Q(x,t) \big)$; \item the \emph{transitive closure} of $R$ is the function $R^{+}$ defined by $R^{+}(s,t) = \bigsqcup_{\pi \in \Pi(S)} R s \pi t$ where $R s \pi t$ is defined as follows: \begin{itemize} \item $R s \emptyset t = R (s, t)$ and \item $R s (\pi^{\frown}u) t = R s \pi u \cdot R(u,t)$. \end{itemize} \end{itemize} We say that $Q$ \emph{extends} $R$, notation $R \sqsubseteq Q$, iff $R(s,t) \sqsubseteq Q(s,t)$ for all $s,t \in S$; $R$ is the \emph{smallest} relation in a set $\{ R_i \}_{i \in I}$ if $R = R_i$ for some $i \in I$ and each $R_i$ extends $R$. $R$ is \emph{transitive} if $R(s,t) \cdot R(t,u) \sqsubseteq R(s,u)$ for all $s,t,u \in S$; and $R$ is \emph{reflexive} if $1 \sqsubseteq R(s,s)$ for all $s \in S$. \end{definition} \noindent Note that we need to assume that all the required joins exist in $\bm{X}$; hence the restriction to finite FL-algebras (however, a restriction to complete $\bm{X}$ is sufficient, as is the assumption that $R, Q$ are ``$\bm{X}$-safe'' \cite[ch.\ 5]{Hajek1998}). \begin{proposition}\label{prop--PRE-closure} Let $\bm{X}$ be a finite FL-algebra and $R$ a binary $\bm{X}$-valued relation on a set $S$. Then $R^{+}$ is the smallest transitive relation extending $R$. For any $R$, define $R^{*}$ as follows: \[ R^{*}(s,t) \: = \: \begin{cases} 1 & \text{if } s = t \\ R^{+}(s,t) & \text{otherwise.} \end{cases} \] Then $R^{*}$ is the smallest reflexive transitive relation extending $R$. \end{proposition} \begin{proof} It is clear that $R^{+}$ is a transitive relation extending $R$. Now assume that so is $Q$. The conclusion that $R^{+} \sqsubseteq Q$ follows from two facts that are easily established by induction on the length of $\pi$: (a) For all $s,t \in S$ and $\pi \in \Pi(S)$, $R s \pi t \sqsubseteq Q s \pi t$ (the assumption that $R \sqsubseteq Q$ is used here); (b) For all $s,t \in S$ and $\pi \in \Pi(S)$, $Q s \pi t \sqsubseteq Q(s,t)$ (the assumption that $Q$ is transitive is used). Since $\bm{X}$ is finite, the two claims imply that, for any given $s$ and $t$, $\bigsqcup_{\pi} Rs\pi t \sqsubseteq \bigsqcup_{\pi} Q s \pi t \sqsubseteq Q(s,t)$. It is clear that $R^{*}$ is a reflexive transitive relation extending $R$. If so is $Q$, then we reason for any given $s$ and $t$ by cases as follows. If $s = t$, then $R^{*}(s,t) \sqsubseteq Q(s,t)$ is equivalent to $1 \sqsubseteq Q(s,s)$, which holds by reflexivity of $Q$. If $s \neq t$, then $R^{*}(s,t) \sqsubseteq Q(s,t)$ is equivalent to $R^{+}(s,t) \sqsubseteq Q(s,t)$, which follows from the assumption that $Q$ is a transitive relation extending $R$. Hence, $R^{*}(s,t) \sqsubseteq Q(s,t)$ for any $s$ and $t$. \end{proof} \begin{lemma}\label{lem--PRE-Identity} Let $\bm{X}$ be a finite FL-algebra and $S$ a set; the $\bm{X}$-valued identity relation on $S$ is defined as follows: \[ Id_{\bm{X}}(s,t) := \begin{cases} 1 & \text{if } s = t \\ \bot^{\bm{X}} & \text{otherwise.} \end{cases} \] If $\bm{X}$ is integral, then $R^{*} = Id_{\bm{X}} \cup R^{+}$ for any binary $\bm{X}$-valued relation on $S$. \end{lemma} \begin{proof} We omit the proof; we just note that if $s = t$, then $R^{*}(s,t) = Id_{\bm{X}}(s,t) \sqcup R^{+}(s,t)$ is equivalent to $R^{+}(s,t) \sqsubseteq 1$, which is guaranteed to hold only if $\bm{X}$ is integral. \end{proof} \begin{definition}\label{def--frame} Let $\bm{X}$ be a finite FL-algebra. An \emph{$\bm{X}$-valued frame for $\mathit{ACT}$} is a pair $\mathfrak{F} = (S, \{ R_{\alpha} \}_{\alpha \in \mathit{ACT}})$ where $S$ is a non-empty set and, for all $\alpha \in \mathit{ACT}$, $R_{\alpha}$ is an $\bm{X}$-valued binary relation on $S$ such that \begin{enumerate*} \item $R_{\alpha \cup \beta} = R_{\alpha} \cup R_{\beta}$; \item $R_{\alpha ; \beta} = R_{\alpha} \circ R_{\beta}$; and \item $R_{\alpha^{+}} = R_{\alpha}^{+}$. \end{enumerate*} \end{definition} \noindent $\bm{X}$-valued frames will also be referred to as $\bm{X}$-frames or simply frames if $\bm{X}$ is clear from the context or immaterial. We will sometimes write $R_{\alpha}st$ instead of $R_{\alpha}(s,t)$. \begin{definition}\label{def--formulas} Formulas of the \emph{language $\mathcal{L}_{\bm{X}}$} are defined as follows: \[ \varphi := \mathrm{p} \mid \bar{c} \mid \varphi \land \varphi \mid \varphi \lor \varphi \mid \varphi \backslash \varphi \mid \varphi \cdot \varphi \mid \varphi \slash \varphi \mid [\alpha] \varphi \, , \] where $\mathrm{p} \in Pr$, $c \in \bm{X}$ and $\alpha \in \mathit{ACT}$. We use $\bot, \top$ instead of $\overline{\bot^{\bm{X}}}$ and $\overline{\top^{\bm{X}}}$, respectively. We often write $\varphi\psi$ instead of $\varphi \cdot \psi$, $\varphi \to \psi$ instead of $\psi \slash \varphi$, $m$ instead of $\mathrm{a}_m$, and $\alpha\beta$ instead of $\alpha; \beta$. We define $\varphi \leftrightarrow \psi := (\varphi \to \psi) \land (\psi \to \varphi)$, $\neg \varphi := \varphi \to \bot$ and $\langle \alpha \rangle \varphi := \neg [\alpha] \neg \varphi$. \end{definition} Note that we use the same symbol $\otimes \in \{ \backslash, \cdot, \slash \}$ for the implication and fusion connectives of the language and for the residuated operations on FL-algebras. We will denote the operations on a given $\bm{X}$ as $\otimes^{\bm{X}}$ in contexts where it is convenient for the reader to distinguish the connectives of the language from the operations on the algebra. (However, $\Rightarrow$ denotes the operation $\slash^{\bm{X}}$ and $\to$ denotes the connective $\slash$ throughout.) \begin{definition}\label{def--model} A \emph{model} based on an $\bm{X}$-frame $(S, \{ R_{\alpha} \}_{\alpha \in \mathit{ACT}})$ is $\mathfrak{M} = (S, \{ R_{\alpha} \}_{\alpha \in \mathit{ACT}}, V)$, where $V $ is a function from $Fm(\mathcal{L}_{\bm{X}}) \times S$ to $\bm{X}$ such that \begin{itemize} \item $V(\bar{c}, s) = c$; \item $V(\varphi \land \psi, s) = V(\varphi, s) \sqcap V(\psi, s)$ and $V(\varphi \lor \psi, s) = V(\varphi, s) \sqcup V(\psi, s)$; \item $V(\varphi \otimes \psi, s) = V(\varphi, s) \otimes^{\bm{X}} V(\psi, s)$ for $\otimes \in \{ \backslash, \cdot, \slash \}$; \item $V([\alpha]\varphi, s) = \bigsqcap_{t \in S} \big ( R_{\alpha} st \Rightarrow V(\varphi, t) \big )$. \end{itemize} \noindent A formula $\varphi$ is \emph{valid in $\mathfrak{M}$} iff $1 \sqsubseteq V(\varphi, s)$ for all $s$ in $\mathfrak{M}$. Validity in frames and classes of frames is defined as expected. The \emph{theory} of a frame is the set of formulas valid in the frame; the theory of a class of frames is the set of formulas valid in each frame in the class. $Th(\bm{X})$ is the theory of the class of all $\bm{X}$-frames. \end{definition} The following addendum to Proposition \ref{prop--PRE-Avoiding Kleene star} suggests that integral FL-algebras are particularly suitable for us. \begin{proposition}\label{prop--PRE-Reflexive transitive closure} Take an arbitrary $\bm{X}$-frame for a finite integral $\bm{X}$. Then $V(\varphi \land [\alpha^{+}]\varphi, s) = \bigsqcap_{t \in S} (R^{*}_{\alpha} st \Rightarrow V(\varphi,t))$. \end{proposition} \begin{proof} The $\sqsubseteq$-inequality is straightforward and the $\sqsupseteq$-inequality follows from Lemma \ref{lem--PRE-Identity}. \end{proof} It is clear that two-valued PDL is a special case of the present framework for $\bm{X} = \bm{2}$. \begin{lemma}\label{lem--PRE-PDL validities hold} The following are valid in each $\bm{X}$-frame: \begin{multicols}{2} \begin{enumerate}[label={(\alph*)}] \item $[\alpha] (\varphi \land \psi) \leftrightarrow ([\alpha]\varphi \land [\alpha]\psi)$ \item $[\alpha \cup \beta]\varphi \leftrightarrow ([\alpha]\varphi \land [\beta]\varphi)$ \item $[\alpha\beta] \varphi \leftrightarrow [\alpha][\beta]\varphi$ \item $[\alpha^{+}]\varphi \leftrightarrow [\alpha](\varphi \land [\alpha^{+}]\varphi)$ \end{enumerate} \end{multicols} \end{lemma} \begin{proof} To prove that $\varphi \leftrightarrow \psi$ is valid if suffices to show that $V(\varphi, s) = V(\psi,s)$ for all $s$ in all models. (a) The proof relies on the fact that $a \Rightarrow (b \sqcap c) = (a \Rightarrow b) \sqcap (a \Rightarrow c)$ in all FL-algebras. (b) The proof relies on the fact that $(a \sqcup b) \Rightarrow c = (a \Rightarrow c) \sqcap (b \Rightarrow c)$ in all FL-algebras. (c) The proof relies on the fact that $a \Rightarrow (b \Rightarrow c) = ab \Rightarrow c$ in all FL-algebras. (Note that composition of relations needs to be defined using monoid multiplication $\cdot$, not lattice meet.) (d) The proof relies on the fact that $R_{\alpha} st \sqsubseteq R_{\alpha^{+}} st$, it also uses simple composition of paths. \end{proof} We will discuss an informal interpretation of a special case of the many-valued framework in the next section. Speaking generally, however, we may adapt the slogan characterizing modal logic as providing languages for talking about relational structures \cite[p.\ viii]{Blackburn2001} and say that many-valued modal logics provide \emph{simple yet expressive languages for talking about many-valued relational structures}. Examples of many-valued relational structures include weighted structures such as weighted graphs etc. Choosing an FL-algebra as the algebra of weights brings the framework closer to substructural logics that include well-known formalisms for reasoning about resources (variants of linear logic) or graded properties and relations (fuzzy logics). Many-valued PDL adds to this the capacity to articulate reasoning about \emph{structured} many-valued relations using the PDL relational operations of choice, composition and iteration. An intriguing connection here is the relation of finitely-valued PDL to weighted automata over finite semirings \cite{Droste2009}, but a more thorough investigation of this connection is left for another occasion. \section{Motivation}\label{sec--MOT} This section discusses the informal interpretation of finitely-valued PDL. We give two general interpretations of the framework first and then we zoom in to PDLs over a specific class of FL-algebras. Our overview is cursory; the present paper is focused more on basic technical results than on informal interpretations and applications. A more thorough exploration of the latter is left for another occasion. We only note here that we consider many-valued PDL to be sufficiently mathematically interesting to be studied independently of informal interpretations and applications. We have mentioned before the slogan that modal logics provide simple yet expressive languages for talking about relational structures \cite[p.\ viii]{Blackburn2001}; by the same token, many-valued modal logics can be seen as providing means of talking about ``weighted'' relational structures. Two-valued PDL has been applied to at least two kinds of relational structures which have very natural weighted generalizations. We discuss these in turn. First, take the interpretation of modal logic that relates it to \emph{description logics} \cite{Baader2007}. Simply put, formulas of a modal language can be seen as expressing ``concepts'', i.e.\ properties of objects, and indices of modal operators as expressing various ``roles'', i.e.\ relations between objects. On this reading, ``states'' in a Kripke model represent arbitrary objects and ``accessibility relations'' between them represent relations between these objects. Structured modal indices that come with PDL (i.e.\ ``action expressions'' as we call them) can be seen as expressing structured relations between objects; union, composition and transitive closure have been found particularly suitable for expressing various important concepts and roles \cite{Baader1991}. Many-valued description logics (see \cite{Straccia2006} for instance) are a generalization of description logics designed for management of \emph{uncertain and imprecise information}. These logics can express the fact that an object is subsumed under a given concept (e.g.\ ``tall'' if the reader will forgive the platitudinous example) only to some degree or that only imprecise information about a relation holding between two objects is available. Finitely-valued PDL as presented here can be seen as a family of many-valued description logics with transitive closure of roles. Second, the original motivation of PDL was reasoning about the behaviour of computer programs \cite{Fischer1979}. From a more general perspective, PDL can be seen as a logic formalising reasoning about types of \emph{structured actions}, represented by ``action expressions''. On this reading, a Kripke frame consists of states and transitions between states labelled by types of action; for instance $R_{\alpha}st$ means that action of type $\alpha$ can be used to get from state $s$ to state $t$. States can be thought of as physical locations, states of a complex system such as a database or states of a computer during the run of a program; but states can also be thought of as ``states of the world'' that can be modified by actions of intelligent agents. PDL can be used to formalize reasoning about properties of actions that modify these kinds of states. One important example is correctness, related to the question if a specific kind of action is guaranteed to lead to a specific outcome when performed under specific circumstances. (This more general perspective makes PDL relevant to automated planning, for example.) Many-valued Kripke models can be seen as transition systems where transitions carry \emph{weights}; these can be costs or resources needed to perform a transition using the given action type. B\v{e}hounek \cite{Behounek2008} suggested a many-valued version of PDL for reasoning about costs of program runs that is close to our framework, but he did not establish completeness or decidability results. Let us now discuss a special case of the finitely-valued PDL framework giving rise to a natural class of weighted relational structures; we show that formulas of the PDL language are able to express interesting features of these structures. Let $\bm{N}$ be the FL-algebra of Example \ref{exam:Luk}, that is, $\bm{N} = (N, max, min, +_N, \to_N )$ where \[ a +_N b = min (a + b, N - 1) \quad \text{and} \quad a \to_N b = max (b - a, 0) \, , \] where $N \in \omega$ is non-empty. The set $N$ is seen as a \emph{weight scale} with $0$ representing zero weight (``for free'') and $N-1$ representing the maximal weight (considered ``infeasible''). The operation $+_{N}$, namely, sum bounded by the maximal weight, represents \emph{weight addition}. $N$ is given a (distributive) lattice structure by including $max$ as meet and $min$ as join; the associated lattice order $\sqsubseteq$ is defined as usual, $a \sqsubseteq b$ iff $min(a,b) = b$. Hence, $a \sqsubseteq b$ (i.e.\ $b \leq a$) means that weight $b$ is at most as big as weight $a$. The choice of $max$ as meet and $min$ as join---not the other way around---may seem unintuitive at first, but it yields the result that $a \sqsubseteq 0$ for all $a \in N$. It is important to note in this respect that $0$ is the identity element with respect to $+_{N}$. (Hence, choosing the natural ordering on $N$ as our lattice ordering would mean that each element of the lattice would be above the monoid identity, which is problematic given our definition of validity.) It is clear that $ a +_{N} b = b +_{N} a$. The residual $\to_{N}$ of $+_{N}$ is truncated subtraction or monus; the crucial feature of $\to_{N}$ is that $a \to_{N} b = 0$ iff $a \sqsubseteq b$ (iff $b \sqsubseteq a$). We note that $\bm{N}$ is isomorphic to the $N$-element {\L}ukasiewicz lattice $\textit{\bm{{\L}}}_{N}$ over $\{ \frac{k}{N-1} \mid k \in N \}$, but we prefer $\bm{N}$ to $\textit{\bm{{\L}}}_{N}$ as a representation of an $N$-element weight scale. $\bm{N}$-frames are weighted relational structures that can be informally interpreted in a number of ways. On the ``description reading'', for instance, states $s \in S$ are objects and $R_{\alpha}$ represent structured weighted relations between these objects. On the ``transition cost reading'', states can be seen as physical locations or states of a system and $R_\alpha st \in N$ is the cost of accessing state $t$ from $s$ by performing action $\alpha$ (hence, frames are weighted labelled transition systems). If $R_{\alpha} st = N - 1$, then we say that $t$ is not in relation $\alpha$ with $s$, or that $t$ cannot be accessed from $s$ by performing $\alpha$; if $R_{\alpha} st = 0$, then $t$ is ``clearly'' in relation $\alpha$ with $s$, or $t$ can be accessed from $s$ by $\alpha$ for free. Let us now discuss some properties of weighted relational structures that can be expressed by PDL formulas. Since $\bm{N}$ is $(N-1)$-involutive, i.e.\ $(a \Rightarrow (N-1)) \Rightarrow (N-1) = a$ for all $a \in \bm{N}$, we have \[ V(\langle \alpha\rangle\bar{0}, s) = \bigsqcup_{t \in S} \big( R_{\alpha} st +_{N} 0\big) = min \big \{ R_{\alpha} st \mid t \in S \big \} \, .\] In other words, $V(\langle \alpha\rangle\bar{0}, s)$ is the minimal guaranteed cost of performing $\alpha$ at $s$ (on the transition cost reading) or the maximal degree to which $s$ is $\alpha$-related to any object (on the description reading). Let us write simply $\alpha$ instead of $\langle\alpha\rangle\bar{0}$ if the context clears up any possible confusion. Note that $a \Rightarrow b$ is the difference between $b$ and $a$ if $a < b$ and $0$ otherwise. The following features of weighted relation structures can be expressed (we use the transition cost reading and the reader is invited to translate to the description reading): \begin{itemize} \item the minimal cost of performing $\alpha$ is at most $m$ (this is true in state $s$ if $V(\bar{m} \to \alpha, s) = 0$); the ``at least'' direction is expressed dually; \item performing $\alpha$ is at least as costly as performing $\beta$ (this is true in state $s$ if $V(\alpha \to \beta, s) = 0$); the ``at most'' direction is expressed dually; \item the difference between the minimal guaranteed cost of $\beta$ and $\alpha$ is at most $m$ (this is true in state $s$ if $V(\bar{m} \to (\alpha \to \beta), s) = 0$). \end{itemize} On the transition cost reading, atomic formulas in $Pr$ can be seen as representing various items that can be obtained at states for a given cost, with $V(\mathrm{p}, s)$ representing the cost of item $\mathrm{p}$ at $s$ (e.g.\ time needed to charge the battery at the charger location). Observe that $V(\langle\alpha\rangle \mathrm{p}, s) = \bigsqcup_{t \in S} (R_{\alpha} st +_{N} V(\mathrm{p}, t))$ is the minimal cost of getting from $s$ to a state $t$ by performing $\alpha$ and obtaining $\mathrm{p}$ at $t$; we may also say that this is the minimal guaranteed cost of obtaining $\mathrm{p}$ by $\alpha$. On the description reading, atomic formulas can be seen as expressing graded, imprecise or vague properties of objects; thus the value of $\langle \alpha\rangle \mathrm{p}$ at $s$ is the ``grade of truth'' of the statement that $s$ is $\alpha$-related to an object with property $\mathrm{p}$. The interesting case obtains where both the relation and the property are graded or vague; think of ``Alice was in contact with a person displaying symptoms of COVID-19''. We write $\varphi^{\alpha}$ instead of $\langle \alpha\rangle \varphi$. The following features of weighted relation structures can be expressed (we use the transition cost reading and the reader is again invited to translate to the description reading): \begin{itemize} \item the minimal cost of obtaining $\mathrm{p}$ by $\alpha$ is at most $m$ (this is true in state $s$ if $V(\bar{m} \to \mathrm{p}^{\alpha}, s) = 0$); the ``at least'' direction is expressed dually; \item obtaining $\mathrm{p}$ by $\alpha$ is at least as costly as obtaining $\mathrm{q}$ by $\beta$ (this is true in state $s$ if $V(\mathrm{p}^\alpha \to \mathrm{q}^\beta, s) = 0$); the ``at most'' direction is expressed dually; \item the difference between the minimal guaranteed cost of obtaining $\mathrm{q}$ by $\beta$ and obtaining $\mathrm{p}$ by $\alpha$ is at most $m$ (this is true in state $s$ if $V(\bar{m} \to (\mathrm{p}^\alpha \to \mathrm{q}^\beta), s) = 0$). \end{itemize} This cursory overview shows that the PDL language provides means to expressing a variety of features of weighted relational structures and so finitely-valued PDL can be used to formalize reasoning about these features. A more thorough exploration of expressivitiy and applications is left for another occasion. \section{Finite model property and decidability}\label{sec--DEC} In this section we prove that $Th(\bm{X})$ is decidable for all finite $\bm{X}$. We prove this by showing that each such $Th(\bm{X})$ has the bounded finite model property. The result is established using a many-valued generalization of the smallest filtration construction; see \cite{ConradieEtAl2017}, where the construction is applied to some many-valued modal logics with $\Box$ and $\Diamond$.\footnote{We are grateful to an anonymous reviewer for pointing the reference out.} Even though the decidability result is not surprising, we consider it to be a ``sanity check'' for the many-valued dynamic framework. We note that presence of canonical constants is not necessary for the decidability result (in contrast to the completeness result of \S\ref{sec--COM}). \begin{definition} The \emph{closure} of a set of formulas $\Psi$ is the smallest $\Phi \supseteq \Psi$ such that \begin{itemize} \item $\Phi$ is closed under subformulas (that is, if $\varphi \in \Phi$ and $\psi$ is a subformula of $\varphi$, then $\psi \in \Phi$); \item $[\alpha \cup \beta]\varphi \in \Phi$ implies $[\alpha]\varphi \in \Phi$ and $[\beta]\varphi \in \Phi$; \item $[\alpha\beta]\varphi \in \Phi$ implies $[\alpha][\beta]\varphi \in \Phi$; \item $[\alpha^{+}]\varphi \in \Phi$ implies $[\alpha][\alpha^{+}]\varphi \in \Phi$ and $[\alpha]\varphi \in \Phi$. \end{itemize} $\Phi$ is \emph{closed} iff $\Phi$ is the closure of $\Phi$. \end{definition} \begin{definition} For each set of formulas $\Phi$ and each model $\mathfrak{M}$, we define the binary two-valued equivalence relation $\approx_{\Phi}$ on states of $\mathfrak{M}$ by \[ s \approx_{\Phi} t \: \iff \: ( \forall \varphi \in \Phi) \big ( V(\varphi, s) = V(\varphi,t)\big ) \, .\] The equivalence class of $s$ under $\approx_{\Phi}$ will be denoted as $[s]_{\Phi}$ or just as $[s]$ if $\Phi$ is clear from the context. \end{definition} \begin{definition} Take an $\bm{X}$-valued model $\mathfrak{M}$ and a finite closed set $\Phi$. The \emph{filtration of $\mathfrak{M}$ through $\Phi$} is the $\bm{X}$-valued model $\mathfrak{M}^{\Phi} = (S^{\Phi}, R^{\Phi}, V^{\Phi})$ such that \begin{itemize} \item $S^{\Phi} = \{ [s] \mid s \in S \}$; \item $R^{\Phi}_{\mathrm{a}_m}([s],[t]) = \bigsqcup \big \{ R_{\mathrm{a}_m}(u,v) \mid s \approx_{\Phi} u \And t \approx_{\Phi} v \big \}$; $R^{\Phi}_{\alpha}$ for $\alpha \notin Ac$ is defined as in models; \item $V^{\Phi}(\mathrm{p}, [s]) = V(\mathrm{p}, s)$ for $\mathrm{p} \in \Phi$; $V^{\Phi}(\mathrm{p}, [s]) = 0^{\bm{X}}$ for $\mathrm{p} \notin \Phi$; $V^{\Phi}(\varphi, [s])$ for $\varphi \notin Pr$ is defined as in models. \end{itemize} \end{definition} It is clear that if $\Phi$ is the closure of a finite set $\Psi$, then $\Phi$ is finite. If $\Phi$ is finite, then so is $\mathfrak{M}^{\Phi}$; in fact, $|S^{\Phi}| \leq |\bm{X}|^{|\Phi|}$. We usually omit reference to $\Phi$ while discussing accessibility relations on $S^{\Phi}$ and we also write $\approx$ instead of $\approx_{\Phi}$. We will write $R_m$ instead of $R_{\mathrm{a}_m}$. In the rest of the section, we fix an $\bm{X}$-model $\mathfrak{M}$ and a finite closed set $\Phi$. \begin{lemma}\label{lem--DEC-filtration lemma} For all $\alpha \in \mathit{ACT}$ and all $x,y \in S$, \begin{enumerate}[label={(\alph*)}] \item $R_{\alpha}xy \sqsubseteq R_{\alpha}[x][y]$; \item For all $[\alpha]\varphi \in \Phi$, $V([\alpha]\varphi, x) \sqsubseteq R_{\alpha}[x][y] \Rightarrow V(\varphi, y)$. \end{enumerate} \end{lemma} \begin{proof} Both claims are established by induction on the complexity of $\alpha$. The base case of (a) holds by definition and the rest is established easily using the induction hypothesis. In the case of $\alpha = \beta^{+}$, we define for each $\pi \in \Pi(S)$ of length $n$ the sequence $[\pi] \in \Pi (S^{\Phi})$ of length $n$ by $[\pi](k) := [\pi(k)]$ for all $k < n$; it is then easy to establish by induction on $n$ that $R_{\beta} x \pi y \sqsubseteq R_{\beta} [x] [\pi] [y]$.) The base case of (b) is follows from the fact that, for all $x' \in [x]$ and $y' \in [y]$, $\bigsqcap_{z \in S} \big( R_{m} x' z \Rightarrow V(\varphi, z) \big) \cdot R_{m} x'y' \sqsubseteq V(\varphi, y')$ using the definition of $\approx_{\Phi}$, closure of $\Phi$ under subformulas and properties of FL-algebras. The fact itself follows easily from properties of FL-algebras. The induction step uses Lemma \ref{lem--PRE-PDL validities hold} and is easy; for instance, in the case $\alpha = \beta^{+}$ we may use the fact that, for all $x$ and $y$, $V([\beta](\varphi \land [\beta^{+}]\varphi, x) \sqsubseteq R_{\beta}[x][y] \Rightarrow V([\beta^{+}]\varphi, y)$ and hence, for all $s,t$ and $\pi \in \Pi(S)$, $V([\beta^{+}]\varphi, s) \sqsubseteq R_{\beta} [s][\pi][t] \Rightarrow V(\varphi, t)$ as required. \end{proof} \begin{lemma}\label{lem--DEC-filtration} For all models $\mathfrak{M}$, all $\varphi \in \Phi$ and $s \in \mathfrak{M}$, $V(\varphi, s) = V^{\Phi}(\varphi, [s])$. \end{lemma} \begin{proof} The proof is by induction on the complexity of $\varphi$. The base case $\varphi \in Pr$ holds by definition, the cases for constants and propositional connectives are trivial and the case $\varphi = [\alpha]\psi$ is established using Lemma \ref{lem--DEC-filtration lemma}. \end{proof} \begin{theorem}\label{thm--decidability} $Th(\bm{X})$ is decidable for each finite $\bm{X}$. \end{theorem} \begin{proof} Lemma \ref{lem--DEC-filtration} implies $\varphi \in Th(\bm{X})$ iff $\varphi$ is valid in all frames where $|S| \leq |\bm{X}|^{|\Phi|}$ where $\Phi$ is the closure of $\{ \varphi \}$. Now $m: = |\bm{X}| = m$, $n : = m^{|\Phi|}$ and let $n$-frames be the frames with $|S| \leq n$. There are at most \[ n \times m^{n^{2}} \] $n$-frames. On each $n$-frame, there are $n \times m^{\omega}$ models, but there are at most $n \times |\Phi| \times m$ possible ways to evaluate elements of $|\Phi|$ on an $n$-frame. Hence, there are at most \[ m^{n^{2} + 1} \times n^{2} \times |\Phi| \] models to check. It is not hard to show that there is an algorithm checking validity of formulas in finite models. \end{proof} \section{Completeness}\label{sec--COM} Bou et al.\ \cite{Bou2011} establish a general weak completeness result for modal logics based on finite commutative integral FL-algebras with canonical constants where $0$ is the bottom element. In this section we build on their work to show how a Hilbert-style axiomatic presentation of any finite commutative integral FL-algebra $\bm{X}$ with canonical constants can be extended to a sound and weakly complete axiomatization of PDL based on $\bm{X}$. The restriction to commutative FL-algebras seems to be necessary for our style of argument to go through and we discuss this at appropriate places in more detail; the restriction to integral FL-algebras is convenient. We leave generalizations of our result as an open problem. Fix a finite commutative integral FL-algebra $\bm{X}$ with canonical constants denoting elements of $\bm{X}$, together with a Hilbert-style axiomatic presentation $\mathsf{Log}(\bm{X})$ in the language $\mathcal{L}_{\bm{X}}$ that is \emph{strongly complete with respect to $\bm{X}$}. That is, we assume that $\varphi \in \mathcal{L}_{\bm{X}}$ is derivable from $\Gamma \subseteq \mathcal{L}_{\bm{X}}$ in $\mathsf{Log}(\bm{X})$, in symbols $\Gamma \vdash_{\mathsf{Log}(\bm{X})} \varphi$, iff each non-modal homomorphism $u : \mathcal{L}_{\bm{X}} \to \bm{X}$ such that $ 1 \sqsubseteq \bigsqcap u[\Gamma] $ satisfies $ 1 \sqsubseteq u(\varphi)$ (values $u([\alpha]\psi)$ of modal formulas under $u$ are arbitrary, so $u$ ``treats'' modal formulas as propositional atoms).\footnote{A function $f : \mathcal{L}_{\bm{X}} \to \bm{X}$ is a non-modal homomorphism iff $f(\bar{c}) = c$ and $f$ commutes with the propositional connectives $\oplus$ of $\mathcal{L}_{\bm{X}}$ and the corresponding operations $\oplus^{\bm{X}}$ on $\bm{X}$; we assume that $\land^{\bm{X}}$ is $\sqcap$ and $\lor^{\bm{X}}$ is $\sqcup$.} For the details on how $\mathsf{Log}(\bm{X})$ looks like, see \cite{Bou2011}. Since $\bm{X}$ is finite, $\vdash_{\mathsf{Log}(\bm{X})}$ is \emph{finitary} in the sense that if $\Gamma \vdash_{\mathsf{Log}(\bm{X})} \varphi$, then there is a finite $\Delta \subseteq \Gamma$ such that $\Delta \vdash_{\mathsf{Log}(\bm{X})} \varphi$. We note that $\vdash_{\mathsf{Log}(\bm{X})}$ is also \emph{monotonic} in the sense that if $\Gamma \vdash_{\mathsf{Log}(\bm{X})} \varphi$ and $\Gamma \subseteq \Delta$, then $\Delta \vdash_{\mathsf{Log}(\bm{X})} \varphi$. Since $\bm{X}$ is commutative, we have $a \backslash b = b \slash a$ and so we use only a single ``official'' implication operator $\to$; see \cite[p.\ 95]{Galatos2007}. Recall that $\varphi \leftrightarrow \psi := (\varphi \to \psi) \land (\psi \to \varphi)$; we define similarly $a \Leftrightarrow b := (a \Rightarrow b) \sqcap (b \Rightarrow a)$. \begin{definition} $\mathsf{PDL}(\bm{X})$ is the Hilbert-style axiom system extending $\mathsf{Log}(\bm{X})$ with the following axioms and rules (for all formulas $\varphi, \psi$, all action expressions $\alpha, \beta \in \mathit{ACT}$ and all canonical constants $\bar{c}$): \begin{center} \begin{minipage}[t]{0.4\linewidth} \begin{tabular}{ll} (A-$1$) & $[\alpha] \bar{1}$\\[1mm] (A-reg) & $[\alpha] \varphi \land [\alpha]\psi \to [\alpha] (\varphi \land \psi)$\\[1mm] (A-$\bar{c}$) & $[\alpha](\bar{c} \to \varphi) \leftrightarrow (\bar{c} \to [\alpha]\varphi)$ \\[2mm] (R-mon) & $\dfrac{\varphi \to \psi}{[\alpha] \varphi \to [\alpha] \psi}$ \end{tabular} \end{minipage} \qquad \begin{minipage}[t]{0.4\linewidth} \begin{tabular}{ll} (A-$\cup$) & $[\alpha \cup \beta]\varphi \leftrightarrow ([\alpha]\varphi \land [\beta]\varphi)$\\[1mm] (A-$;$) & $[\alpha\beta] \varphi \leftrightarrow [\alpha][\beta]\varphi$\\[1mm] (A-$+$) & $[\alpha^{+}]\varphi \leftrightarrow [\alpha](\varphi \land [\alpha^{+}]\varphi)$\\[2mm] (R-$+$) & $\dfrac{\varphi \to [\alpha]\varphi}{\varphi \to [\alpha^{+}]\varphi}$ \end{tabular} \end{minipage} \end{center} The notions of proof, derivability, theorem and a formula derivable from a set of formulas are defined as usual (see \cite{Bou2011}). $\mathsf{Thm}(\mathsf{PDL}(\bm{X}))$ is the set of theorems of $\mathsf{PDL}(\bm{X})$. \end{definition} \noindent Since $\bm{X}$ is fixed, we write $\mathsf{L}$ instead of $\mathsf{Log}(\bm{X})$, $\mathsf{PDL}$ instead of $\mathsf{PDL}(\bm{X})$, $\mathsf{Thm}$ instead of $\mathsf{Thm}(\mathsf{PDL}(\bm{X}))$ and $\mathcal{L}$ instead of $\mathcal{L}_{\bm{X}}$ for the rest of this section. \begin{theorem}\label{thm--soundness} If $\varphi$ is a theorem of $\mathsf{PDL}$, then $\varphi$ is valid in the class of all $\bm{X}$-frames. \end{theorem} \begin{proof} The axioms and the rule in the left column are taken from \cite{Bou2011}. Validity of the axioms in the right column in all FL-algebras was established in Lemma \ref{lem--PRE-PDL validities hold}. To show that the rule (R-$+$) preserves validity in models, assume that $V(\varphi,s) \sqsubseteq V([\alpha]\varphi, s)$ for all $s$ in an arbitrary model. Take some $t$ and assume that $a \sqsubseteq V(\varphi, t)$; we prove that $a \sqsubseteq R_{\alpha^{+}} tu \Rightarrow V(\varphi, u)$ for all $u$. The claim to be proved is equivalent to $(\forall \pi \in \Pi(S))(a \sqsubseteq R_{\alpha} t \pi u \Rightarrow V(\varphi, u))$. This claim is easily established by induction on the length of $\pi$. \end{proof} \noindent We note that, without the assumption of commutativity, versions of (A-$\bar{c}$) are not sound; the axiom is used in the proof of Lemma \ref{lem--COM-Bou} which is in turn applied in most of our arguments below. From now on, let $S$ be the set of non-modal homomorphisms $s: \mathcal{L} \to \bm{X}$ such that $s[\mathsf{Thm}] = \{ 1 \}$ and let $\Phi$ be a fixed finite closed set. \begin{definition} The \emph{$\Phi$-equivalence relation} on $S$ is an $\bm{X}$-valued binary relation $\sim_{\Phi}$ on $S$ defined by \[ s \sim_{\Phi} t \quad := \quad \bigsqcap_{\varphi \in \Phi} \big ( s(\varphi) \Leftrightarrow t(\varphi) \big ) \, . \] \end{definition} \noindent If $\Phi$ is clear from the context, we will write $s \sim t$ or just $st$ instead of $s \sim_{\Phi} t$. \begin{lemma}\label{lem--COM-Phi-equivalence} The relation $\sim_{\Phi}$ is an $\bm{X}$-valued equivalence relation, that is, (a) $1 \sqsubseteq s \sim s$, (b) $s \sim t = t \sim s$ and (c) $(s \sim t)(t \sim u) \sqsubseteq s \sim u$, for all $s,t,u \in S$. \end{lemma} \begin{proof} Claims (a) and (b) are clear; claim (c) follows from Lemma \ref{lem--APP-Properties of L}. \end{proof} Completeness proofs for two-valued PDL typically use a filtration-like construction of the canonical model, where states are (or boil down to) equivalence classes of states taken from some other structure. A natural approach in our case would be to take ``equivalence classes'' of non-modal homomorphisms under $\sim$, where $s \sim t$ expresses ``how much equivalent'' $s$ and $t$ are with respect to $\Phi$. However, in our case a simpler approach is available. We take $S$ itself as the set of states of the canonical model and we refer to $\Phi$ only in the definition of the canonical $R_{\alpha}$, which is a generalization of the definition of accessibility relations in the greatest filtration of a Kripke model. \begin{definition} The \emph{canonical model modulo $\Phi$} is $\mathfrak{M} = (S, R, V)$ where \begin{itemize} \item $S$ is the set of non-modal homomorphisms $s: \mathcal{L} \to \bm{X}$ such that $s[\mathsf{Thm}] = \{ 1 \}$; \item $R_{m} st := \bigsqcap_{[m]\varphi \in \Phi} \big ( s([m]\varphi) \Rightarrow t(\varphi) \big )$ for all $\mathrm{a}_m \in Ac$ and $R_{\alpha}st$ for $\alpha \not\in Ac$ is defined as in models; \item $V(\mathrm{p}, s) := s(\mathrm{p})$ and $V(\varphi, s)$ for $\varphi \not\in Pr$ is defined as in models. \end{itemize} We define for each $\alpha$ the relation $R^{\mathcal{L}}_{\alpha}$ on $S$ by $R^{\mathcal{L}}_{\alpha} st := \bigsqcap_{\varphi \in \mathcal{L}} \big ( s([\alpha]\varphi) \Rightarrow t(\varphi) \big )$. \end{definition} Note that $R^{\mathcal{L}}_{n} st \sqsubseteq R_{n} st$ for all $\mathrm{a}_n \in Ac$ and all $s,t$ since $R^{\mathcal{L}}_{n}$ ``cares'' about more formulas. $R^{\mathcal{L}}_{\alpha}$ is the usual canonical many-valued accessibility relation, see \cite{Bou2011}, but we cannot use it here because of the presence of the Kleene plus iteration operator in $\mathit{ACT}$, similarly as in the case of two-valued PDL. The following lemma states some properties of $R^{\mathcal{L}}_{\alpha}$ that will be useful in our proofs; the proof of the lemma can be found in \cite{Bou2011} (the logics studied there are mono-modal, but the same approach applies here). \begin{lemma}\label{lem--COM-Bou} The following holds for all $\alpha \in \mathit{ACT}$ and all $s \in S$ of the canonical model: \begin{enumerate}[label={(\alph*)}] \item For all $t$, $R^{\mathcal{L}}_{\alpha}st = \bigsqcap_{\varphi \in \mathcal{L}} \big \{ t(\varphi) \mid 1 \sqsubseteq s([\alpha]\varphi) \big \}$ (\cite{Bou2011}, Proposition 4.1.); \item For all $\varphi \in \mathcal{L}$, $s([\alpha]\varphi) = \bigsqcap_{u \in S} \big \{ R^{\mathcal{L}}_{\alpha} su \Rightarrow u(\varphi) \big \}$ (\cite{Bou2011}, Lemma 4.8.). \end{enumerate} \end{lemma} \begin{lemma}\label{lem--COM-Boxed formulas and R} For all $[\alpha]\varphi \in \Phi$ and all $s,t \in S$, $s([\alpha]\varphi) \sqsubseteq R_{\alpha} st \Rightarrow t(\varphi)$. \end{lemma} \begin{proof} The claim is proved by induction on the complexity of $\alpha$. The base case is established as follows. We know that $s([n]\varphi) \cdot (s([n]\varphi) \Rightarrow t(\varphi)) \sqsubseteq t(\varphi)$; from this $s([n]\varphi) \cdot R_{n} st \sqsubseteq t(\varphi)$ follows by the definition of $R_n$. The cases of choice and composition in the induction step are straightforward. The case $\alpha = \beta^{+}$ is established by showing that, for all $\pi \in \Pi(S)$, all $s,t$, and all $\varphi$ such that $[\beta^{+}]\varphi \in \Phi$, $s([\beta^{+}]\varphi) \sqsubseteq R_{\beta} s \pi t \Rightarrow t(\varphi)$. This claim, call it (A), follows from the claims ($s,t$ and $[\beta^{+}]\varphi \in \Phi$ are fixed) \begin{itemize} \item[(B)] $s([\beta]\varphi) \sqsubseteq R_{\beta} st \Rightarrow t(\varphi)$; \item[(C)] for all $\sigma \in \Pi(S)$ and all $u$, $s([\beta^{+}]\varphi) \sqsubseteq R_{\beta} s \sigma u \Rightarrow u([\beta^{+}]\varphi)$. \end{itemize} The proof of (C) is left to the reader; (B) holds by the induction hypothesis. \end{proof} \begin{lemma}\label{lem--COM--R closed under Phi equivalence} For all $\alpha$ and $s,t,u$, $R_{\alpha} su(ut) \sqsubseteq R_{\alpha} st$. \end{lemma} \begin{proof} We argue by induction on the complexity of $\alpha$. The base case is established as follows. If $a \sqsubseteq R_{n} su(ut)$, then, by definition, $a \sqsubseteq \bigsqcap_{[n]\varphi \in \Phi} \big ( s([n]\varphi) \Rightarrow u(\varphi) \big ) (ut)$. Hence, for all $[n]\varphi \in \Phi$, $a \sqsubseteq \big ( s([n]\varphi) \Rightarrow u(\varphi) \big) \big ( u(\varphi) \Rightarrow t(\varphi) \big )$ by the definition of $u \sim t$ and monotonicity of monoid multiplication (also, $[n]\varphi \in \Phi$ implies $\varphi \in \Phi$). It follows by the properties of FL-algebras that $a \sqsubseteq \big ( s([n]\varphi) \Rightarrow t(\varphi) \big )$. Since $[n]\varphi \in \Phi$ was arbitrary, we obtain $a \sqsubseteq R_{n} st$. All cases of the induction step are easy. \end{proof} \begin{definition} For all $\alpha$ and $s$, we define the following formula: \[ R_{\alpha}s \quad := \quad \bigvee_{x \in S} \Big ( \overline{R_{\alpha} s x }\: \cdot \bigwedge_{\varphi \in \Phi} \big ( \overline{x(\varphi)} \leftrightarrow \varphi \big)\Big ) \] \end{definition} \noindent Note that $R_{\alpha} s$ is well defined even though $S$ is infinite -- there are only finitely many possible values of $R_{\alpha} sx$ for $x \in S$, as $\bm{X}$ is finite. Note also that $t( R_{\alpha}s) = \bigsqcup_{x \in S} \big ( R_{\alpha} sx(xt) \big)$. \begin{lemma}\label{lem--COM-RPhi formula} For all $s,t$ and $\alpha$, $t( R_{\alpha} s ) = R_{\alpha} st$. \end{lemma} \begin{proof} First, $R_{\alpha} st \sqsubseteq R_{\alpha} st (tt) $ by Lemma \ref{lem--COM-Phi-equivalence}(a), and $R_{\alpha} st(tt) \sqsubseteq \bigsqcup_{x \in S} \big ( R_{\alpha} sx (xt) \big ) = t(R_\alpha s)$. Second, $R_{\alpha} sx (xt) \sqsubseteq R _{\alpha} st$ for all $x \in S$ by Lemma \ref{lem--COM--R closed under Phi equivalence}. Hence, $\bigsqcup_{x \in S} R_{\alpha} sx(xt)$ and so $t(R_\alpha s) \sqsubseteq R _{\alpha} st$. \end{proof} \begin{lemma}\label{lem--COM-R under RPhi} For all $s,t \in S$ and all $\alpha \in \mathit{ACT}$, $R^{\mathcal{L}}_{\alpha}st \sqsubseteq R_{\alpha} st$. \end{lemma} \begin{proof} Induction on the complexity of $\alpha$. The base case follows from definition. To establish the induction step, we reason by cases. Note that the induction hypothesis is equivalent to the claim that, for all $\alpha,\beta$ and $x$, $1 \sqsubseteq x([\alpha] R_{\alpha} x)$ and $1 \sqsubseteq x([\beta] R_{\beta}x)$ by Lemmas \ref{lem--COM-Bou}(b) and \ref{lem--COM-RPhi formula}. If $a \sqsubseteq R^{\mathcal{L}}_{\alpha \cup \beta} st$, then $a \sqsubseteq \bigsqcap_{\varphi \in \mathcal{L}} \big \{ t(\varphi) \mid 1 \sqsubseteq s([\alpha \cup \beta]\varphi) \big\}$ by Lemma \ref{lem--COM-Bou}(a). By the definition of $S$, this entails $a \sqsubseteq \bigsqcap \big\{ t(\varphi) \mid 1 \sqsubseteq s([\alpha]\varphi) \sqcap s([\beta]\varphi) \big\}$. By the induction hypothesis, $1 \sqsubseteq s([\alpha]R_\alpha s)$ and $1 \sqsubseteq s([\beta]R_\beta s)$. Hence, $1 \sqsubseteq s([\alpha](R_{\alpha}s \lor R_{\beta}s))$ and $1 \sqsubseteq s([\beta](R_{\alpha}s \lor R_{\beta}s))$ by the definition of $S$. It follows that $a \sqsubseteq t(R_{\alpha}s) \sqcup t(R_{\beta}s)$. By Lemma \ref{lem--COM-RPhi formula}, $a \sqsubseteq R_{\alpha}st \sqcup R_{\beta}st$ and so $a \sqsubseteq R_{\alpha \cup \beta} st$. If $a \sqsubseteq R^{\mathcal{L}}_{\alpha\beta} st$, then $a \sqsubseteq \bigsqcap_{\varphi \in \mathcal{L}} \big \{ t(\varphi) \mid 1 \sqsubseteq s([\alpha\beta] \varphi) \big\}$ by Lemma \ref{lem--COM-Bou}(a) and so $a \sqsubseteq \bigsqcap_{\varphi \in \mathcal{L}} \big \{ t(\varphi) \mid 1 \sqsubseteq s([\alpha][\beta] \varphi) \big\}$ by the definition of $S$. For all $x$ and $y$, $R_{\alpha}^{\mathcal{L}}sx R_{\beta}^{\mathcal{L}} xy \sqsubseteq y(R_{\alpha\beta}s)$ by the induction hypothesis, Lemma \ref{lem--COM-RPhi formula} and the definition of $R_{\alpha\beta}$. Hence, for all $x$, $R_{\alpha}^{\mathcal{L}}sx \sqsubseteq x ([\beta]R_{\alpha\beta}s)$ by residuation and Lemma \ref{lem--COM-Bou}(b); from this is follows that $1 \sqsubseteq s ([\alpha][\beta]R_{\alpha\beta}s)$ by another application of residuation and Lemma \ref{lem--COM-Bou}(b). Therefore, $a \sqsubseteq t(R_{\alpha\beta} s)$ and so $a \sqsubseteq R_{\alpha\beta} st$ by Lemma \ref{lem--COM-RPhi formula}. Finally, we discuss the case of $\alpha^{+}$. Fix $s$; we write $F$ instead of $R_{\alpha^{+}} s$. Note that $R _{\alpha^{+}}$ is a transitive relation extending $R^{\mathcal{L}}_{\alpha}$. Hence, for all $t,u \in S$, $u(F) \cdot R^{\mathcal{L}}_{\alpha} ut \sqsubseteq t(F)$ by Lemma \ref{lem--COM-RPhi formula} and the induction hypothesis applied to $R^{\mathcal{L}}_{\alpha}$; we obtain from this $u(F) \sqsubseteq u([\alpha]F)$ for all $u \in S$ by Lemma \ref{lem--COM-Bou}(b). Hence, by definition of $S$, we have $ F \to [\alpha]F \: \in \: \mathsf{Thm} $. Hence, using (R-$+$), we have $F \to [\alpha^{+}]F \: \in \: \mathsf{Thm}$ and, using (R-mon) and (A-$+$), we obtain $[\alpha]F \to [\alpha^{+}]F \in \mathsf{Thm}$. By the induction hypothesis we have $R_{\alpha}^{\mathcal{L}} st \sqsubseteq R_{\alpha} st \sqsubseteq R_{\alpha^{+}} st$ for all $t$ and so $1 \sqsubseteq R_{\alpha}^{\mathcal{L}} st \Rightarrow t(F)$ for all $t$ by Lemma \ref{lem--COM-RPhi formula}. This means that $1 \sqsubseteq s([\alpha]F)$ and so $1 \sqsubseteq s([\alpha^{+}]F)$ which means that $R^{\mathcal{L}}_{\alpha^{+}} st \sqsubseteq t(F)$ for all $t$ by Lemma \ref{lem--COM-Bou}(b). Hence, $R^{\mathcal{L}}_{\alpha^{+}} st \sqsubseteq R _{\alpha^{+}} st$ by Lemma \ref{lem--COM-RPhi formula}. \end{proof} \begin{lemma}\label{lem--COM-Truth lemma} For all $\varphi \in \Phi$ and $s \in S$, $s(\varphi) = V(\varphi, s )$. \end{lemma} \begin{proof} Induction on the complexity of $\varphi$. The base case holds by definition and the cases for non-modal formulas and canonical constants are straightforward. Finally, $s([\alpha]\varphi) \sqsubseteq V([\alpha]\varphi, s )$ holds thanks to Lemma \ref{lem--COM-Boxed formulas and R} and $V([\alpha]\varphi, s ) \sqsubseteq s([\alpha]\varphi)$ holds thanks to Lemma \ref{lem--COM-Bou}(b) and Lemma \ref{lem--COM-R under RPhi}. \end{proof} \begin{theorem}\label{thm--completeness} For all finite commutative integral $\bm{X}$ with canonical constants, $\varphi$ is valid in all $\bm{X}$-frames iff $\varphi$ is a theorem of $\mathsf{PDL}(\bm{X})$. \end{theorem} \begin{proof} Soundness is established by Theorem \ref{thm--soundness}. Completeness is established as usual. If $\varphi$ is not in $\mathsf{Thm}$, then $\mathsf{Thm} \not\vdash_{\mathsf{L}} \varphi$ since $\mathsf{Thm}$ is obviously closed under $\vdash_{\mathsf{L}}$. By strong completeness of $\mathsf{L}$, there is a non-modal homomorphism from $\mathcal{L}$ to $\bm{X}$ such that $s[\mathsf{Thm}] = \{ 1 \}$ and $s(\varphi) \neq 1$. Let $\Phi$ be the closure of $\{ \varphi \}$; $\varphi$ is not valid in the canonical model modulo $\Phi$ by Lemma \ref{lem--COM-Truth lemma}. \end{proof} \section{On Kleene star and test}\label{sec--ISS} Our syntactic presentation of propositional dynamic logic differs from the standard presentation in two important respects, namely, (i) our action operators do not include the \emph{Kleene star}, but rather the Kleene plus operator; (ii) we do not include the \emph{test operator}. Kleene star and test are instrumental in the ability of classical PDL to express standard programming constructs such as while loops and conditionals (test suffices for the latter). In this section we discuss these omissions. Concerning the Kleene star, Proposition \ref{prop--PRE-Reflexive transitive closure} suggests that, working with frames based on finite integral FL-algebras, we can define, for all $\alpha \in \mathit{ACT}$ and $\varphi \in Fm(\mathcal{L}_{\bm{X}})$, \[ [\alpha^{*}]\varphi := [\alpha^{+}]\varphi \land \varphi \] as a semantically equivalent surrogate for formulas with the Kleene star. For instance, $[(\mathrm{a} \cup \mathrm{b})^{*} ; \mathrm{a}^{*}]\mathrm{p}$ is short for $[(\mathrm{a} \cup \mathrm{b})^{+}]([\mathrm{a}^{+}]\mathrm{p} \land \mathrm{p}) \land ([\mathrm{a}^{+}]\mathrm{p} \land \mathrm{p})$. However, it is clear that not all action expressions in $\mathit{STA}$ can be expressed by action expressions in $\mathit{ACT}$. Therefore, for example, $[(\mathrm{a}^{*}; \mathrm{b})^{*}]\mathrm{p}$ is not a well-formed formula since $\mathrm{a}^{*} \not\in \mathit{ACT}$. The technical problem that precluded us from working with Kleene star as a primitive operator is related to Lemma \ref{lem--COM--R closed under Phi equivalence}. Take the reflexive transitive closure $R^{*}_{\alpha}$ of $R_{\alpha}$, defined as in Proposition \ref{prop--PRE-closure}. The issue is that Lemma \ref{lem--COM--R closed under Phi equivalence} fails if Kleene star is a primitive operator and we define $R_{\alpha^{*}} := R_{\alpha}^{*}$. In particular, if $s = u \neq t$, then $R^{*}_{\alpha}su(ut) \sqsubseteq R_{\alpha}^{*}st$ boils down to $s \sim t \sqsubseteq R_{\alpha^{+}}st$, which does not hold in all canonical models. (Take the canonical $\bm{2}$-model modulo the closure $\Phi$ of $\Psi = \{ [\mathrm{a}]\bot \}$. As both $\Psi \cup \{ \mathrm{p}_0 \}$ and $\Psi \cup \{ \mathrm{p}_1 \}$ are consistent, there are two distinct $s,t$ such that $s \sim_{\Phi} t$ equals $1$, but $R_{\mathrm{a}^{+}}st$ equals $0$.) Concerning test, a natural semantic interpretation of $\varphi ?$, endorsed also in \cite{Hughes2006,Liau1999}, is \[ R_{\varphi ?} (s,t) = \begin{cases} V(\varphi, s) & \text{if } s = t \\ \bot^{\bm{X}} & \text{otherwise.} \end{cases} \] However, Lemma \ref{lem--COM--R closed under Phi equivalence} turns out to be problematic for such a relation as well. (Take the model from the previous paragraph and let $\varphi = [\mathrm{a}]\bot$; clearly $R_{\varphi ?} ss (st)$ equals $1$, but $R_{\varphi ?} st$ equals $0$.) It is clear that a more substantial modification of our completeness argument is needed to accommodate logics with Kleene star and test. This is an interesting problem we leave open here. \section{Conclusion}\label{sec--CON} We have studied a general framework for many-valued versions of Propositional Dynamic Logic where both formulas in states and accessibility relations between states of a Kripke model are evaluated in a finite FL-algebra. We established a general decidability result and we provided a general completeness argument for PDLs based on commutative integral FL-algebras with canonical constants. We build on previous work on many-valued modal logic and our techniques are generalizations of the arguments used in the two-valued case; however, to the best of our knowledge, the technical results presented here are the first decidability and completeness results on PDL with many-valued accessibility relations. As our discussion of the informal interpretations of the framework suggests, many-valued PDL has links to existing research in description logics and potential applications in reasoning about weighted labelled transition systems. Our paper also suggests a number of topics for future research. We would like to mention especially the addition of test and further work on the standard version of PDL with primitive Kleene star in the many-valued setting. Another topic are generalizations of our results beyond finite (commutative integral) FL-algebras with canonical constants; in many cases the work here would require modifications of existing techniques used in completeness arguments for many-valued modal logics without ``structured'' modal operators. Finally, informal interpretations and applications of our framework need to be explored in more detail.

2b4b8f37f6aa9900fe773bd8c5d7d5b71c775d38

\section*{Introduction} Characterization of materials often involves investigating their interaction with light. Optical absorption spectroscopy is one of the key experimental techniques for such characterization, and the simulation of optical absorption spectra is essential for interpreting experimental observations and predicting design rules for materials with desired properties. In recent years, absorption spectra of condensed systems have been successfully predicted by solving the Bethe-Salpeter equation (BSE)\cite{salpeter_relativistic_1951,hedin1965new,hanke1980many,onida1995ab,albrecht1997ab,albrecht1998ab,albrecht1998excitonic,benedict1998optical,rohlfing1998electron,rohlfing2000electron,blase2018bethe} in the framework of many-body perturbation theory (MBPT).\cite{strinati1988application,onida_electronic_2002,martin2016interacting,ping_electronic_2013,govoni_2018,golze2019gw} However, for large and complex systems, the use of MBPT is computationally demanding.\cite{govoni_large_2015,seo2016design,gaiduk2016photoelectron,scherpelz_2016,seo2017designing,mcavoy_phonon_2018,smart_fundamental_2018,gaiduk2018electron,gerosa2018role} It is thus desirable to develop methods that can improve the efficiency of optical spectra calculations, especially if results at finite temperature (T) are desired. Simulation of absorption spectra at finite T can be achieved by performing, e.g., first principles molecular dynamics (FPMD)\cite{car1985unified} and by solving the BSE for uncorrelated snapshots extracted from FPMD trajectories. A spectrum can then be obtained by averaging over the results obtained for each snapshot.\cite{garbuio2006ab,lu2008dielectric,bernasconi2010statistical,nguyen_finite-field_2019} Several schemes have been proposed in the literature to reduce the computational cost of solving the BSE,\cite{marsili2017large,elliott2019koopmans,henneke2020fast} including an algorithm that avoids the explicit calculation of virtual single particle electronic states, as well as the storage and inversion of large dielectric matrices.\cite{rocca_ab_2010,rocca_solution_2012} Recently, a so-called finite-field (FF) approach\cite{ma_finite-field_2019,nguyen_finite-field_2019} has been proposed, where the calculation of dielectric matrices is bypassed; rather the key quantities to be evaluated are screened Coulomb integrals, which are obtained by solving the Kohn-Sham (KS) equations\cite{hohenberg1964inhomogeneous,kohn1965self} for the electrons in a finite electric field. The ability to describe dielectric screening through finite field calculations also led to the formulation of GW\cite{ma_finite-field_2019,ma2020correction} and BSE\cite{nguyen_finite-field_2019} calculations beyond the random phase approximation (RPA), and of a quantum embedding approach\cite{ma_quantum_2020,ma2020first} scalable to large systems. From a computational standpoint, one important aspect of solving the Kohn-Sham equations in finite field is that the calculations can be straightforwardly combined with the recursive bisection algorithm\cite{gygi_compact_2009} and thus, by harnessing orbital localization, one may greatly reduce the number of screened Coulomb integrals that need to be evaluated. Importantly, the workload to compute those integrals is of $O(N^4)$, irrespective of whether semilocal or hybrid functionals are used.\cite{nguyen_finite-field_2019} In spite of the improvement brought about by the FF algorithm and the use of the bisection algorithm, the solution of the BSE remains a demanding task. One of the quantities particularly challenging to evaluate is the dielectric matrix of the system, that describes many-body screening effects between the interacting electrons. Intuitively we can understand a dielectric matrix as a complex filter that connects the bare (i.e., unscreened) Coulomb interaction between the electrons to an effective, screened Coulomb interaction. Such screened interaction is used in MBPT to approximately account for electronic correlation effects, when solving the Dyson equation (GW) and the BSE. Here we turn to machine learning (ML), in order to tackle the challenge of evaluating the dielectric matrix. Specifically, for a chosen atomic configuration of a solid or a molecule, we use ML techniques to derive a mapping from the unscreened to the screened Coulomb interaction, thus deriving a model of the dielectric screening. Once such a model is available, it can be re-used for multiple configurations sampled in a FPMD at finite temperature, without the need to recompute a complex dielectric matrix for each snapshot. Hence the use of a ML-derived model may greatly improve the efficiency of the calculation of finite T absorption spectra, provided the dielectric screening is weakly dependent on atomic configurations explored as a function of simulation time. We will show below that this assumption is indeed verified for several disordered systems, including liquid water and Si/water interfaces at ambient conditions and silicon clusters. Importantly, the use of ML-derived models leads to a reduction of 1 to 2 orders of magnitude in the computational workload required to obtain the dielectric screening for the simulation of optical absorption spectra at finite temperature. Another important advantage of the ML-derived dielectric screening is that it provides insight into the approximate screening parameters used in the derivation of hybrid functionals for time-dependent DFT (TDDFT) calculations, including dielectric-dependent hybrid (DDH) functionals. \cite{shimazaki2009first,skone2014self,gerosa2017accuracy,chen2018nonempirical,sun_low-cost_2020} We emphasize that the strategy adopted here is different in spirit from strategies that use ML to infer structure-property relationships\cite{montavon2013machine,brockherde2017bypassing, welborn2018transferability,schleder_dft_2019,ryczko2019deep,noe_machine_2020,hase2020designing,sutton2020identifying,bogojeski2020quantum} or relationships between computational and experimental data\cite{stein_machine_2018}. We do not seek to relate structural properties of a molecule or a solid to its absorption spectrum. Rather, either we consider a known microscopic structure of the system or we determine the structure by carrying out first principles MD (e.g., in the case of liquid water or a solid/liquid interface). Then, for a given atomistic configuration we use ML techniques to obtain the model between the unscreened and the screened Coulomb interaction, and we use such a model in the solution of the BSE for multiple configurations. Hence the method proposed here is conceptually different from the approaches previously adopted to predict the absorption spectra of molecules or materials using ML.\cite{gastegger2017machine,stein_machine_2018,Ye11612,ghosh_deep_2019,carbone2020machine,xue2020machine}. For example, Ghosh et al.\cite{ghosh_deep_2019} predicted molecular excitation spectra from the knowledge of molecular structures at zero T, by using neural networks trained with a dataset of 132531 small organic molecules. Carbone et al.\cite{carbone2020machine} mapped molecular structures to X-ray absorption spectra using message-passing neural networks, and a dataset of $\sim$134000 small organic molecules. Xue et al.\cite{xue2020machine} focused on two specific molecules and used a kernel ridge regression model trained with a minimum of several hundred molecular geometries and their corresponding excitation energies and oscillator strengths computed at the TDDFT\cite{runge_density-functional_1984} level; they then used the results to predict the excitation energies and oscillator strengths of an ensemble of geometries and absorption spectra. All of these methods seek to relate structure to function (absorption spectra). The method presented here uses instead ML to replace a computationally expensive step in first principles simulations, and as we show below, leads to physically interpretable results. The rest of the paper is organized as follows. In the next section, we briefly summarize our computational strategy. We then discuss homogeneous systems, including liquid water and periodic solids, followed by results for heterogeneous and finite systems. We conclude by highlighting the innovation and key results of our work. \section*{Methods} We first briefly summarize the technique used here to solve the BSE, including the use of bisection techniques to improve the efficiency of the method. We then describe the method based on ML to obtain the dielectric screening entering the BSE, including the description of the training set of integrals. These integrals are computed for a chosen configuration of a molecule or a solid. Using the linearized Liouville equation\cite{walker_efficient_2006,rocca_ab_2010,rocca_solution_2012,nguyen_finite-field_2019} and the Tamm-Dancoff approximation\cite{hirata1999time}, the absorption spectrum of a solid or molecule can be computed from DFT\cite{hohenberg1964inhomogeneous,kohn1965self} single particles eigenfunctions as: \begin{equation} S(\omega) \propto \sum_{i=1}^3 \sum_{v=1}^{n_{occ}} \langle \psi_{v} | r_i | a_v^i(\omega) \rangle + c.c. \label{eq:spectrum} \end{equation} where $\omega$ is the absorption energy, $r_i$ are the Cartesian components of the dipole operator, $n_{\text{occ}}$ is the total number of occupied orbitals, and $\mid\psi_{v}\rangle$ is the $v$-th occupied orbital of the unperturbed KS Hamiltonian, $\hat{H}^0$, corresponding to the eigenvalue $\epsilon_v$. The functions $\mid a_v^i \rangle$ are obtained from the solution of the following equation:\cite{rocca_ab_2010,rocca_solution_2012,nguyen_finite-field_2019} \begin{equation} \sum^{n_{\text{occ}}}_{v^\prime=1} (\omega\delta_{vv^\prime}-D_{vv^\prime}-\mathcal{K}^{1e}_{vv^\prime}+\mathcal{K}^{1d}_{vv^\prime}) \mid a^i_{v^\prime}\rangle = \hat{P}_{c} \hat{r}_{i} \mid \psi_{v} \rangle \label{eq:linear_eq} \end{equation} where \begin{equation} D_{vv^\prime}\mid a^i_{v^\prime}\rangle = \hat{P}_{c} (\hat{H}^0-\epsilon_v)\delta_{vv^\prime} \mid a^i_{v^\prime}\rangle, \label{eq:Da} \end{equation} \begin{equation} \mathcal{K}_{vv^\prime}^{1e}\mid a_{v^\prime}^{i}\rangle=2\hat{P}_{c}\left(\int d\mathbf{r^{\prime}}V_c(\mathbf{r},\mathbf{r^{\prime}})\psi_{v^\prime}^{*}(\mathbf{r^{\prime}})a_{v^\prime}^{i}(\mathbf{r^{\prime}})\right)\psi_{v}(\mathbf{r}), \label{eq:K1e_a} \end{equation} \begin{equation} \mathcal{K}_{vv^\prime}^{1d}\mid a_{v^\prime}^{i}\rangle=\hat{P}_{c}\tau_{vv^\prime}(\mathbf{r})a^i_{v^\prime}(\mathbf{r}), \label{eq:K1d_a} \end{equation} $\hat{P}_c=1-\sum^{n_\text{occ}}_{v=1}\mid \psi_{v}\rangle\langle \psi_{v}\mid$ is the projector on the unoccupied manifold, and $V_c=\frac{e^2}{\mid\mathbf{r}-\mathbf{r^{\prime}}\mid}$ is the unscreened Coulomb potential. Following the derivation reported by Nguyen et al.,\cite{nguyen_finite-field_2019} we defined screened Coulomb integrals, $\tau_{vv^\prime}$, entering Eq.~\ref{eq:K1d_a}, as: \begin{eqnarray} \tau_{vv^\prime}(\mathbf{r}) &=& \int W(\mathbf{r},\mathbf{r^{\prime}})\psi_{v}(\mathbf{r^{\prime}})\psi_{v^\prime}^{*}(\mathbf{r^{\prime}})d\mathbf{r}'\label{eq:tau0}\\ &=&\tau^u_{vv^\prime}(\mathbf{r})+\Delta\tau_{vv^\prime}(\mathbf{r}),\label{eq:tau} \end{eqnarray} where the screened Coulomb interaction $W$ is given by $W=\epsilon^{-1} V_c$, and $\epsilon^{-1}$ is the inverse of the dielectric matrix (dielectric screening). Analogously, unscreened Coulomb integrals, $\tau^u_{vv^\prime}$, are defined as: \begin{equation} \tau^u_{vv^\prime}(\mathbf{r})=\int V_c(\mathbf{r,r^{\prime}})\psi_{v}(\mathbf{r^{\prime}})\psi^*_{v^\prime}(\mathbf{r^{\prime}})d\mathbf{r^{\prime}}. \label{eq:tauu} \end{equation} By carrying out finite field calculations\cite{ma_finite-field_2019, ma2020correction, nguyen_finite-field_2019}, one can obtain screened Coulomb integrals without an explicit evaluation of the dielectric matrix (Eq.~\ref{eq:tau0}), but rather by adding to the unscreened Coulomb integrals the second term on the right hand side of Eq.~\ref{eq:tau}, which is computed as: \begin{equation} \Delta\tau_{vv^\prime}(\mathbf{r})=\int V_c(\mathbf{r,r^{\prime}})\frac{\rho^+_{vv^\prime}(\mathbf{r^{\prime}})-\rho^-_{vv^\prime}(\mathbf{r^{\prime}})}{2}d\mathbf{r^{\prime}}. \label{eq:delta_tau_delta_rho} \end{equation} The densities $\rho_{vv^\prime}^{\pm}$ are obtained by solving the KS equations with the perturbed Hamiltonian $\hat{H}\pm\tau^u_{vv^\prime}$; both indexes $v$ and $v^\prime$ run over all occupied orbitals. While all potential terms of $\hat{H}$ may be computed self-consistently\cite{nguyen_finite-field_2019}, in this work the exchange-correlation potential was evaluated for the initial unperturbed electronic density and kept fixed during the self-consistent iterations. This amounts to evaluating the dielectric screening within the RPA. The FF-BSE approach has been implemented by coupling the WEST\cite{govoni_large_2015} and Qbox\cite{gygi_architecture_2008} codes in client-server mode.\cite{ma_finite-field_2019,nguyen_finite-field_2019,Govoni2021} The maximum number of integrals, $n_{\text{int}}=n_{\text{occ}}(n_{\text{occ}}+1)/2$, is determined by the total number of pairs of occupied orbitals. The actual number of integrals to be evaluated can be greatly reduced by using the recursive bisection method,\cite{gygi_compact_2009} which allows one to localize orbitals and consider only integrals generated by pairs of overlapping orbitals\cite{nguyen_finite-field_2019}. The systems studied in this work contain tens to hundreds of atoms, with hundreds to thousands of electrons. For example, for one of the Si/water interfaces discussed below, we considered a slab with 420 atoms, 1176 electrons and each single particle state is doubly occupied. Hence, $n_{\text{occ}}$=588, and $n_{\text{int}}=173166$. Using the recursive bisection method the total number of $vv^\prime$ pairs is reduced to $n_{\text{int}}=5574$ (a reduction factor slightly larger than 30) without compromising accuracy, when a bisection threshold of 0.05 and five bisection levels in each Cartesian direction are adopted\cite{gygi_compact_2009}. We note that the Liouville formalism used in this work (Eq.~\ref{eq:spectrum}) only involves summations over occupied states. Such formalism was shown to yield absorption spectra equivalent to solving the BSE with explicit and converged summations over empty states.\cite{rocca_ab_2010,rocca_solution_2012,nguyen_finite-field_2019} The same formalism may also be used to describe absorption spectra within TDDFT\cite{runge_density-functional_1984}, albeit employing a different definition of the $\mathcal{K}^{1e}$ and $\mathcal{K}^{1d}$ terms. \cite{hutter2003,walker_efficient_2006,rocca2008turbo,malciouglu2011turbotddft,ping_electronic_2013,ge2014turbotddft,nguyen_finite-field_2019} The key point of our work is the use of ML to generate a model for the calculation of screened Coulomb integrals (Eq.~\ref{eq:tau}) that is transferable to multiple atomic configurations; the goal is to reduce the computational cost in the solution of Eq.~\ref{eq:spectrum}. In particular, we consider the mapping between unscreened Coulomb integrals, $\tau^u_{vv\prime}$, and screened Coulomb integrals, $\Delta \tau_{vv\prime}$. Such transformation is mapping $n_\text{int}$ pairs of a 3D array, i.e., $\{ F: \tau^u_{vv^\prime}\to\Delta\tau_{vv^\prime},\,\forall v,v^\prime\in[1,\cdots,n_\text{occ}]\}$ and is similar to 3D image processing. Our objective is to learn the mapping functions and hence it is natural here to use convolutional neural networks (CNN), a widely used technique in image classification. CNNs are artificial neural networks with spatial-invariant features. The screened and unscreened Coulomb integrals are related by the dielectric matrix, which describes a linear response function of the system to an external perturbation. Therefore, the mapping we aim to obtain should follow a linear relationship for physical reasons, and one convolutional layer without nonlinear activation functions should be considered. Here, the surrogate model $F$, used to bypass the explicit calculation of Eq.~\ref{eq:delta_tau_delta_rho}, is represented by a single convolutional layer $K$: \begin{equation} \Delta\tau_{vv^\prime}(x,y,z) = (K * \tau^u_{vv^\prime} )(x,y,z) \label{eq:CNN} \end{equation} where $K$ is the convolutional filter of size $(n_x,n_y,n_z)$ (see the Electronic Supplementary Information (ESI) for details). The filter, $K$, is determined through an optimization procedure that utilizes $n_\text{int}$ pairs of $\tau^u_{vv^\prime}$ and $\Delta\tau_{vv^\prime}$ as the dataset, obtained for one configuration (i.e., one set of atomic positions) using Eq.~\ref{eq:tauu} and Eq.~\ref{eq:delta_tau_delta_rho}, respectively. Therefore this filter captures features in the dielectric screening that are translationally invariant. When the filter size is reduced to $(1,1,1)$, the training procedure is effectively a linear regression and Eq.~\ref{eq:CNN} amounts to applying a global scaling factor to $\tau^u_{vv^\prime}$, which we label $f^{\text{ML}}$. In our calculations, the mapping $F$ corresponds to evaluating the dielectric screening arising from the short-wavelength part (i.e., the body) of the dielectric matrix. The long-wavelength part (i.e., the head of the dielectric matrix) corresponds to the macroscopic dielectric constant $\epsilon_\infty$. The definitions of the head and body of the dielectric matrix are given in Eq.~\ref{eq:dtau_G} of the ESI. One of the main advantages of a ML-based model for the screening is that it may be reused for multiple configurations sampled during a FPMD simulation, thus avoiding the calculations of dielectric matrices for each snapshot, as illustrated in Figure \ref{fig:workflow}.The validity of such an approach and its robustness are discussed below for several systems. In our calculations, we carried out FPMD with the Qbox\cite{gygi_architecture_2008} code and MBPT theory calculations with the WEST\cite{govoni_large_2015} code, coupled in client server mode with Qbox in order to evaluate the screened integrals (Eq.s~\ref{eq:tau}-\ref{eq:delta_tau_delta_rho}), which constitute our training dataset. We implemented an interface between Tensorflow\cite{tensorflow2015-whitepaper} and WEST, including a periodic padding of the data for the convolution in Eq.~\ref{eq:CNN}, in order to satisfy periodic boundary conditions. The computational details of each system investigated here are reported in the ESI. \begin{figure}[htbp] \centering \includegraphics[width=\linewidth]{figs/Figure1.pdf} \caption{Illustration of the strategy to predict absorption spectra at finite temperature based on the solution of the Bethe-Salpeter equation (BSE) and machine learning techniques. $F$ is the mapping obtained by machine learning.} \label{fig:workflow} \end{figure} \section*{Results} We now turn to present our results for several systems, starting from liquid water. \subsection*{Liquids} To establish baseline results with small computational cost, we first considered a water supercell containing 16 water molecules. We tested the accuracy of a single convolutional layer with different filter sizes, from $(1,1,1)$ to $(20,20,20)$. We find that a convolutional model (Eq.~\ref{eq:CNN}) can be used to bypass the calculation of $\Delta\tau$ in Eq.~\ref{eq:delta_tau_delta_rho}, yielding absorption spectra in good agreement with the FF-BSE method. In particular, we find that a filter size of $(1,1,1)$, i.e., a global scaling factor, is sufficient to accurately yield the positions of the lower-energy peaks of the absorption spectra, with an error of only -0.03 eV (see the ESI for a detailed quantification of the error). We then turned to interpret the meaning of the global scaling factor $f^{\text{ML}}$, and we computed the quantity $\epsilon_f^{\text{ML}}=(1+f^{\text{ML}})^{-1}$. For 20 independent snapshots extracted from a FPMD trajectory of the 16-H$_2$O system, we find that $\epsilon_{f}^{\text{ML}}$ ($1.84\pm0.02$ is the same, within statistical error bars, as that of the PBE\cite{perdew_generalized_1996} macroscopic static dielectric constant computed using the polarizability tensor (as implemented in the Qbox code\cite{gygi_architecture_2008}): $\epsilon_\infty^{\text{PT}}=1.83\pm0.01$. Therefore, the global scaling factor that we learned is closely related to the long-wavelength dielectric constant of the system. Interestingly, we obtained similar scaling factors for a simulation using a larger cell, with 64-molecules, e.g., $\epsilon_f^{\text{ML}}=1.83$ for a given, selected snapshot, for which $\epsilon_\infty^{\text{PT}}=1.86$. To further interpret the factor $f^{\text{ML}}$ obtained by ML, we computed the average of $\Delta\tau_{vv^\prime}/\tau_{vv^\prime}^{u}$ over all $vv^\prime$. Specifically, we define $f^{\text{Avg}}=\frac{1}{\Omega}\int f^{\text{Avg}}(\mathbf{r}) d\mathbf{r}$, where $f^{\text{Avg}}(\mathbf{r})=\frac{1}{N_{vv^\prime}}\sum_{v,v^\prime} \Delta\tau_{vv^\prime}(\mathbf{r})/\tau_{vv^\prime}^{u}(\mathbf{r})$, $\Omega$ is the volume of the simulation cell, and $N_{vv^\prime}$ is the total number of ${vv^\prime}$ in the summation. Using one snapshot of the 16-H$_2$O system as an example, we find that $\epsilon_{f}^{\text{Avg}}=(1+f^{\text{Avg}})^{-1}=1.79$, similar to $\epsilon_{f}^{\text{ML}}=1.86$ for the same snapshot. To evaluate how sensitive the peak positions in the absorption spectra of water are to the value of the global scaling factor, we varied $\epsilon_f$ from 1.67 to 1.92. We find that the position of the lowest-energy peak varies approximately in a linear fashion, from 8.69 eV to 8.76 eV. This analysis shows that a global scaling factor is sufficient to represent the average effect of the body (i.e., short-wavelength part) of the dielectric matrix and that this factor is approximately equal to the head of the matrix (related to the long-wavelength dielectric constant). Hence, our results show that a diagonal dielectric matrix is a sufficiently good approximation to represent the screening of liquid water and to obtain its optical spectrum by solving the BSE. This simple finding is in fact an important result, leading to a substantial reduction in the computational time necessary to obtain the absorption spectrum of water at the BSE level of theory. In order to understand how the screening varies over a FPMD trajectory, we applied the global scaling factor $f^{\text{ML}}$ obtained from one snapshot of the 16-H$_2$O system to 10 different snapshots of a 64-H$_2$O system,\cite{dawson_equilibration_2018} at the same T, 400~K, and we computed an average spectrum. As shown in Figure \ref{fig:wat64_ave}, we can accurately reproduce the average spectrum computed with FF-BSE. The RMSE between the two spectra is 0.027. These results show that the global scaling factor is transferable from the 16 to the 64 water cell and that the dependence of the global scaling factor on the atomic positions may be neglected, for the thermodynamic conditions considered here. While it was recognized that the dielectric constant of water is weakly dependent on the cell size, it was not known that the average effect of the body of the dielectric matrix is also weakly dependent on the cell size. In addition, our results show that the dielectric screening can be considered independent from atomic positions for water at ambient conditions. This property of the dielectric screening was not previously recognized; it is not only an important recognition from a physical standpoint, but also from an efficiency standpoint, to improve the efficiency of BSE calculations. \begin{figure}[h] \centering \includegraphics[width=\linewidth]{figs/Figure2.pdf} \caption{Averaged spectra of liquid water obtained by solving the Bethe-Salpeter equation (BSE) in finite field (FF) and using machine learning techniques (ML). Results have been averaged over 10 snapshots obtained from first principles simulations at 400K, using supercells with 64 water molecules. The variability of the FF-BSE spectra within the 10 snapshots is shown in the inset. See also Figure~\ref{fig:wat64_ave_ind} of the ESI for the same variability when using ML-BSE.} \label{fig:wat64_ave} \end{figure} The timing acceleration of ML-BSE compared to FF-BSE is a function of the size of the system (characterized by the number of screened integrals $n_{\text{int}}$ and the number of plane waves (PWs) $n_{\text{pw}}$). We denote by $t_d$ the total number of core hours required to compute the net screening $\Delta\tau$ for all pairs of orbitals. We do not include in $t_d$ the training time, which usually takes only several minutes on one GPU for the systems studied here. Since we perform the training procedure once, we consider the training time to be negligible. We define the acceleration to compute the net effect of the screening as $\alpha_d=t^{\text{FF-BSE}}_d/t^{\text{ML-BSE}}_d$, and we find that $\alpha_d$ increases as $n_{\text{int}}$ and $n_{\text{pw}}$ increase. See the ESI for details. For the 64-H$_2$O system discussed above, we used a bisection threshold equal to 0.05, and a bisection level of 2 for each of the Cartesian direction. This reduces $n_{\text{int}}$ from $256(256+1)/2=32896$ to 3303. In this case, the gain achieved with our machine learning technique is close to two orders of magnitude: $\alpha_d=87$. \subsection*{Solids} We now turn to discussing the accuracy of ML-BSE for several solids, including LiF, MgO, Si, SiC, and C (diamond), for which we found again remarkable efficiency gains, ranging from 13 to 43 times for supercells with 64 atoms. In all cases, we used the experimental lattice constants.\cite{haas2009calculation} Similar to water, we found that a convolutional model (Eq.~\ref{eq:CNN}) can reproduce the absorption spectra of solids at the FF-BSE level, and that global scaling factors, either from linear regression or from averaging $\Delta\tau/\tau^u$ yield similar accuracy (Figures~\ref{fig:si_sf_cnn},\ref{fig:lif_sf_cnn} of the ESI). As shown in Figure~\ref{fig:solids_epsilon}, where we have defined $f^{\text{PT}}=(\epsilon_\infty^{\text{PT}})^{-1}-1$, we found that $f^{\text{ML}}$ is again numerically close to $f^{\text{PT}}$, for $\epsilon_\infty^{\text{PT}}$ computed using the polarizability tensor,\cite{gygi_architecture_2008} and the same level of theory and $k$-point sampling. These results show that, for ordered solids, the average effect of the body (short-wavelength part) of the dielectric matrix, $\epsilon_f^{\text{ML}}$, is similar to that of the head (long-wavelength limit) of the matrix and hence a diagonal screening is sufficient to describe the absorption spectra, similar to the case of water. This is an interesting result that supports the validity of the approximation chosen to derive the DDH functional.\cite{shimazaki2009first,marques2011density,refaely2013gap,skone2014self,skone2016nonempirical,brawand2016generalization,brawand2017performance,pham2017electronic,gerosa2017accuracy} \begin{figure}[h] \centering \includegraphics[width=\linewidth]{figs/Figure3.pdf} \caption{Relationship between the scaling factor obtained by machine learning ($f^{\text{ML}}$) and that obtained by computing the dielectric constant at the same level of theory ($f^{\text{PT}}$) (see text).} \label{fig:solids_epsilon} \end{figure} We note that the FF-BSE algorithm uses the $\Gamma$ point and is efficient and appropriate for large systems. In order to verify that a diagonal dielectric matrix is an accurate approximation also when using unit cells and fine grids of $k$-points, we computed the absorption spectrum of Si with a 2-atom cell and a $12\times12\times12$ $k$-point grid, using the Yambo\cite{marini2009yambo,sangalli2019many} code. We then compared the results with those obtained using a diagonal approximation of the dielectric matrix, and elements derived from the long-wavelength dielectric constant computed with the same cell and $k$-point grid. Fig.~\ref{fig:si_12_yambo} shows that we found an excellent agreement between the two calculations, of the same quality as that obtained for water in the previous section. \begin{figure}[h] \centering \includegraphics[width=\linewidth]{figs/Figure4.pdf} \caption{Absorption spectrum of crystalline Si computed by solving the Bethe-Salpeter equation (BSE) starting from PBE\cite{perdew_generalized_1996} wavefunctions, using a 2-atom cell and $12\times12\times12$ $k$-point sampling (blue line). The orange dashed line (Model-BSE) shows the same spectrum computed using a diagonal dielectric matrix with diagonal elements equal to $\epsilon_\infty=12.21$ (see text). Experimental results\cite{aspnes1983dielectric} are shown by the green dotted line.} \label{fig:si_12_yambo} \end{figure} It is important to note that the method presented here to learn the filter between unscreened and screened integrals represents a way of obtaining a model dielectric function with ML techniques, and without the need of using ad hoc empirical parameters. Several model dielectric functions have been proposed to speed-up the solution of the BSE for solids over the years.\cite{penn1962wave,levine_new_1982,hybertsen_model_1988,baroni_ab_1986,cappellini_model_1993,djurisic_dielectric_2001,bokdam_role_2016,sun_low-cost_2020} Recently, Sun et al.\cite{sun_low-cost_2020} proposed a simplified BSE method that utilizes a model dielectric function (m-BSE). The authors used the model of Cappellini et al.\cite{cappellini_model_1993} with an empirical parameter, which they determined by averaging the values minimizing the RMSE between a model dielectric function and that obtained within the RPA for Si, Ge, GaAs, and ZnSe.\cite{walter1970wave} This simplified BSE method yields good agreement with the results of the full BSE solution. For example, in the case of LiF, the shift between the first peak obtained with m-BSE and BSE is 0.12 eV, to be compared to the shift of 0.04 eV found here, between ML-BSE and FF-BSE. A model dielectric function has been proposed also for 2D semiconductors\cite{trolle2017model} and silicon nanoparticles \cite{wang1994dielectric,tsu1997simple}. However, the important difference between our work and the models just described is that the latter requires empirical parameterization. One of the advantages of the ML approach adopted here is that it does not require the definition of empirical parameters and, importantly, it may also be applied to nanostructures and heterogeneous systems, such as solid/liquid interfaces, as discussed next. \subsection*{Interfaces} We have shown that for solids and liquids, the use of ML leads to the definition of a global scaling factor that, when utilized to model the screened Coulomb interaction, yields results for absorption spectra in very good agreement with those of the full FF-BSE calculations, at a much lower computational cost. We now discuss solid/liquid interfaces as prototypical heterogeneous systems. We considered two silicon/water interfaces modeled by periodically repeated slabs. One is the H-Si/water interface, a hydrophobic interface with 420 atoms (72 Si atoms and 108 water molecules; Si surface capped by 24 H atoms); the other is a COOH-Si/water interface, a hydrophilic interface with 492 atoms (72 Si atoms and 108 water molecules; Si surface capped by 24 -COOH groups).\cite{pham_interfacial_2014} Not unexpectedly, we found that neither a global scaling factor nor a convolutional model is sufficiently accurate to reproduce the spectra obtained with FF-BSE, as shown in Figure~\ref{fig:sihwat_sf_cnn7} of the ESI. Therefore, we have developed a position-dependent ML model to describe the variation of the dielectric properties in the Si, water and interfacial regions. We divided the grid of $\tau_{vv^\prime}$ into slices, each spanning one $xy$ plane parallel to the interface; we then trained for a model on each slice. In this way we describe translationally invariant features along the $x$ and $y$ directions, and we obtain a $z$-dependent convolutional filter $K(z)$ or $z$-dependent scaling factors $f^{\text{ML}}(z)$. We found that a position-dependent filter, $K(z)$, or a scaling factor for each slice, $f^{\text{ML}}(z)$, yield a comparable accuracy, and therefore we focus on the $f^{\text{ML}}(z)$ model, which is simpler. \begin{figure*} \centering \includegraphics[width=17.1cm]{figs/Figure5.pdf} \caption{Comparison of absorption spectra obtained by solving the Bethe-Salpeter equation (BSE) in finite field (FF) and using machine learning (ML) techniques for (a) a H-Si/water interface shown in the lower left panel and (b) a COOH-Si/water interface (shown in the lower right panel). Blue, red and white spheres represent Si, oxygen and hydrogen respectively. C is represented by brown spheres. (See the ESI for results from using a kinetic energy cutoff of 60 Ry for wavefunctions.)} \label{fig:siwat_2interfaces_2par} \end{figure*} We found that the $z$-dependent ML model $f^{\text{ML}}(z)$ is accurate to represent the screening of the Si/water interfaces when computing absorption spectra (Figure~\ref{fig:siwat_2interfaces_2par}). Together with Figure~\ref{fig:sihwat_sf_cnn7} in the ESI, our finding show that a block diagonal dielectric matrix, where all the diagonal elements in the dielectric matrix have the same value, is not a good representation of the screening, unlike the case of water and ordered, periodic solids; instead taking into account the body of the dielectric matrix as in the $f^{\text{ML}}(z)$ model is critical in the case of an interface. Depending on how the grid of $\tau_{vv^\prime}$ are divided, we obtain different $f^{\text{ML}}(z)$ profiles for Si/water interfaces. Figure~\ref{fig:siwat_2interfaces_2par} shows the spectra in the case of $f^{\text{ML}}(z)$ defined by two parameters (a constant value in the Si region, and a different constant value in the water region); we name this profile $f^{\text{ML}}_{p2}(z)$. In Figure~\ref{fig:sihwat_sf_slice}(a) of the ESI, we present the spectra obtained using $f^{\text{ML}}(z)$ in the case of 108 slices evenly spaced in the $z$ direction, which we call $f^{\text{ML}}_{p108}(z)$. The function $\epsilon^{\text{ML}}_f(z)$ corresponding to $f^{\text{ML}}_{p108}(z)$ presents maxima at the interfaces, and minima at the points furthest away from the interface, in the Si and the water regions (Figure~\ref{fig:sihwat_sf_slice}(b) of the ESI). In order to interpret our findings, we express $\Delta\tau$ in terms of projective dielectric eigenpotentials, (PDEP)\cite{wilson2008efficient,wilson2009iterative} and we decompose $f^{\text{Avg}}(\mathbf{r})$ into contributions from each individual PDEP,\cite{zheng2019dielectric} i.e., $f^{\text{Avg}}=\sum_i f^{\text{Avg}}_i$, where \begin{equation} f^{\text{Avg}}_{i}(\mathbf{r})=\frac{1}{N_{v,v^\prime}}\sum_{v,v^\prime}\frac{\phi_{i}(\mathbf{r})(\lambda_{i}/(1-\lambda_{i}))\int\phi_{i}^{*}(\mathbf{r^{\prime\prime}})\tau^u_{vv^\prime}(\mathbf{r^{\prime\prime}})d\mathbf{r^{\prime\prime}}}{\tau^u_{vv^\prime}(\mathbf{r})} \label{eq:f_pdep_i} \end{equation} and $\phi_i$ is the $i$-th eigenpotential of the static dielectric matrix corresponding to the eigenvalue $\lambda_i$. We find that the largest contribution to $f^{\text{Avg}}(\mathbf{r})$ comes from the eigenvectors corresponding to the most negative PDEP eigenvalue. This PDEP component has its maximum near the interfaces, with the square modulus of the corresponding PDEP eigenpotential being localized at the interfaces (Figure~\ref{fig:sihwat_charge_f} of the ESI). This shows that the maximum of $\epsilon^{\text{ML}}_{f}(z)$ at the interfaces stem from the contribution of the PDEP eigenpotential with the most negative eigenvalue. Interestingly, $f^{\text{ML}}_{p2}(z)$ and $f^{\text{ML}}_{p108}(z)$ yield absorption spectra of similar quality. This suggests that the absorption spectrum is not sensitive to the details of the profile at the interface, at least in the case of the H-Si/water interface (Figure~\ref{fig:siwat_2interfaces_2par}(a) and Figure~\ref{fig:sihwat_sf_slice} of the ESI) and the COOH-Si/water interface (Figure~\ref{fig:siwat_2interfaces_2par}(b) and Figure~\ref{fig:sicoohwat_sf_slice} of the ESI) studied here. However, knowing the functional form of $f^{\text{ML}}_{p108}(z)$ is useful to determine the location of the interfaces, and it can be used to define where the discontinuities in $f^{\text{ML}}_{p2}(z)$ are located. We further developed a 3D grid model, $f^{\text{ML}}(\mathbf{r})$. This is a simple extension of the $z$-dependent model, where instead of slicing $\tau_{vv^\prime}$ in only one direction, we equally divided $\tau_{vv^\prime}$ into sub-domains in all three Cartesian directions. We tested cubic sub-domains of side lengths from 0.6 \AA~ to 2.6 \AA, and we found that the accuracy of the resulting spectrum is similar to that obtained with the $z$-dependent model, as shown in Figure~\ref{fig:sihwat_3D} of the ESI. In order to verify the transferability of the position-dependent model derived for one snapshot extracted from FPMD to other snapshots, we computed absorption spectra by using the same $f^{\text{ML}}(z)$ for different snapshots generated at ambient conditions and we found that the screening is weakly dependent on the atomic positions, at these conditions, similar to the case of water discussed above (Figure~\ref{fig:sihwat_5_20} of the ESI). In summary, by obtaining $\epsilon^{\text{ML}}_f(z)$ from machine learning, we have provided a way to define a position-dependent dielectric function for heterogeneous systems. For the Si/water interfaces, the acceleration to compute the net screening effect is $\alpha_d=86$ for H-Si/water if bisection techniques are used ($n_\text{int}=5574$), and $\alpha_d=224$ for COOH-Si/water, again if bisection techniques are used ($n_\text{int}=8919$). \subsection*{Nanoparticles} As our last example we consider nanoparticles, i.e., 0D systems. We focus on silicon clusters Si$_{35}$H$_{36}$ and Si$_{87}$H$_{76}$\cite{govoni2012carrier,govoni_large_2015,brawand2016generalization} but we start from a small cluster Si$_{10}$H$_{16}$ first, to test the methodology. As shown in Figure~\ref{fig:si10h16_40ang_ml}(b), we found that a global scaling factor is not an appropriate approximation of the screening, e.g., for the spectrum of Si$_{10}$H$_{16}$ computed using PW basis set in a simulation cell with a large vacuum (cell length over 25 \AA). This finding points at an important qualitative difference with respect to the case of solids and liquids (condensed systems). Interestingly, we found that convolutional models are instead robust to different sizes of vacuum, and give absorption spectra in good agreement with FF-BSE calculations (Figure~\ref{fig:si10h16_40ang_ml}(a)). The inaccuracy of a global scaling factor stems from two reasons. One is related to the fact that when the volume of the vacuum surrounding the cluster becomes large, the data of the training set is dominated by small matrix elements representing the vacuum region. Because the numerical noise is not translationally invariant, the use of Eq.~\ref{eq:CNN} overcomes this issue, as the noise from vacuum matrix elements is canceled out in the convolution process. We note that the presence of nonzero elements in the vacuum region is due to the choice of the PW basis set, which requires periodic boundary conditions. In the case of isolated clusters, the use of periodic boundary conditions could be avoided by choosing localized basis set. However, there are several systems of interest where using PW basis set is preferable and vacuum regions are present, such as nanoparticles deposited on surfaces. The second reason responsible for the inaccuracy of a global scaling factor, even if the noise arising from vacuum is eliminated, (see Figure~\ref{fig:si10h16_40ang_ml_rhocut100000} of the ESI) is that the mapping between $\tau^u$ and $\Delta\tau$ being is simply more complex in nanoparticles than in homogeneous systems. Such a complexity can be accounted for when using Eq.~\ref{eq:CNN}. \begin{figure}[htbp] \centering \includegraphics[width=\linewidth]{figs/Figure6.pdf} \caption{Comparison of absorption spectra of Si$_{10}$H$_{16}$ (40 \AA~cell) obtained by solving the Bethe-Salpeter equation (BSE) in finite field (FF) and using machine learning (ML) techniques for (a) convolutional layer with filter size $(7,7,7)$ from a cell of 30 \AA, and (b) a global scaling factor. The RMSE value between the FF-BSE and ML-BSE spectra is 0.067 for (a) and 0.141 for (b), respectively. The accuracy of using a convolutional layer with filter size $(7,7,7)$ from the 40 \AA~cell itself is similar to that of (a): RMSE=0.067.} \label{fig:si10h16_40ang_ml} \end{figure} In order to investigate the dependence of the screening of nanoparticles on temperature, we transferedthe ML model trained for one specific snapshot of the Si$_{35}$H$_{36}$ cluster, to different snapshots extracted from a FPMD simulation, in order to predict absorption spectra at finite temperature. We applied the convolutional model with filter size $(7,7,7)$ obtained from the 0~K Si$_{35}$H$_{36}$ cluster to 10 snapshots of Si$_{35}$H$_{36}$ from an FPMD trajectory equilibrated at 500~K. As shown in Figure~\ref{fig:si35h36_ave}, the average ML-BSE spectrum can accurately reproduce the FF-BSE absorption spectrum at 500~K, with a small peak position shift of 0.08 eV. The ML-BSE spectra of individual snapshots is also in good agreement with the corresponding spectra computed with FF-BSE, shown in Figure~\ref{fig:si35h36_10} of the ESI. These results show that for nanoclusters, as for water, the screening is weakly dependent on atomic positions over a 500~K FPMD trajectory; note however that the 0~K spectrum (Figure~\ref{fig:si35h36_25ang_0K_ml} of the ESI) has different spectral features than the one collected at 500~K (Figure~\ref{fig:si35h36_ave}). \begin{figure}[htbp] \centering \includegraphics[width=\linewidth]{figs/Figure7.pdf} \caption{Average spectra of Si$_{35}$H$_{36}$ obtained by solving the Bethe-Salpeter equation (BSE) in finite field (FF) and using machine learning techniques (ML). Results have been averaged over 10 snapshots obtained from first principles simulations at 500~K. The variability of the FF-BSE spectra within the 10 snapshots is shown in the inset. See also Figure~\ref{fig:si35h36_ave_ind} of the ESI for the same variability when using ML-BSE.} \label{fig:si35h36_ave} \end{figure} We also found that the convolutional model trained for Si$_{35}$H$_{36}$ can be applied to Si$_{87}$H$_{76}$ with an error within 0.07 eV for peak positions (Figure~\ref{fig:si87h76_si35model}). The accuracy is comparable to the convolutional model from Si$_{87}$H$_{76}$ itself, as shown in Figure~\ref{fig:si87h76_ml} of the ESI. This shows that the convolutional model captures the nonlocality of the dielectric screening common to Si clusters of different sizes and is transferable from a smaller to a larger nanocluster (Si$_{87}$H$_{76}$) within the size range considered here. The FF-BSE calculation of Si$_{87}$H$_{76}$ is about 6 times more expensive in terms of core hours than that of Si$_{35}$H$_{36}$; hence, being able to circumvent the FF-BSE calculation of Si$_{87}$H$_{76}$ by using the model $K$ computed for Si$_{35}$H$_{36}$ is certainly an advantage. \begin{figure}[htbp] \centering \includegraphics[width=\linewidth]{figs/Figure8.pdf} \caption{Accuracy of the Si$_{87}$H$_{76}$ spectrum obtained from ML-BSE by applying a convolutional model with filter size $(7,7,7)$, trained from Si$_{35}$H$_{36}$. The RMSE value between the FF-BSE and ML-BSE spectra is 0.033.} \label{fig:si87h76_si35model} \end{figure} Conceptually, the convolutional model yields filters that capture the translational invariant features of the dataset, and in our case they capture the nonlocality of the screening. In other words, the convolutional filters represent features in the mapping from $\tau^u_{vv^\prime}$ to $\Delta\tau_{vv^\prime}$ that are invariant across the simulation cell. For Si clusters, we found that the RMSE values between ML-BSE and FF-BSE spectra converges as the size of the filter increases. For example, for Si$_{35}$H$_{36}$, convergence is achieved at the filter size $(7,7,7)$, which corresponds to a cube with side length (2.24 \AA), corresponding approximately to the Si-Si bond length in the cluster (2.35 \AA). This result suggests that the screening of the Si cluster has features of the length of a nearest-neighbor bond that are translationally invariant. The timing acceleration $\alpha_d$ for calculations of the absorption spectra of the Si$_{35}$H$_{36}$ cluster in a cubic cell of 20, 25, or 30 \AA~ in length, is 24, 47, or 90 times, respectively, when using bisection techniques (threshold 0.03, 4 levels in each Cartesian direction), as shown in Figure~\ref{fig:si10_si35_timing} of the ESI. In the case of Si$_{87}$H$_{76}$ cluster, $\alpha_d \simeq 160$. \section*{Conclusions} We presented a method based on machine learning (ML) to determine a key quantity entering many body perturbation theory calculations, the dielectric screening; this quantity determines the strength of the electron-hole interaction entering the BSE. In our ML model, the screening is viewed as a convolutional (linear) filter that transforms the unscreened into the screened Coulomb interaction. Our results show that such a model can be obtained for a chosen atomic configuration and then re-used to represent the screening of multiple configurations sampled in a FPMD at finite temperature for several systems, including water, solid/water interfaces, and silicon clusters. In particular, we found that in the case of homogeneous systems, e.g. liquid water and several insulating and semiconducting solids, absorption spectra can be accurately predicted by using a diagonal dielectric matrix. When using such a diagonal form, we found excellent agreement with spectra computed by the full solution of the BSE in finite field. In addition, our results showed that for liquid water the same diagonal approximation can be used to accurately compute spectra for different configurations from FPMD at ambient conditions, thus easily obtaining a thermal average representing a finite temperature spectrum. In the case of nanostructures and heterogeneous systems, such as solid/liquid interfaces, we found that the use of diagonal matrices or block-diagonal dielectric matrices to describe the two portions of the system (Si and water, in the example chosen here) does not yield accurate spectra; through machine learning of the screening we could define simple models yielding accurate absorption spectra and a simple way of computing thermal averages. For nanostructures, it is necessary to use a convolutional model to properly represent the nonlocality of the dielectric screening. Similar to water and the Si/water interfaces, we found that the function describing the screening for hydrogenated Si-clusters of about 1~nm does not depend in any substantial way on the atomic coordinates of the snapshots sampled during our FPMD simulations, up to the maximum temperature tested here, 500~K. The time savings in the calculations of the screening using ML are remarkable, ranging from a factor of 13 to 87 for the solids and liquids studied here, with cells varying from 64 to 192 atoms. For the clusters and the interface, we obtained time savings ranging from 30 to 224 times, with cells varying from 26 to 492 atoms. Finally, we note that the ML-based procedure presented here, in addition to substantially speeding up the calculation of spectra, especially at finite T, represents a general approach to derive model dielectric functions, which are key quantities in electronic structure calculations, utilized not only in the solution of the BSE. For example, our approach provides a strategy to develop dielectric-dependent hybrid functionals (DDH)\cite{skone2014self,skone2016nonempirical} for TDDFT calculations, as well as an interpretation of the parameters entering model dielectric functions.\cite{penn1962wave,levine_new_1982,hybertsen_model_1988,cappellini_model_1993,tsu1997simple,wang1994dielectric,bokdam_role_2016,sun_low-cost_2020} In particular, for homogeneous systems, our findings points at TDDFT with DDH functionals as an accurate method to obtain absorption spectra, consistent with the results of Sun et al.\cite{sun_low-cost_2020}, which were however derived semi-empirically. Work is in progress to further develop a strategy to develop parameters entering hybrid DFT functionals using machine learning.\cite{dick2020machine} \section*{Conflicts of interest} There are no conflicts to declare. \section*{Acknowledgements} The authors thank Bethany Lusch, He Ma, Misha Salim, and Huihuo Zheng for helpful discussions. The work was supported by Advanced Materials for Energy-Water Systems (AMEWS) Center, an Energy Frontier Research Center funded by the U.S. Department of Energy, Office of Science, Basic Energy Sciences (DOE-BES), and Midwest Integrated Center for Computational Materials (MICCoM) as part of the Computational Materials Science Program funded by DOE-BES. This research used resources of the Argonne Leadership Computing Facility, which is a DOE Office of Science User Facility supported under Contract DE-AC02-06CH11357, and resources of the University of Chicago Research Computing Center (RCC). The GM4 cluster at RCC is supported by the National Science Foundation’s Division of Materials Research under the Major Research Instrumentation (MRI) program award no. 1828629. \section{Computational details} In the following sections we report the computational details used in this work. \subsection{Systems} We considered the systems reported in Table~\ref{tab:systems}. The electronic structure of each system was computed at the density functional theory (DFT) level of theory using plane wave basis sets and the ONCV pseudopotentials,\cite{schlipf_optimization_2015} with the Perdew–Burke–Ernzerhof (PBE)\cite{perdew_generalized_1996} exchange and correlation functional. Quantum Espresso\cite{giannozzi_quantum_2009} (version 6.1.0) and the Qbox\cite{gygi_architecture_2008} (version 1.66.2) codes were used. For each system the macroscopic dielectric constant, $\epsilon_\infty$, was calculated by averaging the diagonal elements of the polarizability tensor computed using the Qbox code, and was labeled $\epsilon_\infty^{\text{PT}}$. The electronic dipole used to compute the polarizability tensor is defined from using the center of charge of maximally localized Wannier functions (MLWF) with the refinement correction by Stengel and Spaldin.\cite{stengel2006accurate} \begin{table} \centering \caption{\label{tab:systems} Systems considered in this work.} \begin{tabular}{|l|l|l|l|} \hline System & Number of atoms & Size of the cell (\AA) & $\epsilon_{\infty}^{\text{PT}}$\tabularnewline \hline \hline 16-H$_{2}$O & 48 & $a=7.82$ & 1.83$\pm$0.01\tabularnewline \hline 64-H$_{2}$O & 192 & $a=12.41$ & 1.87$\pm$0.004\tabularnewline \hline Si & 64 & $a=5.43$ & 10.01\tabularnewline \hline SiC & 64 & $a=4.36$ & 6.12\tabularnewline \hline C & 64 & $a=3.57$ & 5.29\tabularnewline \hline MgO & 64 & $a=4.21$ & 3.16\tabularnewline \hline LiF & 64 & $a=4.03$ & 2.04\tabularnewline \hline Si$_{10}$H$_{16}$ & 26 & $a=20.00$ to $50.00$ & -\tabularnewline \hline Si$_{35}$H$_{36}$ & 71 & $a=25.00$ & -\tabularnewline \hline Si$_{87}$H$_{76}$ & 163 & $a=25.00$ & -\tabularnewline \hline H-Si/water & 420 & $a=11.62,b=13.42,c=33.43$ & 3.34\tabularnewline \hline COOH-Si/water & 492 & $a=11.62,b=13.42,c=35.73$ & 3.84\tabularnewline \hline \end{tabular} \end{table} \subsection{First principles molecular dynamics simulations} First principles molecular dynamics (FPMD) simulations were carried out using the Qbox\cite{gygi_architecture_2008} code (version 1.66.2). For liquid water, we considered unit cells with 16 or 64 water molecules. For 64-water samples, we considered snapshots from each of 10 independent FPMD trajectories (samples s0022-s0031) of the PBE400 dataset.\cite{dawson_equilibration_2018} For 16-water samples, we generated 20 independent MD trajectories starting from 20 independent snapshots of 16 water molecules initiated randomly with the same atomic density as the PBE400 dataset (1.11 g/cm$^3$ D$_2$O). The initiation method is the same as the method used to generate the PBE400 dataset.\cite{dawson_equilibration_2018} FPMD simulations were carried out using the Bussi-Donadio-Parrinello (BDP) thermostat\cite{bussi_canonical_2007} at 400 K with a thermostat time constant of 10000 a.u. and a time step of 10 a.u. ($\sim$0.24 fs). We modeled a hydrophobic Si/water interface with a slab containing 72 Si atoms, 24 H atoms terminating Si, and 108 water molecules. We modeled a hydrophilic Si/water interface with a slab containing 72 Si atoms, 24 -COOH groups terminating Si, and 108 water molecules. The geometrical configurations were taken from Pham et al.\cite{pham_interfacial_2014} The FPMD simulations of the Si$_{35}$H$_{36}$ cluster were carried out using the BDP thermostat at 500 K. The thermostat time constant was 5000 a.u., and the time step was 20 a.u. ($\sim$0.48 fs). Equilibration was reached within the first 20000 time steps ($\sim$9.7 ps). The finite temperature absorption spectrum was obtained by averaging the spectra obtained for ten snapshots extracted every 5000 time steps from the FPMD trajectory, and starting after 10.16 ps. \subsection{Calculations of absorption spectra} Absorption spectra calculations were carried out with the WEST\cite{govoni_large_2015}-Qbox\cite{gygi_architecture_2008} coupled codes using the FF-BSE scheme reported by Nguyen et al.\cite{nguyen_finite-field_2019}. For exited-state energies, a scissor operator was applied to the ground-state PBE energy levels to obtain a band gap corresponding to the value obtained at the G$_0$W$_0$@PBE level. Band gaps at the G$_0$W$_0$@PBE level were taken from the literature\cite{shishkin_self-consistent_2007,pham_probing_2014,pham_interfacial_2014,govoni_large_2015} except for 16-H$_2$O systems and the Si$_{10}$H$_{16}$ clusters, which we computed in this work. The values used are summarized in Table~\ref{tab:BSE_parameters}. For 16-H$_2$O snapshots, the G$_0$W$_0$@PBE band gap was computed using 640 PDEP eigenpotentials. For Si$_{10}$H$_{16}$, the G$_0$W$_0$@PBE band gap was computed using 1024 PDEP eigenpotentials and a cubic unit cell with a side length of 50 \AA. FF-BSE calculations were done at the $\Gamma$ point. Parameters used in FF-BSE simulations are in Table~\ref{tab:BSE_parameters}. The screening in FF-BSE\cite{nguyen_finite-field_2019} in the reciprocal space is expressed as follows: \begin{equation} \Delta\tau_{vv^\prime}(\mathbf{G})=\begin{cases} (\epsilon_{\infty}^{-1}-1)\tau_{vv^\prime}^{u}(\mathbf{G=0}), & \mathbf{G}=\mathbf{0}\\ \frac{4\pi e^{2}}{\mid\mathbf{G}\mid^{2}}\frac{\rho^+_{vv^\prime}(\mathbf{G})-\rho^-_{vv^\prime}(\mathbf{G})}{2}, & \mathbf{G}\neq\mathbf{0} \end{cases} \label{eq:dtau_G} \end{equation} where $\mathbf{G}$ is the reciprocal space lattice vector, and $\rho^\pm_{vv^\prime}(\mathbf{G})$ is the Fourier component of $\rho^\pm_{vv^\prime}(\mathbf{r})$. As described in the main text, we focus on getting a surrogate model corresponding to the $\mathbf{G}\neq\mathbf{0}$ terms; the term corresponding to $\mathbf{G=0}$, the long-wavelength limit, is added separately. \begin{table} \centering \caption{\label{tab:BSE_parameters} Parameters used to obtain the FF-BSE spectra.} \begin{tabular}{|l|P{0.15\linewidth}|P{0.18\linewidth}|P{0.1\linewidth}|p{0.1\linewidth}|p{0.08\linewidth}|} \hline System & G$_{0}$W$_{0}$@PBE band gap (eV) & Scissor operator (eV) & Bisection levels in each Cartesian direction & Bisection threshold & Kinetic energy cutoff (Ry)\tabularnewline \hline \hline 16-H$_{2}$O & 10.41 & - & 2 & 0.02 & 60\tabularnewline \hline 64-H$_{2}$O & 8.1$^{\text{Ref.}}$\cite{pham_probing_2014} & 3.99$\pm$0.16 & 2 & 0.05 & 60\tabularnewline \hline Si & 1.37 ($X_{1c}$-point)$^{\text{Ref.}}$\cite{govoni_large_2015} & 0.70 & 3 & 0.02 & 40\tabularnewline \hline SiC & 2.28 ($X_{1c}$-point)$^{\text{Ref.}}$\cite{govoni_large_2015} & 0.95 & 2 & 0.04 & 70\tabularnewline \hline C & 5.50$^{\text{Ref.}}$\cite{shishkin_self-consistent_2007} & 1.04 & 2 & 0.02 & 80\tabularnewline \hline MgO & 7.25$^{\text{Ref.}}$\cite{shishkin_self-consistent_2007} & 2.48 & 2 & 0.01 & 60\tabularnewline \hline LiF & 13.27$^{\text{Ref.}}$\cite{shishkin_self-consistent_2007} & 4.13 & 2 & 0.02 & 60\tabularnewline \hline Si$_{10}$H$_{16}$ & 8.50 & 3.80 & 4 & 0.03 & 20\tabularnewline \hline Si$_{35}$H$_{36}$ & 6.29$^{\text{Ref.}}$\cite{govoni_large_2015} & 2.80 (0~K); 3.58$\pm$0.17 (500~K) & 4 & 0.03 & 25\tabularnewline \hline Si$_{87}$H$_{76}$ & 4.77$^{\text{Ref.}}$\cite{govoni_large_2015} & 2.21 & 4 & 0.03 & 25\tabularnewline \hline H-Si/water & 1.34$^{\text{Ref.}}$\cite{pham_interfacial_2014} & 0.33 & 5 & 0.05 & 25$^a$\tabularnewline \hline COOH-Si/water & 1.34$^{\text{Ref.}}$\cite{pham_interfacial_2014} & 0.27 & 5 & 0.05 & 25$^a$\tabularnewline \hline \end{tabular} \small \raggedright $^a$Comments on this choice is discussed in Section~\ref{Siwat}. \end{table} For bulk Si, we used the Yambo\cite{marini2009yambo,sangalli2019many} code (version 4.4.0) to compute BSE spectra at the $\Gamma$ point and with $k$-point samplings. The Lanczos-Haydock solver was used, with scissor operators reported in Table~\ref{tab:si_yambo_par}. Details of the calculations with Yambo are reported in Table~\ref{tab:si_yambo_par}. For Model-BSE, the screened Coulomb potential $W=\epsilon^{-1}V_c$ was computed using $\epsilon$ defined as: \begin{equation} \epsilon^{-1}_{\mathbf{G},\mathbf{G^\prime}}=\begin{cases} \epsilon_{\infty}^{-1}, & \mathbf{G}=\mathbf{G^\prime}=\mathbf{0}\\ 1+f, & \mathbf{G}=\mathbf{G^\prime},\mathbf{G}\neq\mathbf{0} \\ 0, & \mathbf{G}\neq\mathbf{G^\prime} \end{cases} \label{eq:epsilon_yambo} \end{equation} where $\epsilon_\infty$ is the dielectric constant computed by Yambo, and $f$ is the scaling factor defined in the main text. $f$ is an input value and is specified in the caption of each Model-BSE spectrum reported in this article. \begin{table}[H] \centering \caption{\label{tab:si_yambo_par} Parameters of bulk Si for BSE calculations with the Yambo code.} \begin{tabular}{|l|p{0.08\linewidth}|p{0.13\linewidth}|p{0.13\linewidth}|p{0.13\linewidth}|p{0.08\linewidth}|p{0.08\linewidth}|} \hline System & Number of atoms & $k$-point & Size of dielectric matrix (in no. PWs) & Size of exchange term in the BSE kernel (in no. PWs) & Number of bands & Scissor operator (eV) \\ \hline \hline Si & 64 & $\Gamma$ & 1000 & 295667 & 320 & 0.70 \\ \hline Si & 2 & $12\times12\times12$ & 80 & 9185 & 20 & 0.85\\ \hline \end{tabular} \end{table} \subsection{Machine learning} We carried out ML-BSE calculations by implementing an interface between WEST\cite{govoni_large_2015} and Tensorflow\cite{tensorflow2015-whitepaper} (version 1.13). The convolutional model used in Eq.~\ref{eq:CNN} of the main text is defined as: \begin{equation} \Delta\tau_{vv^\prime}(x,y,z) = \sum_{i=0}^{n_x-1}\sum_{j=0}^{n_y-1}\sum_{k=0}^{n_z-1} K_{i,j,k} \tau^u_{vv^\prime}[x +(il-m_x) \Delta x, y+(jl-m_y)\Delta y, z+(kl-m_z)\Delta z] \label{eq:CNN_detail} \end{equation} where $K$ is the convolutional filter of size $(n_x,n_y,n_z)$, and $\Delta x$, $\Delta y$, $\Delta z$ are the spacings of the uniform 3D-mesh used to represent periodic functions in real-space. $m_i,i=x,y,z$ is the size of padding in each of the $x,y,z$ directions, and $m_i=\lfloor \frac{p_i}{2} \rfloor$ where $p_i=n_i+(n_i-1)(l-1)-1$, and $l$ is the dilation rate. We performed a hyperparameter search to determine the parameters in the training procedure. In the optimization procedure, we used the Adam optimizer\cite{kingma2014adam} with a learning rate of 0.001. The loss was evaluated by mean squared error (MSE). Early stopping was used to stop training based on validation loss after 25 epochs without an improvement. Periodic boundary conditions are satisfied using padded arrays; we considered the same size for input and output arrays in Eq.~\ref{eq:CNN}. In order to save memory and training time, we have trained parameters for convolutional models skipping every other element of the $\tau^u_{vv^\prime}$ and $\Delta\tau_{vv^\prime}$ arrays in each Cartesian direction (except for 16-H$_2$O samples, where the size of the system allows us to consider all elements). In this way the arrays entering the training procedure have an 8 fold smaller memory footprint when the coarse grid is used. When the trained models are applied, we reconcile the fact that training was done on coarse grids, by applying a dilation rate of 2 in Eq.~\ref{eq:CNN}. In this way, the convolutional filter can be applied, after training, to $\tau^u_{vv^\prime}$ arrays that are defined on the original FFT grid. Therefore, we used $l=2$ in Eq.~\ref{eq:CNN_detail} except for the 16-H$_2$O system, for which we chose $l=1$. For each system considered here, the training and validation data come from one snapshot, and we use data from snapshots different from the training snapshot as the test set. We have trained ML models with either a global scaling factor or a convolutional model with filter size $(n, n, n)$ with $n=3,5,7,9,12,15,20$. To study the effect of the training/validation split, we considered a given snapshot (which we call s00001) of the 16-H$_2$O system as an example, where there are 735 pairs of $\tau^u$ and $\Delta\tau$ arrays. Three different training/validation splits of the data set were considered: (1) all pairs used for training and validation, (2) 80\% pairs used for training and 20\% pairs used for validation, and (3) 60\% pairs used for training and 40\% pairs used for validation. Note that case~(1) does not give the same training and validation loss because minibatches of size 35 were used in training but all data were used in each validation. To evaluate the accuracy of the models, we trained for the models using snapshot s00001, and we used two other snapshots (i.e. s00003 and s00007) as test sets. None of these models have significant differences in the accuracy of reproducing the peak positions of the FF-BSE spectra. For all three split schemes, the average $\Delta\omega$ of the lowest-energy peak over different model architectures is -0.03 eV for s00001, -0.01 eV for s00003, and 0.00 eV for s00007. For the spectrum RMSE from ML-BSE using each split scheme, s00003 is 0.009 greater than s00001, and s00007 is 0.019 greater than s00001. These results suggest that the training/validation split within the range tested here does not significantly impact the accuracy or transferability of the ML models for predicting the absorption spectra. A subset of the dataset ($\tau^u_{vv^\prime}$ and $\Delta\tau_{vv^\prime}$ pairs) can be randomly selected and used as the training set in ML. The size of the subset does have an effect on the accuracy of the prediction. At least 10\% of data are needed in this subset to have a converged $f^{\text{ML}}$ value. When computing $f^{\text{Avg}}$, we note that numerical errors may cause the absolute value of some elements in $\Delta\tau_{vv^\prime}(\mathbf{r})/\tau_{vv^\prime}^{u}(\mathbf{r})$ to be extremely large, and these outliers need to be discarded before computing $f^{\text{Avg}}$. No outliers in data need to be eliminated to carry out ML to obtain $f^{\text{ML}}$. \subsection{Protocol to compute absorption spectra at finite temperature} Below we present the protocol to obtain absorption spectra at finite temperature: \begin{enumerate} \item Obtain representative snapshots of the target system at a finite temperature $T$ using FPMD. \item For one snapshot among those selected in Step 1, perform a FF-BSE calculation to obtain $\tau^u_{vv^\prime}$ and $\Delta\tau_{vv^\prime}$. Note that the selected snapshot may come, in some cases, from a smaller system of similar nature as the target system (see, e.g., the example of water presented in the main text). \item Use the $\tau^u_{vv^\prime}$ and $\Delta\tau_{vv^\prime}$ saved in Step 2 as the training/validation sets to machine learn the mapping between unscreened and screened Coulomb integrals. \item For all snapshots determined in Step 1, except for the one already evaluated in Step 2-3, compute the ML-BSE spectrum using the ML model trained in the previous step as a surrogate model for the calculation of screened Coulomb integrals. \item Obtain the optical absorption spectrum at the finite temperature $T$ by computing the average of the absorption spectra obtained in Step 4. \end{enumerate} \section{Comparison between FF-BSE and ML-BSE absorption spectra} \subsection{Liquid water} To quantify the accuracy of the ML-BSE spectrum, we compare it with the FF-BSE spectrum and compute the change of the energy of individual peaks ($\Delta\omega=\omega^{\text{ML-BSE}}-\omega^{\text{FF-BSE}}$, where $\omega$ is the position of the peak in energy) and the root mean square error (RMSE) of the whole spectrum in a given energy range (the range is 0.0-27.2 eV for all systems except for the interfaces, which is 0.0-13.6 eV). For a representative snapshot of the 16-H$_2$O cell, when using a simple scaling factor model or a convolutional model of filter size $(7,7,7)$, we find that for the lowest-energy peak, $\Delta\omega=-0.03$ eV in both cases, and RMSE is 0.021 or 0.018 when the scaling factor or the convolutional model is used, respectively (Figure~\ref{fig:wat16_cnn7_sf}).This shows that, for the 16-H$_2$O system, the difference introduced by using different ML models (a convolutional model versus a global scaling factor) is negligible. Figure~\ref{fig:wat16_sf_sensitivity} shows the sensitivity of the position of the first peak of the spectrum to the value of the global scaling factor. We used the 16-H$_2$O case as an example. Figure~\ref{fig:wat64_ave_exp_scissor}(a) shows a comparison of the experimentally measured absorption spectrum of liquid water with the one computed for the 64-H$_2$O system using either FF-BSE or ML-BSE. Figure~\ref{fig:wat64_ave_exp_scissor}(b) shows the sensitivity of the spectrum to different choices of the scissor operator. Figure~\ref{fig:wat64_ave_ind} shows the variations of individual snapshots used to calculate the averaged spectrum of liquid water (64-H$_2$O system) in Figure~\ref{fig:wat64_ave}. The same comparison for each individual snapshot is reported in Figure~\ref{fig:wat64_10}. The model used in ML-BSE is the global scaling factor from 16-H$2$O. \begin{figure}[H] \centering \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/wat16_ks7.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/wat16_ks1.pdf} \end{subfigure} \caption{Accuracy of ML-BSE spectra of liquid water (16-H$2$O) obtained using (a) a convolutional model with filter size $(7,7,7)$, (b) a global scaling factor model. RMSE of the spectra is 0.018 for (a) and 0.021 for (b).} \label{fig:wat16_cnn7_sf} \end{figure} \begin{figure}[H] \centering \includegraphics[width=0.6\linewidth]{figs/wat16_sf_sensitivity.pdf} \caption{Sensitivity of the position of the lowest-energy peak ($\omega_1$) of the computed absorption spectrum of water, obtained for a snapshot with 16 H$_2$O molecules), to the global scaling factor ($f$). $\epsilon_f=(1+f)^{-1}$.} \label{fig:wat16_sf_sensitivity} \end{figure} \begin{figure}[H] \centering \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/wat_pbe400_ave_exp.pdf} \label{fig:wat64_ave_exp} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/wat_pbe400_s0031_scissor.pdf} \label{fig:wat64_s0031_scissor} \end{subfigure} \caption{Absorption spectrum of liquid water (64-H$_2$O). (a) FF-BSE and ML-BSE are obtained computing and averaging the spectra of 10 snapshots. The position of the first peak of the experimental spectra from Heller et al.\cite{heller_collective_1974} and from Hayashi et al.\cite{hayashi_complete_2000} is located at 8.18 eV and 7.98 eV, respectively. The position of the first peak of the ML-BSE and FF-BSE spectra is located at 7.40 eV. (b) Sensitivity of the FF-BSE absorption spectrum of a chosen snapshot to the fundamental gap of the system. In (b), the values of the fundamental gap, obtained applying a scissor operator to the computed PBE band gap, were chosen based on the values of the experimental gap of water of 8.7$\pm$0.6 eV,\cite{bernas1997electronic} and the G$_0$W$_0$@PBE band gap of 8.1 eV.\cite{pham_probing_2014} For this particular snapshot, the energy of the first peak is 7.14 eV, 7.85 eV, and 8.47 eV, when the band gaps of the system is equal to 8.1 eV, 8.7 eV, and 9.3 eV, respectively.} \label{fig:wat64_ave_exp_scissor} \end{figure} \begin{figure}[H] \centering \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/wat_pbe400_fluc_ff.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/wat_pbe400_fluc_ml.pdf} \end{subfigure} \caption{The absorption spectra of 10 individual snapshots of 64-H$2$O systems (dotted lines) and their averaged spectrum (solid line) from (a) FF-BSE and (b) ML-BSE calculations. The model used in ML-BSE is a global scaling factor obtained from simulations of a 16-water-molecule cell.} \label{fig:wat64_ave_ind} \end{figure} \begin{figure}[H] \centering \begin{subfigure}{0.45\textwidth} \includegraphics[width=\linewidth]{figs/wat_pbe400_s0022_wat16ks1.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \includegraphics[width=\linewidth]{figs/wat_pbe400_s0023_wat16ks1.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \includegraphics[width=\linewidth]{figs/wat_pbe400_s0024_wat16ks1.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \includegraphics[width=\linewidth]{figs/wat_pbe400_s0025_wat16ks1.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \includegraphics[width=\linewidth]{figs/wat_pbe400_s0026_wat16ks1.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \includegraphics[width=\linewidth]{figs/wat_pbe400_s0027_wat16ks1.pdf} \end{subfigure} \caption{ML-BSE and FF-BSE spectra of 10 snapshots of the 64-H$_2$O system from FPMD trajectories at 400 K. The model used in ML-BSE is a global scaling factor obtained from simulations of a 16-water-molecule cell. The labels on the snapshots on top of each panel follows the labeling of snapshots of the PBE400 set (http://quantum-simulation.org/reference/h2o/pbe400/s32/index.htm).\cite{dawson_equilibration_2018}} \label{fig:wat64_10} \end{figure} \begin{figure}[H] \ContinuedFloat \centering \begin{subfigure}{0.45\textwidth} \includegraphics[width=\linewidth]{figs/wat_pbe400_s0028_wat16ks1.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \includegraphics[width=\linewidth]{figs/wat_pbe400_s0029_wat16ks1.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \includegraphics[width=\linewidth]{figs/wat_pbe400_s0030_wat16ks1.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \includegraphics[width=\linewidth]{figs/wat_pbe400_s0031_wat16ks1.pdf} \end{subfigure} \caption{(Continued from the previous page) ML-BSE and FF-BSE spectra of 10 snapshots of the 64-H$_2$O system. The model used in ML-BSE is a global scaling factor obtained from simulations of a 16-water-molecule cell. The labels on the snapshots on top of each panel follows the labeling of snapshots of the PBE400 set (http://quantum-simulation.org/reference/h2o/pbe400/s32/index.htm).\cite{dawson_equilibration_2018}} \label{fig:wat64_10} \end{figure} \subsection{Solids} Figures~\ref{fig:si_sf_cnn} and \ref{fig:lif_sf_cnn} show that a convolutional model has similar accuracy as a global scaling factor for Si and LiF. The comparisons between FF-BSE and ML-BSE spectrum for C (diamond), SiC, and MgO are reported in Figure~\ref{fig:solids_spec}. \begin{figure}[H] \centering \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/si64_ks1.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/si64_ks7.pdf} \end{subfigure} \caption{Accuracy of ML-BSE spectra of Si obtained using (a) a global scaling factor model and (b) a convolutional model (filter size $(7,7,7)$). Panel (a) also shows the BSE spectrum obtained by using a scaling factor computed from averaging $\Delta\tau/\tau^{u}$, labeled "Avg-BSE". The peak shifts of the ML-BSE spectrum from the FF-BSE spectrum are within 0.01 eV for all cases. The RMSE values between the ML-BSE and FF-BSE spectra are 0.019 for (a) and 0.015 for (b).} \label{fig:si_sf_cnn} \end{figure} \begin{figure}[H] \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/lif222_ks1.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/lif222_ks7.pdf} \end{subfigure} \caption{Accuracy of ML-BSE spectra of LiF obtained using (a) a global scaling factor model and (b) a convolutional model (filter size $(7,7,7)$). The RMSE values between the ML-BSE and FF-BSE spectra are 0.052 for (a) and 0.058 for (b).} \label{fig:lif_sf_cnn} \end{figure} \begin{figure}[H] \centering \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/c64_ks1.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/sic64_ks1.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/mgo64_ks1.pdf} \end{subfigure} \caption{Comparison between FF-BSE and ML-BSE spectrum of (a) diamond, (b) SiC, and (c)MgO. ML-BSE results are obtained from a global scaling factor of the respective system. The RMSE values between the ML-BSE and FF-BSE spectra are 0.005 for (a), 0.027 for (b), and 0.044 for (c).} \label{fig:solids_spec} \end{figure} Figure~\ref{fig:si64_ff_yambo_gamma} shows the comparison between the absorption spectrum of Si computed with Yambo and WEST at the $\Gamma$ point. \begin{figure}[H] \centering \includegraphics[width=0.5\linewidth]{figs/si64_yambo.pdf} \caption{FF-BSE (WEST), ML-BSE (WEST) with a global scaling factor $f^{\text{ML}}=-0.81$, BSE (Yambo), and Model-BSE (Yambo) with $f=-0.81$ for Si with a 64-atom supercell at the $\Gamma$ point. The head of the dielectric matrix used in this figure is $\epsilon_\infty=22.11$, and was obtained for the same cell at the $\Gamma$ point in Yambo.} \label{fig:si64_ff_yambo_gamma} \end{figure} \subsection{H-Si/water interface (hydrophobic interface)} \label{Siwat} \subsubsection*{Comments on the kinetic energy cutoff} For the H-Si/water interface, we tested two different kinetic energy cutoff values for wavefunctions: 25 Ry and 60 Ry. From Figure~\ref{fig:sihwat_ecut}(a), we conclude that 25 Ry is sufficient to obtain the low-energy peaks of the Si/water interface FF-BSE spectrum. Comparing Figure~\ref{fig:siwat_2interfaces_2par}(a) and Figure~\ref{fig:sihwat_ecut}(b), we conclude that using 25 Ry or 60 Ry will not change our conclusions for the Si/water interface. The $f^{\text{ML}}_{p2}(z)$ models for the 25 Ry case and the 60 Ry case are very similar, with $\epsilon^{\text{ML}}_f(z)$ being 2.28 and 2.95 for $z$ in the respective regions of Si and water for the 25 Ry case, and 2.25 and 2.99 for $z$ in the respective regions of Si and water for the 60 Ry case. Therefore, we used 25 Ry for all Si/water interfaces in this work. \begin{figure}[H] \centering \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/sihwat_20_ff_ecut60vs25.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/sihwat_20_ecut60Ry_fz2.pdf} \end{subfigure} \hfill \caption{(a) FF-BSE spectrum from using a kinetic energy cutoff of 25 Ry or 60 Ry for wavefunctions. (b) Accuracy of ML-BSE using a $f^{\text{ML}}_{p2}(z)$ model and a kinetic energy cutoff of 60 Ry for wavefunctions.} \label{fig:sihwat_ecut} \end{figure} \subsubsection*{A position-dependent model $f(\mathbf{r})$ is necessary} From Figure~\ref{fig:sihwat_sf_cnn7}, we can see that when using a model which treats Si and water on the same footing, using average scaling factors, results for the absorption spectrum are not accurate. \begin{figure}[H] \centering \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/sihwat_20_ks1.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/sihwat_20_ks7.pdf} \end{subfigure} \hfill \caption{Accuracy of ML-BSE using (a) a global scaling factor model, and (b) a convolutional model (filter size $(7,7,7)$) for a snapshot representing a H-Si/water interface (see text). The RMSE values between the ML-BSE and FF-BSE spectra are 0.078 for (a) and 0.072 for (b).} \label{fig:sihwat_sf_cnn7} \end{figure} Comparing Figure~\ref{fig:sihwat_sf_cnn7} and Figure~\ref{fig:sihwat_sf_slice}, we find that a position-dependent model is necessary to obtain sufficiently accurate absorption spectra for the H-Si/water interface. \begin{figure}[H] \centering \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/sihwat_20_fz108.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/sihwat_20_fz108_eps.pdf} \end{subfigure} \caption{Accuracy of ML-BSE using (a) a position-dependent model with 108 parameters ($f^{\text{ML}}_{p108}(z)$), and (b) $\epsilon^{\text{ML}}_f(z)$ corresponding to the $f^{\text{ML}}(z)$ profile used to compute (a), for a snapshot representing a H-Si/water interface (see text). The RMSE value between the ML-BSE and FF-BSE spectra is 0.059.} \label{fig:sihwat_sf_slice} \end{figure} \subsubsection*{Interpretation of the $f(\mathbf{r})$ profile} In Figure~\ref{fig:sihwat_charge_f} we show the charge density (a), $f^{\text{Avg}}(\mathbf{r})$ (defined in the main text) (b), PDEP eigenpotential corresponding to the most negative eigenvalue (c), and the component of $f^{\text{Avg}}(\mathbf{r})$ that corresponds to the most negative PDEP eigenvalue ($f_1^{\text{Avg}}(\mathbf{r})$, defined in Eq.~\ref{eq:f_pdep_i} of the main text) (d) for the H-Si/water interface. This shows that the maxima of $\epsilon^{\text{ML}}_{f}(z)$ (Figure~\ref{fig:sihwat_sf_slice}(b)) at the interfaces stem from the contribution of the PDEP eigenpotential with the most negative eigenvalue. \begin{figure}[H] \centering \centering \begin{subfigure}{0.225\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/sihwat_chargedens.pdf} \end{subfigure} \hfill \begin{subfigure}{0.24\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/sihwat_favg_window888.pdf} \end{subfigure} \hfill \begin{subfigure}{0.22\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/sihwat_pdep1.pdf} \end{subfigure} \hfill \begin{subfigure}{0.22\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/sihwat_f1avg_window888.pdf} \end{subfigure} \caption{(a) Charge density, isovalue 0.050; (b) $f^{\text{Avg}}(\mathbf{r})$ from averaging using a moving window of $8\times8\times8$, isovalue -0.644; (c) PDEP eigenpotential corresponding to the most negative eigenvalue, isovalue 3.64; (d) $f^{\text{Avg}}_1(\mathbf{r})$ from averaging using a moving window of $8\times8\times8$, isovalue -0.370. The isosurface at the isovalue (i.e. value that has the greatest absolute value) is in yellow. The 3D visualization was rendered in VESTA (version 3.4.0).\cite{momma2011vesta}} \label{fig:sihwat_charge_f} \end{figure} \subsubsection*{3D grid model from ML} In Figure~\ref{fig:sihwat_3D} we show that a 3D grid model $f^{\text{ML}}(\mathbf{r})$ yields an accurate ML-BSE spectrum for the H-Si/water interface. The 3D grid model has the same RMSE as the $z$-dependent model, and for the lowest-energy peak $\Delta\omega=-0.11$ eV, close to the $-0.08$ eV for the $z$-dependent model. The peak position from 3D grid model is slightly less accurate than the $z$-dependent model largely because the 3D grid model is more susceptible to the local variation in the training data. \begin{figure}[H] \centering \includegraphics[width=0.5\linewidth]{figs/sihwat_20_cg16.pdf} \caption{Accuracy of ML-BSE using the 3D grid model of the H-Si/water interface. Each sub-domain is a cube, and the side length of each sub-domain is 2.6 \AA. The RMSE value between the ML-BSE and FF-BSE spectra is 0.059.} \label{fig:sihwat_3D} \end{figure} \subsubsection*{Transferability of position-dependent models} To test the transferrability of the $f^{\text{ML}}(z)$ model across different snapshots, we consider a $f^{\text{ML}}(z)$ model, $f^{\text{ML}}_{p216}(z)$, more fine-grained than $f^{\text{ML}}_{p2}(z)$ and $f^{\text{ML}}_{p108}(z)$. The use of this more fine-grained model is to test whether overfitted $f^{\text{ML}}(z)$ models can still be transferrable across different snapshots. As shown in Figure~\ref{fig:sihwat_5_20}, the $z$-dependent model for the Si/water interface is transferable between different snapshots of the Si/water interface. \begin{figure}[H] \centering \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/sihwat_05_fz216.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/sihwat_20_fz216.pdf} \end{subfigure} \caption{Accuracy of ML-BSE using the $z$-dependent model $f^{\text{ML}}_{p216}(z)$ for the H-Si/water interface trained from a given snapshot (bottom of panel (b)) applied on a different snapshot (bottom of panel (a)). The spectrum corresponding to each snapshot is above its structural representation. The RMSE values is 0.064 for (a) and 0.059 for (b).} \label{fig:sihwat_5_20} \end{figure} \subsection{COOH-Si/water interface (hydrophilic interface)} For the hydrophilic interface COOH-Si/water, we have found that the performance of the $f^{\text{ML}}_{p108}(z)$ model is also similar to $f^{\text{ML}}_{p2}(z)$ (Figure~\ref{fig:siwat_2interfaces_2par}(b) of the main text), consistent with our results for the H-Si/water interface. \begin{figure}[H] \centering \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/sicoohwat_ddh_fz108.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/sicoohwat_108par_eps.pdf} \end{subfigure} \caption{Accuracy of ML-BSE using (a) a position-dependent model with 108 parameters ($f^{\text{ML}}_{p108}(z)$), and (b) $\epsilon^{\text{ML}}_f(z)$ corresponding to the $f^{\text{ML}}(z)$ profile used to compute (a), for a slab representing a COOH-Si terminated surface interfaced with water. The RMSE value between the ML-BSE and FF-BSE spectra is 0.078.} \label{fig:sicoohwat_sf_slice} \end{figure} \subsection{Si clusters} For Si clusters, we were able to obtain a linear regression model (that results in a global scaling factor $f^{\text{ML}}=-0.28$) and convolutional models in the same way as for homogeneous systems. However, the global scaling factor we obtained from linear regression does not yield accurate spectra, as shown in Figure~\ref{fig:si10h16_40ang_ml} of the main article. In addition, a global scaling factor derived from averaging $\Delta\tau/\tau^{u}$, $f^{\text{Avg}}=-0.54$, is different from from $f^{\text{ML}}$ and yields an even worse spectrum , as shown in Figure~\ref{fig:si10h16_30ang_ML_avg}. The value of the scaling factor may be different for simulation cells of different sizes , even if the absorption spectra have converged with respect to the size of the simulation cell. For a cubic simulation cell 30 (40) \AA~ in length, $f^{\text{ML}}=-0.28$ ($-0.22$). \begin{figure}[H] \centering \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/si10h16_ffks1avg.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/si10h16_ffks1avg_xlim4_9.pdf} \end{subfigure} \caption{Comparison of ML-BSE and Avg-BSE to FF-BSE for Si$_{10}$H$_{16}$ (30 \AA~ cell). ML-BSE spectra are obtained using $f^{\text{ML}}=-0.28$, and Avg-BSE is from using $f^{\text{Avg}}=-0.54$. In (b) we show the same data as in (a) on a smaller energy interval (4 to 9 eV).} \label{fig:si10h16_30ang_ML_avg} \end{figure} The reason behind this behavior is that, although the charge density of the Si cluster decays rapidly as a function of the distance from the cluster, the ratio $\Delta\tau/\tau^{u}$ does not. To have an ML model robust to the size of the simulation cell, we impose a threshold to the Si cluster data so that small elements (e.g. elements smaller than a chosen threshold) in $\tau^u_{vv^\prime}$ are set to 0. Then we use this data to obtain our ML models. The spectra obtained from various models are shown in Figures~\ref{fig:si10h16_40ang_ml_rhocut100000}. Comparing Figures~\ref{fig:si10h16_40ang_ml_rhocut100000} and Figure~\ref{fig:si10h16_40ang_ml} of the main article, it is clear that imposing a threshold on the data improves the accuracy of the spectra. However, a global scaling factor model still performs poorly compared to convolutional models. The convolutional models give highly accurate absorption spectra regardless of whether the data are treated with a threshold. This suggests that convolutional models are robust to sparse data. In the previous two paragraphs, we applied the threshold on the training data, and applied the resulting ML models without imposing a threshold. Another way to consider the effect of the vacuum is to use the original data in training, but use a threshold on $\tau^u_{vv^\prime}$ when the model is applied. Figure~\ref{fig:si10h16_40ang_cg_cnn_rhocut100000}(a) and Figure~\ref{fig:si10h16_40ang_ml}(a) show that, for convolutional models obtained using the original, un-thresholded data, imposing a threshold when applying the model can improve the accuracy of the spectrum. Figure~\ref{fig:si10h16_40ang_cg_cnn_rhocut100000}(b)(c) show that this will not improve much for the model based on scaling factors, even if a position-dependent model is used. This confirms that convolutional models are superior to models based on scaling factors. \begin{figure}[H] \centering \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/si10h16_ks1_rho1e5.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/si10h16_30angrho1e5ks7.pdf} \end{subfigure} \caption{Accuracy of ML-BSE spectra of Si$_{10}$H$_{16}$ (40 \AA~ cell) obtained using (a) a global scaling factor, and (b) a convolutional model (filter size $(7,7,7)$) from a smaller cell (30 \AA~ cell). The models are obtained by considering only regions of $\tau^u_{vv^\prime}$ above a charge density threshold ($10^{-5}$ times the largest charge density). The RMSE value between the FF-BSE and ML-BSE spectra are 0.119 for (a) and 0.061 for (b).} \label{fig:si10h16_40ang_ml_rhocut100000} \end{figure} \begin{figure}[H] \centering \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/si10h16_ks7_rho1e5.pdf} \end{subfigure} \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/si10h16_cg14_rho1e5.pdf} \end{subfigure} \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/si10h16_cg14_rho1e5vs0.pdf} \end{subfigure} \hfill \caption{Accuracy of ML-BSE spectra of Si$_{10}$H$_{16}$ (40 \AA~ cell) obtained from (a) a convolutional model (filter size $(7,7,7)$), and (b) a 3D grid model $f^{\text{ML}}(\mathbf{r})$. When the models are applied to obtain ML-BSE spectra in (a) and (b), only regions of $\tau^u_{vv^\prime}$ above a charge density threshold ($10^{-5}$ times the largest charge density) are considered. The RMSE value between the FF-BSE and ML-BSE spectra are 0.107 for (a) and 0.044 for (b). In (c) we compare ML-BSE spectra from the 3D grid model when the charge density threshold is applied (with cutoff) or not applied (no cutoff).} \label{fig:si10h16_40ang_cg_cnn_rhocut100000} \end{figure} \subsubsection*{Spectrum of Si$_{35}$H$_{36}$ at zero temperature} The 0~K spectrum of Si$_{35}$H$_{36}$ differs from the 500~K spectra, as shown in Figure~\ref{fig:si35h36_25ang_0K_ml} and Figure~\ref{fig:si35h36_ave}. \begin{figure}[H] \centering \includegraphics[width=0.5\linewidth]{figs/si35h36_ks7.pdf} \caption{Accuracy of ML-BSE spectra obtained from a convolutional model (filter size $(7,7,7)$) for the 0 K geometrical configuration of Si$_{35}$H$_{36}$. The RMSE value between the ML-BSE and FF-BSE spectra is 0.029.} \label{fig:si35h36_25ang_0K_ml} \end{figure} \subsubsection*{Finite temperature spectra of Si$_{35}$H$_{36}$} The variations of individual spectra used to calculate the averaged spectrum of Si$_{35}$H$_{36}$ are shown in Figure~\ref{fig:si35h36_ave_ind}. The comparison between FF-BSE and ML-BSE spectra of each snapshot is in Figure~\ref{fig:si35h36_10}. The model used in ML-BSE uses a convolutional model $(7,7,7)$ trained on data obtained for the 0~K snapshot, the same model used in Figure~\ref{fig:si35h36_25ang_0K_ml}. \begin{figure}[H] \centering \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/si35h36_500K_fluc_ff.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/si35h36_500K_fluc_ks7.pdf} \end{subfigure} \caption{The absorption spectra of 10 individual snapshots of Si$_{35}$H$_{36}$ (dotted lines) and their averaged spectrum (solid line) from (a) FF-BSE and (b) ML-BSE calculations.} \label{fig:si35h36_ave_ind} \end{figure} \begin{figure}[H] \centering \begin{subfigure}{0.45\textwidth} \includegraphics[width=\linewidth]{figs/si35h36_021000_0Kks7.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \includegraphics[width=\linewidth]{figs/si35h36_026000_0Kks7.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \includegraphics[width=\linewidth]{figs/si35h36_031000_0Kks7.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \includegraphics[width=\linewidth]{figs/si35h36_036000_0Kks7.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \includegraphics[width=\linewidth]{figs/si35h36_041000_0Kks7.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \includegraphics[width=\linewidth]{figs/si35h36_046000_0Kks7.pdf} \end{subfigure} \caption{ML-BSE and FF-BSE spectra of 10 snapshots of Si$_{35}$H$_{36}$ snapshots extracted from FPMD trajectories at 500~K. The model used in ML-BSE obtained using a convolutional model $(7,7,7)$ trained with data obtained for the 0~K snapshot, the same model used in Figure~\ref{fig:si35h36_25ang_0K_ml}. The 10 snapshots were extracted every 5000 time steps from the FPMD trajectory, starting from the 21000th time step (10.16 ps onwards). We label these snapshots from s021000 to s066000.} \label{fig:si35h36_10} \end{figure} \begin{figure}[H] \ContinuedFloat \centering \begin{subfigure}{0.45\textwidth} \includegraphics[width=\linewidth]{figs/si35h36_051000_0Kks7.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \includegraphics[width=\linewidth]{figs/si35h36_056000_0Kks7.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \includegraphics[width=\linewidth]{figs/si35h36_061000_0Kks7.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \includegraphics[width=\linewidth]{figs/si35h36_066000_0Kks7.pdf} \end{subfigure} \caption{(Continued from the previous page) ML-BSE and FF-BSE spectra of 10 snapshots of Si$_{35}$H$_{36}$ snapshots extracted from FPMD trajectories at 500~K. The model used in ML-BSE obtained using a convolutional model $(7,7,7)$ trained with data obtained for the 0~K snapshot, the same model used in Figure~\ref{fig:si35h36_25ang_0K_ml}. The 10 snapshots were extracted every 5000 time steps from the FPMD trajectory, starting from the 21000th time step (10.16 ps onwards). We label these snapshots from s021000 to s066000.} \label{fig:si35h36_10} \end{figure} \subsubsection*{Transferability of ML models for Si$_{35}$H$_{36}$} As shown in Figure~\ref{fig:si35h36_10} (snapshot s041000) and Figure~\ref{fig:si35h36_0K_500K}, the accuracy of the convolutional model derived from data for the 0~K and 500~K snapshots is similar in predicting the absorption spectrum of the 500 K snapshot. This suggests that the convolutional model is indeed transferable from the 0 K geometry to 500 K geometries for this Si cluster. \begin{figure}[H] \centering \includegraphics[width=0.45\linewidth]{figs/si35h36_041000_ks7.pdf} \caption{Accuracy of the ML-BSE spectrum of Si$_{35}$H$_{36}$ (a 500 K snapshot s041000) obtained by using a convolutional model $(7,7,7)$ derived from the data for snapshot 041000. The RMSE value between the ML-BSE and FF-BSE spectra in this figure is 0.019, to be compared with RMSE=0.020 for the same snapshot's spectrum in Figure~\ref{fig:si35h36_10}.} \label{fig:si35h36_0K_500K} \end{figure} \subsubsection*{Si$_{87}$H$_{76}$} The ML-BSE spectrum of Si$_{87}$H$_{76}$ obtained by applying various models obtained from Si$_{87}$H$_{76}$ is similar to the one obtained by applying the convolutional model obtained from Si$_{35}$H$_{36}$, as shown in Figure~\ref{fig:si87h76_ml} and Figure~\ref{fig:si87h76_si35model}. \begin{figure}[H] \centering \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/si87h76_ks1.pdf} \end{subfigure} \hfill \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/si87h76_ks5.pdf} \end{subfigure} \begin{subfigure}{0.45\textwidth} \caption{} \includegraphics[width=\linewidth]{figs/si87h76_ks7.pdf} \end{subfigure} \caption{Accuracy of the ML-BSE spectrum of Si$_{87}$H$_{76}$ obtained by using (a) a global scaling factor, (b) a convolutional model $(5,5,5)$, and (c) a convolutional model $(7,7,7)$. The RMSE of the spectra are 0.025 for (a), 0.024 for (b), and 0.020 for (c).} \label{fig:si87h76_ml} \end{figure} \subsubsection*{Computational time savings} As demonstrated in Figure~\ref{fig:si10_si35_timing}, the greater the simulation cell is (the larger the number of plane waves $n_{\text{pw}}$ is), the more substantial are the savings in the computational time. \begin{figure}[H] \centering \begin{subfigure}{0.48\textwidth} \includegraphics[width=\linewidth]{figs/si10h16_cell_corehours_ratio.pdf} \end{subfigure} \hfill \begin{subfigure}{0.48\textwidth} \includegraphics[width=\linewidth]{figs/si35h36_cell_corehours_ratio.pdf} \end{subfigure} \caption{Number of core hours saved for the Si$_{10}$H$_{16}$ and Si$_{35}$H$_{36}$ clusters as a function of the size of the simulation cell when using ML-BSE instead of FF-BSE. The numerical values beside each data point is the corresponding $\alpha_d$. Because of the relatively small computational cost of ML-BSE for Si$_{10}$H$_{16}$, its $\alpha_d$ is more sensitive to the machine status than that of Si$_{35}$H$_{36}$ and is not necessarily monotonic as a function of the size of the simulation cell.} \label{fig:si10_si35_timing} \end{figure} \section{Timing and scaling of ML-BSE} The calculation of absorption spectra can be decomposed into the following three steps: \begin{enumerate} \item Calculation of unscreened integrals using Eq.~\ref{eq:tauu}; \item Calculation of screened integrals using Eq.~\ref{eq:delta_tau_delta_rho} (in FF-BSE) or Eq.\ref{eq:CNN} (in ML-BSE); \item Calculation of the frequency-dependent spectrum using Eq.~\ref{eq:spectrum}. \end{enumerate} Here we have replaced Step 2 with a machine learning model. For the largest system considered in this work (COOH-Si/water interface), at least $\sim$50\% of the total computational time is spent in Step 2. We define the speed-up $\alpha_d=t_d^{\text{FF-BSE}}/t_d^{\text{ML-BSE}}$ as the ratio of the core hours needed to carry out Step 2 with FF-BSE or ML-BSE. We found that $\alpha_d$ increases as $n_{\text{int}}$ and the number of plane waves $n_{\text{pw}}$ increase.

9aac22904d77828811a6479dc6a4bce64fee57c8

\section{Introduction} The language $O_n$ is the language built on the alphabet $\Sigma_n = \{a_i,\bar a_i\mid i \in [n]\}$ and that contains exactly all words $w$ which, for every $i$ in $[n]$, have the same number of occurrences of $a_i$ and $\bar a_i$. Writing $|w|_{c}$ the number of occurrences of the letter $c$ in $w$, this condition becomes $|w|_{a_i} = |w|_{\bar a_i}$ for all $i$ in $[n]$. The problem of situating the languages $O_n$ within known classes of languages has been raised at least in two different communities. The first one is that of computational linguistics. This problem has attracted attention with a language called $MIX$ (also called the Bach language as it was introduced by Bach~\cite{bach81:_discontinuous_constituents_generalized_categorioal_grammars, bach88:_categorial_grammar_as_theories_of_languages, pullum83:_context_freeness_and_the_computer_processing_of_human_languages}) that is rationally equivalent to $O_2$. It was related to the research program of Joshi~\cite{joshi85:_tree_adjoining_grammars:_how_much_context_sensitivity_is_required_to_provide_reasonnable_structural_descriptions} which consists in formally describing the class of languages corresponding to Natural Languages. According to Joshi's terminology, this is the class of Mildly Context Sensitive Languages. This program tries to give abstract properties of this class while describing possible candidates \cite{weir_phd,joshi91:_the_convergence_of_mildly_context_sensitive_grammar_formalisms}. Among such candidates, a powerful one is formed by Multiple Context Free Languages (MCFL)~\cite{seki91:_multiple_context_free_grammars}. According to Joshi \textit{et al.}~\cite{joshi91:_the_convergence_of_mildly_context_sensitive_grammar_formalisms} the language ``$MIX$ can be regarded as the extreme case of free word order'' and Mildly Context Sensitive Languages should ``perhaps'' not contain $MIX$ (and thus $O_2$). This paper also mentions that the position of $MIX$ in classes of languages, such as Tree Adjoining Languages, is not known and difficult to establish. In particular, it stresses that it is not known whether $MIX$ is an Indexed Language. The second community that has also shown interest for the problem is that of computational group theory which tries to identify properties of groups by means of properties of their word problem. The word problem for a group consists in describing the language of words that are equal to zero for a given presentation (all presentations giving rise to rationally equivalent languages). For example, Muller and Schuppe have characterized virtually free groups as exactly those groups whose word problems are solved by context free grammars~\cite{muller83:_group_theory_the_theory_of_ends_and_context_free_languages}. A question that has been raised by that community is whether $O_2$---which coincides with the language corresponding to the word problem for the additive group $(\mathbb{Z}^2,{+})$---is an Indexed Language~\cite{gilman05:_formal_languag_applic_combin_group_theor}. This question remains open. MCFL form a natural generalization of ``copyless'' Macro-Languages (see~\cite{DBLP:journals/ieicet/SekiK08} for a discussion) and Macro-Languages are equivalent to Indexed Languages. This makes the question of whether $O_2$ belongs to MCFL relevant. A first important result in that line of research is that $O_2$ is not a well-nested MCFL of dimension $2$~\cite{kanazawa_salvati12:_mix}, which solves a long-standing open problem raised by Joshi~\cite{joshi85:_tree_adjoining_grammars:_how_much_context_sensitivity_is_required_to_provide_reasonnable_structural_descriptions}. Subsequently it has been shown that it is actually an MCFL~\cite{salvati15:mix_2mcfl} and more precisely an MCFL of dimension $2$ (a $2$-MCFL). Nederhof~\cite{DBLP:conf/acl/Nederhof16} has given a similar proof of the same result. Later he conjectured that $O_n$ is an $n$-MCFL~\cite{Nederhof-2017-free-word-orders-and-MCFLs}. As pointed in this work, a simple pumping argument shows that $O_n$ cannot be an MCFL of dimension strictly smaller than $n$. More recently, a breakthrough has been achieved by Ho~\cite{Thewordproblemofnisamultiplecontextfreelanguage} who proved that $O_n$ is an MCFL for every $n$. However the MCFL constructed in the proof is of dimension larger than $n$, namely $8\left\lfloor \frac{n+1}{2} \right\rfloor-2$. All proofs related to these results are based on algebraic topology. While the proofs of \cite{salvati15:mix_2mcfl} and \cite{DBLP:conf/acl/Nederhof16} strongly rely on topological properties of the plane (existence of winding numbers of curves around points), the aforementioned proof by Ho is based on the well-known Borsuk--Ulam theorem, a powerful theorem from algebraic topology which holds in any dimension. More specifically, it uses a combinatorial application of this theorem, due to Alon and West~\cite{alon1986borsuk}: the necklace splitting theorem. In this paper, we also rely on related tools to prove Nederhof's conjecture, namely that $O_n$ is an $n$-MCFL. Nederhof actually conjectures that a particular Multiple Context-Free Grammar (MCFG) of dimension $n$ defines $O_n$. We prove a slightly stronger result by showing that a grammar that uses a more restricted kind of rules is sufficient to define $O_n$. The article is structured as follows: In \Cref{sec:prelim}, we introduce notation regarding formal languages and MCFG as well as the grammar \(G_n\), which is an MCFG of dimension $n$. \Cref{sec:main-result} establishes the main result namely that the language of \(G_n\) is \(O_n\) using a decomposition lemma. The decomposition lemma can on the one hand be derived from a variant of the necklace splitting theorem that we prove in \Cref{sec:necklace}. Alternatively, it can be obtained via purely combinatorial proofs presented in \Cref{sec:combinatorial-proofs}. \section{Background on Multiple Context Free Grammars} \label{sec:prelim} We write $[n]$ for the set $\{1,\dots,n\}$. For a given finite set $\Sigma$, also called \emph{alphabet}, we write $\Sigma^\ast$ for the monoid freely generated by $\Sigma$, and $\Sigma^{+}$ for the free semigroup generated by $\Sigma$. The elements of $\Sigma$ are called \emph{letters} while the elements of $\Sigma^\ast$ and $\Sigma^{+}$ are called \emph{strings} or \emph{words} and we write $\varepsilon$ for the empty word. Given a word $w$, we write $|w|$ for its length, and $|w|_c$ for the number of occurrences of the letter $c$ in $w$. A \emph{language} is a subset of $\Sigma^\ast$. We define the language $O_n$ as $\{w\in \Sigma_n^\ast\mid |w|_{a_i} = |w|_{\bar a_i}\text { for } i \in [n]\}$ where $\Sigma_n$ is the alphabet $\{a_i,\bar a_i \mid i \in [n]\}$. For $\alpha$ in $\Sigma_n$, writing $\bar\alpha$ exchanges $a_i$ and $\bar a_i$: if $\alpha$ is $a_i$, then $\bar \alpha$ is $\bar a_i$; if $\alpha$ is $\bar a_i$, then $\bar \alpha$ is $a_i$. Two letters $\alpha$ and $\beta$ of $\Sigma_n$ are \emph{compatible} when $\alpha = \bar\beta$. We extend the $(\bar \cdot)$ operation to words in $\Sigma_n^\ast$ as follows: the word $\bar w$ is obtained from $w$ by applying $(\bar\cdot)$ to each of its letters. A ranked alphabet $\Omega$ is a pair $(\mathcal{A},\rho)$ where $\mathcal{A}$ is a finite set and $\rho$ is a function from $\mathcal{A}$ to $\mathbb{N}$. For $a$ in $\mathcal{A}$, the integer $\rho(a)$ is the \emph{rank} of $a$. We shall write $\Omega^{(n)}$ for the set $\{a \in \mathcal{A} \mid \rho(a)=n\}$. The \emph{dimension} of a ranked alphabet is the maximal rank of its elements. A Multiple Context Free Grammar (MCFG) $G$ is a tuple $(\Omega, \Sigma, R, S)$ where $\Omega$ is a ranked alphabet of \emph{non-terminals}, $\Sigma$ is a finite set of \emph{letters}, $R$ is a set of \emph{rules} and $S$ is an element of $\Omega^{(1)}$ called \emph{initial non-terminal}. The rules in $R$ are of the form \begin{equation}\label{eq:rule} A(w_1,\ldots, w_n) \Rightarrow B_1 (x_{1,1},\ldots, x_{1,l_1}), \ldots, B_p(x_{p,1},\ldots, x_{p,l_p}) \end{equation} where $A$ is in $\Omega^{(n)}$, $B_k$ is in $\Omega^{(l_k)}$, the $x_{k,j}$ are pairwise distinct variables and the $w_j$ are elements of $(\Sigma\cup X)^\ast$ with $X = \{x_{k,j} \mid k \in [p] \land j\in [l_k]\}$ and with the restriction that each $x_{k,j}$ may have at most one occurrence in the string $w_1\cdots w_n$. Note that $p$ may be equal to $0$, in which case the right part of the rule (the one on the right of the $\Rightarrow$ symbol) is empty. Then we may write the rule by omitting the symbol $\Rightarrow$. The \emph{dimension} of an MCFG is that of its ranked alphabet of non-terminals. An MCFG of dimension at most $n$ is an $n$-MCFG. An MCFG such as $G$ defines \emph{judgments} of the form $\vdash_G A(s_1,\ldots, s_n)$ where $A$ is in $\Omega^{(n)}$ and the $s_j$ belong to $\Sigma^\ast$. The notion of derivable judgments is defined inductively: suppose we are given $p$ derivable judgments $\vdash_{G} B_k(s_{k,1},\ldots, s_{k,l_k})$ where $B_k \in \Omega^{(l_k)}$ for $k$ in $[p]$. For each rule of the form~\eqref{eq:rule}, the judgment $\vdash_G A(s_1, \ldots, s_n)$ is \emph{derivable} when each $s_j$ is obtained from $w_j$ by replacing each occurrence of the variable $x_{k,j}$ by $s_{k,j}$. The language defined by $G$, denoted by $\mathcal{L}(G)$, is the set $\{w \in \Sigma^\ast \mid \vdash_G S(w) \mbox{ is derivable}\}$. The class of languages that are definable by MCFGs is the class of \emph{Multiple Context-Free Languages} (MCFL). Likewise, the class of languages definable by $n$-MCFGs is the class of \emph{$n$-Multiple Context-Free Languages} ($n$-MCFL). We define now $G_n$, the central $n$-MCFG for which we prove that it generates $O_n$. It uses two non-terminals $S$ and $I$ that are respectively of rank $1$ and $n$. The non-terminal $S$ is the initial one. The alphabet of $G_n$ is $\Sigma_n$. The rules of the grammar $G_n$ are the following: \begin{enumerate} \item\label{rule:init} $S(x_1\cdots x_n)\Rightarrow I(x_1,\dots,x_n)$. \item\label{rule:binary} $I(w_1,\dots, w_n)\Rightarrow I(x_1,\dots,x_n), I(y_1,\dots,y_n)$ \qquad for all \(w_1, \ldots, w_n \in \{x_1, \ldots, x_n, y_1, \ldots, y_n\}^\ast\) such that $w_1\cdots w_n = x_1y_1\cdots x_ny_n$. \item\label{rule:const} $I(w_1,\dots,w_n)\Rightarrow I(x_1,\dots,x_n)$ for all \(w_1, \ldots, w_n\) and all $\alpha\in\Sigma_n$, $k,\ell \in [n]$ such that $w_j = x_j$ for $j \neq k,\ell$ and \begin{itemize} \item $k \neq \ell$ implies $w_k \in \{\alpha x_k, x_k \alpha\}$ and $w_{\ell} \in \{\bar\alpha x_\ell, x_\ell \bar\alpha\}$. \item $k = \ell$ implies $w_k =\alpha x_k \bar\alpha$. \end{itemize} \item\label{rule:empty} $I(\varepsilon,\dots,\varepsilon)$. \end{enumerate} Items numbered \eqref{rule:binary} and \eqref{rule:const} describe finite sets of rules. The rules~(\ref{rule:binary}) are parametrized by a particular factorization $(w_1,\dots,w_n)$ of $x_1y_1\cdots x_ny_n$. For example, when $n=3$ letting $w_1 = x_1y_1x_2$, $w_2 = y_2x_3$ and $w_3 =y_3$ is such a factorization; we have $w_1w_2w_3 = x_1y_1x_2y_2x_3y_3$. A rule of the form~(\ref{rule:const}) adds a compatible pair of letters at distinct endpoints of the words. For example, $I(a_1x_1,x_2,\bar a_1x_3,x_4)\Rightarrow I(x_1,x_2,x_3,x_4)$ is such a rule for $n=4$. \begin{example} The grammar $G_2$ contains the following rules: \allowdisplaybreaks \begin{eqnarray*} S(x_1x_2) &\Rightarrow& I(x_1,x_2)\\ I(x_1y_1x_2y_2,\varepsilon)&\Rightarrow & I(x_1,x_2),\,I(y_1,y_2)\\ I(x_1y_1x_2,y_2)&\Rightarrow & I(x_1,x_2),\,I(y_1,y_2)\\ I(x_1y_1,x_2y_2)&\Rightarrow & I(x_1,x_2),\,I(y_1,y_2)\\ I(x_1,y_1x_2y_2)&\Rightarrow & I(x_1,x_2),\,I(y_1,y_2)\\ I(\varepsilon,x_1y_1x_2y_2)&\Rightarrow & I(x_1,x_2),\,I(y_1,y_2)\\ I(\alpha x_1\bar{\alpha},x_2)&\Rightarrow&I(x_1,x_2) \quad \alpha\in \Sigma_2\\ I(\alpha x_1,\bar{\alpha}x_2)&\Rightarrow&I(x_1,x_2) \quad \alpha\in \Sigma_2\\ I(\alpha x_1,x_2\bar{\alpha})&\Rightarrow&I(x_1,x_2) \quad \alpha\in \Sigma_2\\ I(x_1\alpha,\bar{\alpha}x_2)&\Rightarrow&I(x_1,x_2) \quad \alpha\in \Sigma_2\\ I(x_1\alpha,x_2\bar{\alpha})&\Rightarrow&I(x_1,x_2) \quad \alpha\in \Sigma_2\\ I(x_1,\alpha x_2\bar{\alpha})&\Rightarrow&I(x_1,x_2) \quad \alpha\in \Sigma_2\\ I(\varepsilon,\varepsilon)\\ \end{eqnarray*} \end{example} It is usual to present derivations as trees where nodes have the following form: \begin{center} \begin{prooftree} \vdash_G B_1(s_{1,1},\dots,s_{1,l_1}) \dots \vdash_G B_p(s_{p,1},\dots,s_{p,l_p}) \justifies% \vdash_G A(s_1,\dots,s_n) \end{prooftree} \end{center} when from the derivations of the $\vdash_G B_i(s_{i,1},\dots,s_{i,l_i})$ we can derive $\vdash_G A(s_1,\dots,s_n)$ using a rule of $G$. When $p$ equals $0$, there is no derivation above the bar. This means that rules with no right-hand part form the leaves of these trees. Using this notation we can show as follows that $a_1 a_1 \bar{a}_2\bar{a}_1 \bar{a}_1 a_2$ is in the language of $G_2$. We use colors to identify each letter and allow one to infer the rules used in the derivation: \begin{center} \begin{prooftree} \begin{prooftree} \begin{prooftree} \begin{prooftree} \justifies% \vdash_{G_2}I(\varepsilon,\varepsilon) \end{prooftree} \justifies% \vdash_{G_2} I(\textcolor{red}{a_1},\textcolor{red}{\bar a_1}) \end{prooftree} \begin{prooftree} \begin{prooftree} \begin{prooftree} \justifies% \vdash_{G_2}I(\varepsilon,\varepsilon) \end{prooftree} \justifies% \vdash_{G_2}I(\textcolor{green!50!black}{\bar{a}_2}, \textcolor{green!50!black}{a_2}) \end{prooftree} \justifies% \vdash_{G_2}I(\textcolor{blue}{a_1} \textcolor{green!50!black}{\bar a_2}, \textcolor{blue}{\bar a_1}\textcolor{green!50!black}{a_2}) \end{prooftree} \justifies% \vdash_{G_2} I(\textcolor{red}{a_1}, \textcolor{blue}{a_1} \textcolor{green!50!black}{\bar a_2}\textcolor{red}{\bar a_1}\textcolor{blue}{\bar a_1} \textcolor{green!50!black}{a_2}) \end{prooftree} \justifies% \vdash_{G_2}S(\textcolor{red}{a_1} \textcolor{blue}{a_1} \textcolor{green!50!black}{\bar a_2}\textcolor{red}{\bar a_1}\textcolor{blue}{\bar a_1} \textcolor{green!50!black}{a_2}) \end{prooftree} \end{center} \section{Main result} \label{sec:main-result} Our main theorem is that the language of $G_n$ is $O_n$. The inclusion of $\mathcal{L}(G_n)$ into $O_n$ is obvious and the challenge consists in proving the converse inclusion. We prove actually a stronger statement: \emph{$\vdash_{G_n}I(s_1,\dots,s_n)$ is derivable for every $s_1 \cdots s_n$ in $O_n$.} With this statement, the desired inclusion is immediate: for $w\in O_n$, set $s_1=w$ and $s_j=\varepsilon$ for all $j\geq 2$ and apply rule~\eqref{rule:init}. A tuple $(s_1,\dots,s_n)$ is \emph{reducible} if there are at least two compatible letters among the endpoints of the $s_j$'s; otherwise the tuple is \emph{irreducible}. Repeated applications of rules of the form~\eqref{rule:const} allow to derive reducible tuples from irreducible ones. Irreducible tuples are dealt with the following ``decomposition lemma,'' which shows how a rule of the form~\eqref{rule:binary} is applied to derive an irreducible tuple from smaller elements in $O_n$. While rule~\eqref{rule:empty} provides the base case for the induction, this lemma is the main technical result towards the proof of \Cref{thm:o_n_n_mcfl}. It relies on \Cref{thm:multi-neckl}, a result stated and proved in \Cref{sec:necklace}, and formulated with the terminology of the necklace splitting theorem. \begin{lemma}[Decomposition lemma] \label{lem:decomposition} Consider an irreducible tuple $(s_1,\dots, s_n)$ in $(\Sigma_n^{+})^n$, where each $s_j$ is of length at least $2$. If $s_1\cdots s_n$ belongs to $O_n$, then there exist two tuples $(u_1, u_3, \dots,u_{2n-1})$ and $(u_2, u_4, \dots,u_{2n})$ in $(\Sigma_n^{\ast})^n$ and integers \(0 = k_0 \leq k_1 \leq \cdots \leq k_n = 2n\) such that \begin{enumerate}[label=\textup{(\Roman*)}] \item \label{prop:nonempty-O_n}$u_1 u_3 \cdots u_{2n-1}$ and $u_2 u_4 \cdots u_{2n}$ are both nonempty elements of $O_n$, and \item\label{prop:interlacing} $s_j = u_{k_{j -1} + 1} ~ u_{k_{j -1} + 2} \cdots u_{k_j}$ for each \(j \in [n]\). \end{enumerate} In particular, \(s_1 \cdots s_n = u_1 u_2 \cdots u_{2n}\). \end{lemma} \begin{proof} We distinguish the cases $n=1$ and $n\geq 2$. We deal first with the case $n=1$. In that case, $\Sigma_n=\{a_1,\bar a_1\}$. As $s_1$ is irreducible, without loss of generality, we may assume that $s_1 = a_1wa_1$. We consider prefixes $u'$ of $w$ of increasing length, from $|u'|=0$ to $|u'|=|w|$. Since $s_1$ belongs to $O_1$, the quantity $|a_1u'|_{a_1}-|a_1u'|_{\bar a_1}$ starts with the value $1$ and finishes with the value $-1$, and changes by steps of one unit. There is therefore a prefix $u'$ of $w$ such that $a_1u'$ belongs to $O_1$. Setting $u_1=a_1u'$ and $u_2\in\Sigma_1^+$ such that $s_1=u_1u_2$ makes the job. We deal now with the case $n\geq 2$. The proof of \Cref{thm:multi-neckl} builds explicitly $u_{\ell}$'s satisfying the desired properties: property~\ref{prop:nonempty-O_n} is a consequence of the fact that each of $A$ and $B$ are balanced; property~\ref{prop:interlacing} is a consequence of the fact that the endpoints of the $s_j$'s form cuts. \end{proof} From this the inclusion follows easily. \begin{proposition} \label{prop:O_n_included_in_L_G_n} The judgment $\vdash_{G_n}I(s_1,\dots,s_n)$ is derivable for every $s_1\cdots s_n$ in $O_n$. In particular, we have $O_n\subseteq\mathcal{L}(G_n)$. \end{proposition} \begin{proof} As mentioned above, the second part of the statement is a direct consequence of the first part. We proceed by induction on the pairs $(|s_1\cdots s_n|,e)$ ordered lexicographically, where $e$ is the number of $s_j$ equal to $\varepsilon$. Suppose first that $(s_1,\dots,s_n)$ is reducible. A rule of the form~\eqref{rule:const} shows that we can derive the judgment from another judgment $\vdash_{G_n}I(s'_1,\dots,s'_n)$, with $|s'_1\cdots s'_n|<|s_1\cdots s_n|$. The induction hypothesis provides the conclusion. Suppose now that $(s_1,\dots,s_n)$ is irreducible. Four cases are in order, distinguished according to the possible lengths of the $s_j$'s. The first case is when at least one $s_j$ is of length $1$. Without loss of generality we may assume that $s_j = a_1$. As $s_1\cdots s_n$ is in $O_n$, there is a $k$ such that $s_k = v_1\bar a_1 v_2$ for some $v_1$ and $v_2$ in $\Sigma_n^\ast$. The other case being similar, we suppose that $j<k$. Define $t_j = a_1$, $t_k = \bar a_1$, and $t_{\ell} = \varepsilon$ for $\ell\neq j,k$; define also $(u_1,\dots,u_n) = (s_1,\dots, s_{j-1},s_{j+1}, \dots ,s_{k-1},v_1,v_2, s_{k+1}, \dots , s_n)$. Using rule~\eqref{rule:empty} and a rule of the form~\eqref{rule:const}, we get that $\vdash_{G_n}I(t_1,\dots, t_n)$ is derivable. The judgment $\vdash_{G_n}I(u_1,\dots,u_n)$ is derivable by induction. Then using a rule of the form~\eqref{rule:binary} shows that $\vdash_{G_n}I(s_1,\dots, s_n)$ is derivable. More precisely, we instantiate each variable $x_\ell$ with $t_\ell$ and each variable $y_\ell$ with $u_\ell$ in the following rule \begin{multline*} I(x_1y_1,\dots,x_{j-1}y_{j-1},x_{j}, y_j,\dots ,x_{k-2}y_{k-2}x_{k-1},y_{k-1}x_{k}y_k,\dots, x_n y_n)\Rightarrow \\ I(x_1,\dots, x_n), I(y_1,\dots,y_n ) \enspace . \end{multline*} The second case is when all $s_j$ are equal to $\varepsilon$. The conclusion follows from an application of rule~\eqref{rule:empty}. The third case is when some $s_j$ but not all are equal to $\varepsilon$ and no $s_j$ is of length $1$. There is then an $j \in [n-1]$ such that either $s_j=\varepsilon$ and $|s_{j+1}|>1$, or $|s_j|> 1$ and $s_{j+1}=\varepsilon$. By symmetry, we suppose that $s_j=\varepsilon$ and $|s_{j+1}|>1$. As $|s_{j+1}|>1$, we have that $s_{j+1} = v_1v_2$ with $v_1$ and $v_2$ in $\Sigma^+$. Define $(s'_1,\dots,s'_n) = (s_1,\dots,s_{j-1},v_1,v_2, s_{j+2}, \dots,s_n)$. The judgment $\vdash_{G_n}I(s'_1,\dots,s'_n)$ is derivable by induction (we have a smaller $e$). The judgment $\vdash_{G_n}I(\varepsilon,\dots,\varepsilon)$ is derivable from an application of rule~\eqref{rule:empty}. Then using a rule of the form~\eqref{rule:binary} shows that $\vdash_{G_n}I(s_1,\dots, s_n)$ is derivable. More precisely, we instantiate each variable $x_\ell$ with $s'_\ell$ and each variable $y_\ell$ with $\varepsilon$ in the following rule: \[ I(x_1y_1,\dots, x_{j-1}y_{j-1},\varepsilon,x_i y_jx_{j+1}y_{j+1},\dots, x_ny_n)\Rightarrow I(x_1,\dots, x_n), I(y_1,\dots,y_n ) \enspace . \] The fourth case satisfies the conditions of \Cref{lem:decomposition} (``decomposition lemma''), which shows that $\vdash_{G_n}I(s_1,\dots, s_n)$ is derivable by an application of a rule of the form~\eqref{rule:binary} and by induction. \end{proof} From this we can derive our main theorem. \begin{theorem} \label{thm:o_n_n_mcfl} The language $O_n$ is an $n$-MCFL. \end{theorem} \begin{remark} Theorem~\ref{thm:multi-neckl} actually implies a version of \Cref{lem:decomposition} that also holds for reducible tuples albeit only if $n$ is at least $2$ and if we permit to decompose $(s_1, \ldots, s_n)$ in more versatile tuples. This corresponds to a slightly more liberal definition of the rules~\eqref{rule:binary}. More precisely, we may add to the rules~\eqref{rule:binary} for each \(j \in [n]\) the rule: \begin{equation*}\label{rule:binary_liberal} I(x_1 y_1, \ldots, x_{j-1}y_{j-1}, \text{\boldmath\bfseries{$y_j x_j$}},x_{j+1} y_{j+1}, \ldots, x_ny_n) \Rightarrow I(x_1, \ldots, x_n),\, I(y_1, \ldots, y_n) \enspace. \tag{\textasteriskcentered} \end{equation*} We have chosen to exclude rules~\eqref{rule:binary_liberal} to emphasize the surprising simplicity of the rules~\eqref{rule:binary} which allow to decompose every irreducible tuple. Moreover, we want to contrast the grammar we obtain with the one that Nederhof~\cite{Nederhof-2017-free-word-orders-and-MCFLs} conjectures to capture $O_n$. Nederhof proposes binary rules of the following form: \[A(w_1,\dots,w_n)\Rightarrow A(x_1,\dots,x_n),\,A(y_1,\dots,y_n)\] where $w_1\cdots w_n$ is obtained by shuffling the words $x_1\cdots x_n$ and $y_1\cdots y_n$, i.e., $|w_1\cdots w_n|=2n$, removing all occurrences of \(y_j\)'s from $w_1 \cdots w_n$ yields $x_1 \cdots x_n$ and, analogously, removing all occurrences of \(x_j\)'s from $w_1 \cdots w_n$ yields $y_1 \cdots y_n$. Furthermore, a $w_k$ may not contain an occurrence of $x_jx_{j+1}$ or of $y_{j}y_{j+1}$ for some $j \in [n-1]$. The rules~\eqref{rule:binary} may be obtained from Nederhof's by forbidding the occurrence of $x_jx_{j+1}$ and of $y_{j}y_{j+1}$ not only in the $w_k$'s but rather in the combined word $w_1\cdots w_n$. Note however that this additional restriction implies that rules~\eqref{rule:const} need to allow the removal of compatible letters at arbitrary endpoints of the words of a tuple. The corresponding rules of Nederhof's grammar are less liberal. Were we to remove rules~\eqref{rule:const} and treat the elimination of compatible letters as in~\cite{Nederhof-2017-free-word-orders-and-MCFLs}, then we would need to add the rules~\eqref{rule:binary_liberal}. Clearly, these rules form a strict subset of those proposed by Nederhof as $w_k$'s are of length $2$ and at most one occurrence of $x_jx_{j+1}$ is allowed in $w_1\cdots w_n$. In this sense, our grammar is simpler than Nederhof's. Notably, also Nederhof \cite[Sec.~5.3]{Nederhof-2017-free-word-orders-and-MCFLs} conjectures for \(O_3\) that some of the rules of his grammar are redundant. \end{remark} \section{Necklace splitting and proof of the decomposition lemma}\label{sec:necklace} In this section, we prove a combinatorial theorem in the tradition of the necklace splitting problem. The theorem almost implies \Cref{lem:decomposition}, but while it does not require irreducibility, it misses the property~\ref{prop:interlacing}. Its proof however gives the construction for the case $n>1$ of \Cref{lem:decomposition}. This combinatorial theorem is formulated without the terminology of languages and grammars so as to make it understandable easily without background in this area. For readers who are more familiar to manipulating words, the vocabulary of the necklace problem translates easily to that of words: necklaces become words and beads become letters. In the traditional version of the necklace splitting theorem, there is an open necklace with beads of $n$ different types, and an even number of beads of each type. The necklace splitting theorem ensures that such a necklace can always be split between two thieves with no more than $n$ cuts so that each thief gets the same amount of each type. The cuts have to leave the beads untouched. Here, we keep the same setting, except that a bead can be either ``positive'' or ``negative.'' The {\em amount} of a type in a necklace or in a collection of necklaces is the number of positive beads of this type minus the number of negative beads of this type. A necklace or a collection of necklaces is {\em balanced} if the amount of each type is zero. \begin{theorem}\label{thm:multi-neckl} Consider a collection of $n$ open necklaces with positive and negative beads. Suppose that there are $n\geq 2$ types of beads and that each of the necklaces has at least two beads. If the collection is balanced, then there is a way to cut the necklaces using at most $n$ cuts in total and partition the subnecklaces into two parts so that each part is balanced, gets at least one bead (and thus at least two), and is formed by at most $n$ subnecklaces. \end{theorem} The connection to \Cref{lem:decomposition} is as follows: \begin{itemize} \item The $n$ types of beads are the numbers $1, \dots,n$, the letter $a_i$ representing a \emph{positive} bead of type $i$ and the letter $\bar{a}_i$ representing a \emph{negative} bead of type $i$. \item The $n$ open necklaces correspond to the words $s_1, \dots, s_n$. \item The two parts of the at most $n$ subnecklaces are the tuples $(u_1,\dots,u_{2n-1})$ and $(u_2,\dots, u_{2n})$ in the lemma, which, as in the theorem, are required to be balanced and to contain at least one letter each. \end{itemize} The proof of \Cref{thm:multi-neckl} relies on the following proposition, which is actually the traditional necklace splitting theorem extended to negative beads (the relaxation of the parity condition of the number of beads is standard; see, e.g.,~\cite[Section 5.1]{alon2006algorithmic}). \begin{proposition}\label{prop} Consider a necklace with positive and negative beads. Suppose that there are $n\geq 1$ types of beads. Then the necklace can be split between two thieves with no more than $n$ cuts so that the amount of beads received by the thieves differ by at most one unit for each type. It is moreover possible to choose which thief receives an extra bead for each type which requires so. \end{proposition} The types requiring that one of the thieves receives an extra bead are precisely those whose total amount of beads is an odd number. \begin{proof}[Proof of \Cref{prop}] We proceed in two steps. In a first step, we prove a continuous version, in which we are allowed momentarily to locate cuts on beads themselves. In a second step, we show how to move the cuts so as to get a splitting with cuts leaving the beads untouched.\footnote{An alternative purely combinatorial proof is presented in \Cref{sec:alt-proof-prop2}.} We identify the necklace with $[0,1]$ and the beads with intervals included in $[0,1]$, all of same length (this length is the inverse of the total number of beads in the necklace), and with disjoint interiors. Assume that these small intervals are all open. (This assumption is made to ease the proof, but actually whether these small intervals contain or not their boundaries does not matter.) Define $$ g_i(x) \longmapsto \left \{ \begin{array}{ll} +1 & \text{if $x$ is in a interval corresponding to a positive bead of type $i$.} \\ -1 & \text{if $x$ is in a interval corresponding to a negative bead of type $i$.} \\ 0 & \text{otherwise.} \end{array} \right. $$ According to the Hobby--Rice theorem~\cite{hobby1965moment}, there are points $0=x_0 < x_1 < \cdots < x_r < x_{r+1}=1$ with $r \leq n$ such that for all $i \in [n]$ $$\sum_{j=1}^{r+1} (-1)^j \int_{x_{j-1}}^{x_j}g_i(u)\operatorname{d}u=0\enspace.$$ (Here, we take the formulation given by Pinkus~\cite{pinkus1976simple}.) The $r$ points $x_1,\ldots,x_r$ can be interpreted as cuts. The intervals $(x_{j-1},x_j)$ with $j$ odd are given to one thief, and the intervals $(x_{j-1},x_j)$ with $j$ even are given to the other thief. Each thief gets the same amount of each type. The only problem is that some cuts may be located on beads. The second step of the proof consists in explaining how to move these cuts so that none of them touch the beads anymore, without creating a difference of more than one unit between the amounts received by the thieves for each type. We can make that each bead is cut at most twice since moving two cuts inside a bead in the same direction and by the same distance does not change the amounts received by the thieves. If a bead is cut twice, we can similarly move the two cuts until one of them is located between two beads. (Doing this, we can have several cuts located at the same position between two beads, but this is not an issue.) So, we can assume that each bead is cut at most once and that the two parts of a cut bead go to distinct thieves. If a type has two beads or more touched by a cut, then we can move two cuts so that one at least reaches a position between two beads, without changing the amounts received by each thief. Thus, we can assume that each type is cut at most once. We finish the proof by noting that if a type has a bead that is cut, it means that each thief received a non-integral amount of the corresponding type, i.e., a half-integer, and moving the cut arbitrarily leads to the desired splitting. \end{proof} \begin{proof}[Proof of \Cref{thm:multi-neckl}] Denote by $s_1,\ldots,s_n$ the $n$ necklaces. We assume that there are no two beads located at the endpoints of some necklaces and that are of the same type but of opposite signs, for otherwise there would be an easy solution: cut these two beads from the necklaces, form a balanced part with them, and leave the rest of the remaining beads for the second balanced part. The number of cuts would then be $2\leq n$ (or, if these two beads formed a necklace of their own, the solution would need no cut), and the number of subnecklaces in the parts would be $2$ and $n$ (or $1$ and $n-1$). Making this assumption corresponds to considering only the irreducible case in the terminology of \Cref{thm:o_n_n_mcfl}. A natural idea would be to apply a result like \Cref{prop} to the ``big'' necklace $s=s_1\cdots s_n$ obtained by appending the necklaces in their index order. The first issue with this idea is that one thief might get nothing. This can easily be dealt with as done below. The second issue is that, even though there are $2n$ subnecklaces in total, one thief might get more than $n$ subnecklaces.\footnote{ In essence, this is the reason why Ho~\cite{Thewordproblemofnisamultiplecontextfreelanguage} could only show that $O_n$ is an $\left(8\left\lfloor \frac{n+1}{2} \right\rfloor-2\right)$-MCFL.} Instead we consider another big necklace, which we will denote by $s'$ and which we define now. We start by defining $s'_1$ to be $s_1$ from which the left-most bead has been removed, and $s'_n$ to be $s_n$ from which the right-most bead has been removed. Note that without loss of generality, we can assume that the left-most bead of $s_1$ is a positive bead of type $1$ and that the right-most bead of $s_n$ is a positive bead of type $1$ or $n$. We thus consider two cases: \begin{enumerate}[label=(\roman*)] \item\label{case:11} The left-most bead of $s_1$ and the right-most bead of $s_n$ are both positive beads of type $1$. \item\label{case:1n} The left-most bead of $s_1$ is a positive bead of type $1$ and the right-most bead of $s_n$ is a positive bead of type $n$. \end{enumerate} The condition of the theorem ensures that neither $s'_1$ nor $s'_n$ are empty. Consider the ``big'' necklace $s' = s'_1\bar s_2s_3\bar s_4\cdots$. If $n$ is even, the big necklace $s'$ ends with $\bar s'_n$, and if $n$ is odd, it ends with $s'_n$. (Given a sequence $t$ of beads, the notation $\bar t$ means $t$ where all positive beads become negative and conversely, without changing their types.) According to \Cref{prop}, there is a splitting of $s'$ into $n+1$ subnecklaces $t_1,\ldots,t_{n+1}$ (completing with zero-length subnecklaces if necessary) such that we have for $i \in [n]$ in Case~\ref{case:11} and for $i\in\{2,\ldots,n-1\}$ in Case~\ref{case:1n} \begin{align} \displaystyle{\sum_{k\text{ odd}}\mu_i(t_k)} & = \displaystyle{\sum_{k\text{ even}}\mu_i(t_k)} \enspace,\label{eqi} \end{align} and such that in Case~\ref{case:1n} \begin{align} \displaystyle{\sum_{k\text{ odd}}\mu_1(t_k)} & = \displaystyle{\sum_{k\text{ even}}\mu_1(t_k)-1} \enspace,\label{1teq1} \\ \displaystyle{\sum_{k\text{ odd}}\mu_n(t_k)} & = \displaystyle{\sum_{k\text{ even}}\mu_n(t_k)+1} \enspace. \label{1teqn} \end{align} Here, we denote by $\mu_i(t)$ the amount of beads of type $i$ in the subnecklace $t$. We remind the reader that the amount of beads is the number of positive beads minus the number of negative beads of this type. We interpret now the endpoints of the subnecklaces $t_k$ as cuts of the ``big'' necklace $s$ (the one formed by the original necklaces $s_j$). Together with the endpoints of the $s_j$, we get a splitting of $s$ into $2n$ subnecklaces $u_1,\ldots,u_{2n}$ (in this order). Some $u_{\ell}$ can be zero-length subnecklaces. We make two parts: a part $A$ formed by the $u_{\ell}$ with an odd $\ell$, and a part $B$ formed by the $u_{\ell}$ with an even $\ell$. Remark two things: \begin{itemize} \item Since the left-most bead of $s_1$ belongs to $u_1$ and the right-most bead of $s_n$ belongs to $u_{2n}$, each part contains at least one bead \item The number of subnecklaces $u_{\ell}$ is the same in each part, and thus equal to $n$. \end{itemize} We finish the proof by checking that both $A$ and $B$ are balanced. We denote by $q^A_i$ and $q^B_i$ the amount of beads of type $i$ in $A$ and $B$, respectively. Since the original collection is balanced, we have $q^A_i+q^B_i=0$. The end of the proof consists simply in checking that we have also $q^A_i-q^B_i=0$, which implies then immediately that $q^A_i=q^B_i=0$, as desired. Each bead $x$ belongs to exactly one $s_j$. Moreover, apart from the two beads at the endpoints of $s$, any bead $x$ (resp. $\bar x$) belongs also to exactly one $t_k$ when $j$ is odd (resp. even). It is immediate to check that $j+k-1$ and the index $\ell$ of the $u_{\ell}$ to which $x$ belongs have the same parity. Thus if $j+k-1$ is odd, then $x$ belongs to $A$, and if it is even, then $x$ belongs to $B$. We have \[ q^A_1 = 1 + \sum_{j,k\text{ odd}} \mu_1(s_ j\cap t_k) - \sum_{j,k\text{ even}} \mu_1(\bar s_ j\cap t_k) ~~ \mbox{and} ~~ q^B_1 = \delta_{\text{\ref{case:11}}} + \sum_{\substack{j\text{ odd} \\ k\text{ even}}} \mu_1(s_ j\cap t_k) - \sum_{\substack{j\text{ even} \\ k\text{ odd}}} \mu_1(\bar s_ j\cap t_k) \enspace , \] and for $i \neq 1$, we have \[ q^A_i = \sum_{j,k\text{ odd}} \mu_i(s_ j\cap t_k) - \sum_{j,k\text{ even}} \mu_i(\bar s_ j\cap t_k) \quad \mbox{and} \quad q^B_i = \delta^{i = n}_{\text{\ref{case:1n}}} + \sum_{\substack{j\text{ odd} \\ k\text{ even}}} \mu_i(s_ j\cap t_k) - \sum_{\substack{j\text{ even} \\ k\text{ odd}}} \mu_i(\bar s_ j\cap t_k) \enspace , \] where $\delta_{\text{\ref{case:11}}} \in \{0,1\}$ and takes the value $1$ only if we are in Case~\ref{case:11}, and where $\delta^{i=n}_{\text{\ref{case:1n}}} \in \{0,1\}$ and takes the value $1$ only if we are in Case~\ref{case:1n} and $i=n$. For the beads of type $1$, we have \[ \begin{array}{rcl} q^A_1- q^B_1 & = & \displaystyle{1 - \delta_{\text{\ref{case:11}}} + \sum_{k=1}^{n+1}\sum_{j\text{ odd}}(-1)^{k+1}\mu_1(s_j\cap t_k) + \sum_{k=1}^{n+1}\sum_{j\text{ even}}(-1)^{k+1}\mu_1(\bar s_j\cap t_k)} \smallskip\\ & = & \displaystyle{1 - \delta_{\text{\ref{case:11}}} + \sum_{k=1}^{n+1}(-1)^{k+1}\mu_1(t_k)}\smallskip \\ & = & 0 \enspace , \end{array} \] where the last equality is a consequence of Equation~\eqref{eqi} in Case~\ref{case:11} and of Equation~\eqref{1teq1} in Case~\ref{case:1n}. For the beads of type $i \neq 1$, we have \[ \begin{array}{rcl} q^A_i- q^B_i & = & \displaystyle{- \delta^{i = n}_{\text{\ref{case:1n}}} + \sum_{k=1}^{n+1}\sum_{j\text{ odd}}(-1)^{k+1}\mu_i(s_j\cap t_k) + \sum_{k=1}^{n+1}\sum_{j\text{ even}}(-1)^{k+1}\mu_i(\bar s_j\cap t_k)} \smallskip\\ & = & \displaystyle{- \delta^{i = n}_{\text{\ref{case:1n}}}+ \sum_{k=1}^{n+1}(-1)^{k+1}\mu_i(t_k)}\smallskip \\ & = & 0 \enspace , \end{array} \] where the last equality is a consequence of Equation~\eqref{eqi}, except when $i=n$ and we are in Case~\ref{case:1n}, where we use of Equation~\eqref{1teqn} instead. \end{proof} \section{Alternate combinatorial proofs of Proposition~\ref{prop} and Lemma~\ref{lem:decomposition}} \label{sec:combinatorial-proofs} All statements deal with discrete objects and properties. It is thus desirable from a logical point of view that all proofs stay in the ``discrete world'' if possible. Similarly to the traditional proof of the necklace splitting theorem, the proof of \Cref{prop} we gave in \Cref{sec:necklace} is relying on some ``continuous'' notions. P\'alv\H{o}lgyi~\cite{palvolgyi2009combinatorial} showed how such proofs are amenable to the discrete world by using a combinatorial counterpart of the Borsuk--Ulam theorem. We adapt here the approach proposed by P\'alv\H{o}lgyi to provide a purely combinatorial proof of \Cref{prop}, also relying on Tucker's lemma. Moreover, we show how the more general Ky Fan lemma can actually provide a direct combinatorial proof of \Cref{lem:decomposition} itself. \subsection{Combinatorial tools} We start by introducing some notation. Let \(\mathbb{B} = \{-1, 1\}\). For brevity, we may write \(+\) instead of \(+1\) and \(-\) instead of \(-1\). Let \(\mathbb{O} = \{-1,0, 1\}\). We define the partial order \(\prec\) on \(\mathbb{O}\) where \(0 \prec 1\) and \(0 \prec -1\), or, equivalently, \(b \preceq b'\) if \(b = 0\) or \(b = b'\). For every \(m \in \mathbb{N}\), we lift \(\prec\) to \(\mathbb{O}^m\) where for \(x, y \in \mathbb{O}^m\) we have \(x \preceq y\) if for all \(i \in [m]\), \(x(i) \preceq y'(i)\). Given a string \(x\), we denote by \(x(i)\) the \(i\)-th letter of \(x\). We also denote by $-x$ the string obtained from $x$ by replacing $1$'s by $-1$'s and $-1$'s by $1$'s (and the $0$'s are left unchanged). Now the Ky Fan lemma can be stated as follows: \begin{lemma}[Octahedral Ky Fan lemma, {\cite[Lemma 2]{CHEN20111062}}]\label{lemma:ky-fan} Let \(\lambda\colon \mathbb{O}^m \setminus 0^m \to \{\pm1, \ldots, \pm q\}\) such that \begin{enumerate}[label=\textup{(\roman*)}] \item\label{tucker:1} \(\lambda(-x) = -\lambda(x)\) for every \(x\). \item\label{tucker:2} \(\lambda(x) + \lambda(y) \neq 0\) for every \(x \preceq y\). \end{enumerate} Then there is at least one positively alternating \(m\)-chain, i.e., there are \(x_1, \ldots, x_m \in \mathbb{O}^m \setminus 0^m\) and \(j_1, \ldots, j_m \in [q]\) such that \(x_1 \preceq \cdots \preceq x_m\), \(1 \leq j_1 < j_2 < \cdots < j_m\), and \(\lambda(\{x_1, \ldots, x_m\}) = \{j_1, -j_2, j_3, \ldots, (-1)^{m-1} j_m\}\). In particular, \(q \geq m\). \end{lemma} (The statement we provide here is actually a special case of the original Ky Fan lem\-ma~\cite{fan52:_gener_tucker_combin_lemma_topol_applic} when the simplicial complex is the first barycentric subdivision of the octahedron.) The octahedral Tucker lemma~\cite{tucker_lemma_46} is actually the same statement without the existence of the chain, but still with the inequality $q \geq m$. We explain now how strings in $\mathbb{O}^m$ relate to decompositions of strings in $\Sigma_n^m$. We consider here all decompositions of a string $w \in \Sigma_n^m$ into two tuples $(u_1, u_3,\dots)$, $(u_2,u_4,\dots)$ of strings in $\Sigma_n^{+}$ such that $w = u_1u_2\cdots$. Such a decomposition is described as a string $x \in \mathbb{B}^m$ where each maximal segment of consecutive $-1$'s, as well as each maximal segment of consecutive $+1$'s, corresponds to one $u_j$. Each sign change in $x$ corresponds thus to a cut in $w$ and to a transition from a $u_j$ to $u_{j+1}$. A string $x \in \mathbb{O}^m$, containing some $0$'s, can be interpreted as an underspecified decomposition, which can be completed into different decompositions by replacing the $0$'s with $-1$'s or $+1$'s. \subsection{Combinatorial proof of Proposition~\ref{prop}} \label{sec:alt-proof-prop2} We prove \Cref{prop} in a way that is similar to~\cite{palvolgyi2009combinatorial}. For $x$ in $\mathbb{B}^\ast$ we let $\operatorname{alt}(x)$ be the number of sign alternations in $x$. We extend the function $\operatorname{alt}$ to $\mathbb{O}^\ast \setminus 0^\ast$ as follows: $\operatorname{alt}(x) = \max\{\operatorname{alt}(y) \mid y\in\mathbb{B}^\ast,\, x\preceq y\}$. Notice that: \begin{itemize} \item $\operatorname{alt}(-x) = \operatorname{alt}(x)$ for every $x$. \item $\operatorname{alt}(x)\geq \operatorname{alt}(y)$ for every $x\preceq y$. \end{itemize} For $x$ in $\mathbb{O}^\ast \setminus 0^\ast$, we define $\operatorname{sign}(x) \in\{-1,1\}$ as the first letter of some $y$ in $\mathbb{B}^\ast$ so that $x\preceq y$ and $\operatorname{alt}(x) = \operatorname{alt}(y)$. The function $\operatorname{sign}(x)$ is well defined as when $x$ is of the form $0^k \kappa x'$, with $\kappa \in \mathbb{B}$, the only way to replace the first $k$ zeroes so as to maximize $\operatorname{alt}$ consists in changing the sign at each position. From now on we fix the necklace $w$ and assume it is of length $m$ (i.e., $w$ is in $\Sigma_n^m$). We let $E_{\kappa,i}(x)$ for $x$ in $\mathbb{O}^m \setminus 0^m$ and $\kappa \in \{{-1},\,1\}$ be the amount of beads of type $i$ in $w$ that are aligned with the symbol $\kappa$ in $x$. We say that $x$ is $\kappa i$-\emph{unbalanced} if $E_{\kappa,i}(y)>E_{-\kappa,i}(y)$ for every $y$ such that $x\preceq y$. We define $\unb(x)$ to be $\kappa i$ where $i$ is the smallest $i$ so that $x$ is $\kappa i$-unbalanced. When no such $i$ exists, we let $\unb(x) =0$. We have \begin{itemize} \item $\unb(-x) = -\unb(x) $ for every $x$. \item $|{\unb(x)}| \geq |{\unb(y)}| > 0$ for every $x \preceq y$ if $\unb(x) \neq 0$. \end{itemize} Notice that $|{\unb(x)}| = |{\unb(y)}|$ implies $\unb(x) =\unb(y)$ for every $x\preceq y$. Finally we define $\lambda$ as the function from $\mathbb{O}^m$ to $\{-m+1,\dots, 0,\dots,m-1\}$: \[ \lambda (x) = \left\{ \begin{array}{ll} \operatorname{sign}(x)\operatorname{alt}(x)&\text{ when } \operatorname{alt}(x) > n\enspace .\\ \unb(x) & \text{ when } \operatorname{alt}(x) \leq n \enspace . \end{array} \right . \] Because of the properties of $\operatorname{alt}$, $\operatorname{sign}$ and $\unb$, had it no zero, the function $\lambda$ would satisfy the properties \ref{tucker:1} and \ref{tucker:2} of the octahedral Tucker lemma. However, then the conclusion is not satisfied, thus $\lambda$ must have a zero. Take $x$ so that $\lambda(x) = 0$. We must have that $\operatorname{alt}(x)\leq n$. As $\unb(x)=0$, it is possible, for each type $i$, to replace $0$'s in $x$ with $1$'s or $-1$'s so that we eventually obtain $y$ verifying $E_{1,i}(y) = E_{-1,i}(y)$ for every $i$, and such that the remaining $0$'s in $y$ are aligned with at most one unassigned bead of type $i$. We can then choose to replace these $0$'s with either $1$ or $-1$ depending on how we wish to treat the extra bead of a given type. As $\operatorname{alt}(x)$ is smaller than $n$, then no matter how we have conducted the previous changes, we have obtained a way to split the necklace with at most $n$ cuts so that each part contains the same amount of each type, with the extra beads (when the amount is odd) being distributed in any possible way. \begin{remark} The definition of $\unb$ is where the proof departs from the one of~\cite{palvolgyi2009combinatorial}. Indeed P\'alv\H{o}lgyi's proof only considers positive beads and it is then enough to consider that $x$ is unbalanced when one of the thief receives more than half of the beads of some type. In that case the split remains unbalanced for every $y$ so that $x\preceq y$. This property is no longer true with negative beads. Our definition of $\unb$ is taking this into account by imposing that being unbalanced is closed under $\preceq$. \end{remark} \subsection{Combinatorial proof of Lemma~\ref{lem:decomposition}} Lastly, we present a direct combinatorial proof of \Cref{lem:decomposition} that avoids \Cref{thm:multi-neckl}. The method is again inspired by P\'alv\H{o}lgyi~\cite{palvolgyi2009combinatorial}, but instead of constructing a function \(\lambda\) that contradicts the Tucker lemma, if it has no zero, we construct a pair of functions \(\lambda_+\) and \(\lambda_-\), for which the chains guaranteed by the Ky Fan lemma cannot exist for irreducible tuples. In contrast to \Cref{sec:alt-proof-prop2}, we do not consider \(x \in \mathbb{B}^m\) to represent the decomposition of a single string but instead $x$ shall represent the decomposition of an irreducible tuple $(s_1,\dots, s_n)$ in $(\Sigma_n^{+})^n$ where the word $s_1 \cdots s_n$ is in $O_n$ and each \(s_j\) is of length at least \(2\). To account for the arity of the tuple, we need to generalize the notion of \(\mathit{alt}\). Denote the word \(s_1 \cdots s_n\) by \(s\) and the length of \(s\) by \(m\). \begin{figure}[b] \centering \begin{tikzpicture} \matrix[matrix of math nodes, ampersand replacement=\&, every node/.append style={font=\strut}] (m) { \& a_1 \& a_2 \& \bar{a}_2 \& a_3 \& \bar{a}_2 \& \bar{a}_3 \& a_1 \& \bar{a}_1 \& \bar{a_1} \& \bar{a}_2 \\ x = \& + \& - \& - \& - \& - \& - \& - \& + \& - \& - \\ \mathit{alt}(x) = \& \& \& \& \& \& \& \& \& \& \& = 5 \\ ( \& u_1 \& \& u_2 \& \& \& u_4 \& ~ \& u_5 \& u_6 \& \& ) \\ }; \begin{scope}[b/.style={decorate,decoration={brace,amplitude=10pt}, shorten >= 4pt, shorten <= 4pt}] \draw[b] (m-1-2.north west) to node[above=8pt] {$s_1$} (m-1-5.north east); \draw[b] (m-1-6.north west) to node[above=8pt] {$s_2$} (m-1-8.north east); \draw[b] (m-1-9.north west) to node[above=8pt] {$s_3$} (m-1-11.north east); \end{scope} \draw[-latex] (m-2-2.south) to[bend right=40] node[below] {+1} (m-2-3.south); \draw[-latex] (m-2-5.south) to[bend right=40] node[below] {+2} (m-2-6.south); \draw[-latex] (m-2-8.south) to[bend right=40] node[below] {+1} (m-2-9.south); \draw[-latex] (m-2-9.south) to[bend right=40] node[below] {+1} (m-2-10.south); \node[anchor=base] at ($(m-4-4.base)!.4!(m-4-7.base)$) {$u_3,$}; \node[anchor=base] at ($(m-4-8.base)!.5!(m-4-9.base)$) {$,$}; \begin{scope}[on background layer, plus/.style={pattern=north east lines, draw=gray!30, pattern color=gray!30}, minus/.style={pattern=north west lines, draw=gray!30, pattern color=gray!30}] \draw[plus, ] (m-4-2.120) -- (m-1-2.south west) -- (m-1-2.south east) -- (m-4-2.60) --cycle; \draw[minus,] (m-4-4.120) -- (m-1-3.south west) -- (m-1-5.south east) -- (m-4-4.60) --cycle; \draw[minus,] (m-4-7.120) -- (m-1-6.south west) -- (m-1-8.south east) -- (m-4-7.60) --cycle; \draw[plus, ] (m-4-9.120) -- (m-1-9.south west) -- (m-1-9.south east) -- (m-4-9.60) --cycle; \draw[minus,] (m-4-10.120) -- (m-1-10.south west) -- (m-1-11.south east) -- (m-4-10.north east) --cycle; \end{scope} \end{tikzpicture} \caption{A word \(s_1s_2s_3\), a decomposition \(x\) with the corresponding alignment of \(s_j\)'s to \(u_p\)'s (in particular, \(u_3 = \varepsilon\)), and the calculation of \(\mathit{alt}(x)\).} \label{fig:maltXinterlace} \end{figure} Factorize \(x\) such that \(x = \kappa^{\ell_1}_1 \kappa^{\ell_2}_2 \cdots \kappa^{\ell_k}_k\) with \(\kappa_1, \ldots, \kappa_k \in \mathbb{B}\), \(\kappa_p = - \kappa_{p+1}\) and \(\ell_p \in \mathbb{N}\) be such that for each \(j \in [n]\), there is \(q_j \in \{0, \ldots, k\}\) with \(\sum_{p \in [j]} |s_p| = \sum_{p \in [q_j]} \ell_p\), and \(k\) is minimal. Now, if \(k = 2n\), then \(x\) describes a decomposition of $(s_1, \ldots, s_n)$ in two tuples each of size \(n\). Precisely, these tuples are $(u_1, u_3, \ldots, u_{2n-1})$ and $(u_2, u_4, \ldots, u_{2n})$ where the length of $u_j$ is $\ell_j$ for each $j \in [2n]$ and \(s_j = u_{q_{i-1} + 1} u_{q_{i-1} + 2} \cdots u_{q_j}\) for each \(j \in [n]\). If \(k < 2n\), we may add \(\kappa_q\)'s with \(\ell_q = 0\) to increase \(k\) to \(2n\) and proceed similarly. We call the value \(k-1\) the number of \emph{sign alternations of \(x\)} and write \(\mathit{alt}(x) = k - 1\). Hence, in the following we will search for \(x\) with \(\mathit{alt}(x) \leq 2n - 1\). \iffalse Formally \(\mathit{alt} \colon \mathbb{B}^m \to [m + n - 2]\) where \begin{align*} \mathit{alt}(x) = \sum_{i=1}^{m-1} \mathit{alt}'(x,i) \qquad \text{where } \mathit{alt}'(x,i) = \begin{cases} 1 + \delta^{x(i)}_{x(i+1)} & \text{if } \exists k\colon i = \sum_{j=1}^k |s_j| \\ 1 - \delta^{x(i)}_{x(i+1)} & \text{otherwise} \end{cases}\enspace. \end{align*} \fi Crucially, we assume mandatory sign alternations between the \emph{neighboring endpoints} of \(s_j\) and \(s_{j+1}\): if the last position of \(s_j\) and the first position of \(s_{j+1}\) are signed equally in \(x\), then some \(\kappa_p\) with \(\ell_p = 0\) occurs in our factorization of \(x\). In this case $2$ sign alternations are accounted for the transition between neighboring endpoints. In order to maximize \(\mathit{alt}\), we may choose \(x\) to be strictly alternating except for the \(n-1\) transitions between neighboring endpoints. In consequence \(\mathit{alt}(x) \leq m - (n-1) + 2(n-1) - 1 = m + n - 2\). \Cref{fig:maltXinterlace} gives an example: Note that \(\mathit{alt}(x) < 2\cdot3\) and that the subwords in \(+\)-labeled and \(-\)-labeled positions are both in \(O_3\). Thus, \(x\) describes the decomposition of \( (a_1 a_2 \bar{a}_2 a_3, \bar{a}_2 \bar{a}_3 a_1 , \bar{a}_1 \bar{a_1} \bar{a}_2) \) to two strictly smaller \(3\)-tuples \((a_1, \varepsilon, \bar{a}_1)\) and \((a_2 \bar{a}_2 a_3, \bar{a}_2 \bar{a}_3 a_1 , \bar{a_1} \bar{a}_2 )\), for each of which the word obtained by concatenating the tuple's components is in \(O_3\). Again, we consider words in \(\mathbb{O}^m \setminus 0^m\) which may have \emph{unsigned} positions, i.e., positions labeled with \(0\). In this case the decomposition is only partially determined. We define the function \(h\colon \mathbb{O}^m \setminus 0^m \to [m + n - 2]\) by \(h(x) = \max_{y \in \mathbb{B}^m \colon x \preceq y} \mathit{alt}(y)\). Observe that \(\mathit{alt}(x) = \mathit{alt}(-x)\), hence, for each \(x \in \mathbb{O}^m \setminus 0^m\): \(h(x) = h(-x)\). Also, for each \(x, y \in \mathbb{O}^m \setminus 0^m\) with \(x \preceq y\), we have \(h(x) \geq h(y)\). Denote by \(H(x)\) the set \(\arg\max_{y \in \mathbb{B}^m\colon x \preceq y} \mathit{alt}(y)\). All elements of \(H(x)\) may be obtained with the following algorithm: \begin{enumerate}[label = (\roman*)] \item Choose a signed position \(p\) in \(x\) with an unsigned neighbor \(p' \in \{p-1, p+1\}\).\label{step:1}% \item If \(p\) and \(p'\) are neighboring endpoints (in this case we call \(p\) and \(p'\) \emph{internal endpoints}), set \(x'(p') = x(p)\); otherwise set \(x'(p') = - x(p)\). For each \(q \neq p'\), set \(x'(q) = x(q)\). \item If \(x' \in \mathbb{B}^m\) we are done, otherwise, recursively apply the algorithm to \(x'\). \end{enumerate} This algorithm may yield words \(y\), \(y'\) where \(y(p) \neq y'(p)\) if \(p\) is an unsigned position between two signed positions of \(x\). However, for all position \(p'\) smaller than the smallest signed position of \(x\), the signs \(y(p')\) and \(y'(p')\) are equal for all words \(y,y'\) in \(H(x)\). Hence, we may define a function \(\operatorname{sign}\colon \mathbb{O}^m \setminus 0^m \to \mathbb{B}\) to assign to \(x\) the first symbol of some \(y \in H(x)\). Moreover, for each \(x \in \mathbb{O}^m \setminus 0^m\): \(\operatorname{sign}(x) = -\operatorname{sign}(-x)\). To prove \Cref{lem:decomposition}, we need to show that there is \(x \in \mathbb{B}^m\) with \(\mathit{alt}(x) < 2n\) and that both words \(u_1u_3\cdots\) and \(u_2u_4\cdots\) described by \(x\) are in \(O_n\) and non-empty. To assure non-emptiness, we just need that both signs occur in \(x\). To ensure that the words are in \(O_n\), we rephrase the notions \(E_{\kappa, i}\) and \(\unb\) from \Cref{sec:alt-proof-prop2} in terms of tuples of words instead of necklaces. Now \(E_{\kappa, i}(x)\) is formulated as the difference of the amount of \(a_i\)'s and \(\bar{a}_i\)'s aligned with \(\kappa\) in \(x\). Formally: \[E_{\kappa,i}(x) = |\{p \mid s(p) = a_i \land x(p) = \kappa\}|-|\{p \mid s(p) = \bar{a}_i \land x(p) = \kappa\}|\enspace.\] Observe that \(E_{+,i}(x) + E_{-,i}(x) = 0\) for \(s \in O_n\) and \(x \in \mathbb{B}^m\). We define two functions \(\lambda_+\) and \(\lambda_-\), where for each \(b \in \mathbb{B}\), \(\lambda_b\colon \mathbb{O}^m \setminus 0^m \to \{\pm1, \ldots, \pm m\} \cup \{0\}\) is such that \newsavebox{\mycases \tagsleft@false\let\veqno\@@eqno \begin{alignat}{2} \sbox{\mycases}{$\displaystyle \lambda_b(x) =\left\{% \begin{array}{@{}c@{}}% \vphantom{\operatorname{sign}(x) \cdot (h(x) - n + 2)\ \operatorname{sign}(x) \cdot (h(x) - n + 2) \quad \text{if } h(x) \geq 2n}\\% \vphantom{\unb(x) \quad \text{if } h(x) < 2 n \land |x|_{+} > 0 \land |x|_{-} > 0} \\ \vphantom{b \cdot (-1) \cdot b' \cdot (n + 1) \text{if } h(x) < 2n \land \exists b' \in \mathbb{B} \colon |x|_{b'} = 0 }\\ \vphantom{\cdot}% \end{array}% \right.\kern-\nulldelimiterspace$} \raisebox{-.5\ht\mycases}[0pt][0pt]{\usebox{\mycases}} &\operatorname{sign}(x) \cdot (h(x) - n + 2) && \quad \text{if } h(x) \geq 2n\enspace. \label{case:hg2n} \tag{C1}\\ &b \cdot (-1) \cdot b' \cdot (n + 1) && \quad \text{if } h(x) < 2n \land \exists b' \in \mathbb{B} \colon |x|_{b'} = 0 \enspace.\label{case:singlesigned} \tag{C2}\\ &{\unb(x)}&& \quad \text{if } h(x) < 2 n \land |x|_{+} > 0 \land |x|_{-} > 0 \enspace.\label{case:unbal} \tag{C3} \end{alignat} \tagsleft@true\let\veqno\@@leqno If there are \(x \in \mathbb{O}^m \setminus 0^m\) and \(b \in \mathbb{B}\) such that \(\lambda_b(x) = 0\), then \Cref{lem:decomposition} holds: Note that case \ref{case:unbal} of \(\lambda_b\) applies. Thus, for each \(i \in [n]\) there is \(y \in \mathbb{B}^m\) such that \(E_{+,i}(y) = E_{-,i}(y) = 0\). The positions relevant to balance \(a_i\) and \(a_j\) where \(i \neq j\) are distinct. Hence, we choose \(\hat{y} \in \mathbb{B}^m\) with \(x \preceq \hat{y}\) such that all symbols are balanced. Since \(\hat{y} \succeq x\) we have that \(\mathit{alt}(\hat{y}) \leq h(x) < 2n\), \(|\hat{y}|_{+} > 0\), and \(|\hat{y}|_- > 0\). Thus, \(\hat{y}\) encodes a decomposition with the desired properties. Otherwise, if \(\lambda_+\) and \(\lambda_-\) have no zero, we want to apply the Ky Fan lemma. Clearly, \(\lambda_b\) satisfies property \ref{tucker:1} of the octahedral Ky Fan lemma if cases \ref{case:hg2n} and \ref{case:unbal} apply. For case \ref{case:singlesigned}, note that \(|x|_{+} = 0 = |{-x}|_{-}\) implies \[-\lambda_b(x) = -(b \cdot (-1) \cdot (+1) \cdot (n+1)) = (b \cdot (-1) \cdot (-1)\cdot (n+1)) = \lambda_b(-x) \enspace. \] It remains to show property \ref{tucker:2}, i.e., that the sum \(\lambda_b(x) + \lambda_b(y)\) is not zero if \(x \preceq y\) and \(\lambda_b(x) \neq 0 \neq \lambda_b(y)\). Assume \(\lambda_b(x) + \lambda_b(y) = 0\). By the definition of \(\lambda_b\), there are three cases: \begin{enumerate} \item[\ref{case:hg2n}:] \(h(x) = h(y) \geq 2n\) and \(\operatorname{sign}(x) = -\operatorname{sign}(y)\). But then \(H(x) \supseteq H(y)\) and, thus, \(\operatorname{sign}(x) = \operatorname{sign}(y)\), a contradiction. \item[\ref{case:singlesigned}:] \(h(x), h(y) < 2n\) and either \(|x|_+ = 0 = |y|_-\) or \(|x|_- = 0 = |y|_+\). W.l.o.g.\@ assume that \(|x|_- = 0 = |y|_+\). Then there is \(k \in [m]\) such that \(x(k) = +\). But \(x \preceq y\) implies \(y(k) = +\), a contradiction to \(|y|_+ = 0\). \item[\ref{case:unbal}:] \(h(x), h(y) < 2n\), \(|x|_+, |y|_+ > 0\), \(|x|_-,|y|_- > 0\) and \(\unb(x) = -\unb(y)\). This contradicts properties of \(\unb\) shown in \Cref{sec:alt-proof-prop2}: for every \(x \preceq y\) with \(\unb(x) \neq 0\) we have $|{\unb(x)}| \geq |{\unb(y)}|$, where $|{\unb(x)}| = |{\unb(y)}|$ implies $\unb(x) =\unb(y)$. \end{enumerate} \begin{proposition} If \(\lambda_+\) and \(\lambda_-\) have no zero, then every pair of compatible symbols \(a_i, \bar{a}_i \in \Sigma_n\) can be reduced from \((s_1, \ldots, s_n)\).\label{lemma:no-zero-implies-twisting-or-interlacing} \end{proposition} \noindent \Cref{lem:decomposition} follows directly, as whenever \((s_1, \ldots, s_n)\) is irreducible, \(\lambda_+\) and \(\lambda_-\) have a zero. To show \Cref{lemma:no-zero-implies-twisting-or-interlacing} assume that for either \(b \in \mathbb{B}\) there is no \(x \in \mathbb{O}^m \setminus 0^m\) such that \(\lambda_b(x) = 0\). By \Cref{lemma:ky-fan}, for each \(b \in \mathbb{B}\), there is a positively alternating \(m\)-chain, i.e., there are \(x_1^b \preceq x_2^b \preceq \cdots \preceq x_m^b \) and \(1 \leq j_1^b < j_2^b < \cdots < j_m^b \leq m\) such that \(\lambda_b(\{x_1^b, \ldots, x_m^b\}) = \{(-1)^{k-1} j_k^b \mid k \in [m]\}\). We derive the following properties of the chains and \((s_1, \ldots, s_n)\): \begin{enumerate}[label=(\alph*), leftmargin=2.5em] \item \label{item:onto} The values \(j_1^b, \ldots, j_m^b\) are pairwise distinct and in \([m]\). Thus, \(j_k^b = k\) for each \(k \in [m]\). \item By the definition of \(\prec\), every strictly increasing chain in \(\mathbb{O}^m \setminus 0^m\) has at most length \(m\). Specifically, \(|x_1^b|_+ + |x_1^b|_{-} = 1\), \(|x_2^b|_+ + |x_2^b|_{-} = 2\), $\ldots$, and \(|x_m^b|_+ + |x_m^b|_{-} = m\). Also, for each \(k \in [m-1]\), \(x^b_k\) and \(x^b_{k+1}\) differ in exactly one position \(p\), i.e., \(x^b_k(p) = 0\) and \(x^b_{k+1}(p) \in \mathbb{B}\) and for each \(p' \neq p\): \(x^b_k(p') = x^b_{k+1}(p')\). \item \label{item:signedness:1} By \ref{item:onto}, all cases (\ref{case:hg2n}, \ref{case:singlesigned}, \ref{case:unbal}) of \(\lambda_b\) apply. Because \(h\) is antitone, \ref{case:hg2n} applies to \(x_1^b, \ldots, x_{m-n-1}^b\) where \(h(x^b_{k}) > h(x^b_{k+1})\) for each \(k \in [m-n-1]\). On the other hand, for \(k \in \{0, 1, \ldots, n\}\), we have that \(h(x^b_{m-n+k}) < 2n\). Case \ref{case:singlesigned} applies to exactly one of the words \(x^b_{m-n}, \ldots, x^b_{m}\) while to the remaining words case \ref{case:unbal} applies. For \ref{case:singlesigned} to apply, only one sign $\top_b \in \mathbb{B}$ may occur in the word, while \ref{case:unbal} requires that both signs occur. Hence \ref{case:singlesigned} applies to \(x^b_{m-n}\) as the words, to which \ref{case:unbal} apply, need to be larger with respect to \(\preceq\). Moreover, as \(x^b_{m-n}\) is larger than \(x_1^b, \ldots, x^b_{m-n-1}\), the latter are in \(\{\top_b,0\}^m\), too. \(\top_b\) is determined by the sign of \(\lambda_b(x^b_{m-n}) = (-1)^n \cdot (n+1)\), i.e., it depends on \(b\) and the parity of \(n\). Precisely: \(\top_b = b \cdot (-1)^n\). \item \label{item:signedness:2} \ref{case:unbal} applies to \(x^b_{m-n+k}\) for each \(k \in [n]\) and yields values \(\{1, -2, 3, \ldots, (-1)^{n-1}n\}\). As \(\unb(x^b_{m-n+k}) \neq 0\), we have that \({|{\unb(x^b_{m-n+k})}|} > {|{\unb(x^b_{m-n+k+1})}|}\) for all \(k \in [n-1]\). Thus, \(\lambda(x^b_{m-n+k}) = (-1)^{n-k}\cdot(n-k+1)\). Denote \(n-k+1\) by \(i\). Observe that \(\unb(x^b_{m-n}) = 0\) because the word \({\top_b}^m \succeq x^b_{m-n}\) trivially balances every symbol. Hence, to ultimately unbalance $a_{i}$, a position \(p^b_{i}\) is newly signed in \(x^b_{m-n+k}\) with \(-\top_b\) where \(s(p^b_{i})\) is either \(a_{i}\) or \(\bar{a}_{i}\). The exact choice of $p^b_{i}$ depends on \(\top_b\) and the parity of \(i\): If \((-1)^{i-1} = -\top_b\), then \(p^b_i\) is such that \(s(p^b_{i}) = a_{i}\), otherwise, \(p^b_i\) is such that \(s(p^b_{i}) = \bar{a}_{i}\). \end{enumerate} \iffalse \begin{figure}[t] \resizebox{\linewidth}{!}{ \begin{tikzpicture} \matrix[matrix of math nodes, ampersand replacement=\&, every node/.append style={font=\strut, inner sep = 0.15em}, column sep = .08em] (m) { w \& a_1 \& \bar{a}_3 \& a_2 \& \bar{a}_2 \& a_3 \& a_3 \& \bar{a}_3 \& \bar{a}_4 \& \bar{a}_4\& a_4 \& a_4 \& \bar{a}_1 \& \operatorname{sign}(\cdot) \& h(\cdot)\& \lambda_+(\cdot) \\ x^+_1 \& \& \& \& \& \& \& \& \& \& \& + \& \& - \& 14 \& -12 \\ x^+_2 \& \& \& \& \& \& \& \& \& + \& \& + \& \& + \& 13 \& 11 \\ x^+_3 \& \& \& \& \& \& \& \& + \& + \& \& + \& \& - \& 12 \& -10 \\ x^+_4 \& \& \& \& \& \& + \& \& + \& + \& \& + \& \& + \& 11 \& 9 \\ x^+_5 \& \& \& \& \& + \& + \& \& + \& + \& \& + \& \& - \& 10 \& -8 \\ x^+_6 \& \& \& \& + \& + \& + \& \& + \& + \& \& + \& \& + \& 9 \& +7 \\ x^+_7 \& \& + \& \& + \& + \& + \& \& + \& + \& \& + \& \& - \& 8 \& -6 \\ x^+_8 \& + \& + \& \& + \& + \& + \& \& + \& + \& \& + \& \& \& 7 \& +5 \\ x^+_9 \& + \& + \& \& + \& + \& + \& \& + \& + \& - \& + \& \& \& 7 \& -4 \\ x^+_{10} \& + \& + \& \& + \& + \& + \& - \& + \& + \& - \& + \& \& \& 7 \& +3 \\ x^+_{11} \& + \& + \& - \& + \& + \& + \& - \& + \& + \& - \& + \& \& \& 7 \& -2 \\ x^+_{12} \& + \& + \& - \& + \& + \& + \& - \& + \& + \& - \& + \& - \& \& 7 \& +1 \\ }; \begin{scope}[b/.style={decorate,decoration={brace,amplitude=10pt}, shorten >= 4pt, shorten <= 4pt}] \draw[b] (m-1-2.north west) to node[above=8pt] {$s_1$} (m-1-4.north east); \draw[b] (m-1-5.north west) to node[above=8pt] {$s_2$} (m-1-7.north east); \draw[b] (m-1-8.north west) to node[above=8pt] {$s_3$} (m-1-10.north east); \draw[b] (m-1-11.north west) to node[above=8pt] {$s_4$} (m-1-13.north east); \end{scope} \draw (m-1-1.south west) to (m-1-16.south east); \draw (m-13-1.east|-m-1-1.north east) to (m-13-1.south east); \draw (m-1-13.north east) to (m-1-13.east|-m-13-13.south east); \end{tikzpicture} \qquad \begin{tikzpicture} \matrix[matrix of math nodes, ampersand replacement=\&, every node/.append style={font=\strut, inner sep = 0.15em}, column sep = .08em] (m) { w \& a_1 \& \bar{a}_3\& a_2 \& \bar{a}_2 \& a_3 \& a_3\& \bar{a}_3 \& \bar{a}_3\& \bar{a}_4\& a_4 \& a_4 \& \bar{a}_1 \& \operatorname{sign}(\cdot) \& h(\cdot)\& \lambda_-(\cdot) \\ x^-_1 \& \& \& \& \& \& \& \& \& \& \& \& - \& - \& 14 \& -8 \\ x^-_2 \& \& \& \& \& \& \& \& \& \& \& - \& - \& + \& 13 \& -8 \\ x^-_3 \& \& \& \& \& \& \& \& \& \& - \& - \& - \& - \& 12 \& -8 \\ x^-_4 \& \& \& \& \& \& \& \& - \& \& - \& - \& - \& + \& 11 \& -8 \\ x^-_5 \& \& \& \& \& \& \& - \& - \& \& - \& - \& - \& - \& 10 \& -8 \\ x^-_6 \& \& \& \& \& - \& \& - \& - \& \& - \& - \& - \& + \& 9 \& +7 \\ x^-_7 \& \& \& - \& \& - \& \& - \& - \& \& - \& - \& - \& - \& 8 \& -6 \\ x^-_8 \& \& - \& - \& \& - \& \& - \& - \& \& - \& - \& - \& \& 7 \& +5 \\ x^-_9 \& \& - \& - \& \& - \& \& - \& - \& + \& - \& - \& - \& \& 7 \& -4 \\ x^-_{10}\& \& - \& - \& \& - \& + \& - \& - \& + \& - \& - \& - \& \& 7 \& +3 \\ x^-_{11}\& \& - \& - \& + \& - \& + \& - \& - \& + \& - \& - \& - \& \& 7 \& -2 \\ x^-_{12}\& + \& - \& - \& + \& - \& + \& - \& - \& + \& - \& - \& - \& \& 7 \& +1 \\ }; \begin{scope}[b/.style={decorate,decoration={brace,amplitude=10pt}, shorten >= 4pt, shorten <= 4pt}] \draw[b] (m-1-2.north west) to node[above=8pt] {$s_1$} (m-1-4.north east); \draw[b] (m-1-5.north west) to node[above=8pt] {$s_2$} (m-1-7.north east); \draw[b] (m-1-8.north west) to node[above=8pt] {$s_3$} (m-1-10.north east); \draw[b] (m-1-11.north west) to node[above=8pt] {$s_4$} (m-1-13.north east); \end{scope} \draw (m-1-1.south west) to (m-1-16.south east); \draw (m-13-1.south east) to (m-13-1.south east|-m-1-1.north east); \draw (m-1-13.north east) to (m-1-13.east|-m-13-13.south east); \end{tikzpicture} } \caption{Positively alternating \(m\)-chains for \(\lambda_+\) and \(\lambda_-\) where \(n = 4\), \(m=12\), and \((s_1,s_2,s_3) = (a_1 \bar{a}_3 a_2, \bar{a}_2 a_3 a_3, \bar{a}_3 \bar{a}_4 \bar{a}_4, a_4 a_4 \bar{a}_1)\). \(0\)-s in \(x^b_i\) were omitted.} \label{fig:alternating-chains} \end{figure} \fi Let \(U = \{p \in [m] \mid \exists b \in \mathbb{B}\colon x^b_{m-n}(p) = 0\} = \{p^b_i \mid b \in \mathbb{B}, i \in [n]\}\). From \[\top_+= (+1) \cdot (-1)^n = (-1) \cdot (-1) \cdot (-1)^n = (-1) \cdot \top_- \] and \ref{item:signedness:2} it follows that \(p^+_i \neq p^-_i\) and \(a_i, \bar{a}_i \in \{s(p) \mid p \in U\}\) for each \(i \in [n]\). Hence, \(|U| = 2n\) and \(\Sigma_n \subseteq \{s(p) \mid p \in U\}\). To complete our proof, we further characterize the set \(U\). \begin{property}\label{lemma:at-most-one-position-signed-in-boundary-cluster} Let \(p\) and \(p+1\) be neighboring endpoints. At most one of \(p\) and \(p+1\) is signed in \(x^b_{m-n}\). \end{property} \begin{proof} Assume the contrary. Let \(k \in [m-n]\) be minimal such that \(x^b_{k}(p) = \top_b = x^b_{k}(p+1)\), i.e., either \(x^b_{k-1}(p) = 0\) or \(x^b_{k-1}(p+1) = 0\). W.l.o.g.\@ assume that \(x^b_{k-1}(p+1) = 0\). Hence, using our algorithm to construct an element of \(H(x^b_{k-1})\), we may set \(p+1\) to \(\top_b\), i.e., we obtain \(x^b_k\). But then \(h(x^b_k) = h(x^b_{k-1})\), contradicting \ref{item:signedness:1}. \end{proof} Let \(p\) and \(p+1\) be neighboring endpoints. We denote \(C(p) = C(p+1) = \{p, p+1\}\). For each position \(p' \in [m]\) that is not an internal endpoint, we denote \(C(p') = \{p'\}\). \begin{property}\label{lemma:unsigned-boundary-cluster-not-embraced-by-signs} Let \(1 < p < m\) be an unsigned position in \(x^b_{m-n}\) that is not an internal endpoint. Then \(C(p-1)\) or \(C(p+1)\) contain only unsigned positions in \(x^b_{m-n}\). \end{property} \begin{proof} Assume the contrary. Let \(k \in [m-n]\) be minimal such that both \(C(p-1)\) and \( C(p+1)\) contain a signed position in \(x^b_{k}\). W.l.o.g.\@ assume that \(C(p-1)\) contains a signed position already in \(x^b_{k-1}\). Using the algorithm to construct an element of \(H(x^b_{k-1})\), we may first set the other position in \(C(p-1)\) to \(\top_b\) (if it exists), then set \(p\) to \(-\top_b\), then set each position of \(C(p+1)\) to \(\top_b\) and call the intermediate result \(y\). We may also obtain \(y\) from \(x^{b}_{k}\) by first setting each unsigned position of \(C(p+1)\) to \(\top_b\), then \(p\) to \(-\top_b\), and then the remaining position of \(C(p-1)\) to \(\top_b\), if it exists. Thus, \(h(x^b_{k}) = h(x^b_{k-1})\), contradicting \ref{item:signedness:1}. \end{proof} There are \(n-1\) pairs of neighboring endpoints, i.e., in total \(2n-2\) internal endpoints. \(x^+_{m-n}\) and \(x^-_{m-n}\) each have \(n\) unsigned positions, of which, by \Cref{lemma:at-most-one-position-signed-in-boundary-cluster}, at least \(n-1\) are internal endpoints. \Cref{lemma:unsigned-boundary-cluster-not-embraced-by-signs} implies that the only unsigned position that is not an internal endpoint, if one exists, is \(1\) or \(m\). Thus, \(U\) consists solely of endpoints of the \(s_j\)'s. % However, all symbols of \(\Sigma_n\) are distributed at the positions in \(U\). Hence, \((s_1, \ldots, s_n)\) is such that every pair of compatible letters of \(\Sigma_n\) can be reduced. This concludes the proof of \Cref{lemma:no-zero-implies-twisting-or-interlacing}.

2395b02afb5683e253d77df7b958070e3be024f0

\section{\label{sec:Indroduction}Introduction} High Energy Physics (HEP) communities have a long history of working with large data and applying advanced statistics techniques to analyze experimental data in the energy, intensity, and cosmic frontiers. With ever-increasing data volumes, the HEP community needs a significant computational breakthrough to continue this trajectory, and tools developed in Quantum Information Science (QIS) could provide a viable solution. Quantum advantage is the potential to solve problems faster than any classical methods \cite{arute2019quantum, harrow2017quantum}. In computational-complexity-theoretic terms, this generally means providing a superpolynomial speedup over the best known or possible classical algorithm \cite{nielsen2002quantum}. Machine learning methods promise great benefits for scalable data analytics. The big wave of deep learning algorithm development stems from recent advances in convolutional neural networks (CNNs) \cite{lecun1998gradient}, which can effectively capture spatial dependencies within an image, as well as automatically learn important features from them \cite{krizhevsky2012imagenet, szegedy2015going, simonyan2014very, lecun2015deep, goodfellow2016deep}. Along with big data and graphics processing unit (GPU) processing capabilities, deep learning has significantly improved the ability to analyze large volumes of images. There are several examples where CNNs have been successfully applied to HEP challenges using classical computers~\cite{hep-ml}. However, in quantum computing, no significant progress has been made toward implementing such robust representation learning methods to date. In this work, we present a new hybrid Quantum Convolutional Neural Network (QCNN) framework to demonstrate the quantum advantage versus corresponding classical algorithms. We simulate its performance for classification of HEP events from the simulated data in neutrino experiments. We show that with a similar number of parameters in QCNN and classical CNNs, the QCNN can learn faster or reach better testing accuracy with fewer training epochs. Thus, our simulations empirically demonstrate the quantum advantage of QCNN over CNN in terms of testing accuracy. This paper is organized as follows: Section~\ref{sec:hep-data} introduces the HEP experimental data used in this work. In Section~\ref{sec:vqc} and \ref{sec:qcnn}, we describe the new QCNN architecture in detail. Section~\ref{sec:results} shows the performance of the QCNN on the experimental data. Section~\ref{sec:conclusion} concludes the overall discussion. \section{High Energy Physics Data} \label{sec:hep-data} In this work, we use simulated data from the Deep Underground Neutrino Experiment (DUNE)~\cite{dune} to develop and test our QCNN algorithms on high-energy experiments. Hosted in the United States, DUNE is the next-generation, international, world-class experiment~to reveal new symmetries of nature. DUNE's primary goals include searching for CP violation in the lepton sector, determining neutrino mass ordering, performing precision tests of the three-neutrino paradigm, detecting supernova neutrino bursts, and searching for nucleon decays beyond the Standard Model. DUNE is an excellent test case for the QCNN because its main detector technology, the Liquid Argon Time Projection Chamber (LArTPC), effectively provides high-resolution images of particle activities as the ionized electrons drift toward the multiple sensing wire planes~\cite{rubbia77,Chen:1976pp,willis74,Nygren:1976fe}. An advanced LArTPC simulation package, the Wire-Cell Toolkit~\cite{wct} and LArSoft software~\cite{larsoft}, is used to generate realistic single-particle images in a LArTPC detector. The simulation implements a chain of algorithms, including: 1) generating single-particle kinematics, 2) applying LArTPC detector response, 3) adding realistic electronic noise, and 4) performing digital signal processing. Details about the LArTPC simulation can be found in Ref~\cite{lartpc-sp}. Four different types of particles ($\mu^+$, $e^-$, $\pi^+$, and $p$) are simulated. \figureautorefname{\ref{simu_particle}} shows sample images of simulated particle activities on the collection wire plane. The images have a resolution of 480 x 600 pixels, where each pixel in the x-axis represents a single wire and each pixel in the y-axis represents a sampling time tick. In this work, the goal of the QCNN is to predict the types of different particles by analogy with those performed via the classical CNN~\cite{lartpc-cnn}. Each particle's momentum is set such that the mean range of the particle is about 2 meters, so the classification is not sensitive to the image size. \begin{figure}[htbp] \centering \includegraphics[width=0.85\linewidth]{figures/particles-4-white-bg.png} \caption{Example images of simulated particle activities ($\mu^+$, $e^-$, $\pi^+$, $p$) in a LArTPC detector. Colors in the images represent the sizes of the ionization energy loss along the particle trajectories when measured by LArTPC's wire planes.} \label{simu_particle} \end{figure} As visualized in~\figureautorefname{\ref{simu_particle}}, the classification of the four different particles primarily is a pattern recognition problem. A positively charged muon ($\mu^+$) is a track-like particle, while an electron ($e^-$) produces electromagnetic showers that are spatially extended. A muon is a minimum ionizing particle in terms of energy loss along its trajectory, which translates into the intensity of the pixels. It experiences Multiple Coulomb Scattering (MCS) when passing the detector, causing its trajectory to deviate from a straight line. It also decays into a low-energy positron after it loses most of the kinetic energy and stops in the detector, leading to another short track segment near the end of the main track. A positively charged pion ($\pi^+$) looks similar to a muon in terms of energy loss, MCS, and decay, but it experiences additional nuclear interactions during its passage in the detector, often leading to a hard scattering (represented as a ``kink'') along its main trajectory. Finally, a proton ($p$) also is a track-like particle. However, because a proton's mass is much heavier than a muon or pion, it has higher energy loss and encounters less MCS during travel. Consequently, a proton's track has higher intensity and is straighter than those from muons and pions. These diverse features in detector images make the LArTPC data analysis well suited for CNN-type machine learning algorithms rather than hand-crafted feature extraction methods. Previous work with LArTPC has shown excellent performance from single-particle classification~\cite{lartpc-cnn} to the more complicated neutrino interaction classifications~\cite{dune-cnn}. In this work, we perform a quantum implementation of the classical CNN through the variational quantum circuits for the first time in LArTPC data analysis. By comparing the performance to the classical CNN, we explore possible quantum acceleration and advantage in machine learning for HEP data analysis. \section{Variational Quantum Circuits}\label{sec:vqc} Variational quantum circuits (VQC) are quantum circuits that have \emph{tunable} or \emph{adjustable} parameters subject to classical iterative optimizations, which are commonly based on gradient descent and its variants \cite{schuld2019evaluating, benedetti2019parameterized}. The general structure of VQC is presented in~\figureautorefname{\ref{Fig:GeneralVQC}}. Here, the $F(\mathbf{x})$ block is for the state preparation that encodes the classical data $\mathbf{x}$ into the quantum state for the circuit to operate on and is not subject to optimization. This state preparation part is designed according to the given research problem. The $V(\boldsymbol{\theta})$ block represents the variational or \emph{learning} part. The \emph{learnable} parameters labeled with $\boldsymbol{\theta}$ will be optimized through gradient-based methods. For example, the commonly used gradient-based optimizers are Adam \cite{kingma2014adam} and RMSProp \cite{Tieleman2012}. In concept, these parameters are comparable to the \emph{weights} in classical deep neural networks (DNNs). In the final part of this VQC block, we perform the quantum measurement on a subset (or all) of the qubits to retrieve the information. If we run the circuit once and perform a single quantum measurement, it will yield a bit string, such as $0010$, and it generally differs from what we will get if we prepare the circuit again and perform another quantum measurement due to the stochastic nature of quantum systems. However, if we prepare the same circuit and perform the quantum measurement several times, e.g., $1000$ times, we will get the expectation values on each qubit, which should be quite close to the results from theoretical calculation. For example, consider a two-qubit system $\ket{\Psi}$, in every single measurement, the result is one of the following: $00$, $01$, $10$, and $11$. If we prepare $\ket{\Psi}$ and measure it $1000$ times, we will get numerous $00, \cdots ,11$. We can count the frequencies of the appearance of $0$ and $1$ in each qubit and use them to estimate the expectation values. For example, after $1000$ times of repeated measurement, we get $600$ times of $0$ and $400$ times of $1$ in the first qubit. Therefore, the expectation value of the first qubit is $0.4$. In an $N$-qubit system, we place the expectation values of all qubits into a $N$-dimensional vector, which can be processed further in classical or quantum neural networks. We may choose different bases for the measurement. For example, in this work, we exclusively use the Pauli-$Z$ expectation values at the end of VQC. Although VQCs are simple in concept, they are successful in machine learning tasks. Recent studies have reported the application of such variational architectures in the field of classification \cite{mitarai2018quantum, schuld2018circuit, Farhi2018ClassificationProcessors, benedetti2019parameterized, mari2019transfer, abohashima2020classification, easom2020towards, sarma2019quantum, stein2020hybrid,chen2020hybrid}, function approximation \cite{chen2020quantum, mitarai2018quantum,kyriienko2020solving}, generative machine learning \cite{dallaire2018quantum, stein2020qugan, zoufal2019quantum, situ2018quantum,nakaji2020quantum}, metric learning \cite{lloyd2020quantum, nghiem2020unified}, deep reinforcement learning \cite{chen19, lockwood2020reinforcement, jerbi2019quantum, Chih-ChiehCHEN2020}, sequential learning \cite{chen2020quantum, bausch2020recurrent}, and speech recognition \cite{yang2020decentralizing}. \begin{figure}[htbp] \begin{center} \begin{minipage}{10cm} \Qcircuit @C=1em @R=1em { \lstick{\ket{0}} & \multigate{3}{F(\mathbf{x})} & \qw & \multigate{3}{V(\boldsymbol{\theta})} & \qw & \meter \qw \\ \lstick{\ket{0}} & \ghost{F(\mathbf{x})} & \qw & \ghost{V(\boldsymbol{\theta})} & \qw & \meter \qw \\ \lstick{\ket{0}} & \ghost{F(\mathbf{x})} & \qw & \ghost{V(\boldsymbol{\theta})} & \qw & \meter \qw \\ \lstick{\ket{0}} & \ghost{F(\mathbf{x})} & \qw & \ghost{V(\boldsymbol{\theta})} & \qw & \meter \qw \\ } \end{minipage} \end{center} \caption[General structure for the variational quantum circuit.]{{\bfseries General structure for the variational quantum circuit (VQC).} The $F(\mathbf{x})$ is the quantum operation for encoding the classical data into the quantum state and $V(\boldsymbol{\theta})$ is the variational quantum circuit block with the adjustable parameters $\boldsymbol{\theta}$. } \label{Fig:GeneralVQC} \end{figure} \section{Quantum CNN}\label{sec:qcnn} \emph{CNNs} have been tremendously successful in a wide spectrum of modern machine learning tasks, especially in the area of computer vision \cite{krizhevsky2012imagenet, szegedy2015going, simonyan2014very, lecun2015deep, lecun1998gradient, goodfellow2016deep}. Such methods also afford new insights and progress in scientific research, for example, in HEP event classification \cite{lartpc-cnn, aurisano2016convolutional} and phase transition studies \cite{tanaka2017detection}. With recent advances in quantum computing hardware \cite{preskill2018quantum, cross2018ibm, arute2019quantum}, it is interesting to study the potential advantages and application scenarios for CNNs in the quantum regime. \emph{QCNN} is the framework that uses VQCs to perform the convolutional operations. In this work, we replace the classical neural-network-based convolutional filters, or \emph{kernels}, with VQCs to harvest the expressive power granted by quantum entanglements. The quantum convolutional kernels will sweep through the input image pixels and transform them into a \emph{representation vector} of lower dimensions by performing measurements (see~\figureautorefname{\ref{QCNN_Operation}}). A stack of VQCs will ensure features of varied length scales are captured in different layers. \subsection{Quantum Convolutional Filters} This section describes the VQC building blocks for the QCNN architecture. \subsubsection{Data Encoding Layer} In this layer, we first \emph{encode} a classical input vector into a quantum state, which is necessary for additional processing. A general $N$-qubit quantum state can be represented as: \begin{equation} \label{eqn:quantum_state_vec} \ket{\psi} = \sum_{(q_1,q_2,...,q_N) \in \{ 0,1\}^N}^{} c_{q_1,...,q_N}\ket{q_1} \otimes \ket{q_2} \otimes \ket{q_3} \otimes ... \otimes \ket{q_N}, \end{equation} where $ c_{q_1,...,q_N} \in \mathbb{C}$ is the \emph{amplitude} of each quantum state and $q_i \in \{0,1\}$. The square of the amplitude $c_{q_1,...,q_N}$ is the \emph{probability} of measurement with the post-measurement state in $\ket{q_1} \otimes \ket{q_2} \otimes \ket{q_3} \otimes ... \otimes \ket{q_N}$, and the total probability should sum to $1$, i.e., \begin{equation} \label{eqn:quantum_state_vec_normalization_condition} \sum_{(q_1,q_2,...,q_N) \in \{ 0,1\}^N}^{} ||c_{q_1,...,q_N}||^2 = 1. \end{equation} In the proposed framework, the input to the VQC is a matrix with a dimension $n \times n$, where $n$ is the filter/kernel size. The input will first be flattened and transformed into \emph{rotation angles} for the quantum gates. In general, the input values of pixels are not in the interval of $[-1, 1]$. We use the arc tangent function to transform these input values into rotation angles. For each of the $x_i$ in the $n \times n$ input, there will be two rotation angles generated, $\arctan(x_i)$ and $\arctan(x_i^2)$. The ($n \times n$)-dimensional vector will then be transformed into $2\times n \times n$ angles for the single-qubit rotation. \subsubsection{Variational Layer} After encoding the classical values into a quantum state, it will be subject to a series of unitary transformations. The variational layer (grouped in a dashed-line box in~\figureautorefname{\ref{Fig:Basic_VQC}}) consists of two parts. One is the \emph{entanglement part}, which is a group of CNOT gates. The other is the \emph{rotation part} that includes several single-qubit unitary rotations parameterized by $3$ parameters $\alpha_i$, $\beta_i$, and $\gamma_i$, where $i$ represents the index of qubits. The parameters labeled by $\alpha_i$, $\beta_i$, and $\gamma_i$ are the ones that will be updated by the optimization procedure. \subsubsection{Quantum Measurement Layer} To obtain the transformed data from VQC blocks, we perform quantum measurements. We consider the ensemble samplings (expectation values) of the VQCs. If working with quantum simulation software (for example, PennyLane~\cite{bergholm2018pennylane} or IBM Qiskit), this value can be calculated deterministically. While implementing on a real quantum computer, it is required to prepare the same system and carry out the measurements repeatedly to gain enough statistics. In the proposed QCNN architecture, the quantum convolutional filter will output a single value for each sweep step. Here, we perform the quantum measurements on the first qubit to get the expectation value. \subsection{Quantum Convolutional Operations} Both of the classical and quantum convolutional operations follow the following rule: \begin{equation} W_{out}=\frac{\left(W_{in}-F+2 P\right)}{S}+1, \end{equation} where \begin{itemize} \item $W_{out}$ : the output dimension of the convolutional layer \item $W_{in}$ : the output dimension of the convolutional layer \item $F$ : filter size \item $P$ : padding size. \end{itemize} To capture the spatial dependency of the input data (e.g., images), the convolutional filter will sweep across the pixels and output the corresponding values on each location (see~\figureautorefname{\ref{QCNN_Concept}}). In the QCNN, the filter itself is a VQC, which will transform an $n \times n$ dimensional vector into a single value (see ~\figureautorefname{\ref{QCNN_Operation}}). The circuit component for the quantum convolutional filter is in the~\figureautorefname{\ref{Fig:Basic_VQC}}. In general, each of the quantum convolutional filters captures a single kind of feature. We may place several filters in a convolutional layer to extract multiple features. In addition, we can \emph{stack} multiple convolutional layers to extract different levels of features. \begin{figure}[htbp] \centering \includegraphics[width=1.0\linewidth]{figures/QCNN_Concept.pdf} \caption[Concepts: Quantum CNN.]{{\bfseries Quantum CNN Architecture.} In the proposed hybrid quantum-classical model, the \emph{filter} or \emph{kernel} is a variational quantum circuit as shown in \figureautorefname{\ref{Fig:Basic_VQC}}. Classical pooling and nonlinear activation functions can be optionally added between the convolutional layers analogous to the classical CNN.} \label{QCNN_Concept} \end{figure} \begin{figure}[htbp] \centering \includegraphics[width=1.0\linewidth]{figures/QCNN_FlowDiagram.pdf} \caption[QCNN Operation.]{{\bfseries QCNN Operation.} In the QCNN operation, the input pixel values $(x_1, x_2, x_3, x_4)$ will first be encoded into a quantum state via the variational encoding method. Each value $x_i$ is mapped into two values $\arctan(x_i)$ and $\arctan(x_i^2)$ for the $R_y$ and $R_z$ rotation angles, respectively. The quantum gates parameterized by $\alpha_i, \beta_i, \gamma_i$ then act on this encoded state. At the end of the circuit, Pauli-$Z$ expectation values are retrieved. The retrieved values can then be processed with another layer of quantum convolutional layer or other classical operations (e.g. pooling, nonlinear activation functions or dropout). } \label{QCNN_Operation} \end{figure} \begin{figure}[htbp] \begin{center} \begin{minipage}{10cm} \Qcircuit @C=1em @R=1em { \lstick{\ket{0}} & \gate{R_y(\arctan(x_1))} & \gate{R_z(\arctan(x_1^2))} & \ctrl{1} & \qw & \qw & \targ & \gate{R(\alpha_1, \beta_1, \gamma_1)} & \meter \qw \\ \lstick{\ket{0}} & \gate{R_y(\arctan(x_2))} & \gate{R_z(\arctan(x_2^2))} & \targ & \ctrl{1} & \qw & \qw & \gate{R(\alpha_2, \beta_2, \gamma_2)} & \qw \\ \lstick{\ket{0}} & \gate{R_y(\arctan(x_3))} & \gate{R_z(\arctan(x_3^2))} & \qw & \targ & \ctrl{1} & \qw & \gate{R(\alpha_3, \beta_3, \gamma_3)} & \qw \\ \lstick{\ket{0}} & \gate{R_y(\arctan(x_4))} & \gate{R_z(\arctan(x_4^2))} & \qw & \qw & \targ & \ctrl{-3}& \gate{R(\alpha_4, \beta_4, \gamma_4)} & \qw \gategroup{1}{4}{4}{8}{.7em}{--}\qw } \end{minipage} \end{center} \caption[Variational quantum circuit component for QCNN kernel (filter).]{{\bfseries Variational quantum circuit component for QCNN kernel (filter).} The QCNN kernel (filter) includes three components: \emph{encoding}, \emph{variational} and \emph{quantum measurement}. The encoding component consists of several single-qubit gates $R_y(\arctan(x_i))$ and $R_z(\arctan(x_i^2))$ which represent rotations along $y$-axis and $z$-axis by the given angle $\arctan(x_i)$ and $\arctan(x_i^2)$, respectively. These rotation angles are derived from the input pixel values $x_i$ and are not subject to iterative optimization. The choose of arc tangent function is that in general the input values are not in the interval of $[-1, 1]$. The variational component consists of CNOT gates between each pair of neighbouring qubits which are used to entangle quantum states from each qubit and general single qubit unitary gates $R(\alpha,\beta,\gamma)$ with three parameters $\alpha,\beta,\gamma$. Parameters labeled $\alpha_i$, $\beta_i$ and $\gamma_i$ are the ones for iterative optimization. The quantum measurement component will output the Pauli-$Z$ expectation values of designated qubits. The number of qubits and the number of measurements can be adjusted to fit the problem of interest. In this work, we use the VQC as a convolutional kernel (filter), therefore the number of qubits equals to the square of kernel (filter) size and we only consider the measurement on the first qubit. The grouped box in the VQC may repeat several times to increase the number of parameters, subject to the capacity and capability of the available quantum computers or simulation software used for the experiments.} \label{Fig:Basic_VQC} \end{figure} \subsection{Classical Post-processing} The output from the last quantum convolutional layer then will be flattened and processed by a single layer of a fully connected classical neural network. To represent the output values as the probabilities of each class label, we further employ the softmax function on the post-processed output. \subsection{Loss Function and Optimization} In this classification task, we use \emph{cross-entropy} loss, which can be written in the following formulation: \begin{equation} L(\hat{\bm{y}}, \bm{y}) = -\sum_{c=1}^{M} y_{o, c} \log \left(\hat{y}_{o, c}\right), \end{equation} where \begin{itemize} \item $M$ - the number of classes \item $log$ - the natural log \item $y_{o, c}$ - the binary indicator ($0$ or $1$) if class label $c$ is the correct classification for observation $o$ \item $\hat{y}_{o, c}$ - predicted probability observation $o$ is of class $c$. \end{itemize} In this work, we use the gradient-based method to update the circuit parameters. The first problem is to calculate the gradients of quantum functions. The quantum functions are a series of operations with quantum gates, which are not the same as the layer operations in classical DNNs. In addition, the quantum functions typically are measured to retrieve the expectation values, which are stochastic by nature. In our work, we adopt the \emph{parameter-shift} rule~\cite{schuld2019evaluating, bergholm2018pennylane} to perform all of the quantum gradient calculation. For example, if we know how to calculate the expectation value of an observable $\hat{P}$ on our quantum function, \begin{equation} f\left(x ; \theta_{i}\right)=\left\langle 0\left|U_{0}^{\dagger}(x) U_{i}^{\dagger}\left(\theta_{i}\right) \hat{P} U_{i}\left(\theta_{i}\right) U_{0}(x)\right| 0\right\rangle=\left\langle x\left|U_{i}^{\dagger}\left(\theta_{i}\right) \hat{P} U_{i}\left(\theta_{i}\right)\right| x\right\rangle, \end{equation} where $x$ is the input value (e.g., pixel values); $U_0(x)$ is the state preparation routine to transform or encode $x$ into a quantum state; $i$ is the circuit parameter index for which the gradient is to be evaluated; and $U_i(\theta_i)$ represents the single-qubit rotation generated by the Pauli operators $X, Y$, and $Z$. It can be shown \cite{mitarai2018quantum} that the gradient of this quantum function $f$ with respect to the parameter $\theta_i$ is \begin{equation} \nabla_{\theta_i} f(x;\theta_i) = \frac{1}{2}\left[ f\left(x;\theta_i + \frac{\pi}{2}\right) - f\left(x;\theta_i - \frac{\pi}{2}\right)\right]. \label{eq:quantum gradient} \end{equation} Here, we have the recipe to calculate the quantum gradients. However, it still is not clear how to update the circuit parameters. In the simplest form of the gradient-descent method, the parameters are updated according to: \begin{equation} \theta \leftarrow \theta - \eta \nabla_{\theta} L(x;\theta), \end{equation} where the $\theta$ is the model parameter, $L$ is the loss function, and $\eta$ is the learning rate. However, this vanilla form does not always work. For example, it may be easily stuck in local optimum \cite{ruder2016overview}, or it can make the model difficult to train. There are several gradient-descent variants that are successful \cite{ruder2016overview, Tieleman2012, kingma2014adam}. Based on previous works \cite{chen2020quantum, chen19}, we use the RMSProp optimizer to optimize our hybrid quantum-classical model. RMSProp \cite{Tieleman2012} is a special kind of gradient-descent method with an adaptive learning rate that updates the parameters $\theta$ as: \begin{subequations} \begin{align} E\left[g^{2}\right]_{t} &= \alpha E\left[g^{2}\right]_{t-1}+ (1 - \alpha) g_{t}^{2}, \\ \theta_{t+1} &= \theta_{t}-\frac{\eta}{\sqrt{E\left[g^{2}\right]_{t}}+\epsilon} g_{t}, \end{align} \end{subequations} where $g_t$ is the gradient at step $t$ and $E\left[g^{2}\right]_{t}$ is the weighted moving average of the squared gradient with $E[g^2]_{t=0} = g_0^2$. The hyperparameters are set for all experiments in this paper as follows: learning rate $\eta =0.01$, smoothing constant $\alpha = 0.99$, and $\epsilon = 10^{-8}$. \subsection{Dropout} \emph{Overfitting} is a phenomenon where a machine learning model learned the statistical noise in the training data. This will cause poor performance when the models are tested against the unseen data or testing data. In other words, the model is not well generalizable. Such difficulties often emerge when training a classical DNN on a relatively small training set, and a QCNN is no exception. One potential method to reduce overfitting is to train all possible neural network architectures on the given dataset and average the predictions from each model. However, this is impractical because it will require unlimited computational resources. \emph{Dropout} is a method that approximates the effect of training a large number of neural networks with different architectures simultaneously~\cite{srivastava2014dropout}. The dropout operation entails (as follows): during the \emph{training} phase, on each of the forward passes, some of the output values from the specified layer will become zeroes. Each one of the values from the specified layer will be zeroed independently with probability $p$ from a Bernoulli distribution. This dropout procedure will not be performed in the \emph{testing} phase. \section{Experiments and Results}\label{sec:results} This section presents the numerical simulation of QCNN on the task of classifying different HEP events. The input data are in the dimension of $30 \times 30$. To demonstrate the possible quantum advantage, the classical and quantum CNN have a comparable number of parameters. \figureautorefname{\ref{scaled_examples}} shows the examples of the data used to train and test the QCNN models. For fair comparison, we arrange the experiment of QCNN and CNN to have similar numbers of parameters. In the classical CNN part, there are $4$ channels in the first convolutional layer with the filter size as $5 \times 5$ and $2$ channels in the second convolutional layer with the filter size as $5 \times 5$. Finally, there is a fully connected layer, which features $7 \times 7 \times 2 \times 2 + 2 = 198$ parameters. Therefore, the total number of parameters in the classical CNN is $4 \times 5 \times 5 + 4 \times 2 \times 5 \times 5 + 198 = 498$. For the QCNN, there is $1$ channel in the first quantum convolutional layer with the filter size of $3 \times 3$ and another single channel in the second quantum convolutional layer with the filter size of $2 \times 2$. Finally, there is a classical fully connected layer with $ 14 \times 14 \times 1 \times 2 + 2 = 394$ parameters. As such, the total number of the hybrid quantum-classical CNN is $54 + 24 + 394 = 472$. \begin{figure}[htbp] \centering \includegraphics[width=1.\linewidth]{figures/DUNE_DATA_DEMO_DOWNSCALE_for_QCNN_full.pdf} \caption{Examples of scaled images of simulated particle activities ($\mu$, $\pi^+$, $p$, $e^-$) in a LArTPC detector. These are the images used in the QCNN experiments. The dimension of these images is $30 \times 30$ pixels.} \label{scaled_examples} \end{figure} The software used for this work are PyTorch \cite{paszke2019pytorch}, PennyLane \cite{bergholm2018pennylane}, and Qulacs \cite{suzuki2020qulacs}. \subsection{Muon versus Electron} \figureautorefname{\ref{comparison_results_mu_electron}} and Table~\ref{tab:results_comparison_mu_e} depict the results of the classification between $\mu^+$ and $e^-$. As mentioned in Section~\ref{sec:hep-data}, A $\mu^+$ is a track-like particle, while an $e^-$ produces electromagnetic showers that are spatially extended. This is a relatively straightforward pattern recognition problem when the particle tracks are more than a few meters ($\sim$2 meters in this simulation). In fact, we see that the test accuracy in QCNN (92.5\%) and CNN (95\%) is comparable to each other with a comparable number of parameters. On the other hand, the QCNN converges to its optimal accuracy much faster with a fewer number of epochs. \begin{figure}[htbp] \centering \includegraphics[width=1.0\linewidth]{comparison_results/QCNN_vs_CNN_mu+_vs_e-_full.pdf} \caption[Result: QCNN on binary classification of muon versus electron]{{\bfseries Result: QCNN on binary classification of muon versus electron} Training QCNN on the classification of $\mu^+$ and $e^-$. The filter size is $3$ in the first convolutional layer and $2$ in the second convolutional layer. There is $1$ channel in both Conv Layers. The number of parameters in this setting: $9 \times 3 \times 2 = 54$ in the first convolutional layer, $4 \times 3 \times 2 = 24$ in the second convolutional layer, and $14 \times 14 \times 1 \times 2 + 2= 394$ in the fully connected layer. Total number of parameters is $54 + 24 + 394 = 472$.} \label{comparison_results_mu_electron} \end{figure} \begin{table}[htbp] \centering \begin{tabular}{|l|l|l|l|l|} \hline & Training Accuracy & Testing Accuracy & Training Loss & Testing Loss \\ \hline QCNN & $100\%$ & $92.5\%$ & $0.017$ & $0.13$ \\ \hline CNN & $99.38\%$ & $95\%$ & $0.0002$ & $0.0046$ \\ \hline \end{tabular} \caption{Performance comparison between QCNN and CNN on the binary classification between $\mu^+$ and $e^-$.} \label{tab:results_comparison_mu_e} \end{table} \subsection{Muon versus Proton} \figureautorefname{\ref{comparison_results_mu_proton}} and Table~\ref{tab:results_comparison_mu_p} show the results of the classification between $\mu^+$ and $p$. As described in Section~\ref{sec:hep-data}, a proton is a track-like particle akin to a muon. However, because a proton's mass is much heavier than a muon or pion, it has higher energy loss and encounters less MCS when it passes the detector. As a result, a proton's track has higher intensity and is straighter than that of a muon. This classification is more difficult than the previous case detailing muon versus electron, which is evident from the CNN's test accuracy of 80\%. In this case, we show that with a comparable number of parameters, the QCNN outperforms the classical CNN, both in test accuracy and learning speed. The QCNN reaches $97.5$\% test accuracy at around $10$ epochs, while the classical CNN plateaus at a test accuracy of $80\%$ at roughly $75$ epochs. \begin{figure}[htbp] \centering \includegraphics[width=1.\linewidth]{comparison_results/QCNN_vs_CNN_mu+_vs_proton_full.pdf} \caption[Result: QCNN on binary classification of muon versus proton]{{\bfseries Result: QCNN on binary classification of muon versus proton} Training QCNN on the classification of $\mu^+$ and $p$. The filter size is $3$ in the first convolutional layer and $2$ in the second convolutional layer. There is $1$ channel in both Conv Layers. The number of parameters in this setting: $9 \times 3 \times 2 = 54$ in the first convolutional layer, $4 \times 3 \times 2 = 24$ in the second convolutional layer, and $14 \times 14 \times 1 \times 2 + 2= 394$ in the fully connected layer. Total number of parameters is $54 + 24 + 394 = 472$.} \label{comparison_results_mu_proton} \end{figure} \begin{table}[htbp] \centering \begin{tabular}{|l|l|l|l|l|} \hline & Training Accuracy & Testing Accuracy & Training Loss & Testing Loss \\ \hline QCNN & $100.00\%$ & $97.5\%$ & $0.041$ & $0.087$ \\ \hline CNN & $91.25\%$ & $80\%$ & $0.002$ & $0.01$ \\ \hline \end{tabular} \caption{Performance comparison between QCNN and CNN on the binary classification between $\mu^+$ and $p$.} \label{tab:results_comparison_mu_p} \end{table} \subsection{Muon versus Charged Pion} \figureautorefname{\ref{comparison_results_mu_charged_pion}} and Table~\ref{tab:results_comparison_mu_pi} illustrate the results of the classification between $\mu^+$ and $\pi^+$. This is another difficult classification problem because a charged pion behaves much like a muon in terms of energy loss, MCS, and decay. As described in Section~\ref{sec:hep-data}, the main difference is that the $\pi^+$ experiences additional nuclear interactions during its passage in the detector, often leading to a hard scattering (represented as a ``kink'') along its main trajectory. In this case, we show that with a comparable number of parameters, the QCNN outperforms the classical CNN, both in test accuracy and learning speed. The QCNN reaches $97.5$\% test accuracy at the first few epochs, while the classical CNN flattens at the test accuracy of $82.5\%$ at around $100$ epochs. In this case, we add a dropout layer in the QCNN with dropout rate $=0.3$ to improve its robustness against overfitting. \begin{figure}[htbp] \centering \includegraphics[width=1.\linewidth]{comparison_results/QCNN_vs_CNN_mu+_vs_charged_pion_dropout_03_full.pdf} \caption[Result: QCNN on binary classification of muon versus charged pion]{{\bfseries Result: QCNN on binary classification of muon versus charged pion} Training QCNN on the classification of $\mu^+$ and $\pi^+$. The filter size is $3$ in the first convolutional layer and $2$ in the second convolutional layer. There is $1$ channel in both Conv Layers. The number of parameters in this setting: $9 \times 3 \times 2 = 54$ in the first convolutional layer, $4 \times 3 \times 2 = 24$ in the second convolutional layer, and $14 \times 14 \times 1 \times 2 + 2= 394$ in the fully connected layer. Total number of parameters is $54 + 24 + 394 = 472$. In this experiment, we add a dropout layer with dropout rate of $0.3$ to the QCNN.} \label{comparison_results_mu_charged_pion} \end{figure} \begin{table}[htbp] \centering \begin{tabular}{|l|l|l|l|l|} \hline & Training Accuracy & Testing Accuracy & Training Loss & Testing Loss \\ \hline QCNN & $96.88\%$ & $97.5\%$ & $0.1066$ & $0.1121$ \\ \hline CNN & $97.5\%$ & $82.5\%$ & $0.0006$ & $0.0116$ \\ \hline \end{tabular} \caption{Performance comparison between QCNN and CNN on the binary classification between $\mu^+$ an $\pi^+$.} \label{tab:results_comparison_mu_pi} \end{table} \section{Related Works} The concept of QCNNs has been discussed recently. In~\cite{cong2019quantum}, the authors propose an architecture based on the Multiscale Entanglement Renormalization Ansatz (MERA) tensor network to perform the classification of quantum states. Our approach differs from this work as we focus on the classical input data. In~\cite{kerenidis2019quantum, li2020quantum}, the authors propose a QCNN framework to deal with the classical data, which is like our method in the sense of targets. However, those works require the operation of quantum random access memory (QRAM), which is difficult to implement on physical devices in the near term. In~\cite{oh2020tutorial, liu2019hybrid}, the authors consider a more realistic architecture that also is hybrid quantum-classical. While in a similar vein, our work differs from that research because it implements input data with much larger dimensions. In~\cite{liu2019hybrid}, the data are in the dimension of $3 \times 3$, while in~\cite{oh2020tutorial}, the data are in the dimension of $10 \times 10$. Our architecture is capable of dealing with dimensions up to $30 \times 30$. We also noted the work in \cite{henderson2020quanvolutional}. In~\cite{henderson2020quanvolutional,yang2020decentralizing}, quantum circuits are randomly sampled and not subject to iterative optimization. In our work, the quantum and classical parts are trained in an end-to-end fashion. The trainability of QCNN is studied in the recent work, \cite{pesah2020absence}, indicating that QCNN optimization is more viable than other quantum neural network architectures. \section{Conclusion and Outlook} \label{sec:conclusion} In this work, we propose a quantum machine learning framework for learning HEP events. Specifically, we demonstrate the QCNN architecture with significant learning capacity in terms of learning speed and testing accuracy compared to the classical CNN when both use a comparable number of parameters. We expect the proposed framework will have a wide range of applications in the era of Noisy Intermediate-Scale Quantum (NISQ) and beyond, as well as in more HEP experiments. Looking ahead, there are several research areas where we could extend our QCNN framework. First, in this work, we perform experiments using the input dimension of $30 \times 30$ and a single input channel. In comparison, multiple input channels (e.g., different color channels) with higher dimensions are rather common in a classical CNN. Future studies to increase the data complexity in QCNNs are expected as the speed of quantum simulators improves. Second, this work presents a noise-free simulation to demonstrate proof-of-concept QCNN experiments on HEP event classification. Our framework is based on VQC, which has the potential to be robust against device noise. However, applying parameter-shift methods to calculate quantum gradients requires a large amount of quantum circuit evaluations, which is infeasible at this time. For example, given a filter with the size $N \times N$ and the number of operations needed to scan over a single input images $S$, the contribution to the total number of circuit evaluation is, at least, $\mathcal{O}(N^2S)$. Therefore, the number of total circuit evaluations grows as the circuit goes deeper (with more convolutional layers) or expands wider (with more filters in each layer). We reserve pursuing this area of study until the appropriate quantum computing resources are available. Finally, as the convolutional operation is quite versatile, it has been widely used beyond computer vision in classical machine learning, for example, in modeling data with temporal or sequential dependencies \cite{kalchbrenner2014convolutional, hu2014convolutional, acharya2017deep, lai2015recurrent, yin2016abcnn}. As such, our proposed general QCNN architecture is not limited to image classification tasks and can be extended to other application domains. \begin{acknowledgments} This work is supported by the U.S.\ Department of Energy, Office of Science, Office of High Energy Physics program under Award Number DE-SC-0012704 and the Brookhaven National Laboratory Directed Research and Development (LDRD) Program \#20-024. \end{acknowledgments}

431b5f7a4e2ad81c4ac98b0c99735117b71f2bd7

\section{Introduction} The novel coronavirus disease (COVID-19), which has been officially announced as a pandemic by the World Health Organization (WHO), is perhaps the most serious public health emergency over the past decades. The coronavirus continues to spread around the globe, which challenges the governments and medical systems all over the world. While metropolitan-wide lockdown has demonstrated its effectiveness as a nonpharmaceutical intervention in several countries, the cost of such measures, including unemployment, economic crash, and social anxiety, makes it a tough decision on behalf of administrators. A compromised solution is to place containment measures onto a subset of areas in a city to stop or slow down the spread of COVID-19 while minimizing the social and economic cost. To precisely distinguish the high-risk areas from the city, the spatial distribution of confirmed cases has been used as the key criterion for the potential containment measures in a data-driven fashion. While the spatial statistics of confirmed cases work, the time consumption and the granularity of data acquisition significantly lower the efficiency and effectiveness of such methods. For example, the incubation period of COVID-19 is around 5--6 days on average, but it could last as long as 14 days. During such a period, a community or neighborhood would have been already invaded by a small number of asymptotic carriers who might not be with any clinical symptoms. When a small number of cases being confirmed, the community and its surrounding neighborhoods may have fallen into COVID-19 for a long time. Furthermore, though the mobility of confirmed patients is usually well restricted, the asymptomatic carriers with no symptoms would still spread the virus to where he/she has gone. When fine-grained mobility traces are not available, the administrators can only place containment measures to places in coarse-grained. To tackle the technical issues, in this paper, we propose C(OVID)-Watcher, a novel framework to support early detection of high-risk residential neighborhoods for fighting against the spread of COVID-19. Our intuition is to incorporate human mobility data, so as to (1) characterize the socioeconomic and demographic status of every neighborhood~\cite{borjas2020demographic,huang2020kddtransp} based on ``how residents move''~\cite{renso2013you} and (2) unfold the spatial interactions~\cite{jiang2020spatial} and potential influences on COVID-19 caused by the mobility of massive asymptotic carriers. Specifically, C-Watcher includes a mobility data-driven machine learning model that screens every neighborhood in a target city and predicts the infection risks, prior to the spread of COVID-19 from epicenters to the city. In addition to the use of mobility-related features, we also hope to generalize the evidence already witted in the epicenter for screening the risk of neighborhoods in the target city, prior to or in the early stage of local outbreaks. To achieve the goal, a core component of C-Watcher is a novel cross-city transfer learning model that transfers knowledge about COVID-19 infections from the epicenter to the target city, while cities are quite different in a large number of domains, ranging from living, foods, transportation, and residences. All in all, we have made three contributions as follows. \begin{itemize} \item Through extensive data analytics, we explore a set of empirical features related to long-term/regular human mobility patterns (before the COVID-19). With such long-term mobility features, the socioeconomic and demographic status, as well as the spatial interactions among neighborhoods could be well characterized. In this way, one can easily distinguish high-risk neighborhoods from the urban area and predict the potential risks for infection, with respect to the two factors. \item While these features are with certain discriminative information for risk prediction, they also involve some city-specific characteristics. For example, the popular choices for transport modes in different cities vary. Such city-specific characteristics burden the use of mobility-related features to transfer the knowledge obtained in the epicenter to the target city. To generalize the knowledge transfer, C-Watcher adopts a novel adversarial encoder-decoder framework to learn the ``city-invariant'' representations from the mobility-related features for prediction. \item To validate C-Watcher, we collect and construct real-world datasets for high-risk neighborhood detection based on the publicly available information from the web and human mobility traces from the largest online map service in China. We conduct extensive experiments for evaluation. The results demonstrate that C-Watcher can accurately predict the potential risk of massive residential neighborhoods in a large number of Chinese cities. With large datasets, C-Watcher makes insightful suggestions on preventing the epidemic of COVID-19 alike for different residential neighborhoods via feature importance. \end{itemize} \section{Notations and Related Work} In this section, we first introduce the basic notations used throughout this paper, and then we formally formulate the research problem for early detection of high-risk neighborhoods. Last, we review the studies that are relevant to our work with the most related work discussed. \paragraph{Notations and Formulation.} We use $\bm{n}$ to denote the features of a residential neighborhood which will be presented in Section \ref{sec:feature}, and use $y$ to denote the binary label meaning whether the neighborhood is high risky ($y=1$) or not ($y=0$). The detection problem can be defined as: $f(\bm{n}) \rightarrow y$ where the function $f(\cdot)$ can be any machine learning model like Multi-Layer Perceptron (MLP). The objective of C-Watcher is to make early detection of high-risk neighborhoods without epidemic outbreaks. Instead of relying on the confirmed infection cases to make a prediction which deemed to be time-delayed like \cite{fu2020aware}, we assume that the COVID-19 epidemic only outbroke in epicenter cities (such as Wuhan in China) and no prior knowledge of confirmed cases, spreading trend or known hazard neighborhoods in target cities can be referred to. Such a cross-city prediction problem of latent high risky residential neighborhoods can be formulated as: \begin{equation} \label{eq-1} f_{cross}(\bm{n}^{T}|\,\{(\bm{n}_i^{E}\!, y_i^{E})\}) \rightarrow y^{T} \end{equation} where $\bm{n}^T$ and $y^T$ denote the features and binary label of a residential neighborhood in the target city. $\bm{n}_i^E$ and $y_i^E$ denote the features and label of a neighborhood from epicenter cities set. Hereafter, we omit subscript $i$ for simplicity. The $f_{cross}(\cdot)$ is a cross-city transfer learning model which is trained without ground-truth information in the target city. \paragraph{Related Work.} Aiming to fight against the COVID-19 pandemic, researchers in the computer science community carried out many studies from several perspectives recently. For instance, \citet{huang2020kddtransp} exhibit that user transportation-related behaviors in China have indeed been impacted by the containment measures during the COVID-19 pandemic. There are also a few studies \cite{huang2020quantifying,xiong2020understanding,liu2020Investigation} investigating the human mobility, the local economy, and the information acquisition during the COVID-19 outbreak in China while most of these studies remain at city level. Some studies also demonstrate the effectiveness of mobility data for controlling the spread of COVID-19. \citet{vollmer2020using} exploit a Bayesian semi-mechanism model with mobility data to show the effectiveness to slow down the spread of the virus by constraints on individual movements and social interactions. Based on the integration of mobility data and the global epidemic model \cite{balcan2009multiscale}, a study also reveals the effectiveness of fine-grained targeted mobility control policies towards the COVID-19 pandemic \cite{hao2020understanding}. The mobility data can also be integrated with compartmental models in epidemiology (like Susceptible-Exposed-Infected-Recovered (SEIR) model) \cite{ghamizi2020data} to better predict the epidemic dynamics. \paragraph{Discussion.} From the problems and methodologies perspectives, the most relevant work to our study includes~\cite{fu2020aware} and~\cite{xu2019deep,peng2019cm,mai2020modality}. Compared to~\cite{fu2020aware}, which smooths the confirmed cases of infections over spatial domains and predicts hazard areas during the COVID-19 outbreaks using simple spatial features like distance, C-Watcher system tackles the time delay and coarse-grained granularity issues and can early detect the high-risk residential neighborhoods even before the outbreaks, through leveraging features derived from long-term/regular human mobility patterns. In terms of methodologies, though a great number of algorithms have been proposed for adversarial representation learning~\citet{makhzani2015adversarial}, adversarial metric learning~\cite{xu2019deep} and cross-modalities~\cite{mai2020modality}, our work is the first to study the city-invariant representation learning through Generative Adversarial Networks \cite{goodfellow2014generative} in the context of urban computing and COVID-19 prediction. \section{Features for Neighborhood Detection}\label{sec:feature} In this section, we present how to construct features from mobility data to characterize a residential neighborhood for early risk detection. We first introduce the data source used in our framework, and then three groups of constructed features are briefly discussed which are Point of Interest (POI) radius features (see Section \ref{sec:poi_radius}), demographic features (see Section \ref{sec:demographic features}) and transportation-related features (see Section \ref{sec:Transport-related features}), respectively. More details about the feature construction can be found in the Appendix \ref{apx:feacon}. The feature construction is mainly based on three data sources: POI basic property data, user profile data and human mobility data. POI basic property data contains the basic information of a POI, such as name, coordinates and types, which provides many semantic information for a POI \cite{10.1145/3394486.3403318,yuan2020spatio,hu2019we}. This data enable us to analyze the spatial relationship between neighborhoods and different types of POIs, such as hospitals, schools and bus stops \cite{Li2020competitive}. The user profile data are obtained from a user profile platform of Baidu which can return profile features for almost all internet users in China, such as gender, age and educational level. Human mobility data, collected from Baidu Maps in China, record search and transportation behaviors of the map users. \subsection{POI Radius Features}\label{sec:poi_radius} Here we introduce how to compute a group of POI radius features for a residential neighborhood based on POI basic property data. The intuition for this feature group is that basic living facilities around a residential neighborhood may have a correlation with the probability of its residents being infected by COVID-19. For example, a neighborhood lacking basic living facilities may face a high risk, for the residents may passively go further away for basic living needs and face greater infection risks. Moreover, neighborhoods with poor living facilities often lack good property management, which may also lead to high infection risks. To describe these living facilities related characteristics, we construct 15 POI radius features. Each of them is defined as the shortest distance between the neighborhood and one certain type of POIs. All the used types of POIs are listed in the Appendix \ref{apx:feacon}. Meanwhile, we define an additional binary feature to directly represent the perfect degree of living facilities. The value of this feature will be assigned as ``perfect'' if a set of basic living facilities (e.g. hospital, bus stop and so on) are all within $1 km$ of the given neighborhood. Otherwise, it is assigned ``poor''. The list of basic living facilities is also shown in the Appendix \ref{apx:feacon}. We collect the high-risk and low-risk neighborhoods data in Wuhan city in China which is officially announced by the local government. Figure \ref{infrastructure} presents the ratio distribution of high-risk and low-risk neighborhoods grouping by this ``perfect-poor'' facility label in Wuhan data. As we can see from Figure \ref{infrastructure}, for the neighborhoods with feature value as ``perfect'', the ratio of low-risk neighborhoods and high-risk ones is $0.57:0.44$; whereas the ratio of them for ``poor'' ones is $0.43:0.56$. It indicates that more high-risk neighborhoods have poor living facilities, while the low-risk neighborhoods are just the opposite. \begin{figure} \centering \subfigure[]{ \label{infrastructure} \includegraphics[width=0.48\columnwidth]{figs/infrastructure.pdf}} \hspace{-1mm} \subfigure[]{ \label{population_density} \includegraphics[width=0.48\columnwidth]{figs/population_density_norm.pdf}} \caption{Features of living facilities and population density visual analysis.} \label{feature_distribution} \end{figure} \subsection{Demographic Features}\label{sec:demographic features} Next, we present the demographic features of a residential neighborhood. At first, given that the COVID-19 is easy to transmit in a person-to-person way \cite{liu2020aerodynamic}, it is necessary to take into account population density for infection risks prediction. As Figure \ref{population_density} illustrates, on average high-risk neighborhoods do have a higher population density than low-risk neighborhoods in Wuhan city. We also compute average commute distance as a feature for each neighborhood since residents with long commutes have high infection risks. Moreover, different groups of residents may face different risk levels in a neighborhood. For example, old people and children are easier to be infected. And residents with higher educational levels may pay more attention to scientific prevention. Hence, we construct 11 features based on the distribution of residents according to different human attributes. We present each of these features as a vector of histogram statistics of residents' distribution. The full list of such attributes is provided in the Appendix \ref{apx:feacon}. \subsection{Transportation-Related Features}\label{sec:Transport-related features} We also extract features of transportation-related behaviors from human mobility data to help predict infection risks. There have been some studies to prove that transportation-related behaviors have a close relationship with COVID-19 contagion spreading \cite{huang2020kddtransp}. The transportation-related behaviors typically are recognized as the origin-transportation-destination (OTD) information\cite{8708931,xu2016taxi,Polestar}. Thus, we consider detailed features from the perspectives of T (transportation), OD (origin \& destination venues) and OTD (origin-transportation-destination pattern). All the features are extracted from the search and transportation data of Baidu Maps in a certain time period. Previous studies have shown that map search behavior is a leading indicator and predictor for crowd dynamics \cite{zhou2018early}. A vector in which the value of each element equals the corresponding ratio of transportation means, mainly including walk, bicycle, public transit and private vehicle, is used to depict the ``T feature'' for a residential neighborhood. The ``OD feature'' consists of types of visit venues and the distance between origin and destination venues. We classify the destination venues according to their types (e.g. hospital, restaurant, hotel, and school) and compute the proportion of each type. We also extract origin-destination distance and categorize it into different distance buckets. The proportions of different distance buckets for the neighborhood are also formed as a feature vector. Moreover, since the OTD (origin-transportation-destination) patterns most directly reflect human mobility, we collect the top-20 hottest travel patterns from all the cities in our dataset, which is a triplet tuple composed of the type of origin venue (residential area), means of transportation and type of destination venue. The histogram distribution of these top-20 OTD travel patterns of each neighborhood is treated as ``OTD'' features. More details about the transportation-related features can be found in the Appendix \ref{apx:feacon}. \section{Cross-City Transfer Learning}\label{sec:transer frame work} \begin{figure}[t] \centering \includegraphics[width=0.48\textwidth]{figs/preprint-C-Watcher-frame.png} % \caption{Illustration of cross-city transfer learning model of C-Watcher.} \label{model-frame} \end{figure} In this section, we present the cross-city transfer learning model, which is a core component of C-Watcher, to improve the performance of early detection of high-risk neighborhoods via transferring the knowledge about COVID-19 infection from the epicenter to the target city. Usually, discrepancies always exist between different cities. Thus, we intend to learn city-invariant knowledge applicable to both epicenters and target cities, instead of those characteristics unique to epicenter cities. \subsection{Overview}\label{sec:framework overview} An overview of our proposed cross-city transfer learning model with four components is given in Figure \ref{model-frame}. The first component is a neural network encoder used to learn the representation of a neighborhood on the basis of three groups of features introduced in Section \ref{sec:feature}. However, the discrepancy of input distributions in different cities may lead to the gap between encoded representations of epicenter cities and target cities, which may severely disrupt the detection ability in target cities. Thus, we adopt adversarial learning by adding a discriminator component to identify whether the output of the encoder belongs to the target city or not. In addition, to ensure that the embedded feature of the encoder still keeps the ability to depict the residential neighborhoods, we exert two decoders which recover features of epicenter cities and target city respectively from the output of the encoder. Moreover, to achieve the prediction goal of C-Watcher, a classifier is also added to optimize the learned representations space and make it more related to the prediction of COVID-19 infection risks. In the following sections, we present our cross-city transfer learning model in detail by introducing how each component works. Here comes the notation of model input first. For both residential neighborhoods in epicenter cities and target cities, we have a feature vector consist of three groups: \begin{equation} \begin{split} \label{eq-2} \bm{n}^{E} &= cat(\bm{n}_{r}^{E}, \bm{n}_{h}^{E}, \bm{n}_{t}^{E})\\ \bm{n}^{T} &= cat(\bm{n}_{r}^{T}, \bm{n}_{h}^{T}, \bm{n}_{t}^{T}) \end{split} \end{equation} where $\bm{n}^E$ denotes all of the features of a residential neighborhood in epicenter cities. $\bm{n}_r^E$, $\bm{n}_h^E$ and $\bm{n}_t^E$ respectively denote the POI radius features, demographic features and transportation-related features of that residential neighborhood in epicenter cities. $\bm{n}^T$, $\bm{n}_r^T$, $\bm{n}_h^T$ and $\bm{n}_t^T$ similarly denote the corresponding features of residential ones in target city. The function $cat(\cdot)$ is the concatenating operation. \subsection{City-Invariant Representation Learning}\label{asversarial learning} Since learning unique characteristics of neighborhoods in epicenter cities brings little benefit for early risk detection in target cities, we propose a city-invariant representation learning method, which is inspired by the multi-mode adversarial representation learning methods \cite{mai2020modality,makhzani2015adversarial}. Here the encoder is a transformer of data distribution. Given the input vectors $\bm{n}^E$ and $\bm{n}^T$, we use $\widetilde{\bm{n}}^E$ and $\widetilde{\bm{n}}^T$ to denote the outputs of the encoder. Similar to \cite{mai2020modality} and \cite{makhzani2015adversarial}, the distributions transformation from the inputs to encoded representations can be presented as: \begin{equation} \begin{split} \label{eq-3} p(\widetilde{\bm{n}}^E,\Phi_e) &= \int_{\bm{n}^E} e(\widetilde{\bm{n}}^E |\,\bm{n}^E, \Phi_e)\,p(\bm{n}^E)\,d\bm{n}^E\\ p(\widetilde{\bm{n}}^T,\Phi_e) &= \int_{\bm{n}^T} e(\widetilde{\bm{n}}^T |\,\bm{n}^T, \Phi_e)\,p(\bm{n}^T)\,d\bm{n}^T \end{split} \end{equation} where $p(\cdot)$ denotes the data distribution and $e(\cdot \,,\Phi_e)$ represents the encoding distribution. $\Phi_e$ are parameters of the encoder, determining the projection space where the distributions of input data $p(\bm{n}^E)$ and $p(\bm{n}^T)$ are transformed into that of encoded representations $p(\widetilde{\bm{n}}^E,\Phi_e)$ and $p(\widetilde{\bm{n}}^T,\Phi_e)$. In general, $p(\widetilde{\bm{n}}^E,\Phi_e)$ and $p(\widetilde{\bm{n}}^T,\Phi_e)$ are different distributions characterizing different cities. It means that some features helpful in predicting COVID-19 infection risk may be unique to neighborhoods in epicenter cities unless we impose constraints on the encoder. To this end, we use adversarial learning to narrow the discrepancies between distributions $p(\widetilde{\bm{n}}^E,\Phi_e)$ and $p(\widetilde{\bm{n}}^T,\Phi_e)$ by adding a discriminator to distinguish whether the neighborhood comes from epicenter cities or the target city. In this way, the discriminator needs to do a binary classification task, in which it takes the encoded representations as inputs and aims to identify the inputs $\widetilde{\bm{n}}^E$ from epicenter cities as true but the inputs $\widetilde{\bm{n}}^T$ from target cities as false, while the encoder tries its best to confuse the discriminator to classify both of them as true. We can formulate the function of discriminator as: \begin{equation} \begin{split} \label{eq-4} D(\widetilde{\bm{n}}^E,\Phi_D) \rightarrow 1\\ D(\widetilde{\bm{n}}^T,\Phi_D) \rightarrow 0 \end{split} \end{equation} where $D(\,\cdot,\Phi_D)$ denotes the function of the discriminator which can be an MLP model that outputs the probability from 0 to 1. On the contrary, the encoder competes against the discriminator by: \begin{equation} \begin{split} \label{eq-5} D(\widetilde{\bm{n}}^E,\Phi_D) \rightarrow 1\\ D(\widetilde{\bm{n}}^T,\Phi_D) \rightarrow 1 \end{split} \end{equation} For this adversarial learning procedure, we use binary cross entropy (BCE) to define the loss function: \begin{align} \label{eq-6} \mathcal{L}_{al} &= \mathcal{L}_{diff}(\widetilde{\bm{n}}^E,\widetilde{\bm{n}}^T) + \mathcal{L}_{ch}(\widetilde{\bm{n}}^E,\widetilde{\bm{n}}^T)\\ \mathcal{L}_{diff} &= -[log(D(\widetilde{\bm{n}}^E)) + log(1-D(\widetilde{\bm{n}}^T)]\\ \mathcal{L}_{ch} &= -[log(D(\widetilde{\bm{n}}^E)) + log(D(\widetilde{\bm{n}}^T)] \end{align} where $D(\widetilde{\bm{n}}^E)$ is used to represent $D(\widetilde{\bm{n}}^E, \Phi_D)$ in simplicity and so does $D(\widetilde{\bm{n}}^T)$. The differentiation loss $\mathcal{L}_{diff}$ guides discriminator to predict $\widetilde{\bm{n}}^E$ as true (epicenter cities) but $\widetilde{\bm{n}}^T$ as false (target city), while the encoder tries to learn features that are common between epicenter cities and target city to hinder discriminator from distinguishing successfully, under the effects of cheat loss $\mathcal{L}_{ch}$. The adversarial procedure will finally reach an equilibrium situation where the discriminator could no longer distinguish whether the encoded representations come from epicenter cities or target city, then the encoder is able to extract ``city-invariant'' features from raw inputs $\bm{n}^E$ and $\bm{n}^T$. In this case, discrepancies between cities decrease and the experience which helps predict infection risks in epicenter cities can make more sense in target cities. \subsection{Embedding Space Constraints}\label{space constrains} A problem about city-invariant representation learning is that, if no regulations and restrictions are imposed on the embedding space of the encoder, the encoded representations of epicenter cities and target cities may only be similar in distribution but fail to retain useful information for identifying high-risk neighborhoods. We solve this problem with multi-task learning strategy by additionally exerting an auto encoder-decoder features reconstruction component, as well as a COVID-19 infection risks prediction component. The reconstruction component consists of two decoders (one for residential neighborhoods in epicenter cities and another one for residential neighborhoods in the target city) which take the encoded representations as inputs. The decoding operation can also be considered as a distribution transformation like encoding: \begin{equation} \begin{split} \label{eq-9} p(\widehat{\bm{n}}^E,\Phi_d^E) &= \int_{\widetilde{\bm{n}}^E} d^E(\widehat{\bm{n}}^E |\,\widetilde{\bm{n}}^E, \Phi_d^E)\,p(\widetilde{\bm{n}}^E)\,d\widetilde{\bm{n}}^E\\ p(\widehat{\bm{n}}^T,\Phi_d^T) &= \int_{\widetilde{\bm{n}}^T} d^T(\widehat{\bm{n}}^T |\,\widetilde{\bm{n}}^T, \Phi_d^T)\,p(\widetilde{\bm{n}}^T)\,d\widetilde{\bm{n}}^T \end{split} \end{equation} where $\widehat{\bm{n}}^E$, $\widehat{\bm{n}}^T$ denote the reconstructed outputs of decoders from $\widetilde{\bm{n}}^E$ and $\widetilde{\bm{n}}^T$ respectively, and $d^E(\widehat{\bm{n}}^E |\,\widetilde{\bm{n}}^E,\Phi_d^E)$ represents the epicenter cities decoder function with parameters $\Phi_d^E$, while $d^T(\widehat{\bm{n}}^T |\,\widetilde{\bm{n}}^T,\Phi_d^T)$ is similar but for the target city. Aiming to approximate decoded representations to the original inputs ($\widehat{\bm{n}}^E \rightarrow \bm{n}^E$ and $\widehat{\bm{n}}^T \rightarrow \bm{n}^T$), we use mean square error to define reconstruction loss function: \begin{align} \label{eq-10} \mathcal{L}_{rec} &= \mathcal{L}_{rec}^E + \mathcal{L}_{rec}^T\\ &= \|\,\widehat{\bm{n}}^E,\bm{n}^E\|_2 + \|\,\widehat{\bm{n}}^T,\bm{n}^T\|_2 \end{align} Optimized by the reconstruction loss above, the encoder-decoder framework ensures that the embedding space is still characterizing a residential neighborhood. Moreover, considering that our ultimate objective is to detect latent high-risk neighborhoods, we add a classifier to identify COVID-19 infection risks in epicenter cities upon the learned encoded representations. The classification problem can be defined as : \begin{equation} \label{eq-11} C(\widetilde{\bm{n}}^E,\Phi_c) \rightarrow y^E,\, y^E \in \{0,1\} \end{equation} where $C(\cdot,\Phi_c)$ denotes the function of MLP classifier with parameter $\Phi_c$. This is also a binary classification task and we use BCE to define the classification loss function: \begin{equation} \label{eq-12} \mathcal{L}_{cl} \!=\! -y^{E}\!log(C(\widetilde{\bm{n}}^E\!\!,\Phi_c)) \!-\! (1 \!-\! y^{E})log(1 \!-\! C(\widetilde{\bm{n}}^E\!\!,\Phi_c)) \end{equation} The classification loss transmits the known information carried by label $y^E$ to encoder and classifier, which is COVID-19 infection risks of neighborhoods in epicenter cities. It achieves the goals to restrict the encoded representations to be instructive in high-risk neighborhood identification. All in all, loss functions generated from all the three components of discriminator, decoders and classifier will act on the encoder and optimize the embedding space in our proposed cross-city transfer model. The total loss function can be expressed qualitatively as: \begin{equation} \label{eq-13} \mathcal{L} = \lambda_{diff}\mathcal{L}_{diff} + \lambda_{ch}\mathcal{L}_{ch} + \lambda_{rec}\mathcal{L}_{rec} + \lambda_{cl}\mathcal{L}_{cl} \end{equation} In model training, the adversarial model is optimized in an alternate mode. We use differentiation loss $\mathcal{L}_{diff}$ to optimize the discriminator first to improve its discriminatory ability to neighborhoods in both epicenter cities and that target city, which also leads to the rise of cheat loss. Then we apply cheat loss $\mathcal{L}_{ch}$ combined with $\mathcal{L}_{rec}$ and $\mathcal{L}_{cl}$, to guide the encoder to optimize its parameters in a direction where demands to learn city-invariant, informative and risk-discriminative features are all taken into consideration. Together with the encoder, the decoders and classifier update themselves based on $\mathcal{L}_{rec}$ and $\mathcal{L}_{cl}$, respectively. \subsubsection{Reference City Validation Mechanism.}\label{sec: anchor} \begin{figure}[t] \centering \includegraphics[width=0.3\textwidth]{figs/anchor.pdf} \caption{Diagram of reference city validation mechanism.} \label{anchor-frame} \end{figure} Another problem of C-Watcher is how to select the best hyperparameters to train the model. Here we build a reference city validation mechanism to tune hyperparameters. The illustrated diagram is shown in Figure \ref{anchor-frame}. Reference city in our paper can be epicenter cities, and can also be some cities with COVID-19 outbreak but not so serious as epicenters. We train the C-Watcher model on epicenter cities set, and use ground truth data of the reference city as validation data to choose the hyperparameters. Then we evaluate the early detection performance in target cities geographically close to that reference city. In this case, we ensure that our trained model to detect latent high-risk neighborhoods in a target city with best hyperparameters, without any prior information related to COVID-19 confirmed cases and spreading trend. \input{sec/exp} \section{Conclusion} In this paper, we study the problem of predicting infection risks of COVID-19 in urban neighborhoods. We first construct a set of features incorporating human mobility data to characterize the demographic/socioeconomic status and spatial interactions of a residential neighborhood, then propose C-Watcher, a data-driven framework based on these features to early detect high-risk neighborhoods in a city ahead of local COVID-19 outbreaks. To improve infection risks identification in target cities, C-Watcher adopts adversarial learning algorithms that learn ``city-invariant'' features to boost generalizing knowledge witted in epicenter and build a reference city validation mechanism for hyperparameters selection. We conduct extensive experiments upon real-world data in the early stage of COVID-19 outbreaks from China to demonstrate the advantages of C-Watcher to early detect high-risk neighborhoods across cities and analyze the importance and effectiveness of explored features. \section*{Acknowledgment} We thank all reviewers for insightful comments. This work is supported in part by National Key R\&D Program of China (No. 2018YFB1402600), and in part by grant from the National Natural Science Foundation of China (Grant No. 91746301 and No. 71531001). Part of experiments in this paper was carried out using anonymous data and secure data analytics provided by Baidu Data Federation Platform (Baidu FedCube). For data accesses and usages, check http://fedcube.baidu.com/page\_en.html. \section{Appendix} \subsection{Feature Constructions}\label{apx:feacon} \subsubsection{Types of POIs} As mention in section \ref{sec:poi_radius}, we construct 15 POI radius features defined as the shortest distance between a neighborhood and one certain type POIs. The 15 types of POIs are as follows: \emph{hospital}, \emph{clinic}, \emph{campus}, \emph{kindergarten \& primary school \& secondary school} (we see these three types as a whole), \emph{bus stop}, \emph{subway station}, \emph{airport} , \emph{train station}, \emph{coach station}, \emph{shopping mall}, \emph{supermarket}, \emph{market}, \emph{shop}, \emph{police station}, \emph{scenic spots}. We also define another binary feature to directly reflect the perfect degree of basic living facilities and it will be assign “perfect” if there are all the following types of living facilities shown in Table \ref{table-apdx-facilities} within 1 km of the given neighborhood. The types of living facilities within 1 km are mainly selected according to a official document released by Ministry of Housing and Urban-Rural Development of China \footnote{http://www.mohurd.gov.cn/wjfb/201811/W020181130044801\\.pdf}. \subsubsection{Human Attributes Features} In section \ref{sec:demographic features}, we introduce 11 demographic features of a neighborhood based on distributions of residents according to the attributes shown in Table \ref{table-apdx-profile}. \input{sec/apdx-userprofile} \subsubsection{Transportation-Related Features} In section \ref{sec:Transport-related features}, we discuss transportation-related features from the perspectives of \textbf{T} (means of transportation) features, \textbf{OD} (types of visit venues and the distance between origin and destination venues) features and \textbf{OTD} (origin-transportation-destination pattern) features. The details of each feature are shown in Table \ref{table-apdx-OTD}. \subsection{Statistics of Datasets}\label{apx:sod} Table \ref{table-apdx-dataset} shows the statistics of high-risk neighborhoods of datasets used in our experiments. Note that the number of high-risk and low-risk neighborhoods of Wuhan is officially announced by local government. Thus, the total number of neighborhoods of Wuhan is small. \input{sec/apdx-datasets} \subsection{Hyperparameters}\label{apx:hyp} We evaluate cross-city early detection performance of our C-Watcher and baselines in section \ref{sec: hyperparameters}. Table \ref{table-apdx-hyper-baseline} presents the hyperparameters of each baseline. \subsection{Results of Early Detection}\label{apx:red} Table \ref{table-apdx-results} shows the high-risk neighborhoods detecting performance our C-Watcher and baselines. The ``target city - reference city'' relationship are ``Huizhou \& Guangzhou -- Shenzhen'', ``Shaoyang \& Yiyang -- Changsha'', ``Lianyungang \& Nanjing -- Shanghai'', ``Xuchang \& Anyang -- Zhengzhou'' and `` Chongqing \& Kunming -- Chengdu''. \subsection{Features Abbreviation and Full Name} \label{apx:fafn} In the analysis of features importance in section \ref{features-importances}, we use abbreviation to name our constructed features (see section \ref{sec:feature}) for simplicity. The abbreviation -- full name matches are presented in Table \ref{table-apdx-fullname}. \input{sec/apdx-facilities} \input{sec/apdx-OTD} \input{sec/apdx-results} \input{sec/apdx-hyper-baseline} \input{sec/apdx-fullname} \section{Cross-City Transfer Learning}\label{sec:transer frame work} \begin{figure}[t] \centering \includegraphics[width=0.48\textwidth]{figs/preprint-C-Watcher-frame.png} % \caption{Illustration of cross-city transfer learning model of C-Watcher.} \label{model-frame} \end{figure} In this section, we present the cross-city transfer learning model, which is a core component of C-Watcher, to improve the performance of early detection of high-risk neighborhoods via transferring the knowledge about COVID-19 infection from the epicenter to the target city. Usually, discrepancies always exist between different cities. Thus, we intend to learn city-invariant knowledge applicable to both epicenters and target cities, instead of those characteristics unique to epicenter cities. \subsection{Overview}\label{sec:framework overview} An overview of our proposed cross-city transfer learning model with four components is given in Figure \ref{model-frame}. The first component is a neural network encoder used to learn the representation of a neighborhood on the basis of three groups of features introduced in Section \ref{sec:feature}. However, the discrepancy of input distributions in different cities may lead to the gap between encoded representations of epicenter cities and target cities, which may severely disrupt the detection ability in target cities. Thus, we adopt adversarial learning by adding a discriminator component to identify whether the output of the encoder belongs to the target city or not. In addition, to ensure that the embedded feature of the encoder still keeps the ability to depict the residential neighborhoods, we exert two decoders which recover features of epicenter cities and target city respectively from the output of the encoder. Moreover, to achieve the prediction goal of C-Watcher, a classifier is also added to optimize the learned representations space and make it more related to the prediction of COVID-19 infection risks. In the following sections, we present our cross-city transfer learning model in detail by introducing how each component works. Here comes the notation of model input first. For both residential neighborhoods in epicenter cities and target cities, we have a feature vector consist of three groups: \begin{equation} \begin{split} \label{eq-2} \bm{n}^{E} &= cat(\bm{n}_{r}^{E}, \bm{n}_{h}^{E}, \bm{n}_{t}^{E})\\ \bm{n}^{T} &= cat(\bm{n}_{r}^{T}, \bm{n}_{h}^{T}, \bm{n}_{t}^{T}) \end{split} \end{equation} where $\bm{n}^E$ denotes all of the features of a residential neighborhood in epicenter cities. $\bm{n}_r^E$, $\bm{n}_h^E$ and $\bm{n}_t^E$ respectively denote the POI radius features, demographic features and transportation-related features of that residential neighborhood in epicenter cities. $\bm{n}^T$, $\bm{n}_r^T$, $\bm{n}_h^T$ and $\bm{n}_t^T$ similarly denote the corresponding features of residential ones in target city. The function $cat(\cdot)$ is the concatenating operation. \subsection{City-Invariant Representation Learning}\label{asversarial learning} Since learning unique characteristics of neighborhoods in epicenter cities brings little benefit for early risk detection in target cities, we propose a city-invariant representation learning method, which is inspired by the multi-mode adversarial representation learning methods \cite{mai2020modality,makhzani2015adversarial}. Here the encoder is a transformer of data distribution. Given the input vectors $\bm{n}^E$ and $\bm{n}^T$, we use $\widetilde{\bm{n}}^E$ and $\widetilde{\bm{n}}^T$ to denote the outputs of the encoder. Similar to \cite{mai2020modality} and \cite{makhzani2015adversarial}, the distributions transformation from the inputs to encoded representations can be presented as: \begin{equation} \begin{split} \label{eq-3} p(\widetilde{\bm{n}}^E,\Phi_e) &= \int_{\bm{n}^E} e(\widetilde{\bm{n}}^E |\,\bm{n}^E, \Phi_e)\,p(\bm{n}^E)\,d\bm{n}^E\\ p(\widetilde{\bm{n}}^T,\Phi_e) &= \int_{\bm{n}^T} e(\widetilde{\bm{n}}^T |\,\bm{n}^T, \Phi_e)\,p(\bm{n}^T)\,d\bm{n}^T \end{split} \end{equation} where $p(\cdot)$ denotes the data distribution and $e(\cdot \,,\Phi_e)$ represents the encoding distribution. $\Phi_e$ are parameters of the encoder, determining the projection space where the distributions of input data $p(\bm{n}^E)$ and $p(\bm{n}^T)$ are transformed into that of encoded representations $p(\widetilde{\bm{n}}^E,\Phi_e)$ and $p(\widetilde{\bm{n}}^T,\Phi_e)$. In general, $p(\widetilde{\bm{n}}^E,\Phi_e)$ and $p(\widetilde{\bm{n}}^T,\Phi_e)$ are different distributions characterizing different cities. It means that some features helpful in predicting COVID-19 infection risk may be unique to neighborhoods in epicenter cities unless we impose constraints on the encoder. To this end, we use adversarial learning to narrow the discrepancies between distributions $p(\widetilde{\bm{n}}^E,\Phi_e)$ and $p(\widetilde{\bm{n}}^T,\Phi_e)$ by adding a discriminator to distinguish whether the neighborhood comes from epicenter cities or the target city. In this way, the discriminator needs to do a binary classification task, in which it takes the encoded representations as inputs and aims to identify the inputs $\widetilde{\bm{n}}^E$ from epicenter cities as true but the inputs $\widetilde{\bm{n}}^T$ from target cities as false, while the encoder tries its best to confuse the discriminator to classify both of them as true. We can formulate the function of discriminator as: \begin{equation} \begin{split} \label{eq-4} D(\widetilde{\bm{n}}^E,\Phi_D) \rightarrow 1\\ D(\widetilde{\bm{n}}^T,\Phi_D) \rightarrow 0 \end{split} \end{equation} where $D(\,\cdot,\Phi_D)$ denotes the function of the discriminator which can be an MLP model that outputs the probability from 0 to 1. On the contrary, the encoder competes against the discriminator by: \begin{equation} \begin{split} \label{eq-5} D(\widetilde{\bm{n}}^E,\Phi_D) \rightarrow 1\\ D(\widetilde{\bm{n}}^T,\Phi_D) \rightarrow 1 \end{split} \end{equation} For this adversarial learning procedure, we use binary cross entropy (BCE) to define the loss function: \begin{align} \label{eq-6} \mathcal{L}_{al} &= \mathcal{L}_{diff}(\widetilde{\bm{n}}^E,\widetilde{\bm{n}}^T) + \mathcal{L}_{ch}(\widetilde{\bm{n}}^E,\widetilde{\bm{n}}^T)\\ \mathcal{L}_{diff} &= -[log(D(\widetilde{\bm{n}}^E)) + log(1-D(\widetilde{\bm{n}}^T)]\\ \mathcal{L}_{ch} &= -[log(D(\widetilde{\bm{n}}^E)) + log(D(\widetilde{\bm{n}}^T)] \end{align} where $D(\widetilde{\bm{n}}^E)$ is used to represent $D(\widetilde{\bm{n}}^E, \Phi_D)$ in simplicity and so does $D(\widetilde{\bm{n}}^T)$. The differentiation loss $\mathcal{L}_{diff}$ guides discriminator to predict $\widetilde{\bm{n}}^E$ as true (epicenter cities) but $\widetilde{\bm{n}}^T$ as false (target city), while the encoder tries to learn features that are common between epicenter cities and target city to hinder discriminator from distinguishing successfully, under the effects of cheat loss $\mathcal{L}_{ch}$. The adversarial procedure will finally reach an equilibrium situation where the discriminator could no longer distinguish whether the encoded representations come from epicenter cities or target city, then the encoder is able to extract ``city-invariant'' features from raw inputs $\bm{n}^E$ and $\bm{n}^T$. In this case, discrepancies between cities decrease and the experience which helps predict infection risks in epicenter cities can make more sense in target cities. \subsection{Embedding Space Constraints}\label{space constrains} A problem about city-invariant representation learning is that, if no regulations and restrictions are imposed on the embedding space of the encoder, the encoded representations of epicenter cities and target cities may only be similar in distribution but fail to retain useful information for identifying high-risk neighborhoods. We solve this problem with multi-task learning strategy by additionally exerting an auto encoder-decoder features reconstruction component, as well as a COVID-19 infection risks prediction component. The reconstruction component consists of two decoders (one for residential neighborhoods in epicenter cities and another one for residential neighborhoods in the target city) which take the encoded representations as inputs. The decoding operation can also be considered as a distribution transformation like encoding: \begin{equation} \begin{split} \label{eq-9} p(\widehat{\bm{n}}^E,\Phi_d^E) &= \int_{\widetilde{\bm{n}}^E} d^E(\widehat{\bm{n}}^E |\,\widetilde{\bm{n}}^E, \Phi_d^E)\,p(\widetilde{\bm{n}}^E)\,d\widetilde{\bm{n}}^E\\ p(\widehat{\bm{n}}^T,\Phi_d^T) &= \int_{\widetilde{\bm{n}}^T} d^T(\widehat{\bm{n}}^T |\,\widetilde{\bm{n}}^T, \Phi_d^T)\,p(\widetilde{\bm{n}}^T)\,d\widetilde{\bm{n}}^T \end{split} \end{equation} where $\widehat{\bm{n}}^E$, $\widehat{\bm{n}}^T$ denote the reconstructed outputs of decoders from $\widetilde{\bm{n}}^E$ and $\widetilde{\bm{n}}^T$ respectively, and $d^E(\widehat{\bm{n}}^E |\,\widetilde{\bm{n}}^E,\Phi_d^E)$ represents the epicenter cities decoder function with parameters $\Phi_d^E$, while $d^T(\widehat{\bm{n}}^T |\,\widetilde{\bm{n}}^T,\Phi_d^T)$ is similar but for the target city. Aiming to approximate decoded representations to the original inputs ($\widehat{\bm{n}}^E \rightarrow \bm{n}^E$ and $\widehat{\bm{n}}^T \rightarrow \bm{n}^T$), we use mean square error to define reconstruction loss function: \begin{align} \label{eq-10} \mathcal{L}_{rec} &= \mathcal{L}_{rec}^E + \mathcal{L}_{rec}^T\\ &= \|\,\widehat{\bm{n}}^E,\bm{n}^E\|_2 + \|\,\widehat{\bm{n}}^T,\bm{n}^T\|_2 \end{align} Optimized by the reconstruction loss above, the encoder-decoder framework ensures that the embedding space is still characterizing a residential neighborhood. Moreover, considering that our ultimate objective is to detect latent high-risk neighborhoods, we add a classifier to identify COVID-19 infection risks in epicenter cities upon the learned encoded representations. The classification problem can be defined as : \begin{equation} \label{eq-11} C(\widetilde{\bm{n}}^E,\Phi_c) \rightarrow y^E,\, y^E \in \{0,1\} \end{equation} where $C(\cdot,\Phi_c)$ denotes the function of MLP classifier with parameter $\Phi_c$. This is also a binary classification task and we use BCE to define the classification loss function: \begin{equation} \label{eq-12} \mathcal{L}_{cl} \!=\! -y^{E}\!log(C(\widetilde{\bm{n}}^E\!\!,\Phi_c)) \!-\! (1 \!-\! y^{E})log(1 \!-\! C(\widetilde{\bm{n}}^E\!\!,\Phi_c)) \end{equation} The classification loss transmits the known information carried by label $y^E$ to encoder and classifier, which is COVID-19 infection risks of neighborhoods in epicenter cities. It achieves the goals to restrict the encoded representations to be instructive in high-risk neighborhood identification. All in all, loss functions generated from all the three components of discriminator, decoders and classifier will act on the encoder and optimize the embedding space in our proposed cross-city transfer model. The total loss function can be expressed qualitatively as: \begin{equation} \label{eq-13} \mathcal{L} = \lambda_{diff}\mathcal{L}_{diff} + \lambda_{ch}\mathcal{L}_{ch} + \lambda_{rec}\mathcal{L}_{rec} + \lambda_{cl}\mathcal{L}_{cl} \end{equation} In model training, the adversarial model is optimized in an alternate mode. We use differentiation loss $\mathcal{L}_{diff}$ to optimize the discriminator first to improve its discriminatory ability to neighborhoods in both epicenter cities and that target city, which also leads to the rise of cheat loss. Then we apply cheat loss $\mathcal{L}_{ch}$ combined with $\mathcal{L}_{rec}$ and $\mathcal{L}_{cl}$, to guide the encoder to optimize its parameters in a direction where demands to learn city-invariant, informative and risk-discriminative features are all taken into consideration. Together with the encoder, the decoders and classifier update themselves based on $\mathcal{L}_{rec}$ and $\mathcal{L}_{cl}$, respectively. \subsubsection{Reference City Validation Mechanism.}\label{sec: anchor} \begin{figure}[t] \centering \includegraphics[width=0.3\textwidth]{figs/anchor.pdf} \caption{Diagram of reference city validation mechanism.} \label{anchor-frame} \end{figure} Another problem of C-Watcher is how to select the best hyperparameters to train the model. Here we build a reference city validation mechanism to tune hyperparameters. The illustrated diagram is shown in Figure \ref{anchor-frame}. Reference city in our paper can be epicenter cities, and can also be some cities with COVID-19 outbreak but not so serious as epicenters. We train the C-Watcher model on epicenter cities set, and use ground truth data of the reference city as validation data to choose the hyperparameters. Then we evaluate the early detection performance in target cities geographically close to that reference city. In this case, we ensure that our trained model to detect latent high-risk neighborhoods in a target city with best hyperparameters, without any prior information related to COVID-19 confirmed cases and spreading trend. \section{Experiments}\label{sec:experiments} \subsection{Datasets and Settings}\label{sec:datasets and settings} \subsubsection{Dataset construction.}\label{sec: dataset construction} The constructed datasets simulate a common outbreak pattern in a country. In the scenarios, there is a set of epicenter cities in the country (like Wuhan in China), and a few reference cities (see Section \ref{sec: anchor}) which have some confirmed cases. The C-Watcher can be trained on the epicenter cities and reference cities datasets, and then be used to make early detection of high-risk residential neighborhoods in the rest of cities in the country.\footnote{The code can be found at \url{https://github.com/PaddlePaddle/Research/tree/master/ST_DM/AAAI2021-CWatcher/}.} In this evaluation, all the datasets are built based on 16 cities in China which consists of one dataset from epicenter city, 5 evaluation datasets from selected reference cities, and 10 test datasets from other cities. The epicenter city dataset is constructed based on Wuhan, which has the largest number of confirmed cases in China and is well-recognized as the epicenter of the COVID-19 outbreak in China. The five selected reference cities, Shenzhen, Changsha, Chengdu, Shanghai and Zhengzhou, are key cities in their provinces and they are also evenly situated in different geographical regions of China. For each reference city, we also construct two test datasets from two cities geographically closed to them. The full list of test cities is in the Appendix \ref{apx:sod}. The POI data and user profile data of all the cities are both collected by the first week of March 2020. The human mobility data are collected from January 1, 2020 to March 3, 2020. We also make a great effort to build the ground-truth dataset. For the Wuhan dataset, we manually collected all the high-risk residential neighborhoods (released on February 24, 2020) and low-risk residential neighborhoods (released on March 6, 2020) which are officially published by the local government. After data cleaning and feature alignment, there are 336 high-risk neighborhoods and 715 low-risk neighborhoods. The statistics of high-risk neighborhoods in other cities datasets are listed in Appendix \ref{apx:sod}. For datasets of other cities, we label the neighborhoods with at least one confirmed case as high-risk while others as low-risk, based on the public COVID-19 patients dataset by \cite{fu2020aware}. In order to tune hyperparameters for baselines, we split Wuhan dataset into three folds as train, validation and test data by a 0.7:0.15:0.15 ratio. The hyperparameters tuning for C-Watcher is done by reference city validation mechanism (see Section \ref{sec: anchor}). \subsubsection{Baselines.}\label{sec:baseline} Since we are the first to study the COVID-19 high-risk neighborhoods early detection problem, there is no direct competitor of C-Watcher. Thus, we compare C-Watcher with classical machine learning methods of Multi-Layer Perceptron (MLP), Support Vector Machine (SVM), XGBoost (XGB) and Lasso Logistic Regression (Lasso-R). We use the dataset from epicenter city to train the baselines, and make a prediction on the datasets of test cities. \subsubsection{Metrics.}\label{sec:metrics} Since the detection of high-risk neighborhoods is an imbalance binary classification task (high-risk neighborhoods are much less than low-risk ones), we mainly evaluate the performance by AUC (Area under the ROC Curve), which reflects model performance within different discrimination thresholds \cite{manning2008introduction, fu2020aware}. In addition, we also calculate the p-value by pairwise t-test between baselines and C-Watcher to show the statistical significance of the evaluation results. \subsubsection{Optimization and hyperparameters tuning.} \label{sec: hyperparameters} We optimize C-Watcher by Adam optimizer. The main hyperparameters of C-Watcher, including weights of the loss function ($\lambda_{ch}$, $\lambda_{diff}$, $\lambda_{rec}$ and $\lambda_{cl}$ ), learning rate and hidden size of the neural network of each component are determined by grid search method, with batch size fixed as 64. \subsection{Performance Evaluation of Early Detection}\label{sec: main-exp} \input{sec/table_main_exp} We evaluate the performance of C-Watcher and its baselines for early detection of high-risk neighborhoods on test datasets of the 10 target cities. In Table \ref{table_main_exp}, the overall column shows the average AUC of the 10 cities. We can see that C-Watcher can improve the AUC by 8.18\% over the best baseline (SVM and MLP). We also conduct pairwise t-test between the C-Watcher and each baseline. The p-values in Table \ref{table_main_exp} demonstrate that C-Watcher can achieve significantly better performance than other baselines. We also show the prediction performance on test datasets of five target cities in Table \ref{table_main_exp}. Each target city corresponds to one reference city. We can see that the improvement by C-Watcher over baselines in different cities is different. For example, the improvement by C-Watcher over the best baseline on Shaoyang is 14.57\% (i.e., C-Watcher (0.6433) vs. SVM (0.5615)); but the one by C-Watcher over the best baseline on Xuchang is about 0\% (the AUC of C-Watcher is almost the same with other baselines). It is an interesting problem to investigate what factors impact the performance of transfer learning of C-Watcher. A possible reason is that some geographically closed cities are not similar, thus the reference city cannot help to select the best hyperparameters for transfer learning. We leave this problem as a further research investigation. We put the prediction performance of all the ten cities in the Appendix \ref{apx:red}. \subsection{Feature Importance}\label{sec:feature importance} \begin{figure}[b!] \centering \includegraphics[width=0.45\textwidth]{figs/importance_barh1.pdf} \caption{The top-20 most important features for high-risk neighborhoods detection.} \label{features-importances} \end{figure} Here we conduct a feature importance analysis to discuss possible characteristics of neighborhoods leading to the high risk for infection. We use Lasso Logistic Regression (Lasso-R) on epicenter Wuhan dataset to select the top-20 important features according to the absolute coefficient value, which is illustrated in Figure \ref{features-importances}. The full name of each feature is listed in the Appendix \ref{apx:fafn}. The feature importance analysis reveals several insightful and interesting points for preventing the epidemic. For POI radius features, except the effect of perfect and poor living facility of a neighborhood (which are denoted by ``P:PFLF'' and ``P:PRLF'' in Figure \ref{features-importances}), the coefficient of ``P:RTS'' indicates that the long distance to a train station can reduce the risk of the neighborhood. For the demographic features, except the high population density (denoted by ``D:PD''), the long average commute distance (denoted by ``D:ACD'') also increases the risk of the neighborhood. For the transportation-related features, we find that the percentage of travelling on walk (denoted by ``T:TW'') can reduce the risk of the neighborhood by a large margin. We believe such analysis can help us identify factors for high-risk neighborhoods, and provide insightful suggestions on preventing the epidemic of COVID-19 in future. \subsection{Effectiveness of Feature Groups}\label{sec:feature effectiveness} In this section, we verify the effectiveness of 3 groups of hand-crafted features. In specific, we separately evaluate the performance of each group of features in detecting high/low-risk neighborhoods, then we compare them with the performance of taking all 3 groups of features together as inputs. All the comparative experiments for feature effectiveness are conducted by MLP on the epicenter Wuhan dataset. As we can see from Table \ref{table:feature effectiveness}, all the three groups of features can positively classify the high-risk and low-risk neighborhoods. More importantly, the combination of these three groups certainly improves the model's overall performance, which proves complementary among the three groups of features. \begin{table}[htb] \centering \begin{tabular}{c|c} \hline \rule{0pt}{10pt} Feature groups & AUC \\ \hline \rule{0pt}{10pt} POI Radius & 0.8033 \\ \hline \rule{0pt}{10pt} Demographic & 0.7579 \\ \hline \rule{0pt}{10pt} Transportation-Related & 0.7414 \\ \hline \rule{0pt}{10pt} All three Groups & \textbf{0.8458} \\ \hline \end{tabular} \caption{Detection performance comparison of MLP with different feature groups on Wuhan dataset.} \label{table:feature effectiveness} \end{table} \section{Features for Neighborhood Detection}\label{sec:feature} In this section, we present how to construct features from mobility data to characterize a residential neighborhood for early risk detection. We first introduce the data source used in our framework, and then three groups of constructed features are briefly discussed which are Point of Interest (POI) radius features (see Section \ref{sec:poi_radius}), demographic features (see Section \ref{sec:demographic features}) and transportation-related features (see Section \ref{sec:Transport-related features}), respectively. More details about the feature construction can be found in the Appendix in our submitted supplementary materials. The feature construction is mainly based on three data sources: POI basic property data, user profile data and human mobility data. POI basic property data contains the basic information of a POI, such as name, coordinates and types, which provides many semantic information for a POI \cite{10.1145/3394486.3403318,yuan2020spatio,hu2019we}. This data enable us to analyze the spatial relationship between neighborhoods and different types of POIs, such as hospitals, schools and bus stops \cite{Li2020competitive}. The user profile data are obtained from a user profile platform of Baidu which can return profile features for almost all internet users in China, such as gender, age and educational level. Human mobility data, collected from Baidu Maps in China, record search and transportation behaviors of the map users. \subsection{POI Radius Features}\label{sec:poi_radius} Here we introduce how to compute a group of POI radius features for a residential neighborhood based on POI basic property data. The intuition for this feature group is that basic living facilities around a residential neighborhood may have a correlation with the probability of its residents being infected by COVID-19. For example, a neighborhood lacking basic living facilities may face a high risk, for the residents may passively go further away for basic living needs and face greater infection risks. Moreover, neighborhoods with poor living facilities often lack good property management, which may also lead to high infection risks. To describe these living facilities related characteristics, we construct 15 POI radius features. Each of them is defined as the shortest distance between the neighborhood and one certain type of POIs. All used POI types are listed in the Appendix. Meanwhile, we define an additional binary feature to directly represent the perfect degree of living facilities. The value of this feature will be assigned as ``perfect'' if a set of basic living facilities (e.g. hospital, bus stop and so on) are all within $1 km$ of the given neighborhood. Otherwise, it is assigned ``poor''. The list of basic living facilities is also shown in the Appendix. We collect the high-risk and low-risk neighborhoods data in Wuhan city in China which is officially announced by the local government. Figure \ref{infrastructure} presents the ratio distribution of high-risk and low-risk neighborhoods grouping by this ``perfect-poor'' facility label in Wuhan data. As we can see from Figure \ref{infrastructure}, for the neighborhoods with feature value as ``perfect'', the ratio of low-risk neighborhoods and high-risk ones is $0.57:0.44$; whereas the ratio of them for ``poor'' ones is $0.43:0.56$. It indicates that more high-risk neighborhoods have poor living facilities, while the low-risk neighborhoods are just the opposite. \begin{figure} \centering \subfigure[]{ \label{infrastructure} \includegraphics[width=0.48\columnwidth]{figs/infrastructure.pdf}} \hspace{-1mm} \subfigure[]{ \label{population_density} \includegraphics[width=0.48\columnwidth]{figs/population_density_norm.pdf}} \caption{Features of living facilities and population density visual analysis.} \label{feature_distribution} \end{figure} \subsection{Demographic Features}\label{sec:demographic features} Next, we present the demographic features of a residential neighborhood. At first, given that the COVID-19 is easy to transmit in a person-to-person way \cite{liu2020aerodynamic}, it is necessary to take into account population density for infection risks prediction. As Figure \ref{population_density} illustrates, on average high-risk neighborhoods do have a higher population density than low-risk neighborhoods in Wuhan city. We also compute average commute distance as a feature for each neighborhood since residents with long commutes have high infection risks. Moreover, different groups of residents may face different risk levels in a neighborhood. For example, old people and children are easier to be infected. And residents with higher educational levels may pay more attention to scientific prevention. Hence, we construct 11 features based on the distribution of residents according to different human attributes. We present each of these features as a vector of histogram statistics of residents' distribution. The full list of such attributes is provided in the Appendix. \subsection{Transportation-Related Features}\label{sec:Transport-related features} We also extract features of transportation-related behaviors from human mobility data to help predict infection risks. There have been some studies to prove that transportation-related behaviors have a close relationship with COVID-19 contagion spreading \cite{huang2020kddtransp}. The transportation-related behaviors typically are recognized as the origin-transportation-destination (OTD) information\cite{8708931,xu2016taxi,Polestar}. Thus, we consider detailed features from the perspectives of T (transportation), OD (origin \& destination venues) and OTD (origin-transportation-destination pattern). All the features are extracted from the search and transportation data of Baidu Maps in a certain time period. Previous studies have shown that map search behavior is a leading indicator and predictor for crowd dynamics \cite{zhou2018early}. A vector in which the value of each element equals the corresponding ratio of transportation means, mainly including walk, bicycle, public transit and private vehicle, is used to depict the ``T feature'' for a residential neighborhood. The ``OD feature'' consists of types of visit venues and the distance between origin and destination venues. We classify the destination venues according to their types (e.g. hospital, restaurant, hotel, and school) and compute the proportion of each type. We also extract origin-destination distance and categorize it into different distance buckets. The proportions of different distance buckets for the neighborhood are also formed as a feature vector. Moreover, since the OTD (origin-transportation-destination) patterns most directly reflect human mobility, we collect the top-20 hottest travel patterns from all the cities in our dataset, which is a triplet tuple composed of the type of origin venue (residential area), means of transportation and type of destination venue. The histogram distribution of these top-20 OTD travel patterns of each neighborhood is treated as ``OTD'' features. More details about the transportation-related features can be found in the Appendix. \section{Introduction} The novel coronavirus disease (COVID-19), which has been officially announced as a pandemic by the World Health Organization (WHO), is perhaps the most serious public health emergency over the past decades. The coronavirus continues to spread around the globe, which challenges the governments and medical systems all over the world. While metropolitan-wide lockdown has demonstrated its effectiveness as a nonpharmaceutical intervention in several countries, the cost of such measures, including unemployment, economic crash, and social anxiety, makes it a tough decision on behalf of administrators. A compromised solution is to place containment measures onto a subset of areas in a city to stop or slow down the spread of COVID-19 while minimizing the social and economic cost. To precisely distinguish the high-risk areas from the city, the spatial distribution of confirmed cases has been used as the key criterion for the potential containment measures in a data-driven fashion. While the spatial statistics of confirmed cases work, the time consumption and the granularity of data acquisition significantly lower the efficiency and effectiveness of such methods. For example, the incubation period of COVID-19 is around 5--6 days on average, but it could last as long as 14 days. During such a period, a community or neighborhood would have been already invaded by a small number of asymptotic carriers who might not be with any clinical symptoms. When a small number of cases being confirmed, the community and its surrounding neighborhoods may have fallen into COVID-19 for a long time. Furthermore, though the mobility of confirmed patients is usually well restricted, the asymptomatic carriers with no symptoms would still spread the virus to where he/she has gone. When fine-grained mobility traces are not available, the administrators can only place containment measures to places in coarse-grained. To tackle the technical issues, in this paper, we propose C(OVID)-Watcher, a novel framework to support early detection of high-risk residential neighborhoods for fighting against the spread of COVID-19. Our intuition is to incorporate human mobility data, so as to (1) characterize the socioeconomic and demographic status of every neighborhood~\cite{borjas2020demographic,huang2020kddtransp} based on ``how residents move''~\cite{renso2013you} and (2) unfold the spatial interactions~\cite{jiang2020spatial} and potential influences on COVID-19 caused by the mobility of massive asymptotic carriers. Specifically, C-Watcher includes a mobility data-driven machine learning model that screens every neighborhood in a target city and predicts the infection risks, prior to the spread of COVID-19 from epicenters to the city. In addition to the use of mobility-related features, we also hope to generalize the evidence already witted in the epicenter for screening the risk of neighborhoods in the target city, prior to or in the early stage of local outbreaks. To achieve the goal, a core component of C-Watcher is a novel cross-city transfer learning model that transfers knowledge about COVID-19 infections from the epicenter to the target city, while cities are quite different in a large number of domains, ranging from living, foods, transportation, and residences. All in all, we have made three contributions as follows. \begin{itemize} \item Through extensive data analytics, we explore a set of empirical features related to long-term/regular human mobility patterns (before the COVID-19). With such long-term mobility features, the socioeconomic and demographic status, as well as the spatial interactions among neighborhoods could be well characterized. In this way, one can easily distinguish high-risk neighborhoods from the urban area and predict the potential risks for infection, with respect to the two factors. \item While these features are with certain discriminative information for risk prediction, they also involve some city-specific characteristics. For example, the popular choices for transport modes in different cities vary. Such city-specific characteristics burden the use of mobility-related features to transfer the knowledge obtained in the epicenter to the target city. To generalize the knowledge transfer, C-Watcher adopts a novel adversarial encoder-decoder framework to learn the ``city-invariant'' representations from the mobility-related features for prediction. \item To validate C-Watcher, we collect and construct real-world datasets for high-risk neighborhood detection based on the publicly available information from the web and human mobility traces from the largest online map service in China. We conduct extensive experiments for evaluation. The results demonstrate that C-Watcher can accurately predict the potential risk of massive residential neighborhoods in a large number of Chinese cities. With large datasets, C-Watcher makes insightful suggestions on preventing the epidemic of COVID-19 alike for different residential neighborhoods via feature importance. \end{itemize} \section{Notations and Related Work} In this section, we first introduce the basic notations used throughout this paper, and then we formally formulate the research problem for early detection of high-risk neighborhoods. Last, we review the studies that are relevant to our work with the most related work discussed. \paragraph{Notations and Formulation.} We use $\bm{n}$ to denote the features of a residential neighborhood which will be presented in Section \ref{sec:feature}, and use $y$ to denote the binary label meaning whether the neighborhood is high risky ($y=1$) or not ($y=0$). The detection problem can be defined as: $f(\bm{n}) \rightarrow y$ where the function $f(\cdot)$ can be any machine learning model like Multi-Layer Perceptron (MLP). The objective of C-Watcher is to make early detection of high-risk neighborhoods without epidemic outbreaks. Instead of relying on the confirmed infection cases to make a prediction which deemed to be time-delayed like \cite{fu2020aware}, we assume that the COVID-19 epidemic only outbroke in epicenter cities (such as Wuhan in China) and no prior knowledge of confirmed cases, spreading trend or known hazard neighborhoods in target cities can be referred to. Such a cross-city prediction problem of latent high risky residential neighborhoods can be formulated as: \begin{equation} \label{eq-1} f_{cross}(\bm{n}^{T}|\,\{(\bm{n}_i^{E}\!, y_i^{E})\}) \rightarrow y^{T} \end{equation} where $\bm{n}^T$ and $y^T$ denote the features and binary label of a residential neighborhood in the target city. $\bm{n}_i^E$ and $y_i^E$ denote the features and label of a neighborhood from epicenter cities set. Hereafter, we omit subscript $i$ for simplicity. The $f_{cross}(\cdot)$ is a cross-city transfer learning model which is trained without ground-truth information in the target city. \paragraph{Related Work.} Aiming to fight against the COVID-19 pandemic, researchers in the computer science community carried out many studies from several perspectives recently. For instance, \citet{huang2020kddtransp} exhibit that user transportation-related behaviors in China have indeed been impacted by the containment measures during the COVID-19 pandemic. There are also a few studies \cite{huang2020quantifying,xiong2020understanding,liu2020Investigation} investigating the human mobility, the local economy, and the information acquisition during the COVID-19 outbreak in China while most of these studies remain at city level. Some studies also demonstrate the effectiveness of mobility data for controlling the spread of COVID-19. \citet{vollmer2020using} exploit a Bayesian semi-mechanism model with mobility data to show the effectiveness to slow down the spread of the virus by constraints on individual movements and social interactions. Based on the integration of mobility data and the global epidemic model \cite{balcan2009multiscale}, a study also reveals the effectiveness of fine-grained targeted mobility control policies towards the COVID-19 pandemic \cite{hao2020understanding}. The mobility data can also be integrated with compartmental models in epidemiology (like Susceptible-Exposed-Infected-Recovered (SEIR) model) \cite{ghamizi2020data} to better predict the epidemic dynamics. \paragraph{Discussion.} From the problems and methodologies perspectives, the most relevant work to our study includes~\cite{fu2020aware} and~\cite{xu2019deep,peng2019cm,mai2020modality}. Compared to~\cite{fu2020aware}, which smooths the confirmed cases of infections over spatial domains and predicts hazard areas during the COVID-19 outbreaks using simple spatial features like distance, C-Watcher system tackles the time delay and coarse-grained granularity issues and can early detect the high-risk residential neighborhoods even before the outbreaks, through leveraging features derived from long-term/regular human mobility patterns. In terms of methodologies, though a great number of algorithms have been proposed for adversarial representation learning~\citet{makhzani2015adversarial}, adversarial metric learning~\cite{xu2019deep} and cross-modalities~\cite{mai2020modality}, our work is the first to study the city-invariant representation learning through Generative Adversarial Networks \cite{goodfellow2014generative} in the context of urban computing and COVID-19 prediction.

ce460c4421d73e33fd931fdf73a48fc22ee7eaf8

\section{Introduction} \label{intro} Systems with time-delays, or delay-differential equations (DDE), play an important role in modeling various natural phenomena and technological processes \cite{Stepan1989,Hale1993,Diekmann1995a,Erneux2009,Smith2010,Erneux2017,Yanchuk2017,Krisztin2008}. In optoelectronics, delays emerge due to finite optical or electric signal propagation time between the elements \cite{Vladimirov2005,Erzgraber2006,DHuys2008,Vicente2008,Fiedler2008,Wolfrum2010,Yanchuk2010a,Soriano2013,Oliver2015,Marconi2015,Puzyrev2016,Yanchuk2019}. Similarly, in neuroscience, propagation delays of the action potentials play a crucial role in information processing in the brain \cite{Foss2000,Wu2001,Izhikevich2006,Stepan2009,Deco2009a,Perlikowski2010a,Popovych2011,Kantner2013}. Machine Learning is another rapidly developing application area of delay systems \cite{Paugam-Moisy2008,Appeltant2011,Martinenghi2012,Appeltant2012a,Larger2012,Brunner2013,SCH13l,Toutounji2014,Grigoryeva2015,Penkovsky2017,Larger2017,Harkhoe2019,Stelzer2019,Hart2019,Koster2020,Koester2020,Argyris2020,Goldmann2020,Sugano2020}. It is shown recently that DDEs can successfully realize a reservoir computing setup, theoretically \cite{Hart2017,Keuninckx2017,Hart2019,Stelzer2019,Koster2020,Stelzer2020,Koester2020,Goldmann2020}, and implemented in optoelectronic hardware \cite{Appeltant2011,Appeltant2012a,Larger2017}. In time-delay reservoir computing, a single DDE with either one or a few variables is used for building a ring network of coupled maps with fixed internal weights and fixed input weights. In a certain sense, the network structure emerges by properly unfolding the temporal behavior of the DDE. In this paper, we explain how such an unfolding appears, not only for the ring network as in reservoir computing but also for arbitrary networks of coupled maps. In \cite{Hermans2015}, a training method is proposed to modify the input weights while the internal weights are still fixed. Among the most related previous publications, Hart and collaborators unfold networks with arbitrary topology from delay systems \cite{Hart2017,Hart2019}. Our work extends their results in several directions, including varying coupling weights and applying it to a broader class of delay systems. The networks constructed by our method allow for a modulation of weights. Hence, they can be employed in Machine Learning applications with weight training. In our recent paper \cite{Stelzer2020}, we show that a single DDE can emulate a deep neural network and perform various computational tasks successfully. More specifically, the work \cite{Stelzer2020} derives a multilayer neural network from a delay system with modulated feedback terms. This neural network is trained by gradient descent using back-propagation and applied to machine learning tasks. As follows from the above-mentioned machine learning applications, delay models can be effectively used for unfolding complex network structures in time. Our goal here is a general description of such networks. While focusing on the network construction, we do not discuss details of specific machine learning applications such as e.g., weights training by gradient descent, or specific tasks. The structure of the paper is as follows. In Sec.~\ref{sec:general-case} we derive a feed-forward network from a DDE with modulated feedback terms. Section~\ref{sec:recurrent-network} describes a recurrent neural network. In Sec.~\ref{sec:semilinear-case}, we review a special but practically important case of delay systems with a linear instantaneous part and nonlinear delayed feedback containing an affine combination of the delayed variables; originally, these results have been derived in \cite{Stelzer2020}. \section{From delay systems to multilayer feed-forward networks} \label{sec:general-case} \subsection{Delay systems with modulated feedback terms} \label{subsec:delay-system} Multiple delays are required for the construction of a network with arbitrary topology by a delay system \cite{Hart2017,Hart2019,Stelzer2020}. In such a network, the connection weights are emulated by a modulation of the delayed feedback signals \cite{Stelzer2020}. Therefore, we consider a DDE of the following form \begin{align}\label{eq:general-system} \dot{x}(t) = f(x(t), z(t), \mathcal{M}_1(t)x(t-\tau_1),\ldots , \mathcal{M}_D(t)x(t-\tau_D)), \end{align} with $D$ delays $\tau_1,\dots,\tau_D$, a nonlinear function $f$, a time-dependent driving signal $z(t)$, and modulation functions $\mathcal{M}_1(t), \ldots , \mathcal{M}_D(t)$. System \eqref{eq:general-system} is a non-autonomous DDE, and the properties of the functions $\mathcal{M}_d(t)$ and $z(t)$ play an important role for unfolding a network from \eqref{eq:general-system}. To define these properties, a time quantity $T>0$ is introduced, called the \textit{clock cycle}. Further, we choose a number $N$ of grid points per $T$-interval and define $\theta := T/N$. We define the clock cycle intervals $$I_\ell := ((\ell-1)T, \ell T],\ \ell = 1, \ldots ,L,$$ which we split into smaller sub-intervals $$I_{\ell, n} := ((\ell -1)T + (n-1)\theta, (\ell - 1)T + n\theta], \ n=1,\ldots ,N,$$ see Fig.~\ref{fig:first}. We assume the following properties for the delays and modulation functions:\\ \textbf{Property (I):} The delays satisfy $\tau_d = n_d \theta,\ d=1,\ldots ,D$ with natural numbers $0 < n_1 < \cdots < n_D < 2N$. Consequently, it holds $0 < \tau_1 < \cdots < \tau_D < 2T$. \\ \textbf{Property (II):} The functions $\mathcal{M}_d(t)$ are step-functions, which are constant on the intervals $I_{\ell,n}$. We denote these constants as $v^\ell_{d,n}$, i.e. $$ \mathcal{M}_d(t) = v^\ell_{d,n} \quad \text{for} \quad t\in I_{\ell,n}. $$ \begin{figure} \centering \includegraphics{solution-and-modulation} \caption{ Illustration of the clock cycle intervals $I_\ell$ and sub-intervals $I_{\ell,n}$. The node $x^\ell_n$ (blue dot) is defined by the value of the solution $x(t)$ of system~\eqref{eq:general-system} (blue line) at the time point $t=(\ell -1)T+n\theta$. The modulation function $\mathcal{M}_d(t)$ is a step function with constant values $v^\ell_{d,n}$ on the intervals $I_{\ell ,n}$. } \label{fig:first} \end{figure} In the following sections, we show that one can consider the intervals $I_\ell$ as layers with $N$ nodes of a network arising from the delay system~\eqref{eq:general-system} if the modulation functions $\mathcal{M}_d(t)$ fulfill certain additional requirements. The $n$-th node of the $\ell$-th layer is defined as \begin{align} \label{eq:nodes} x^\ell_n := x((\ell - 1)T + n\theta), \quad n = 1,\ldots ,N, \ \ell =1,\ldots ,L, \end{align} which corresponds to the solution of the DDE \eqref{eq:general-system} at time point $(\ell - 1)T + n\theta$. The solution at later time points $x^{\ell'}_{n'}$ with either $\ell'>\ell$ or $n'>n$ for $\ell'=\ell$ depends, in general, on $x^\ell_n$, thus, providing the interdependence between the nodes. Such dependence can be found explicitly in some situations. The simplest way is to use a discretization for small $\theta$, and we consider such a case in the following Sec.~\ref{subsec:discretization}. Another case, when $\theta$ is large, can be found in \cite{Stelzer2020}. Let us remark about the initial state for DDE~\eqref{eq:general-system}. According to the general theory~\cite{Hale1993}, in order to solve an initial value problem, an initial history function $x_0(s)$ must be provided on the interval $s \in [-\tau_D, 0]$, where $\tau_D$ is the maximal delay. In terms of the nodes, one needs to specify $x_n^\ell$ for $n_D$ ``history'' nodes. However, the modulation functions $\mathcal{M}_d(t)$ can weaken this requirement. For example, if $\mathcal{M}_d(t) = 0$ for $t \leq \tau_d$, then it is sufficient to know the initial state $x(0) = x_0^1 =x_0$ at a single point, and we do not require a history function at all. In fact, the latter special case has been employed in \cite{Stelzer2020} for various machine learning tasks. \subsection{Disclosing network connections via discretization of the DDE} \label{subsec:discretization} Here we consider how a network of coupled maps can be derived from DDE \eqref{eq:general-system}. Since the network nodes are already introduced in Sec.~\ref{subsec:delay-system} as $x_n^\ell$ by Eq.~\eqref{eq:nodes}, it remains to describe the connections between the nodes. Such links are functional connections between the nodes $x_n^\ell$. Hence, our task is to find functional relations (maps) between the nodes. For simplicity, we restrict ourselves to the Euler discretization scheme since the obtained network topology is independent of the chosen discretization. Similar network constructions by discretization from ordinary differential equations have been employed in \cite{Haber2017,Lu2018,Chen2018}. We apply a combination of the forward and backward Euler method: the instantaneous system states of~\eqref{eq:general-system} are approximated by the left endpoints of the small-step intervals of length $\theta$ (forward scheme). The driving signal $z(t)$ and the delayed system states are approximated by the right endpoints of the step intervals (backward scheme). Such an approach leads to simpler expressions. We obtain \begin{align}\label{eq:euler} x^\ell_n = x^\ell_{n-1} + \theta f(x^\ell_{n-1},z(t^\ell_n), \mathcal{M}_1(t^\ell_n)x(t^\ell_n-\tau_1),\ldots ,\mathcal{M}_D(t^\ell_n)x(t^\ell_n-\tau_D)) \end{align} for $n = 2,\ldots,N$, where $t^\ell_n := (\ell - 1)T + n\theta$, and \begin{align}\label{eq:euler1} x^\ell_1 = x^{\ell-1}_N + \theta f(x^{\ell-1}_N, z(t^\ell_1), \mathcal{M}_1(t^\ell_1)x(t^\ell_1-\tau_1),\ldots ,\mathcal{M}_D(t^\ell_1)x(t^\ell_1-\tau_D)) \end{align} for the first node in the $I_\ell$-interval. According to Property (I), the delays satisfy $0<\tau_d<2T$. Therefore, the delay-induced feedback connections with target in the interval $I_\ell$ can originate from one of the following intervals: $I_\ell$, $I_{\ell-1}$, or $I_{\ell-2}$. In other words: the time points $t^\ell_n - \tau_d$ can belong to one of these intervals $I_{\ell}$, $I_{\ell -1}$, $I_{\ell -2}$. Formally, it can be written as \begin{equation} \label{eq:inIell} t_{n}^{\ell}-\tau_{d}=t_{n}^{\ell}-n_{d}\theta=\begin{cases} t_{n-n_{d}}^{\ell}\in I_{\ell}, & \text{if}\quad n_{d}<n,\\ t_{N+n-n_{d}}^{\ell-1}\in I_{\ell-1}, & \text{if}\quad n\leq n_{d}<N+n,\\ t_{2N+n-n_{d}}^{\ell-2}\in I_{\ell-2}, & \text{if}\quad N+n\leq n_{d}. \end{cases} \end{equation} We limit the class of networks to multilayer systems with connections between the neighboring layers. Such networks, see Fig.~\ref{fig:network-from-delay-system}b, are frequently employed in machine learning tasks, e.g. as deep neural networks \cite{Bishop2006,Goodfellow2016,Lecun2015,Schmidhuber2015,Stelzer2020}. Using \eqref{eq:inIell}, we can formulate a condition for the modulation functions $\mathcal{M}_d(t)$ to ensure that the delay terms $x(t-\tau_d)$ induce only connections between subsequent layers. For this, we set the modulation functions' values to zero if the originating time point $t^\ell_n - \tau_d$ of the corresponding delay connection does not belong to the interval $I_{\ell-1}$. This leads to the following assumption on the modulation functions:\\ \textbf{Property (III):} The modulation functions $\mathcal{M}_d(t)$ vanish at the following intervals: \begin{align}\label{eq:M-condition} \mathcal{M}_d(t) = v_{d,n}^\ell = 0 \quad \text{for} \quad t\in I_{\ell,n} \quad \text{if} \quad (n_d < n) \quad \text{or} \quad (N+n \leq n_d). \end{align} In the following, we assume that condition (III) is satisfied. Expressions \eqref{eq:euler}--\eqref{eq:euler1} contain the interdependencies between $x_n^\ell$, i.e., the connections between the nodes of the network. We explain these dependencies and present them in a more explicit form in the following. Our goal is to obtain the multilayer network shown in Fig.~\ref{fig:network-from-delay-system}b. \subsection{Effect of time-delays on the network topology} \label{subsec:network-topology} \begin{figure} \centering \includegraphics{net-topo} \caption{Network connections induced by one time-delay $\tau_d$. Panel (a): connections induced by $\tau_d<T$. Panel (b): $\tau_d = T$. Panel (c): $\tau_d > T$. Multiple delays $\tau_1, \ldots ,\tau_D$ result in a superposition of parallel patterns as shown in Fig.~\ref{fig:network-from-delay-system}b. } \label{fig:net-topo} \end{figure} Taking into account property (III), the node $x^\ell_n$ of layer $I_{\ell}$ receives a connection from a node $x^{\ell -1}_{n-n'_d}$ of layer $I_{\ell -1}$, where $n'_d:=n_d-N$. Two neighboring layers are illustrated in Fig.~\ref{fig:net-topo}, where the nodes in each layer are ordered vertically from top to bottom. Depending on the size of the delay, we can distinguish three cases. \begin{itemize} \item[(a)] For $\tau_d<T$, there are $n_d$ ``upward'' connections as shown in panel Fig.~\ref{fig:net-topo}a. \item[(b)] For $\tau_d=T$, there are $n_d=N$ ``horizontal'' delay-induced connections, i.e. connections from nodes of layer $\ell -1$ to nodes of layer $\ell$ with the same index, see Fig.~\ref{fig:net-topo}b. \item[(c)] For larger delays $\tau_d>T$, there are $2N-n_d$ ``downward'' delay-induced connections, as shown in Fig.~\ref{fig:net-topo}c. \end{itemize} In all cases, the connections induces by one delay $\tau_d$ are parallel. Since the delay system possesses multiple delays $0 < \tau_1 < \ldots < \tau_D < 2T$, the parallel connection patterns overlap, as illustrated in Fig.~\ref{fig:network-from-delay-system}b, leading to a more complex topology. In particular, a fully connected pattern appears for $D = 2N-1$ and $\tau_d = \theta d$. \subsection{Modulation of connection weights} \label{subsec:weight-modulation} With the modulation functions satisfying property (III), the Euler scheme~\eqref{eq:euler}--\eqref{eq:euler1} simplifies to the following map \begin{align}\label{eq:general-map1} x^\ell_1 &= x^{\ell-1}_N + \theta f(x^{\ell-1}_N,z(t^\ell_1), v^\ell_{1,1} x^{\ell -1}_{1-n'_1},\ldots ,v^\ell_{D,1} x^{\ell -1}_{1-n'_D}),\\ x^\ell_n &= x^\ell_{n-1} + \theta f(x^\ell_{n-1}, z(t^\ell_n), v^\ell_{1,n} x^{\ell -1}_{n-n'_1},\ldots ,v^\ell_{D,n} x^{\ell -1}_{n-n'_D}), \quad n=2,\ldots ,N,\label{eq:general-map} \end{align} where Eq.~\eqref{eq:M-condition} implies $v^{\ell}_{d,n} = 0$ if $n-n'_d < 1$ or $n-n'_d > N$. In other words, the dependencies at the right-hand side of \eqref{eq:general-map1}--\eqref{eq:general-map} contain only the nodes from the $\ell-1$-th layer. Moreover, the numbers $v^\ell_{d,n}$ determine the strengths of the connections from $x^{\ell -1}_{n-n'_d}$ to $x^\ell_n$ and can be considered as network weights. By reindexing, we can define weights $w^\ell_{nj}$ connecting node $j$ of layer $\ell -1$ to node $n$ of layer $\ell$. These weights are given by the equation \begin{align}\label{eq:weight-matrix1} w^\ell_{nj} := \sum_{d =1}^D \delta_{n-n'_d,j} v^\ell_{d,n} = \begin{cases} 0 & \text{if } \forall d \colon j \neq n - n'_d,\\ v^\ell_{d,n} & \text{if } \exists d \colon j = n - n'_d, \end{cases} \end{align} and define the entries of the weight matrix $W^\ell = (w^\ell_{nj}) \in \mathbb{R}^{N\times (N+1)}$, except for the last column, which is defined below and contains bias weights. The symbol $\delta_{nj}$ is the Kronecker delta, i.e. $\delta_{nj} = 1$ if $n=j$, and $\delta_{nj} = 0$ if $n\neq j$. \begin{figure}[t] \centering \includegraphics{matrix12} \caption{ Coupling matrix $W^\ell$ between the hidden layers $\ell - 1$ and $\ell$, see Eq.~\eqref{eq:weight-matrix1}--\eqref{eq:weight-matrix2}. The nonzero weights are arranged along the diagonals, and equal $v^\ell_{d,n}$. The position of the diagonals is determined by the corresponding delay $\tau_d$. If $\tau_d = T = N\theta$, then the main diagonal contains the entries $v^\ell_{d,1},\ldots ,v^\ell_{d,N}$ (shown in yellow). If $\tau_d = n_d\theta < T$, then the corresponding diagonal lies above the main diagonal and contains the values $v^\ell_{d , 1}, \ldots , v^\ell_{d , n_d}$ (red). If $\tau_d = n_d\theta > T$, then the corresponding diagonal lies below the main diagonal and contains the values $v^\ell_{d , n_d-N+1}, \ldots , v^\ell_{d, N}$ (blue). The last column of the matrix contains the bias weights (gray). } \label{fig:weight-matrix} \end{figure} The time-dependent driving function $z(t)$ can be utilized to realize a bias weight $b^\ell_n$ for each node $x^\ell_n$. For details, we refer to Sec.~\ref{subsec:mulilayer-network}. We define the last column of the weight matrix $W^\ell$ by \begin{align}\label{eq:weight-matrix2} w^\ell_{n,N+1} := b^\ell_n. \end{align} The weight matrix is illustrated in Fig.~\ref{fig:weight-matrix}. This matrix $W^\ell$ is in general sparse, where the degree of sparsity depends on the number $D$ of delays. If $D=2N-1$ and $\tau_d = d\theta, \ d=1,\ldots ,D$, we obtain a dense connection matrix. Moreover, the positions of the nonzero entries and zero entries are the same for all matrices $W^2, \ldots , W^L$, but the values of the nonzero entries are in general different. \subsection{Interpretation as multilayer neural network} \label{subsec:mulilayer-network} The map~\eqref{eq:general-map1}--\eqref{eq:general-map} can be interpreted as the hidden layer part of a multilayer neural network provided we define suitable input and output layers. \begin{figure}[t] \centering \includegraphics{network-from-delay-system} \caption{Implementing a multilayer neural network by delay system~\eqref{eq:general-system}. Panel (a): The system state is considered at discrete time points $x^\ell_{n} := x((\ell - 1)T +n\theta)$. The intervals $I_\ell$ correspond to layers. Due to delayed feedback, non-local connections emerge (color lines). Panel (b) shows the resulting neural network. } \label{fig:network-from-delay-system} \end{figure} The input layer determines how a given input vector $u\in \mathbb{R}^{M+1}$ is transformed to the state of the first hidden layer $x(t),\ t\in I_1$. The input $u\in \mathbb{R}^{M+1}$ contains $M$ input values $u_1,\ldots ,u_M$ and an additional entry $u_{M+1}=1$. In order to ensure that $x(t),\ t\in I_1$ depends on $u$ and the initial state $x(0)=x_0$ exclusively, and does not depend on a history function $x(s),\ s < 0$, we set all modulation functions to zero on the first hidden layer interval. This leads to the following\\ \textbf{Property (IV):} The modulation functions satisfy \begin{align} \mathcal{M}_d(t) = 0, \quad t\in I_1, \ d=1,\ldots ,D. \end{align} The dependence on the input vector $u\in\mathbb{R}^{M+1}$ can be realized by the driving signal $z(t)$. \\ \textbf{Property (V):} The driving signal $z(t)$ on the interval $I_1$ is the step function given by \begin{align}\label{eq:J-1} z(t) = & J(t) \quad \text{for } \quad t\in I_1, \\ & J(t) = J_n = \left[ f^\mathrm{in}(W^{\mathrm{in}} u) \right]_n \quad \text{for} \quad t\in I_{1,n},\label{eq:J-2} \end{align} where $f^\mathrm{in}(W^{\mathrm{in}} u)\in\mathbb{R}^N$ is the preprocessed input, $W^{\mathrm{in}}\in\mathbb{R}^{N\times (M+1)}$ is an input weight matrix, and $f^\mathrm{in}$ is an element-wise input preprocessing function. For example, $f^\mathrm{in}(a)=\tanh (a)$ was used in~\cite{Stelzer2020}. As a result, the following holds for the first hidden layer \begin{align}\label{eq:general-system-first-layer} \dot{x}(t) = f(x(t), J(t), 0,\ldots ,0), \quad t\in I_1, \end{align} which is just a system of ordinary differential equations, which requires an initial condition at a single point $x(0)=x_0$ for solving it in positive time. This yields the coupled map representation \begin{align}\label{eq:general-map1-first-layer} x^\ell_1 &= x_0 + \theta f(x_0, J_{1},0,\ldots ,0),\\ x^1_n &= x^1_{n-1} + \theta f(x^1_{n-1}, J_{n},0,\ldots ,0), \quad n=2,\ldots ,N.\label{eq:general-map-first-layer} \end{align} For the hidden layers $I_2,I_3,\dots$, the driving function $z(t)$ can be used to introduce a bias as follows. \\ \textbf{Property (VI):} The driving signal $z(t)$ on the intervals $I_\ell$, $\ell\ge 2$, is the step function given by \begin{align} z(t) = & b(t) \quad \text{for } \quad t>T, \\ & b(t) = b_n^\ell \quad \text{for} \quad t\in I_{\ell,n},\quad \ell \ge 2. \end{align} Assuming the properties (I)--(VI), Eqs.~\eqref{eq:general-map1}--\eqref{eq:general-map} imply \begin{align}\label{eq:general-map1-hidden-layer} x^\ell_1 &= x^{\ell-1}_N + \theta f(x^{\ell-1}_N,b^\ell_1, v^\ell_{1,1} x^{\ell -1}_{1-n'_1},\ldots ,v^\ell_{D,1} x^{\ell -1}_{1-n'_D}),\\ x^\ell_n &= x^\ell_{n-1} + \theta f(x^\ell_{n-1}, b^\ell_n, v^\ell_{1,n} x^{\ell -1}_{n-n'_1},\ldots ,v^\ell_{D,n} x^{\ell -1}_{n-n'_D}), \quad n=2,\ldots ,N.\label{eq:general-map-hidden-layer} \end{align} Let us finally define the output layer, which transforms the node states $x^L_1,\ldots , x^L_n$ of the last hidden layer to an output vector $\hat{y} \in \mathbb{R}^P$. For this, we define a vector $x^L := (x^L_1, \ldots , x^L_N, 1)^\mathrm{T} \in \mathbb{R}^{N+1}$, an output weight matrix $W^\mathrm{out}\in\mathbb{R}^{P\times (N+1)}$, and an output activation function $f^\mathrm{out}\colon \mathbb{R}^P \to \mathbb{R}^P$. The output vector is then defined as \begin{align}\label{eq:general-output} \hat{y} = f^\mathrm{out}(W^\mathrm{out}x^L). \end{align} Figure~\ref{fig:network-from-delay-system} illustrates the whole construction process of the coupled maps network; it is given by the equations~\eqref{eq:general-map1-first-layer}--\eqref{eq:general-output}. We summarize the main result of section 2. \begin{tcolorbox} Under assumptions (I)--(VI) and for small $\theta$, DDE \eqref{eq:general-system} describes the multilayer network of coupled maps shown in Fig.~\ref{fig:network-from-delay-system}, with the specific dependencies given by Eqs.~\eqref{eq:general-map1-first-layer}, \eqref{eq:general-map-first-layer}, \eqref{eq:general-map1-hidden-layer}, \eqref{eq:general-map-hidden-layer}, and \eqref{eq:general-output}. \end{tcolorbox} \section{Constructing a recurrent neural network from a delay system} \label{sec:recurrent-network} System~\eqref{eq:general-system} can also be considered as recurrent neural network. To show this, we consider the system on the time interval $[0,KT]$, for some $K\in\mathbb{N}$, which is divided into intervals $I_k :=((k-1)T,kT], \ k = 1, \ldots ,K$. We use $k$ instead of $\ell$ as index for the intervals to make clear that the intervals do not represent layers. The state $x(t)$ on an interval $I_k$ is interpreted as the state of the recurrent network at time $k$. More specifically, \begin{align} x^k_n := x((k - 1)T + n\theta), \quad n = 1,\ldots ,N, \ k = 1, \ldots ,K \end{align} is the state of node $n$ at the discrete time $k$. The driving function $z(t)$ can be utilized as an input signal for each $k$-time-step.\\ \textbf{Property (VII):} $z(t)$ is the $\theta$-step function with \begin{align} z(t)= z_n^k & \quad \text{for} \quad t\in I_{k,n}, \\ & (z_1^k,\dots,z_N^k)^T = f^\mathrm{in}(W^\mathrm{in} u(k)), \end{align} where $u(k),\ k = 1, \ldots ,K$ are $(M+1)$-dimensional input vectors, $W^\mathrm{in} \in \mathbb{R}^{N\times (M+1)}$ is an input weight matrix, $f^\mathrm{in}$ is an element-wise input preprocessing function. Each input vector $u(k)$ contains $M$ input values $u_1(k),\ldots ,u_M(k)$ and a fixed entry $u_{M+1}(k):=1$ which is needed to include bias weights in the last column of $W^\mathrm{in}$. The main difference of the Property (VII) from (VI) is that it allows for the information input through $z(t)$ in all intervals $I_k$. Another important difference is related to the modulation functions, which must be $T$-periodic in order to implement a recurrent network. This leads to the following assumption. \\ \textbf{Property (VIII):} The modulation functions $\mathcal{M}_d(t)$ are $T$-periodic $\theta$-step functions with \begin{align} \mathcal{M}_d(t)= v_{d,n}\quad \text{for} \quad t\in I_{k,n}. \end{align} Note that the value $v_{d,n}$ is independent on $k$ due to periodicity of $\mathcal{M}_d(t)$. When assuming the Properties (I), (III), (IV), (VII), and (VIII), the map equations~\eqref{eq:general-map1}--\eqref{eq:general-map} become \begin{align}\label{eq:general-map1-recurrent} x^k_1 &= x^{k-1}_N + \theta f(x^{k-1}_N, z^k_1, v_{1,1} x^{k -1}_{1+n'_1},\ldots ,v_{D,1} x^{k -1}_{1+n'_D}),\\ x^k_n &= x^k_{n-1} + \theta f(x^k_{n-1}, z^k_n, v_{1,n} x^{k -1}_{n+n'_1},\ldots ,v_{D,n} x^{k -1}_{n+n'_D}), \quad n=2,\ldots ,N,\label{eq:general-map-recurrent} \end{align} and can be interpreted as a recurrent neural network with the input matrix $W^\mathrm{in}$ and the internal weight matrix $W = (w_{nj}) \in \mathbb{R}^{N\times N}$ defined by \begin{align}\label{eq:recurrent-weight-matrix} w_{nj} := \sum_{d =1}^D \delta_{n-n'_d,j} v_{d,n} = \begin{cases} 0 & \text{if } \forall d \colon j \neq n - n'_d,\\ v_{d,n} & \text{if } \exists d \colon j = n - n'_d. \end{cases} \end{align} \begin{figure} \centering \includegraphics{rnn} \caption{Recurrent network obtained from DDE \eqref{eq:general-system} with two delays. The delays $\tau_1<T$ and $\tau_2>T$ induce connections with opposite direction (color arrows). Moreover, the nodes of the recurrent layer a linearly locally coupled (black arrows). All nodes of the recurrent layers are connected to the input and output layer. } \label{fig:recurrent-network} \end{figure} When we choose the number of delays to be $D=2N-1$, we can realize any given connection matrix $W\in \mathbb{R}^{N\times N}$. For that we need to choose the delays $\tau_d = d\theta , \ d=1,\ldots ,2N-1$. Consequently there are $D=2N-1$ modulation functions $\mathcal{M}_d(t)$ which are step functions with values $v_{d,n}$. In this case Eq.~\eqref{eq:recurrent-weight-matrix} provides for all entries $w_{nj}$ of $W$ exactly one corresponding $v_{d,n}$. Therefore, the arbitrary matrix $W$ can be realized by choosing appropriate step heights for the modulation functions. In the setting of Sec.~\ref{sec:recurrent-network} the resulting network is an arbitrary recurrent network. Summarizing, the main message of Sec. 3 is as follows. \begin{tcolorbox} Under assumptions (I), (III), (IV), (VII), and (VIII), and for small $\theta$, DDE \eqref{eq:general-system} describes the recurrent network shown in Fig.~\ref{fig:recurrent-network}, with the specific dependencies given by Eqs.~\eqref{eq:general-map1-recurrent}--\eqref{eq:general-map-recurrent} and an internal weight matrix $W$ given by \eqref{eq:recurrent-weight-matrix}. \end{tcolorbox} \section{Networks from delay systems with linear instantaneous part and nonlinear delayed feedback} \label{sec:semilinear-case} \label{subsec:semilinear-delay-system} Particularly suitable for the construction of neural networks are delay systems with a stable linear instantaneous part and a feedback given by a nonlinear function of an affine combination of the delay terms and a driving signal. Such DDEs are described by the equation \begin{align}\label{eq:semilinear-system} \dot{x}(t) &= -\alpha x(t) + f(a(t)), \end{align} where $\alpha > 0$ is a constant time scale, $f$ is a nonlinear function, and \begin{align}\label{eq:activation-signal} a(t) &= z(t) + \sum_{d=1}^D \mathcal{M}_d(t)x(t-\tau_d). \end{align} Ref.~\cite{Krisztin2008} studied this type of equation for the case $D=1$, i.e. for one delay. An example of \eqref{eq:semilinear-system} is the Ikeda system~\cite{Ikeda1979} where $D=1$, i.e. $a(t)$ consists of only one scaled feedback term $x(t-\tau)$, signal $z(t)$, and the nonlinear function $f(a)=\sin(a)$. This type of dynamics can be applied to reservoir computing using optoelectronic hardware~\cite{Larger2012}. Another delay dynamical system of type~\eqref{eq:semilinear-system}, which can be used for reservoir computing, is the Mackey-Glass system~\cite{Appeltant2011}, where $D=1$ and the nonlinearity is given by $f(a)=\eta a /(1+|a|^p)$ with constants $\eta , p > 0$. In the work~\cite{Stelzer2020}, system~\eqref{eq:semilinear-system} is used to implement a deep neural network. Even though the results of the previous sections are applicable to \eqref{eq:semilinear-system}--\eqref{eq:activation-signal}, the special form of these equations allows for an alternative, more precise approximation of the network dynamics. \subsection{Interpretation as multilayer neural network} \label{subsec:ddnn-network} It is shown in \cite{Stelzer2020} that one can derive a particularly simple map representation for system~\eqref{eq:semilinear-system} with activation signal~\eqref{eq:activation-signal}. We do not repeat here the derivation, and only present the resulting expressions. By applying a semi-analytic Euler discretization and the variation of constants formula, the following equations connecting the nodes in the network are obtained: \begin{align}\label{eq:first-hidden-layer-first-node} x^1_1 &= e^{-\alpha \theta} x_0 + \alpha^{-1}(1-e^{-\alpha \theta}) f( a^1_1), \\ x^1_n &= e^{-\alpha \theta} x^1_{n-1} + \alpha^{-1}(1-e^{-\alpha \theta}) f( a^1_n), \quad n = 2, \ldots , N, \end{align} for the first hidden layer. The hidden layers $\ell = 2, \ldots ,L$ are given by \begin{align}\label{eq:hidden-layer-first-node} x^\ell_1 &= e^{-\alpha \theta} x^{\ell -1}_N + \alpha^{-1}(1-e^{-\alpha \theta}) f( a^\ell_1), \\ x^\ell_n &= e^{-\alpha \theta} x^\ell_{n-1} + \alpha^{-1}(1-e^{-\alpha \theta}) f( a^\ell_n), \quad n = 2, \ldots , N.\label{eq:hidden-layer} \end{align} The output layer is defined by \begin{align}\label{eq:output-layer} \hat{y}_p := f^\mathrm{out}_p (a^\mathrm{out}), \quad p=1,\ldots, P, \end{align} where $f^\mathrm{out}$ is an output activation function. Moreover, \begin{align} a^\mathrm{in}_n &:= \sum_{m=1}^{M+1} w^\mathrm{in}_{nm} u_m, & & n = 1, \ldots , N, \label{eq:activation-in}\\ a^1_n &:= g(a^\mathrm{in}_n), & & n = 1, \ldots , N, \label{eq:activation-first}\\ a^\ell_n &:= \sum_{j=1}^{N+1} w^\ell_{nj} x^{\ell -1}_j, & & n=1, \ldots , N, \ \ell = 2, \ldots , L,\label{eq:activation}\\ a^\mathrm{out}_p &:= \sum_{n=1}^{N+1} w^\mathrm{out}_{pn} x^L_n, & & p = 1, \ldots , P,\label{eq:activation-output} \end{align} where $u_{M+1}:=1$ and $x^\ell_{N+1} := 1$, for $\ell=1,\ldots ,L$. One can also formulate the relation between the hidden layers in a matrix form. For this, we define \begin{align} A := \begin{pmatrix} 0 & \cdots & \cdots & \cdots & 0 \\ e^{-\alpha\theta} & \ddots & & & \vdots \\ 0 & \ddots & \ddots & & \vdots \\ \vdots & \ddots & \ddots & \ddots &\vdots \\ 0 & \cdots & 0 & e^{-\alpha\theta} & 0 \end{pmatrix}. \end{align} Then, for $\ell = 2,\ldots ,L$, the equations~\eqref{eq:hidden-layer-first-node}--\eqref{eq:hidden-layer} become \begin{align}\label{eq:dnn-matrix-form-1} x^\ell = A x^\ell + \begin{pmatrix}e^{-\alpha\theta}x^{\ell -1}_N\\0\\ \vdots \\ 0\end{pmatrix} + \alpha^{-1}(1-e^{-\alpha \theta})f(W^\ell x^{\ell -1}). \end{align} where $f$ is applied component-wise. By subtracting $A x^\ell$ from both sides of Eq.~\eqref{eq:dnn-matrix-form-1} and multiplication by the matrix \begin{align} E := (\mathrm{Id} - A)^{-1} = \begin{pmatrix} 1 & 0 & \cdots & \cdots & 0 \\ e^{-\alpha\theta} & 1 & \ddots & & \vdots \\ e^{-2\alpha\theta} & \ddots & \ddots & \ddots & \vdots \\ \vdots & \ddots & \ddots & \ddots & 0 \\ e^{-(N-1)\alpha\theta} & \cdots & e^{-2\alpha\theta} & e^{-\alpha\theta} & 1 \end{pmatrix}, \end{align} we obtain a matrix equation describing the $\ell$-th hidden layer \begin{align} x^\ell = \begin{pmatrix}e^{-\alpha\theta}x^{\ell -1}_N\\e^{-2\alpha\theta}x^{\ell -1}_N\\ \vdots \\ e^{-N\alpha\theta}x^{\ell -1}_N\end{pmatrix} + \alpha^{-1}(1-e^{-\alpha \theta}) E f(W^\ell x^{\ell -1}). \end{align} The neural network~\eqref{eq:first-hidden-layer-first-node}--\eqref{eq:activation-output} obtained from delay system~\eqref{eq:semilinear-system}--\eqref{eq:activation-signal} can be trained by gradient descent~\cite{Stelzer2020}. The training parameters are the entries of the matrices $W^\mathrm{in}$ and $W^\mathrm{out}$, the step heights of the modulation functions $\mathcal{M}_d(t)$, and the bias signal $b(t)$. \subsection{Network for large node distance $\theta$} \label{subsec:largetheta} In contrast to the general system~\eqref{eq:general-system}, the semilinear system~\eqref{eq:semilinear-system} with activation signal~\eqref{eq:activation-signal} does not only emulate a network of nodes for small distance $\theta$. It is also possible to choose large $\theta$. In this case, we can approximate the nodes given by Eq.~\eqref{eq:nodes} by the map limit \begin{align}\label{eq:map-limit} & x^\ell = \alpha^{-1} f(a^\ell), \\ & \text{where} \quad a^\ell = W^\ell x^{\ell-1} \ \text{for} \ \ell>1 \quad \text{and} \quad a^1 = g(W^\mathrm{in} u), \end{align} up to exponentially small terms. The reason for this limit behavior lies in the nature of the local couplings. Considering Eq.~\eqref{eq:semilinear-system}, one can interpret the parameter $\alpha$ as a time scale of the system, which determines how fast information about the system state at a certain time point decays while the system is evolving. This phenomenon is related to the so-called instantaneous Lyapunov exponent \cite{Heiligenthal2011,Heiligenthal2013,Kinzel2013}, which equals $-\alpha$ in this case. As a result, the local coupling between neighboring nodes emerges when only a small amount of time $\theta$ passes between the nodes. Hence, increasing $\theta$ one can reduce the local coupling strength until it vanishes up to a negligibly small value. For a rigorous derivation of Eq.~\eqref{eq:map-limit}, we refer to \cite{Stelzer2020}. The apparent advantage of the map limit case is that the obtained network matches a classical multilayer perceptron. Hence, known methods such as gradient descent training via the classical back-propagation algorithm \cite{Rumelhart1986} can be applied to the delay-induced network~\cite{Stelzer2020}. The downside of choosing large values for the node separation $\theta$ is that the overall processing time of the system scales linearly with $\theta$. We need a period of time $T=N\theta$ to process one hidden layer. Hence, processing a whole network with $L$ hidden layers requires the time period $LT=LN\theta$. For this reason, the work~\cite{Stelzer2020} provides a modified back-propagation algorithm for small node separations to enable gradient descent training of networks with significant local coupling. \section{Conclusions} We have shown how networks of coupled maps with arbitrary topology and arbitrary size can be emulated by a single (possibly even scalar) DDE with multiple delays. Importantly, the coupling weights can be adjusted by changing the modulations of the feedback signals. The network topology is determined by the choice of time-delays. As shown previously \cite{Appeltant2011,Larger2012,Brunner2013,Larger2017,Stelzer2020}, special cases of such networks are successfully applied for reservoir computing or deep learning. As an interesting conclusion, it follows that the temporal dynamics of DDEs can unfold arbitrary spatial complexity, which, in our case, is reflected by the topology of the unfolded network. In this respect, we shall mention previously reported spatio-temporal properties of DDEs \cite{Arecchi1992,Giacomelli1994,Giacomelli1996,Bestehorn2000,Giacomelli2012,Yanchuk2014,Yanchuk2015a,Yanchuk2015b,Kashchenko2016,Yanchuk2017}. These results show how in some limits, mainly for large delays, the DDEs can be approximated by partial differential equations. Further, we remark that similar procedures have been used for deriving networks from systems of ordinary differential equations \cite{Haber2017,Lu2018,Chen2018}. However, in their approach, one should use an $N$-dimensional system of equations for implementing layers with $N$ nodes. This is in contrast to the DDE case, where the construction is possible with just a single-variable equation. As a possible extension, a realization of adaptive networks using a single node with delayed feedback would be an interesting open problem. In fact, the application to deep neural networks in \cite{Stelzer2020} realizes an adaptive mechanism for the adjustment of the coupling weights. However, this adaptive mechanism is specially tailored for DNN problems. Another possibility would be to emulate networks with dynamical adaptivity of connections \cite{Berner}. The presented scheme can also be extended by employing delay differential-algebraic equations \cite{Mehrmann,Unger}. \section*{Acknowledgements} This work was funded by the ``Deutsche Forschungsgemeinschaft'' (DFG) in the framework of the project 411803875 (S.Y.) and IRTG 1740 (F.S.).

bc07ac44092bf42c587cc380e20037c27e2b68a8

\section{Introduction} Although the Standard Model (SM) of particle physics was built on the basis of massless neutrinos, a plethora of data collected by neutrino oscillation experiments, using solar, atmospheric, reactor and accelerator neutrino sources, had provided uncontroversial evidence that neutrinos are actually massive and light~\cite{nuoscPDG}. Most observed oscillation phenomena is well explained if standard weak neutrino flavors are not the actual mass egienstates, but rather a combination of the last, such that $\nu_{\ell L}= U_{\ell i}~\nu_{iL}$, for $\ell = e,\mu,\tau$ and $i=1,2, 3$. The mechanism that provides neutrino masses is still unknown, yet, Majorana neutrinos seem the most natural choice, since the simple addition of missing right handed neutrinos to the SM would give an explanation to the lightness of the active neutrinos, through the seesaw mechanism~\cite{seesaw}. In any case, effective Majorana mass terms written in the weak basis, $(M_\nu)_{\ell\ell^\prime} \bar{\nu}_{\ell L}^c \nu_{\ell^\prime L}$, should be non diagonal. Diagonalization of these would proceed through an orthogonal transformation with the unitary mixing matrix $U$, so that $M_{diag} = U^T M_\nu U$, where $M_{diag} = \text{Diag}\{ m_1 ,m_2 , m_3 \}$, with $m_i$ the real masses. The $U$ matrix can be written in terms of three complex rotations by using Pontecorvo-Maki-Nakagawa-Sakata (PMNS)~\cite{pontecorvo,MNS} parameterization, where $U = V\cdot K$, and $$ V= \left( \begin{array}{ccc} c_{12} c_{13} & s_{12}c_{13} & z \\ - s_{12} c_{23} - c_{12}s_{23}\bar{z} & c_{12}c_{23} - s_{12}s_{23}\bar{z} & s_{23}c_{13} \\ s_{12}s_{23} - c_{12}c_{23}\bar{z} & - c_{12}s_{23} -c_{23}s_{12}\bar{z} & c_{23}c_{13} \end{array} \right)~, $$ with $c_{ij}$ and $s_{ij}$ standing for $\cos \theta_{ij}$ and $\sin \theta_{ij}$, respectively, of the mixing angles given as $\theta_{12}$, $\theta_{13}$, and $\theta_{23}$. Here, $z= s_{13}e^{-i\delta_{CP}}$, where $\delta_{CP}$ is the Dirac $CP$ phase, whereas $K$ is a diagonal matrix containing two Majorana phases which do not contribute to neutrino oscillations. In the basis where charge lepton masses are diagonal, same that we assume hereafter, $U$ also describes the mixings involved in charged weak currents, $W_\mu \bar \ell_L \gamma^\mu U_{\ell i}\nu_{iL}$. Neutrino oscillation experiments are actually sensible to above mixings and to mass squared differences, $\Delta m^2_{ij} = m_i^2 - m_j^2$, identified as solar $\Delta m^2_{sol} = \Delta m^2_{21}$ and atmospheric $\Delta m^2_{ATM}=|\Delta m^2_{31}|$ scales. Due to matter effects within the sun, it is know that $\Delta m^2_{21}>0$, whereas the sign of $\Delta m^2_{31}$ and therefore the hierarchy of masses is still unknown. If $m^2_3>m^2_1$ we would have a normal hierarchy (NH), otherwise it is called inverted. Global fits of oscillation data with all three neutrinos~\cite{nuglobal} indicate that in NH, for instance, at one sigma level, $\Delta m_{sol}^2 = 7.42^{+0.21}_{-0.20}\times 10^{-5}~eV^2$ and $\Delta m_{ATM}^2 = 2.517^{+0.026}_{-0.028}\times 10^{-3}~eV^2$, $\sin^2\theta_{12} = \sin^2\theta_{sol}=0.304^{+0.012}_{-0.012}$, $\sin^2\theta_{23}= \sin^2\theta_{ATM}=0.573^{+0.016}_{-0.020}$, and $\sin^2\theta_{13}= 0.02219^{+0.00062}_{-0.00063}$, corresponding to reactor oscillation mixings. On the other hand, $\delta_{CP}$ should be within the interval $[120^o,369^o]$ at three sigma level, which still allows for non CP violation ($\delta_{CP}=180^o$ or $360^o$). In the attempt to understand such parameter pattern, flavor symmetries had long been advocated as a possible explanation of the observed mass hierarchies and mixings among fundamental fermions. Amid the possibilities, $\mu-\tau$ exchange symmetry stands out since it naturally appears in $M_\nu$ when maximal $\theta_{23}$ and zero $\theta_{13}$ are taken for $U$, and a large number of studies had been dedicated to explore this symmetry and its possible realization in extended fermion mass models~\cite{mutau,others,mtreviews}. However, all such studies always assume that a null $\theta_{13}$ is a necessary outcome of $\mu-\tau$ symmetry and its non zero observed value as the signature of the breaking of the symmetry. Also, this approach leaves out solar mixing, offering no understanding for its observed value whatsoever. It turns out, however, that above claims are not compulsory. As I will discuss along this paper, the actual signature of exact $\mu-\tau$ symmetry is solely: (i) the maximality of $\theta_{23}$, (ii) a zero Dirac CP phase, and (iii) the existence of an null mixing.\\ Identifying the last with $\theta_{13}$ is just a particular case subjected to a specific ordering of mass eigenstates. As a matter of fact, there is a possible ordering where $\theta_{13}$ arises with a non zero value in the exact symmetry limit. In such a scenario it is rather $\theta_{12}$ which results null. Although such an outcome seems unwanted, given the observed large value of solar mixing, the right value can be obtained from the breaking of $\mu-\tau$ symmetry, to the cost of starting with a degenerate spectrum in the one-two sector before considering symmetry breaking corrections. With this mechanism, the emergence of the solar mixing gets attached to the origin of $\Delta m^2_{sol}$. To support these claims, I shall present a mass matrix structure for normal hierarchy where both $\tan\theta_{13}$ and the breaking of $\mu-\tau$ are governed by a single parameter, the ratio $\sqrt{\Delta m^2_{sol}/\Delta m^2_{ATM}}$. As I will show, the proposed $M_\nu$ successfully predicts the oscillation mass scales, a small reactor mixing, an atmospheric mixing that is larger than $\pi/4$, and a large solar mixing, all consistent with observed values. Besides, it also allows to incorporate CP violation without losing above features. The discussion is organized by starting with a brief review of the general $\mu-\tau$ symmetric predictions to state the basis of the case that will be considered along the paper. Then, the texture under study and their predictions shall be introduce in section III. Stability under renormalization effects and the exploration of CP violation are addressed in sections IV and V, respectively, and the paper is closed with some remarks and conclusions. \section{masses and mixings from $\mu-\tau$ symmetry} Assuming $\mu-\tau$ symmetry in the Majorana neutrino mass matrix means that the matrix elements, $m_{\ell,\ell^\prime}$ should comply with the identities $m_{e\mu} = m_{e\tau}$ and $m_{\mu\mu} = m_{\tau\tau}$. On the other hand, in order to determine the mixing matrix we should consider the hermitian squared matrix $H = MM^\dagger$, which would explicitly exhibit $\mu-\tau$ symmetry too. $H$ can in general be written as \begin{equation} H= \left( \begin{array}{ccc} \alpha & \omega & \omega \\ \bar\omega & \beta & \rho \\ \bar\omega & \rho & \beta \end{array} \right) \label{H} \end{equation} where $\alpha = |m_{ee}|^2 + 2 |m_{e\mu}|^2$, $\beta = |m_{\mu\mu}|^2 + |m_{e\mu}|^2 + |m_{\mu\tau}|^2$, and $\rho= 2Re(m_{\mu\mu}\bar m_{\mu\tau}) + |m_{e\mu}|^2$, with the only complex component being $\omega=m_{ee}\bar m_{e\mu} + m_{e\mu}(\bar m_{\mu\mu} + \bar m_{\mu\tau})$. The mixing matrix $U$ should then diagonalize $H$ through the unitary transformation $U^T HU^* = M_{diag}^2$. Thus, the complex conjugated orthonormal eigenvectors of $H$ do correspond to the columns of the mixing matrix. Furthermore, given that the characteristic polynomial can be factored as $p_H(\lambda) = [(\beta -\rho) - \lambda][(\alpha - \lambda)(\beta +\rho - \lambda) - 2|\omega|^2]$, the squared eigenmasses are $\lambda_0 = \beta-\rho$ and \begin{equation} \lambda_{\pm} = \frac{1}{2}\left[\alpha + \beta + \rho \pm \sqrt{\left(\alpha - (\beta+\rho)\right)^2 + 8|\omega|^2}\right]~. \end{equation} The corresponding eigenvector for $\lambda_0$ is \begin{equation} \nu_0 = \frac{1}{\sqrt{2}}\left(\begin{array}{r} 0\\ -1\\ 1 \end{array} \right)~. \end{equation} Clearly, this implies $\mu$ and $\tau$ maximal mixing as the content of such eigenstate, which means a maximal $\theta_{23}$ on PMNS matrix, regardless of mass ordering. Moreover, and yet more interestingly, other mixings can be pined down right out of this unique eigenstate, as I discus next. By construction, $\lambda_+>\lambda_-$, thus, mass ordering is defined by the sole location of $\lambda_0$ within the spectrum. Either, (a) $\lambda_0>\lambda_+$, (b) $\lambda_+> \lambda_0>\lambda_-$, or (c) $\lambda_->\lambda_0$. Nonetheless, neutrino hierarchy would also be associated to the value of $\Delta\lambda = \lambda_+-\lambda_-$, which could either be about solar or atmospheric scale. Because of this, all above possible orderings may result phenomenologically viable, and so, $\lambda_0$ could play the role of any $m_i^2$ in the spectrum, with $\nu_0$ realizing any of the columns on PMNS mixing matrix. Hence, $\mu-\tau$ symmetry would have, besides a maximal $\theta_{23}$, any of the following outcomes. (i) For $\lambda_0= m_3^2$ one gets $z=0$, and thus a null $\theta_{13}$ and non Dirac CP phase. This is the case implicitly assumed in most $\mu-\tau$ models studied so far in the literature. (ii) For $\lambda_0= m_2^2$ one rather gets $\sin\theta_{12}= 0$, whereas, (iii) $\lambda_0= m_1^2$, implies $\cos\theta_{12} = 0$. On last two cases it is obvious that orthogonality of the eigenstates means that $|z|\neq 0$, and hence a non zero $\theta_{13}$. A calculation gives the general formula \begin{equation} \tan\theta_{13}=\frac{|\rho+\beta-m_3^2|}{\sqrt{2}\, |\omega|}~. \label{symmtheta13} \end{equation} Also, it is not difficult to check that in latter cases the Dirac CP phase becomes $\delta_{CP} = \arg(\omega)$, but it can always be factored out from the mixing matrix and absorbed within the Majorana phases. Thus, it can consistently be taken as zero at this level. Note that if $\lambda_0$ is degenerated with any other eigenvalue, scenarios (ii) and (iii) reduce to a single one where $\theta_{12}=0$. None of the above scenarios is consistent with data, yet closeness to maximal mixing in atmospheric oscillations suggest that $\mu-\tau$ could be treated as a slightly broken symmetry. Indeed, as the many previous studies show, symmetry breaking can eventually explain the non zero value of $\theta_{13}$ in the first scenario. I will not elaborate on such a case any further in here. Rather, I shall address the more interesting question raised by the other possible scenarios. In both of them, a small reactor mixing comes at the symmetric limit as a consequence of the interplay of matrix elements. Therefore, one could naturally expect that a small breaking of $\mu-\tau$ would provide small corrections to $\theta_{23}$ and $\theta_{13}$, so to explain the experimental values. Nevertheless, given that the input value from the symmetric limit for $\theta_{12}$ is either $0$ or $\pi/2$, it might appear challenging to understand from the same perspective the rather large measured value of solar neutrino mixing. It turns out that such is not the case. As a matter of fact, a large $\theta_{12}$ mixing can be obtained from zero mixing through perturbations, provided the initial spectrum is degenerate ($\Delta m^2_{12} = 0$). To illustrate this point, consider the following toy matrix, \begin{equation} A= \left (\begin{array}{cc} m &\varepsilon\\ \varepsilon & m \end{array} \right)~. \end{equation} Clearly, at the limit when $\varepsilon=0$, there is no mixing at all, and states are degenerated. However, regardless of its value, as soon as we take $\varepsilon\neq 0$, the mixing becomes maximal, meaning $\tan\theta = 1$, which happens also regardless of the value of $m$. Accommodating a value in between for $\theta$ becomes a matter of fixing $A$. Besides, the same perturbation also breaks the degeneracy, introducing a mass gap $\Delta m^2 = 4m\varepsilon$. In this line of thought, given that $\Delta m_{sol}^2$ is the smallest of the scales, it comes natural to think of it as indicating an initial degeneracy on the $m_1$-$m_2$ sector that sources the large solar mixing when lifted. That is indeed possible, and as I discuss next, one can use a realization of this mechanism in the context of three neutrinos to get a working structure for $M_\nu$ . \section{Generating solar oscillation parameters} For the rest of our discussion let us consider the following mass matrix structure for normal neutrino hierarchy, \begin{equation} \label{texture} M_\nu= \left (\begin{array}{ccc} d\epsilon^2 &\epsilon &a\epsilon\\ \epsilon & b & c \\ a\epsilon & c & 1 \end{array} \right)m_0~. \end{equation} that explicitly realizes $\mu-\tau$ symmetry for $a=b=1$. For the overall neutrino scale one naturally expects $m_0\sim\sqrt{\Delta m^2_{ATM}}$, and I also assume \begin{equation}\label{epsilon} \epsilon\sim\sqrt{{\Delta m^2_{sol}}/{\Delta m^2_{ATM}}}= 0.172~. \end{equation} The hierarchies within $M_\nu$ could be realized in flavor models with an approximated $U(1)_{L_e}$ symmetry (additional to $\mu-\tau$), where $\epsilon$ encodes the amount of breaking of $L_e$, but I will not pursue such models here. Upon weak neutrino phase redefinition, one can take $b$ and $\epsilon$ as real numbers, while $a$, $c$ and $d$ remain complex. Their phases would non trivially combine to generate observable phases. However, in order to explain the way neutrino oscillation parameters arise from above ansatz, while keeping the discussion simple, I will assume for the moment non CP violation, and take all matrix elements as real. I shall come back to the complex case latter on. Note that for $a=b=c=d=1$, $M_\nu$ reduces to \begin{equation}\label{M0} \left (\begin{array}{ccc} \epsilon^2 &\epsilon &\epsilon\\ \epsilon & 1 & 1 \\ \epsilon & 1 & 1 \end{array} \right)m_0~, \end{equation} which exhibits $\mu-\tau$ symmetry. Furthermore, its eigenvalues become $m_1 = m_2 = 0$ and $m_3 = (2+\epsilon^2)m_0$, which suggests that $m_0^2\sim \Delta m^2_{ATM}/4$, a value that is consistent with NH, but for the fact that non solar neutrino oscillations scale exist at this level. It is easy to see that this does correspond to the scenario at hand, since at this level $\sin\theta_{12}=0$, whereas for the reactor mixing one gets \begin{equation} \tan\theta_{13} = \frac{\epsilon}{\sqrt{2}}~. \end{equation} Notice that if $\epsilon$ gets the exact value in RHS of Eq.~(\ref{epsilon}), one gets the $\mu-\tau$ symmetric prediction $\tan\theta_{13}^{\tiny{(s)}} = 0.1215$, that results a bit smaller than the central measured value, $\tan\theta_{13}= 0.1506$. Of course, this can be solved with a bit larger value for $\epsilon$. Its actual value, though, would emerge once $\mu-\tau$ gets broken. As already stated above, generating the missing solar neutrino parameters requires to work on the broken symmetry case, although, to avoid large corrections to $\theta_{ATM}$ and $\theta_{13}$ one has to do it perturbatively, in terms of $\epsilon$. To such a purpose, let us come back to $M_\nu$ given in Eq.~(\ref{texture}), and subject its parameters to the convenient condition, \begin{equation} b= 1+\delta_b \epsilon~,\quad\text{and}\quad c= 1- \delta_c \epsilon^2~, \end{equation} with $d$ and $\delta_{b,c}$ order one parameters. On the other hand, $a$ should range within order one to $10^{-1}$. This parameter should be set later on to a value that properly fixes the neutrino observable outcomes. After some algebra, a perturbative diagonalization of the so given $M_\nu$ shows that the predicted mixings, at leading order in $\epsilon$, are \begin{equation} \tan\theta_{ATM}\approx 1 + \frac{\delta_b}{2}\epsilon~,\quad \tan\theta_{13}\approx \frac{\epsilon}{2\sqrt{2}}(1+a)~, \label{thetaatm} \end{equation} and \begin{equation} \tan\theta_{sol}\approx \frac{2\sqrt{2}(1-a)}{\delta_b + \sqrt{\delta_b^2 + 8(1-a)^2}} \label{tansol} \end{equation} Above expressions show that, indeed, the proposed mass matrix provides a correction on the atmospheric mixing in the right direction, tending to increase the angle above $\pi/4$, whereas adding a small factor correction to reactor mixing, as given in the $\mu-\tau$ limit, provided $\delta_b>0$ and $a$ is a small number. More importantly, the given $M_\nu$ predicts a solar mixing that does not depend on the expansion parameter, $\epsilon$, as expected, but links its value with the atmospheric mixing through $\delta_b$ and to reactor mixing through $a$, the actual $\mu-\tau$ breaking parameters. Additionally, the perturbation does lift the degeneracy on the lightest sector, and one gets \begin{equation}\label{deltasol} \Delta m^2_{sol}\approx \epsilon^2 m_0^2\left(\frac{\delta_b}{4}\right) \sqrt{\delta_b^2 + 8(1-a)^2} ~, \end{equation} which is given in terms of the very same parameters that provide solar mixing. On the other hand, for the atmospheric scale one gets, \begin{equation}\label{deltaatm} \Delta m^2_{ATM}\approx (4 + 2\delta_b\epsilon+\frac{1}{8}C\epsilon^2)m_0^2 \end{equation} with $C= 8(1+a)^2 - 4(1-a)^2 +5\delta_b^2 - 32\delta_c - \delta_b\sqrt{\delta_b^2+8(1-a)^2}$. The set of equations (\ref{thetaatm}) - (\ref{deltaatm}) provides a unique solution for the involved mass matrix parameters. Observed mixings set the numerical values for $\delta_b$, $a$ and $\epsilon$, whereas oscillation scales fix the values for $\delta_c$ and $m_0$. A numerical solution, with central values of the oscillation neutrino parameters, gives $a=0.237 $, $\delta_b = 0.92$, $\epsilon =0.34$, $\delta_c = 5.49$, and $m_0 =0.034~eV$~. This outcomes validate our approximation. Notice that $d$ has no relevance in this calculation. It only enters at the $\epsilon^3$ order. Of course, small deviations on above values would still provide neutrino oscillation parameters well within accepted experimental ranges. The small value of $a$ signals that the breaking of $\mu-\tau$ in $m_{e\tau}$ is the strongest one. The correction should amount to reduce an entrance of order $\epsilon$, as in Eq.~(\ref{M0}), down to a value about $\epsilon^2$. Yet, this is still a small correction with respect to the overall $m_0$ scale. One has to keep in mind that this conclusion holds for the non CP violationg case. As it is discussed below, the addition of CP phases would also provide a way to overcome the condition of an small $a$, providing the possibility to account for the neutrino observables even with an small amount of breaking for $\mu-\tau$ symmetry. \section{Stability under RGE evolution} As it s well known~\cite{RGE}, renormalization of neutrino mass may substantially alter mass textures and their predictions when these last heavily relay on the degeneracy of the eigenvalues. (For a recent discussion about the integral solutions to RGEs for neutrino masses and flavor mixing parameters in both the Majorana and Dirac cases see Ref.~\cite{zhang}.) At first glance this could be considered a matter of concern in the present case, where the mass matrix in Eq.~(\ref{texture}) is assumed to be realized in a flavor model that includes $\mu-\tau$ symmetry at some higher energy $\Lambda$, and it has to be addressed by explicitly including the effect or renormalization to run the mass matrix elements down to a lower scale, $\Lambda_0$, usually taken as the electroweak scale. Nonetheless, it should be stressed that even though we started the discussion by requiring degeneracy on two neutrino states, the proposal to understand the solar mixing does also require to break such degeneracy down by opening the mass gap associated to solar neutrino scale, $\Delta m^2_{sol}$. Thus, the texture (\ref{texture}), expected to emerge from the flavor model at the scale $\Lambda$, does not produce degenerated neutrinos, but rather a hierarchical spectrum. Henceforth, one should not expect any mayor distortion to the proposed solution to neutrino masses and mixings from renormalization group equations (RGE). This is explicitly shown next. At one-loop order, the low energy effective neutrino mass is defined from the high energy one by the scaling induced by the RGE, as follows~\cite{RGE}, \begin{equation} M_\nu(\Lambda_0) = I_\gamma~ {\cal I}M_\nu(\Lambda) {\cal I}~, \end{equation} where the diagonal matrix ${\cal I} = Diag\{I_e,I_\mu,I_\tau \}$, whose entries are given by the integral expression \begin{equation} I_\ell = exp\left[-\frac{f}{16\pi^2}\int_{\ln\Lambda_0}^{\ln\Lambda}y_\ell^2~dt\right]~, \end{equation} whereas the overall scaling factor is written as \begin{equation} I_\gamma = exp\left[-\frac{1}{16\pi^2}\int_{\ln\Lambda_0}^{\ln\Lambda}\gamma(t)~dt\right]~. \end{equation} In above, considering the Standard Model, the scale function $\gamma =-3g_2^2 + \lambda + 6 y_t^2 $, where $g_2$ stands for the $SU(2)_L$ guage coupling, $\lambda$ for the Higgs self coupling and $y_{t,\ell}$ for the top and lepton Yukawa couplings, whereas $f = -3/2$. The overall factor $I_\gamma$ contributes by only redefining the overall neutrino scale, and thus, it has to be considered to fix the correct value that should provide the right neutrino oscillation scales at low energy. It has no relevance for renormalization of the mixing angles, though. Flavor dependent corrections, $I_\ell$, on the other hand, break by themselves $\mu-\tau$ symmetry and constitute an additional source that alters the mixings defined at the symmetric limit. However, since lepton Yukawas are rather small and hierarchical, with $y_e<<y_\mu<<y_\tau$, it is expected that $I_\ell$ would be close to unity, with the main contributions coming from tau lepton. In general, one can write $I_\ell=e^{r_\ell}\approx 1 + r_\ell$, where \begin{equation} r_\ell \approx \frac{3}{32\pi^2}y^2_\ell(\Lambda_0)\ln\left(\frac{\Lambda_0}{\Lambda}\right)~. \end{equation} By assuming $\Lambda = \Lambda_{GUT}\sim 10^{16}~GeV$ and $\Lambda_0$ the electroweak scale, one gets that $r_\tau\sim 10^{-5}$, whereas $r_{\mu}\sim 10^{-7}$ and $r_e\sim 10^{-8}$. Under this prescription, taking $M_\nu(\Lambda)$ as the mass matrix given in Eq.~(\ref{texture}), the effective mass matrix at low energy becomes \begin{equation} M_{\nu}(\Lambda_0) = \left (\begin{array}{ccc} d_R\epsilon_R^2 &\epsilon_R &a_R\epsilon_R\\ \epsilon_R & b_R & c_R \\ a_R\epsilon_R & c_R & 1 \end{array} \right)m_{0,R}~, \end{equation} where the index $R$ refers to the renormalized quantities, given as \begin{eqnarray} a_R = \left(\frac{I_\tau}{I_\mu}\right)a~, &\qquad & b_R = \left(\frac{I_\mu}{I_\tau}\right)^2 b~, \\ c_R = \left(\frac{I_\mu}{I_\tau}\right) c~, && d_R = \left(\frac{I_\tau}{I_\mu}\right)^2 d~, \end{eqnarray} \begin{equation} \epsilon_R = \left(\frac{I_eI_\mu}{I_\tau^2}\right)\epsilon~, \end{equation} and where the overall neutrino scale $m_{0R} = (I_\gamma I_\tau^2)\,m_0$. Due to the smallness of $r_{e,\mu}$ with respect to $r_\tau$, it is enough to consider only the effect of the tau lepton corrections in $M_\nu(\Lambda_0)$. In this approach, it is straightforward to see that the renormalized parameters involved in the mixings just get an small factor correction, of order ${\cal O} (10^{-5})$, with respect to those defined at high energy. Explicitly, one gets $a_R \approx (1+r_\tau)a$, $b_R \approx (1-2r_\tau) b$, $c_R \approx (1-r_\tau) c$, $d_R \approx (1+2r_\tau) d$, and $\epsilon_R \approx (1-2r_\tau)\epsilon$. As expected, these effect should not alter the texture itself, nor the predictions of the theory, in any sensible way. As a matter of fact, the formulae (\ref{thetaatm}) and (\ref{tansol}) for the mixing angles would still be valid for the renormalized parameters. This allow to express, to the lowest order, the renormalized mixings as \begin{eqnarray} \tan\theta_{ATM}^R &\approx& \tan\theta_{ATM} - r_\tau~,\\ \tan\theta_{13}^R&\approx&\tan\theta_{13}\left(1-\frac{2+a}{1+a}r_\tau\right) ~, \end{eqnarray} and \begin{equation} \tan\theta_{sol}^R = \tan\theta_{sol}\left(1+\Delta R\right)~, \end{equation} where, to smplify matters, I have used a short hand notation for $\Delta R = \tan\theta_{sol}\left(1+(1+x^2)^{-\frac{1}{2}}\right)\Delta s$, with $x=\delta b/\sqrt{8}(1+a)$ and \begin{equation} \Delta s = \frac{(2-a)\delta_b-2b(1-a)}{(1-a)^2}\frac{r_\tau}{\epsilon}~. \end{equation} This correction term remains small, about $\Delta R\sim 10^{-5}$. Hence, all corrections induced on the mixing angles by running under RGE do came out to be very small. \section{CP violation} Setting back CP phases in $M_\nu$ would slightly modify above formulae for mass scales and mixings, and provide observable phases. As mentioned, relevant phases arise on $a$, $c$ and $d$ parameters, called $\phi_{a,c,d}$ respectively. In order to maintain the predictions given above one should now assume $|b|= 1+\delta_b\epsilon$ and $|c|= 1- \delta_c\epsilon^2$, to get $$ \tan\theta_{ATM}\approx 1+\frac{\delta_b\epsilon}{2|\cos\phi_{c}|}, \quad \text{and}\quad \tan\theta_{13}\approx \frac{\epsilon}{2\sqrt{2}}\frac{|1+a|}{\cos\frac{\phi_c}{2}} $$ Both expressions, as expected, reduce to those in Eq.~(\ref{thetaatm}) for null phases. The small deviation of ATM mixing from $\pi/4$ suggests to take $\phi_c\approx 0$, in which case one only has to replace $1\pm a$ by $|1\pm a|$ in our previous formulae. Finally, Dirac CP phase can be directly calculated from $M_\nu$ by using the Jarlskog invariant~\cite{jarlskog} $J =\sin2\theta_{12}\sin2\theta_{23}\sin\theta_{13}\cos\theta_{13}\sin\delta_{CP}/8$, which can also be written as~\cite{branco} \begin{equation} J = - \frac{Im\left(H_{21}H_{32}H_{13}\right)}{\Delta m^2_{21}\Delta m^2_{31}\Delta m^2_{32}}~. \end{equation} Hence, at order ${\cal{O}} (\epsilon^2)$, setting in the experimental values of neutrino oscillation parameters, one gets \begin{equation} \sin\delta_{CP}\approx -1.886~h_j~, \end{equation} where, $h_j = 2\delta_b\epsilon|a|\sin(\phi_a +\phi_c) -(1+\delta_b\epsilon-2\delta_c\epsilon^2-|a|^2)\sin(2\phi_c) -|d|\epsilon^2 (\sin\phi_d+\sin(\phi_d+2\phi_c))$. This expression depends on the three free phases of $a$, $c$ and $d$. It vanishes when they are null, as it should be, and it can easily accommodate possible values for $\delta_{CP}$ within its expected experimental range. Numerically scanning the parameter space that is constrained by the condition to provide mass scales and mixings within their current 1$\sigma$ values, one gets the exact predictions for $\delta_{CP}$ depicted in figure 1, in terms of the main parameter associated to the breaking of $\mu-\tau$ symmetry, $a$. As it is notorious, a rather order one $a$ can actually account for a CP violating phase well within the present bounds at $3\sigma$ level. This indicates that a solution to neutrino oscillation parameters out of the proposed texture is possible without relaying on a strong breaking of the symmetry in the electron sector. All the points in the plot correspond to relatively small values of $\epsilon$, within the range of $(0.2,0.35)$, and are consistent with order one $\delta_{b,c}$, within the range $(0.6,1)$. Also, they all were found to have a value of $\phi_c$ very close to $0$, confirming the findings of our previous analytical analysis. \begin{figure}[h] \includegraphics[width=.5\textwidth]{figure.pdf} \caption{CP violation phase predictions in terms of the $|a|$ parameter. The values currently excluded at $3\sigma$ level are within the shadowed region. } \label{delta3_CP} \end{figure} \section{Concluding remarks} In summary, approximated $\mu-\tau$ symmetry in the neutrino sector appears as an interesting approach to the understanding of the way all neutrino oscillation parameters arise from neutrino mass matrix. As discussed above, in the exact symmetric limit, there is an appealing and general scenario where $\theta_{23}$ is maximal, and $\theta_{13}$ is already non zero and small. Unlike many other approaches, in this scenario solar neutrino oscillation parameters emerge from the breaking of $\mu-\tau$ symmetry, which also contributes to fix other mixings to their observed values through out small corrections. The presented mass matrix realizes these predictions in the context of normal neutrino hierarchy using a single perturbation parameter, $\epsilon\sim\sqrt{\Delta m^2_{sol}./\Delta m^2_{ATM}}$ and may provide the right amount of CP violation. One interesting implication of this is the fact that $|m_{ee}|\approx |d|\epsilon^2 m_0\sim 2|d|\times 10^{-3}~eV$, and thus, the expected amplitude for neutrinoless double beta decay is predicted to be small. A positive observation of such events in the ongoing experiments would rule out the proposed $M_\nu$. Furthermore, as the present analysis shows, the proposed texture is stable under radiative corrections. Thus, its realization under a possible high energy flavor theory may also deserve further attention. Additionally, it is worth mentioning that the main observations regarding mixing outcomes from $\mu-\tau$ exchange symmetry made in here would also be reached for Dirac neutrinos, since in that case the calculation of the mixing matrix also uses an hermitian squared matrix, $H= M_DM_D^\dagger$, which has a similar form as the one given en Eq.~(\ref{H}) (see for instance Ref.~\cite{luna}). In contrast, these conclusions do not hold for $\mu-\tau$ reflection symmetry~\cite{mtreflection,xing} which adds CP conjugation to $\mu-\tau$ exchange, with Dirac neutrinos, to predict non zero mixings without breaking the symmetry. A way to understand this is to notice that reflection symmetry do break the exchange symmetry through the phase difference among $H_{e\mu}$ and $H_{e\tau} = H_{e\mu}^*$. Of course, since switching off the phases trivially reduces one case into the other, the proposed connection of solar mixing to the origin of solar scale could still be realized in $\mu-\tau$ reflection symmetry models, interestingly, trough CP phases alone. Finally, it is worth mentioning that there are also appropriate mass structures that work for inverted hierarchy, but they will be presented in a forthcoming extended study. \vspace*{1ex} \section*{Acknowledgments} Work partially supported by Conacyt, Mexico, under FORDECYT-PRONACES grant No. 490769.

003700bc07ce0616eccf6dcc22ffb03ac6038b7e

\section{Introduction} Consider a forward contract on electricity or some other non-storable commodity. The forward price is often represented as a random field $f:[0,T]\times \mathbb{R}^+\times\Omega\to\mathbb{R}$, where $T<\infty$ and $(\Omega,\mathcal{F},\mathbb{P})$ is a complete probability space equipped with a $\mathbb{P}$-augmented filtration $\mathbb{F}=\{\mathcal{F}_t\}_{0\leq t \leq T}$. In the Musiela notation, $t\in[0,T]$ is the time horizon and $x=T-t\in\mathbb{R}^+$ represents the time to maturity. Clearly, we can regard the forward price as an infinite dimensional stochastic process $f:[0,T]\times\Omega\to H$, where $H$ is a suitable space of real-valued functions on $\mathbb{R}^+$. This is the approach taken to describe the dynamics of evolution of the forward prices. In this paper we consider an infinite dimensional stochastic volatility model for the forward price. Our interests are motivated by the power markets where there are clear signs of intertemporal correlation structures across time to maturity (see e.g. \cite{BSBK}) and also signs of non-Gaussianity (see e.g. \cite{BP}). We choose to consider the Heston-type infinite dimensional volatility model proposed in \cite{BS}, and study both the pricing of options written on the forwards and provide tools for the sensitivity analysis. For this last one, we have to introduce the adequate concepts in the infinite dimensional setting. We resolve this by exploiting some ideas in the approach of \cite{BDHP}, using the interplay of functional derivatives and the Malliavin calculus via a form of randomization. We remark that there is a structural difference in the application perspectives between our paper and \cite{BDHP}. While we face an infinite dimensional problem all the way through, the authors in \cite{BDHP} consider a finite dimensional noise and path-dependent coefficients. Our paper is organized as follows. In Section \ref{section:vol} we introduce the forward price model. In Section \ref{section:sensitivity} we consider prices of options written on such forward contracts and study the sensitivity of these prices with respect to different model parameters. \section{Stochastic volatility forward price model} \label{section:vol} Let $H$ be a separable Hilbert space with inner product $\<\cdot,\cdot\>_H$ and associated norm $\|\cdot\|_H$. The space of bounded linear operators from $H$ into itself is denoted by $L(H)$. It is a Banach space with the operator norm denoted by $\|\cdot\|_{\text{op}}$. Furthermore, let $\mathcal{H}\subseteq L(H)$ denote the space of Hilbert-Schmidt operators on $H$. Recall that $\mathcal{H}$ is a separable Hilbert space. The inner product and associated norm in $\mathcal{H}$ are denoted by $\<\cdot,\cdot\>_{\mathcal{H}}$ and $\|\cdot\|_{\mathcal{H}}$, respectively. We start by introducing the stochastic volatility model from \cite{BS}. Let $W=\{W_t\}_{t\geq 0}$ and $B=\{B_t\}_{t\geq 0}$ be independent $\mathbb{F}$-adapted Wiener processes in $H$ with covariance operators $Q_W$ and $Q_B$. Let $\mathcal{A}$ and $\mathcal{C}$ be densely defined operators on $H$ which generate $C_0$-semigroups $\{\mathcal{U}_t\}_{t\geq 0}$ and $\{\mathcal{S}_t\}_{t\geq 0}$ respectively, and let $\eta\in L(H)$. Let $Z=\{Z_t\}_{t\geq 0}$ be an $\mathbb{F}$-adapted stochastic process in $H$ satisfying $\|Z_t\|_H=1$ for all $t\geq 0$. The process $X=\{X_t\}_{t\geq 0}$, which will later be used to model the forward price, is defined by the following equation: \begin{equation}\label{YX} \left\{ \begin{alignedat}{2} &dX_t = \mathcal C X_tdt + \Gamma^Z_t dB_t, && \mspace{40mu} X_0=x_0 \in H, \\ &dY_t = \mathcal{A} Y_t\, dt + \eta\, dW_t, && \mspace{40mu} Y_0=y_0 \in H, \end{alignedat} \right. \end{equation} where $\Gamma^Z_t := Z_t\otimes Y_t$ defines the volatility process $\Gamma_Z=\{\Gamma^Z_t\}_{t\geq 0}$ which is $\mathbb{F}$-adapted and takes values in $\mathcal{H}$. From Peszat and Zabczyk~\cite[Sect.~9.4]{PZ} we know that \eqref{YX} has a unique mild solution given by \begin{equation}\label{mild-sol} \left\{ \begin{alignedat}{2} &X_t = \mathcal{S}_t x_0 + \int_0^t \mathcal{S}_{t-s}\Gamma^Z_s\,dB_s, \qquad &t\geq 0, \\ &Y_t = \mathcal{U}_t y_0 + \int_0^t \mathcal{U}_{t-s}\eta\, dW_s, \qquad &t\geq 0, \end{alignedat} \right. \end{equation} since the stochastic integral $\int_0^t \Gamma^Z_s dB_s$ is well-defined by \cite{BS}. The stochastic integral $\int_0^t \mathcal{S}_{t-s}\Gamma^Z_s\,dB_s$ is also well-defined, since $\mathcal{S}_t \in L(H)$ and its operator norm grows at most exponentially by the Hille-Yoshida Theorem, see Engel and Nagel~\cite[Prop.~I.5.5]{EN}. The variance process $\mathcal{V}=\{\mathcal{V}_t\}_{t\geq 0}$ is defined by $\mathcal{V}_t\coloneqq Y_t^{\otimes 2} = Y_t\otimes Y_t = \<Y_t,\,\cdot\,\>_H Y_t$, $t\geq 0$. It is a process of symmetric and positive definite operators in $\mathcal{H}$, and hence it has a unique square root process, which is given by \begin{equation*} \mathcal{V}^{1/2}_t=\left\{\begin{array}{cl}\|Y_t\|_H^{-1}\mathcal{V}_t, & Y_t\neq 0 \\ 0, & Y_t=0.\end{array}\right. \end{equation*} The variance process can be decomposed as $\mathcal{V}_t = \Gamma^Z_t(\Gamma^Z_t)^*$ for all $t\geq 0$, where $(\Gamma^Z_t)^* = Y_t\otimes Z_t$. Clearly, several choices of $\{Z_t\}_{t\geq 0}$ are possible. A simple choice of $\{Z_t\}_{t\geq 0}$ is $Z_t=\gamma\in H$ for all $t\geq 0$, where $\|\gamma\|_H=1$. We could also define $\{Z_t\}_{t\geq 0}$ as $Z_t=Y_t/\|Y_t\|_H$ when $Y_t\neq 0$ and $Z_t=0$ when $Y_t=0$ to get $\Gamma^Z_t = \mathcal{V}^{1/2}_t$ for all $t\geq 0$. We have a full description of the characteristic functional of $X_t$ for $t\geq 0$ (result presented in \cite[Proposition 9]{BS}): \begin{proposition} \label{charfunc_X} Assume that $Z$ is $\mathbb{F}^Y$-adapted, where $\mathbb{F}^Y$ is the filtration generated by $Y$. Then, for any $h\in H$ and $t\geq 0$, \begin{align*} \varphi_{X_t}(h) &\coloneqq\mathbb{E}\left[{\mathrm{e}}^{\mathrm{i}\left<X_t,h\right>_H}\right] \\ &= {\mathrm{e}}^{\mathrm{i}\left<\mathcal{S}_t x_0,h\right>_H}\mathbb{E}\left[\exp\left(-\frac{1}{2}\left\<\left(\int_0^t\|Q_B^{1/2}Z_s\|_H^2\mathcal{S}_{t-s}\mathcal{V}_s\mathcal{S}^*_{t-s}\,ds\right)h,h\right\>_H\right)\right], \end{align*} where the integral above is a Bochner integral in $L(H)$. \end{proposition} From the proposition above, we see that for any $t\geq 0$, $X_t$, conditional on $\mathcal{F}_t^Y$, is a Gaussian $H$-valued random variable. The expectation of $X_t$ is $\mathbb{E}[X_t]=\mathcal{S}_t x_0$ and its covariance operator $Q_{X_t}$ is found in \cite{BS}: \begin{equation} \label{eq:covar-X_t} Q_{X_t}=\int_0^t\mathcal{S}_{t-s}\mathbb{E}\left[\|Q^{1/2}_BZ_s\|_H^2\mathcal{V}_s\right]\mathcal{S}^*_{t-s}\,ds. \end{equation} Hence $X$ is an $H$-valued conditionally Gaussian process. In the particular case when $Z\equiv\gamma\in H$ with $\|\gamma\|_H=1$, we have that \begin{equation*} Q_{X_t}=\int_0^t\mathcal{S}_{t-s}\left((\mathcal{U}_s y_0)^{\x2}+\int_0^s\mathcal{U}_u\eta Q_W\eta^*\mathcal{U}^*_u\,du\right)\mathcal{S}^*_{t-s}\,ds, \end{equation*} which is obtained by applying Theorem 8.7(iv) of Peszat and Zabczyk~\cite{PZ}, see \cite{BS}. \subsection{The forward price model} We shall model the price $F(t,T)$ at time $t$ of a forward contract maturing at time $T$ by choosing a specific Hilbert space $H_w$ and evaluating $X_t$ at the time to maturity. The price $F(t,T)$ will be called the forward price. We will use the Musiela notation in which the forward price is written as $f(t,x)=F(t,t+x)$ where $x=T-t$ is the time to maturity. The Filipovi\'{c} space $H_w$ was first introduced in \cite{filipovic} and is a Hilbert space consisting of real-valued measurable functions. Let $w:\mathbb{R}_+\rightarrow[1,\infty)$ be a measurable and increasing weight function with $w(0)=1$ and $\int_0^{\infty} w^{-1}(x)\,dx<\infty$. Then $H_w$ is defined as the space of functions $h\in L^1_{\text{loc}}(\mathbb{R}^+)$ which possess a weak derivative $h'\in L^1_{\text{loc}}(\mathbb{R}^+)$ such that $\int_0^{\infty} |h'(x)|^2 w(x)\,dx<\infty$. We know that every such function has an absolutely continuous version. The norm on $H_w$ is defined as \begin{equation*} \|h\|_w^2 \coloneqq |h(0)|^2 + \int_0^{\infty} |h'(x)|^2 w(x)\,dx. \end{equation*} The evaluation functional $\delta_x:H_w\rightarrow\mathbb{R}_+$ defined by $\delta_x(h)\coloneqq h(x)$ is a continuous linear functional on $H_w$ for all $x\in\mathbb{R}_+$, see \cite{filipovic}. This means that $\delta_x\in H_w^*$, and hence by the Riesz representation theorem $\delta_x = \<\cdot,h_x\>_w$ for some element $h_x\in H_w$. Let $\|\cdot\|_*$ denote the norm on $H_w^*$. From \cite[Lemma 3.11]{filipovic}, we have \begin{equation} \label{h_x} h_x(y) = 1 + \int_0^{x\wedge y} \frac{1}{w(s)}ds, \quad y\in\mathbb{R}_+, \end{equation} and we see that for $h\in H_w$, \begin{align*} \<h,h_{x}\>_w &= h(0)h_{x}(0) + \int_0^{\infty} w(y)h'(y)h_{x}'(y)\,dy = h(0) + \int_0^{\infty} w(y)h'(y)\frac{1}{w(y)}1_{\{y\leq x\}}\,dy \\ &= h(0) + \int_0^{x} h'(y)\,dy = h(x) = \delta_{x}(h). \end{align*} The following lemma is part of a result in Benth and Kr\"uhner~\cite[Lemma 3.1]{BK-HJM}. \begin{lemma} Let $\delta_x$ be the evaluation functional on $H_w$. Then $\|\delta_x\|_*^2 = h_x(x)$, where $h_x$ is given by equation \eqref{h_x}. \end{lemma} Consider now equation \eqref{YX} with $H$ being the Filipovi\'{c} space $H_w$ and $\mathcal{C}$ being the derivative operator $\partial/\partial x$. The $C_0$-semigroup generated by the derivative operator on $H_w$ is the semigroup of left-shift operators, i.e. $\mathcal{S}_x(h)=h(\cdot+x)$ for $x\geq 0$. It is shown in Filipovi\'{c}~\cite[Equation (5.10)]{filipovic} that for $x\geq 0$, $\|\mathcal{S}_x\|_{\text{op}}\leq C$ for some constant $C$. A value for the constant $C$ is found in Benth and Kr\"uhner~\cite[Lemma 3.4]{BK-2015}: \begin{lemma}\label{shift-op-bound} For $x\geq 0$, it holds that $\|\mathcal{S}_x\|_{\text{op}}\leq \sqrt{2 \max \left(1,\int_0^\infty \frac{1}{w(s)}\,ds\right)}$. \end{lemma} While on $L^2(\mathbb{R})$ the adjoint of the left-shift operator is the right-shift operator, this is not the case on $H_w$. The following lemma gives the adjoint of the left-shift operator on $H_w$. \begin{lemma} For $x\geq 0$, let $\mathcal{S}_x$ denote the left-shift operator on $H_w$, defined by $\mathcal{S}_x h=h(\cdot+x)$. The adjoint operator $\mathcal{S}_x^*$ of $\mathcal{S}_x$, defined by the relation $\<\mathcal{S}_x f,g\>_w = \<f,\mathcal{S}_x^* g\>_w$ for all $f,g\in H_w$, is given by \begin{equation} \label{shift-adjoint} \mathcal{S}_x^* g(y) = \left\{ \begin{alignedat}{2} &g(0)\left(1 + \int_0^y \frac{1}{w(s)}\,ds\right), \qquad &0\leq y \leq x, \\ &g(0)\left(1 + \int_0^x \frac{1}{w(s)}\,ds\right) + \int_0^{y-x} \frac{w(s)}{w(s+x)}g'(s)\,ds, \qquad &y>x. \end{alignedat} \right. \end{equation} \end{lemma} \begin{proof} From Proposition 3.8 in \cite{BK-2015}, we have that for an operator $\mathcal{T}\in L(H_w)$, the adjoint operator $\mathcal{T}^*$ is given by \begin{equation*} \mathcal{T}^* g(y) = g(0)\eta(y) + \int_0^\infty q(y,s)g'(s)\,ds, \end{equation*} where \begin{equation*} \eta(y)\coloneqq (\mathcal{T} h_y)(0) \end{equation*} and \begin{equation*} q(y,s)\coloneqq (\mathcal{T} h_y)'(s)w(s). \end{equation*} For the left-shift operator we find that \begin{equation*} \eta(y) = (\mathcal{S}_x h_y)(0) = h_y(x) = \int_0^{y\wedge x}\frac{1}{w(s)}\,ds,\qquad y\geq 0. \end{equation*} From \cite[p. 78]{filipovic} we have that $(\mathcal{S}_x h)'=\mathcal{S}_x h'$. Hence, for $y,s\geq 0$, \begin{align*} q(y,s) &= (\mathcal{S}_x h_y)'(s) w(s) = \mathcal{S}_x h_y'(s) w(s) \\ &= \mathcal{S}_x\left(\frac{1}{w(s)}1_{\{s\leq y\}}\right) w(s) = \frac{1}{w(s+x)}1_{\{s+x\leq y\}} w(s)\\ &= \frac{w(s)}{w(s+x)}1_{\{s\leq y-x\}}. \end{align*} This gives \begin{equation*} \mathcal{S}_x^* g(y) = g(0)\left(1+\int_0^{y\wedge x}\frac{1}{w(s)}\,ds\right) + \int_0^{(y-x)\vee 0} \frac{w(s)}{w(s+x)}g'(s)\,ds, \end{equation*} and the result follows. \end{proof} We define the forward price as \begin{equation}\label{musiela} f(t,x)\coloneqq \delta_x(X_t), \qquad t\in[0,T],\,x\in\mathbb{R}^+. \end{equation} Observe that the covariance between the prices of two forward contracts with different times to delivery follows from Proposition \ref{charfunc_X}, see \cite{BS}. \begin{corollary} For all $x,y\in\mathbb{R}_+$, we have \begin{align*} \textnormal{Cov}\big(f(t,x),\, f(t,y)\big)=\mathbb{E}\left[\int_0^t\|Q_B^{1/2}Z_s\|_w^2\,\delta_{x+t-s}(Y_s)\delta_{y+t-s}(Y_s)\,ds\right]. \end{align*} \end{corollary} In the special case when $Z_t=\gamma\in H_w$ with $\|\gamma\|_w=1$, it is shown in \cite{BS} that \begin{align*} \textnormal{Cov}(f(t,x),f(t,y))&=\|Q_B^{1/2}\gamma\|_w^2\int_0^t\delta_{y+t-s}(\mathcal{U}_sy_0)^{\otimes 2}\delta_{x+t-s}^*(1)\,ds \\ &\qquad+\|Q_B^{1/2}\gamma\|_w^2\int_0^t\delta_{y+t-s}\left(\int_0^s\mathcal{U}_u\eta Q_W\eta^*\mathcal{U}^*(u)\,du\right)\delta^*_{x+t-s}(1)\,ds, \end{align*} by applying Theorem 8.7(iv) of Peszat and Zabczyk~\cite{PZ}. \subsection{Commodities with delivery period} In the case of commodities with delivery over a period of time $[T_1, T_2]$, as for electricity, the forward price $G(t,T_1,T_2)$ at time $t\leq T_1$ is modelled by \begin{equation*} G(t,T_1,T_2) \coloneqq \frac{1}{T_2-T_1} \int_{T_1}^{T_2} F(t,T)\,dT. \end{equation*} Hereafter, we study different representations of the forward price. For this we introduce two integration functionals, $\mathcal{I}_d$ and $\mathcal{J}_{x,d}$. For $d\geq 0$, $\mathcal{I}_d:H_w\rightarrow \mathbb{R}$ is defined as \begin{equation}\label{I} \mathcal{I}_d(h)\coloneqq \frac{1}{d}\int_0^d h(u)\,du, \qquad h\in H_w. \end{equation} For $x,d\geq 0$, $\mathcal{J}_{x,d}:H_w\rightarrow \mathbb{R}$ is defined as \begin{equation}\label{J} \mathcal{J}_{x,d}(h)\coloneqq \frac{1}{d}\int_0^d \delta_{x+u}(h)\,du, \qquad h\in H_w. \end{equation} We can see that $\mathcal{J}_{x,d} = \mathcal{I}_d\circ \mathcal{S}_x$, where $\mathcal{S}_x$ is the left-shift operator on $H_w$. Using the Musiela notation we let $x$ denote the time to the start of the delivery period and $d$ denote the length of the delivery period. Then $[T_1,T_2]=[t+x,t+x+d]$, and the forward price can be written as \begin{equation}\label{g} g(t,x,d)\coloneqq \mathcal{J}_{x,d}(X_t) = \frac{1}{d} \int_0^d f(t,x+u)\,du, \qquad t\in[0,T_1],\, x\in\mathbb{R}^+, d\in\mathbb{R}^+. \end{equation} The covariance between the prices of two forward contracts with different times $x$ and $y$ to the start of the delivery period and different lengths $d_1$ and $d_2$ of the delivery period was found in \cite{BS}: \begin{equation*} \textnormal{Cov}\big(g(t,x,d_1),\, g(t,y,d_2)\big)=\frac{1}{d_1 d_2}\int_0^{d_1}\int_0^{d_2}\int_0^t\mathbb{E}\left[\|Q_B^{1/2}Z_s\|_w^2\,\delta_{x+t-s+u}(Y_s)\delta_{y+t-s+v}(Y_s)\right]\,ds\,dv\,du. \end{equation*} The next two results present some properties of the functionals $\mathcal{I}_d$ and $\mathcal{J}_{x,d}$ defined in \eqref{I} and \eqref{J}. \begin{lemma} The integration functionals $\mathcal{I}_d$ and $\mathcal{J}_{x,d}$ defined in \eqref{I} and \eqref{J} are well-defined continuous linear functionals on $H_w$, i.e. $\mathcal{I}_d, \mathcal{J}_{x,d}\in H_w^*$. \end{lemma} \begin{proof} Since $h\in H_w$ is locally integrable, $\mathcal{I}_d$ is a well-defined linear functional on $H_w$. We show that it is bounded. For $h\in H_w$, \begin{equation*} |\mathcal{I}_d(h)| \leq \frac{1}{d}\int_0^d|\delta_{u}(h)|\,du \leq \frac{1}{d}\left(\int_0^d \|\delta_{u}\|_*\,du\right)\|h\|_w \end{equation*} holds. Since $x\mapsto \|\delta_x\|_*$ is locally integrable, it follows that $\mathcal{I}_d$ is bounded, i.e. $\mathcal{I}_d\in H_w^*$. Since $\mathcal{S}_t$ is continuous, it follows that $\mathcal{J}_{x,d}\in H_w^*$. \end{proof} \begin{lemma} \label{h_x,d} For any $h\in H_w$, we have that \begin{equation*} \mathcal{J}_{x,d}(h) = \<h, h_{x,d}\>_w, \end{equation*} where \begin{equation} \label{eq:h_x,d} h_{x,d}(y) = \left\{ \begin{alignedat}{3} &1+\int_0^y\frac{1}{w(s)}\,ds, \qquad &0\leq y \leq x, \\ &1+\int_0^x\frac{1}{w(s)}\,ds + \int_0^{y-x} \frac{d-s}{dw(s+x)} \,ds, \qquad &x<y\leq x+d, \\ &1+\int_0^x\frac{1}{w(s)}\,ds + \int_0^d \frac{d-s}{dw(s+x)} \,ds, \qquad &y>x+d. \end{alignedat} \right. \end{equation} \end{lemma} \begin{proof} Since $\mathcal{I}_d\in H_w^*$, we have that $\mathcal{I}_d = \<\cdot,h_d^{\mathcal{I}}\>_w$ for some element $h_d^{\mathcal{I}}\in H_w$. From Filipovi\'{c} \cite[Lemma 3.11]{filipovic} we have \begin{equation} \label{h_d^I} h_d^{\mathcal{I}}(y) = 1 + \frac{1}{d}\int_0^y \frac{d-s\wedge d}{w(s)}\,ds. \end{equation} Since also $\mathcal{J}_{x,d}\in H_w^*$, we have that $\mathcal{J}_{x,d} = \<\cdot,h_{x,d}\>_w$ for some $h_{x,d}\in H_w$. We see that \begin{equation*} \mathcal{J}_{x,d}(h) = \mathcal{I}_d(\mathcal{S}_x(h)) = \left\<\mathcal{S}_x(h),h_d^{\mathcal{I}}\right\>_w = \left\<h,\mathcal{S}_x^*(h_d^{\mathcal{I}})\right\>_w, \end{equation*} which gives $h_{x,d} = \mathcal{S}_x^*(h_d^{\mathcal{I}})$. Since $h_d^{\mathcal{I}}(0)=1$ and \begin{equation*} (h_d^{\mathcal{I}})'(y) = \frac{1}{d}\frac{d-y\wedge d}{w(y)} = \frac{d-y}{dw(y)} 1_{\{0\leq y\leq d\}}, \end{equation*} we get by \eqref{shift-adjoint} that \begin{align*} h_{x,d}(y) &= \mathcal{S}_x^* h_d^{\mathcal{I}}(y) \\ &= h_d^{\mathcal{I}}(0)\left(1+\int_0^{y\wedge x}\frac{1}{w(s)}\,ds\right) + \int_0^{(y-x)\vee 0} \frac{w(s)}{w(s+x)}(h_d^{\mathcal{I}})'(s)\,ds\\ &= 1+\int_0^{y\wedge x}\frac{1}{w(s)}\,ds + \int_0^{(y-x)\vee 0} \frac{w(s)}{w(s+x)}\frac{d-s}{dw(s)} 1_{\{0\leq s\leq d\}}\,ds\\ &= 1+\int_0^{y\wedge x}\frac{1}{w(s)}\,ds + \int_0^{((y-x)\vee 0)\wedge d} \frac{d-s}{dw(s+x)} \,ds. \end{align*} \end{proof} The following corollary gives a representation for the forward price as an inner product between $X_t$ and an element of $H_w$. \begin{corollary} The forward price can be expressed as \begin{equation*} g(t,x,d) = \<X_t, h_{x,d}\>_w, \end{equation*} where the function $h_{x,d}$ is given in \eqref{eq:h_x,d}. \end{corollary} \section{Sensitivity analysis} \label{section:sensitivity} Consider an option written on a forward contract with delivery period $[T_1, T_2]$ as described in the previous section. Let $\tau$ denote the exercise time of the option, where $0\leq \tau\leq T_1$. Let $\Phi:\mathbb{R}\rightarrow\mathbb{R}^{+}$ denote the payoff function, which is assumed to be measurable and of at most linear growth. The price of the option at time $t\leq \tau$ is represented as \begin{equation*} \Pi_t = e^{-r(\tau-t)} \mathbb{E}\left[\Phi(G(\tau,T_1,T_2))\,|\,\mathcal{F}_{t}\right], \end{equation*} where $r$ is a constant instantaneous interest rate. Note that we are taking the expectation under the market probability measure $\mathbb{P}$ and not a risk-neutral probability measure. The reason for this is that electricity is not storable, hence it is not a tradeable asset in the usual sense, see \cite{BSBK}. The standard argument of using a risk-neutral probability measure to prevent arbitrage opportunities is therefore not valid in this case, and any equivalent martingale measure can be used for option pricing. As before we adopt the Musiela notation with $x=T_1-\tau$ and $d=T_2-T_1$. Recalling \eqref{g}, the option price can then be written as \begin{equation} \label{option-price} \Pi_t = e^{-r(\tau-t)} \mathbb{E}\left[\Phi(g(\tau,x,d))\,|\,\mathcal{F}_{t}\right] = e^{-r(\tau-t)} \mathbb{E}\left[(\Phi\circ\mathcal{J}_{x,d})(X_\tau)\,|\,\mathcal{F}_{t}\right], \end{equation} where the functional $\mathcal{J}_{x,d}$ is defined in \eqref{J}. The option price at time $t=0$ is given by \begin{equation} \Pi_0 = {\mathrm{e}}^{-r\tau}\mathbb{E}\left[(\Phi\circ\mathcal{J}_{x,d})(X_\tau)\right]. \end{equation} In sensitivity analysis we are interested in the derivatives of the option price with respect to different parameters of the underlying price model. In finance, these derivatives are called \emph{Greeks} since they are denoted by greek letters. Examples are the delta, which is the derivative with respect to the initial value of the underlying asset, and the vega, which is the derivative with respect to the volatility. Our framework presents "parameters" in infinite dimensions, e.g. the initial forward price is a function in $H_w$. Thus we have to reinterpret the meaning of the Greeks. For the delta, a natural choice is to take inspiration from \cite{BDHP} and interpret it as a directional derivative. For the vega however, there is no natural generalization to our framework. We choose to compute the directional derivatives with respect to the "parameters" $y_0$ and $\eta$ of the stochastic volatility model instead. We now consider the option price at time $t=0$ as a function of the initial value $x_0$ of $X_t$, the initial value $y_0$ of $Y_t$ and the volatility $\eta$ of $Y_t$. In the previous sections we assumed that $\eta\in L(H_w)$. We recall that the space of bounded linear operators on a Hilbert space is not reflexive. In fact, it contains a subspace (the diagonal operators with respect to a given orthonormal basis) which is isomorphic to $\ell^{\infty}$, and $\ell^{\infty}$ is not reflexive, see \cite[Theorem 1.11.16, Proposition 1.11.18 and Example 1.11.23]{Megginson}. Since $L(H_w)$ is not reflexive, it is not a UMD Banach space \cite[p.5]{PV}. We therefore need to have stronger assumptions on $\eta$ to be able to apply the Malliavin calculus in infinite dimensions and in particular to use the chain rule \cite[Proposition 3.8]{PV} for the Malliavin derivative. Hence, in this section, we assume that $\eta\in\mathcal{H}_w\coloneqq L_{HS}(H_w)\subset L(H_w)$. The option price at time $t=0$ is then a functional on $H_w\times H_w \times \mathcal{H}_w$ which takes the form \begin{equation} \label{Pi_0} \Pi_0(x_0,y_0,\eta) = {\mathrm{e}}^{-r\tau}\mathbb{E}\left[(\Phi\circ\mathcal{J}_{x,d})(X_{\tau}(x_0,y_0,\eta))\right] = \mathbb{E}\left[\Psi(X_{\tau}(x_0,y_0,\eta))\right], \end{equation} where the notation $X_{\tau}(x_0,y_0,\eta)$ means that we regard the random variable $X_\tau$ as a function of $x_0$, $y_0$ and $\eta$, and to ease the notation, we have introduced the functional $\Psi$ on $H_w$ as $\Psi\coloneqq{\mathrm{e}}^{-r\tau}(\Phi\circ\mathcal{J}_{x,d})$. We are going to consider $\Pi_0$ as a function of one of the parameters keeping the two others fixed. For this purpose we will use the notation $\Pi_0^{y_0,\eta}$, $\Pi_0^{x_0,\eta}$ and $\Pi_0^{x_0,y_0}$, where the variables in superscript are the ones we keep fixed. The directional derivative $\partial_h\Pi_0^{y_0,\eta}(x_0)$ of $\Pi_0^{y_0,\eta}$ at $x_0\in H_w$ in direction $h\in H_w$ is defined as \begin{equation}\label{directional-derivative} \partial_{h} \Pi_0^{y_0,\eta}(x_0) \coloneqq \frac{d}{d\varepsilon} \Pi_0^{y_0,\eta}(x_0+\varepsilon h) \bigg{|}_{\varepsilon=0} = \lim_{\varepsilon\rightarrow 0} \frac{\Pi_0^{y_0,\eta}(x_0 + \varepsilon h) - \Pi_0^{y_0,\eta}(x_0)}{\varepsilon}, \quad h\in H_w. \end{equation} This will be our interpretation of the Greek delta. The directional derivative $\partial_h\Pi_0^{x_0,\eta}(y_0)$ is defined similarly. The directional derivative of $\Pi_0^{x_0,y_0}$ at $\eta\in \mathcal{H}_w$ in direction $\zeta\in \mathcal{H}_w$ is defined as \begin{equation}\label{directional-derivative-V3} \partial_{\zeta} \Pi_0^{x_0,y_0}(\eta) \coloneqq \frac{d}{d\varepsilon} \Pi_0^{x_0,y_0}(\eta+\varepsilon \zeta) \bigg{|}_{\varepsilon=0} = \lim_{\varepsilon\rightarrow 0} \frac{\Pi_0^{x_0,y_0}(\eta + \varepsilon \zeta) - \Pi_0^{x_0,y_0}(\eta)}{\varepsilon}, \quad \zeta\in \mathcal{H}_w. \end{equation} Before proceeding further, we briefly review some elements of Malliavin calculus on Hilbert spaces. We refer to \cite{N} and \cite{DPO} for an introduction to Malliavin calculus, and to \cite{CT} and \cite{PV} for more details on Malliavin calculus on Hilbert and Banach spaces. \subsection{Some elements of Malliavin calculus} Let us consider an isonormal process $\mathbb{W}$ on a filtered probability space $(\Omega,\mathcal{F},\mathbb{F}, \mathbb{P})$ and some Hilbert space $H$, where $\mathbb{F}=\{\mathcal{F}_t\}_{0\leq t\leq T}$ is the filtration generated by $\mathbb{W}$ and $\mathcal{F}=\mathcal{F}_T$. For all $h_1,\ldots,h_n\in H$, $\mathbb{W}(h_1),\ldots,\mathbb{W}(h_n)$ are jointly normally distributed real-valued random variables with mean zero and $\mathbb{E}[\mathbb{W}(h_i)\mathbb{W}(h_j)] = \<h_i,h_j\>_H$. Let $E$ be another Hilbert space, which is assumed to be separable. An $E$-valued random variable $F$ is called \emph{smooth} if there exists $h_1,\dots,h_n\in H$ such that $F$ can be written on the form $F = f(\mathbb{W}(h_1),\dots,\mathbb{W}(h_n))$, where $f:\mathbb{R}^n\rightarrow E$ is infinitely differentiable and all derivatives are polynomially bounded. The set of smooth $E$-valued random variables is dense in $L^p(\Omega;E)$ for $1\leq p<\infty$. The Malliavin derivative $\mathscr{D} F$ of a smooth $E$-valued random variable $F$ is a random variable taking values in $E\otimes H$, and it is defined as \begin{equation*} \mathscr{D} F \coloneqq \sum_{i=1}^n \frac{\partial f}{\partial x_i}(\mathbb{W}(h_1),\ldots,\mathbb{W}(h_n)) \otimes h_i. \end{equation*} The space $E\otimes H$ can be identified with the space $L_{HS}(H,E)$ of Hilbert-Schmidt operators from $H$ to $E$. In the special case when $E=\mathbb{R}$, $L_{HS}(H,\mathbb{R})$ can be identified with $H$. The Malliavin derivative of a real-valued random variable is therefore an $H$-valued random variable. The Malliavin derivative is closable as an unbounded operator from $L^p(\Omega;E)$ to $L^p(\Omega; E\otimes H)$, and the closure will also be denoted by $\mathscr{D}$. The domain of the closure is denoted by $\mathbb{D}^{1,p}(E)$, and it becomes a Banach space if we endow it with the following norm: \begin{equation*} \| F \|_{\mathbb{D}^{1,p}(E)} \coloneqq \left(\| F \|_{L^p(\Omega;E)}^p + \| \mathscr{D} F \|_{L^p(\Omega;E\otimes H)}^p\right)^{1/p}. \end{equation*} The space $\mathbb{D}^{1,2}(E)$ is a Hilbert space. We will need the following chain rule. We state it here for Hilbert spaces, but it is also valid in the more general case of UMD Banach spaces, see \cite[Proposition 3.8]{PV}. \begin{lemma}[Chain rule]\label{chain-rule} Let $E_1$ and $E_2$ be Hilbert spaces, and suppose $\varphi:E_1 \rightarrow E_2$ is Fr\'{e}chet differentiable with a continuous and bounded derivative. If $F\in \mathbb{D}^{1,p}(E_1)$, then $\varphi(F)\in \mathbb{D}^{1,p}(E_2)$ and \begin{equation*} \mathscr{D}(\varphi(F)) = D\varphi(F) \circ \mathscr{D} F. \end{equation*} \end{lemma} \begin{proof} See \cite[Proposition 3.8]{PV}. \end{proof} The adjoint operator of the Malliavin derivative is denoted by $\delta$ and is often called the \emph{Skorohod integral}. We remark that this is not to be confused with the evaluation functional for which we used the same notation. The domain of $\delta$ is the set of random variables $u\in L^2(\Omega;E\otimes H)$ for which there exists a constant $C\geq 0$ such that \begin{equation*} \left|\mathbb{E}\left[\<\mathscr{D} F,u\>_{E\otimes H}\right]\right| \leq C \left(\mathbb{E}\left[\|F\|_E^2\right]\right)^{1/2}, \end{equation*} for all $F\in \mathbb{D}^{1,2}(E)$. For all $u\in\text{dom}(\delta)$ and $F\in \mathbb{D}^{1,2}(E)$, the following relation holds \begin{equation}\label{duality-formula} \mathbb{E}\left[\<\mathscr{D} F,u\>_{E\otimes H}\right] = \mathbb{E}\left[\<F,\delta(u)\>_E\right]. \end{equation} The following lemmas are stated for the case $E=\mathbb{R}$, since we will only need them in this setting. In this case, $\delta$ is an unbounded operator from $L^2(\Omega;H)$ to $L^2(\Omega)$. The first lemma gives a formula for the Skorohod integral of the product of a real-valued random variable and an $H$-valued random variable. For a more general formulation of this lemma, see \cite[Lemma 4.9]{PV}. \begin{lemma}[Integration by parts]\label{multiplying by scalar} Let $F\in \mathbb{D}^{1,2}(\mathbb{R})$ and $u\in \text{dom}\,\delta$ such that $Fu\in L^2(\Omega;H)$. Then $Fu\in \text{dom}\,\delta$ and the following holds: \begin{equation*} \delta(Fu) = F\delta(u) -\<\mathscr{D} F,u\>_H. \end{equation*} \end{lemma} \begin{proof} See \cite[Proposition 1.3.3]{N}. \end{proof} \begin{lemma}\label{estimate} Let $u\in\mathbb{D}^{1,2}(H)$. Then $u\in \text{dom}\,\delta$ and the following estimate holds: \begin{equation*} \|\delta(u)\|_{L^2(\Omega)}^2 \leq \|u\|_{\mathbb{D}^{1,2}(H)}^2 = \| u \|_{L^2(\Omega;H)}^2 + \| \mathscr{D} u \|_{L^2(\Omega;H\otimes H)}^2. \end{equation*} \end{lemma} \begin{proof} See \cite[Proposition 1.3.1]{N}. \end{proof} \subsection{Sensitivity of $\Pi_0$ with respect to $x_0$, $y_0$ and $\eta$} We now return to the computation of the directional derivatives of $\Pi_0$ defined in \eqref{directional-derivative} and \eqref{directional-derivative-V3}. In the framework of the previous subsection, we choose $H$ to be the Filipovi\'{c} space $H_w$, and as the Hilbert space $E$ we will use either $\mathbb{R}$, $H_w$ or $\mathcal{H}_w$, depending on the context. Let $(\Omega^{B,W},\mathcal{F}^{B,W},\mathbb{P}^{B,W})$ denote the probability space on which $B$ and $W$ are defined. To apply the Malliavin calculus, we introduce an isonormal Gaussian process $\mathbb{W}$ on a filtered probability space $(\Omega^{\mathbb{W}}, \mathcal{F}^{\mathbb{W}}, \mathbb{F}^{\mathbb{W}}, \mathbb{P}^{\mathbb{W}})$ and $H_w$, where $\mathbb{P}^{\mathbb{W}}$ is independent of $\mathbb{P}^{B,W}$, $\mathbb{F}^{\mathbb{W}}=\{\mathcal{F}^{\mathbb{W}}_t\}_{0\leq t\leq T}$ is the filtration generated by $\mathbb{W}$, $\mathcal{F}^{\mathbb{W}}=\mathcal{F}_T^{\mathbb{W}}$. Note that $\mathbb{W}$ is independent of $B$ and $W$. For a random variable $F$, we can write $F(\omega) = F(\omega^{B,W},\omega^{\mathbb{W}})$. When we talk about the Malliavin derivative of $F$, we mean the Malliavin derivative of $F$ with respect to $\mathbb{W}$. For fixed $\omega^{B,W}\in\Omega^{B,W}$ and $\omega^{\mathbb{W}}\in\Omega^{\mathbb{W}}$, we consider $X_{\tau}(\omega^{B,W},\omega^{\mathbb{W}}):H_w\times H_w \times \mathcal{H}_w\rightarrow H_w$ as a function of the initial value $x_0$ of $X_{\tau}$, the initial value $y_0$ of $Y_t$ and the volatility $\eta$ of $Y_t$. Recall that \begin{equation*} X_{\tau}(x_0,y_0,\eta) = \mathcal{S}_\tau x_0 + \int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \left(\mathcal{U}_s y_0 + \int_0^s \mathcal{U}_{s-u}\eta\, dW_u\right)\right)\,dB_s \quad \omega-a.e, \end{equation*} where $\mathcal{S}_t$ is the right-shift operator. Following the notation introduced previously, we let $X_{\tau}^{y_0,\eta}$, $X_{\tau}^{x_0,\eta}$ and $X_{\tau}^{x_0,y_0}$ denote $X_{\tau}$ as a function of $x_0$, $y_0$ and $\eta$ respectively, keeping the two other variables fixed. The Fr\'{e}chet derivatives of $X_{\tau}$ are given in the following lemma. \begin{lemma}[Fr\'{e}chet derivatives of $X_{\tau}$] The Fr\'{e}chet derivatives of $X_{\tau}$ with respect to $x_0$, $y_0$ and $\eta$ are given by: \begin{itemize} \item[(i)] $DX_{\tau}^{y_0,\eta}(x_0)(h) = \mathcal{S}_\tau(h)$, \quad $h\in H_w$. \item[(ii)] $DX_{\tau}^{x_0,\eta}(y_0)(h) = \int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \mathcal{U}_s h\right)\,dB_s$, \quad $\omega-a.e.$, $h\in H_w$. \item[(iii)] $DX_{\tau}^{x_0,y_0}(\eta)(\zeta) = \int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right)\,dB_s$, \quad $\omega-a.e.$, $\zeta\in \mathcal{H}_w$. \end{itemize} \end{lemma} \begin{proof} $ $ \begin{itemize} \item[(i)] The Fr\'{e}chet derivative $DX_{\tau}^{y_0,\eta}(x_0)$ of $X_{\tau}^{y_0,\eta}$ at $x_0\in H_w$ is defined as the bounded linear operator $L:H_w\rightarrow H_w$ such that \begin{equation*} \lim_{\|h\|_w\rightarrow 0} \frac{\|X_{\tau}^{y_0,\eta}(x_0+h)-X_{\tau}^{y_0,\eta}(x_0)-Lh\|_w}{\|h\|_w} = 0\quad \omega-a.e. \end{equation*} Since $X_{\tau}^{y_0,\eta}(x_0+h)-X_{\tau}^{y_0,\eta}(x_0) = \mathcal{S}_\tau(h)$, the result follows. \item[(ii)] Since \begin{equation*} X_{\tau}^{x_0,\eta}(y_0+h)-X_{\tau}^{x_0,\eta}(y_0) = \int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \mathcal{U}_s h\right)\,dB_s\quad \omega-a.e., \end{equation*} for $y_0,h\in H_w$, the result follows from the definition of the Fr\'{e}chet derivative. \item[(iii)] Since \begin{equation*} X_{\tau}^{x_0,y_0}(\eta+\zeta)-X_{\tau}^{x_0,y_0}(\eta) = \int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right)\,dB_s\quad \omega-a.e., \end{equation*} for $\eta,\zeta\in\mathcal{H}_w$, the result follows from the definition of the Fr\'{e}chet derivative. \end{itemize} \end{proof} Assuming that $\Psi$ (see \eqref{Pi_0}) is Fr\'{e}chet differentiable, the composition $\Psi\circ X_{\tau}(\omega):H_w\times H_w \times \mathcal{H}_w\rightarrow\mathbb{R}$ is also Fr\'{e}chet differentiable with respect to each of the variables. For $h\in H_w$, it holds that \begin{equation*} D(\Psi\circ X_{\tau}^{y_0,\eta})(x_0)(h) = \partial_h(\Psi\circ X_{\tau}^{y_0,\eta})(x_0) = \lim_{\varepsilon\rightarrow 0} \frac{\Psi(X_{\tau}^{y_0,\eta}(x_0 + \varepsilon h)) - \Psi(X_{\tau}^{y_0,\eta}(x_0))}{\varepsilon} \quad\omega-a.e. \end{equation*} Taking expectations, we have that \begin{equation*} \mathbb{E}\left[D(\Psi\circ X_{\tau}^{y_0,\eta})(x_0)(h)\right] = \mathbb{E}\left[\lim_{\varepsilon\rightarrow 0} \frac{\Psi(X_{\tau}^{y_0,\eta}(x_0 + \varepsilon h)) - \Psi(X_{\tau}^{y_0,\eta}(x_0))}{\varepsilon}\right]. \end{equation*} From \eqref{directional-derivative} we recall that \begin{equation*} \partial_h \Pi_0^{y_0,\eta}(x_0) = \lim_{\varepsilon\rightarrow 0} \frac{\Pi_0^{y_0,\eta}(x_0 + \varepsilon h) - \Pi_0^{y_0,\eta}(x_0)}{\varepsilon} = \lim_{\varepsilon\rightarrow 0} \mathbb{E}\left[\frac{\Psi(X_{\tau}^{y_0,\eta}(x_0 + \varepsilon h)) - \Psi(X_{\tau}^{y_0,\eta}(x_0))}{\varepsilon}\right]. \end{equation*} The following lemma shows that we can move the limit inside the expectation, and hence express the directional derivatives of $\Pi_0$ in terms of the Fr\'{e}chet derivatives of $\Psi\circ X_{\tau}$. \begin{lemma}\label{step1} Assume that $\Psi$ is Fr\'{e}chet differentiable. Then the following holds: \begin{itemize} \item[(i)] $\partial_h\Pi_0^{y_0,\eta}(x_0) = \mathbb{E}\left[D(\Psi\circ X_{\tau}^{y_0,\eta})(x_0)(h)\right]$, \quad $h\in H_w$. \item[(ii)] $\partial_h\Pi_0^{x_0,\eta}(y_0) = \mathbb{E}\left[D(\Psi\circ X_{\tau}^{x_0,\eta})(y_0)(h)\right]$, \quad $h\in H_w$. \item[(iii)] $\partial_{\zeta}\Pi_0^{x_0,y_0}(\eta) = \mathbb{E}\left[D(\Psi\circ X_{\tau}^{x_0,y_0})(\eta)(\zeta)\right]$, \quad $\zeta\in\mathcal{H}_w$. \end{itemize} \end{lemma} \begin{proof} Let $L_{\Psi}$ denote the Lipschitz constant of $\Psi$. (i): For each $\varepsilon\in\mathbb{R}\backslash\{0\}$ and for almost all $\omega\in\Omega$, we have that \begin{align*} \left|\frac{\Psi(X_{\tau}^{y_0,\eta}(x_0 + \varepsilon h)) - \Psi(X_{\tau}^{y_0,\eta}(x_0))}{\varepsilon}\right| &\leq \frac{L_{\Psi}\|X_{\tau}^{y_0,\eta}(x_0 + \varepsilon h) - X_{\tau}^{y_0,\eta}(x_0)\|_w}{|\varepsilon|}\\ &= \frac{L_{\Psi}\|\mathcal{S}_\tau(x_0 + \varepsilon h) - \mathcal{S}_\tau(x_0)\|_w}{|\varepsilon|}\\ &= \frac{L_{\Psi}\|\mathcal{S}_\tau(\varepsilon h)\|_w}{|\varepsilon|}\\ &\leq L_{\Psi}\|\mathcal{S}_\tau(h)\|_w\\ &\leq L_{\Psi}\|\mathcal{S}_\tau\|_{\text{op}}\|h\|_w\\ &\leq L_{\Psi}C\|h\|_w, \end{align*} where the constant $C$ is given in Lemma \ref{shift-op-bound}. The desired result then follows by the bounded convergence theorem. (ii): For each $\varepsilon\in\mathbb{R}\backslash\{0\}$, we have that \begin{equation*} \left|\frac{\Psi(X_{\tau}^{x_0,\eta}(y_0 + \varepsilon h)) - \Psi(X_{\tau}^{x_0,\eta}(y_0))}{\varepsilon}\right| \leq \frac{L_{\Psi}\|X_{\tau}^{x_0,\eta}(y_0 + \varepsilon h) - X_{\tau}^{x_0,\eta}(y_0)\|_w}{|\varepsilon|}\eqqcolon A_{\varepsilon}, \end{equation*} which means that \begin{equation*} -A_{\varepsilon} \leq \frac{\Psi(X_{\tau}^{x_0,\eta}(y_0 + \varepsilon h)) - \Psi(X_{\tau}^{x_0,\eta}(y_0))}{\varepsilon} \leq A_{\varepsilon}. \end{equation*} Since \begin{equation*} DX_{\tau}^{x_0,\eta}(y_0)(h) = \lim_{\varepsilon\to 0} \frac{X_{\tau}^{x_0,\eta}(y_0 + \varepsilon h) - X_{\tau}^{x_0,\eta}(y_0)}{\varepsilon}\quad\omega-a.e., \end{equation*} it follows by the continuity of the norm $\|\cdot\|_w$ that \begin{equation*} \|DX_{\tau}^{x_0,\eta}(y_0)(h)\|_w = \lim_{\varepsilon\to 0} \frac{\|X_{\tau}^{x_0,\eta}(y_0 + \varepsilon h) - X_{\tau}^{x_0,\eta}(y_0)\|_w}{|\varepsilon|}. \end{equation*} This gives \begin{equation*} \lim_{\varepsilon\to 0} A_{\varepsilon} = L_{\Psi} \|DX_{\tau}^{x_0,\eta}(y_0)(h)\|_w \eqqcolon A\quad\omega-a.e. \end{equation*} If we can show that $A_{\varepsilon}\rightarrow A$ in $L^1(\Omega)$ when $\varepsilon\rightarrow 0$, the desired result will follow by Pratt's lemma, see \cite{Pratt}. To show convergence in $L^1(\Omega)$ we will use Vitali's theorem. We first show that $\|A_{\varepsilon}\|_{L^2(\Omega)}\leq M$ for all $\varepsilon>0$ for some constant $M$. By the It\^{o} isometry and Lemma \ref{shift-op-bound} we have that \begin{align*} \mathbb{E}\left[|A_{\varepsilon}|^2\right] &= \mathbb{E}\left[\frac{L_{\Psi}^2\|X_{\tau}^{x_0,\eta}(y_0 + \varepsilon h) - X_{\tau}^{x_0,\eta}(y_0)\|_w^2}{\varepsilon^2}\right] \\ &= \frac{L_{\Psi}^2}{\varepsilon^2}\mathbb{E}\left[\bigg\|\int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \mathcal{U}_s (\varepsilon h)\right)\,dB_s\bigg\|_w^2\right]\\ &= L_{\Psi}^2\mathbb{E}\left[\bigg\|\int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \mathcal{U}_s h\right)\,dB_s\bigg\|_w^2\right]\\ &= L_{\Psi}^2\mathbb{E}\left[\int_0^\tau \| \mathcal{S}_{\tau-s}\left(Z_s\otimes \mathcal{U}_s h\right)Q_B^{1/2}\|_{\mathcal{H}_w}^2\,ds\right]\\ &\leq L_{\Psi}^2\mathbb{E}\left[\int_0^\tau \|\mathcal{S}_{\tau-s}\|_{\text{op}}^2 \|\left(Z_s\otimes \mathcal{U}_s h\right)Q_B^{1/2}\|_{\mathcal{H}_w}^2\,ds\right]\\ &\leq L_{\Psi}^2\mathbb{E}\left[\int_0^\tau C^2 \|\left(Z_s\otimes \mathcal{U}_s h\right)Q_B^{1/2}\|_{\mathcal{H}_w}^2\,ds\right], \end{align*} since $\|\mathcal{S}_t\|_{\text{op}} \leq C$ where the constant $C$ is given in Lemma \ref{shift-op-bound}. Let $\{e_n\}_{n\in\mathbb{N}}$ be an orthonormal basis of $H_w$. By Parseval's identity we have that \begin{align*} \|\left(Z_s\otimes \mathcal{U}_s h\right)Q_B^{1/2}\|_{\mathcal{H}_w}^2 &= \sum_{n=1}^{\infty} \|\left(Z_s\otimes \mathcal{U}_s h\right)Q_B^{1/2} e_n\|_w^2 \\ &= \sum_{n=1}^{\infty} \|\<Z_s,Q_B^{1/2}e_n\>_w \,\mathcal{U}_s h\|_w^2 \\ &= \|\mathcal{U}_s h\|_w^2 \sum_{n=1}^{\infty} |\<Z_s,Q_B^{1/2}e_n\>_w|^2 \\ &= \|\mathcal{U}_s h\|_w^2 \sum_{n=1}^{\infty} |\<Q_B^{1/2}Z_s,e_n\>_w|^2 \\ &= \|\mathcal{U}_s h\|_w^2 \|Q_B^{1/2}Z_s\|_w^2. \end{align*} By the Hille-Yosida theorem we have that $\|\mathcal{U}_{t}\|_{\text{op}} \leq K{\mathrm{e}}^{kt}$ for some constants $K$ and $k$. Hence \begin{equation*} \|\mathcal{U}_s h\|_w^2 \leq \|\mathcal{U}_{s}\|_{\text{op}}^2 \|h\|_w^2 \leq K^2{\mathrm{e}}^{2ks} \|h\|_w^2. \end{equation*} Let $\{v_n\}_{n\in\mathbb{N}}$ be the orthonormal basis of $H_w$ consisting of eigenvectors of $Q_B$ with corresponding eigenvalues $\{\lambda_n\}_{n\in\mathbb{N}}$. We know that such a basis exists since $Q_B$ is a symmetric, positive definite trace class operator. By Parseval's identity and Cauchy-Schwarz inequality, we find that \begin{align*} \|Q_B^{1/2}Z_s\|_w^2 &= \<Q_B^{1/2}Z_s,Q_B^{1/2}Z_s\>_w = \<Q_B Z_s,Z_s\>_w = \sum_{n=1}^{\infty} \<Z_s,Q_B v_n\>_w \<Z_s,v_n\>_w \\ &= \sum_{n=1}^{\infty} \lambda_n \<Z_s,v_n\>_w^2 \leq \sum_{n=1}^{\infty} \lambda_n \|Z_s\|_w^2 \|v_n\|_w^2 = \sum_{n=1}^{\infty} \lambda_n = \text{Tr}(Q_B). \end{align*} Hence we have that \begin{equation*} \mathbb{E}\left[|A_{\varepsilon}|^2\right] \leq L_{\Psi}^2 C^2 K^2\text{Tr}(Q_B) \|h\|_w^2 \int_0^\tau {\mathrm{e}}^{2c(\tau-s)}e^{2ks}\,ds =: M. \end{equation*} For a fixed $\delta>0$, it then holds by H\"{o}lder's inequality and Markov's inequality that \begin{align*} \lim_{N\rightarrow\infty} \sup_{|\varepsilon|<\delta} \mathbb{E}\left[|A_{\varepsilon}| 1_{\{|A_{\varepsilon}|>N\}}\right] &\leq \lim_{N\rightarrow\infty} \sup_{|\varepsilon|<\delta} \mathbb{E}\left[|A_{\varepsilon}|^2\right]^{1/2} \mathbb{E}\left[1_{\{|A_{\varepsilon}|>N\}}\right]^{1/2}\\ &= \lim_{N\rightarrow\infty} \sup_{|\varepsilon|<\delta} \|A_{\varepsilon}\|_{L^2(\Omega)} \sqrt{\mathbb{P}(|A_{\varepsilon}|>N)}\\ &\leq \lim_{N\rightarrow\infty} \sup_{|\varepsilon|<\delta} \|A_{\varepsilon}\|_{L^2(\Omega)} \frac{\|A_{\varepsilon}\|_{L^2(\Omega)}}{N}\\ &= \lim_{N\rightarrow\infty} \sup_{|\varepsilon|<\delta} \frac{\|A_{\varepsilon}\|_{L^2(\Omega)}^2}{N}\\ &\leq \lim_{N\rightarrow\infty} \frac{M^2}{N}\\ &=0. \end{align*} By Vitali's theorem it follows that $\mathbb{E}[|A_{\varepsilon}-A|]\rightarrow 0$ when $\varepsilon\rightarrow 0$, and by Pratt's lemma it then follows that \begin{equation*} \partial_h V_2(y_0) = \mathbb{E}\left[D(\Psi\circ X_{\tau}^{x_0,\eta})(y_0)(h)\right]. \end{equation*} (iii): This proof follows the same approach as the proof of (ii). For each $\varepsilon\in\mathbb{R}\backslash\{0\}$, we have that \begin{equation*} \left|\frac{\Psi(X_{\tau}^{x_0,y_0}(\eta + \varepsilon \zeta)) - \Psi(X_{\tau}^{x_0,y_0}(\eta))}{\varepsilon}\right| \leq \frac{L_{\Psi}\|X_{\tau}^{x_0,y_0}(\eta + \varepsilon \zeta) - X_{\tau}^{x_0,y_0}(\eta)\|_w}{|\varepsilon|}\eqqcolon A_{\varepsilon}, \end{equation*} which gives \begin{equation*} -A_{\varepsilon} \leq \frac{\Psi(X_{\tau}^{x_0,y_0}(\eta + \varepsilon \zeta)) - \Psi(X_{\tau}^{x_0,y_0}(\eta))}{\varepsilon} \leq A_{\varepsilon}. \end{equation*} By the continuity of the norm $\|\cdot\|_w$ it follows that \begin{equation*} \lim_{\varepsilon\to 0} A_{\varepsilon} = L_{\Psi} \|DX_{\tau}^{x_0,y_0}(\eta)(\zeta)\|_w \eqqcolon A\quad\omega-a.e. \end{equation*} We show that $\|A_{\varepsilon}\|_{L^2(\Omega)}\leq M$ for all $\varepsilon>0$ for some constant $M$. By the It\^{o} isometry, the Hille-Yosida thorem and Fubini we have that \begin{align*} \mathbb{E}\left[|A_{\varepsilon}|^2\right] &= \mathbb{E}\left[\frac{L_{\Psi}^2\|X_{\tau}^{x_0,y_0}(\eta + \varepsilon \zeta) - X_{\tau}^{x_0,y_0}(\eta)\|_w^2}{\varepsilon^2}\right] \\ &= \frac{L_{\Psi}^2}{\varepsilon^2}\mathbb{E}\left[\bigg\|\int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \int_0^s \mathcal{U}_{s-u}\varepsilon\zeta\, dW_u\right)\,dB_s\bigg\|_w^2\right]\\ &= L_{\Psi}^2\mathbb{E}\left[\bigg\|\int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right)\,dB_s\bigg\|_w^2\right]\\ &= L_{\Psi}^2\mathbb{E}\left[\int_0^\tau \left\| \mathcal{S}_{\tau-s}\left(Z_s\otimes \int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right)Q_B^{1/2}\right\|_{\mathcal{H}_w}^2\,ds\right]\\ &\leq L_{\Psi}^2\mathbb{E}\left[\int_0^\tau \|\mathcal{S}_{\tau-s}\|_{\text{op}}^2 \left\|\left(Z_s\otimes \int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right)Q_B^{1/2}\right\|_{\mathcal{H}_w}^2\,ds\right]\\ &\leq L_{\Psi}^2\mathbb{E}\left[\int_0^\tau C^2{\mathrm{e}}^{2c(\tau-s)} \left\|\left(Z_s\otimes \int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right)Q_B^{1/2}\right\|_{\mathcal{H}_w}^2\,ds\right]\\ &= L_{\Psi}^2 C^2 \int_0^\tau {\mathrm{e}}^{2c(\tau-s)} \mathbb{E}\left[\left\|\left(Z_s\otimes \int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right)Q_B^{1/2}\right\|_{\mathcal{H}_w}^2\right]\,ds. \end{align*} Let $\{e_n\}_{n\in\mathbb{N}}$ be an orthonormal basis of $H_w$. By Parseval's identity we have that \begin{align*} \left\|\left(Z_s\otimes \int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right)Q_B^{1/2}\right\|_{\mathcal{H}_w}^2 &= \sum_{n=1}^{\infty} \left\|\left(Z_s\otimes \int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right)Q_B^{1/2} e_n\right\|_w^2 \\ &= \sum_{n=1}^{\infty} \left\|\<Z_s,Q_B^{1/2}e_n\>_w \,\int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right\|_w^2 \\ &= \left\|\int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right\|_w^2 \sum_{n=1}^{\infty} |\<Z_s,Q_B^{1/2}e_n\>_w|^2 \\ &= \left\|\int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right\|_w^2 \sum_{n=1}^{\infty} |\<Q_B^{1/2}Z_s,e_n\>_w|^2 \\ &= \left\|\int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right\|_w^2 \left\|Q_B^{1/2}Z_s\right\|_w^2. \end{align*} By the It\^{o} isometry and the Hille-Yosida theorem we have that \begin{align*} \mathbb{E}\left[\left\|\int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right\|_w^2\right] &= \mathbb{E}\left[\int_0^s \left\|\mathcal{U}_{s-u}\zeta Q_W^{1/2}\right\|_{\mathcal{H}_w}^2 du\right]\\ &\leq \mathbb{E}\left[\int_0^s \|\mathcal{U}_{s-u}\|_{\text{op}}^2 \|\zeta\|_{\text{op}}^2 \|Q_W^{1/2}\|_{\mathcal{H}_w}^2 du\right]\\ &= \text{Tr}(Q_W) \|\zeta\|_{\text{op}}^2 \int_0^s \|\mathcal{U}_{s-u}\|_{\text{op}}^2 du\\ &\leq \text{Tr}(Q_W) \|\zeta\|_{\text{op}}^2 K^2 \int_0^s e^{2k(s-u)} du. \end{align*} Since $\|Q_B^{1/2}Z_s\|_w^2 \leq \text{Tr}(Q_B)$, it follows that \begin{equation*} \mathbb{E}\left[|A_{\varepsilon}|^2\right] \leq L_{\Psi}^2 C^2 K^2\text{Tr}(Q_B)\text{Tr}(Q_W) \|\zeta\|_{\text{op}}^2 \int_0^\tau \int_0^s {\mathrm{e}}^{2c(\tau-s)}e^{2k(s-u)}\,duds =: M. \end{equation*} We have now shown that there exists a constant $M$ such that $\|A_{\varepsilon}\|_{L^2(\Omega)}\leq M$ for all $\varepsilon>0$. By the same argument as in the proof of (ii), it then follows from Vitali's theorem that $A_{\varepsilon}\rightarrow A$ in $L^1(\Omega)$ when $\varepsilon\rightarrow 0$. The desired result then follows by Pratt's lemma. \end{proof} The following lemma gives expressions for the expectation of the Fr\'{e}chet derivatives of $\Psi\circ X_{\tau}$. The trick we use to compute these Fr\'{e}chet derivatives is to randomize the parameter that we want to differentiate with respect to with an $H_w$-valued noise independent of $B$ and $W$. By applying the chain rule we can then express the Malliavin derivative with respect to this noise in terms of the Fr\'{e}chet derivative that we want to compute. \begin{lemma}\label{technical-lemma} Let $\xi$ be a real-valued random variable on $(\Omega^{\mathbb{W}}, \mathcal{F}^{\mathbb{W}}, \mathbb{P}^{\mathbb{W}})$. Assume that $\xi\in \mathbb{D}^{1,2}(\mathbb{R}^+)$ and that $\mathscr{D}\xi(\omega,x)\neq 0$ for all $x\in\mathbb{R}^+$ and almost all $\omega\in\Omega$. Let $x\in\mathbb{R}^+$, and assume that $\frac{\xi}{\mathscr{D}\xi(x)} h_x$ is Skorohod integrable and that the evaluation of the Skorohod integral at $\lambda=\frac{1}{\xi}$ is well defined. Assume also that the Skorohod integrals below and their evaluations at $\lambda=\frac{1}{\xi}$ are well defined. Then it holds that \begin{itemize} \item[(i)] $\mathbb{E}\left[D(\Psi\circ X_{\tau}^{y_0,\eta})(x_0)(h)\right] = -\mathbb{E}\left[\left\{\delta\left(\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda\xi h)) \frac{\xi}{\mathscr{D}\xi(x)}h_x\right)\right\}\bigg{|}_{\lambda=\frac{1}{\xi}}\right]$, \item[(ii)] $\mathbb{E}\left[D(\Psi\circ X_{\tau}^{x_0,\eta})(y_0)(h)\right] = -\mathbb{E}\left[\left\{\delta\left(\Psi(X_{\tau}^{y_0,\eta}(y_0-h+\lambda\xi h)) \frac{\xi}{\mathscr{D}\xi(x)}h_x\right)\right\}\bigg{|}_{\lambda=\frac{1}{\xi}}\right]$, \item[(iii)] $\mathbb{E}\left[D(\Psi\circ X_{\tau}^{x_0,y_0})(\eta)(\zeta)\right] = -\mathbb{E}\left[\left\{\delta\left(\Psi(X_{\tau}^{x_0,y_0}(\eta-\zeta+\lambda\xi \zeta)) \frac{\xi}{\mathscr{D}\xi(x)}h_x\right)\right\}\bigg{|}_{\lambda=\frac{1}{\xi}}\right]$. \end{itemize} \end{lemma} \begin{proof} We only prove (i) here. The proofs of (ii) and (iii) follow the same approach. For (iii), note that we can still use the chain rule in Theorem \ref{chain-rule} since we assumed $\eta\in \mathcal{H}_w$. For a general $\eta\in L(H_w)$ we can not use the chain rule \cite[Proposition 3.8]{PV} since $L(H_w)$ is not reflexive and hence not a UMD Banach space. Define first a function $\theta:\mathbb{R}^+\rightarrow H_w$ by $\theta(y) = x_0-h+\lambda y h$ for some $\lambda\in\mathbb{R}$ and $h\in H_w$. The randomized initial path is then defined as $X_0\coloneqq \theta(\xi)$. The Fr\'{e}chet derivative $D\theta(y)$ of $\theta$ at $y$ is the bounded linear operator $D\theta(y):\mathbb{R}^+\rightarrow H_w$ given by $D\theta(y)(x) = \lambda x h$. By the chain rule in Lemma \ref{chain-rule} it holds that $\theta(\xi)\in\mathbb{D}^{1,2}(H_w)$ and \begin{equation}\label{chain-rule-eta} \mathscr{D} X_0 = \mathscr{D}(\theta(\xi)) = D\theta(\xi) \circ \mathscr{D} \xi \qquad\omega-a.e. \end{equation} The Malliavin derivative $\mathscr{D}\xi$ is an $H_w$-valued random variable. For any $x\in\mathbb{R}^+$ we have that \begin{equation}\label{chain-rule-eta2} \mathscr{D} X_0(\omega,x) = D\theta(\xi(\omega))(\mathscr{D}\xi(\omega,x)) = \lambda \mathscr{D}\xi(\omega,x) h \qquad\omega-a.e. \end{equation} By the chain rule in Lemma \ref{chain-rule} it holds that $X_{\tau}^{y_0,\eta}(X_0)\in\mathbb{D}^{1,2}(H_w)$. Since the Fr\'{e}chet derivative of $\Psi$ is also bounded and continuous by assumption, we also have $\Psi(X_{\tau}^{y_0,\eta}(X_0))\in\mathbb{D}^{1,2}(\mathbb{R}^+)$ by the chain rule. The chain rule in Lemma \ref{chain-rule} also gives that \begin{equation}\label{chain-rule-X_t} \mathscr{D}(\Psi(X_{\tau}^{y_0,\eta}(X_0))) = D\Psi(X_{\tau}^{y_0,\eta}(X_0)) \circ DX_{\tau}^{y_0,\eta}(X_0) \circ D\theta(\xi) \circ \mathscr{D}\xi \qquad\omega-a.e., \end{equation} and therefore for any $x\in\mathbb{R}^+$ and for almost all $\omega\in\Omega$ it holds that \begin{align*} \mathscr{D} (\Psi(X_{\tau}^{y_0,\eta}(X_0)))(\omega,x) &= D\Psi(X_{\tau}^{y_0,\eta}(X_0(\omega)))\big(DX_{\tau}^{y_0,\eta}(X_0(\omega))(D\theta(\xi(\omega))(\mathscr{D} \xi(\omega,x)))\big)\\ &= D\Psi(X_{\tau}^{y_0,\eta}(X_0(\omega)))\big(DX_{\tau}^{y_0,\eta}(X_0(\omega))(\lambda \mathscr{D}\xi(\omega,x) h)\big)\\ &= \lambda \mathscr{D}\xi(\omega,x) D\Psi(X_{\tau}^{y_0,\eta}(X_0(\omega)))\big(DX_{\tau}^{y_0,\eta}(X_0(\omega))(h)\big), \end{align*} where we used \eqref{chain-rule-eta2} and then factored out the scalar $\lambda\mathscr{D}\xi(\omega,x)$ since the Fr\'{e}chet derivative is linear. By assumption, $\mathscr{D}\xi(\omega,x)\neq 0$ for all $x\in\mathbb{R}^+$ and almost all $\omega\in\Omega$. Multiplying with $\frac{\xi(\omega)}{\mathscr{D}\xi(\omega,x)}$ in the above equation, we have that for any $x\in\mathbb{R}^+$ it holds that \begin{equation*} \frac{\xi(\omega)}{\mathscr{D}\xi(\omega,x)} \mathscr{D} (\Psi(X_{\tau}^{y_0,\eta}(X_0)))(\omega,x) = \lambda\xi(\omega) D\Psi(X_{\tau}^{y_0,\eta}(X_0(\omega)))(DX_{\tau}^{y_0,\eta}(X_0(\omega))(h)) \qquad\omega-a.e. \end{equation*} For a fixed pair $(\omega, x)$ we evaluate the above expression at $\lambda=\frac{1}{\xi(\omega)}$, recalling that $X_0 = x_0-h+\lambda\xi h$. It holds for almost all $\omega\in\Omega$ that \begin{align} \label{*} &\left\{\frac{\xi(\omega)}{\mathscr{D}\xi(\omega,x)} \mathscr{D} (\Psi(X_{\tau}^{y_0,\eta}(X_0)))(\omega,x)\right\}\bigg{|}_{\lambda=\frac{1}{\xi(\omega)}} \\ \notag &= \lambda\xi(\omega) D\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda\xi(\omega) h))(DX_{\tau}^{y_0,\eta}(x_0-h+\lambda\xi(\omega) h)(h))\bigg{|}_{\lambda=\frac{1}{\xi(\omega)}} \\ \notag &= D\Psi(X_{\tau}^{y_0,\eta}(x_0))(DX_{\tau}^{y_0,\eta}(x_0)(h)). \end{align} By the chain rule for Fr\'{e}chet derivatives, equation \eqref{*} and the fact that evaluating a function in the Filipovi\'{c} space $H_w$ at $x\in\mathbb{R}^+$ corresponds to taking the inner product with the function $h_x$ defined in \eqref{h_x}, we get that for almost all $\omega\in\Omega$, \begin{align*} D(\Psi\circ X_{\tau}^{y_0,\eta})(x_0)(h) &= D\Psi(X_{\tau}^{y_0,\eta}(x_0))(DX_{\tau}^{y_0,\eta}(x_0)(h))\\ &= \left\{\frac{\xi}{\mathscr{D}\xi(x)} \mathscr{D} (\Psi(X_{\tau}^{y_0,\eta}(X_0)))(x)\right\}\bigg{|}_{\lambda=\frac{1}{\xi}}\\ &= \left\{\frac{\xi}{\mathscr{D}\xi(x)}\Big\<\mathscr{D} (\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda\xi h))),h_x\Big\>_w\right\}\bigg{|}_{\lambda=\frac{1}{\xi}}\\ &= \left\<\mathscr{D} (\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda\xi h))),\frac{\xi}{\mathscr{D}\xi(x)} h_x\right\>_w\bigg{|}_{\lambda=\frac{1}{\xi}}. \end{align*} Finally we compute the expectation of the Fr\'{e}chet derivative above. Note that since we have an evaluation in the expression above, we can not apply the duality formula \eqref{duality-formula} here. We apply instead Lemma \ref{multiplying by scalar} with $F = \Psi(X_{\tau}^{y_0,\eta}(X_0))$ and $u = \frac{\xi}{\mathscr{D}\xi(x)} h_x$, and get that \begin{align*} &\mathbb{E}\left[D(\Psi\circ X_{\tau}^{y_0,\eta})(x_0)(h)\right] \\ &= \mathbb{E}\left[\left\<\mathscr{D} (\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda\xi h))),\frac{\xi}{\mathscr{D}\xi(x)} h_x\right\>_w\bigg{|}_{\lambda=\frac{1}{\xi}}\right]\\ &= \mathbb{E}\left[\left\{\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda\xi h)) \delta \left(\frac{\xi}{\mathscr{D}\xi(x)} h_x\right) - \delta\left(\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda\xi h)) \frac{\xi}{\mathscr{D}\xi(x)} h_x\right)\right\}\bigg{|}_{\lambda=\frac{1}{\xi}}\right]\\ &= \mathbb{E}\left[\Psi(X_{\tau}^{y_0,\eta}(x_0)) \delta \left(\frac{\xi}{\mathscr{D}\xi(x)} h_x\right) - \left\{\delta\left(\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda\xi h)) \frac{\xi}{\mathscr{D}\xi(x)} h_x\right)\right\}\bigg{|}_{\lambda=\frac{1}{\xi}}\right]\\ &= -\mathbb{E}\left[\left\{\delta\left(\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda\xi h)) \frac{\xi}{\mathscr{D}\xi(x)} h_x\right)\right\}\bigg{|}_{\lambda=\frac{1}{\xi}}\right], \end{align*} where we in last equality used that $\Psi(X_{\tau}^{y_0,\eta}(x_0))$ and $\delta \left(\frac{\xi}{\mathscr{D}\xi(x)} h_x\right)$ are independent since $\Psi(X_{\tau}^{y_0,\eta}(x_0))$ is $\mathcal{F}^B$-measurable and $\delta \left(\frac{\xi}{\mathscr{D}\xi(x)} h_x\right)$ is $\mathcal{F}^{\mathbb{W}}$-measurable, and $\mathbb{E}\left[\delta \left(\frac{\xi}{\mathscr{D}\xi(x)} h_x\right)\right]=0$. \end{proof} Our final step is to choose a specific $\xi$ which satisfies the assumptions in Lemma \ref{technical-lemma}. \begin{theorem} \label{thm:sensitivity:x0} Assume that $\Psi$ is Lipschitz continuous and Fr\'{e}chet differentiable, and that the Fr\'{e}chet derivative of $\Psi$ is bounded and Lipschitz continuous. Let $\xi = \exp(\mathbb{W}(1_{[0,\tau]}))$, and let $x\in\mathbb{R}^+$. \begin{itemize} \item[(i)] For all $h\in H_w$ it holds that \begin{equation*} \partial_h\Pi_0^{y_0,\eta}(x_0) = -\mathbb{E}\left[\Big\{\delta\left(\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda\xi h)) h_x \right)\Big\}\bigg{|}_{\lambda=\frac{1}{\xi}}\right]. \end{equation*} \item[(ii)] For all $h\in H_w$ it holds that \begin{equation*} \partial_h\Pi_0^{x_0,\eta}(y_0) = -\mathbb{E}\left[\Big\{\delta\left(\Psi(X_{\tau}^{x_0,\eta}(y_0-h+\lambda\xi h)) h_x \right)\Big\}\bigg{|}_{\lambda=\frac{1}{\xi}}\right]. \end{equation*} \item[(iii)] For all $\zeta\in \mathcal{H}_w$ it holds that \begin{equation*} \partial_h\Pi_0^{x_0,y_0}(\eta) = -\mathbb{E}\left[\Big\{\delta\left(\Psi(X_{\tau}^{x_0,y_0}(\eta-\zeta+\lambda\xi \zeta)) h_x \right)\Big\}\bigg{|}_{\lambda=\frac{1}{\xi}}\right]. \end{equation*} \end{itemize} \end{theorem} \begin{proof} Choose $\xi = \exp(\mathbb{W}(1_{[0,T]}))$ for some $T>0$. Then $\xi\in \mathbb{D}^{1,2}(\mathbb{R}^+)$ and $\mathscr{D}\xi(y) = \xi$ for $y\in[0,T]$. \noindent\textbf{Proof of (i):} Let \begin{equation*} u(z,\lambda) \coloneqq \Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda\xi h)) h_x(z) \end{equation*} for $z\in\mathbb{R}^+$ and $\lambda\in\mathbb{R}$. Since $\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda\xi h))\in\mathbb{D}^{1,2}(\mathbb{R})$ and $h_x$ is deterministic, $u(\cdot,\lambda)$ is Skorohod integrable for all $\lambda\in\mathbb{R}$. We need to show that the evaluation of the Skorohod integral at $\lambda=\frac{1}{\xi}$ is well defined. If for all $\Lambda>0$, there exists a $C>0$ such that \begin{equation} \label{kolmogorov-condition} \mathbb{E}\left[\left|\delta\big(u(\cdot,\lambda_1)-u(\cdot,\lambda_2)\big)\right|^2\right] < C|\lambda_1 - \lambda_2|^2, \end{equation} for all $\lambda_1$, $\lambda_2$ $\in$ supp $(\xi^{-1})$ with $|\lambda_1|$, $|\lambda_2|<\Lambda$, then the process $\lambda\mapsto\delta(u(\cdot,\lambda))$ has a continuous version by Kolmogorov's continuity theorem, and the evaluation at $\lambda = \frac{1}{\xi}$ is well-defined. We will now verify that \eqref{kolmogorov-condition} holds. Choose $\Lambda>0$ and $\lambda_1$, $\lambda_2$ $\in$ supp $(\xi^{-1})$ with $|\lambda_1|$, $|\lambda_2|<\Lambda$. By Lemma \ref{estimate} we have that \begin{equation} \label{1.47} \mathbb{E}\left[\left|\delta\big(u(\cdot,\lambda_1)-u(\cdot,\lambda_2)\big)\right|^2\right] \leq \|u(\cdot,\lambda_1)-u(\cdot,\lambda_2)\|_{\mathbb{D}^{1,2}(H_w)}^2, \end{equation} where we recall that \begin{equation*} \|u(\cdot,\lambda_1)-u(\cdot,\lambda_2)\|_{\mathbb{D}^{1,2}(H_w)}^2 = \| u(\cdot,\lambda_1)-u(\cdot,\lambda_2) \|_{L^2(\Omega;H_w)}^2 + \| \mathscr{D} \left(u(\cdot,\lambda_1)-u(\cdot,\lambda_2)\right) \|_{L^2(\Omega;H_w\otimes H_w)}^2. \end{equation*} We have that \begin{align*} &\|X_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h) - X_{\tau}^{y_0,\eta}(x_0-h+\lambda_2\xi h)\|_{L^2(\Omega;H_w)}^2\\ &= \|\mathcal{S}_\tau(x_0-h+\lambda_1\xi h) - \mathcal{S}_\tau(x_0-h+\lambda_2\xi h)\|_{L^2(\Omega;H_w)}^2\\ &= \|(\lambda_1-\lambda_2)\xi\mathcal{S}_\tau(h) \|_{L^2(\Omega;H_w)}^2\\ &\leq |\lambda_1-\lambda_2|^2 \|\mathcal{S}_\tau(h) \|_w^2 \mathbb{E}[|\xi|^2]\\ &= \|h \|_w^2 \mathbb{E}[|\xi|^2] |\lambda_1-\lambda_2|^2. \end{align*} Since $\Psi$ is Lipscitz with Lipschitz constant $L_{\Psi}$, it follows that \begin{align*} &\mathbb{E}\left[\big|\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h)) - \Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_2\xi h))\big|^2\right] \\ &\leq L_{\Psi}^2 \|X_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h) - X_{\tau}^{y_0,\eta}(x_0-h+\lambda_2\xi h)\|_{L^2(\Omega;H_w)}^2\\ &\leq L_{\Psi}^2 \|h \|_w^2 \mathbb{E}[|\xi|^2] |\lambda_1-\lambda_2|^2. \end{align*} Then we have \begin{align*} \| u(\cdot,\lambda_1)-u(\cdot,\lambda_2) \|_{L^2(\Omega;H_w)}^2 &= \|\left(\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h)) - \Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_2\xi h)) \right) h_x \|_{L^2(\Omega;H_w)}^2\\ &= \|h_x\|_w^2 \mathbb{E}\left[\big|\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h)) - \Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_2\xi h))\big|^2\right] \\ &\leq L_{\Psi}^2 \|h_x\|_w^2 \|h \|_w^2 \mathbb{E}[|\xi|^2] |\lambda_1-\lambda_2|^2. \end{align*} By the chain rule in Lemma \ref{chain-rule} and using that we have chosen $\xi$ such that $\mathscr{D}\xi(y) = \xi$ for all $y\in[0,T]$, the following holds for all $y\in[0,T]$ \begin{align*} &\big|\mathscr{D}\big(\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h)) - \Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_2\xi h))\big)(y)\big|^2 \\ &= \big|\mathscr{D}\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h))(x) - \mathscr{D}\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_2\xi h))(y)\big|^2 \\ &= \big|D\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h))(DX_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h)(\lambda_1\mathscr{D}\xi(y) h)) \\ &\qquad - D\Psi(X_{\tau}^{y_0,\eta}(x_0-h+ \lambda_2\xi h))(DX_{\tau}^{y_0,\eta}(x_0-h+\lambda_2\xi h)(\lambda_2\mathscr{D}\xi(y) h))\big|^2 \\ &= \big|\lambda_1\xi D\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h))(DX_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h)(h)) \\ &\quad - \lambda_2\xi D\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_2\xi h))(DX_{\tau}^{y_0,\eta}(x_0-h+\lambda_2\xi h)(h))\big|^2. \end{align*} Next, we use that $A_1 x_1 - A_2 x_2 = (A_1-A_2)x_1 + A_2 (x_1-x_2)$ with $A_i = D\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_i\xi h))$ and $x_i = \lambda_i\xi DX_{\tau}^{y_0,\eta}(x_0-h+\lambda_i\xi h)(h)$ for $i=1,2$. We also use the property $|a+b|^2\leq 2|a|^2 + 2|b|^2$, the assumption that $D\Psi$ is Lipschitz continuous with Lipschitz constant $L_{D\Psi}$ and that $DX_{\tau}^{y_0,\eta}(h)(g) = \mathcal{S}_\tau(g)$. \begin{align*} &\mathbb{E}\Big[\big|\mathscr{D}\big(\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h)) - \Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_2\xi h))\big)(y)\big|^2\Big] \\ &= \mathbb{E}\Big[ \big|\big(D\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h))-D\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_2\xi h))\big)(\lambda_1\xi DX_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h)(h)) \\ &+ D\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_2\xi h))\big(\lambda_1\xi DX_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h)(h)-\lambda_2\xi DX_{\tau}^{y_0,\eta}(x_0-h+\lambda_2\xi h)(h)\big)\big|^2\Big]\\ &\leq 2 \mathbb{E}\Big[\big|\big(D\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h))-D\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_2\xi h))\big)(\lambda_1\xi DX_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h)(h))\big|^2 \Big]\\ &+ 2 \mathbb{E}\Big[\big|D\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_2\xi h))\big(\lambda_1\xi DX_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h)(h)-\lambda_2\xi DX_{\tau}^{y_0,\eta}(x_0-h+\lambda_2\xi h)(h)\big)\big|^2\Big]\\ &\leq 2\mathbb{E}\Big[\big\|D\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h))-D\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_2\xi h))\big\|_{\text{op}}^2\big\|\lambda_1\xi DX_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h)(h)\big\|_w^2 \Big]\\ &+ 2 \mathbb{E}\Big[\big\|D\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_2\xi h))\big\|_{\text{op}}^2\big\|\lambda_1\xi DX_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h)(h)-\lambda_2\xi DX_{\tau}^{y_0,\eta}(x_0-h+\lambda_2\xi h)(h)\big\|_w^2\Big]\\ &\leq 2 L_{D\Psi}^2 \mathbb{E}\Big[\big\|X_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h)-X_{\tau}^{y_0,\eta}(x_0-h+\lambda_2\xi h)\big\|_w^2\big\|\lambda_1\xi DX_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h)(h)\big\|_w^2 \Big]\\ &+ 2 L_{\Psi}^2\mathbb{E}\Big[\big\|\lambda_1\xi DX_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h)(h)-\lambda_2\xi DX_{\tau}^{y_0,\eta}(x_0-h+\lambda_2\xi h)(h)\big\|_w^2\Big]\\ &\leq 2 L_{D\Psi}^2 \mathbb{E}\Big[\big\|(\lambda_1-\lambda_2)\xi\mathcal{S}_\tau(h)\big\|_w^2\big\|\lambda_1\xi \mathcal{S}_\tau(h)\big\|_w^2 \Big] + 2 L_{\Psi}^2\mathbb{E}\Big[\big\|(\lambda_1-\lambda_2)\xi\mathcal{S}_\tau(h)\big\|_w^2\Big]\\ &\leq 2 L_{D\Psi}^2 |\lambda_1-\lambda_2|^2|\lambda_1|^2 \|\mathcal{S}_\tau(h)\|_w^4\mathbb{E}\big[|\xi|^4\big] + 2 L_{\Psi}^2 |\lambda_1-\lambda_2|^2 \|\mathcal{S}_\tau(h)\|_w^2 \mathbb{E}\Big[|\xi|^2\Big]\\ &\leq 2 L_{D\Psi}^2 |\lambda_1-\lambda_2|^2|\lambda_1|^2 \|h\|_w^4\mathbb{E}\big[|\xi|^4\big] + 2 L_{\Psi}^2 |\lambda_1-\lambda_2|^2 \|h\|_w^2 \mathbb{E}\big[|\xi|^2\big]\\ &= 2 \|h\|_w^2 \mathbb{E}\big[|\xi|^2\big] \left(L_{D\Psi}^2 |\lambda_1|^2 \|h\|_w^2\mathbb{E}\big[|\xi|^2\big] + L_{\Psi}^2 \right)|\lambda_1-\lambda_2|^2. \end{align*} For all $y\in\mathbb{R}^+$, we then have \begin{align*} &\| \mathscr{D} \left(u(\cdot,\lambda_1)-u(\cdot,\lambda_2)\right)(y) \|_{L^2(\Omega;H_w)}^2 \\ &= \big\| \mathscr{D} \big(\left(\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h)) - \Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_2\xi h)) \right) h_x\big) (y)\big\|_{L^2(\Omega;H_w)}^2\\ &= \|h_x\|_w^2 \mathbb{E}\left[\big|\mathscr{D} \left(\Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_1\xi h)) - \Psi(X_{\tau}^{y_0,\eta}(x_0-h+\lambda_2\xi h)) \right)(y) \big|^2\right] \\ &\leq 2 \|h_x\|_w^2 \|h\|_w^2 \mathbb{E}\big[|\xi|^2\big] \left(L_{D\Psi}^2 |\lambda_1|^2 \|h\|_w^2 \mathbb{E}\big[|\xi|^2\big] + L_{\Psi}^2 \right)|\lambda_1-\lambda_2|^2. \end{align*} Hence \eqref{kolmogorov-condition} is satisfied and the result follows by Lemma \ref{technical-lemma}. \noindent\textbf{Proof of (ii):} We define \begin{equation*} u(z,\lambda) \coloneqq \Psi(X_{\tau}^{x_0,\eta}(y_0-h+\lambda\xi h)) h_x(z) \end{equation*} for $z\in\mathbb{R}^+$ and $\lambda\in\mathbb{R}$. Following the argument in the proof of (i), we need to show that \begin{equation}\label{u-lip} \| u(\cdot,\lambda_1)-u(\cdot,\lambda_2) \|_{L^2(\Omega;H_w)}^2 < C_1|\lambda_1 - \lambda_2|^2 \end{equation} and \begin{equation}\label{Du-lip} \| \mathscr{D} \left(u(\cdot,\lambda_1)-u(\cdot,\lambda_2)\right) \|_{L^2(\Omega;H_w\otimes H_w)}^2 < C_2|\lambda_1 - \lambda_2|^2 \end{equation} for some constants $C_1$ and $C_2$. Since $\Psi$ is Lipscitz with Lipschitz constant $L_{\Psi}$, we have that \begin{align*} \| u(\cdot,\lambda_1)-u(\cdot,\lambda_2) \|_{L^2(\Omega;H_w)}^2 &= \mathbb{E}\left[\|\left(\Psi(X_{\tau}^{x_0,\eta}(y_0-h+\lambda_1\xi h)) - \Psi(X_{\tau}^{x_0,\eta}(y_0-h+\lambda_2\xi h)) \right) h_x \|_w^2\right]\\ &= \|h_x\|_w^2 \mathbb{E}\left[\big|\Psi(X_{\tau}^{x_0,\eta}(y_0-h+\lambda_1\xi h)) - \Psi(X_{\tau}^{x_0,\eta}(y_0-h+\lambda_2\xi h))\big|^2\right] \\ &\leq \|h_x\|_w^2 L_{\Psi}^2 \mathbb{E}\left[\|X_{\tau}^{x_0,\eta}(y_0-h+\lambda_1\xi h) - X_{\tau}^{x_0,\eta}(y_0-h+\lambda_2\xi h)\|_w^2\right]. \end{align*} By Hölder's inequality, the Burkholder-Davis-Gundy inequality, the It\^{o} isometry, the Hille-Yosida theorem and the calculations in Step 1, it holds that \begin{align*} &\mathbb{E}\left[\|X_{\tau}^{x_0,\eta}(y_0-h+\lambda_1\xi h) - X_{\tau}^{x_0,\eta}(y_0-h+\lambda_2\xi h)\|_w^2\right]\\ &= \mathbb{E}\left[\Big\|\int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \mathcal{U}_s ((\lambda_1-\lambda_2)\xi h)\right)\,dB_s\Big\|_w^2\right]\\ &= |\lambda_1-\lambda_2|^2 \mathbb{E}\left[|\xi|^2\Big\|\int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \mathcal{U}_s h\right)\,dB_s\Big\|_w^2\right]\\ &\leq |\lambda_1-\lambda_2|^2 \mathbb{E}\left[|\xi|^4\right]^{1/2}\mathbb{E}\left[\Big\|\int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \mathcal{U}_s h\right)\,dB_s\Big\|_w^4\right]^{1/2}\\ &\leq C_{BDG} |\lambda_1-\lambda_2|^2 \mathbb{E}\left[|\xi|^4\right]^{1/2}\mathbb{E}\left[\left(\int_0^\tau \| \mathcal{S}_{\tau-s}\left(Z_s\otimes \mathcal{U}_s h\right)Q_B^{1/2}\|_{\mathcal{H}_w}^2\,ds\right)^2\right]^{1/2}\\ &\leq C_{BDG} |\lambda_1-\lambda_2|^2 \mathbb{E}\left[|\xi|^4\right]^{1/2} \mathbb{E}\left[\left(\int_0^\tau C^2{\mathrm{e}}^{2c(\tau-s)}\|(Z_s\otimes \mathcal{U}_s h) Q_B^{1/2}\|_{\mathcal{H}_w}^2\,ds\right)^2 \right]^{1/2}\\ &\leq C_{BDG} |\lambda_1-\lambda_2|^2 \mathbb{E}\left[|\xi|^4\right]^{1/2} \mathbb{E}\left[\left(\int_0^\tau C^2{\mathrm{e}}^{2c(\tau-s)}K^2{\mathrm{e}}^{2ks} \|h\|_w^2 \text{Tr}(Q_B)\,ds\right)^2 \right]^{1/2}\\ &= C_{BDG} |\lambda_1-\lambda_2|^2 \mathbb{E}\left[|\xi|^4\right]^{1/2} C^2 K^2 \|h\|_w^2 \text{Tr}(Q_B)\int_0^\tau {\mathrm{e}}^{2c(\tau-s)} e^{2ks}\,ds. \end{align*} This shows that \eqref{u-lip} holds. To show that \eqref{Du-lip} holds, we need to show that \begin{equation}\label{eq1} \mathbb{E}\left[\|X_{\tau}^{x_0,\eta}(y_0-h+\lambda_1\xi h) - X_{\tau}^{x_0,\eta}(y_0-h+\lambda_2\xi h)\|_w^2 \|\lambda_1 \xi DX_{\tau}^{x_0,\eta}(y_0-h+\lambda\xi h)(h)\|_w^2\right] \leq C|\lambda_1-\lambda_2|^2 \end{equation} and \begin{equation}\label{eq2} \mathbb{E}\left[\|\lambda_1 \xi DX_{\tau}^{x_0,\eta}(y_0-h+\lambda\xi h)(h) - \lambda_2 \xi DX_{\tau}^{x_0,\eta}(y_0-h+\lambda\xi h)(h)\|_w^2\right] \leq C|\lambda_1-\lambda_2|^2. \end{equation} We see that \begin{align*} &\mathbb{E}\left[\|\lambda_1 \xi DX_{\tau}^{x_0,\eta}(y_0-h+\lambda\xi h)(h) - \lambda_2 \xi DX_{\tau}^{x_0,\eta}(y_0-h+\lambda\xi h)(h)\|_w^2\right] \\ &= \mathbb{E}\left[\Big\|(\lambda_1-\lambda_2)\xi \int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \mathcal{U}_s h\right)\,dB_s \Big\|_w^2\right] \\ &= |\lambda_1-\lambda_2|^2\mathbb{E}\left[|\xi|^2 \Big\| \int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \mathcal{U}_s h\right)\,dB_s \Big\|_w^2\right], \end{align*} hence \eqref{eq2} follows from the calculations above. By Hölder's inequality, the Burkholder-Davis-Gundy inequality, the It\^{o} isometry and the Hille-Yosida theorem, it holds that \begin{align*} &\mathbb{E}\left[\|X_{\tau}^{x_0,\eta}(y_0-h+\lambda_1\xi h) - X_{\tau}^{x_0,\eta}(y_0-h+\lambda_2\xi h)\|_w^2 \|\lambda_1 \xi DX_{\tau}^{x_0,\eta}(y_0-h+\lambda\xi h)(h)\|_w^2\right]\\ &=\mathbb{E}\left[\Big\|\int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \mathcal{U}_s ((\lambda_1-\lambda_2)\xi h)\right)\,dB_s\Big\|_w^2 \Big\|\lambda_1 \xi \int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \mathcal{U}_s h\right)\,dB_s\Big\|_w^2\right]\\ &= \lambda_1^2 |\lambda_1-\lambda_2|^2 \mathbb{E}\left[|\xi|^4\Big\|\int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \mathcal{U}_s h\right)\,dB_s\Big\|_w^4\right]\\ &\leq \lambda_1^2 |\lambda_1-\lambda_2|^2 \mathbb{E}\left[|\xi|^8\right]^{1/2}\mathbb{E}\left[\Big\|\int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \mathcal{U}_s h\right)\,dB_s\Big\|_w^8\right]^{1/2}\\ &\leq C_{BDG} \lambda_1^2 |\lambda_1-\lambda_2|^2 \mathbb{E}\left[|\xi|^8\right]^{1/2}\mathbb{E}\left[\left(\int_0^\tau \|\mathcal{S}_{\tau-s}\left(Z_s\otimes \mathcal{U}_s h\right)Q_B^{1/2}\|_w^2\,ds\right)^4 \right]^{1/2}. \end{align*} Hence \eqref{eq1} holds from the calculations above. Equation \eqref{Du-lip} then follows from the same argument as in the proof of (i). The result then follows by Lemma \ref{technical-lemma}. \noindent\textbf{Proof of (iii):} We define \begin{equation*} u(z,\lambda) \coloneqq \Psi(X_{\tau}^{x_0,y_0}(\eta-\zeta+\lambda\xi \zeta)) h_x(z) \end{equation*} for $z\in\mathbb{R}^+$ and $\lambda\in\mathbb{R}$. Following the argument in the proof of (i), we need to show that \begin{equation}\label{u-lip2} \| u(\cdot,\lambda_1)-u(\cdot,\lambda_2) \|_{L^2(\Omega;H_w)}^2 < C_1|\lambda_1 - \lambda_2|^2 \end{equation} and \begin{equation}\label{Du-lip2} \| \mathscr{D} \left(u(\cdot,\lambda_1)-u(\cdot,\lambda_2)\right) \|_{L^2(\Omega;H_w\otimes H_w)}^2 < C_2|\lambda_1 - \lambda_2|^2 \end{equation} for some constants $C_1$ and $C_2$. Since $\Psi$ is Lipscitz with Lipschitz constant $L_{\Psi}$, we have that \begin{align*} \| u(\cdot,\lambda_1)-u(\cdot,\lambda_2) \|_{L^2(\Omega;H_w)}^2 &= \mathbb{E}\left[\|\left(\Psi(X_{\tau}^{x_0,y_0}(\eta-\zeta+\lambda_1\xi \zeta)) - \Psi(X_{\tau}^{x_0,y_0}(\eta-\zeta+\lambda_2\xi \zeta)) \right) h_x \|_w^2\right]\\ &= \|h_x\|_w^2 \mathbb{E}\left[\big|\Psi(X_{\tau}^{x_0,y_0}(\eta-\zeta+\lambda_1\xi \zeta)) - \Psi(X_{\tau}^{x_0,y_0}(\eta-h+\lambda_2\xi \zeta))\big|^2\right] \\ &\leq \|h_x\|_w^2 L_{\Psi}^2 \mathbb{E}\left[\|X_{\tau}^{x_0,y_0}(\eta-\zeta+\lambda_1\xi \zeta) - X_{\tau}^{x_0,y_0}(\eta-\zeta+\lambda_2\xi \zeta)\|_w^2\right]. \end{align*} Since $\xi$ is independent of $B$ and $W$, we have by the It\^{o} isometry that \begin{align*} &\mathbb{E}\left[\|X_{\tau}^{x_0,y_0}(\eta-\zeta+\lambda_1\xi \zeta) - X_{\tau}^{x_0,y_0}(\eta-\zeta+\lambda_2\xi \zeta)\|_w^2\right]\\ &= \mathbb{E}\left[\Big\|\int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \int_0^s \mathcal{U}_{s-u}(\lambda_1-\lambda_2)\xi\zeta\, dW_u\right)\,dB_s\Big\|_w^2\right]\\ &= |\lambda_1-\lambda_2|^2 \mathbb{E}\left[|\xi|^2\Big\|\int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right)\,dB_s\Big\|_w^2\right]\\ &= |\lambda_1-\lambda_2|^2 \mathbb{E}\left[|\xi|^2\right]\mathbb{E}\left[\Big\|\int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right)\,dB_s\Big\|_w^2\right]\\ &= |\lambda_1-\lambda_2|^2 \mathbb{E}\left[|\xi|^2\right]\mathbb{E}\left[\int_0^\tau \left\| \mathcal{S}_{\tau-s}\left(Z_s\otimes \int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right)Q_B^{1/2}\right\|_{\mathcal{H}_w}^2\,ds\right]. \end{align*} From the calculations in Step 1 it follows that \eqref{u-lip2} holds. To show that \eqref{Du-lip2} holds, we need to show that \begin{equation}\label{eq3} \mathbb{E}\left[\|X_{\tau}^{x_0,y_0}(\eta-\zeta+\lambda_1\xi \zeta) - X_{\tau}^{x_0,y_0}(\eta-\zeta+\lambda_2\xi \zeta)\|_w^2 \|\lambda_1 \xi DX_{\tau}^{x_0,y_0}(\eta-\zeta+\lambda\xi \zeta)(\zeta)\|_w^2\right] \leq C|\lambda_1-\lambda_2|^2 \end{equation} and \begin{equation}\label{eq4} \mathbb{E}\left[\|\lambda_1 \xi DX_{\tau}^{x_0,y_0}(\eta-\zeta+\lambda\xi \zeta)(\zeta) - \lambda_2 \xi DX_{\tau}^{x_0,y_0}(\eta-\zeta+\lambda\xi \zeta)(\zeta)\|_w^2\right] \leq C|\lambda_1-\lambda_2|^2. \end{equation} We see that \begin{align*} &\mathbb{E}\left[\|\lambda_1 \xi DX_{\tau}^{x_0,y_0}(\eta-\zeta+\lambda\xi \zeta)(\zeta) - \lambda_2 \xi DX_{\tau}^{x_0,y_0}(\eta-\zeta+\lambda\xi \zeta)(h\zeta)\|_w^2\right] \\ &= \mathbb{E}\left[\Big\|(\lambda_1-\lambda_2)\xi \int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right)\,dB_s \Big\|_w^2\right] \\ &= |\lambda_1-\lambda_2|^2\mathbb{E}\left[|\xi|^2 \Big\| \int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right)\,dB_s \Big\|_w^2\right], \end{align*} hence \eqref{eq4} follows from the calculations above. Since $\xi$ is independent of $B$ and $W$, it follows from the Burkholder-Davis-Gundy inequality, Jensen's inequality and Fubini's theorem that \begin{align*} &\mathbb{E}\left[\|X_{\tau}^{x_0,y_0}(\eta-\zeta+\lambda_1\xi \zeta) - X_{\tau}^{x_0,y_0}(\eta-\zeta+\lambda_2\xi \zeta)\|_w^2 \|\lambda_1 \xi DX_{\tau}^{x_0,y_0}(\eta-\zeta+\lambda\xi \zeta)(\zeta)\|_w^2\right]\\ &=\mathbb{E}\left[\Big\|\int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \int_0^s \mathcal{U}_{s-u}(\lambda_1-\lambda_2)\xi\zeta\, dW_u \right)\,dB_s\Big\|_w^2 \Big\|\lambda_1 \xi \int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right)\,dB_s\Big\|_w^2\right]\\ &= \lambda_1^2 |\lambda_1-\lambda_2|^2 \mathbb{E}\left[|\xi|^4\Big\|\int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \mathcal{U}_s h\right)\,dB_s\Big\|_w^4\right]\\ &= \lambda_1^2 |\lambda_1-\lambda_2|^2 \mathbb{E}\left[|\xi|^4\right]\mathbb{E}\left[\Big\|\int_0^\tau \mathcal{S}_{\tau-s}\left(Z_s\otimes \int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right)\,dB_s\Big\|_w^4\right]\\ &\leq C \lambda_1^2 |\lambda_1-\lambda_2|^2 \mathbb{E}\left[|\xi|^4\right]\mathbb{E}\left[\left(\int_0^\tau \left\|\mathcal{S}_{\tau-s}\left(Z_s\otimes \int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right)Q_B^{1/2}\right\|_w^2\,ds\right)^2 \right]\\ &\leq C \lambda_1^2 |\lambda_1-\lambda_2|^2 \mathbb{E}\left[|\xi|^4\right]\mathbb{E}\left[\int_0^\tau \left\|\mathcal{S}_{\tau-s}\left(Z_s\otimes \int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right)Q_B^{1/2}\right\|_w^4\,ds \right]\\ &\leq C \lambda_1^2 |\lambda_1-\lambda_2|^2 \mathbb{E}\left[|\xi|^4\right]\mathbb{E}\left[\int_0^\tau \|\mathcal{S}_{\tau-s}\|_{\text{op}}^4\left\|\left(Z_s\otimes \int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right)Q_B^{1/2}\right\|_w^4\,ds \right]\\ &= C \lambda_1^2 |\lambda_1-\lambda_2|^2 \mathbb{E}\left[|\xi|^4\right]\int_0^\tau \|\mathcal{S}_{\tau-s}\|_{\text{op}}^4 \mathbb{E}\left[\left\|\left(Z_s\otimes \int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right)Q_B^{1/2}\right\|_w^4 \right]\,ds. \end{align*} From the calculations in Step 1, we have that \begin{align*} \left\|\left(Z_s\otimes \int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right)Q_B^{1/2}\right\|_w^4 \leq \text{Tr}(Q_B)^2 \left\|\int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right\|_w^4. \end{align*} The Burkholder-Davis-Gundy inequality, Jensen's inequality and the Hille-Yosida theorem give that \begin{align*} \mathbb{E}\left[\left\|\int_0^s \mathcal{U}_{s-u}\zeta\, dW_u\right\|_w^4\right] &\leq C \mathbb{E}\left[\left(\int_0^s \|\mathcal{U}_{s-u}\zeta Q_W^{1/2}\|_{\mathcal{H}_w}^2\,ds\right)^2\right]\\ &\leq C \mathbb{E}\left[\int_0^s \|\mathcal{U}_{s-u}\zeta Q_W^{1/2}\|_{\mathcal{H}_w}^4\,ds\right]\\ &\leq C \mathbb{E}\left[\int_0^s \|\mathcal{U}_{s-u}\|_{\text{op}}^4 \|\zeta\|_{\text{op}}^4 \text{Tr}(Q_W)^2\,ds\right]\\ &\leq C \text{Tr}(Q_W)^2 \|\zeta\|_{\text{op}}^4 \mathbb{E}\left[\int_0^s K^4 e^{4k(s-u)} \,ds\right]. \end{align*} Hence \eqref{eq3} holds. Equation \eqref{Du-lip2} then follows from the same argument as in the proof of (i). The result then follows by Lemma \ref{technical-lemma}. \end{proof} In \cite[Section 3.3]{BDHP} the authors generalize the expression for the delta in their model to payoff functions which are not Fr\'{e}chet differentiable. In this case they replace the payoff function by a Moreau-Yosida approximation and take the limit to obtain the delta. It is expected that the same argument can be used to generalize our results to non-smooth payoff functions in a similar way. \section*{Acknowledgement} The research leading to this work has received support from The Research Council of Norway via the project STORM: Stochastics for Time-Space Risk Models (nr. 274410). We are grateful to Espen Sande for many useful discussions.

de3e5466e431c77b508ede4799f02e0efaacddfe

\section{Introduction} Machine understanding of natural language queries is of paramount importance to automate different workflows. The natural language query can be in the form of text or speech. Processing of query in the form of text is more popular and easy than directly processing the raw speech waveform. The text-based Natural Language Processing (NLP) include tasks like classification, token tagging, summarization, and translation. Machine translation is an NLP technique to translate a sentence from a source language to a target language. The Neural Machine Translation (NMT) is a recent approach to translation producing state of the art results \cite{bahdanau2014neural}. NMT defines translation as a sequence to sequence task and uses sequence-based neural architectures like Long Short Term Memory (LSTM) \cite{hochreiter1997long} and Transformer \cite{vaswani2017attention}. Traditional techniques like rule-based translation and Statistical Machine Translation (SMT) have been outperformed by NMT models achieving significant improvements on MT tasks. In this work, we are specifically concerned with English-Hindi neural translation. The Hindi language is one of the most popular languages in India and the fourth most spoken language in the world. Hindi is native to India and is spoken by more than 550 million total speakers worldwide. However, the number is much less as compared to global languages like English. On similar lines, the training data for the Hindi language that is publicly available for MT tasks is relatively less as compared to other highly popular languages worldwide like English, French, and German. This is important as MT tasks require a huge amount of training data to produce remarkable results using NMT models. Hindi being a relatively low resource and morphologically rich language, the amount of research in MT tasks for the Hindi language is limited \cite{philip2019baseline}. As Hindi is the most widely spoken language in the Indian subcontinent and the majority of content across the globe is published in English, the research in MT tasks for English-Hindi language pair becomes highly important. Domain adaptation of translation systems to specific domains is a common practice for low resource language pairs. The adaptation is relevant as the text in different domains can vary widely \cite{luong2015stanford}. For example, social media text and the text in literary work will be quite different from style, grammar, and abbreviations perspective. The domains can be distinguished based on topics like politics, life science, news, etc, or the style of writing like formal and informal. A translation model trained on one domain may not work well on other domains. The problem is more severe in models that use word-based representation as most of the domain-specific words will be out of vocabulary \cite{sennrich2015neural}. In this work, we explore ideas for domain adaptation for English-Hindi translation on the AdapMT Shared Task ICON 2020 data sets. The AdapMT Shared Task ICON 2020 aims to evaluate the capability of general domain machine translation for Low Resource Indian Languages. Indian languages considered in AdapMT Shared Task ICON 2020 for translation are English-Hindi, English-Telugu, and Hindi-Telugu. The shared task also focuses on Low Resource domain adaptation of machine translation systems. The adaptation is done with the use of already publicly available parallel corpora and some small in-domain parallel data for AI and Chemistry domains. The creation of a publicly available parallel corpus for low resource Indian languages is another important goal of this task. This paper describes the system built for the English-Hindi general MT and domain adaptation tasks held under AdapMT Shared Task ICON 2020. We experimented with two popular NMT models namely attention-based LSTM encoder-decoder architecture and the Transformer architecture. For domain adaption, we explore fine-tuning and mixed domain training approaches. We show that the mixed domain training performs better than the fine-tuning based approach for the datasets used in this work. \section{Architecture} In this section, we describe the two popular seq2seq neural architectures for machine translation used in this work. The encoder-decoder architecture consisting of a source side encoder and a target side decoder is used for the sequence to sequence tasks \cite{sutskever2014sequence}. The encoder encodes the text in a source language into a latent representation which is consumed by the decoder to generate the text in the target language. The decoder acts like a contextual language model generating target text by attending to the source representations. The attention mechanism is thus an integral part of encoder-decoder models which allows the decoder to focus on the right context while generating the corresponding target token. \subsection{LSTM model} The LSTM based encoder-decoder models use stacked LSTM layers on both encoder and decoder sides. The LSTM and GRU are commonly used recurrent neural network architectures for machine translation. In this work, we use LSTM based recurrent architecture as it is shown to give slightly better results \cite{britz2017massive}. The series of stacked LSTM layers encode the source text. The hidden state of the last LSTM layer is used as the encoded output. Subsequently, the target sequence is decoded sequentially using stacked LSTM layers. The decoder also makes use of an attention mechanism to attend to the encoder's hidden state. The additive attention and dot product attention are widely used attention mechanisms \cite{bahdanau2014neural, luong2015effective}. In this work, we restrict ourselves to the use of additive attention. \subsection{Transformer model} The recently introduced Transformer model has found a home in almost all NLP tasks starting with neural machine translation \cite{vaswani2017attention}. It has helped advance the state of the art in NLP and even employed for speech and vision tasks \cite{karita2019comparative, ramachandran2019stand}. The Transformer uses the self-attention mechanism as the single most important component. For the task of translation, the Transformer is used on both the encoder and decoder side. It comprises various encoders and decoders stacked over each other. The main advantage of Transformer over LSTM is the parallelism on the encoder side which helps us fully exploit the underlying hardware. The multi-headed self-attention is another architectural change that helps in providing superior results as compared to LSTM. On the encoder side, the input words are converted to vector embeddings and positional encoding is added to those embeddings so that the transformer gets the sense of the order or position of words. These embeddings are then passed on to the first encoder layer of the Transformer. The encoder consists of multi-head self-attention and a feed-forward neural network. The output from one encoder layer is given as input to the next encoder layer. The output of the final encoder layer is sent to the decoder. The decoder consists of masked-multi head attention, multi-head Attention, and a feed-forward neural network. The embeddings along with positional encodings are passed on to the first layer of the decoder. The masked multi-head Attention mechanism only pays attention to the previous words. Then, it is passed through the multi-head attention mechanism attending to the encoder state and a feed-forward neural network. The output of the decoder is passed to the linear and the softmax layer where the vector scores are turned into probabilities and the word with the highest probability is chosen as output. \section{Domain Adaptation} While generalization is always desirable the machine learning systems are often biased towards the domain of the training data. Each domain has a different distribution and different domain data are mixed while building general systems. In very basic terms, the vocabulary of the different domains is mostly different. Table \ref{present_percentage_table} shows the percentage of in-domain tokens that are present in publicly available English-Hindi parallel training corpus. Almost 20-40\% of the tokens are specific to the target domain and not present in the general corpus. Some terms are specific to and most frequently used in a particular domain. For some words, the meaning may be different across domains. For example, \textit{"As I said that here we have one hidden layer, you can have multiple hidden layers also"} is a sentence from AI domain. Whereas, \textit{"Triacylglycerol contains three fatty acids that are esterified to the glycerol backbone"} is from the Chemistry domain. These two sentences are very specific to their domain and rarely use in real-life conversations. To interpret them in the best way we need a domain expert or subject matter expert. Similarly to build a system that works best on a particular domain, we need to make use of domain-specific data. Now because we have the same underlying language rules irrespective of the domain we can make use of out of domain data to enhance our systems if in-domain data is less. This is exactly where domain adaptation comes into the picture. Domain adaptation is a form of transfer learning where we adapt a general system for a specific domain. That is we tune the model to adapt to the distribution of the target domain. It has been widely studied in the context of machine translation \cite{chu2018survey}. The adaptation techniques can either be data or model-centric. The data related approaches try to exploit the monolingual corpus of the target domain \cite{domhan2017using}. A commonly used technique is to use back-translation to expand the parallel corpus of the in-domain data \cite{sennrich2015improving}. The model-based approaches also make use of monolingual corpus from the target domain to train a language model and then do a shallow or deep fusion \cite{gulcehre2015using}. There is another set of training based technique which also go into model-centric approaches. In these approaches, the model is first trained on large out of domain parallel corpus and then re-trained or fine-tuned on the small in-domain parallel corpus. There are different variations proposed in literature where the second fine-tuning is done on a mixed parallel corpus instead of only using the in-domain corpus \cite{chu2017empirical}. The concept of domain tag was introduced in \cite{sennrich2016controlling}. The model is passed the domain label along with each training sample so that it learns to distinguish between the domains. The under-represented domains are oversampled. In this work, we evaluate the domain data fine-tuning approach and mixed-data training approach. In the first approach, we train the model on general corpus followed by in-domain corpus. In the second approach, we mixed the in-domain corpus with the general data and do a single training. Since the amount of in-domain data is very less as compared to the overall general or mixed-domain data we oversample in-domain examples while training. \section{Experimental Setup} \subsection{Dataset Details} In our English to Hindi machine translation experiments, we have used the publicly available IIT Bombay (IITB) English-Hindi Parallel Corpus \cite{kunchukuttan2017iit}. The training data in the IITB corpus consists of nearly 1.5M training samples. The IITB training data consists of sentences from the various domain. In addition to this, we have also used the AI and Chemistry in-domain parallel corpus provided by AdapMT Shared Task ICON 2020 organizers for training and testing the models for respective domains. The AI in-domain corpus contains 4872, 400, and 401 sentences in the train, validation, and test set, respectively. The Chemistry in-domain corpus contains 4984, 300, and 397 sentences in the train, validation, and test set, respectively. The data set details are described in Table \ref{statistics_table}. \begin{table} \centering \begin{tabular}{ccc} \hline \textbf{Data} & \textbf{Sentences} & \textbf{\texttildelow Tokens} \\ \hline IIT Bombay Train & 1561840 & 19.85M / 21.4M \\ General Test & 507 & 9k / - \\ AI Train & 4872 & 77k / 83k \\ AI Dev & 400 & 6k / 6k \\ AI Test & 401 & 7k / - \\ Chemistry Train & 4984 & 125k / 139k \\ Chemistry Dev & 300 & 7k / 8k \\ Chemistry Test & 397 & 7k / - \\ \hline \end{tabular} \caption{\label{font-table} Statistics of the Data (En / Hi)} \label{statistics_table} \end{table} \begin{table} \centering \begin{tabular}{cccc} \hline \textbf{Data} & \textbf{AI} & \textbf{Chemistry} & \textbf{General}\\ \hline Train (U) & 47 / 68 & 64 / 60 & - \\ Dev (U) & 58 / 80 & 78 / 76 & - \\ Test (U) & 59 / - & 55 / - & 56 / - \\ Train & 78 / 90 & 81 / 86 & - \\ Dev & 77 / 90 & 81 / 87 & - \\ Test & 78 / - & 77 / - & 76 / - \\ \hline \end{tabular} \caption{\label{font-table} Approx. \% of AdapMT domain dataset tokens (En / Hi) present in IITB Train data. Rows with a suffix 'U' indicates unique tokens, while data with no suffix indicates all tokens} \label{present_percentage_table} \end{table} \subsection{Data Processing} The individual data samples are lowercased followed by the removal of all the special characters. For training purposes, we exempted all the sentences from IITB English-Hindi Parallel Corpus with a length greater than 20 words. This was mainly done because of resource constraints to speed up training. After pre-processing, we train a sentence piece sub-word tokenizer to tokenize the English, as well as Hindi sentences \cite{kudo2018sentencepiece}. We train a unigram based tokenizer with a vocab size 32k \cite{kudo2018subword}. The source and target corpus of the IITB parallel corpus was used to train the individual sentence piece models. For experiments involving domain adaptation, the domain data from the train set was also included in the sentence piece training data. \subsection{Training Details} In this paper, we used the LSTM and Transformers based models for the English to Hindi machine translation task. For the LSTM model-based experiments, we used an attention-based encoder-decoder LSTM architecture. The encoder side of LSTM is bi-directional and the decoder side of LSTM is unidirectional with Bahdanau additive attention mechanism. The number of layers on the encoder and decoder side is set to 1 with 512 hidden units in each layer. We have used a batch size of 128 and an embedding size of 256. Adam optimizer was used as an optimizer \cite{kingma2014adam}. The subword tokenizer is used to get the subword tokens as it is known to handle the OOV problem well. For the Transformer model, the encoder and decoder have 6 layers each and the number of hidden layers in each layer is set to 512. The batch size was set to 128. The number of heads used is 8 with a word embedding size of 512. The optimizer used was Adam. The models were implemented in Tensorflow 2.0 and trained for a maximum of 10 epochs. The validation loss was used to pick the best epoch. The standard greedy decoding was used for all the experiments. For longer sentences, during decoding, a simple heuristic to split the data at comma was used followed by separate translations. While this approach may not be well suited to the translation as the alignment is not always monotonous, it worked decently well given the nature of the in-domain sentences. For our experiments with LSTM and Transformer models, we first trained the models on the IITB training corpus. The models are then retrained on in-domain AI and Chemistry parallel corpus to see the improvements in the machine translation model with the inclusion of small in-domain parallel data. In the second approach, the IITB corpus is mixed with the in-domain corpus individually and, a single training is performed. The in-domain corpus in oversampled 10 times to account for a very low in-domain corpus as compared to the general corpus. \begin{table} \centering \begin{tabular}{ccc} \hline \textbf{Model} & \textbf{AI dev} & \textbf{Che dev} \\ \hline LSTM (only IITB) & 11.54 & 8.13 \\ Transformer (only IITB) & 10.66 & 4.73 \\ LSTM (mixed) & \textbf{16.53} & \textbf{9.86} \\ Transformer (mixed) & 12.68 & 5.07 \\ LSTM (fine-tuning) & 10.62 & 5.63 \\ Transformer (fine-tuning) & 11.60 & 4.88 \\ \hline \end{tabular} \caption{\label{font-table} BLEU scores on in-domain dev data (model with the suffix 'only IITB' indicates that model is trained on samples from IITB train examples only, the model with the suffix 'mixed' indicates that the model is trained on data that is obtained by mixing oversampled in-domain training data with IITB training data, the model with suffix 'fine-tuning' indicates that the model is first trained on samples from IITB training data and then re-trained on in-domain corpus)} \label{dev_result_table} \end{table} \begin{table} \centering \begin{tabular}{cccc} \hline \textbf{Data} & \textbf{General} & \textbf{AI} & \textbf{Chemistry}\\ \hline Test Data & 14.81 & 19.08 & 13.95 \\ \hline \end{tabular} \caption{\label{font-table} BLEU scores on test data as reported by AdapMT Shared Task ICON 2020 organizers} \label{test_result_table} \end{table} \section{Results and Discussion} We evaluate the mixed data and fine-tuning approaches on LSTM and Transformer NMT models. To compare the models Bilingual Evaluation Understudy (BLEU) score is used \cite{papineni2002bleu}. We report the BLEU score on validation data of AI and Chemistry in-domain corpus. Table \ref{dev_result_table} shows the results for de-tokenized validation data. The mixed data training approach performs the best in comparison to the no-domain data and fine-tuning approach. The no-domain data approach performs better as compared to the fine-tuning approach. This indicates that the simple fine-tuning approach is not suited to the very small in-domain corpus and is susceptible to catastrophic forgetting. We see that although the Transformer based models perform well on the IITB test data they do not generalize well on the domain tasks. However, we feel that the low numbers with the Transformer can be enhanced using appropriate hyper-parameters and modifying the training approach. The system submitted for evaluation was LSTM based model trained on a mixed corpus which was giving the best validation scores. The results of the test system are shown in Table \ref{test_result_table}. The translations for the general test set were generated using the LSTM model trained only on the IITB parallel corpus. Some sample example translations using no-domain data and mixed-training approach is shown in Figure \ref{fig:sample1} and Figure \ref{fig:sample2}. \begin{figure}[b] \includegraphics[scale=0.6]{adapmt_ai_new.png} \caption{Sample sentence from AI domain} \label{fig:sample1} \end{figure} \begin{figure}[b] \includegraphics[scale=0.6]{adapmt_che_new.png} \caption{Sample sentence from Chemistry domain} \label{fig:sample2} \end{figure} \section{Conclusion} In this paper, we evaluated the effectiveness of attention-based encoder-decoder LSTM and Transformer models on a low resource English to Hindi Translation Task held under AdapMT Shared Task ICON 2020. Our experiments showed that mixed domain training works well as compared to the fine-tuning approach for domain adaptation. The addition of small in-domain parallel data can indeed improve the results on AI and Chemistry domains provided in the shared task. \section*{Acknowledgements} This work was done under the L3Cube Pune mentorship program. We would like to thank L3Cube and our mentors for the end to end guidance and encouragement to participate in the shared task. \bibliographystyle{acl_natbib}

7403aa6217702543154c915d0df5752b60379a7d

\subsection{Overview of AirSim} We conduct our simulation experiments using the high fidelity simulator, Microsoft AirSim. AirSim acts as a plugin to Unreal Engine, which is a AAA videogame engine providing access to high fidelity graphics features such as high resolution textures, realistic lighting, soft shadows etc. making it a good choice for rendering for computer vision applications. AirSim internally provides physics models for a quadrotor vehicle, which we leverage for performing autonomous drone landing. As a plugin, AirSim can be paired with any Unreal Engine environmnent to simulate autonomous vehicles that can be programmed with an API both in terms of planning/control as well as obtaining camera images. AirSim also allows for controlling environmental features such as time of day, dynamically adding/removing objects, changing object textures and so on. \subsection{3D Boosters Classification Experiment} \paragraph{Format of 3D models} To evaluate the performance of pretrained ImageNet classifiers at detecting 3D unadversarial/boosted objects (e.g. the jet shown in the main paper) among realistic settings, we set up an experiment using AirSim for image classification of common classes (warplane, car, truck, ship, etc). We pick the class of `warplane' as our object class of interest download publicly available 3D meshes for this class from \url{www.sketchfab.com}. Using the open source 3D modeling software Mitsuba, we modify the object texture to match the boosted texture for the corresponding class, and then export these meshes into the GLTF format for ingestion into Unreal Engine/AirSim. This allows us to import the boosted objects into the AirSim framework, and spawn them as objects in any of the environments being created. \paragraph{Environment screenshots and description} Within AirSim, in the interest of generating realistic imagery, we simulate a city environment (\autoref{fig:city}). For this experiment, we use the ComputerVision mode of AirSim, which does not simulate a vehicle but rather, gives the user control of a free moving camera, allowing us to generate data at ease from various locations and varying camera and world parameters. \paragraph{Sampling and evaluation} Once the 3D objects (unadversarial or normal) are present in AirSim's simulated world, the next step is to evaluate the classification of these objects from different camera angles, weather conditions etc. Given the location of a candidate object (which we randomize and average over five locations), we sample a grid ($10\times10\times10$) of camera positions in 3D around the object. For each of these positions, we move AirSim's main camera and orient it towards the object, resulting in images from various viewpoints. At runtime, each of these images are immediately processed by a pretrained ResNet-18 ImageNet classifier, which reports the top 5 class predictions. We average the accuracies across the five different locations in the scene and the $1000$ grid points around the object at each location. Along with this variation in camera angles and thereby, object pose in the frame; we also evaluate the performance of of the various 3D objects given environmental perturbations. We achieve this through the AirSim's weather conditions feature, using which we simulate weather conditions such as dust and fog dynamically with varying levels of severity of these conditions. We will open-source binaries for the AirSim code and environments that we use which will allow people to replicate our results, and investigate more scenarios of interest. \begin{figure*}[!htbp] \begin{subfigure}[t]{0.22\textwidth} \centering \includegraphics[width=\textwidth]{assets/airsim_figs/city.png} \caption{City environment in AirSim used for detection experiment} \label{fig:city} \end{subfigure} \quad \begin{subfigure}[t]{0.22\textwidth} \centering \includegraphics[width=\textwidth]{assets/airsim_figs/jet1.png} \caption{Boosted `jet' model in the City environment.} \label{fig:jet_city} \end{subfigure} \quad \begin{subfigure}[t]{0.23\textwidth} \centering \includegraphics[width=\textwidth]{assets/airsim_figs/pad1.png} \caption{Sample landing pads atop buildings in the City environment.} \label{fig:pad_city} \end{subfigure} \quad \begin{subfigure}[t]{0.23\textwidth} \centering \includegraphics[width=\textwidth]{assets/airsim_figs/eval.png} \caption{Drone in test environment used for the landing experiments.} \label{fig:pad_grass} \end{subfigure} \end{figure*} \subsection{Drone Landing Experiment} In this experiment, we evaluate how unadversarial/boosted objects can help robustify perception-action loops that are driven by vision-based pose estimation. Perception-action loops are at the heart of many robotics tasks, and accurate perception is imperative for safe, efficient navigation of robots. We choose the scenario of autonomous drone landing as our experiment, and simulate it within AirSim. For this experiment, we create assets of landing pads that are similar to helipads on top of buildings in the city environment(\autoref{fig:pad_city}). We also use a test environment with a single landing pad located on a patch of grass. An example of such a landing pad can be seen in \autoref{fig:pad_grass}. We use AirSim to simulate a quadrotor drone in these worlds, which can be programmatically controlled using a Python API. AirSim allows us to equip a downward facing, gimballed camera on this drone in order to obtain RGB images, which are then processed by our landing pad pose estimation (regression) model. Given an RGB image, the regression model outputs a 6 degree of freedom pose for the landing pad. We use/optimize only the first two enteries of this output corresponding to the relative x and y location of the landing pad w.r.t the drone. We formulate the drone landing experiment as a visual servoing task: a perception action loop that involves estimating the relative location of the pad from the image frame captured by the downward facing camera of the drone, and sending an appropriate velocity command in order to align the camera center with that of the pad. We achieve these through the following steps: \paragraph{Data Collection.} We use AirSim's inbuilt data collection API for this step. Given the location of the pad in the world, we sample various feasible locations for the drone in an imaginary cone whose vertex aligns with the center of the landing pad. We then spawn the drone in these randomly sampled positions, and obtain the RGB and segmentation views of the pad as generated by AirSim, along with the relative ground truth position of the landing pad w.r.t the drone, and repeat this process to create a dataset. The collected dataset contains 20000 images and is split 80-20 between train and evaluation sets. \paragraph{Landing pad pose estimator.} We train a model that maps top view images of a scene with a landing pad, to the relative 2D location of the landing pad w.r.t the drone in the camera frame. We use a ResNet-18 pretrained on ImageNet as the backbone for the pose regressor, and we replace the last classification layer with a regression layer outputting the $(x,y)$ relative location of the pad w.r.t drone. The model is trained end-to-end by minimizing the mean squared error (MSE) loss between the predicted location and the ground truth location. The ground truth is collected along with the images using the AirSim City simulation environment as describe before. We train the model for 10 epochs using SGD with a fixed learning rate of 0.001, a batch size of 512, a weight decay of 1e-4, and with MSE as the objective function. The model converges fairly quickly (within the first few epochs). \paragraph{Drone Landing.} To use the pose estimator's predictions and send appropriate actions, we utilize the Multirotor API of AirSim. This allows us to control the drone by setting the desired velocity commands along all the axes (translational/rotational). Given the position of the landing pad in the scene relative to that of the drone( as output by the pose regressor) we execute the landing operation by sending appropriate velocity commands to the drone. To generate the right velocity commands, given the relative position of the landing pad, we use a standard PID controller that computes corrective velocity values until the position of the drone matches that of the landing pad. For a pose output by the regressor treated as the setpoint $P_{set}$, and current drone pose $P_{curr}$ and at any point at time $t$, the appropriate velocity command $v(t)$ can simply be computed as follows: \begin{align} e(t) &= P_{set} - P_{curr} \nonumber \\\nonumber v(t) &= K_p e(t) + K_d \frac{d}{dt}e(t) + K_i * \int_{0}^{t} e(t)dt \end{align} where $K_p, K_d, and K_i$ are the hyperparameters of the PID controller and are manually tuned. We find that $K_p = K_d = 5$ and $K_i = 0$ to be reasonable for our task. For realistic perturbations to the scene, similar to the 3D boosters classification experiment, we continue making use of the weather API to generate weather conditions in AirSim. This results in a variation of factors such as amount of dust or fog in the scene, allowing us to evaluate the performance of landing under various realistic conditions. \section{Pseudocode for Unadversarial Example Generation} \label{app:algorithms} \input{appendix/algorithms} \clearpage \section{3D Simulation Details} \label{app:airsim} \input{appendix/airsim} \clearpage \section{Experimental Setup} \label{app:exp_setup} \input{appendix/exp_details} \clearpage \section{Omitted Results} \label{app:detailed_results} \input{sections/detailed_results} \subsection{Pretrained vision models we evaluate} \label{app:pretrained-models} Here we present details of the different vision models we use in our paper. For more details on all of these, please check the README of our code at \xspace{\url{https://git.io/unadversarial}}. \paragraph{Corruption benchmark experiments:} We use pretrained ResNet-18 and ResNet-50 (both standard and $\ell_2$-robust with $\varepsilon = 3$) architectures from \cite{salman2020adversarially}: \url{https://github.com/microsoft/robust-models-transfer}. \paragraph{3D object classification in AirSim:} We use an ImageNet pretrained ResNet-18 architecture from the PyTorch's Torchvision\footnote{These models can be found here \url{https://pytorch.org/docs/stable/torchvision/models.html}} to classify all the boosted and non-boosted versions of the jets, cars, ships etc in AirSim. \paragraph{Drone landing experiment in AirSim:} We finetune an ImageNet pretrained ResNet-18 model on the regression task of drone landing. The last layer of the pretrained model is replaced with a 2D linear layer estimating the relative pad location w.r.t the drone. We collect a 20k sample dataset for training the pad pose estimation in AirSim with an $80-20$ train-val spilt. We use a learning rate of $0.001$, a batch size of $512$, a weight decay of $1e-4$. We train for $10$ epochs. \paragraph{Physical world unadversarial examples experiment:} Similar to the 3D object classification experiment in AirSim, we use an ImageNet pretrained ResNet-18 architecture from Torchvision. \subsection{Unadversarial patch/texture training details} \paragraph{Patches training details} We fix the training procedure for all of the 2D patches we optimize in our paper. We train all the patches starting from random initialization with batch size of 512, momentum of $0.9$, and weight decay of $1e-4$. We train all the patches for 30 epochs (which is more than enough as we observe that for both ImageNet and CIFAR-10, the patch converges within the first 10 epochs) with a learning rate of $0.1$ {We sweep over three learning rates $\in \{0.1, 0.01, 0.001\}$ but we find that all of these obtain very similar results. So we stick with a learning rate of $0.1$ for all of our experiments.}. For the classification tasks (i.e., everything but drone landing) we use the standard cross-entropy loss. For the drone landing task (landing pad pose estimation), we use the standard mean squared error loss. \paragraph{Texture training details} We now outline the process for constructing adversarial textures. We implemented a custom PyTorch module with a distinct forward and backward pass; on the forward pass (i.e., during evaluation), the module takes as input an ImageNet image, and a 200px by 200px texture; using the Python bindings for Mitsuba~\citep{mitsuba} 3D renderer, the module returns a rendering of the desired 3D object, overlaid onto the given ImageNet image. On the backwards pass (i.e., when computing gradients), we use the 3D model's UV map\footnote{Mitsuba provides direct access to the UV map through the \texttt{aov} integrator; see the code release for more details.}---a linear transformation from $(x, y)$ locations on the texture to $(x, y)$ locations in the rendered image---to approximate gradients through the rendering process. This is the same procedure used by~\citep{athalye2018synthesizing} for constructing physical adversarial examples. Note that this is a simple approximation that only accounts for the location of pixels in the rendered image (i.e., ignores the effects of lighting, warping, etc.). However, \subsection{Details of the physical world experiment} To conduct the physical-world experiments, we used a toy racecar\footnote{https://www.amazon.com/gp/product/B07T5X69TZ/}, a toy warplane\footnote{\url{https://www.amazon.com/CORPER-TOYS-Pull-Back-Aircraft-Birthday/dp/B07DB3839X/}} (both from \url{amazon.com}) as well as two household objects: a coffeepot and eggnog container. We then printed the unadversarial patches corresponding to classes ``racer,'' ``warplane,'' ``coffeepot,'' and ``eggnog'' on an HP DeskJet 2700 InkJet printer, at 250\% scale. We adhere the patches to the top of their respective objects with clear tape (the results are shown in Figure~\ref{fig:physical_world}). We choose 18 distinct poses (camera positions), and for each pose took one picture of the object with the patch attached, and one picture without (keeping the location of the patch constant throughout the experiment). Example photographs are shown in Figure \ref{fig:more_physical_examples}. We evaluated a pre-trained ResNet-18 classifier on the resulting images. \begin{figure} \centering \includegraphics[width=\textwidth]{assets/main_figs/more_3d_objects.png} \caption{Photographs in different poses of the four physical objects we experimented on, with and without an unadversarial patch.} \label{fig:more_physical_examples} \end{figure} \subsection{Replicate our results} We desired simplicity and kept reproducibility in our minds when conducting our experiments, so we use standard hyperparameters and minimize the number of tricks needed to replicate our results. Our code is available at \xspace{\url{https://git.io/unadversarial}}. \section{Introduction} \label{sec:intro} \input{sections/intro} \section{Motivation and Approach} \label{sec:motivation} \input{sections/motivation} \subsection{Constructing unadversarial objects} \label{sec:algorithms} \input{sections/methods} \section{Experimental Evaluation} \label{sec:experiments} \input{sections/experiments} \section{Related Work} \label{sec:related} \input{sections/related} \section{Discussion and Conclusions} \label{sec:conclusions} \input{sections/conclusions} \section*{Acknowledgements} We are grateful to Ian Engstrom for helping take the photographs for the physical-world experiments. Work supported in part by the NSF grants CCF-1553428 and CNS-1815221, and the Microsoft Corporation. This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Contract No. HR001120C0015. Research was sponsored by the United States Air Force Research Laboratory and was accomplished under Cooperative Agreement Number FA8750-19-2-1000. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the United States Air Force or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein. \clearpage \printbibliography \clearpage \input{appendix/appendix} \clearpage \end{document} \subsubsection{QR-Code} We compare our unadversarial patches to the well-known QR-Code patches. We create a QR-Code for each class of the ImageNet dataset using Python's \texttt{qrcode} package(we avoid using CIFAR-10 since the images are too small for QR-Codes to be visible and detected at all). We overlay the QR-Codes over the ImageNet validation set according in accordance to what label each image has. We add the various ImageNet-C corruption on top of the resulting images, then we use python's \texttt{Pyzbar}\footnote{We experiment with \texttt{OpenCV for detecting the QR-Codes but find that \texttt{Pyzbar} leads to better performance}.} package to detect the QR-Codes. The results are shown in \autoref{fig:qrcode-results}. The performance of QR-Codes is not comparable to what we obtain with unadversarial patches (see \autoref{fig:2d_results}). \begin{figure}[!htbp] \centering \includegraphics[width=.8\linewidth]{assets/result_figs/imagenet_booster_vs_qrcode.pdf} \\ \includegraphics[width=.8\linewidth]{assets/result_figs/example_transform_true_translate_true.PNG} \caption{QR-Code boosted ImageNet results under various corruptions.} \label{fig:qrcode-results} \end{figure} \clearpage \subsubsection{Best training image per class as patch} Another natural baseline that we compare with is using the best images per class in the training set of the task of interest as patches for boosting the performance of pretrained models. For example, for ImageNet classification, we simply evaluate the loss of each training image using a pretrained ImageNet model (ResNet-18 in our case), and we the image with the lowest loss per class as the patch for that class. Now we overlay these found patches onto the ImageNet validation set with random scaling, rotation, and translation (as shown in \autoref{fig:bestimages-imagenet}), we add ImageNet-C corruptions, and we evaluate this new dataset using the same pretrained model we used to extract the patches. The results for ImageNet and CIFAR-10 are shown in \autoref{fig:bestimages-imagenet} and \autoref{fig:bestimages-cifar}, respectively. \begin{figure}[!htbp] \centering \includegraphics[width=.45\linewidth]{assets/result_figs/imagenet_booster_vs_bestloss.pdf} \\ \includegraphics[width=.4\linewidth]{assets/result_figs/example_imagenet_transform_true_translate_true.PNG} \caption{Best training image with translation, rotation, and scaling for ImageNet.} \label{fig:bestimages-imagenet} \end{figure} \begin{figure}[!htbp] \centering \includegraphics[width=.45\linewidth]{assets/result_figs/cifar_booster_vs_bestloss.pdf} \\ \includegraphics[width=.4\linewidth]{assets/result_figs/example_cifar10.PNG} \caption{Best training image with translation, rotation, and scaling for CIFAR-10.} \label{fig:bestimages-cifar} \end{figure} \clearpage \subsubsection{Best training image vs random training image as patch} Here we investigate whether using a random image from the training set does any better than using the best-loss image as a patch. The results are shown in the below Figures. As one would expect, using a random image from the training set leads to strictly worse performance. This holds for both ImageNet and CIFAR-10 as shown in \autoref{fig:imagenet-rand-image-res} and \autoref{fig:cifar-rand-image-res}. \begin{figure}[!htbp] \centering \includegraphics[width=.55\linewidth]{assets/result_figs/bestimagebaseline/imagenet_tranform_1_translate_True_best_vs_random.pdf} \caption{Best training image vs random training image with translation, rotation, and scaling.} \label{fig:imagenet-rand-image-res} \end{figure} \begin{figure}[!htbp] \centering \includegraphics[width=.55\linewidth]{assets/result_figs/bestimagebaseline/cifar_tranform_1_translate_True_best_vs_random.pdf} \caption{Best training image vs random training image with translation, rotation, and scaling.} \label{fig:cifar-rand-image-res} \end{figure} \clearpage \subsubsection{Predefined fixed-pattern unadversarial patches} This baselines is slightly different than the previous baselines since it allows the underlying classification model to be changed. Basically, we fix the a set of patches to predefined pattern (here a fixed random gaussian noise for each class), and we train a classifier on an undversarial/boosted dataset with these patches. The resulting models are consistently weaker on all corruptions of ImageNet-C and CFAR-10-C as shown in \autoref{fig:imagenet-predefined-patch} and \autoref{fig:cifar-predefined-patch} compared to our trained patches the main paper in \autoref{fig:2d_results}. \begin{figure}[!htbp] \centering \includegraphics[width=.6\linewidth]{assets/result_figs/imagenet_booster_vs_model.pdf} \caption{Robustness of an ImageNet ResNet-18 model trained on a predefined patch.} \label{fig:imagenet-predefined-patch} \end{figure} \begin{figure}[!htbp] \centering \includegraphics[width=.6\linewidth]{assets/result_figs/cifar_booster_vs_model.pdf} \caption{Robustness of a CIFAR-10 ResNet-50 model trained on a predefined patch.} \label{fig:cifar-predefined-patch} \end{figure} \subsection{Corruption benchmark main results: additional results to \autoref{fig:2d_results}} Here we show the detailed main results for boosting ImageNet and CIFAR-10 with unadverasarial patches. \begin{figure}[!htbp] \centering \includegraphics[width=.415\linewidth]{assets/result_figs/booster_mode_imagenet.pdf} \includegraphics[width=.49\linewidth]{assets/result_figs/booster_mode_imagenet_per_severity.pdf} \caption{Robustness of a trained 2D booster over pretrained ImageNet ResNet-18 model.} \label{fig:imagenet-main-results-app} \end{figure} \begin{figure}[!htbp] \centering \includegraphics[width=.415\linewidth]{assets/result_figs/booster_mode_cifar.pdf} \includegraphics[width=.49\linewidth]{assets/result_figs/booster_mode_cifar_per_severity.pdf} \caption{Robustness of a trained 2D booster over pretrained CIFAR-10 ResNet-50 model.} \label{fig:cifar10-main-results} \end{figure} \clearpage \subsection{Baselines} \input{sections/baselines} \subsection{Clean data and synthetic corruptions} \label{sec:eval_benchmark} We first test whether unadversarial examples improve the performance of image classifiers on benchmark datasets. Using the algorithm described in Section \ref{sec:algorithms}, we construct unadversarial patches of varying size for pre-trained ResNet-50 classifiers on the CIFAR \citep{krizhevsky2009learning} and ImageNet \citep{russakovsky2015imagenet} datasets. For evaluation, we add these patches at random positions, scales, and orientations to validation set images (see Appendix~\ref{app:exp_setup} for the exact protocol). As shown in Figure \ref{fig:main_imagenet_clean}, the pre-trained ImageNet classifier is more consistently more accurate on the augmented ImageNet images. For example, an unadversarial patch 20 times smaller than ImageNet images boosts accuracy by $26.3\%$ (analogous results for CIFAR are given in Appendix \ref{app:detailed_results}). \begin{figure}[h] \centering \includegraphics[width=.45\textwidth]{assets/main_figs/boosted_500.jpg} \quad\quad\quad \includegraphics[width=.45\textwidth]{assets/main_figs/adv_500.jpg} \caption{Clean (left) and corresponding corrupted (right) ImageNet images augmented with an unadversarial patch---we use such images to evaluate the efficacy of unadversarial patches in Section \ref{sec:eval_benchmark}.} \begin{subfigure}{0.29\textwidth} \includegraphics[width=\linewidth]{assets/result_figs/main_imagenet_clean.pdf} \caption{Performance on ImageNet} \label{fig:main_imagenet_clean} \end{subfigure} \begin{subfigure}{0.7\textwidth} \centering \includegraphics[width=\linewidth]{assets/result_figs/main_imagenet.pdf} \caption{Performance on synthetically corrupted data (ImageNet-C)} \label{fig:2d_results} \end{subfigure} \caption{Accuracy on (a) clean ImageNet images and (b) synthetically corrupted ImageNet-C images as a function of patch size (given as a percentage of image area). In (b), each bar denotes the average accuracy over the five severities in ImageNet-C, and the horizontal dashed lines report the accuracy on the original (non-patched) datasets. Unadversarial patches consistently boost performance for both clean and corrupted images, with accuracy monotonically increasing with patch size. The patches were trained without any corruptions or non-standard data augmentation in-the-loop (we train with the same augmentations that the pre-trained model itself was trained with).} \end{figure} \paragraph{Robustness to synthetic corruptions.} Next, we use the CIFAR-C and ImageNet-C datasets \citep{hendrycks2019benchmarking} (consisting of the CIFAR and ImageNet validation sets corrupted in 15 systematic ways) to see whether the addition of unadversarial patches to images confers any robustness to image corruptions. We use the same patches and evaluation protocol that we used when looking at clean data (to ensure a fair evaluation, we apply corruptions to boosted images only {\em after} the unadversarial patches have been applied). As a consequence, at test time neither model nor patch has been exposed to any image corruptions beyond standard data augmentation. As a result, this experiment tests the ability for unadversarially boosted images to withstand completely unforeseen corruptions; we also avoid any potential biases from training on (and thus ``overfitting'' to \cite{kang2019testing}) a specific type of corruption. The results (cf. Figure~\ref{fig:2d_results} for ImageNet and Appendix~\ref{app:detailed_results} for CIFAR) indicate that unadversarial patches do improve performance across corruption types; for example, applying an unadversarial patch 5\% the size of a standard ImageNet image boosts accuracy by an average of $31.7\%$ points across corruptions \footnote{Since the original corruption benchmarks proposed by \citep{hendrycks2019benchmarking} are only available as pre-computed JPEGs (for which we cannot apply a patch pre-corruption) or CPU-based Python image operations (which were prohibitively slow), we re-implemented all 15 corruptions as batched GPU operations; we verified that model accuracies on our corruptions mirrored the original CPU counterparts (i.e., within $1\%$ accuracy). For more details about our reimplementation, see our code release.}. \paragraph{Baselines.} We also compare our results to a variety of natural baselines; the most relevant of these is the ``best loss image patch,'' where we use the minimum-loss training image in place of a patch. We compare with this baseline to ensure that our method is doing something beyond this naive way to add signal to an image. The results are shown in Appendix~\ref{app:detailed_results}, along with comparisons to less sophisticated baselines, such as QR Codes and predefined random Gaussian noise patches. \begin{figure*}[t] \centering \includegraphics[width=.9\linewidth]{assets/result_figs/jetplots.pdf} \caption{The jet unadversarial example task. We show example conditions under which we evaluate the objects, along with aggregate statistics for how well an ImageNet classifier classifies the objects in different conditions. We find that the classifiers perform consistently better on the unadversarial jet texture over the standard jet texture in both standard and distributionally shifted conditions. We also give a baseline of a white jet with a lighter texture because of the poorly visibility inherent in the simulator; we find it performed worse than even the standard jet.} \label{fig:jetfighter} \end{figure*} \begin{figure*}[t] \centering \includegraphics[width=1\linewidth]{assets/main_figs/extra_3D_boosters.pdf} \caption{Additional examples reporting aggregate statistics for how well an ImageNet classifier classifies various objects in different conditions. Again, we find that the classifiers perform consistently better on the unadversarial objects texture over the standard objects.} \label{fig:other-3D-boosters} \end{figure*} \subsection{Classification in 3D simulation} We now test unadversarial examples in a more practical setting: recognizing three-dimensional objects in a high-fidelity similar. We collect a set of three-dimensional meshes corresponding to four ImageNet classes: ``warplane'', ``minibus'', ``container ship'', and ``trailer truck'' , sourced from \url{sketchfab.com}. We generate a texture for each object using the unadversarial texture algorithm outlined in Section~\ref{sec:algorithms}, using the ImageNet validation set as the set of backgrounds for the algorithm, and a pre-trained ResNet-50 as the classifier. To evaluate the resulting textures, we import each mesh into Microsoft AirSim, a high-fidelity three-dimensional simulator; we then test pre-trained ImageNet models' ability to recognize each object with and without the unadversarial texture applied in a variety of surroundings. We also test each texture's robustness to more realistic weather corruptions (snow and fog) built directly into the simulator (rather than applied as a post-processing step). We provide further detail on AirSim and our usage of it in Appendix~\ref{app:airsim}. Examples of the images used to evaluate the unadversarial textures, as well as our main results for one of the meshes are shown in Figure~\ref{fig:jetfighter}. We find that in both standard and simulated adverse weather conditions, the model consistently performs better on the mesh wrapped in the unadversarial texture than on the original. We present similar results for the other three meshes in Figure~\ref{fig:other-3D-boosters}. \subsection{Localization for (Simulated) Drone Landing} We then assess whether unadversarial examples can help outside of the classification setting. Again using AirSim, we set up a drone landing task with a perception module that receives as input an axis-aligned aerial image of a landing pad, and is tasked with outputing an estimate of the camera's $(x,y)$-position relative to the pad. While this task is quite basic, we are particularly interested in studying performance in the presence of heavy (simulated) weather-based corruptions. The drone is equipped with a pretrained regression model that localizes the landing pad (described in detail in Appendix~\ref{app:airsim}). We optimize an unadversarial texture for the surface of the landing pad to best help the drone's regression model in localization. Figure~\ref{fig:drone} shows an example of the landing pad localization task, along with the performance of the unadversarial landing pad compared to the standard pad. The drone landing on the unadversarial pad consistently lands both more reliably. \begin{figure*} \centering \includegraphics[width=.9\linewidth]{assets/result_figs/landing.pdf} \label{fig:droneeval} \caption{Drone landing task. On the left we show the unadversarial versus standard landing pads. On the right we show the results for the task when both the standard and unadversarial landing pads are used. We find that the drone consistently takes less time to land, and has a higher chance of landing correctly, when detecting the unadversarial landing pad.} \label{fig:drone} \end{figure*} \subsection{Physical World Unadversarial Examples} Finally, we move out of simulation and test whether the unadversarial patches that we generate can survive naturally-arising distribution shift from effects such as real lighting, camera artifacts, and printing imperfections. We use four household objects (a toy racecar, miniature plane, coffeepot, and eggnog container), and print out (on a standard InkJet printer) the adversarial patch corresponding to the label of each object. We take pictures of the toy with and without the patch taped on using an ordinary cellphone camera, and count the number of poses for which the toy is correctly classified by a pre-trained ImageNet classifier. Our results are in Table~\ref{tab:physical_world}, and examples of patches are in Figure~\ref{fig:physical_world}. Classifying both patched and unpatched images over a diverse set of poses, we find that the adversarial patches consistently improve performance even at uncommon object orientations. \begin{figure} \begin{subtable}{0.33\linewidth} {\begin{tabular}{@{}lcccc@{}} \toprule Class & No Patch & Patch \\ \midrule ``racer'' & 22\% & 83\% \\ ``eggnog'' & 22\% & 44\% \\ ``coffee pot'' & 39\% & 56\% \\ ``warplane'' & 67\% & 83\% \\ \bottomrule \end{tabular}} \caption{Accuracy of pre-trained ResNet-18 on photographs of real world objects with and without patches.} \label{tab:physical_world} \end{subtable} \hspace{1em} \begin{subfigure}{0.6\linewidth} \centering \includegraphics[width=\linewidth]{assets/result_figs/physical_world_results.png} \caption{Examples photos of the ``warplane'' and ``racer'' physical objects taken with (top) and without (bottom) an unadversarial patch.} \label{fig:physical_world} \end{subfigure} \caption{Physical-world experiments. We take pictures of objects at diverse orientations while varying the presence of a patch on the object. Note that we don't do any additional data augmentation on the patches, which are the same used in our previous ImageNet benchmark experiment.} \end{figure} \subsection{Leveraging more controlled vision settings} \label{subsec:controlled_settings_motivation} Consider the vision tasks of detecting a landing pad from a drone, or classifying manufacturing components from a factory robot. In both these tasks, reliable in-distribution performance is a necessity; still, a number of possible distribution shifts may occur at deployment time. For example, the drone might approach the landing pad at an atypical angle, or have a view obstructed by snow, smoke, or rain. Similarly, the factory robot may encounter objects in unfamiliar poses, or using a low-quality/noisy camera. At first glance, dealing with these issues seems to require tackling the difficult problem of general distribution shift robustness discussed earlier in this section. However, there is in fact a critical distinction between the scenarios considered above and vision tasks in their full generality. In particular, in these scenarios and many others, the system designer has control over not only the model that is used but also the physical objects that the model operates on. As we will demonstrate, the designer can use this capability to modify these objects to majorly boost the model's ability to solve the problem at hand. For instance, the designer of the drone's landing algorithm could, in addition to training a detection model, also paint the landing pad bright yellow. A machine learning model trained to detect this custom landing pad might then be more effective than a model trained to detect a standard grey pad, especially in low-visibility conditions. Still, the particular choice to paint the landing pad yellow is rather ad hoc, and likely rooted in the way {\em humans} recognize objects. Meanwhile, an abundance of prior work (e.g., \citep{jacobsen2019excessive, geirhos2018imagenettrained, jetley2018friends, ilyas2019adversarial}) demonstrates that humans and machine learning models tend to use different sets of features to make their decisions. This suggests that rather than relying on human priors, we should instead be asking: {\em how can we build objects that are easily detectable by machine learning models?} \subsection{Unadversarial examples} The task of making inputs {\em less} recognizable by computer vision systems has been a focus of research in {\em adversarial examples}. Adversarial examples are small, carefully constructed perturbations to natural images that can induce arbitrary (mis)behavior from machine learning models~\citep{biggio2013evasion,szegedy2014intriguing}. These perturbations are typically constructed as the result of an optimization problem that maximizes the loss of a machine learning model with respect to the input, i.e., by solving the optimization problem \begin{equation} \label{eq:adv_example} \delta_{adv} = \arg\max_{\delta \in \Delta} L(f_{\theta}(x + \delta), y), \end{equation} where $f_{\theta}$ is a parameterized model (e.g., a neural network with weights $\theta$); $x$ is a natural input; $y$ is the corresponding correct label; $L$ is the loss function used to train $\theta$ (e.g., cross-entropy loss) and $\Delta$ is a class of permissible perturbations (e.g., norm-bounded perturbations: $\Delta = \{\delta: \|\delta\|_p \leq \epsilon\}$ for some small $\epsilon > 0$). Adversarial perturbations are typically crafted via projected gradient descent (PGD) \citep{nesterov2003introductory} in input space, a standard iterative first-order optimization method---prior work in adversarial examples has shown that even a few iterations of PGD suffice to completely change the prediction of many state-of-the-art machine learning systems \citep{madry2018towards}. \paragraph{From adversarial examples to unadversarial objects.} The goal of this work is to modify the design of objects so that they are more easily recognizable by computer vision systems. If we could specify every pixel of every image that a model encounters at test time, we could draw on the effectiveness of adversarial examples, and construct image perturbations (using PGD) that {\em minimize} the loss of the system, e.g., \begin{equation} \label{eq:unadv_example} \delta_{unadv} = \arg\min_{\delta \in \Delta} L(\theta; x + \delta, y). \end{equation} In our setting of interest, however, having such fine-grained access to the test inputs is unrealistic (presumably, if we had precise control over every pixel in the input, we could just directly encode the ground-truth label directly in it). Instead, we have {\em limited} control over some physical objects; these objects are in turn captured within image inputs, along with many signals that are out of our control, such as camera artifacts, weather effects, or background scenery. It turns out that we can still draw on techniques from adversarial examples research in this limited-control setting. Specifically, a recent line of work \citep{kurakin2017adversarial, sharif2016accessorize, evtimov2018robust, athalye2018synthesizing} concerns itself with constructing {\em robust adversarial examples}~\citep{athalye2018synthesizing}, i.e., physically realizable objects that act as adversarial examples when introduced into a scene in any one of a variety of ways. For example, \citet{sharif2016accessorize} design glasses frames that cause facial recognition models to misclassify faces, \citet{athalye2018synthesizing} design custom-textured 3D models that are misclassified by state-of-the-art ImageNet classifiers from many angles and viewpoints, and~\citep{brown2018adversarial} design adversarial patches: stickers that can be placed anywhere on objects causing them to be misclassified. In this paper, we leverage the techniques developed in the above line of work to constructi robust un-adversarial objects---physically realizable {\em objects} optimized to minimize (rather than maximize) the loss of a target classifier. In the next section, we will more concretely discuss our methods for generating unadversarial objects, then outline our evaluation setup.

9e5532a761ca023cd50531e122a9c84023af31c5

\section{Background}\label{sec:background} \textbf{Trusted Execution Environments (TEEs)}. A trusted execution environment is a hardware-protected part of the processor. Depending on the specific version and implementation, a TEE can guarantee confidentiality, integrity and protection against several types of attacks for code and data executed and processed within it. Currently, there exist several hardware-based technologies that enable physical isolation of different execution environments available in a wide range of CPUs, including \textsc{Arm}\xspace \textsc{TrustZone}\xspace~\cite{trustzone,pinto2019demystifying,amacher2019performance}, Intel\textregistered SGX\xspace~\cite{costan2016intel}, AMD SEV~\cite{kaplan2016amd}, and RISC-V Keystone~\cite{lee2020keystone}. While we expect more TEE options to surface in the coming years, we focus on \textsc{Arm}\xspace\textsc{TrustZone}\xspace in the reminder of the paper, highlighting its main features and programming framework. \textbf{\textsc{TrustZone}\xspace}. \textsc{TrustZone}\xspace is a hardware feature implemented in \textsc{Arm}\xspace processors since 2004~\cite{alves2004tz}. It enables physical separation between two different execution environments: the trusted side (known as the TEE or \emph{secure world}) and the untrusted side (known as the REE or \emph{normal world}). The \textsc{TrustZone}\xspace protects the integrity and confidentiality of the code run inside the \emph{secure world} from an attacker with physical access to the device, a malicious kernel or a high-privileged software. Programs hosted inside the \textsc{TrustZone}\xspace, known as Trusted Applications, can leverage additional \textsc{TrustZone}\xspace functionalities such as secure persistent storage with the use of APIs. \begin{figure}[!t] \centering \includegraphics[scale=0.7]{figures/OPTEE-v3.pdf} \caption{Architecture of \textsc{Op-Tee}\xspace to realize trusted applications leveraging \textsc{TrustZone}\xspace. GP: GlobalPlatform~\cite{globalplatform}, a standardization effort for TEEs.} \label{fig:optee} \end{figure} \textbf{\textsc{Op-Tee}\xspace}. The Open Portable Trusted Execution Environment (\textsc{Op-Tee}\xspace) is an open source operating system with native support for the \textsc{TrustZone}\xspace. \textsc{Op-Tee}\xspace implements two APIs compliant with the GlobalPlatform~\cite{globalplatform} specifications: the TEE Internal Core API~\cite{teeinternalapi}, which is exposed to the Trusted Applications, and the TEE Client API~\cite{teeclientapi}, which defines how a client in the REE should communicate with the TEE. The TEE can run alongside a Linux-based operating system (such as a GNU/Linux distribution or AOSP) as the untrusted OS. \textbf{Trusted Application}. Trusted Applications (TAs) run inside the \emph{secure world}, making use of the TEE kernel to access system resources. TAs can act as a service for applications running on the \emph{normal world} as well as for other TAs. When using \textsc{Op-Tee}\xspace, Trusted Applications are implemented in C and they can leverage the TEE Internal Core API implemented by \textsc{Op-Tee}\xspace, which offers several services, including trusted storage and cryptographic, time and arithmetical operations. \textsc{KeVlar-Tz}\xspace is an application that runs on the TEE, so it is a Trusted Application. When using \textsc{KeVlar-Tz}\xspace, we can do so from another TA (if we are running it on the TEE) or from a normal application running in the REE. The design of trusted applications for \textsc{Op-Tee}\xspace is depicted in Figure~\ref{fig:optee}. \textbf{Trusted Persistent Storage}. \textsc{Op-Tee}\xspace provides the Trusted Storage API for Data and Keys as part of the TEE Internal Core API~\cite{teeinternalapi}. This API can be used by Trusted Applications to access a secure storage which is only accessible to that particular TA and that is persistent between reboots. The data is stored encrypted and signed on the disk, to prevent it from being accessed or tampered with by any other application. The data can later be transparently accessed in cleartext by the TA. \textsc{KeVlar-Tz}\xspace exploits this by saving the encryption keys using a public ID, which are the value and key (respectively) in the key-value storage. When an untrusted application needs to use an encryption key, it sends the key's ID and \textsc{KeVlar-Tz}\xspace retrieves it. \textbf{\textsc{Mqtt}\xspace \& \texttt{mosquitto}\xspace}. The Message Queuing Telemetry Transport (\textsc{Mqtt}\xspace) is a lightweight, publish-subscribe network protocol, suited for communication in environments with few resources and low network bandwidth. MQTT has two types of entities: the broker and the clients. The clients can publish messages to a topic or subscribe to one of them, while the broker is a server that forwards each incoming message to all the subscribers of its topic. \texttt{mosquitto}\xspace is an open source implementation of the \textsc{Mqtt}\xspace broker developed and maintained by the Eclipse Foundation, which also provides a C library for implementing MQTT clients, as well as one implementation of both a subscriber client and a publisher client. \textbf{\textsc{Mqt-Tz}\xspace}. \textsc{Mqt-Tz}\xspace \cite{segarra2020mqttz-srds} is a fork of \texttt{mosquitto}\xspace, a topic-based publish-subscribe framework for IoT. It allows brokers and the clients to leverage the \textsc{TrustZone}\xspace TEE, by encrypting the messages sent on the network to prevent the broker from being able to read them, while maintaining the publish-subscribe pattern. Similarly it allows full decoupling between publishers and subscribers, shielding the subscriptions inside the TEE. \section{Evaluation} \label{sec:evaluation} This section presents our experimental evaluation of \textsc{KeVlar-Tz}\xspace using both micro- and macro-benchmarks. Our goal is to define the overheads of running \textsc{KeVlar-Tz}\xspace to further assess whether the trade-offs to have a secure storage system are reasonable for a real-world scenario. \textbf{Evaluation settings}. We deploy \textsc{KeVlar-Tz}\xspace on a Raspberry Pi 3 Model B+ as well as on an emulated environment using QEMU version 8~\footnote{\url{https://www.qemu.org}} to test the application. QEMU is a tool that has been proven useful, despite its limitations, in validating design and implementation in \textsc{Arm}\xspace processors without having to deploy large (and potentially) expensive testbeds. The QEMU runtime is deployed on a Lenovo ThinkPad with Intel\textregistered{} Core\texttrademark{} i7-5600U CPU @ 2.60GHz. We rely on OP-TEE version 3.11.0. \subsection{Micro-benchmarks} We begin by micro-benchmarking two of the subcomponents of the \textsc{KeVlar-Tz}\xspace TA, namely the Base64 encoder and the one in charge of cryptographic operations. \begin{figure}[!t] \centering \includegraphics[scale=0.7]{plots/combined-base64/combined-base64-bars.pdf} \caption{Base64 encoding and decoding throughput for randomly generated data.} \label{fig:base64} \end{figure} \textbf{Base64 encoding and decoding}. We measure the throughput of the base64 encoding and decoding operations. The measurements have been done by measuring the encoding and decoding of randomly generated data of 1KB and 100KB, both for the hardware deployment as well as under emulation. In both cases, the component is deployed in the TEE. We show average and standard deviation for each configuration, which is executed 200 times. Our results are shown in Figure~\ref{fig:base64}. First, we observe the size of the data payload does not negatively affect the observed throughput, whereas we do observe differences between the emulated and hardware environment. For instance, encoding 100KB of data reaches 43 MB/s, while the QEMU is approximately 4$\times$ faster, reaching 130MB/s. Similar differences can be observed for smaller data and decoding. We also report the results obtained when executing the same operations in the REE. As we see, encoding and decoding throughputs are similar, due to the operations being GPU-intense, instead of memory-intense. \begin{figure}[!t] \centering \includegraphics[scale=0.7]{plots/combined-cipher/combined-cipher-bars.pdf} \caption{Encryption and decryption throughput. We compare the built-in \textsc{Op-Tee}\xspace library for cryptographic operations against OpenSSL. For both cases we deploy them in TEE, and compare the throughput to encode and decode randomly generated data.} \label{fig:crypto} \end{figure} \textbf{Cryptographic operations}. Next, we measure the throughput of the cryptographic operations run by \textsc{KeVlar-Tz}\xspace, \ie symmetric encryption and decryption. We generate random data of different sizes: 128B, 1kB and 4kB. Figure~\ref{fig:crypto} depicts our results. We observe that encryption and decryption achieve similar encoding and decoding throughputs in each of the two environments (QEMU and the Raspberry Pi). For instance, we observe an average of 2 MB/s encrypting a payload of 4kB in QEMU, and a 25\% improvement for the same test in hardware. Expectedly, decryption operations are slightly faster (by 6\% on average). We compare our results with OpenSSL version 1.1.1f, running on the REE of both QEMU and Raspberry Pi. We observe that OpenSSL in the REE is much faster, especially for decryption operations which can be parallelized: this is expected, as OpenSSL optimizes the compiled binary for the underlying hardware. We leave as future work further investigation and porting of a (subset of) OpenSSL to run in the TEE. \textbf{TCP communication}. The TPC sockets handle the communication between \textsc{KeVlar-Tz}\xspace and untrusted applications. In this benchmark, we measure the throughput of our trusted TCP channels, whose endpoints terminate into the \textsc{TrustZone}\xspace area. We measure the throughput for messages of different sizes: 1B, 245B (\ie, 128B once encrypted and encoded in base64), and 757B (\ie, 512B plus AES encryption and base64 encoding), and 1024B. We use these values since they represent a reasonable range of values found in real-world deployments. Figure~\ref{fig:tcp} reports our results for the two testing environments. We observe that the throughput is significantly higher for larger amounts of data. Concerning the system used, we see that the Raspberry Pi is much faster than the emulated environment. \begin{figure}[!t] \centering \includegraphics[scale=0.7]{plots/combined-tcp/combined-tcp.pdf} \caption{Trusted TCP server: incoming throughput.} \label{fig:tcp} \end{figure} \begin{figure}[!t] \centering \includegraphics[scale=0.7]{plots/combined-tcp/combined-tcp-ree.pdf} \caption{TCP throughput REE to REE \ob{prettify plot and caption}} \label{fig:tcp} \end{figure} \textbf{Cache module}. The last in our series of micro-benchmarks focus on the throughput of the cache module itself. First, we fill up the persistent storage with 200 keys, using a payload of 32B. Figure~\ref{fig:persistent-store} indicates that as the number of keys increase, there is a corresponding increase in the time to insert the key and value in the persistent storage. \begin{figure}[!t] \centering \includegraphics[scale=0.7]{plots/both-persistent-storage/both-store-key.pdf} \caption{Time to store a 32B encryption key (value) with a 12B ID (key) to persistent storage.} \label{fig:persistent-store} \end{figure} Finally, we query the previously stored keys by issuing request queries for random keys to the cache module. Each cache request can result in a \emph{hit} or \emph{miss}. These results are shown in Figure~\ref{fig:cache-get}. We can observe 5 orders of magnitude between the hits and miss throughputs, both in the case of emulated as well as hardware deployments. \begin{figure}[!t] \centering \includegraphics[scale=0.7]{plots/both-cache/both-cache.pdf} \caption{Throughput of queries. We compare the difference between volatile memory (cache hits) and persistent storage (cache miss) on both hardware (Raspberry Pi) and emulated (QEMU) environments. The percentiles of the distribution are represented with shades of gray. From the brightest to the darkest: minimum, 25th, the 50th (median), the 75th and the maximum percentile.} \label{fig:cache-get} \end{figure} \subsection{Macro-benchmark: Digital Health} We conclude our evaluation by demonstrating the overall performance of \textsc{KeVlar-Tz}\xspace. We setup the Digital Health scenario (Section~\ref{ssec:health}), where heart-rate monitoring data streams are pushed toward \textsc{KeVlar-Tz}\xspace, leaving the Smart Building use-case to future work. For the considered workload, we use a database obtained from CSEM’s proprietary wrist-located sensors and chest-located dry electrodes~\cite{chetelat2015}. In particular, cardiac data is obtained following a standardized protocol in which they perform a range of physical activities from sedentary to vigorous~\cite{delgado2014}. A 5-second sample of the ECG used is shown on Figure~\ref{fig:macro}. In this scenario, the sensors inject 10 electrocardiogram data points every 93.4ms. Each data point embeds the following information: the time when it was taken and the voltage measured. Plot~\ref{fig:macro} compares the time taken to process 60 seconds of such streaming for different amounts of clients. \vs{maybe this plot has to be changed a bit, to be discussed tomorrow. Since this is a 'scalability' plot, we should show plot a bit differently the data. On the y-axis you can put the average time to process 1 second of message. Maybe it is sufficient to divide /60 the y-axis values.} We conclude that the Raspberry Pi takes approximately 0.064 seconds to process 1 second of streaming of 1 client, while the emulated environment takes 0.107 seconds for the same amount of data. \begin{figure}[t] \centering \includegraphics[width=0.7\linewidth]{plots/macro/macro.pdf} \includegraphics[width=0.29\linewidth]{plots/ecg/ecg.pdf} \caption{Time to process a 60-second data stream from ECG sensors.} \label{fig:macro} \end{figure} \vspace{-10pt} \section{Conclusion and Future Work} \label{sec:conclusion} \vspace{-6pt} Motivated by the increasing attack surface of today's edge devices, \textsc{KeVlar-Tz}\xspace addresses an integral part for storing and performing computations over the collected/transmitted data in a privacy-preserving fashion. This becomes critical when the data is highly sensitive and personal, which is the case for nowadays medical implantables, wearables, and nearables. For instance, in scenarios where such IoT devices interact by means of publish/subscribe frameworks, as is typical in real-world deployments, protecting the brokers with a minimal increase in power consumption is necessary in order to preserve the ubiquity of such network of sensing devices. We intend to extend \textsc{KeVlar-Tz}\xspace along the following lines. First, we intend to extend the API exposed to the Trusted Application so that it is easier to implement new functions (similar to the already implemented reencryption. Second, standardizing the length of the messages going through TCP when communicating with untrusted applications, so that binary data can be sent and easily parsed, this will reduce the amount of bytes sent as well as eliminate the base64 dependency. Finally, we intend to compare \textsc{KeVlar-Tz}\xspace to other \textsc{TrustZone}\xspace cache implementations. \vspace{-10pt} \section{Future Work}\label{sec:futurework} \section{Acknowledgements}\label{sec:acks} This work is supported in part by Moore4Medical, which has received funding within the Electronic Components and Systems for European Leadership Joint Undertaking (ECSEL JU) in collaboration with the European Union's H2020 framework Programme (H2020/2014-2020) and National Authorities, under grant agreement H2020-ECSEL-2019-IA-876190. Moreover, this project has received funding from the European Union's Horizon 2020 research and innovation programme under grant agreement No 766733. \section{Related Work}\label{sec:related-work} The problem of executing software inside TEEs in general, and \textsc{TrustZone}\xspace in particular, has attracted several research groups. CaSE~\cite{zhang2016case} is a cache-assisted secure execution framework for the \textsc{TrustZone}\xspace to defend against multiple attacks. Others~\cite{havet2017securestreams,park2019streamboxtz,segarra2019meddata} have implemented frameworks for TEEs to securely process data streams that could benefit from \textsc{KeVlar-Tz}\xspace. However, while such projects have implemented full-fledged frameworks, \textsc{KeVlar-Tz}\xspace provides a lean and resource-efficient cache with an easy-to-use API for applications that need fast access to persistent data. A service like \textsc{KeVlar-Tz}\xspace is available within \textsc{Mqt-Tz}\xspace~\cite{segarra2020mqttz-srds}, a publish-subscribe framework optimized for IoT and \textsc{TrustZone}\xspace deployments and backward compatible with \texttt{mosquitto}\xspace, a well-known MQTT messaging framework supporting TLS. In that context, a secure cache like the one developed in \textsc{KeVlar-Tz}\xspace protects data against eavesdroppers or untrusted brokers. \textsc{KeVlar-Tz}\xspace offers a more generic approach, including an API to use it from inside the TEE and a modular design to choose a specific cache eviction policy or some of its internal subcomponents. Recently~\cite{sasaki2020secure,wan2020rustee}, authors tried to hardening TEE applications against a broad set of attacks, including side-channels or against known weaknesses of the implementation language. While some of the countermeasures developed there could be beneficial for \textsc{KeVlar-Tz}\xspace, we consider those out of scope. In~\cite{gentilal2017bitcoin}, authors implement a cache to speed up operations on a secure Bitcoin wallet, while using the \textsc{TrustZone}\xspace's persistent storage. While the focus is on the security of the private keys used to unlock the cryptocurrency wallet, the approach is similar to \textsc{KeVlar-Tz}\xspace. To the best of our knowledge, \textsc{KeVlar-Tz}\xspace is the first application specifically designed to run on a \textsc{TrustZone}\xspace and provide a lightweight cache to leverage the \textsc{TrustZone}\xspace's persistent storage while maintaining a minimal read/write latency. \textsc{KeVlar-Tz}\xspace implements a generic cache that can be easily embedded into other Trusted Applications or used as a secure storage for an untrusted application in the \emph{normal world} without significantly increasing the trusted computing base. \section{Introduction} \label{sec:intro} \vspace{-4pt} Wearable and Internet-of-Things (IoT) devices are becoming increasingly pervasive in modern society. It is predicted that by the year 2025 there will be more than 600 million wearable devices deployed and connected worldwide~\cite{global-wearable}, and according to Cisco up to 500 billion IoT devices by 2030~\cite{cisco-iot}. These devices continuously produce data from a wide range of sensor types: inertial sensors (\eg accelerometers, gyroscopes)~\cite{Bennett2016}, biopotential (\eg electrocardiography)~\cite{Chaudhuri2009}, optical (\eg photopletysmography)~\cite{Tamura2014}, biochemical (\eg pH)~\cite{Coyle2014}, \etc. Combinations of such sensors allow for the monitoring of the health statuses of the users, ranging from the user's physical activity~\cite{DelgadoGonzalo2017} to the detection of cardiac abnormalities~\cite{Faraone2020}. The nature of this data is intrinsically privacy-sensitive. Applications and system designers must protect it from malicious attackers, including those with physical access, from accessing and possibly unveiling them. Similarly, IoT devices are regularly used to monitor and record privacy-related data. Examples include motion sensors (\eg, in the case of a smart-home deployment, revealing for instance the presence of humans indoors~\cite{lin2016iot}), power-consumption meters (\eg, potentially revealing the habits of a household), weather sensors (\eg, a key asset in farming used to decide on optimal irrigation levels~\cite{padalalu2017smart}), \etc. The vast majority of such applications deal with the insertion and retrieval of data from/to a dedicated, and preferably persistent, memory area. The mentioned operations are typically offered by key-value stores, \eg, software libraries or services that allow to \texttt{put} and \texttt{get} values associated with unique identifiers (\ie, the keys), for later retrieval, similar to a \emph{caching} mechanism. Note that such libraries are vastly known in literature (\ie \cite{246158,han2011survey}, \etc), extensively studied~\cite{gokhale2010kvzone} and find usage in several and diverse application domains. Noteworthy, the result of confidential computations (\eg, edge-based privacy-preserving machine-learning model training, just to name one) must also be stored and retrieved following the same access patterns. Hence, the content of such memory area must be shielded. The need for stringent data privacy guarantees, such as the mentioned shielding, usually comes at the cost of computational overhead. This is the case of full homomorphic encryption (HE)~\cite{gentry2009fully}, a purely software-based approach to compute and operate over encrypted data. However, recent work~\cite{gottel2018security} has shown how state-of-the-art HE implementations~\cite{halevi2013design} still result in orders of magnitude slowdown even for simple arithmetical operations, and major breakthroughs are yet to be seen for HE to become a viable solution. The introduction and widespread adoption in the last few years of trusted execution environments (TEE) for consumer- and server-grade devices offers an opportunity to combine the need for privacy with the ones of viable performance. TEEs provide a hardware-supported mechanism to maintain the privacy and integrity of data while allowing for efficient and transparent protection from malicious attackers or compromised operating systems. Such protected areas are commonly referred to as \emph{enclaves}, and they represent the main programming abstraction supported by the large majority of available TEEs. Notable examples include Intel\textregistered SGX\xspace~\cite{costan2016intel}, AMD Secure Encrypted Virtualization (SEV)~\cite{kaplan2016amd} for server-grade as well as cloud-based deployments~\cite{gce-sev,sgx-aws} and \textsc{Arm}\xspace \textsc{TrustZone}\xspace~\cite{amacher2019performance,pinto2019demystifying} for more edge-centric scenarios, the focus of this work. In this practical experience report paper, we present \textsc{KeVlar-Tz}\xspace, an efficient trusted cache application for \textsc{Arm}\xspace \textsc{TrustZone}\xspace with support for non-volatile secure storage. \textsc{KeVlar-Tz}\xspace exposes an easy-to-use REST interface to facilitate the integration with existing systems, protocols, and third-party devices. The network connection endpoints are established within the \textsc{TrustZone}\xspace enclave. Finally, \textsc{KeVlar-Tz}\xspace is designed to exploit the secure storage implemented by some \textsc{TrustZone}\xspace-enabled systems, allowing for secure data durability. The main \textbf{contributions} of this work are twofold. First, we present the design and implementation of \textsc{KeVlar-Tz}\xspace, a secure cache for \textsc{Arm}\xspace \textsc{TrustZone}\xspace. Second, we describe in detail our implementation and evaluate it using real-world data, including a performance comparison with an emulator, showcasing the performance-tradeoffs that practitioners must face. \textbf{Roadmap.} The rest of this paper is organized as follows. Section~\ref{sec:background} provides relevant background material on TEEs, \textsc{TrustZone}\xspace and trusted applications in general, including some of the underlying libraries and systems used in our evaluation. We present the architecture of \textsc{KeVlar-Tz}\xspace in Section~\ref{sec:architecture}, detailing some of its implementation details in Section~\ref{sec:implementation}. Section~\ref{sec:evaluation} presents our in-depth performance evaluation of the \textsc{KeVlar-Tz}\xspace prototype, using micro- and macro-benchmarks as well as real-world data. We survey related work in Section~\ref{sec:related-work}, before concluding and devising future work in Section~\ref{sec:conclusion}. \section{Architecture} \label{sec:architecture} \begin{figure}[!t] \centering \includegraphics[scale=0.7]{figures/kevlar-v3.pdf} \caption{Architecture of the \textsc{KeVlar-Tz}\xspace TA.} \label{fig:kevlar:arch} \end{figure} \textsc{KeVlar-Tz}\xspace implements a secure key-value storage with non-volatile entries (\ie, available across reboots). To do so, we leverage \textsc{Op-Tee}\xspace's Trusted Storage API~\cite{optee-securestorage} to store keys to a secure persistent storage, while implementing a cache in volatile memory to minimize the number of requests made to the persistent storage. The key idea is to limit as much as possible operations (\ie, \texttt{read}/\texttt{write}) involving the persistent storage, as they are considerably slower (see our evaluation in Section~\ref{sec:evaluation}). The architecture of \textsc{KeVlar-Tz}\xspace is depicted in Figure~\ref{fig:kevlar:arch}. In the remainder of this section, we describe the designing principles behind its various components as well as their interaction. Finally, we detail the typical workflow of a single \texttt{write} operation, a keystone operation of \textsc{KeVlar-Tz}\xspace. \textbf{Secure Persistent Storage}. The persistent storage area is a dedicated hardware component that guarantees data durability, confidentiality and integrity. \textsc{Op-Tee}\xspace supports two modes for secure storage: \emph{(1)} using the REE file-system (the default option), or \emph{(2)} relying on a Replay Protected Memory Block (RPMB) partition of an eMMC device~\cite{reddy2015mobile}. \textsc{KeVlar-Tz}\xspace uses the REE file-system. \textsc{KeVlar-Tz}\xspace implements a wrapper around the Trusted Storage API to access directly to writing and reading operations, which otherwise requires the management of several internal \textsc{Op-Tee}\xspace components (omitted from the Figure~\ref{fig:kevlar:arch} for the sake of clarity). This wrapper exposes two functions: \begin{itemize} \item \texttt{read\_ss(const char *key, char *value, uint32\_t *value\_sz)}: reads the data mapped to a given \texttt{key}, which is bound to the array pointed by \texttt{value}; \item \texttt{write\_ss(const char *key, const char *value, uint32\_t value\_sz)}: writes the data in \texttt{value} mapped to the \texttt{key} into persistent storage. \end{itemize} Finally, we note that the available storage memory dedicated to this component is only limited by the underlying hardware. \textbf{Volatile memory -- Cache}. This component is the secure caching component of \textsc{KeVlar-Tz}\xspace. Our design supports a few cache eviction policies (currently limited to Least Recently Used LRU and FIFO). The implementation uses structs inserted in a queue and a hash table. These are used to handle the key and value of each entry. The queue is used to remember the order of deletion of entries when new entries are to be added to a full cache; the hash table is used to access entries in average constant time. The cache is \texttt{write-through}~\cite{jouppi1993cache}, so that if the trusted application is stopped unexpectedly, no data is lost. \textbf{API for Trusted Applications}. \textsc{KeVlar-Tz}\xspace provides a very simple API for applications running inside the TEE with four operations: \begin{itemize}[noitemsep,topsep=0pt,parsep=0pt,partopsep=0pt] \item Initialize a cache with a given configuration consisting of cache size, hash output size and eviction policy; \item Delete a cache, freeing all space used in volatile memory. Objects in persistent storage are left untouched. \item Query a cache, to fetch the value associated to a given key. For instance, when using \textsc{Mqt-Tz}\xspace~\cite{segarra2020mqttz-srds}, for a given ID, the cache will return the corresponding encryption key. \item Save a new key/value pair in volatile and persistent memory. \end{itemize} \textbf{TCP interface for applications of REE}. The TEE and REE are two different systems and, as such, programs can't communicate (\ie, share data) between each other as if they where running on the same machine. However, \textsc{KeVlar-Tz}\xspace can be useful as a secure cache service to an application running in the \emph{normal world} (\eg, in the \textsc{Mqt-Tz}\xspace broker scenario~\cite{segarra2020mqttz-srds}). To expose \textsc{KeVlar-Tz}\xspace to the \emph{normal world}, we designed and implemented a TCP interface, protected by \textsc{TrustZone}\xspace, that allows to communicate \textsc{KeVlar-Tz}\xspace with any other application reachable on the network. The establishment of the TCP connection works as follows. First, an application in the REE opens a server TCP socket. Secondly, \textsc{KeVlar-Tz}\xspace connects to such socket and waits (\ie, blocks) for new messages. Once a new message is received, \textsc{KeVlar-Tz}\xspace will execute the requested operation and return the desired value \textbf{The workflow of a write}. To conclude the description of the architecture, we take a step-by-step walkthrough for a \texttt{write} operation to insert a new key/value pair into \textsc{KeVlar-Tz}\xspace. When an REE-based application needs to store a new key/value, it must first connect to \textsc{KeVlar-Tz}\xspace via TCP, and pass over the content of the key/value pair. For the sake of simplicity, we assume those to be encoded using \texttt{base64}. Once received by the \textsc{KeVlar-Tz}\xspace TA, they get base64-decoded, and saved to the persistent storage. The architecture allows to attach additional application-specific processing operations to the inserted key/value pairs, both before or after the value is retrieved. For instance, one might send a cipher that will be decrypted with one value and encrypted with another~\cite{segarra2020mqttz-srds}, securely changing the encryption key of a cipher without the REE ever getting ahold of any of them. This post-retrieve operations can be changed to any operation needed for the application that is using \textsc{KeVlar-Tz}\xspace. If an application in the \emph{secure world} uses \textsc{KeVlar-Tz}\xspace to store a new key, it can directly leverage the functions exposed by the \textsc{KeVlar-Tz}\xspace API (\texttt{cache\_save\_object(Cache *cache, char *id, char *data)}), which takes the binary values and stores them to the persistent storage. \section{Appendix: \textsc{KeVlar-Tz}\xspace API}\label{appendix:api} \vs{Add some text describing this} \begin{itemize} {\color{red} \item \texttt{init\_cache(int max\_size, int hash\_size, int id\_sz, int data\_sz, int policy)} \item \texttt{void free\_cache(Cache *cache)} \item \texttt{char * cache\_query(Cache *cache, char *id)} \item \texttt{TEE\_Result cache\_save\_object(Cache *cache, char *id, char *data)} } \end{itemize} \section{Implementation}\label{sec:implementation} This section describes some of the internal details and implementation \sloppy{choices of \textsc{KeVlar-Tz}\xspace}. The system itself is implemented in C, and consists of 791 LoC, released as open-source from \url{https://github.com/mqttz/kevlar-tz/}. We note that applications implemented using the \textsc{Op-Tee}\xspace framework are basically organized as two distinct components: the Host Application (HA) and the corresponding Trusted Application (TA). The host application runs on the \emph{normal world} and initilizes and finalizes the TEE context using the TEE Client API. Moreover, the HA is in charge of invoking functions over the Trusted Application, and can do so multiple times, dividing the work between the \emph{normal} and \emph{secure world}. However, the TEE's volatile memory is lost between calls, hence \textsc{KeVlar-Tz}\xspace's Host Application only invokes the TA once. The TA acts as a daemon that receives queries. We detail the TA components next. \subsection{\textsc{KeVlar-Tz}\xspace Trusted Application} The \textsc{KeVlar-Tz}\xspace trusted application is split into several modules. The implementation of a particular module is independent of the rest, to facilitate future evolutions of the code in a loosely coupled manner (\eg you can change the communication module to work with UDP instead of TCP, different symmetric encryption algorithm, \etc). The \textsc{KeVlar-Tz}\xspace TA is composed of the following modules, which we evaluate individually and as a part of micro-benchmarks in Section~\ref{sec:evaluation}. \textbf{Persistent storage module}. This module implements the functions \texttt{read\_ss} and \texttt{write\_ss}, which read and write persistent storage, respectively. Our implementation follows the guidelines from the Linaro Security Working Group.~\footnote{\url{https://github.com/linaro-swg/optee_examples/tree/master/secure_storage}} \textbf{Cache module}. The \textsc{KeVlar-Tz}\xspace cache module directly implements the proper cache API, namely: \texttt{init\_cache}, \texttt{free\_cache}, \texttt{cache\_query}, and \texttt{cache\_save\_object}. In our implementation, the cache is made of nodes that are part of both a queue and a hash table, which enables accessing objects in constant time (on average). The cache module also interacts with the persistent storage (using the pertinent module). Any access to the key-value storage can be done through the cache, whether the value was stored on volatile memory or not. \textbf{AES module}. Our prototype includes a symmetric cipher module on top of AES. It directly exposes two functions: \texttt{encrypt}, which encrypts data with a given key, and \texttt{reencrypt}, which given two keys and a cipher, decrypts it with one key and encrypts it with the other. Our implementation uses the Cryptographic Operation API implemented by \textsc{Op-Tee}\xspace to encrypt and decrypt the data. We extended it to support PCKS padding, \ie the default padding used by OpenSSL and \textsc{Mqt-Tz}\xspace. \textbf{Base64 module}. This module implements standard Base64 encoding/decoding operations (\ie, \texttt{base64\_encode}, \texttt{base64\_decode}), as well as auxilary ones (\ie, \texttt{base64\_decode\_length} to return the length an encoded string after the decoding). The encoding and decoding implementations leverage an open-source library.\footnote{\url{https://web.mit.edu/freebsd/head/contrib/wpa/src/utils/base64.c}} This module is used to encode and decode data transmitted to other applications over the network layer to simplify the parsing of data, in particular when dealing with multipart binary messages. \textbf{Trusted TCP module}. \textsc{KeVlar-Tz}\xspace uses TCP to communicate with untrusted applications. It exposes functions to initialize and close a connection (\ie, \texttt{net\_connect} and \texttt{net\_disconnect}) as well as send and receive packets (\ie, \texttt{net\_send} and \texttt{net\_receive}). The implementation uses the socket library that \textsc{Op-Tee}\xspace exposes, and while we use TCP, the code can easily be adapted to use other protocols without any changes on any other part of the application. \section{Motivating Scenarios} \label{sec:scenarios} In this section, we describe our two main real-world scenarios behind \textsc{KeVlar-Tz}\xspace, also depicted in Figure~\ref{fig:kevlar:scenarios}. The use-cases originate from two ongoing EU H2020 projects, in collaboration with industry-leading companies, which we details next. \subsection{Digital Health}\label{ssec:health} The first scenario stems from the H2020 project MOORE4MEDICAL.\footnote{\url{https://moore4medical.eu/}} One of its objectives is to use wearable sensors and remote sensing technologies to reduce hospitalization, resulting in more comfort for the patient and less costly clinical trials in drug development. In this context, the monitoring of vital signs is increasingly off-loaded and out-sourced to third-party untrusted data centers. The main reason for such off-load is to exploit the economy of scale that comes with cloud computing. The flow of data is mainly generated from smart medical devices and sensors and it is composed of a mix of physiological signals (\eg, electrocardiograms, photopletymograms) and vital signs (\eg heart rare, respiration rate, stress levels). The data streams are highly heterogeneous, since the physiological signals can reach high sampling rates (\eg, Holter operate at 1~kHz) and the vital signs have typically much lower sampling rates ($\sim$1~Hz). \subsection{Smart Building Management} The second scenario stems from TABEDE\footnote{\url{http://www.tabede.eu/}}, an EU H2020 project with the aim to integrate energy grid demand-response schemes into buildings through low-cost extenders for Building Management Systems or as a standalone system, which is independent from communication standards and integrates innovative flexibility algorithms. The flow of information relies on MQTT brokers deployed at the edge to minimize latency and limits the physical access from untrusted entities. However, it directly raises several privacy and security concerns. The flow of data is mainly generated from home appliances and sensors and it is composed of physical magnitudes such as electric current, temperature, or humidity. The data streams are generated at a slow frequency ($<$1~Hz) and are transfered via a large variety of communication protocols (\eg, EnOcean~\cite{li2014bacnet}, KNX~\cite{lee2009implementation}, Zigbee~\cite{farahani2011zigbee}).

5b61d009ef01b7e90f8ba4ec9f3bdc333ba0b938

\section{Introduction} \input{sections/introduction} \input{sections/preliminaries} \section{Algorithmic framework} \label{sec:framework} \input{sections/algorithm} \subsection{Convergence analysis} \label{sec:convergence} \input{sections/convergence} \section{Convergence results under constraint qualifications} \label{sec:convergenceKKT} \input{sections/convergenceKKT} \section{Numerical Experiments} \label{sec:experiments} \input{sections/numerical_results} \section{Conclusions} \label{sec:conclusions} \input{sections/conclusions} \section{On the relationship between stationarity conditions and KKT conditions} \label{sec:appendix} Consider the continuous optimization problem \begin{equation} \label{eq:convex_prob} \begin{aligned} \min_{x}\;& f(x)\\\text{s.t. }& x\in X, \end{aligned} \end{equation} where $X=\{x\in\mathbb{R}^n\mid h(x) = 0,\;g(x)\le 0\}$ is a convex set ($h_i$, $i=1,\ldots,p$ are affine functions, $g_i$, $i=1,\ldots,m$, are convex functions). We assume $f$ and $g$ to be continuously differentiable; $h$ is differentiable, being affine. \begin{definition} A point $x^*\in X$ is a \textit{stationary point} for problem \eqref{eq:convex_prob} if, for any direction $d$ feasible at $x^*$, we have \begin{equation*} \nabla f(x^*)^{\top}d\ge 0. \end{equation*} \end{definition} It can be shown that a point $x^*$ is stationary for problem \eqref{eq:convex_prob} if and only if \begin{equation} \label{eq:proj-stat} x^* = \Pi_X[x^*-\nabla f(x^*)], \end{equation} where $\Pi_X$ denotes the orthogonal projection operator. Stationarity is a necessary condition of optimality for problem \eqref{eq:convex_prob}. It is possible to show that a point satisfying the KKT conditions is always a stationary point. Viceversa is true by stronger assumptions on the set of feasible directions. \begin{proposition} \label{prop:appendix} Let $x^*\in X$ satisfy KKT conditions for problem \eqref{eq:convex_prob}. Then, $x^*$ is stationary for problem \eqref{eq:convex_prob}. \end{proposition} \begin{proof} Assume $x^*$ satisfies KKT conditions with multipliers $\lambda$ and $\mu$. Let $d$ be any feasible direction at $x^*$. Since $X$ is convex, we know that: \begin{gather} \label{eq:app_h} \nabla h_i(x^*)^{\top}d=0\quad \forall i=1,\ldots,p,\\ \label{eq:app_g} \nabla g_i(x^*)^{\top}d\le 0 \quad \forall i:g_i(x^*)=0. \end{gather} Moreover, from KKT conditions we know that \begin{gather} \label{eq:app_compl} \lambda_i=0 \quad \forall\, i:g_i(x^*)<0. \end{gather} We know that $$\nabla f(x^*) + \sum_{i=1}^{m}\lambda_i\nabla g_i(x^*) + \sum_{i=1}^{m}\mu_i\nabla h_i(x^*)=0,$$ hence $$\left(\nabla f(x^*) + \sum_{i=1}^{m}\lambda_i\nabla g_i(x^*) + \sum_{i=1}^{p}\mu_i\nabla h_i(x^*)=0\right)^{\top}d=0,$$ and then $$\nabla f(x^*)^{\top}d + \sum_{i=1}^{m}\lambda_i\nabla g_i(x^*)^{\top}d + \sum_{i=1}^{m}\mu_i\nabla h_i(x^*)^{\top}d=0.$$ From equations \eqref{eq:app_h} and \eqref{eq:app_compl}, we get $$\nabla f(x^*)^{\top}d + \sum_{i:g_i(x^*)=0}^{}\lambda_i\nabla g_i(x^*)^{\top}d=0,$$ thus, recalling \eqref{eq:app_g} and $\lambda\ge0$, $$\nabla f(x^*)^{\top}d =- \sum_{i:g_i(x^*)=0}^{}\lambda_i\nabla g_i(x^*)^{\top}d\ge 0.$$ Since $d$ is an arbitrary feasible direction, we get the thesis. \end{proof} \begin{proposition} \label{prop2:appendix} Let $x^*\in X$ be a stationary point for problem \eqref{eq:convex_prob}. Assume that one of the following conditions holds: \begin{itemize} \item [(i)] the set of feasible direction $D(x^*)$ is such that $$ D(x^*)=\{d\in \mathbb{R}^n:\nabla g_i(x^*)^{\top}d\le 0, \ i\in I(x^*), \nabla h_i(x^*)^{\top}d=0, i=1,\ldots ,p\} $$ \item [(ii)] the set of feasible direction $D(x^*)$ is such that \begin{gather*} D(x^*)=\{d\in \mathbb{R}^n\mid\nabla g_i(x^*)^{\top}d<0, \ i: g_i(x^*) = 0,\ \nabla h_j(x^*)^{\top}d=0, j=1,\ldots ,p\}, \end{gather*} and a constraint qualification holds. \end{itemize} Then, $x^*$ is a KKT point. \end{proposition} \begin{proof} Assertion (i). Let $x^*$ be a stationary point. Then, there does not exist a direction $d\in D(x^*)$ such that $$ \nabla f(x^*)^{\top}d<0. $$ This implies that the system $$ \begin{array}{ccc} \nabla f(x^*)^{\top}d&<0&\\ \nabla g_i(x^*)^{\top}d & \le 0 & \ \ i: g_i(x^*) = 0\\ \nabla h_i(x^*)^{\top}d&\le 0 & i=1,\ldots ,p\\ -\nabla h_i(x^*)^{\top}d&\le 0 & i=1,\ldots ,p\\ \end{array} $$ does not admit solution. By Farkas' Lemma we get the thesis. \par\medskip\noindent Assertion (ii). Let $x^*$ be a stationary point. Then, there does not exist a direction $d\in D(x^*)$ such that $$ \nabla f(x^*)^{\top}d<0. $$ This implies that the system $$ \begin{array}{ccc} \nabla f(x^*)^{\top}d&<0&\\ \nabla g_i(x^*)^{\top}d & < 0 & \ \ i: g_i(x^*) = 0\\ \nabla h_i(x^*)^{\top}d&= 0 & i=1,\ldots ,p\\ \end{array} $$ does not admit solution. By Motzkin’s theorem we get that $x^*$ satisfies the Fritz-John conditions and hence, by assuming a constraint qualification, the thesis is proved. \end{proof} \par\medskip\noindent Condition (i) of \cref{prop2:appendix} holds if the functions $g_i$, $i=1,\ldots ,m$, $h_j$, $j=1,\ldots ,p$ are affine. \par\medskip\noindent Condition (ii) of \cref{prop2:appendix} holds by assuming that the convex functions $g_i$, for $i=1,\ldots ,m$ are such that \begin{equation}\label{d1} g_i(x+td)\ge g_i(x)+t\nabla g_i(x)^{\top}d+\frac{1}{2}\gamma t^2\|d\|^2 \end{equation} with $\gamma >0$. Indeed, in this case it is easy to see that a direction $d$ is a feasible direction at $x^*$ if and only if $$ \nabla g_i(x^*)^{\top}d<0\quad \ \ i: g_i(x^*) = 0 \quad\quad \nabla h_j(x^*)^{\top}d=0\quad i=1,\ldots ,p $$ \par\medskip\noindent Condition \eqref{d1} is satisfied by assuming that the functions $g_i$ are twice continuosly differentiable and the Hessian matrix is positive definite. \par\medskip\noindent Condition \eqref{d1} holds also for continuously differentiable functions $g_i$ assuming that they are {\it strongly convex with constant $\mu_i>0$}, i.e., that for $i=1,\ldots ,m$ the functions $$ g_i(y)\ge g_i(x)+\nabla g_i(x)^\top(y-x)+\frac{\mu_i}{2}\|y-x\|^2, \quad \forall\ x,y. $$ \subsection{Generality of Discrete Stationarity} In this Section, we show that our definition of discrete stationarity is rather general and that, by a suitable discrete neighborhood $\mathcal{N}(x,y)$ most of the optimality conditions summarized in Section \ref{sec:opt_cond} can be recovered. \begin{itemize} \item As shown in \cref{prop:dic_impl_s}, by simply using KKT as continuous stationarity and $\mathcal{N}(x,y)=\{(x,y)\}$ in \cref{def: stationary}, we obtain a stronger condition than S-stationarity and therefore also stronger than M-stationarity. \item Strong Lu-Zhang conditions can be retrieved as follows: KKT-stationarity is considered; the neighborhood $$\mathcal{N}(x^*,y^*)=\{(x,y)\mid x = x^*,\;e^{\top}y=n-s,\; y_ix^*_i=0\;\forall\,i=1,\ldots,n\}.$$ Note that, since $x$ is constant in the discrete neighborhood, (KKT) stationarity has to be checked for all $y$ defining a super support set. \item Lu-Zhang conditions cannot be defined by using discrete neighborhoods, because it is not possible to know a priori which particular super support set has to be checked. \item Basic feasibility can be obtained by using stationarity \eqref{eq:proj-stat} and, again, the neighborhood $$\mathcal{N}(x^*,y^*)=\{(x,y)\mid x = x^*,\;e^{\top}y=n-s,\; y_ix^*_i=0\;\forall\,i=1,\ldots,n\}.$$ Note that BF stationarity requires that, for any $y_J$ defining a super support set, $x^*=\Pi_{\mathcal{X}(y_J)}[x^*+d]$, where $d_J=-\frac{1}{L}\nabla_J f(x^*)$ and $d_{\bar{J}}=0,$ whereas the conditions in \cref{def: stationary} require $x^*=\Pi_{\mathcal{X}(y_J)}[x^*-\nabla f(x^*)]$. In fact, in the case of our problem the two conditions are equivalent, as we show hereafter. Let us consider $$\hat{x} = \Pi_{\mathcal{X}(y)}[x^*-\nabla f(x^*)],\qquad \tilde{x} = \Pi_{X(y)}[x^*+d].$$ Let us denote $\hat{x}^p = x^k-\nabla f(x^k)$ and $ \tilde{x}^p = x^k+d$ Since both $\hat{x}$ and $\tilde{x}$ belong to $\mathcal{X}(y_J)$, we have $\hat{x}_{\bar{J}}=0$ and $\tilde{x}_{\bar{J}}=0$. From the well-known properties of the projection operator, we get: $$(\hat{x}-\hat{x}^p)^\top(\hat{x}-x)\le 0\quad\forall\,x\in X(y_J),$$ $$(\tilde{x}-\tilde{x}^p)^\top(\tilde{x}-x)\le 0\quad\forall\,x\in X(y_J),$$ hence $$(\hat{x}-\hat{x}^p)^\top(\hat{x}-\tilde{x})\le 0,\qquad (\tilde{x}-\tilde{x}^p)^\top(\tilde{x}-\hat{x})\le 0.$$ Taking into account that $\tilde{x}_{\bar{J}}=\hat{x}_{\bar{J}} = 0$ and $\hat{x}_J^p=\tilde{x}_J^p=x^k-\nabla_J F(x^k)$, we get $$(\hat{x}_J-\hat{x}^p_J)^\top(\hat{x}_J-\tilde{x}_J)\le 0,\qquad (\tilde{x}_J-\hat{x}^p_J)^\top(\tilde{x}_J-\hat{x}_J)\le 0,$$ i.e., $$\|\hat{x}_J\|^2 - \tilde{x}_J^\top\hat{x}_J - \hat{x}_J^\top\hat{x}_J^p + \tilde{x}_J^\top\hat{x}_J^p\le0,\qquad\|\tilde{x}_J\|^2-\tilde{x}_J^\top\hat{x}_J-\tilde{x}_J^\top\hat{x}_J^p+\hat{x}_J^\top\hat{x}_J^p\le 0$$ Summing the two inequalities, we get $$\|\hat{x}_J\|^2+\|\tilde{x}_J\|^2-2\hat{x}_J^\top\tilde{x}_J\le0,$$ i.e., $$\|\hat{x}_J-\tilde{x}_J\|\le0,$$ from which we get $\tilde{x}_J=\hat{x}_J$ and hence $\hat{x} = \tilde{x}.$ \item $L$-stationarity cannot be obtained in our framework, since it is based on the sparse projection operation. \item The definition of CW-optimality is based on argmin operators, i.e., on global information, and thus cannot be directly encapsulated in a concept of stationarity; however, if we relax the definition, we can think of a CW-stationarity concept, i.e., we can replace the argmin operation w.r.t.\ a variable by stationarity w.r.t.\ the same variable; when $\|x^*\|<s$, we require $\nabla_i f(x^*) = 0$, i.e., since we are in the ``unconstrained'' case, $\nabla f(x^*) = 0$; when $\|x^*\|=s$, we require $\nabla f(\hat{x}) = 0$ for all $\hat{x}=x^*-x_i^*e_i$, $i=1,\ldots,n$. This definition can be obtained, in the mixed-integer setting, by using the discrete neighborhood $$\mathcal{N}(x^*,y^*) = \left\{\!\!\begin{array}{ll} \{(x,y)\mid x=x^*,\;e^{\top}y=n-s,\; y_ix^*_i=0\;\forall\,i\}&\text{if }\|x^*\|_0<s,\\ \{(x,y)\mid x=x^*-x_j^*e_j,\;e^{\top}y=n-s,\; y_ix^*_i=0\;\forall\,i\}&\text{if }\|x^*\|_0=s. \end{array}\right.$$ \end{itemize} \subsection{Implementation details} Algorithms SNS, PD and GSS have been implemented in Python 3.7, mainly exploiting libraries \texttt{numpy} and \texttt{scipy}. The convex subproblems of both PD and GSS have been solved up to global optimality by using the L-BFGS algorithm (in the implementation from \cite{liu1989limited}, provided by \texttt{scipy}). We also employed L-BFGS for the local optimization steps in SNS. All algorithms start from the feasible initial point $x^0=0\in\mathbb{R}^n$. For the PD algorithm, we set the starting penalty parameter to 1 and its growth rate to 1.05. The algorithm stops when $\|x^k-y^k\|<0.0001$, as suggested in \cite{Lu2013}. AS for the GSS, we stop the algorithm as soon as $\|x^{k+1}-x^k\|\le0.0001$. Concerning our proposed \cref{alg:MISO}, the parameters have been set as follows: \begin{itemize} \item $\xi = 10^{3}$, \item $\theta = 0.5$, \item $\eta_0 = 10^{-5}$. \end{itemize} For what concerns $\mu_0$ and $\delta$, we actually keep the value of $\mu$ fixed to $10^{-6}$. We again employ the stopping criterion $\|x^{k+1}-x^k\|\le0.0001$. For all the algorithms, we have also set a time limit of $10^4$ seconds. All the experiments have been carried out on an Intel(R) Xeon E5-2430 v2 @2.50GHz CPU machine with 6 physical cores (12 threads) and 16 GB RAM. As benchmark for our experiments, we considered 18 problems, obtained from the 6 datasets in \cref{tab:logistic_datasets} and setting $s$ to 3, 5 and 8 in \eqref{eq:sparse_logistic}. For SNS and GSS we consider the computational time employed to find the best solution. We take into account four versions of \cref{alg:MISO}, with neighborhood radius $\rho\in\{1,2,3,4\}$. In \cref{fig:pp} the performance profiles \cite{Dolan2002} w.r.t.\ the objective function values and the runtimes (intended as the time to find the best solution) attained by the different algorithms are shown. We do not report the runtime profile of SNS(1) since it is much faster than all the other methods and thus would dominate the plot, making it poorly informative. We can however note that unfortunately its speed is outweighed by the very poor quality of the solutions. We can observe that increasing the size of the neighborhood consistently leads to higher quality solutions, even though the computational cost grows. We can see that SNS (with a sufficiently large neighborhood) has better performances than the other algorithms known from the literature; in particular, while the neighborhood radius $\rho=1$ only allows to perform forward selection, with poor outcomes, $\rho\ge 2$ makes swap operations possible, with a significant impact on the exploration capabilities. The GSS has worse quality performance than SNS(2), which is reasonable, since its move set is actually smaller and optimization is always carried out w.r.t.\ a single variable and not the entire active set. However, it proved to also be slower than the SNS, mostly because of two reasons: it always tries all feasible moves, not necessarily accepting the first one that provides an objective decrease, and it requires many more iterations to converge, since it considers one variable at a time. Finally, the PD method appears not to be competitive from both points of view: it is slow at converging to a feasible point and it has substantially no global optimization features that could guide to globally good solutions. \begin{figure}[htbp] \centering \subfloat[objective value]{\includegraphics[width=0.48\textwidth]{images/logistic_obj.pdf}} \subfloat[time]{\includegraphics[width=0.48\textwidth]{images/logistic_time_to_opt.pdf}} \caption{Performance profiles for the considered algorithms on 18 sparse logistic regression problems.} \label{fig:pp} \end{figure} It is interesting to remark how considering larger neighborhoods appears to be particularly useful in problems where the sparsity constraint is less strict and thus combinatorially more challenging. As an example, we show the runtime-objective tradeoff for the \texttt{breast}, \texttt{spam} and \texttt{a2a} problems for $s=3$ and $s=8$ in \cref{fig:tradeoffs}. We can observe that for $s=3$, SNS finds good, similar solutions for either $\rho=2,3$ or $4$, with a similar computational cost. On the other hand, as $s$ grows to 8, using $\rho=4$ allows to significantly improve the quality of the solution without a significant increase in terms of runtime. \begin{figure}[htbp] \centering \subfloat[\texttt{breast} - $s=3$]{\includegraphics[width=0.48\textwidth]{images/plot_breast_3_tradeoff.pdf}} \subfloat[\texttt{breast} - $s=8$]{\includegraphics[width=0.48\textwidth]{images/plot_breast_8_tradeoff.pdf}} \subfloat[\texttt{spam} - $s=3$]{\includegraphics[width=0.48\textwidth]{images/plot_spam_3_tradeoff.pdf}} \subfloat[\texttt{spam} - $s=8$]{\includegraphics[width=0.48\textwidth]{images/plot_spam_8_tradeoff.pdf}} \subfloat[\texttt{a2a} - $s=3$]{\includegraphics[width=0.48\textwidth]{images/plot_a2a_3_tradeoff.pdf}} \subfloat[\texttt{a2a} - $s=8$]{\includegraphics[width=0.48\textwidth]{images/plot_a2a_8_tradeoff.pdf}} \caption{Quality/cost trade-off for the algorithms on sparse logistic regression problems from datasets \texttt{breast}, \texttt{spam} and \texttt{a2a}.} \label{fig:tradeoffs} \end{figure} \FloatBarrier \section{Basic definitions and preliminary results} \label{sec:opt_cond} Even though problem \eqref{prob: cc_prob} is a continuous optimization problem, it has an intrinsic combinatorial nature and in applications the interest often lies in finding a good, possibly globally optimal configuration of active variables. Being \eqref{prob: cc_prob} a continuous problem, $x^*\in \mathcal{X}$ is a local minimizer if there exists an open ball $\mathcal{B}(x^*,\epsilon)$ such that $f(x^*)=\min\{f(x)\mid x\in \mathcal{X}\cap \mathcal{B}(x^*,\epsilon)\}$. In some works from the literature (e.g., \cite{Burdakov2016,Lu2013}) necessary conditions of local optimality have been proposed. However, for this particular problem every local minimizer for a fixed active set of $s$ variables is a local minimizer of the given problem. Hence the number of local minimizers grows as fast as $\binom{n}{s}$ and is thus of low practical usefulness. In \cite{Beck2013,Beck2016}, the authors propose necessary conditions for global optimality that go beyond the concept of local minimum described above, thus allowing to consider possible changes to the structure of the support set, and reducing the pool of optimal candidates. However, these conditions are either tailored to the ``unconstrained case'', or limited to moderate changes in the support, or involve hard operations, such as exact minimizations or projections onto nonconvex sets. In order to introduce a general and affordable necessary optimality condition that also takes into account the combinatorial nature of the problem, we consider in our analysis the equivalent reformulation of problem \eqref{prob: cc_prob} described in \cite{Burdakov2016}: \begin{equation}\label{prob:mi_prob} \begin{aligned} \min_{x,y}\;&f(x)\\ \text{s.t. }& e^\top y \geq n - s, \\ & x_i y_i = 0, \quad \forall\, i=1,\ldots,n,\\ & x \in X,\\ & y \in \{ 0, 1 \}^n. \end{aligned} \end{equation} From here onwards, we will use the following notation: \begin{equation*} \arraycolsep=1.4pt \begin{array}{rl} \mathcal{Y} &= \{y \mid y \in \{0,1\}^n, e^\top y \geq n - s\},\\ \mathcal{X}(y) &= \{ x \in X \mid x_i y_i = 0 \ \forall\, i=1,\ldots,n \}. \end{array} \end{equation*} We further define the support set of a vector $z$ by \[ I_1(z) = \{ i \mid z_i \neq 0 \}, \] while its complement is defined by \[ I_0(z) = \{ i \mid z_i = 0 \}. \] Moreover, we recall the concept of super support set \cite{Beck2016}: \begin{definition} Let us consider a feasible point $z$ for problem \eqref{prob: cc_prob}. A set $J\subset \{1,\ldots ,n\}$ is called super support of $z$ if it is such that $|J|=s$ and $I_1(z)\subseteq J$. \end{definition} We denote by $z_I$ the subvector of $z$ identified by the components contained in an index set $I$. We also denote by $\Pi_C$ the orthogonal projection operator over the closed convex set $C$. We notice that given a feasible point $(x,y)$ of problem \eqref{prob:mi_prob}, the components $I_0(y)$ give an \emph{active subspace} for $x$, i.e., those components identify the subspace where the nonzero components of $x$ lay. We thus have that $I_1(x)\subseteq I_0(y)$. Nonlinear mixed-integer programs have been characterized exploiting the notion of \textit{neighborhood} \cite{li2006nonlinear,Lucidi2005}. Given a feasible point $(x, y)$, a discrete neighborhood $\mathcal{N} (x,y)$ is a set of feasible points that are close, to some extent, to $(x,y)$ and that contains $(x,y)$ itself. We introduce here an example of tailored neighborhood for problem \eqref{prob:mi_prob} that can be implemented at a reasonable computational cost. Such a neighborhood will also help us to relate our analysis to the other theoretical tools available in the literature. \begin{definition} \label{ex:discrete_neighborhood} Let $d_H:\{0,1\}^n\times\{0,1\}^n\to \mathbb{N}$ denote the Hamming distance. Moreover, let $J(y,\hat{y})=\{i\mid y_i\neq \hat{y}_i\}$ and let ${H}_{J(y,\hat{y})}(\cdot)$ be a function such that $\hat{x}={H}_{J(y,\hat{y})}(x)$ is defined as $$\left\{\!\!\begin{array}{ll} \hat{x}_h= 0 &\text{if }h\in J(y,\hat{y})\\ \hat{x}_h=x_h&\text{otherwise} \end{array}\right.$$ Then, given $\rho\in\mathbb{N}$, the neighborhood is \begin{equation} \label{D_N} \mathcal{N}_\rho(x,y)=\left\{(\hat{x},\hat{y})\mid e^\top\hat{y}\ge n-s,\;d_H(\hat{y},y)\le \rho,\; \hat{x}=H_{J(y,\hat{y})}(x)\right\}. \end{equation} \end{definition} We notice that this particular definition of neighborhood allows to take into account the potential ``change of status'' of up to $\rho$ variables in the vector $\hat y$ defining an active subspace. \begin{example} \label{ex:discrete_neighborhood_hamming} Consider the problem \eqref{prob:mi_prob} with $n=3$ and $s=2$ and let $\rho = 2$. Let $(x,y)$ be a feasible point defined as follows $$ (x,y)= \left( \begin{array}{l} 1\\ 2\\ 0 \end{array} \right) \left( \begin{array}{l} 0\\ 0\\ 1 \end{array} \right) $$ The neighborhood ${\cal N}_\rho(x,y)$ is given by \begin{gather*} {\cal N}_2(x,y)=\left\{ \left(\begin{array}{l} 1\\2\\0 \end{array}\right) \left(\begin{array}{l} 0\\0\\1 \end{array}\right), \; \left(\begin{array}{l} 1\\0\\0 \end{array}\right) \left(\begin{array}{l} 0\\1\\0 \end{array}\right), \; \left(\begin{array}{l} 0\\2\\0 \end{array}\right) \left(\begin{array}{l} 1\\0\\0 \end{array}\right), \; \left(\begin{array}{l} 1\\0\\0 \end{array}\right) \left(\begin{array}{l} 0\\1\\1 \end{array}\right), \; \left(\begin{array}{l} 0\\2\\0 \end{array}\right) \left(\begin{array}{l} 1\\0\\1 \end{array}\right), \; \left(\begin{array}{l} 0\\0\\0 \end{array}\right) \left(\begin{array}{l} 1\\1\\1 \end{array}\right) \right\} \end{gather*} \end{example} Now, a notion of local optimality for problem \eqref{prob:mi_prob}, depending on the neighborhood $\mathcal{N} (x,y)$, can be introduced: \begin{definition} A point $(x^*,y^*)\in \mathcal{X}(y^*)\times \mathcal{Y}$ is a local minimizer of problem \eqref{prob:mi_prob} if there exists an $\epsilon> 0$ such that and for all $(\hat{x},\hat{y})\in\mathcal{N}(x^*,y^*)$ it holds \begin{equation*} f(x^*)\le f(x)\quad \forall\, x \in \mathcal{B}(\hat{x},\epsilon )\cap X(\hat{y}). \end{equation*} \end{definition} Note that in the above definition the continuous nature of the problem, expressed by the variables $x$, is taken into account by means of the standard ball $\mathcal{B} (\hat{x}, \epsilon)$. The given definition clearly depends on the choice of the discrete neighborhoods. A larger neighborhood $\mathcal{N} (x^*, y^*)$ should give a better local minimizer, but the computational effort needed to locate the solution may increase. Inspired by the definition of local optimality for problem \eqref{prob:mi_prob}, we introduce a necessary condition of global optimality for problem \eqref{prob: cc_prob} that allows to take into account possible, beneficial changes of the support and that hence properly captures, from an applied point of view, the essence of the problem. Such a condition relies on the use of stationary points related to continuous problems obtained by fixing the binary variables in problem \eqref{prob:mi_prob}, i.e., for a fixed $\bar y \in \mathcal{Y}$, \begin{equation}\label{subprob_cont} \begin{aligned} \min \;& f(x) \\ \text{s.t. } & x \in \mathcal{X}(\bar y). \end{aligned} \end{equation} \begin{definition} \label{def: stationary} A point $x^*\in {\cal X}$ is called an $\mathcal{N}$-stationary point, if there exists an $y^*\in {\cal Y}$ such that \begin{enumerate}[label=(\roman*)] \item $(x^*, y^*)$ is feasible for problem \eqref{prob:mi_prob}; \item the point $x^*$ is a stationary point of the continuous problem \begin{equation*} \begin{aligned} \min \;& f(x) \\ \text{s.t. } & x \in \mathcal{X}(y^*); \end{aligned} \end{equation*} \item every $(\hat{x}, \hat{y}) \in \mathcal{N}_\rho(x^*, y^*)$ satisfies $f(\hat{x}) \geq f(x^*)$ and if $f(\hat{x}) = f(x^*)$, the point $\hat{x}$ is a \mbox{stationary} point of the continuous problem \begin{equation*} \begin{aligned} \min \;& f(x) \\ \text{s.t. } & x \in \mathcal{X}(\hat{y}). \end{aligned} \end{equation*} \end{enumerate} \end{definition} It is easy to see that the following result stands: \begin{theorem}\label{th:nec_cond} Let $x^*$ be a minimum point of problem \eqref{prob: cc_prob}. Then $x^*$ is an $\mathcal{N}$-stationary point. \end{theorem} We will show later in this work that the definition of $\mathcal{N}$-stationariy allows to retrieve in a unified view most of the known optimality conditions, if a suitable neighborhood $\mathcal{N}$ is employed. In \cref{def: stationary} we generically refer to stationary points of problem \eqref{subprob_cont}, namely, to points satisfying suitable optimality conditions. Then, concerning the assumptions on the feasible set $\mathcal{X}(\bar y)$, we distinguish the two cases: \begin{itemize} \item [(i)] no constraint qualifications hold; \item [(ii)] constraint qualifications are satisfied and the usual KKT theory can be applied. \end{itemize} In case (i), we will refer to the following definition of stationary point of problem \eqref{subprob_cont}. \begin{definition}\label{stat_1} Given $\bar y \in \mathcal{Y}$ and $\bar x\in \mathcal{X}(\bar y)$, we say that $\bar x$ is a stationary point of problem \eqref{subprob_cont} if and only if $$ \bar x=\Pi_{\mathcal{X}(\bar y)}\left[\bar x-\nabla f(\bar x)\right].$$ \end{definition} We notice that $\mathcal{X}({\bar y})$ is a convex set when $X$ is convex, then the condition given above is a classic stationarity condition for the problem \eqref{subprob_cont}. Case (ii) will be considered later.

4668d9d15059995acb76f4a4af1e9d308dabcd71

\section{Omitted Proofs} \label{app:omit} \begin{proof}[Proof of \Cref{lem:skewintcor}] (If) Suppose $x_i\in X,y_j\in Y,z_k\in Z$ and $t_{(i,j,k)}\in T$ where $t_{(i,j,k)}=(x_i,y_j,z_k)$ then it is easy to verify that indeed \[ t'_{(i,j,k)} + x'_i + y'_j + z'_k + \sum_{l=4}^{m-1}c'_l = b \] (Only if) Conversely, suppose that there is a subset $S\subseteq U'$ with $|S|=m$ which sums to $b$. Let $S=\{a'_1,\dots,a'_m\}$. Considering the equation modulo $r$ and using the fact $\sum_{l=1}^{m}2^l$ is the only possible way of obtaining $2^{m+1}-1$ as a sum of $m$ elements (possibly with repetition) from the set $\{1,2,\dots,2^m\}$. Therefore, we conclude there are some elements $x_i\in X,y_j\in Y,z_k\in Z$ and a tuple $t_{(i_0,j_0,k_0)}\in T$ such that $x'_i,y'_j,z'_k,t'_{(i_0,j_0,k_0)}\in S$. Now, considering the equation modulo $r^2,r^3,r^4$ gives us $i=i_0$,$j=j_0$, and $k=k_0$. \end{proof} \begin{proof}[Proof of \Cref{lem:skewbinsize}] The first part follows from the fact that the first component of any vector is strictly greater than $\frac{1}{m+1}$. The second part of the claim follows from the fact any vector in the instance has first component greater than $\frac{1}{m+1}$ and the dummy vector has first component equal to $\frac{m-1}{m+1}$. For the third part, observe if there are two dummy vectors then both the components would add up to $2\cdot \frac{m-1}{m+1}>1$. However, if one of them is a non-dummy vector, then notice that both components of a non-dummy vector are less than $\frac{2}{m+1}$, and both components of any vector are less than or equal to $\frac{m-1}{m+1}$. Hence, both components of the sum are less than 1. \end{proof} \begin{proof}[Proof of \Cref{lem:skewvectorcor}] (If) For a tuple $t_{(i,j,k)}=(x_i,y_j,z_k)$, we have $t_{(i,j,k)}'+x'_i+y'_j+z'_k+\sum_{l=4}^{m-1}c'_l=b$ by \Cref{lem:skewintcor}. So, we have \begin{align*} &\mathbf t_{(i,j,k)} + \mathbf x_i + \mathbf y_j + \mathbf z_k + \sum_{l=4}^{m-1}\mathbf c_l\\ &= \left(\frac{m}{m+1}+\frac{t'_{(i,j,k)}+x'_i+y'_j+z'_k+\sum_{l=4}^{m-1}c'_l}{(m+1)b},\frac{m+2}{m+1}-\frac{t'_{(i,j,k)}+x'_i+y'_j+z'_k+\sum_{l=4}^{m-1}c'_l}{(m+1)b}\right)\\ &= (1,1). \end{align*} (Only if) Suppose there is a set $S=\{\mathbf a_i\}$ of $m$ vectors which fit in a bin. By \Cref{lem:skewbinsize} all the vectors are non-dummy vectors. Hence, each vector can be written as: \[ \mathbf a_i = \left(\frac{1}{m+1}+\frac{a'_i}{(m+1)b},\frac{m+2}{m(m+1)}-\frac{a'_i}{(m+1)b}\right). \] As they fit in a bin we get the following two conditions, \begin{align*} \frac{m-1}{m+1}+\frac{\sum\limits_{i=1}^ma'_i}{(m+1)b}\leq 1,\\ \frac{m+2}{m+1}-\frac{\sum\limits_{i=1}^ma'_i}{(m+1)b}\leq 1, \end{align*} which simplify to $\sum\limits_{i=1}^ma'_i\leq b$ and $\sum\limits_{i=1}^ma'_i\geq b$. Combining the inequalities we get $\sum\limits_{i=1}^ma'_i=b$. Therefore, by \Cref{lem:skewintcor} we have shown what we set out to show. \end{proof} \begin{proof}[Proof of \Cref{lem:vectorcor2}] (If) For a tuple $t_{(i,j,k)}=(x_i,y_j,z_k)$, we have $t'_{(i,j,k)}+x'_i+y'_j+z'_k=b$ by \Cref{lem:intcor}. So, we have \[ \mathbf t_{(i,j,k)} + \mathbf x_i + \mathbf y_j + \mathbf z_k = \left(\frac{4}{5}+\frac{t'_{(i,j,k)}+x'_i+y'_j+z'_k}{5b}, \frac{6}{5}-\frac{t'_{(i,j,k)}+x'_i+y'_j+z'_k}{5b}\right) = (1,1). \] (Only if) Suppose there are four vectors $\mathbf a_1,\mathbf a_2,\mathbf a_3,\mathbf a_4$ which cover a non-D-bin. By our assumption all the vectors are non-dummy vectors. Hence each vector can be written as, \[ \mathbf a_i = \left(\frac{1}{5}+\frac{a'_i}{5b},\frac{3}{10}-\frac{a'_i}{5b}\right). \] As they fit in a bin we get the following two conditions, \begin{align*} \frac{4}{5}+\frac{\sum\limits_{i=1}^4a'_i}{5b}\geq 1,\\ \frac{6}{5}-\frac{\sum\limits_{i=1}^4a'_i}{5b}\geq 1, \end{align*} which simplify to $\sum\limits_{i=1}^4a'_i\geq b$ and $\sum\limits_{i=1}^4a'_i\leq b$. Combining the inequalities, we get $\sum\limits_{i=1}^4a'_i=b$. Therefore, by \Cref{lem:intcor} the vectors correspond to a tuple. \end{proof} \section{Introduction} In the $d$-dimensional vector bin packing problem we are given a set of $d$-dimensional vectors (say $S$) each of whose components belong to $(0,1]$, i.e., $S\subseteq (0,1]^d$. The aim is to find the minimum cardinality partition of $S$ into bins $\{B_i|i\in [m]\}$ such that $\left\|\sum_{v\in B_i}v\right\|_\infty\leq 1$, for any $i\in [m]$. Whenever the above condition holds for a set of vectors $B$ we say that the vectors in $B$ fit in a bin. This is a natural generalization of the classical bin packing problem, which can be obtained by setting $d=1$. We also study related problem called the vector bin covering problem. This is a generalization of another classical problem called the bin covering problem. In the $d$-dimensional vector bin covering problem we are again given a set of $d$-dimensional vectors (say $S$) each of whose components belong to $(0,1]$. The aim is to obtain a family of disjoint subsets (also called unit cover or bin) with the maximum cardinality such that for each bin $B$, we have $\sum_{v\in B}v\geq \mathbf 1$. There is a well-known reduction from the partition problem to the classical bin packing showing that it is NP-hard to approximate bin packing within a factor of $3/2$. This guarantee holds even for vector bin packing. However, note that this does not even rule out extremely strong guarantees like $\text{OPT}+1$. Nevertheless, if we take a closer look at the reduction, we can conclude that the above lower bound holds only for instances with solutions with small values (e.g., a small number of bins in bin packing), which is not typical of most instances. Hence, it is instructive to look at the asymptotic approximation ratio ($R_\infty$) for a minimization problem is defined by, \begin{align*} R_\infty&=\limsup_{n\to \infty} R_n\\ R_n &= \max_I\left\{\frac{A(I)}{\text{OPT}(I)}|\text{OPT}(I)=n\right\} \end{align*} where $I$, $A(I)$, $\text{OPT}(I)$ denote the instances of the problem, the value of the output of an algorithm $A$, and the optimal value for the instance $I$, respectively. An asymptotic polynomial time approximation scheme (APTAS) for a problem is a class of algorithms for the problem in which there is an $1+\epsilon$ approximation algorithm for each $\epsilon>0$. For a maximization problem, we have $R_n = \max\left\{\frac{\text{OPT}(I)}{A(I)}|\text{OPT}(I)=n\right\}$ and $R_\infty=\limsup R_n$ while the definition of APTAS remains the same. In fact, even in case of bin covering there is another well-known reduction from the partition problem to bin covering showing it is NP-hard to get an absolute approximation ratio of $2-\epsilon$. Hence, we will be looking at the asymptotic approximation ratio for both these problems. For the remainder of this paper, we will refer to the asymptotic approximation ratio simply as the approximation ratio. \subsection{Related Works} For vector bin packing, in the case where $d$ has been supplied as part of the input, de la Vega and Lueker gave $(d+\epsilon)$-approximate algorithm in \cite{DBLP:journals/combinatorica/VegaL81}. This algorithm is almost optimal as there is a reduction from the vertex coloring problem known, which shows a $d^{1-\epsilon}$ hardness (see \cite{DBLP:conf/soda/BansalE016}). If $d$ is kept fixed, i.e., it is not supplied as part of the input, then the above lower bound does not hold, and in fact, much better approximation factors are known for this case. The barrier of $d$ was broken by Chekuri and Khanna \cite{DBLP:journals/siamcomp/ChekuriK04} by obtaining a $1+\epsilon d + H_{\epsilon^{-1}}$ approximation where $H_k=\sum_{m=1}^k \frac{1}{m}$. Notice that taking $\epsilon = \frac{1}{d}$, this implies a $\mathcal O(\ln d)$ approximation (in fact\footnote{$\gamma\approx 0.57721$ is the Euler-Mascheroni constant} $\ln d + 2 + \gamma$). This was further improved to $\ln d + 1$ by Bansal, Caprara, and Sviridenko~\cite{DBLP:journals/siamcomp/BansalCS09} and then to $\ln (d+1) + 0.807$ by Bansal, Eli\'a\v{s} and Khan \cite{DBLP:conf/soda/BansalE016}. Recently Sandeep \cite{DBLP:journals/eccc/Sandeep21} showed that the best approximation ratio any vector bin packing algorithm, for high enough dimensions, can have is $\Omega (\ln d)$. In the $d=2$ case, Bansal, Caprara, and Sviridenko \cite{DBLP:journals/siamcomp/BansalCS09} gave a 1.693 approximation algorithm which was later improved to 1.406 by Bansal, Eli\'a\v{s}, and Khan \cite{DBLP:conf/soda/BansalE016}. Bansal, Eli\'a\v{s}, and Khan \cite{DBLP:conf/soda/BansalE016} also note that skewed instance constitute the hard instances for rounding based algorithms for 2-dimensional vector bin packing. An item is called $\delta$-\emph{large} if at least two dimensions are larger than $\delta$, for some constant $\delta>0$, otherwise it is called $\delta$-\emph{skewed}. In fact, the case where all items are skewed forms an important sub-case for many packing problems. G\'alvez et al. \cite{DBLP:conf/approx/Galvez0AJ0R20} study the strip packing problem in this context and give a $(\frac{3}{2}+\epsilon)$-approximation and also show an (almost) matching $(\frac{3}{2}-\epsilon)$ lower bound. Recently, Khan and Sharma \cite{DBLP:journals/corr/abs-2105-02827} gave an APTAS for 2-dimensional geometric bin packing with skewed items. They also note that it is possible to solve Maximum Indepent Set of Rectangle (MISR) and 2-dimensional geometric knapsack (2GK) exactly in polynomial time if all items are $\delta$-large. Finally, not only are the skewed instances of packing problems of theoretical interest but they occur in practice as well, e.g., in scheduling, it captures scenarios where no job can consume a significant amount of a shared resource (energy, memory space, etc.) for a significant amount of time. For vector bin covering, when $d$ is supplied as part of the input the best known algorithm by Alon et al. \cite{DBLP:journals/algorithmica/AlonACESVW98} gives an approximation ratio of $O(\ln d)$. In the same paper, they also give another algorithm which gives $d$-approximation which is better than their $O(\log d)$ algorithm for small values of $d$. Finally, Sandeep \cite{DBLP:journals/eccc/Sandeep21} also gave a lower bound of $\Omega(\frac{\log d}{\log \log d})$ for arbitrary value of dimension $d$. For further information on approximation and online algorithms for multidimensional variants of the bin packing and bin covering problems, we refer the reader to the survey \cite{DBLP:journals/csr/ChristensenKPT17} by Christensen et al. \subsection{Our Results} It was believed that, \cite{DBLP:journals/ipl/Woeginger97} shows there is no APTAS for vector bin packing with $d\geq 2$. However, as we show in \cref{sec:original}, there was a minor oversight in the original proof. Specifically, an essential claim made in the proof fails to hold for a few special cases. We also examine some natural modifications to the claim, which exclude those special cases. We conclude that it is impossible to obtain the final result if we try to use the same arguments. Unfortunately, this oversight is also present in the $\frac{391}{390}$ lower bound for vector bin packing by Chleb\'ik and Chlebikov\'a \cite{DBLP:journals/eccc/ECCC-TR06-019}. Hence, we present a revision to the original proof in \cref{sec:revised}. Although our proof uses essentially the same construction as the original proof, the final analysis is slightly different, and the main ideas for the analysis are borrowed from \cite{DBLP:journals/mor/BansalCKS06,DBLP:journals/jda/ChlebikC09}. Specifically, we have obtained a gap reduction instead of an approximation ratio preserving reduction. The APX-hardness of vector bin packing, though not stated explicitly, was considered to be a simple corollary of the original result (see \cite{DBLP:journals/mor/BansalCKS06}). Although the revised proof gives us a constant lower bound on the approximation ratio, we cannot conclude that the vector bin packing problem is APX-hard. We note that Sandeep's lower bound of $\Omega(\ln d)$ does not hold for low dimensions, and hence, it does not even rule out the possibility of APTAS in the 2-dimensional case. Our second result is concerning vector bin packing with skewed items. Extending the proof of our non-existence of APTAS for vector bin packing we show that for $\epsilon\in (0,\frac{1}{2500})$ we need $\delta\leq 20\sqrt \epsilon$ to obtain a $(1+\epsilon)$-approximation for $\delta$-skewed $d$-dimensional vector bin packing. This rules out a APTAS for a fixed $\delta$ for $d$-dimensional vector bin packing (for $d\geq 2$). Finally, we adapt our proof for vector bin covering to show that there is no APTAS for vector bin covering with dimension $d\geq 2$. \subsection{Preliminaries} As is the case in \cite{DBLP:journals/ipl/Woeginger97} and \cite{DBLP:journals/mor/BansalCKS06}, we start with maximum 3-dimensional matching (denoted by MAX-3-DM) and reduce it to an instance of 4-Partition and finally reduce it to a vector bin packing instance. In a 3-dimensional matching instance we have three sets $X=\{x_1,x_2,\dots,x_q\}$,$Y=\{y_1,y_2,\dots,y_q\}$, and $Z=\{z_1,z_2,\dots,z_q\}$ and we are given a set of tuples $T\subseteq X\times Y\times Z$. The aim is to find the maximum cardinality subset $T'\subseteq T$ such that no element from $X,Y,$ or $Z$ occurs in more than one tuple. In the bounded variant of MAX-3-DM (denoted by MAX-3-DM-$B$), it is assured that any element which belongs to either $X,Y,$ or $Z$ will appear in at most $B$ tuples. Kann \cite{DBLP:journals/ipl/Kann91} showed that bounded maximum matching with bound 3 is MAX SNP-hard which in turn implies it is APX-hard. Later, it was shown by Petrank \cite{DBLP:journals/cc/Petrank94} that it is NP-hard to distinguish between instances where there is a solution $T'$ with $|T'|=q$ and from instances for which every solution $T'$, $|T'|\leq (1-\epsilon)q$ for a constant $\epsilon$. There is also a more restricted variant of the problem, which is frequently studied where there are exactly $B$ tuples for each element of the sets $X$,$Y$, and $Z$ called the exact maximum 3-dimensional matching (denoted by MAX-3-DM-E$B$). In case of MAX-3-DM-E2 it was shown by Berman and Karpinski \cite{DBLP:journals/eccc/ECCC-TR03-008} that it is NP-hard to approximate with ratio better than $\frac{98}{97}$, which Chleb\'ik and Chleb\'ikov\'a \cite{DBLP:journals/tcs/ChlebikC06} improved to $\frac{95}{94}$. Finally, Chleb\'ik and Chleb\'ikov\'a \cite{DBLP:journals/jda/ChlebikC09} also note an useful corollary of their $\frac{95}{94}$ bound, which is the following result for the promise variant of MAX-3-DM-E2, \begin{theorem}{\cite{DBLP:journals/jda/ChlebikC09}} \label{thm:matching} Let $I_M$ be a MAX-3-DM-E2 instance comprising of sets $X,Y,Z$ and tuples $T\subseteq X\times Y\times Z$ with $|X|=|Y|=|Z|=q$. Then, it is NP-hard to distinguish between the case with $\text{OPT}(I_M)\geq \lceil \beta_0 q \rceil$ and $\text{OPT}(I_M)\leq \lfloor \alpha_0 q \rfloor$, where $\alpha_0 = 0.9690082645$ and $\beta_0 = 0.979338843$. \end{theorem} For our result for the skewed item case we note that the size of the items in the reduction from MAX-3-DM to 2-dimensional vector bin packing can be reduced by going through $m$-Partition instead of 4-Partition. Finally, for our vector bin covering result we use the same reduction to 4-Partition to finally obtain a reduction to a vector bin covering instance. \section{The Main Result} \label{sec:revised} In this section, we prove our main result, i.e., there is no APTAS for vector bin packing. We do so by modifying the construction in the original proof given in \cite{DBLP:journals/ipl/Woeginger97} by adding a set of dummy vectors.\footnote{We have tried to stay as close to the original proof as possible in terms of notations and arbitrary choices. Yet we have made two notable changes, (i) using $r=64q$, and (ii) using $t_{(i,j,k)}$ to denote a tuple.} The final analysis is based on the analysis in \cite{DBLP:journals/mor/BansalCKS06} for the geometric bin packing lower bound. We start by defining a few integers based on the given MAX-3-DM instance $I_M$. Let $r=64q$, where $q=|X|=|Y|=|Z|$ and $b=r^4+15$. Define integers $x'_i,y'_i,z'_i$ corresponding to ${x_i\in X},{y_i\in Y}, {z_i\in Z}$ to be, \begin{align*} x'_i=ir+1,\\ y'_i=ir^2+2,\\ z'_i=ir^3+4, \end{align*} and for $t_{(i,j,k)}=(x_i,y_j,z_k)\in T$ define $t'_{(i,j,k)}$ as, \[ t'_{(i,j,k)}=r^4-kr^3-jr^2-ir+8. \] Let $U'$ be the set of integers constructed as above. Also, note that for any integer $a'\in U'$ constructed as above we have $0<a'<b$. These integers were constructed so that the following statement holds (cf. Observation 2 in \cite{DBLP:journals/ipl/Woeginger97}). \begin{lemma} \label{lem:intcor} A set of four integers from $U'$ add up to $b$ if and only if they correspond to some elements $x_i\in X,y_j\in Y,z_k\in Z$ and tuple $t_{(i,j,k)}\in T$ where $t_{(i,j,k)}=(x_i,y_j,z_k)$. \end{lemma} \begin{proof} (If) Suppose $x_i\in X,y_j\in Y,z_k\in Z$ and $t_{(i,j,k)}\in T$ where $t_{(i,j,k)}=(x_i,y_j,z_k)$ then it is easy to verify that indeed \[ t'_{(i,j,k)} + x'_i + y'_j + z'_k = b \] (Only if) Conversely, suppose that four integers $a'_1,a'_2,a'_3,a'_4$ sum to $b$. Considering the equation modulo $r$ and using the fact $1+2+4+8$ is the only possible way of obtaining 15 as a sum of four elements (possibly with repetition) from the set $\{1,2,4,8\}$ and therefore we conclude the integers must correspond one element each from $X,Y,Z,T$. This means we can write $a'_1,a'_2,a'_3,a'_4$ as $x'_i,y'_j,z'_k,t'_{(i_0,j_0,k_0)}$. Now, considering the equation modulo $r^2,r^3,r^4$ gives us $i=i_0$,$j=j_0$, and $k=k_0$. \end{proof} To obtain a vector bin packing instance for each integer $a'$ constructed above construct the following vector, \[ \mathbf a = \left(\frac{1}{5}+\frac{a'}{5b},\frac{3}{10}-\frac{a'}{5b}\right). \] We also construct additional $|T|+3q-4\beta(I_M)$ dummy vectors as follows, \[ \mathbf d = \left(\frac{3}{5},\frac{3}{5}\right), \] where $\beta(\cdot)$ is an arbitrary function which will be fixed later. We now note a few properties of the vectors. First of these pertains to how many vectors can fit in a bin (cf. Observation 4, Lemma 2.5 from \cite{DBLP:journals/ipl/Woeginger97} and \cite{DBLP:journals/mor/BansalCKS06}). \begin{lemma} \label{lem:binsize} A bin can contain at most $4$ vectors. If a bin contains a dummy vector it can contain at most one more vector. Furthermore, two dummy vectors will not fit in a bin while any other set of two vectors fit in a bin. \end{lemma} \begin{proof} The first part follows from the fact that the first component of any vector is strictly greater than $\frac{1}{5}$. The second part of the claim follows from the fact any vector in the instance has first component greater than $\frac{1}{5}$ and the dummy vector has first component equal to $\frac{3}{5}$. For the third part, observe if there are two dummy vectors then both the components would add up to $2\cdot \frac{3}{5}>1$. However, if one of them is a non-dummy vector, then notice that both components of a non-dummy vector are less than $\frac{2}{5}$, and both components of any vector are less than $\frac{3}{5}$. Hence, both components of the sum are less than 1. \end{proof} The following lemma shows that a configuration corresponding to a tuple is an optimal configuration (cf. Observation 3 from \cite{DBLP:journals/ipl/Woeginger97}). \begin{lemma} \label{lem:vectorcor} A set of four vectors fits in a bin if and only if it corresponds to a tuple. \end{lemma} \begin{proof} (If) For a tuple $t_{(i,j,k)}=(x_i,y_j,z_k)$, we have $t_{(i,j,k)}'+x'_i+y'_j+z'_k=b$ by \Cref{lem:intcor}. So, we have \[ \mathbf t_{(i,j,k)} + \mathbf x_i + \mathbf y_j + \mathbf z_k = \left(\frac{4}{5}+\frac{t'_{(i,j,k)}+x'_i+y'_j+z'_k}{5b}, \frac{6}{5}-\frac{t'_{(i,j,k)}+x'_i+y'_j+z'_k}{5b}\right) = (1,1). \] (Only if) Suppose there are four vectors $\mathbf a_1,\mathbf a_2,\mathbf a_3,\mathbf a_4$ which fit in a bin. By \Cref{lem:binsize} all the vectors are non-dummy vectors. Hence, each vector can be written as: \[ \mathbf a_i = \left(\frac{1}{5}+\frac{a'_i}{5b},\frac{3}{10}-\frac{a'_i}{5b}\right). \] As they fit in a bin we get the following two conditions, \begin{align*} \frac{4}{5}+\frac{\sum\limits_{i=1}^4a'_i}{5b}\leq 1,\\ \frac{6}{5}-\frac{\sum\limits_{i=1}^4a'_i}{5b}\leq 1, \end{align*} which simplify to $\sum\limits_{i=1}^4a'_i\leq b$ and $\sum\limits_{i=1}^4a'_i\geq b$. Combining the inequalities we get $\sum\limits_{i=1}^4a'_i=b$. Therefore, by \Cref{lem:intcor} the vectors correspond to a tuple. \end{proof} Now, we show that the above construction is a gap reduction from MAX-3-DM to 2-dimensional vector packing (cf. Theorem 2.1 from \cite{DBLP:journals/mor/BansalCKS06}), \begin{lemma} \label{lem:main} If a MAX-3-DM instance $I_M$ has a solution with $\beta(I_M)$ tuples the constructed vector bin packing instance has a solution with $|T|+3q-3\beta(I_M)$ bins. Otherwise, if all the solutions of the MAX-3-DM instance have at most $\alpha(I_M)$ tuples then the constructed instance needs at least $|T|+3q-\frac{\alpha(I_M)}{3}-\frac{8\beta(I_M)}{3}$ bins where $\alpha(\cdot)$ is an arbitrary function. \end{lemma} \begin{proof} First, we show that if a MAX-3-DM instance has a matching consisting of $\beta(I_M)$ tuples, then the vector bin packing instance has a solution of $|T|+3q-3\beta(I_M)$ bins. Using Lemma~\ref{lem:vectorcor}, the $4\beta(I_M)$ vectors corresponding to the $\beta(I_M)$ tuples and their elements can be packed into $\beta(I_M)$ bins. Each of the $|T|+3q-4\beta(I_M)$ non-dummy vectors can be packed along with a dummy vector into $|T|+3q-4\beta(I_M)$ bins by \Cref{lem:binsize}. Now, suppose that for a given instance, all the solutions have at most $\alpha(I_M)$ tuples. Let $n_g$ be the number of bins with 4 vectors, $n_d$ be the number of bins with dummy vectors, and $n_r$ be the rest of the bins. Now, since any solution to the bin packing instance must cover all the non-dummy vectors, \begin{enumerate}[(a)] \item any bin containing four vectors consists of only non-dummy vectors by Lemma~\ref{lem:vectorcor}, \item any bin containing a dummy vector contains at most one non-dummy vector, by Lemma~\ref{lem:binsize}, and \item any other bin can contain at most 3 vectors by Lemma~\ref{lem:binsize}. \end{enumerate} Therefore, we have \begin{align*} 4n_g+3n_r+n_d&\geq 3q+|T|. \end{align*} Now, by \Cref{lem:binsize} we have $n_d=|T|+3q-4\beta(I_M)$. Hence, the above inequality simplifies to, \begin{align*} 4n_g+3n_r&\geq 4\beta(I_M)\\ \Rightarrow n_g+n_r&\geq \frac{4}{3}\beta(I_M)-\frac{n_g}{3}\\ \Rightarrow n_g+n_r+n_d&\geq |T|+3q-\frac{n_g}{3}-\frac{8}{3}\beta(I_M)&[\text{Since, } n_d = |T|+3q-4\beta(I_M)]. \end{align*} Since there are at most $\alpha(I_M)$ triples in the MAX-3-DM instance, by Lemma~\ref{lem:vectorcor} we have $n_g\leq \alpha(I_M)$. Therefore, the number of bins is at least \begin{align*} &&|T|+3q-\frac{\alpha(I_M)}{3}-\frac{8\beta(I_M)}{3}. \end{align*} \end{proof} The following inaproximability for vector bin packing directly follows from \Cref{lem:main}. \begin{theorem} There is no APTAS for the $d$-dimensional vector bin packing with $d\geq 2$ unless P=NP. Furthermore, for the 2-dimensional case there is no algorithm with asymptotic approximation ratio better than $\frac{600}{599}$. \end{theorem} \begin{proof} Suppose that there is an algorithm with approximation ratio $1+\frac{\beta_0-\alpha_0}{15-9\beta_0}$. Then we can distinguish between MAX-3-DM-E2 instances (i) having a solution of $\lceil \beta_0 q \rceil$ triples and (ii) having no solutions with more than $\lfloor \alpha_0 q \rfloor$ tuples using \Cref{lem:main} with $\alpha(I_M)=\lfloor \alpha_0 q \rfloor$ and $\beta(I_M)=\lceil \beta_0 q \rceil$. By \Cref{thm:matching}, we know it is NP-hard to distinguish between these two types of MAX-3-DM-E2 instances with $\beta_0=0.979338843$, and $\alpha_0=0.9690082645$. Hence, we obtain the bound of $1+\frac{\beta_0-\alpha_0}{15-9\beta_0}$. Simple calculations will show this is at least $1+\frac{1}{599}$. \end{proof} \section{The Original Proof} \label{sec:original} The original proof uses essentially the same reduction as ours, i.e., there we had $r=32q$, $b=r^4+15$ and then for each $x_i\in X,y_i\in Y,z_i\in Z$ we had, \begin{align*} x'_i=ir+1,\\ y'_i=ir^2+2,\\ z'_i=ir^3+4, \end{align*} and for $t_l\in T$ was $t'_l$ defined by, \[ t'_l=r^4-kr^3-jr^2-ir+8. \] And finally, to obtain a vector bin packing instance for each integer $a'$ constructed above construct the following vector, \[ \mathbf a = \left(\frac{1}{5}+\frac{a'}{5b},\frac{3}{10}-\frac{a'}{5b}\right) \] The above set of vectors forms a 2-dimensional vector bin packing instance $\mathbf U$. A noticeable difference from our reduction being the absence of dummy vectors.\footnote{Also notice that $r=32q$ and tuples are denoted by $t_l$.} In \cite{DBLP:journals/ipl/Woeginger97}, Woeginger claimed that, \begin{claim*}[Observation 4 in \cite{DBLP:journals/ipl/Woeginger97}] Any set of 3 vectors in $\mathbf U$ can be packed in a unit-bin. No set of 5 vectors in $\mathbf U$ can be packed into a unit-bin. \end{claim*} We show that this claim does not hold in general. In particular, all sets of 3 vectors can not be packed into a unit-bin. Consider the tuple vectors for the tuples $t_1={(x_1,y_1,z_1)},t_2={(x_2,y_1,z_1)}$, and $t_3={(x_3,y_1,z_1)}$. According to the claim, the vectors $\mathbf t_1,\mathbf t_2, \mathbf t_3$ corresponding to the above tuples can be packed in a bin. Suppose $\mathbf t_1,\mathbf t_2,\mathbf t_3$ can indeed be packed in a bin. This implies that the first components of the vectors do not exceed 1, i.e., \[ \frac{3}{5}+\frac{t'_1+t'_2+t_3}{5b}\leq 1, \] which simplifies to, \[ t'_1+t'_2+t'_3\leq 2b. \] Finally, using \begin{align*} t'_1=r^4-r^3-r^2-r+8,\\ t'_2=r^4-r^3-r^2-2r+8,\\ t'_3=r^4-r^3-r^2-3r+8, \end{align*} and \[ b=r^4+15, \] along with further simplification we get, \[ r^4\leq 3r^3+3r^2+6r+6. \] But this inequality does not even hold for $r\geq 32$ whereas 32 is the smallest value for $r=32q$. Thus the claim is incorrect. This implies his main lemma (Lemma 5 in \cite{DBLP:journals/ipl/Woeginger97}) fails to hold. \begin{claim*}[Lemma 5 in \cite{DBLP:journals/ipl/Woeginger97}] Let $\alpha>0$ be an integer such that $|T|-\alpha$ is divisible by 3. Then there exists a feasible solution for the instance $I_M$ for MAX-3-DM that contains at least $\alpha$ triples if and only if there exists a feasible packing for the instance $\mathbf U$ of the 2-dimensional vector packing problem that uses at most $|T|+\frac{1}{3}(q-\alpha)$ unit bins. \end{claim*} We now consider, and rule out, a few natural attempts at fixing the proof. As a first attempt, we modify the claim of Observation 4 from \cite{DBLP:journals/ipl/Woeginger97} to exclude the case of 3 tuples. In this case, we see that it is not possible prove the original claim of Lemma 5 from \cite{DBLP:journals/ipl/Woeginger97}. Another attempt inspired by the failures of the above attempt is to consider modifying the claim of Observation 4 to apply to any set of 2 vectors. In this case, it is clear that the original claim of Lemma 5 can not be proven and even the natural modification to the claim of Lemma 5 to consider at most $\frac{1}{2}(3q+|T|)-\alpha$ bins instead of $q+\frac{1}{3}(q-\alpha)$ bins can not be proven. It seems that it may not be possible to show that the construction given in \cite{DBLP:journals/ipl/Woeginger97} is an approximation preserving reduction. Finally, the analysis in \cite{DBLP:journals/eccc/ECCC-TR06-019} also uses the relation in Lemma 5 of Woeginger's proof to obtain their $\frac{391}{390}$ inapproximability which unfortunately means that their analysis also suffers from the same oversight. Hence, the need for the revision. \section{Vector Bin Packing with skewed items} \input{skew} \section{Vector Bin Covering has no APTAS} \label{sec:vbc} \input{vbc} \section*{Concluding Remarks} In the original proof of the non-existence of an APTAS for vector bin packing reduction used claimed to be an approximation preserving reduction, and hence, the APX-hardness of the vector bin packing problem was a simple corollary. This is no longer the case now as we could only show that our reduction is a gap preserving reduction, and hence, it is not known if 2-dimensional vector bin packing is indeed APX-hard. More importantly, there is still a considerable gap between the best-known approximation ratio for $d=2$ case (1.406) and our lower bound (1.00167). We also make similar observation in case of the vector bin covering problem, i.e., we have not showed that it is APX-hard, and the gap between the best-known algorithm for the two-dimensional case (2) and our lower bound (1.001). As Sandeep observed, his reduction (in~\cite{DBLP:journals/eccc/Sandeep21}) does not give the exact lower bounds on the approximation ratio. Hence, the problem of finding the optimal lower bound on the approximation ratio in the exact sense is still open for vector bin packing with fixed dimension. Also, note that in case of vector bin covering there is a gap of factor $c\log \log d$ between the lower bound and upper bound on the approximation ratio. Finally, we observe that \Cref{lem:intcor} implies that the intermediate reduction to 4-Partition is still approximation preserving, and hence, 4-Partition is indeed APX-hard.

a19e01a3e3091bdd6ea13dc70de45bab660c2fdc

\section{\label{sec:1} Introduction} Sound sources localization (SSL) is a research field in acoustics that has only recently made use of machine learning approaches. Even if research works on the subject can be found as early as in the 1990s\cite{steinberg1991neural} -- and sporadically until 2015\cite{Guentchev1998learning, weng2001three, talmon2011supervised, deleforge2015acoustic} -- it is mainly from 2017 onwards, in particular with the development of deep learning, that the scientific community proposed to use advanced machine learning techniques for SSL tasks. The present paper proposes a supervised deep learning approach for the localization of acoustic sources in aerial environment, using raw, unprocessed time-domain acoustic pressure measurements on a compact array of microphones. Although not aiming at giving a systematic and exhaustive bibliographical review on this subject, this section presents an overview of the methods recently proposed in the scientific literature on this subject in which motivates the development of the BeamLearning approach. The growing interest in using deep learning techniques for SSL has been recently illustrated by the publication of the work of several authors\cite{chakrabarty2019multi, perotin2019crnn, adavanne2018sound, brendel2019distributed} in a special issue on acoustic source localization\cite{gannot2019introduction}. The LOCATA challenge\cite{lollmann2018locata}, which brings together a corpus of data recorded by different microphone arrays for different scenarios (single or multiple, moving or static sound sources combinations recorded on static or moving arrays), has also motivated the proposal of the use of a deep neural network \cite{pak2018locata}, which has been shown to outperform both the baseline method and the conventional MUSIC method. While most publications on the subject use supervised learning strategies\cite{liu2020source, pak2019sound, comanducci2020source, yasuda2020sound, varzandeh2020exploiting, chakrabarty2019multi, perotin2019crnn, brendel2019distributed,gemba2019robust, liu2020multi, chakrabarty2017broadband, adavanne2018sound, Adavanne2019_DCASE, hirvonen2015classification,he2018deep,sundar2020raw,yalta2017sound,ma2017exploiting,nguyen2018autonomous,sivasankaran2018keyword,vesperini2016neural,ferguson2018sound, tang2019regression,salvati2018exploiting,takeda2016sound,yiwere2017distance, xiao2015learning,suvorov2018deep, vera2018towards, huang2020time,comminiello2019quaternion} for audible acoustic sources in air and for underwater localization\cite{gemba2019robust, liu2020multi}, unsupervised or semi-supervised learning approaches have also recently been successfully applied for SSL tasks in reverberant and noisy environments \cite{hu2020semi, hu2020unsupervised}. Among these studies, the majority of authors apply the SSL task to a single active source -- this is also the case in this paper -- while others evaluate their methods for multiple sources localization\cite{chakrabarty2017broadband, adavanne2018sound, Adavanne2019_DCASE, hirvonen2015classification,he2018deep, sundar2020raw} or propose specific deep learning approaches to count the number of active sources\cite{stoter2018countnet,GrumiauxKGG20}.\\ The rise of deep learning algorithms based on convolutional neural networks for machine hearing and acoustics\cite{bianco2019machine} -- along with their ability to improve the robustness of direction of arrival (DoA) estimation techniques in presence of noise and reverberation -- led to the use of various kinds of inputs to the proposed neural network architectures. Among them, the most common choices are based on time-frequency representations based on short-time Fourier transforms\cite{chakrabarty2017broadband,chakrabarty2019multi,adavanne2018sound,he2018deep}. Using this bidimensional representation, some authors proposed to exploit the phase information only\cite{chakrabarty2017broadband}, others used both amplitude and phase\cite{adavanne2018sound, Adavanne2019_DCASE}, and some used the power computed from the STFT matrix\cite{yalta2017sound, hirvonen2015classification, salvati2018exploiting}. Still, other kinds of imput features have been proposed in the litterature, such as the GCC features\cite{xiao2015learning}, the eigenvectors of the spatial covariance matrix\cite{takeda2016sound, yiwere2017distance}, or ambisonics signals\cite{adavanne2018sound,comminiello2019quaternion, perotin2019crnn, tang2019regression}. However, there is still no consensus on the best representation to use in order to better encode the information needed to localize sound sources.\\ Since the unprocessed, time-domain audio signals contain all the information to be extracted, the scientific community has recently put some efforts to directly use the raw waveforms as inputs for deep learning models, either for machine hearing tasks\cite{dieleman2014end,dai2017very,sainath2015learning,lee2018samplecnn,sainath2015learning,bavu2019timescalenet,ravanelli2018speaker} or for SSL tasks\cite{suvorov2018deep, vera2018towards, huang2020time, sundar2020raw}. Joint acoustic model learning from the raw waveform has therefore emerged as an active area of research in the last few years, and recent works have shown that this approach, also known as \textit{end-to-end learning}, allows to successfully learn the temporal and spatial dynamics scales of the waveforms. These studies, along with recent advances in machine learning architectures for one-dimensional signals\cite{oord2016wavenet,kaiser2017depthwise,rethage2018wavenet} and the work of the authors on a similar approach for speech and environmental sounds recognition\cite{bavu2019timescalenet} has motivated the present work, which aims at showing the benefit of a deep-learning multi-resolution approach for the SSL task, that allows to avoid the need to pre-process the multichannel waveforms in order to encode the relevant information contained in raw acoustic measurements.\\ Among the already published studies on SSL using deep learning techniques, the problem is the most commonly treated in a classification framework, where the space is divided into distinct zones. These zones can be angular sectors, and thus represent portions of azimuthal angles\cite{yalta2017sound, ma2017exploiting, nguyen2018autonomous, sivasankaran2018keyword}, or combine azimuthal and elevation angles to define portions of a sphere\cite{Adavanne2019_DCASE}. Perotin et al.\cite{perotin2019crnn} proposed to use a classification task to also estimate the distance between the source and the microphone array. However, for a SSL task, a regression approach seems as legitimate as classification\cite{perotin2019regression}. When a high angular precision is required, a classification approach may also not be longer sufficient, and a regression approach, giving a continuous numerical value of the source's angular position\cite{vesperini2016neural,ferguson2018sound,tang2019regression} even appears to be more accurate in scenarios with diffuse interference\cite{perotin2019regression}. In a recent study\cite{sundar2020raw}, the authors propose to combine the two approaches in order to determine the position of several speakers in a room. The position is first roughly inferred using a classification approach, and then refined using a regression approach. With regard to the BeamLearning network detailed in the present paper, both approaches will be studied in order to compare the observed results.\\ Machine learning approaches for SSL localization require a large amount of training data. Since signals captured by microphone arrays with different geometries are radically different, individual experimental data collection is required for each specific type of microphone array. This is why most SSL studies use simulated data, which therefore need to be as realistic as possible in order to adapt to real recording conditions. Some solutions have been proposed in the litterature using domain adaptation\cite{he2019adaptation} when the mismatch between training and testing conditions is too high. We therefore also developed an efficient tensor GPU-based computation of synthetic room impulse responses using fractional delays for image source models.\\ In this article, we present the Beamlearning approach, which relies on the use of a deep neural network (DNN) DoA estimation system, using raw multichannel measurements. The BeamLearning neural network architecture has been specifically developed in order to be insensitive to the microphone array geometry and to the number of sensors involved in the measurement array, and to be compatible with both classification and regression approaches. The obtained results are evaluated for both experimental and synthetic data using speech signals, in anechoic and reverberant environments, with strong noisy conditions. These results are compared on the same data with two state of the art model-based SSL methods (MUSIC and SRP-PHAT). Section II provides details on the method used to build the synthetic datasets, and the BeamLearning neural network architecture is extensively detailed. The different datasets used in this paper are described in Section III, along with the evaluation methods and the training procedure. The ensemble of three multichannel, multiresolution learnable stacked filterbanks, which constitutes the core of the first subnet of the BeamLearning network is analyzed in detail in Section IV. Finally, several SSL experiments are presented and analyzed in Section V, before drawing our conclusions in Section VI. \section{\label{Methods}Methods} \subsection{\label{dataset_generation}Synthetic and experimental datasets generation} For both the dataset generation and the neural network architecture, the BeamLearning approach proposed in the present paper has been tailored and extensively tested in order to accomodate to any microphone array configurations in arbitrary reverberant rooms. However, in this paper, for sake of clarity and concision, we chose to illustrate the results using only a particular microphone array. For the experimental dataset, we used the MiniDSP\textsuperscript{\small{\textregistered}} UMA-8 compact microphone array (\url{https://www.minidsp.com/products/usb-audio-interface/uma-8-microphone-array}). This compact array is composed of $N_c = 7$ digital MEMS microphones. The first one is placed at the center of the circular array, and the 6 others microphones are evenly distributed over the perimeter of a 8 cm diameter circle. This geometry is mostly similar to the microphone arrays used in personal assistants such as Amazon Echo\textsuperscript{\small{\textregistered}}, or Apple HomePod\textsuperscript{\small{\textregistered}}, where the estimation of a speaker's direction could represent an important pre-requisite in the processing chain for hands-free communication during voice interaction with these devices in noisy and reverberating environments.\\ The experimental datasets have been generated using the UMA-8 microphone array, placed at the ''sweet spot'' of the laboratory's 5-th order ambisonic playback system\cite{lecomte2016fifty}, which allows a great flexibility and reproducibility for the generation and labelling task of experimental datasets. This sound field synthesis system consists in a 1.07 m radius spherical loudspeaker array following a 50-node Lebedev grid. The loudspeakers are made with a tubular plastic cabinets and 2''~drivers (AuraSound\textsuperscript{\small{\textregistered}} NSW2-380-8A). Datasets of 72000 sources positions (80\% for learning, 10\% evaluation and 10\% for testing) have been encoded in the ambisonic domain in order to drive the 50 loudspeakers\cite{lecomteambitools}. These sources positions are randomly generated using the procedure detailed in \ref{subsec:Datasets}.\\ In this paper, the BeamLearning network has also been trained and tested using synthetic datasets corresponding to different room configurations. For these datasets, the microphone array geometry matches the one used in the experimental datasets. Since there is a need to model accurately a large number of sound sources positions and signals in various simulated environments, we specifically developed for this purpose a GPU-based computation of synthetic room impulse responses (RIRs) using fractional delays\cite{pujol2019source}. This computation of realistic RIRs relies on the use of the image source method\cite{allen1979image}. In this process, all the reflections on the room's boundaries are computed, for any time of arrival on the array smaller than the reverberation time (RT) of the room. As an example, for a typical classroom size of $7{ }\times10{}\times3.70$~m with a RT of 0.5~s, the whole number of image sources requires a high order of reflections (up to the 60th order), and represents more than 80000 image sources for each of RIR of the synthetic dataset. For each simulated environment, 48000 primary source positions are used in the RIR computations. These sources positions are randomly generated using the procedure detailed in \ref{subsec:Datasets}. Using the UMA-8 microphone array, the dataset therefore corresponds to the computation of 336000 RIRs for each room. For each of these RIRs, the whole number of image sources positions and corresponding attenuations that contribute to the acoustic field for a time interval of $[0;T_R]$ are computed using the Pyroomacoustics library\cite{pyroomacoustics}. For such a large number of RIRs and image sources, the existing frameworks did not allow to perform the RIR computation efficiently and accurately enough. This is the reason why we developed for this specific application a fast parallel batch RIR computation performed on the GPU using the Tensorflow APIs\cite{tensorflow2015-whitepaper}. This implementation is achieved using sparse tensors computations and an efficient fractional delay filtering implementation\cite{pujol2019source}. For a compact microphone such as the UMA-8, there is indeed a critical need to keep the precision of the time delays corresponding to the distance between each of the image sources involved in the RIR computation and the microphones. In order to implement the RIR with non integer sample delays, we therefore used a 7\textsuperscript{th} order Lagrange interpolation\cite{smithdigitalfilters}, that allows to perform a better accuracy than standard truncated sinc interpolations, with a much lower computation cost. The integer part of the sample delay is represented as a sparse array, and the 8 coefficients of each fractional delay filters along with the individual dampings corresponding to the cumulative effect of each wall of the room and the distance between the image source and the microphones are then used to compute the RIRs using every source image contributions\cite{pujol2019source}. Using this large number of RIRs, the complete dataset is built using batch convolutions with real-life recordings for each simulated environments. \subsection{\label{subsec:beamlearning_architecture}Proposed neural network architecture} The proposed BeamLearning neural network is composed of three subnetworks, which are further detailed in the following subsections. This DNN architecture is fed with raw multichannel audio data measured by a microphone array (see Fig.~\ref{fig:Reseau}), without any further preprocessing. This allows to perform a joint feature learning along with the SSL task, and to avoid manual input features choices that may not be well adapted to particular microphone array geometries. We also specifically developed this architecture in order to allow real-time SSL inference after convergence of the training process. This led to the use of input frames of length $N_t = 1024$ samples. Since most conventional beamforming methods can be thought as filter and sum approaches\cite{brandstein2013microphone}, the global neural architecture depicted on Fig.~\ref{fig:Reseau} has been developed in a similar way. The DNN architecture is mainly based on the use of a learnable filterbank, with specific operations allowing the neural network to be expressive enough to achieve the SSL task in reverberant and noisy environments using raw multichannel audio input data of dimension $N_c \times N_t$, with $N_c = 7$ for the particular array used in this paper.\\ \begin{figure*}[ht!] \centering \includegraphics[width=0.99\textwidth]{Fig_01.pdf} \caption{\label{fig:Reseau}{Schematical representation of the global architecture of the BeamLearning network. This neural network takes a raw multichannel waveform measured by a microphone array, and outputs the DOA estimation of the emitting source.}} \end{figure*} The first subnetwork can be thought as a set of finite impulse-response (FIR) filters with learnable coefficients, that self-adapts to the multichannel audio input data for the SSL task. The convolutional neural cells, which are widely used in DNN architectures, can indeed be thought as a strict equivalent to learnable finite impulse response (FIR) filters in standard digital signal processing (DSP). For this purpose, we propose the use of residual networks of one-dimensional depthwise separable atrous convolutions, which allow to operate in a computationally efficient way, with a similar strategy to the one used in the FrameNet subnet that we proposed in a previous published study for sound recognition\cite{bavu2019timescalenet}. A depthwise separable convolution consists in a convolution performed independently over each channel of an input, followed by a pointwise convolution, \emph{i.e.} a 1x1 convolution, projecting the channels output by the depthwise convolution onto a new channel space. This process allows to operate a convolution on data, faster, with much less parameters than standard convolutions\cite{chollet2017xception}. This residual subnetwork of depthwise separable atrous convolutions is described in detail in \ref{subsec:filterbank_subnetwork}. The overall output of these three filterbanks -- which are composed of $3\times768$ cascaded learnable filters, each of them having a channel multiplier of 4 -- is a two-dimensional map, where the first dimension is $N_t$, and the second dimension represents the $N_f$ outputs of the cascaded learnable filters. In the present paper, for a 2D-DoA finding task, $N_f = 128$. This hyperparameter value has been determined after extensive preliminary training and testing phases for the proposed DOA task, in order to optimize the neural network performances while limiting the impact on the computational cost on a single GPU. The obtained data representation is then fed to the second subnetwork depicted on Fig.~\ref{fig:Reseau}, which aims at computing an energy-like representation in $N_f$ dimensions. This subnet is inspired by Beamforming approaches, where the sound source position is determined using a maximization of the Steered Response Power (SRP) of a beam former\cite{brandstein2013microphone}. The last part of the BeamLearning network aggregates the energy of the $N_f$ channels using a full connected subnetwork in order to infer the sound source position. In the following, the architecture of each of these sub-networks is presented and discussed in relation to the physics of the associated SSL problem.\\ \subsubsection{\label{subsec:filterbank_subnetwork}Learnable filterbanks : raw waveform processing}\ As introduced in the previous subsection, from a machine-learning point of view, the first subnetwork in the BeamLearning DNN consists of three stacked modules, which are residual networks of one-dimensional depthwise separable atrous convolutions. The proposed filterbank module architecture is detailed on Fig.~\ref{fig:Filtre}. The use of convolutional layers is a common operation in deep learning for audio applications, that allows to extract features from the data. From a signal processing point of view, the kernel of the convolutional cells can be seen as the Finite Impulse Response (FIR) of a learnable filter. However, acoustic model learning from the raw waveform using dense kernels can lead to large FIR sizes in order to be efficient at filtering at low frequency\cite{bavu2019timescalenet,ravanelli2018speaker}. As a consequence, rather than making convolutions with large and dense kernels, the proposed filterbank is implemented using one-dimensional atrous convolutional kernels, which have a long history in classical signal processing, and have originally been developed for the efficient computation of the undecimated wavelet transform\cite{holschneider1990real}. The use of this kind of convolution with upsampled filters has been revisited in the context of Deep Learning and have already been shown as an efficient architecture for audio generation\cite{oord2016wavenet}, denoising\cite{rethage2018wavenet}, neural machine translation\cite{kaiser2017depthwise}, and word recognition\cite{bavu2019timescalenet}. Similarly to previous published studies\cite{oord2016wavenet,rethage2018wavenet,bavu2019timescalenet}, we use dilation rates which are multiplied by a factor of two for each successive layers and very short kernel widths of 3 samples (early experiments with larger filters of length 5, 7 and 9 showed inferior performances). This allows to achieve a large receptive field (127 samples for a single residual subnetwork of depthwise separable atrous convolutions, and a total receptive field of 379 samples for the three stacked learnable filterbanks), with only $3\times6$ sets of one-dimensional convolutions with kernels of size $1\times3$, corresponding to dilation rates ranging from 1 to 32.\\ The proposed network topology has been developed in order to optimize the SSL performances in the frequency range of [100~Hz ; 4000~Hz]. For such a large frequency range, using raw audio data, it is useful to offer different timescales of analysis in order to extract the relevant information for the SSL task. The use of three stacked residual atrous convolutional filterbanks depicted on Fig.~\ref{fig:Reseau} and Fig.~\ref{fig:Filtre} -- which share the same architecture but do not operate on the same data, being in cascade -- allow the network to operate on multiple time scales in the range of [68 $\mu$s~;~8.6 ms]. The corresponding filter impulse response lengths can thus take values of one period for frequencies ranging from 115~Hz to 14.7~kHz without impacting the computational efficiency. Since the frequency response of a FIR filter is highly dependent on its impulse response length, the proposed network topology -- along with the skip connections offered by the residual network -- allows for a wide range of filtering possibilities in the targeted frequency range.\\ \begin{figure} \centering \includegraphics[width=0.9\columnwidth]{Fig_02.pdf} \caption{\label{fig:Filtre}{Inner architecture of a learnable filterbank module, showing the residual connections (dotted lines) between the successive layers. Each layer corresponds to a different dilation rate D, taking values 1,2,4,8,16, and 32. In each depthwise-separable convolutional layer, the depthwise convolution is achieved using a channel multiplier $m=4$, and a pointwise convolution mapping the outputs of the depthwise convolution into $N_f = 128$ pooled channels.}} \end{figure} In our approach, we use non-causal depthwise separable convolutions, which present the considerable advantage of making a much more efficient use of the parameters available for representation learning than standard convolutions\cite{kaiser2017depthwise}. The convolutions are performed independently over channels (depthwise separable convolutions), with a channel multiplier $m$. These computed depthwise convolutions are then projected onto a new channel space of dimension $N_f$ for each layer using a pointwise convolution, which is a type of convolution that uses a kernel of size $1\times1$ : this kernel iterates through every single time sample of the input tensor. This kernel has a depth of however many channels its input tensor has. From a signal processing point of view, this approach aims at pooling together the contents in the filtered multichannel soundwave that share similar spatial and spectral features, in order to ease the sound localization task: the pointwise convolution therefore helps combining the pooled channels in order to enhance the expressivity of the network.\\ As shown on Fig.~\ref{fig:Reseau}, three of these filterbank modules are stacked, and residual connections are added between each layers of the three successive filterbanks (see Fig.~\ref{fig:Filtre}), in order to allow shortcut connections between layers and offer increased representation power by circumventing some of the learning difficulties introduced by deep layers\cite{he2016identity}. These skip connections allow the information to flow across the layers easier by bypassing the activations from one layer to the next. This limits the probability of saturation or deterioration phenomenons of the learning process, both for forward and backward computations in deep neural networks\cite{he2016identity,he2016deep,srivastava2015training}.\\ As depicted on Fig.~\ref{fig:Filtre}, each atrous convolutional layer is followed by a layer-normalization layer (LN)\cite{ba2016layer}, which allows to compute layer-wise statistics and to normalize the nonlinear tanh activation accross all summed inputs within the layer, instead of within the batch with a standard batch-normalization\cite{ioffe2015batch, ioffe2017batch} process. The layer-normalization approach has the great advantage of being insensitive to the mini-batchsize\cite{ba2016layer}, and to be usable even if the batch dimension is reduced to one, when the DNN is used for realtime inference. Each LN layer is then followed in this module by a tanh nonlinear activation function, which was selected after extensive comparison with other standard activation functions for the SSL task using the BeamLearning architecture.\\ \subsubsection{Pseudo-energy computation}\ In the following, the set of $N_f$ time-domain outputs $s^{(i)}[n]$ of the filterbank subnetwork will be denoted as $\mathbf{S}_{i,n}$ - where $i$ stands for the filter channel index, and $n$ for the time sample - the bold notation signifying that this a two-dimensional tensor. $\mathbf{S}_{i,n}$ is fed to the second subnetwork of the BeamLearning DNN, which allows to compute a pseudo-energy for each of the $N_f$ pooled channels. As shown in Fig.~\ref{fig:Reseau}, from a machine learning point of view, $\mathbf{S}_{i,n}$ is squared, and cropped in the time dimension -- therefore only keeping the time frames corresponding to the valid part for all the atrous convolutional layers used in the filterbanks subnetwork -- and undergoes an averagepool operation. This pooling operation acts as a dimensionality reduction and has been preferred to standard maxpool operation, since the proposed pseudo-energy can be seen as an equivalent to a mean quadratic pressure, which is similar to the value that is maximized by standard model-based Beamforming algorithms\cite{brandstein2013microphone}. The output of this deterministic pseudo-energy computation is then fed to a LN layer\cite{ba2016layer} in order to normalize the Selu\cite{klambauer2017self} nonlinear activations, which in this case helps to speed up the convergence process.\\ \subsubsection{\label{subsec:output}Matching the pseudo-energy features to the source angular output}\ The last layer of the BeamLearning network is a standard full connected layer, which aims at combining the previously computed pseudo energy features in order to infer the source DoA. In our experiments, the BeamLearning approach is evaluated for 2D DoA determination tasks using either a classification framework or a regression framework. For the classification problem experiments, the BeamLearning DNN output is a n-dimensional vector representing the probabilities of the source belonging to $n$ different angular partitions. For a regression problem, and in agreement with Adavanne et al.\cite{adavanne2018sound} proposal, the output corresponds to the projection of the source position on the unit circle (in 2D) or on the unit sphere (in 3D). This output vector therefore corresponds to the unit vector pointing towards the source DOA. Because of the angular periodicity of the DoA determination problem, this allows an easier retro-propagation process of errors from these projections, than directly from the error of a periodic angle. The angular output DoA $\theta$ (and if needed $ \phi$ in 3D) is then deduced from this projected output.\\ \subsubsection{\label{subsec:trainig} Training procedure}\ Using a classification framework, the BeamLearning DNN is trained with one-hot encoded labels corresponding to the angular area where the source is located, therefore allowing to compute the cross-entropy loss function between estimated labels and ground truth labels. For the regression problem experiments, the BeamLearning DNN is trained with labels corresponding to the projection of the source position on the unit circle, therefore allowing to compute a standard L2-norm loss function between estimated labels and ground truth labels, which represents a geometrical distance between the prediction and the true DOA using cartesian coordinates. \\ The DNN variables updates and the back propagation of errors through the neural network are optimized using the Adaptive Moment Estimation (Adam\cite{kingma2014adam}) algorithm, which performs an exponential moving average of the gradient and the squared gradient, and allows to control the decay rates of these moving averages. In addition to the natural decay of the learning rate that Adam performs during the learning process, we set an exponential learning rate decay. The learning rate $\lambda$ for each step is set to decrease exponentially at each learning iteration $k$, from a maximum value $\lambda_{Max} = 10^{-3}$ to a minimum value $\lambda_{min} = 10^{-6} $ : $$\lambda = \lambda_{min} + (\lambda_{Max} - \lambda_{min}) * e^{-\dfrac{k}{N_{iteration}}}$$ The models have been implemented and tested using the Python Tensorflow\cite{tensorflow2015-whitepaper} open source software library, and computations were carried out on a Nvidia\textsuperscript{\small{\textregistered}} GTX 1080Ti GPU card, using mini-batches of 100 raw multichannel waveforms for each training steps. All the network variables were initialized using a centered truncated normal distribution with a standard deviation of 0.1. Using the proposed datasets, the convergence of the training process requires approximately 150 epochs using a Nvidia\textsuperscript{\small{\textregistered}} GTX 1080Ti GPU card.\\ \section{\label{sec:Evaluation}Evaluation} \subsection{\label{subsec:Datasets}Datasets} In the present paper, we evaluate the performances of the proposed BeamLearning DNN for a 2D DoA SSL task, using experimental and synthetic datasets. Both dataset generation methods have been detailed in subsection \ref{dataset_generation}. In the present section, the source positions around the microphone array used in the datasets are detailed, and the different propagating environments are also presented.\\ In the following, the source positions are denoted using spherical coordinates (R, $\theta, \phi$) centered on the microphone array, where R is the distance to the central microphone, and ($\theta, \phi$) are the azimuthal and elevation angles defined in Fig.~\ref{fig:Pos_src}. In the present study, since we evaluate the BeamLearning approach for a 2D DoA determination task, the $\theta$ coordinate is the only DNN global output. However, in order to increase the robustness to to spatial variations, the source positions are randomly picked from a uniform distribution in a torus of section $ R \times (2 \Delta R) \times \sin(2 \Delta \phi)$ and of central radius $R$, as depicted in Fig.~\ref{fig:Pos_src}. \begin{figure}[ht] \centering \includegraphics[width=0.9\columnwidth]{Fig_03.png} \caption{\label{fig:Pos_src}{Sound sources positioning strategy around the microphone array for the generation of the experimental and synthetic datasets.}} \end{figure} For each synthetic dataset, a set of 48000 sources positions is randomly generated using this procedure. For the experimental dataset, a set of 72000 sources positions has also been randomly generated using the same geometry. The datasets are split into training, validation and test sets in the standard ratios of 80:10:10.\\ In this following, the BeamLearning DNN is trained and tested using synthetic data in two different reverberating environments, and using experimental data in a real room. The first synthetic dataset is generated in a simulated room corresponds to a typical classroom size of $7\times10\times3.7$~m and a reverberation time of 0.5~s. In this case, the microphone array is positioned at coordinates $[4,6,1.5]$~m, the origin of the Cartesian coordinate system corresponding to a corner of the room. The second simulated room exhibits an almost cubic shape, with dimensions $5\times5\times4$~m, and a reverberation time $T_r = 0.5$~s. For both simulated rooms, the room boundaries exhibit homogeneous absorbing coefficients computed using the method proposed by Lehmann et al.\cite{lehmann2007reverberation,lehmann2008prediction} in order to ensure a simulated reverberation time of 0.5~s. In order to also study the localization performances offered by the proposed BeamLearning approach in presence of measurement noise independantly of reverberation, a free field dataset has also been generated for this particular experiment.\\ For the synthetic datasets, a dataset of multichannel RIRs is first computed for each propagating environment using the method detailed in \ref{dataset_generation}. These sets of multichannel RIRs are then convoluted with several kinds of signals emitted from the position of the simulated sound sources. This common method\cite{vera2018towards, perotin2019crnn, Adavanne2019_DCASE} allows a great flexibility in terms of emitted signals in a memory-efficient way. In the present study, the signals used to generate the synthetic datasets using the computed RIRs are composed of anechoic recordings of symphonic music\cite{lokki2008recording} (\url{https://users.aalto.fi/~ktlokki/Sinfrec/sinfrec.html}), outside recordings of multiple women talking in Danish (\url{https://odeon.dk/downloads/anechoic-recordings/}), and car horn honks from the UrbanSound8K dataset\cite{Salamon:UrbanSound:ACMMM:14}.\\ During the training phase, each multichannel recordings of the training mini batches are added to a random multichannel Gaussian noise with randomly picked signal-to-noise ratio (SNR) values in the range of $[x,+\infty[$ dB. This process can be interpreted as a data augmentation strategy, which allows to ease the generalization capabilities of the proposed method during the learning process. In the following, the presented results will include data augmentation strategies during the training phase with $x = 5$~dB (section \ref{sec:Results}), $x = 20$~dB (section \ref{sec:Output}), $x = 25$~dB (section \ref{subsec:Robustness_noise}), and $x = +\infty$~dB (without any data augmentation, section \ref{subsec:Robustness_noise}). The proposed data augmentation strategy will be shown to offer an increase of robustness to measurement noise in comparison to state of the art SSL algorithms.\\ In order to evaluate the 2D-DoA SSL performances of the BeamLearning approach and to compare it in a reproducible way to model-based algorithms (MUSIC\cite{schmidt1986multiple} and SRP-PHAT\cite{dibiase2001robust}), each method is tested in section \ref{sec:Results} using the exact same testing datasets. These testing datasets consist in 3600 sources, randomly distributed in the torus depicted on Fig~\ref{fig:Pos_src}, emitting unseen signals during the training phase. In order to evaluate the robustness of each methods to noisy measurements in reverberating environments, the statistics of the DoA mismatch are derived from these 3600 positions, for each proposed propagating environments, and for different deterministic (SNR) values, ranging from $+40$~dB to $-1$~dB.\\ \subsection{\label{subsec:eval}Evaluation metrics} In the present paper, the BeamLearning approach is evaluated both in a classification framework and in a regression framework. As a consequence, in order to analyze precisely the performances of the proposed DNN for the task of 2D-DoA SSL, several evaluation metrics will be used in the following. For the task of supervised multi-class angular classification, the overall accuracy is computed during the training phase and for testing using the number of correctly recognized class examples (true positives, $t_{p_i}$ ), the number of correctly recognized examples that do not belong to the class (true negatives, $t_{n_i}$ ), and examples that either were incorrectly assigned to the class (false positives, $f_{p_i}$) or that were not recognized as class examples (false negatives, $f_{n_i}$)\cite{sokolova2009systematic}. Other angular mismatch estimators will also be detailed in section \ref{subsec:number_classes}. For the task of supervised angular regression, the DoA mismatch between the estimated angular positions $\tilde{\theta(k)}$ and the groundtruth angular positions $\theta(k)$ is evaluated for each sources using statistics derived from the absolute angular mismatch $\vert \tilde{\theta(k)} - \theta(k)\vert$ over all the sources in the testing dataset. \section{BeamLearning network analysis \label{sec:Output}} \subsection{\label{subsec:filterbank_analysis}Learnable filterbanks inner working analysis} In this subsection, we analyse the variables learnt in the first subnetwork described in \ref{subsec:filterbank_subnetwork}, in order to give further insight on the learning process involved. The architecture of this first subnetwork of the BeamLearning DNN has been specifically developed to automatically build a 2D map $\mathbf{S}_{i,n}$ that can be interpreted as the filtering of raw input multichannel measurements by 128 different tunable filters. On contrary to conventional Beamforming, these 128 time domain outputs cannot be interpreted as beams, but rather as 128 optimized channels which aim at achieving a joint filtering in space and time. Since this subnetwork is composed of 3 filterbanks constituted by 6 depthwise separable atrous convolutions with a channel multiplier of 4 and a number of output channels of 128, it is important to note that the global output of this first subnet results from the combination of nonlinear filtering achieved by the equivalent $3 \times 6 \times 4 \times 128 = 9216 $ learnable filters followed by mathematical operations operated by the LN layer, and their nonlinear activation functions. Rather than looking at each of these filters independantly, this analysis aims at giving insightful representations of the spatio-temporal filtering achieved when the multichannel input data has partially passed through the $n$ first layers of the learnable filterbank subnetwork. This analysis will allow to better understand the multiresolution offered by the residual networks of one-dimensional depthwise separable atrous convolutions that constitute the main ingredient of this subnetwork.\\ \begin{figure*}[ht] \centering \subfloat[\label{fig:Directivite_1_1}]{% \includegraphics[width=0.333\linewidth]{Fig_04_a.pdf}} \hfill \subfloat[\label{fig:Directivite_1_8}]{% \includegraphics[width=0.333\linewidth]{Fig_04_b.pdf}} \hfill \subfloat[\label{fig:Directivite_1_32}]{% \includegraphics[width=0.333\linewidth]{Fig_04_c.pdf}} \caption{\label{fig:Directivite}Directivity maps (in dB) of one of the 128 equivalent filters obtained by passing through the first layers of the network, for the 1\textsuperscript{st} (a), 4\textsuperscript{th} (b), and 6\textsuperscript{th} (c) depthwise-separable atrous convolutional layers of the filterbank subnetwork. These 3 layers correspond respectively to a dilatation factor of 1, 8 and 32.} \end{figure*} In order to perform this analysis, the BeamLearning network is trained for a regression problem using the synthetic dataset described in \ref{subsec:Datasets}. This training dataset, will be referred in the following as $\mathcal{D}_{\text{train}}^{\text{ Room1}}$. In this dataset, each of the sources located in a room of dimensions $7\times10\times3.7$~m and a reverberation time of 0.5~s. emit ambient speech noise signals, car horn signals, and symphonic music. The data augmentation procedure detailed with added measurement noise detailed in \ref{subsec:Datasets} is used with a $SNR \geq 20$ dB. Following the procedure detailed in \ref{subsec:Datasets}, the 48000 sources positions included in $\mathcal{D}_{\text{train}}^{\text{ Room1}}$ are randomly distributed in a torus of section $ R \times (2 \Delta R) \times \sin(2 \Delta \phi)$ and of central radius $R=2$~m, and with $\Delta R = 0.5$~m and $\Delta \phi = 7$ degrees (see Fig~\ref{fig:Pos_src}). All the parameters of the BeamLearning network are then frozen after convergence of the learning process. This frozen BeamLearning network is then used to build 2 dimensional $(\theta,f)$ magnitude maps of the $128\times18$ spatio-temporal filterings offered at the output of the 18 successive depthwise-separable atrous convolutional layers used in the learnable filterbank subnetwork. Among these 2304 possible $(\theta,f)$ transfer function magnitude maps offered at increasing depths of the subnetwork, Fig.~\ref{fig:Directivite} shows three of them. These three maps are selected from the first filterbank, and corresponding to dilatation factors of 1, 8, and 32 in order to illustrate the multiresolution capabilities offered by the BeamLearning approach.\\ The representations in Fig.~\ref{fig:Directivite} are obtained using unseen input data during the DNN training. These unseen testing input data correspond to raw measurements on the microphone array using 360 sources uniformly sampled on a circle of radius 2~m at $\phi = \pi/2$, emitting monochromatic signals ranging from 100~Hz to 4000~Hz. This allows to compute the magnitude of the filtering process for each frequency and for each azimuthal angle at any of the layer composing the learnable filterbank subnetwork (including nonlinear activations and normalization layers), which is a strict equivalent of representing the directivity maps obtained at different depths of the subnetwork.\\ When analysing the obtained directivity maps on Fig.~\ref{fig:Directivite}, one can observe that the frequency and spatial selectivity of the learnt filters increase with the depth through which the input data passes into the DNN. For the very first layer of the filterbank subnetwork corresponding to Fig.~\ref{fig:Directivite_1_1}, the obtained directivity obtained for this particular filter corresponds to a dipolar response above 1000~Hz, with directivity notches at $\theta = 150$~degree and $\theta = 330$~degree. At low frequencies, this particular filter (out of the 128 possible filters for this layer) exhibits a global attenuation of -20~dB, without any strong directivity pattern. It is also interesting to note that among the 128 filters obtained for this particular first convolutional layer (data not shown), most filters exhibit such a dipolar angular response at high frequencies, with various spatial notches values. This behaviour can be explained by the fact that for this layer, the filter kernels have a very small size of 3 samples, with a dilation factor of 1. \\ However, when the data passes through more depthwise-separable atrous convolutional layers in the filterbank (Fig.~\ref{fig:Directivite_1_8} and Fig.~\ref{fig:Directivite_1_32}), the combination of optimized filters begin to be more and more selective, both in the frequency domain and in the angular domain. This is even observed at low frequencies, thanks to the kernel widthening offered by the increased dilation factors and the combination of filters at previous layers offered by the residual connections. This observation suggests that the stacking of the layers composing the three successive filterbanks converge to a global nonlinear filtering process into 128 channels, where each output channel specializes in several frequency bands and several angle incidences. When the source signal exhibits a particular spectrum in a reverberating environment, the energies carried by these 128 channels are then combined by the full connected layer in order to infer the source position. \subsection{\label{subsec:classif_directivity} BeamLearning output analysis for a classification problem} Most published studies on SSL using a deep learning approach treat the localization problem as an angular classification problem, where the objective is to determine to which discrete portion of the angular space the emitting source belongs to\cite{perotin2019crnn, chakrabarty2017broadband, he2018deep, takeda2016sound}. When seeking a better angular resolution for the SSL problem, these approaches require to increase the number of classes. \\ In this n-class classification approach, the source angular position $\theta \in [-180,180]$ is first converted to an integer $i$ representing the belonging of the source to the i\textsuperscript{th} angular partition among the n possibilities : $$i = \lfloor (\theta + 180) \frac{n}{360} \rfloor $$ Using this approach, $\theta$ is then approximated to a corresponding discrete $\theta_i$ (the center of the i\textsuperscript{th} angular partition), which is equivalent at discretizing the space on an angular grid : $$\theta_i = (i+\frac{1}{2}) \times \frac{360}{n} - 180 $$ The label used for classification can then be converted using a one-hot encoded target strategy using the integer $i$\cite{perotin2019regression,chakrabarty2017broadband}. Some authors also proposed the use of a soft Gibbs target\cite{perotin2019regression,he2018deep} exploiting the absolute angular distance between the true angular position $\theta$ and the attributed discrete value $\theta_i$ on the grid.\\ In the following, we study the behaviour of the DNN output neurons for a 8-class classification problem. The analysis of the BeamLearning DNN outputs for a regression problem will be carried out similarily in \ref{subsec:regression_directivity}. For the classification problem, the BeamLearning DNN is trained with a cross-entropy loss function preceded by a softmax activation function and with one-hot encoded labels. Using the same training dataset $\mathcal{D}_{\text{train}}^{\text{ Room1}}$ as in \ref{subsec:filterbank_analysis} with data augmentation procedure with $SNR \geq 20$~dB and after convergence of the BeamLearning network, the frozen DNN is fed with unseen input data during the DNN training. These unseen testing inputs corresponds to 4800 sources randomly distributed over a torus of mean radius $R=2$~m, emitting unseen speech signals. This testing dataset, which will be used several times throughout the present paper, will be referred in the following as $\mathcal{D}_{\text{test speech}}^{\text{Room1}}$. The use of such a testing dataset allows to represent quantitatively the probability value obtained using the softmax function applied to the estimated multiclass label for each of the 8 directions as a polar representation, which can be interpreted as a directivity diagram (see Fig.\ref{fig:Directivite_classif}).\\ \begin{figure}[ht] \centering \subfloat[]{% \includegraphics[width=0.333\linewidth]{Fig_05_a.pdf}} \hfill \subfloat[]{% \includegraphics[width=0.333\linewidth]{Fig_05_b.pdf}} \hfill \subfloat[]{% \includegraphics[width=0.333\linewidth]{Fig_05_c.pdf}} \caption{\label{fig:Directivite_classif}{Directivity diagrams of 3 of the 8 output neurons for the SSL task seen as a classification problem. The diagrams are obtained by plotting the output probability (solid blue) obtained for each output neuron when testing the BeamLearning network with the 4800 sources positions in $\mathcal{D}_{\text{test speech}}^{\text{Room1}}$. The sources positions where the predicted output is the i\textsuperscript{th} angular partition are depicted using light blue shaded areas.}} \end{figure} The analysis of the obtained directivity diagrams on Fig.~\ref{fig:Directivite_classif} allows to highlight the fact that during learning, each output neuron specializes to its corresponding angular partition, with a sharp directivity pattern. However, a precise analysis of these directivity pattern shows that some of these directivity patterns slightly overlap adjacent angular partitions that define the angular classes used for the SSL problem. As a consequence, a loss of accuracy can be observed for sound sources located in the vicinity of angular partitions. This also clearly shows that the misclassified sources mostly belong to contiguous angular sectors to the estimated angular sector. This suggest that for a rather coarse angular grid, the use of a soft Gibbs target or a Gibbs-weighted cross-entropy loss such as proposed in\cite{perotin2019regression,he2018deep} may only offer small accuracy improvements for the proposed BeamLearning approach over the simple one-hot encoded labelling strategy we use. However, for finer angular grids and with a higher number of angular partitions, the number of boundaries between classes increase, which could lead to a loss of global classification accuracy, which may justify the use of a regression approach. \subsection{\label{subsec:number_classes}Influence of the number of angular classes} Since an angular classification problem is directly linked to the choice of a discrete angular grid to define angular partitions, the present subsection aims at giving further insight on the influence of the number of classes when aiming at obtaining a better angular resolution for the SSL problem. As observed in the previous section, for a small number of classes, most misclassified source positions are located in the vicinity of the chosen angular partitions. In order to verify if this assumption still stands for a high count of classes, we conducted several n-class classification experiments, with $n$ ranging from 8 to 128. Those experiments have been conducted both with the experimental dataset $\mathcal{D}_{}^{\text{exp}}$ detailed in \ref{subsec:Datasets} with $\Delta R = 0.5~m$ and $\Delta \phi = 7^{\circ}$ partitionned in $\mathcal{D}_{\text{train}}^{\text{exp}}$ and $\mathcal{D}_{\text{test}}^{\text{exp}}$, and the same training and testing synthetic datasets $\mathcal{D}_{\text{train}}^{\text{ Room1}}$ and $\mathcal{D}_{\text{test speech}}^{\text{Room1}}$ than those used in \ref{subsec:filterbank_analysis}. Though, it is interesting to note that for the experimental dataset $\mathcal{D}_{}^{\text{exp}}$, the BeamLearning network has been trained using 72000 acoustic sources generated using the laboratory's higher order ambisonics sound field synthesis system\cite{lecomte2016fifty}, each source emitting a simple set of 6 pure sinusoidal signals, whose frequencies correspond to the central frequencies of the octave bands between 125 Hz and 4000 Hz. Since the sound field synthesis system is located in a rather moderately reverberating environment ($T_R \approx 0.2$~s) and since the frequency content and the dynamics of those emitted signals are much less diverse than in $\mathcal{D}_{\text{train}}^{\text{ Room1}}$, this represents an easier learning task than the one achieved using the synthetic dataset.\\ For both training datasets and for each number of classes $n$, after convergence of the learning process, the frozen Beamlearning DNNs are fed with unseen input data during the training phase. These unseen testing inputs corresponds to 4800 sources randomly distributed over a torus of mean radius $R=2$~m around the microphone array emitting unseen signals during the DNN training phase. Table~\ref{tab:Prec_classif} presents the obtained performances in terms of multi-class accuracy and in terms of angular accuracy. Since the SSL determination is treated here as a classification problem, two kinds of angular accuracies can be calculated, which both carry some insightful information about the localization performances of the algorithm : $$ \Delta \theta_{class} = \overline{\vert \bm{\tilde{\theta_i}} - \bm{\theta_i} \vert} $$ $$ \Delta \theta = \overline{\vert \bm{\tilde{\theta_i}} - \bm{\theta} \vert} $$ , where $\bm{\theta}$ is the set of continuous ground truth positions of the 4800 testing sources, $\bm{\theta_i}$ is the set of corresponding groundtruth grid positions defined in \ref{subsec:classif_directivity}, and $\bm{\tilde{\theta_i}}$ is the set of discrete grid positions estimated by the BeamLearning network.\\ \renewcommand{\arraystretch}{1.5} \begin{table}[ht] \caption{Multiclass classification performances (accuracy, and mean absolute angular errors) with increasing number $n$ of classes. Results are presented both for an experimental dataset and a synthetic dataset in a reverberating environment, with a $SNR = 20$~dB. The testing dataset is composed of 4800 positions randomly distributed around the microphone array. Results with both $\Delta \theta \leq 5^{\circ}$ and $\frac{\Delta \theta}{\alpha} \leq 1$ are in bold.} \begin{tabularx}{\linewidth}{|X|X||X|X|X|X|X|} \hline\hline \multicolumn{2}{|c||}{$n$} & 8 & 16 & 32 & 64 & 128 \\ \hline \multicolumn{2}{|c||}{Half angular width $\alpha$} & $22.5^{\circ}$ & $11.25^{\circ}$ & $5.63^{\circ}$ & $2.81^{\circ}$ & $1.4^{\circ}$ \\ \hline \hline & Accuracy & 99.2\% & 97.1\%& \textbf{94.3\%} & \textbf{91.3\%} & \textbf{87.7\%} \\ \raisebox{-.5\normalbaselineskip}[0pt][0pt]{\rotatebox[origin=c]{90}{\parbox{0.8cm}{\centering{Experim. dataset}}}} & $\Delta \theta_{class}$ & $0.3^{\circ}$ & $0.6^{\circ}$ & $\mathbf{0.6^{\circ}}$ & $\mathbf{0.5^{\circ}}$ & $\mathbf{0.4^{\circ}}$ \\ & $\Delta \theta$ & $11.1^{\circ}$ & $5.8^{\circ}$ & $\mathbf{3.0^{\circ}}$ & $\mathbf{1.5^{\circ}}$ & $\mathbf{0.9^{\circ}}$ \\ \hline & Accuracy & 92.1\% & 87.6\%& \textbf{79.5\%} & 61.4\% & 37.9\% \\ \raisebox{-.5\normalbaselineskip}[0pt][0pt]{\rotatebox[origin=c]{90}{\parbox{0.8cm}{\centering{Synthetic dataset}}}} & $\Delta \theta_{class}$ & $5.9^{\circ}$ & $5.1^{\circ}$ & $\mathbf{3.0^{\circ}}$ & $4.1^{\circ}$ & $4.7^{\circ}$ \\ & $\Delta \theta$ & $14.9^{\circ}$ & $8.8^{\circ}$ & $\mathbf{4.3^{\circ}}$ & $4.5^{\circ}$ & $4.8^{\circ}$ \\ \hline\hline \end{tabularx} \label{tab:Prec_classif} \end{table} The analysis of the results in Table~\ref{tab:Prec_classif} allow to show that even with a rather low classification accuracy for a high count of angular classes, the mean absolute angular errors on the grid $\Delta \theta_{class}$ remain very low and almost constant across the number of classes. This suggests that for a high count of angular classes, the growing number misclassified sources are attributed to grid points that remain close to their ground truth positions. Since the angular grid refines when $n$ increases, the observed decrease of classification accuracy is therefore compensated by the increasing of the angular resolution offered by a higher count of angular classes. However, there is a clear influence of this grid resolution when estimating $\Delta \theta$, which also includes the uncertainty due to the grid resolution. Indeed, the computation of $\Delta \theta$ also includes the absolute angular error for sources that are classified in the right angular partition. One can observe that for the synthetic dataset in a reverberating environment, $\Delta \theta$ remains smaller than the half angular width $\alpha$ of the angular partitions for $n \leq 32$. However, for $n=64$ and $n=128$, the mean absolute angular error $\Delta \theta$ exceed $\alpha$, which signifies that for such a angular resolution, a regression approach could be more appropriate.\\ \begin{table}[ht] \caption{Statistics over the misclassified populations $\Omega_m$ with increasing number of classes $n$ (percentage of misclassified sources , and mean absolute angular errors) with increasing number $n$ of classes. Results are presented both for an experimental dataset and a synthetic dataset in a reverberating environment, with a $SNR = 20$~dB. The testing dataset is composed of 4800 positions randomly distributed around the microphone array. Results with both $P_{adj} \geq 85 \%$ and $\frac{\beta}{\alpha} \leq 1$ are in bold.} \begin{tabularx}{\linewidth}{|X|X||X|X|X|X|X|} \hline\hline \multicolumn{2}{|c||}{$n$} & 8 & 16 & 32 & 64 & 128 \\ \hline \multicolumn{2}{|c||}{Half angular width $\alpha$} & $22.5^{\circ}$ & $11.25^{\circ}$ & $5.63^{\circ}$ & $2.81^{\circ}$ & $1.4^{\circ}$ \\ \hline \hline & Card($\Omega_m$) & \textbf{34} & \textbf{139} & \textbf{273} & \textbf{417} & \textbf{591} \\ \raisebox{-.5\normalbaselineskip}[0pt][0pt]{\rotatebox[origin=c]{90}{\parbox{0.8cm}{\centering{Experim. dataset}}}} & $P_{adj}$ & $ \mathbf{100 \%}$ & $ \mathbf{100 \%}$ &$ \mathbf{100 \%}$ &$ \mathbf{100 \%}$ & $\mathbf{100 \%}$ \\ & $\beta$ & $\mathbf{0.01^{\circ}}$ & $\mathbf{0.03^{\circ}}$ & $\mathbf{0.09^{\circ}}$ & $\mathbf{0.17^{\circ}}$ & $\mathbf{0.37^{\circ}}$ \\ \hline & Card($\Omega_m$) & \textbf{379} & \textbf{595} & \textbf{984} & \textbf{1853} & 2981 \\ \raisebox{-.5\normalbaselineskip}[0pt][0pt]{\rotatebox[origin=c]{90}{\parbox{0.8cm}{\centering{Synthetic dataset}}}} & $P_{adj}$ & $ \mathbf{88.9 \%} $ & $ \mathbf{82.2 \%} $ & $ \mathbf{94.6 \%}$ & $ \mathbf{88.1 \%} $ & $ 57.8 \%$ \\ & $\beta$ & $\mathbf{1.8^{\circ}}$ & $\mathbf{2.0^{\circ}}$ & $\mathbf{1.0^{\circ}}$ & $\mathbf{2.4^{\circ}}$ & $3.0^{\circ}$ \\ \hline\hline \end{tabularx} \label{tab:Stats_errors_classif} \end{table} It is also interesting to study the spatial repartition of the misclassified sources. Table~\ref{tab:Stats_errors_classif} presents two insightful statistics over the population $\Omega_{m}$ of misclassified sources. $P_{adj}$ represents the percentage of $\Omega_{m}$ elements whose groundtruth position is located in an contiguous angular partition to the estimated angular partition. $\beta$ represents the mean absolute angular distance between the ground truth position of misclassified sources and the boundary of the mis-estimated angular partition.\\ The analysis of these metrics -- and in particular of $P_{adj}$ -- allows to conclude that even with a high count of classes and a refined discrete grid of possible angles, the great majority of misclassified sources mostly belong to contiguous angular sectors to the estimated angular sector. The only situation where misclassified sources begin to spread on non contiguous angular partitions corresponds to a 128-class classification problem in a reverberating and noisy environment. Using the proposed BeamLearning network, this observation again advocates for the limited interest of the use of a proximity-aware Gibbs weighing strategy proposed in\cite{perotin2019regression,he2018deep} for labelling the sound sources positions. The proposed BeamLearning network is only trained for classification with one-hot encoded labels that do not take into account any proximity weighting. This result suggests that the obtained representation at the output of the learnable filterbanks and the pseudo-energy network already encodes efficiently this information. \\ Furthermore, the analysis of $\beta$ shows that even with a high count of classes, the misclassified sources even remain very close to the boundary of the estimated angular section (less than $3^{\circ}$ in all cases) . Except for $n=64$ and $n=128$ classification problems in a reverberating and noisy environment, the mean distance to the boundary remains smaller than the quarter width of the angular sectors. However, when seeking an angular resolution of less than $5^{\circ}$ (corresponding to $n > 32$), these results suggest that a regression approach could be more appropriate and less dependent to the choice of the angular partitions, which confirms the results obtained in\cite{perotin2019regression} for a similar problem in 3D, but with a completely different DNN architecture and different kind of input signals representation. \subsection{\label{subsec:regression_directivity} BeamLearning output analysis for a regression problem} In this subsection, the BeamLearning DNN is used for DoA determination, in a regression scenario. As explained in sections \ref{subsec:output} and \ref{subsec:trainig}, in this case, the sound source angular position is inferred from the output vector of the network, which represents the estimated cartesian coordinates of the unit vector pointing to the source DoA. The whole BeamLearning DNN architecture remains unchanged with respect to the classification problem, except for the last layer and the cost function described in section \ref{subsec:trainig}. In a regression framework though, the DNN output cannot be used to derive directivity patterns as it was in \ref{subsec:classif_directivity}, since the output is directly linked to the continuous DoA space. However, a polar representation of the absolute angular errors committed by the DNN over a large testing dataset (see Fig.~\ref{fig:Err_circ}) allows to represent the SSL performances of the proposed approach in a compact way.\\ \begin{figure}[ht] \centering \includegraphics[width=0.75\columnwidth]{Fig_06.pdf} \caption{\label{fig:Err_circ} Angular diagram of statistical performance metrics derived from the absolute angular mismatches obtained by the BeamLearning DNN using an unseen testing dataset $\mathcal{D}_{\text{test speech}}^{\text{Room1}}$ with a fixed $SNR = 20$~dB. Global median over the whole dataset (dashed black line), local sliding median over an angular window of 5 degrees (solid blue line), and local interquartile range (light blue shaded area).} \end{figure} The polar diagram on Fig.~\ref{fig:Err_circ} has been obtained by freezing the BeamLearning network after convergence of the training process for a DoA regression task in a reverberating environment, trained with a variable $SNR \geq 20$ dB. The training dataset used for this task is the same synthetic dataset $\mathcal{D}_{\text{train}}^{\text{ Room1}}$ as in \ref{subsec:filterbank_analysis}, \ref{subsec:classif_directivity}, and \ref{subsec:number_classes}. The frozen BeamLearning DNN is then fed with input data $\mathcal{D}_{\text{test speech}}^{\text{Room1}}$ (unseen during the training phase), with a fixed SNR~$= 20$~dB. For each of the 4800 testing positions, 50 different draws of random noise have been added to the microphone measurements. Using the $4800\times50$ obtained DoA estimations, statistics of the absolute angular errors are computed, including the global median over $\mathcal{D}_{\text{test speech}}^{\text{Room1}}$ ($1.9^{\circ}$), the local median computed using a sliding average over a window of 5 degrees (ranging from $1.5^{\circ}$ to $2.2^{\circ}$), and the local interquartile range (ranging from $0.7^{\circ}$ to $3.9^{\circ}$) which represents the confidence interval of the DoA mismatch estimations. Since this experiment uses the same training and testing datasets as those used for classification experiments in previous subsections, this allows to compare the obtained performances for this particular 2D DoA task. The obtained results show that the proposed regression approach with the BeamLearning DNN is more accurate than any of the classification test cases, even with a high count of classes. As a consequence, in the following, all the results and experiments will be conducted using a regression framework. \section{Results and discussion using a regression approach \label{sec:Results}} In this section, the SSL performances of the BeamLearning approach will be extensively studied and systematically compared with the performances of two state-of-the-art model-based SSL algorithms (MUSIC and SRP-PHAT) for a 2D DoA determination problem. This choice has been made using extensive preliminary comparisons between the TOPS, CSSM, wideband MUSIC, SRP-PHAT, and FRIDA algorithms included in the Pyroomacoustics library\cite{pyroomacoustics} and by selecting the most robust methods among them. \\ SRP-PHAT\cite{dibiase2001robust} is a widely used algorithm, which is often considered by the scientific community as one of the ideal strategies for SSL systems\cite{do2010srp,badali2009evaluating,cobos2010modified}. This algorithm aims at increasing robustness to signal and ambient conditions by combining the robustness of the steered response power\cite{dmochowski2010steered} (SRP) approach with phase transformation filtering\cite{zhang2008does} (PHAT).\\ The MUSIC algorithm\cite{schmidt1986multiple} and its wideband extensions \cite{argentieri2007broadband} is also a widely used and tested algorithm\cite{argentieri2007broadband,dmochowski2007broadband,ishi2009evaluation,zhang2009robust} for DoA estimation in noisy and reverberant environments, which also makes it an excellent choice for comparative tests. This algorithm, which is able to handle arbitrary geometries and multiple simultaneous sources, belongs to the family of subspace approaches, and depends on the eigen-decomposition of the covariance matrix of the observation vectors. In the present paper, we used the broadband implementation of the MUSIC algorithm proposed in \cite{pyroomacoustics,argentieri2007broadband}, which is obtained by decomposing the wideband signal into multiple frequency bins, and then applying the narrowband algorithm for each frequency bin in order to get the source direction. In the following, for each methods, frequency content within $[100~Hz;4~kHz]$ is used for DOA estimations. This frequency range corresponds approximately to 90 frequency bins for the wideband MUSIC and SRP-PHAT methods.\\ It is also important to note that the MUSIC and SRP-PHAT methods are both based on an angular grid search. Therefore, we used a 1-degree angular grid resolution, which is smaller than the observed interquartile ranges of the three compared SSL methods. Although this affects the computational cost of the two traditional DOA estimation algorithms, it ensures a fair comparison of DOA estimation performance between BeamLearning, MUSIC, and SRP-PHAT.\\ In the following, a particular attention will be paid to the robustness to measurement noise in reverberating environments, to changes in the propagation environment or in the characteristics of the emitted signals. The computational efficiency of the three localization methods will also be compared for realtime 2D-DoA determination. \subsection{Robustness to noisy measurements \label{subsec:Robustness_noise}} In this subsection, we evaluate the SSL performances of the BeamLearning DNN along with those of the MUSIC and SRP-PHAT algorithms, which both have been developed to accomodate to noisy measurement conditions, ranging from sensor ego-noise and acquisition boards electronic noise, to residual acoustic noise due to other interference sources. As explained in \ref{subsec:Datasets}, during the training phase, each multichannel recordings of the training mini batches are added to a random multichannel Gaussian noise with randomly picked signal-to-noise ratio (SNR) values in the range of $[x,+\infty[$~dB. This data augmentation procedure intends at offering an improved robustness to noisy measurements for SSL.\\ \begin{figure}[ht] \includegraphics[width=\columnwidth]{Fig_07.pdf} \caption{\label{fig:Libre}{(Color online) Mean absolute angular mismatch obtained in an anechoïc environment for different SNRs, using an unseen speech testing dataset (3600 positions for each SNR). The DoA errors curves are plotted for wideband MUSIC (dashed), SRP-PHAT (dotted), and the proposed BeamLearning approach (solid blue lines). In order to emphasize the influence of the proposed data augmentation strategy, the BeamLearning DNN has been pre-trained three different times using noisy measurements, with random SNR~$\geq$~5~dB (thick blue line), with random SNR~$\geq$~25~dB (medium blue line), and without any noise added (thin blue line). The interquartile range of absolute angular mismatch is also represented for BeamLearning trained with SNR~$\geq$~5~dB (light blue shaded area).}} \end{figure} The BeamLearning DNN has been trained and frozen after convergence for two different environments: a perfectly anechoïc environment, and the same reverberating room than in previous subsections. In the anechoïc environment, the training procedure has been performed with three different minimum SNR values ($x = 5$~dB, $x = 25$~dB, and $x = +\infty$~dB, \emph{i.e.} noiseless data) for data augmentation. In the reverberating environment, results are presented with a data augmentation procedure with $x = 5$~dB. This allows to observe the robustness to measurement noise only (see Fig.~\ref{fig:Libre}), and to demonstrate the capabilities of the proposed approach in both noisy and reverberating environments (see Tab.~\ref{tab:Prec_loca_reverb}).\\ For both environments, the testing datasets used for the evaluation of the MUSIC and SRP-PHAT algorithms and the frozen BeamLearning DNNs are composed of noisy measurements of 3600 sources positions emitting speech signals at a distance of $2\pm0.5$~m from the microphone array, with a varying SNR ranging from -1~dB to $+\infty$~dB. Both the sources positions and the emitted signals were unseen during the training phases of the BeamLearning DNN. The analysis of the obtained results presented on Fig.~\ref{fig:Libre} and Tab.~\ref{tab:Prec_loca_reverb} shows that the proposed BeamLearning approach is significantly more insensitive to measurement noise than MUSIC and SRP-PHAT, both in a free field situation and in presence of reverberation. \\ \newcolumntype{b}{X} \newcolumntype{s}{>{\hsize=.4\hsize}X} \begin{table}[ht] \caption{Mean absolute angular mismatch obtained in a reverberating environment (RT~=~0.5~s), for different SNRs, using an unseen speech testing dataset (3600 positions). The DNN has been pre-trained using noisy measurements, with random SNR~$\geq$~5~dB. The best results are in bold.} \begin{tabularx}{\linewidth}{|b||s|s|s|s|s|s|} \hline\hline SNR (dB) & -1 & 2 & 5 & 10 & 15 & 20 \\ \hline SRP-PHAT & $41.6^{\circ}$ & $28.7^{\circ}$ & $17.4^{\circ}$ & $7.9^{\circ}$ & $4.6^{\circ}$ & $3.4^{\circ}$ \\ \hline MUSIC & $27.5^{\circ}$ & $21.3^{\circ}$ & $17.8^{\circ}$ & $13.1^{\circ}$ & $11.0^{\circ}$ & $10.9^{\circ}$ \\ \hline BeamLearning & $\bm{21.1}^{\circ}$ & $\bm{13.0}^{\circ}$ & $\bm{8.2}^{\circ}$ & $\bm{4.8}^{\circ}$ & $\bm{3.4}^{\circ}$ & $\bm{2.7}^{\circ}$ \\ \hline\hline \end{tabularx} \label{tab:Prec_loca_reverb} \end{table} A comparison of the results obtained with the BeamLearning approach on Fig.~\ref{fig:Libre} allows to conclude that the proposed data augmentation process offers a very efficient way to improve SSL estimations. Over all the proposed DOA estimation methods in Fig.~\ref{fig:Libre}, the best results are obtained using the BeamLearning approach trained using a data augmentation process with a random SNR~$\geq~5$~dB. This result still stands when the testing dataset is used with SNR as small as $-1$~dB. For a moderate noise data augmentation (a random SNR training larger than 25~dB), the BeamLearning approach offer very similar SSL performances than both MUSIC and SRP-PHAT for unseen testing data with a SNR~$\geq~10$~dB. Without any data augmentation, the BeamLearning approach only outperforms MUSIC and SRP-PHAT for unseen testing data with a SNR~$\geq~30$~dB. It is also interesting to note that the use of the proposed data augmentation mechanism does not degrade the localization accuracy at high SNR, even if the data augmentation uses very noisy data. This property is directly linked to the fact that the training process is performed with randomly picked signal-to-noise ratio (SNR) values in the range of $[x,+\infty[$ dB, which does not give more importance to noisy data than to noise-free data when training the BeamLearning DNN.\\ In free-field (Fig.~\ref{fig:Libre}) and for low-noise measurements, the three methods offer very similar SSL performances. However, as expected, the MUSIC algorithm performs significantly better than SRP-PHAT for heavily noisy measurements (SNR~$\leq$~5~dB). It is although interesting to note that the proposed BeamLearning approach performs even better than MUSIC at low SNRs in free field. Since the mean absolute angular errors obtained using wideband MUSIC and SRP-PHAT stand outside the interquartile range of the BeamLearning approach trained with a random SNR~$\geq~5$~dB, one can also conclude that the improvement offered by the proposed BeamLearning approach is statistically significant over those two methods at low SNR : the DOA estimation precision is improved over MUSIC by 30\% (from 10.0\textsuperscript{o} to 7.0\textsuperscript{o} at -1~dB).\\ The same trend is also observed in a reverberating environment (Tab.~\ref{tab:Prec_loca_reverb}), where the BeamLearning approach offers better angular estimations for all tested SNRs. The improvements in terms of SSL accuracies range from 23\% at -1 dB to 75\% at 20~dB when compared to MUSIC, and from 49\% at -1dB to 21\% at 20~dB when compared to SRP-PHAT. It is also interesting to note that for low-noise measurements (SNR~$\geq$~15 dB) in this reverberating environment, the SSL performances offered by MUSIC are degraded, since this subspace-based algorithm relies on the assumption that signal paths are uncorrelated or can be fully decorrelated, which does not hold true with early reflections in room acoustics, leaving DoA estimation a difficult problem in these situations for the wideband MUSIC method. \subsection{Robustness to mismatch between the propagating environments in the learning and testing phases} One of the main common pitfalls when using machine learning approaches resides in overfitting training data\cite{roelofs2019meta}. In order to avoid this phenomenon, the authors in\cite{perotin2019crnn} proposed the use of a training dataset composed by a large number of randomly sized rooms, which is an excellent way of circumventing these potential problems.\\ Still, it is interesting to study the robustness of the BeamLearning DNN, when tested in a totally different propagating environment than the one used for training. Using the same frozen network as in the previous subsection, trained using $\mathcal{D}_{\text{train}}^{\text{ Room1}}$, the frozen BeamLearning DNN is tested using 3600 sources positions in $\mathcal{D}_{\text{test speech}}^{\text{Room2}}$ in a $5 \times 5 \times 4$~m room, and the results are compared to those obtained using MUSIC and SRP-PHAT (see Tab.~\ref{tab:Prec_loca_salle}), for different SNR situations.\\ \begin{table}[ht] \caption{Mean absolute angular mismatch obtained for 3600 sources positions in $\mathcal{D}_{\text{test speech}}^{\text{Room2}}$, for different SNRs. The DNN has been pre-trained using $\mathcal{D}_{\text{train}}^{\text{ Room1}}$, with random SNR~$\geq$5~dB. The best results are in bold.} \begin{tabularx}{\linewidth}{|b||s|s|s|s|s|s|} \hline\hline SNR (dB) & -1 & 2 & 5 & 10 & 15 & 20 \\ \hline SRP-PHAT & $53.7^{\circ}$ & $42.0^{\circ}$ & $30.5^{\circ}$ & $15.5^{\circ}$ & $9.2^{\circ}$ & $\bm{6.6}^{\circ}$ \\ \hline MUSIC & $\bm{41.9}^{\circ}$ & $36.7^{\circ}$ & $30.8^{\circ}$ & $26.9^{\circ}$ & $24.6^{\circ}$ & $25.0^{\circ}$ \\ \hline BeamLearning & $\bm{35.7}^{\circ}$ & $\bm{25.9}^{\circ}$ & $\bm{18.7^{\circ}}$ & $\bm{11.1}^{\circ}$ & $\bm{8.0}^{\circ}$ & $\bm{6.8}^{\circ}$ \\ \hline\hline \end{tabularx} \label{tab:Prec_loca_salle} \end{table} This testing room exhibits a volume reduced by half, when compared to the one used during the training process. However, with a same reverberation time ($T_r = 0.5$~s), the testing room walls are therefore much more reflective, and the positions of the sources are much closer to the room boundaries than during the training process, which is commonly recognized as a difficult task for DoA estimation\cite{shen2020voice}. The degraded performances obtained using MUSIC in this testing room illustrates very well this observation. On the other hand, despite this significant change in the properties of the room between the training dataset and the testing dataset, the BeamLearning approach still offers significantly enhanced performances when compared to MUSIC and SRP-PHAT, except for a high SNR of 20~dB, where SRP-PHAT and BeamLearning allow equivalent SSL accuracies. \subsection{Robustness to change of signal characteristics} Similarly, we also investigate the influence of a mismatch between test signals emitted by the sources and the signals used during the training procedure. We use the same frozen network as in the previous subsection. As detailed in \ref{subsec:Datasets}, the signals emitted by the labelled sources in $\mathcal{D}_{\text{train}}^{\text{ Room1}}$ during the training process are composed of speech signals, symphonic music, and car horn honks. Even if the speech signals used for testing in the previous subsections were unseen during training, it is important to check if the performances remain satisfying when using a signal whose temporal and spectral dynamics are dissimilar to the ones used during the training process. This is the reason why Tab.~\ref{tab:Sig_rob} shows the localization performances obtained for SNRs of 5~dB and 15~dB using sound sources emitting dog bark excerpts from the UrbanSound8K dataset\cite{Salamon:UrbanSound:ACMMM:14}, along with the SSL results obtained with unseen voice signals.\\ The wideband MUSIC method allows a better localization of dog barks when compared to voice signals, mainly due to a smaller bandwidth occupied by the signal subspace when compared to the noise subspace. Still, the BeamLearning approach offers again better localization performances than both SRP-PHAT and wideband MUSIC in this situation. The analysis of these results allows to conclude that in each case, the BeamLearning approach outperforms MUSIC and SRP-PHAT in terms of SSL accuracy, and that the SSL performances observed with the frozen DNN do not degrade significantly when a signal with a different spectral and temporal characteristics is emitted by the sources to be located. \begin{table}[ht] \caption{Mean absolute angular mismatch obtained using several unseen signals in an reverberant environment (RT~=~0.5~s), for different SNRs for 3600 sound sources positions. The DNN has been pre-trained using $\mathcal{D}_{\text{train}}^{\text{ Room1}}$, with random SNR~$\geq$5~dB. The best results are in bold.} \begin{tabularx}{\linewidth}{|s||b|s|s|} \hline\hline SNR & Localization approach & Voice signals & Dog bark \\ \hline & MUSIC & $17.8^{\circ}$ & $\bm{8.9}^{\circ}$ \\ 5 dB & SRP-PHAT & $17.4^{\circ}$ & $16.9^{\circ}$ \\ & BeamLearning & $\bm{8.2}^{\circ}$ & $\bm{6.7}^{\circ}$ \\ \hline & MUSIC & $11.0^{\circ}$ &$6.4^{\circ}$\\ 15 dB & SRP-PHAT & $4.6^{\circ}$ &$6.1^{\circ}$\\ & BeamLearning & $\bm{3.4}^{\circ}$ & $\bm{4.2}^{\circ}$ \\ \hline\hline \end{tabularx} \label{tab:Sig_rob} \end{table} \subsection{Computational efficiency for realtime inference} Since the proposed SSL method is a learning based method, the optimization of the $1.1\times10^6$ variables of the BeamLearning DNN represents a great amount of computational time (approx. 7 days on a Nvidia\textsuperscript{\small{\textregistered}} GTX 1080TI GPU card using the Tensorflow framework). This corresponds to a total of $10^6$ successive iterations for learning (approx. 434 epochs). Each of these learning iterations includes the calculation of the feed forward propagation, cross entropy loss, back-propagation, gradients computations, and variables updates using Adam, for each mini-batch. On this GPU architecture, and for each of the audio excerpts of 23.2 ms (1024 samples), the mean computation time for the learning process is therefore only 6 ms for the whole learning process involved. Given that some amount of these 6 ms are dedicated to the gradient optimizer operations, which are not needed for the inference with a frozen model, this gives confidence for a realtime inference.\\ \begin{table}[ht] \caption{Computation times for 3600 sources localizations} \begin{tabularx}{\linewidth}{|X|X|X|} \hline\hline & Mean time (CPU)& Mean time (GPU)\\ \hline MUSIC & 45 min 40 s & \textit{not implemented} \\ \hline SRP-PHAT & 3 min 01 s & \textit{not implemented} \\ \hline BeamLearning & \textbf{2 min 47 s } & \textbf{16.2 s} \\ \hline\hline \end{tabularx} \label{tab:Calc_time} \end{table} In order to compare the computational time required by the three algorithms for an inference task, we benchmarked the SSL tasks detailed in section \ref{sec:Results} (see Table~\ref{tab:Calc_time}). The BeamLearning DNN has been tested both on a Nvidia\textsuperscript{\small{\textregistered}} GTX 1080TI GPU and a i7-6900K CPU (3.20GHz) with a mini-batch of size 1. On the other hand, MUSIC and SRP-PHAT have only been tested on the same CPU, since this is the only possible implementation proposed by\cite{pyroomacoustics}. In both cases, the BeamLearning approach allows to estimate the position of the sources in a faster way than the model approaches tested under the same conditions, with drastic improvements using a GPU computation. \section{Conclusion} In this paper, we have presented BeamLearning, a new deep learning approach for the localization of acoustic sources. The proposed DNN allows to achieve a 2D-DoA task in real-time from raw multichannel measurements on a microphone array. The proposed DNN architecture is partly inspired by the principle of filter and sum beamforming. In particular, the use of depthwise separable atrous convolutions combined with pointwise convolutions allows to process the input audio data at multiple timescales. The BeamLearning network acts as a joint-feature learning process, that efficiently encodes the spatio-temporal information contained in raw measurements, and accomodates heavy measurement noise situations and reverberation. An extensive analysis of the BeamLearning approach using both a classification framework and a regression framework also allowed to show that the regression approach seems better suited when a fine angular resolution is seeked, with computation times that open up the possibility of precise and real-time localization using a deep learning approach, without any pre-processing of microphone array measurements. The analysis of the BeamLearning approach for SSL tasks in noisy and reverberating environments also proves that BeamLearning outperforms state of the art SRP-PHAT and MUSIC in almost each test situations.\\ In addition to the proposed network architecture, we also proposed the use of a fast and accurate RIR generator on GPU in order to build large realistic training datasets in an efficient way, using an image source method. Using these datasets, the proposed data augmentation process allowed to obtain increased robustness to measurement noise when compared to state of the art model-based SSL algorithms that were specifically crafted to handle this kind of degraded SNR situations. We also demonstrate that even if the propagation environment in which the localization is carried out significantly differs from the one used during the training phase, the BeamLearning approach remains quite robust.\\ The good SSL performances obtained for a 2D-DoA task motivates under development extensions of this work, which include realtime 3D-DoA and simultaneous localization / sound source recognition using the BeamLearning approach. Further developments may include the use of multi-room datasets and multi-source localization. \section*{Acknowledgments} This work has been partially supported through the Deeplomatics project funded by DGA/AID (ANR-18-ASTR-0008). \typeout{}

36e997c0b9f5f0a9cd69ef0f34295f040d39e0b8

\section{Introduction}\label{sec:intro} Modern spintronic materials are being developed for high-speed, non-volatile data storage products such as magnetic random-access memory. An ideal material would be one which has high spin-polarisation, yet little or no magnetic moment. Mn$_2$Ru$_x$Ga (MRG) is a ferrimagnetic inverse Heusler alloy, having an XA structure with two inequivalent antiferromagnetically aligned Mn sublattices \cite{kurt2014} and almost complete filling of the four lattice sites \cite{siewierska2021}, as shown in Figure \ref{fig:crystal}. It is considered to be a zero-moment half-metal (ZHM) when its net moment vanishes at the compensation temperature, $T_{\text{comp}}$, and it is a potential candidate for spintronic applications owing to its immunity to magnetic fields, reasonably high Curie temperature and large perpendicular anisotropy field close to compensation. \hl{The possibility of tunable anti-ferromagnetic oscillation modes in the {\si{\tera\hertz}} region} \cite{troncoso2019} \hl{is key for future applications in high-frequency computing and communications, and has already been observed in the similar Mn$_{3-x}$Ga materials} \cite{awari2016}. Perpendicular magnetic anisotropy (PMA) due to biaxial strain is observed in films grown on MgO. The ruthenium $4d$ site occupancy allows the magnetic properties to be tuned according to the application; in particular, $T_{\text{comp}}$ increases with increasing Ru concentration and valence electron count, $n_v$, in the unit cell \cite{siewierska2021,nivetha2015}. \begin{figure}[h \centering \resizebox{0.7\linewidth}{!}{\input{./crystal}} \caption{Crystal structure of Mn$_2$Ru$_x$Ga is a distorted XA Heusler structure with two inequivalent, antiferromagnetically coupled sublattices, Mn$^{4a}$ and Mn$^{4c}$, leading to zero moment at compensation \cite{kurt2014}} \label{fig:crystal} \end{figure} Annealing of magnetoresistive devices can have varying outcomes - \hl{the crystallisation of the MgO/CoFeB bilayer results in a dramatic increase of tunnelling magnetoresistance (TMR) ratio due to coherent tunnelling and perpendicular anisotropy in the free layer}, while the diffusion of Ta and other species into the active portion of the device can have a negative impact on functionality \cite{kurt2010,Ikeda2008,fukumoto2004,Yang2012}. This MgO/CoFeB electrode structure is currently the industry choice for TMR applications, and annealing temperatures for crystallisation of amorphous CoFeB into a body centered cubic structure typically occurs above \SI{320}{\degreeCelsius}, depending on boron content. The use of thin interlayers to enhance performance of MnGa-based magnetic tunnel junctions has proved effective \cite{ma2012,kubota2011,liang2014}. A major concern with Mn-based materials, however, is diffusion which can affect the both crystal structure as well as the quality of the spin-polariser/barrier interfaces \cite{suzuki2018,mao2017}. We have previously demonstrated that an ultrathin dusting layer of Ta \cite{titova2019} or Al \cite{borisov2016} can be used to mitigate the effects of diffusion between MRG and an MgO tunnel barrier. \hl{Using this technique, a TMR of {\SI{40}{\percent}} has been achieved at {\SI{10}{\kelvin}} and at low bias in a MRG based magnetic tunnel junction using a CoFeB free layer} \cite{borisov2016}. Similarly, we have shown that Hf can magnetically couple MRG and CoFeB thin films, after annealing at high temperatures \cite{borisov2017}. Hf and Al are not known for preventing diffusion, so the mechanism by which they act to maintain performance is not clear, compared to other similar diffusively mobile materials. A previous study of MRG, annealed at various temperatures up to \SI{400}{\degreeCelsius} and capped with AlO$_x$, shows that the magnetic and crystalline properties are stable up to \SI{350}{\degreeCelsius}, beyond which the crystal structure relaxes \cite{kat_2017}. There are several types of defects that occur in the MRG lattice - Mn-Ga ($4a/4c \leftrightarrow 4b$), Mn-Ru ($4a/4c \leftrightarrow 4d$), and Ru-Ga ($4d \leftrightarrow 4b$) antisites - as well as excess Mn and Ga on ${4d}$ sites when $x < 1$ . We can quickly identify Ru-based defects in X-ray crystallography due to the much stronger atomic scattering factor of Ru compared to Mn and Ga. Mn-Ga antisite defects have previously been identified as the source of additional electronic pressure that modifies the position of the Fermi level, increasing the moment and spin polarisation \cite{zic_2016}. Here we correlate magnetic and transport properties with the presence of crystalline defects. \begin{table}[t] \caption{\hl{Spacer material selection}} \begin{tabularx}{\linewidth}{c X} \hline \textbf{Material ($y$)} & \textbf{Reason}\\\hline TiN & Known diffusion barrier; close lattice match with MgO should allow for crystalline growth on MRG at high temperature \\ Hf & Ultrathin layers promote coupling between MRG and CoFeB\\ HfO$_x$ & Used as high-$k$ barrier (gate dielectric)\\ V & A known Al getter\\ Mo & Low affinity for alloying with Mn\\ \end{tabularx} \label{table:spacers} \end{table} To investigate, we use a spin-valve structure of the form MgO//Mn$_2$Ru$_0._7$Ga(\num{30})/X$(t)$/[Co(\num{0.4})/Pt(\num{0.8})]$_8$/Co(\num{0.4})-/Pt(\num{3}) where the spacer layer, X, has a thickness, $t$, of \SIlist{1.4;2}{\nano\metre} (parenthetical values in \si{\nano\metre}). It will act as a barrier preventing diffusion, or as a source of mobile atoms. The stack uses a Co/Pt superlattice as the free layer due to its perpendicular anisotropy \hl{as deposited, independent of seed layer or annealing requirements as with a MgO/CoFeB free layer}. The five materials selected as X in this study are listed in Table \ref{table:spacers}. TiN thin films are widely used as diffusion barriers and serve as a baseline comparison where we expect no diffusion. Hf has also been chosen to serve as a comparison due to our previous experience with it \cite{borisov2017}. Diffusion-driven defects are expected to result in a reduction of crystalline order in MRG. Any change in $c/a$ ratio will change the magneto-crystalline anisotropy in MRG, while the perpendicular anisotropy in Co/Pt multilayers is primarily caused by interfacial anisotropy driven by spin-orbit hybridisation of Co and Pt\cite{nakajima1998}. This means that a change in the interfacial roughness or strain will have an impact on the Co/Pt magnetic properties \cite{qiu_2016,nakagawa2005}. We measured crystallographic and magnetotransport characteristics before and after annealing at \SI{350}{\degreeCelsius}, a standard industrial practice for MgO/CoFeB-based devices. In this way we aimed to identify the influence of annealing the structure based on the interlayer material used, and its effect on the properties of both electrodes. \section{Methods}\label{sec:method} All films except TiN were deposited by \hl{direct current} sputtering onto \hl{$\SI{10x10}{\milli\metre}$} single crystal MgO \hkl(001) substrates in a SFI Shamrock sputter tool (base pressure \hl{$\SI{< 2.5e-6}{\pascal}$}); TiN was deposited by \hl{radio frequency} sputtering. MRG is deposited at \SI{320}{\degreeCelsius} by co-sputtering from a Mn$_2$Ga target and a Ru target, \hl{both {$\SI{76.2}{\milli\metre}$} diameter, at a target-substrate distance of {$\SI{100}{\milli\metre}$}. Argon gas flow was {$\SI{38}{\cubic\centi\metre\per\minute}$}}. All other materials were deposited at room temperature, except for two spacer layers of TiN grown at \SI{320}{\degreeCelsius}. HfO$_x$ was formed by natural oxidation of metallic Hf films \hl{in O$_2$ atmosphere at {$\SI{2}{\pascal}$} for {$\SI{60}{\second}$}.} Annealing is carried out \hl{under vacuum} at \SI{350}{\degreeCelsius} for 60 minutes in a perpendicular field of \SI{800}{\milli\tesla}. X-ray diffraction \hl{(XRD) and X-ray reflection (XRR) were measured} with a Panalytical X'Pert tool using Cu K$_\alpha$ radiation, and magnetotransport data was collected by the Van der Pauw method in a \SI{1}{\tesla} GMW electromagnet. Crystallographic data were analysed \hl{following the method of }\cite{Nandi1978} by fitting a Voigt function to the peaks, taking into consideration K$\alpha_2$ and instrumental broadening, characterised using an Al$_2$O$_3$ standard, NIST 1976b. \section{Results}\label{sec:results} \begin{figure \resizebox{\linewidth}{!}{\input{./Hf_XRD_in}} \caption{XRD of a spin-valve structure with a \SI{1.4}{\nano\metre} Hf spacer layer, before and after annealing. Peak labels are: $\bullet$ MgO substrate, $\bigtriangleup$ MRG, and $\circ$ CoPt$_3$} \label{fig:Hf_XRD} \end{figure} \begin{figure \resizebox{\linewidth}{!}{\input{./Hf_1_4nm_in}} \caption{AHE of $\SI{1.4}{\nano\metre}$ Hf spacer stack - Co/Pt and MRG layers are able to switch independently. \hl{Remanence (Rem.), coercivity ($H_c$), and saturation (Sat.) are indicated for each layer. Remanence for MRG is given as the remanent state after the Co/Pt layer has switched, and it's coercivity is given as the point in the hysteresis halfway between the remanent and fully saturated states.} Co/Pt has lost some perpendicular anisotropy and appears partially coupled to the MRG after annealing} \label{fig:Hf_EHE} \end{figure} \begin{figure \resizebox{\linewidth}{!}{\input{./xrr}} \caption{XRR and model fitting of a spin-valve structure with a \SI{1.4}{\nano\metre} V spacer layer, (a) before and (b) after annealing. $\ast$ indicates the Co/Pt \hkl(001) superlattice peak position for $d_{Co}+d_{Pt}=\SI{1.19}{\nano\metre}$.} \label{fig:XRR} \end{figure} \begin{figure* \resizebox{\textwidth}{!}{\input{./MRG_combined_in}} \caption{Relative change of tetragonal distortion ($\delta_c$), coherence length ($L_c$), strain ($\epsilon$), intensity ratio ($S$), coercivity ($H_c$), and squareness in MRG after annealing. The grey box indicates a threshold of \SI{10}{\percent} change in squareness as a metric for potential use in applications. The yellow ellipse indicates the best two candidate spacer materials.} \label{fig:MRG_combined} \end{figure*} As-deposited MRG is textured, with the crystallographic $c$-axis perpendicular to the substrate surface. Two diffraction peaks are observed in a symmetric $\theta\text{-}2\theta$ scan, namely the \hkl(002) and \hkl(004) (Figure \ref{fig:Hf_XRD}). In addition to the substrate and MRG peaks, we also note the Co/Pt superlattice \hkl(002) peak. As mentioned in section \ref{sec:intro}, the magnetic properties are strongly dependent on the crystal characteristics. We determine the out-of-plane lattice parameter $c$, crystal coherence length $L_c$ perpendicular to the film (using the Scherrer formula), \hl{the perpendicular strain $\epsilon$,} and the ratio of intensities of the two peaks $S_{\hkl(002)/\hkl(004)}$ signifying the Ru/Mn order as Ru occupying the $4a$ or $4b$ sites will increase the intensity of the \hkl(002) peak. Higher $S_{\hkl(002)/\hkl(004)}$ ratios indicates disorder. By comparison $L_c$ \hl{and $\epsilon$} give us a metric for the long-range disorder in the film, due to grain boundaries and stacking faults, where larger grain size or less strain indicates \hl{a more homogenous film}. MRG has a coercivity in a perpendicular magnetic field that diverges as the temperature approaches $T_{\text{comp}}$, where the anomalous Hall effect (AHE) signal changes sign \cite{nivetha2015}. The composition of MRG used here, $x = 0.7$, was chosen to give $T_{\text{comp}}=\SI{165}{\kelvin}$\hl{ }\cite{siewierska2021}\hl{,} with an as-deposited coercivity of $H_c \approx \SI{380}{\milli\tesla}$ at room temperature. Squareness of the Hall signal, i.e. the ratio of Hall voltage at saturation to that at remanence, is used an indicator of perpendicular anisotropy. The main parameters from the AHE hysteresis are labeled in Figure \ref{fig:Hf_EHE}. The as-deposited MRG exhibits perpendicular magnetic anisotropy (PMA) in all cases, irrespective of spacer material or thickness; magnetic properties upon annealing \textit{are} spacer material dependent. It is typically reported that perpendicular anisotropy in Co/Pt is accompanied by a \hkl(111) texture \cite{lin1991}, however we note here that all as-deposited Co/Pt films exhibited perpendicular anisotropy regardless of \hl{initial texture or phases present}. After annealing with a $\SI{1.4}{\nano\metre}$ Hf spacer, the Co/Pt layer loses much of its perpendicular anisotropy, indicated by the loss of squareness as demonstrated in Figure \ref{fig:Hf_EHE}. This can be attributed to \hl{an increase in average interfacial roughness, from {\SI{0.46}{\nano\metre}} to {\SI{0.50}{\nano\metre}} in Co and from {\SI{0.53}{\nano\metre}} to {\SI{0.99}{\nano\metre}} in Pt. This} reduces the interfacial anisotropy, bringing the moment in-plane, \hl{and is caused by intermixing} as evidenced by the weak crystallisation of \hl{CoPt$_3$} with \hkl(111) texture after annealing, as seen in Figure \ref{fig:Hf_XRD}. \hl{The Co and Pt interfaces of the free layer are well defined before annealing, as can be seen from XRR data for a hetereostructure with a {$\SI{1.4}{\nano\metre}$} V spacer layer in Figure {\ref{fig:XRR}}(a). A clear and intense Co/Pt {\hkl(001)} superlattice peak is visible which disappears upon annealing, shown in Figure {\ref{fig:XRR}}(b), a result of intermixing and alloying between the layers forming a CoPt$_3$ phase as shown in Figure {\ref{fig:Hf_XRD}}. XRD and roughness data for the Co/Pt multilayers is given in Table {\ref{table:copt}}. Of interest is that in each case, the average Co roughness is approximately equivalent to the desired layer thickness ({\SI{0.4}{\nano\metre}}), indicating incomplete coverage.} The anisotropy of MRG appears to be maintained, with minimal change in coercivity or squareness. Changes in Co/Pt properties are considered independent of changes in MRG. As discussed in Section \ref{sec:intro}, the magneto-crystalline anisotropy \hl{of MRG is largely} dependent on the $c/a$ ratio, or more specifically the tetragonal distortion of the cubic cell - $\delta_c=(c-a)/a$. \hl{The properties of MRG are highly dependent on the spacer layer used, and the} relative change of each parameter in MRG after annealing has been collated in Figure \ref{fig:MRG_combined}. Below, we summarise the results for each spacer material. For reference, Table \ref{table:stats} gives a summary of the mean and standard deviation for the various properties measured in MRG before annealing. As a basic metric, we aim to maintain squareness of magnetic hysteresis in MRG to within \SI{90}{\percent} of initial value after annealing \hl{in order to preserve functionality of active devices}. Giant magnetoresistance was not measured in these devices as a current-perpendicular-to-plane geometry would be required to prevent shunting effects of currents through highly resistive layers, necessitating nanoscale lithographic patterning of the metallic films. \begin{table \caption{\hl{Statistics for MRG properties before annealing. $a$ is {\SI{5.96}{\angstrom}}, defined by growth on the substrate}} \begin{tabularx}{\columnwidth}{l X X} \hline & \textbf{Mean} & \textbf{$\sigma$}\\\hline $c$ & \SI{6.0390}{\angstrom} & \SI{0.0108}{\angstrom}\\ $\delta_c$ & \SI{1.326}{\percent} & \SI{0.181}{\percent}\\ $L_c$ & \SI{14.66}{\nano\metre} & \SI{1.93}{\nano\metre}\\ $\epsilon$ & \SI{0.511}{\percent} & \SI{0.056}{\percent}\\ $S_{\hkl(002)/\hkl(004)}$ & 0.0758 & 0.0056\\ $H_c$ & \SI{371.5}{\milli\tesla} & \SI{71.7}{\milli\tesla}\\ Squareness & \SI{92.1}{\percent} & \SI{7.5}{\percent}\\ \end{tabularx} \label{table:stats} \end{table} \subsection{Vanadium}\label{subsec:V} Annealing of a vanadium spacer stack results in large reductions of the MRG coercivity ($\sim \SI{-90}{\percent}$) and squareness ($\sim \SI{-75}{\percent}$), along with a significant ($\sim \SI{-80}{\percent}$) reduction in $\delta_c$. This indicates the anisotropy has a primarily in-plane contribution, though a small fraction of grains are still perpendicular. This is accompanied by an extremely large ($\SI{> 500}{\percent}$) increase of $S$, \hl{as well as a moderate decrease of $L_c$ and increase in $\epsilon$} indicating complete reordering of the crystal structure. Additional peaks near the MRG \hkl(002) and \hkl(004) peaks with $d$-spacing of $\sim$ \SIlist{0.3058;0.1511}{\nano\metre} are present after annealing. These $d$-spacings correspond well to either RuV$\hkl(100) (\SI{0.298}{\nano\metre})$, RuV$\hkl(200) (\SI{0.148}{\nano\metre})$ or Ga$_5$V$_2 \hkl(220) (\SI{0.317}{\nano\metre})$, Ga$_5$V$_2\hkl(530) (\SI{0.1537}{\nano\metre})$. Given the phase diagrams for Ga-V and Ru-V \cite{okamoto_bapd}, we attribute this phase to a Ga-V binary such as Ga$_5$V$_2$ or Ga$_{41}$V$_8$, both of which are peritectics with relatively low decomposition temperatures. Based on the substantial increase in $S$, reordering of Ru from the Heusler $4d$ site occurs forming a highly disordered cubic alloy, with an A2 structure. There is a moderate \hl{CoPt$_3$} \hkl(111) and \hkl(200) texture before annealing. After annealing, only a strong \hl{CoPt$_3$}\hkl(200) peak is present, and magnetic anisotropy is also in-plane. \hl{In addition, the average roughness for Co increases moderately, while the roughness of the Pt layers approximately doubles in both cases.} \subsection{Titanium Nitride}\label{subsec:TiN} When using a TiN spacer, we observe a collapse in the tetragonality of MRG while $L_c$ increases and \hl{$\epsilon$ decreases}, indicating that defects are being removed. This is accompanied by a loss of coercivity and squareness in the magnetic hysteresis. TiN is a well known diffusion barrier, and defects introduced by diffusion are not considered as the cause. We also see that there is no direct correlation with Ru-Mn disorder, as there are three cases with small increase in $S$ contrasted by the \SI{1.4}{\nano\metre} TiN grown at high temperature which has a lower $S$ ratio. \hl{However, there is a correlation between strain $\epsilon$ and chemical ordering $S$.} This means that defects such as stacking faults and grain boundaries induce strain on the crystal structure, which contributes strongly to the magneto-crystalline anisotropy of the film and has a more prominent effect than Ru-Mn anti-sites. The effect is independent of growth temperature. From the Co/Pt superlattice, we observe a very weak Pt\hkl(111) peak, which is distinct from the \hl{CoPt$_3$}\hkl(111), along with a strong \hl{CoPt$_3$}\hkl(200). The high intensity, as well as the Laue fringes observed around the peak indicates that the Pt\hkl(111) is due to the thick Pt capping layer, caused by relaxation at a critical thickness. The close lattice matching of TiN\hkl(100) to MRG\hkl(110) means the TiN\hkl(001) out-of-plane texture is expected, which helps to induce a \hl{CoPt$_3$}\hkl(200) \hl{phase} in the Co/Pt \cite{hongyu_2015}. However, PMA is lost after annealing. \hl{Average roughness of Co and Pt increase moderately in all cases. Pt layers begin with a higher roughness in the case of high temperature grown TiN spacer layers. By comparison, the roughness after annealing is equal or greater when using room temperature grown TiN. This indicates that a high temperature grown TiN seed layer is less suitable, but is also more stable.} \subsection{Hafnium Oxide}\label{subsec:HfO$_x$} As HfO$_x$ is formed by natural oxidation of a Hf metal layer, there is a gradient of oxygen content that depends on the film thickness. MRG performs poorly with a $\SI{1.4}{\nano\metre}$ thick HfO$_x$ spacer resulting in near total loss of perpendicular anisotropy upon annealing, whereas with a $\SI{2}{\nano\metre}$ spacer it maintains both coercivity and squareness with reductions of \SI{11}{\percent} and \SI{14}{\percent} respectively, \hl{despite a significant increase in strain ($\sim {\SI{25}{\percent}}$)}. This is correlated with a lesser reduction in $c$-spacing for the $\SI{2}{\nano\metre}$ spacer stack. Similarly to the case of TiN, the reduction in $c$ comes despite an improvement in the coherence length. If the $\SI{1.4}{\nano\metre}$ HfO$_x$ is fully oxidised ($x \sim 2$), then diffusion into MRG, either across the barrier or from the barrier itself, will be reduced. This will result in oxidation of MRG at the interface, however. The effect is then similar to that using TiN. \hl{As the HfO$_x$ films were treated for equal time, the} thicker HfO$_x$ is not fully oxidised due to passivation, which leads to some interdiffusion. There is no observable texture in the Co/Pt superlattice before annealing. Weak \hl{CoPt$_3$}\hkl(111) and \hkl(200) texture is observed after annealing, but PMA is lost. \hl{As with the TiN layers, we see a moderate increase in Co and Pt roughness, with a higher as-deposited roughness.} \subsection{Hafnium}\label{subsec:Hf} Hf spacers work well, with the $\SI{1.4}{\nano\metre}$ spacer showing only a small change in all properties. In contrast to other materials, the $\SI{2}{\nano\metre}$ Hf spacer increases the $\delta_c$ of the MRG unit cell. Furthermore, it raises $T_{\text{comp}}$ to a value slightly above RT, which causes $H_c$ to increase significantly due to its divergent nature close to $T_{\text{comp}}$. The increase in $c$ is accompanied by a decrease in $L_c$ \hl{and an increase in $\epsilon$}, indicating an increase of line and planar defects. According to the binary phase diagrams \cite{okamoto_bapd}, Hf absorption into both Mn and Ga is limited to $\sim \SI{1}{at\ldotp\thinspace\percent}$ while maintaining crystal structure. This limited solubility accounts for the reduced effects seen in the HfO$_x$ $\SI{2}{\nano\metre}$ and Hf $\SI{1.4}{\nano\metre}$ based heterostructures, where the counter-diffusion into the Hf layer limits the available soluble material. We see different effects on the Co/Pt layer depending on the thickness of the Hf spacer, with the $\SI{1.4}{\nano\metre}$ thick spacer inducing no discernable texture in the superlattice. After annealing there is a weak \hl{CoPt$_3$}\hkl(111) texture. The $\SI{2}{\nano\metre}$ thick spacer promotes a \hl{CoPt$_3$}\hkl(111) orientation which is maintained after annealing, as well. In neither case is there any PMA after annealing, although there is a clear preferred orientation. \hl{The roughness of Co layers increases moderately. Annealing causes a substantial increase in roughness of Pt layers, to {$\sim$} {\SI{1}{\nano\metre}}, with the {\SI{2}{\nano\metre}} Hf spacer giving a higher initial roughness.} \subsection{Molybdenum}\label{subsec:Mo} Mo spacers have similar properties to Hf spacers, and the phase diagrams show a greater affinity of Mn for Mo ($\SI{4}{at\ldotp\thinspace\percent}$) and Ga ($\SI{16}{at\ldotp\thinspace\percent}$) Counter-diffusion is limited to Mn, as the Mo does not absorb much Ga at \SI{350}{\degreeCelsius}. The increase in $S$ for the \SI{1.4}{\nano\metre} Mo heterostructure, indicates reordering of the Ru within MRG which should result in a decrease in magnetic properties. However, the squareness is still within acceptable limits and the coercivity increases, indicating that $T_{\text{comp}}$ has increased despite a lower $c$ parameter. This is corroborated by the change in sign of the MRG switching direction in EHE, showing that $T_{\text{comp}}$ is now above RT. This effect is less prominent with the \SI{2}{\nano\metre} Mo heterostructure, but still indicates an increase in $T_{\text{comp}}$. In this case $L_c$ \hl{and $\epsilon$ have both} increased without any change in $S$. With more material to diffuse, the additional Mo \hl{inclusions maintain chemical ordering of the Ru $4d$ site.} With the thin $\SI{1.4}{\nano\metre}$ spacer, a weak \hl{CoPt$_3$}\hkl(200) peak is present, which becomes stronger after annealing. By contrast, there is a clear \hl{CoPt$_3$}\hkl(220) peak present when using a $\SI{2}{\nano\metre}$ spacer layer, which remains after annealing. This is the only case we witnessed where magnetic anisotropy of the Co/Pt was still perpendicular to the film surface after annealing. There is no clear Mo texture visible in XRD. \hl{Co and Pt layers both have moderate increases of roughness, however the heterostructure with a {$\SI{2}{\nano\metre}$} spacer layer has higher initial and final roughness.} \begin{table}[t] \caption{\hl{Data for Co/Pt layers before and after annealing - XRD peaks visible are given by CoPt$_3${\hkl(111)} $\vartriangle$, CoPt$_3${\hkl(200)} $\circ$, CoPt$_3${\hkl(220)} $\square$, Pt{\hkl(111)} $\triangledown$. Filled shapes indicate a strong peak, underlined shapes indicate a weak peak. Average interfacial roughness of Co and Pt layers with the superlattice are given based on fitting of XRR data. A $\checkmark$ indicates that PMA was still present after annealing}} \begin{tabularx}{255pt}{l X c c X c c } \hline & \multicolumn{3}{c}{\textbf{As-deposited}} & \multicolumn{3}{c}{\textbf{Annealed}}\\ & & \multicolumn{2}{c}{$\mathbf{R_a}$ (\si{\nano\metre})} & & \multicolumn{2}{c}{$\mathbf{R_a}$ (\si{\nano\metre)}} \\ \textbf{Spacer} & \textbf{XRD} & Co & Pt & \textbf{XRD}& Co & Pt\\ \hline V \SI{1.4}{\nano\metre} & $\vartriangle \circ$ &0.44 &0.39 & $\bullet$ & 0.51 & 0.80 \\ V \SI{2}{\nano\metre} & $\vartriangle \circ$ &0.43 &0.42 & $\bullet$ &0.52 &0.79 \\ TiN HT \SI{1.4}{\nano\metre} & $\circ$ \underline{$\triangledown$} &0.40 &0.70 & $\bullet$ \underline{$\triangledown$} &0.45 &0.80 \\ TiN HT \SI{2}{\nano\metre} & $\circ$ \underline{$\triangledown$} &0.43 &0.65 & $\bullet$ \underline{$\triangledown$} &0.50 &0.79 \\ TiN RT \SI{1.4}{\nano\metre} & $\circ$ \underline{$\triangledown$} &0.41 &0.54 & $\bullet$ \underline{$\triangledown$} &0.47 &0.91 \\ TiN RT \SI{2}{\nano\metre} & $\circ$ \underline{$\triangledown$} &0.44 &0.57 & $\bullet$ \underline{$\triangledown$} &0.50 &0.75 \\ HfO$_x$ \SI{1.4}{\nano\metre} & &0.44 &0.77 & \underline{$\vartriangle$} \underline{$\circ$} &0.49 &0.90 \\ HfO$_x$ \SI{2}{\nano\metre} & &0.44 &0.80 & \underline{$\vartriangle$} \underline{$\circ$} &0.55 &0.91 \\ Hf \SI{1.4}{\nano\metre} & &0.46 &0.53 & \underline{$\vartriangle$} &0.50 &0.99 \\ Hf \SI{2}{\nano\metre} & $\vartriangle$ &0.44 &0.73 & $\vartriangle$ &0.50 &1.00 \\ Mo \SI{1.4}{\nano\metre} & \underline{$\vartriangle$} &0.43 &0.43 & $\vartriangle$ &0.48 &0.50 \\ Mo \SI{2}{\nano\metre} $\checkmark$ & $\square$ &0.39 &0.59 & $\blacksquare$ &0.45 &0.70 \\ \end{tabularx} \label{table:copt} \end{table} \section{Discussion}\label{sec:disc} Strain plays a significant role in the properties of MRG thin films, and crystalline point defects have previously been discussed as a source of electronic doping that maintains strain within the lattice \cite{zic_2016}. As demonstrated by use of a TiN spacer, annealing of MRG removes line and planar-type crystalline defects, which relaxes the lattice. \hl{These types of defects have previously been identified in MRG from HRTEM} \cite{zic_2016,teichert2021}. This results in a \SIrange{60}{70}{\percent} loss of tetragonal distortion of the unit cell (where a \SI{100}{\percent} reduction would indicate a fully cubic cell). It is clear here that this is responsible for the reduction of coercivity and squareness in magnetic hysteresis due to the reorientation of the magneto-crystalline anisotropy into the plane. This effect is much stronger than could be attributable to Ru-based anti-sites, so it is not possible here to determine \hl{their contribution}. \hl{Where diffusion is discounted, there is a correlation between strain and chemical ordering within MRG, consistent with }\cite{zic_2016}. The diffusion of different materials has varying effect on MRG, \hl{for example} with V causing significant disorder of the crystal structure, ultimately resulting in the formation of additional intermetallic phases. \hl{The choice of a suitable spacer material is based on a requirement to maintain magnetic properties of MRG after processing.} Hf and Mo are strong candidates for future applications, \hl{as in both cases, the magnetic hysteresis is reliable with a change in squareness {\SI{<10}{\percent}} from the as-deposited state. This is especially important in magnetoresistive devices which rely upon the relative orientation of magnetic moments between the free and reference layers.} $T_{\text{comp}}$ is sensitive to the number of valence electrons within MRG, and increases linearly with $n_v$ \cite{siewierska2021}. Both Hf and Mo appear to donate valence electrons to the film, as indicated by the increase in $T_{\text{comp}}$. Based on Slater-Pauling rules, both Hf and Mo would donate $\sim 0.2$ electrons per unit cell of MRG for the considered solubility. This contextualises the same effect seen in our previous work using thin interlayers \cite{borisov2017,titova2019}. \hl{The results for these materials indicate a complex interplay between strain and film defects when we consider incorporation of additional valence electrons into the unit cell, that results in retention of perpendicular anisotropy.} The strongly bonded compounds TiN and HfO$_x$ do not diffuse. However, the heterostructure using the thicker, partially oxidised HfO$_x$ spacer layer more resembles the structures using a Hf spacer layer in behaviour due to diffusion of metallic species closer to the MRG interface where the layer is not fully oxidised. Typically we see either \hl{CoPt$_3$}\hkl(200) or \hl{CoPt$_3$}\hkl(111) texture in the free layer, however in the case of the \SI{2}{\nano\metre} Mo heterostructure there appears a \hl{CoPt$_3$}\hkl(220) peak. It is of interest that of all the structures, this is the only one that maintained perpendicular anisotropy in the Co/Pt after annealing. We can surmise that a mixture of weak \hkl(111) and \hkl(200) indicates an amorphous interface between spacer and the first Co layer, which is then crystallised and induces a preferential texture on the superlattice during annealing. Annealing results in an increase in interfacial roughness within the superlattice. \hl{PMA after annealing is independent on the degree of roughness measured, and is instead correlated with the appearance of CoPt$_3$ phases which are the result of intermixing. The superlattice which forms a CoPt$_3$}\hkl(220) phase does not loose perpendicular anisotropy, which we attribute here to the additional strain on the \hl{remaining Co/Pt interfaces} induced by the crystal texture. As Mo is an element of comparable atomic weight to Pt, the additional orbital contribution at the interface between spacer and superlattice likely contributes to the anisotropy rather than any strain induced by epitaxial growth, considering that the Mo has no visible texture itself. \hl{PMA after annealing when using a {$\SI{2}{\nano\metre}$} Mo spacer, when there is none present using a {$\SI{1.4}{\nano\metre}$} Mo spacer, indicates that counter diffusion of MRG has an effect on the interfacial relationship between the spacer and the initial Co layer of the superlattice.} \section{Conclusions}\label{sec:conclusions} \hl{This} study has enabled us to characterise the effects of \hl{various} spacer layers on the structural and magnetic properties of both the MRG and the Co/Pt superlattice, which will be useful for developing perpendicular GMR and TMR structures that can withstand annealing at \SI{350}{\degreeCelsius}. The PMA of MRG is more robust, since it depends on strain imposed by the MgO substrate. That of the Co/Pt multilayer is only maintained when a \hl{CoPt$_3${\hkl(220)} phase is developed by intermixing}, with the help of Mo. \hl{Based on the effect of annealing MRG films capped with a TiN spacer, line and planar defects help to maintain the substrate-induced strain and the tetragonal distortion of the MRG crystal structure, which the perpendicular magnetic anisotropy is dependent on.} Annealing causes these defects to be removed, which results in the partial collapse of tetragonal distortion and therefore magnetic properties dependent on the perpendicular anisotropy. There was no identifiable link between these properties and the presence of Ru-based anti-site disorder. \hl{The effect of strain on anisotropy is counteracted by the addition of valence electrons in the MRG unit cell} In Co/Pt multilayers, interface roughness is the main antagonist of \hl{PMA due to the reduction of interfacial anisotropy. Intermixing results in the formation of a CoPt$_3$ phase, even before annealing. However, the development of CoPt$_3$} with \hkl(220) texture retains perpendicular anisotropy after a \SI{350}{\degreeCelsius} anneal. A Mo underlayer is proposed here as a viable solution. Hf and Mo \hl{will} be useful as thin protective layers in future active devices \hl{based on MRG}, as they \hl{are able to offset changes in strain and defect density to maintain the PMA of the MRG layer}. The solubility of both materials within MRG provides additional valence electrons, modifying the Fermi level position, and therefore $T_{\text{comp}}$ \hl{as well as the overall magnetic properties}. This is in contrast to V, which prefers to form intermetallic phases instead. Considering the enhanced high-temperature stability of the crystal, doping of MRG with these or similar materials, such as Zr, should be investigated. By tuning the Ru/dopant composition, desired magnetic properties should be easily achievable. \section{Acknowledgements} This work was partly funded by Science Foundation Ireland under grant 12/RC/2278 (AMBER), and by the European Commission under Grant Agreement No. DLV-737038 (TRANSPIRE). The work of K. S. was supported by the Irish Research Council under Grant GOIPG/2016/308. \bibliographystyle{elsarticle-num}

4649b4aeba0d8ec6bbd5d8703285f145a191cfa6

\section{Introduction} \label{sec:intro} Decoherence in quantum systems arises as an inevitable consequence of interactions with the environment. Theoretically, it has been used to understand the measurement problem and provide a gateway to the classical world \cite{joos2013decoherence}. From a practical perspective, decoherence poses a challenge to developing large-scale quantum computers, whose efficacy degrades with the loss of coherence. Quantum error correction meets this challenge by providing a scalable means of correcting a continuum of possible errors, representable by a finite set of operators, using only a discrete number of components \cite{shor1995scheme,gottesman2009introduction}. In this paper, we will investigate whether classical analog systems can exhibit similar behavior. We shall begin in Sec.\ \ref{sec:qed} with a description of a classical emulation of a gate-based quantum computer. There, we describe how arbitray quantum states may be represented by classical signals conforming to the mathematical tensor-product structure of a multi-qubit Hilbert space. We go on to describe how one- and two-qubit gate operations can be performed on these representative states and even how quantum measurements can be faithfully emulated. Finally, describe briefly a hardware implementation of a device capable of emulating a two-qubit quantum computer. We go on, in Sec.\ \ref{sec:qst}, to describe how one may use such a device to perform quantum state tomography and, therefore, infer the equivalent mixed quantum state from an ensemble of imperfect state preparations. In Sec.\ \ref{sec:qpt} we extend our investigation to modeling gate operations in our classical device in terms of quantum operations. This provides the natural mathematical framework for studying classical performance degradation in terms of quantum decoherence. By applying a sequence of gate operations iteratively, we are able to measure the systematic falloff in performance of the device and model it as a parameterized quantum channel. Our conclusions are summarized in Sec.\ \ref{sec:end}. \section{Classical Emulation of a Quantum Device} \label{sec:qed} In previous work, we have described the use of classical signals and analog electronics to emulate or ``mimic'' the behavior of a gate-based quantum computer \cite{LaCour&Ott2015}. The basic idea is quite simple. Given an abstract Hilbert space $\mathcal{H}$ used to represent an $n$-qubit quantum state $\ket{\psi} \in \mathcal{H}$, we seek a classical representation of these abstract mathematical objects. Many possibilities suggest themselves. The one we shall choose is motivated by our desire to easily perform operations on it. To that end, we adopt sinusoidal analog signals as a convenient physical representation of a pure quantum state. If we denote by $\ket{x}$, where $x \in \{0, \ldots, 2^n-1\}$, a basis function in the so-called computational basis of $\mathcal{H}$, then its classical representation is defined as follows. Let $[x_{n-1} \cdots x_0]$ be the little endian binary representation of $x$. The corresponding classical representation is then written as a complex signal $\phi_x$ defined such that \begin{equation} \phi_x(t) = \exp[(-1)^{x_{n-1}} i \omega_{n-1} t] \cdots \exp[(-1)^{x_{0}} i \omega_{0} t] \; , \end{equation} where the frequencies $\omega_{n-1} > \cdots > \omega_0$ correspond to each of the $n$ qubits. In particular, taking $\omega_k = 2^k \omega_0$ allows for a uniform spacing among the $2^n$ different combinations of sums and differences. Clearly, the required bandwidth for such a representation grows exponentially with the number of qubits. Linear combinations of basis signals provide a representation for an state in $\mathcal{H}$. Thus, if $\braket{x|\phi} = \alpha_x$, then \begin{equation} \psi(t) = \sum_{x=0}^{2^n-1} \alpha_x \, \phi_x(t) \end{equation} provides a classical representation of the quantum state $\ket{\psi}$. If $T = 2\pi/\omega_0$ is the period of the signal, then an inner product may be defined as \begin{equation} \braket{\phi_x|\psi} = \frac{1}{T} \int_0^T \phi_x(t)^* \psi(t) \, dt \; . \end{equation} Note that the set of computational basis signals forms an orthonormal basis. Gate operations are performed using projection operations. Given a quantum state $\ket{\psi}$ and a qubit $i$ to be addressed, we can formally decompose it into projections onto the subspaces in which the qubit is either 0 or 1. Thus, we may write \begin{equation} \ket{\psi} = \Pi_0^{(i)} \ket{\psi} + \Pi_1^{(i)} \ket{\psi} = \ket{0}_i \ket{\psi_0^{(i)}} + \ket{1}_i \ket{\psi_1^{(i)}} \; , \end{equation} where $\ket{\psi_0^{(i)}}$ and $\ket{\psi_1^{(i)}}$ will be called partial projection states. As described in Ref.\ \cite{LaCour&Ott2015}, the corresponding signals $\psi_0^{(i)}$ and $\psi_1^{(i)}$ may be produced from the signal $\psi$ using classical analog signal processing devices. In this manner, the nonseparable subspace projections may be separated and operated upon individually. Doing so allows one to perform a single-qubit gate operation $U$ on qubit $i$ by noting that \begin{equation} U_i \ket{\psi} = U\ket{0}_i \ket{\psi_0^{(i)}} + U\ket{1}_i \ket{\psi_1^{(i)}} \; . \end{equation} So, simple multiplication and addition of analog signals is all that is needed to effect this transformation. Measurements are performed in a similar manner. Given the partial projection signals $\psi_0^{(i)}$ and $\psi_1^{(i)}$ we may perform a measurement on qubit $i$ by first measuring the root-mean-square (RMS) voltages $v_0 = \|\psi_0^{(i)}\|$ and $v_1 = \|\psi_0^{(i)}\|$ of each signal and, from these, computing the probability \begin{equation} p = \frac{v_0^2}{v_0^2 + v_1^2} \; . \end{equation} A random variable $u_i \in [0,1]$, which serves the role of a hidden variable, is drawn such that $u_i \le p$ indicates an outcome of 0 and $u_i > p$ indicates an outcome of 1. Upon measurement, the state ``collapses'' to form the new signal $\psi' \propto \psi_0^{(i)}$ or $\psi' \propto \psi_1^{(i)}$, depending upon the measurement outcome. Additional qubits may be measured sequentially in this manner. This procedure, then, faithfully reproduces the quantum statistics dictated by the Born rule. We have implemented a small-scale quantum emulation device in hardware using breadboards and analog electronic components interfaced to a digital desktop computer. The device can be operated in one- or two-qubit mode at frequencies of 1000 Hz and 2000 Hz, respectively. Arbitary one-qubit gate operations can be performed as well as controlled gate operations using qubit 0 as the control and qubit 1 as the target. Typical gate fidelities are found to be over 99\% \cite{LaCour&al2016}. Measurement gates are performed using true-RMS voltage chips and digital switching. Sequential gate operations can be performed using a software interface. For example, a simple implementation of Deutch's algorithm can be programmed for an unknown Boolean function. In practice, we find that the device is able to correctly identify whether the function is constant or balanced about 96\% of the time. Since the algorithm should, ideally, produce the correct answer every time, the nonzero error rate must be due to device imperfections. In the following sections, we will investigate whether these imperfections can be modeled as quantum decoherence. \section{Application of Quantum State Tomography} \label{sec:qst} Using the representation and measurement procedure described in Sec.\ \ref{sec:qed}, we are able to use our device to prepare and measure any quantum state and observable of up to two qubits. Of course, imperfections in the device itself can give only a limited approximation of the ideal mathematical operations. This situation is quite similar to that found in actual quantum devices or experiments, and we may use similar tools to study it. We have heretofore discussed the representation of \emph{pure} quantum states using our device, but the more general quantum description is that of a \emph{mixed} state. Mixed quantum states may be thought of as an ensemble of pure states. Equivalently, in our classical representation, they may be thought of as noisy signals. For example, it can be shown that additive Gaussian white noise is equivalent to a mixture of the ideal state and a fully mixed state \cite{LaCour&Ostrove2017}. Noise can also be added intentionally to reproduce certain quantum measurement effects, as was done with so-called dressed states \cite{LaCour&al2016}. For the purposes of the present study, however, we are simply interested in estimating the equivalent quantum mixed state given a set of measurement outcomes. This may be done using the technique of quantum state tomography (QST). A general (mixed) $n$-qubit state $\rho$ may be decomposed into a basis of $4^n$ separable orthonormal operators using the Hilbert-Schmidt inner product, as follows: \begin{equation} \rho = \sum_{j_{n-1}=1}^{4} \cdots \sum_{j_0=1}^{4} \textcolor{black}{\mathrm{Tr}\left[ \rho \; \frac{\sigma^{(n-1)}_{j_{n-1}} \otimes \cdots \otimes \sigma^{(0)}_{j_0}}{2^n} \right]} \frac{\sigma^{(n-1)}_{j_{n-1}} \otimes \cdots \otimes \sigma^{(0)}_{j_0}}{2^n} \; , \end{equation} where \begin{equation} \sigma_1^{(k)} = \mtx{I}_k \; , \quad \sigma_2^{(k)} = \mtx{X}_k \; , \quad \sigma_3^{(k)} = \mtx{Y}_k \; , \quad \sigma_4^{(k)} = \mtx{Z}_k \end{equation} are the four Pauli spin operators applied to qubit $k$. Since the trace represents an expectation value under the Born rule \cite{vonNeumann}, we may use this decomposition to empirically determine the quantum state by measuring each of the basis operators. Let $\bar{B}_{j_{n-1},\ldots,j_0} \in \mathbb{R}$ denote the mean value obtained from a finite sample of measurements of the operator $\sigma^{(n-1)}_{j_{n-1}} \otimes \cdots \otimes \sigma^{(0)}_{j_0} / 2^n$. From these results, one may estimate the quantum state to be \begin{equation} \bar{\rho} = \sum_{j_{n-1}=1}^{4} \cdots \sum_{j_0=1}^{4} \textcolor{black}{\bar{B}_{j_{n-1},\ldots,j_0}} \frac{\sigma^{(n-1)}_{j_{n-1}} \otimes \cdots \otimes \sigma^{(0)}_{j_0}}{2^n} \; . \end{equation} In practice, $\bar{\rho}$ will not be a valid quantum state, since the numerical coefficients are not guaranteed to yield an operator that is both positive definite and of unit trace. A better procedure is to restrict one's search to valid quantum states and find the maximum likelihood estimate (MLE) of the quantum state, here denoted $\hat{\rho}$, that both satisfies this constraint and best fits the measured mean values. To this end, we use an MLE procedure developed by Altepeter, Jeffrey, and Kwiat under the assumption of independent Gaussian errors \cite{Kwiat2006}. Once a valid estimate $\hat{\rho}$ of the quantum state is obtained, it may be compared to the ideal quantum state $\ket{\psi}$ by computing the fidelity $F$ of the former to the latter using the expression \cite{Jozsa1994} \begin{equation} F = \sqrt{\langle \psi | \hat{\rho} | \psi \rangle} \; . \end{equation} Note that $F$ is bounded between zero and one. If $\hat{\rho} = \mtx{I} \otimes \cdots \otimes \mtx{I} / 2^n$ (a completely mixed state), then $F = 1/2^n$. Thus, in practice we expect to find intermediate values of $F$ such that $1/2^n < F < 1$. As an example, we considered the pure entangled state $\ket{\psi} = \frac{1}{\sqrt{2}} [ \ket{01} - \ket{10} ]$. So, in matrix form in the computational basis, the ideal quantum state is \begin{equation} \rho = \ket{\psi}\bra{\psi} = \frac{1}{2} \begin{pmatrix} 0 & 0 & 0 & 0 \\ 0 & +1 & -1 & 0 \\ 0 & -1 & +1 & 0 \\ 0 & 0 & 0 & 0 \end{pmatrix} \; . \end{equation} The estimated quantum state was found to be \begin{equation*} \resizebox{\textwidth}{!}{$\hat{\rho}= \begin{pmatrix} 0.0001 & -0.0011-0.0082i & \;0.0008+0.0075i & -0.0006+0.0004i \\ -0.0011+ 0.0082i & 0.5412 & -0.4968-0.0082i & 0.0019-0.0364i \\ 0.0008-0.0075i & -0.4968+0.0082i & 0.4562 & -0.0012+0.0335i \\ -0.0006+0.0004i & 0.0019 + 0.0364i & -0.0012-0.0335i & 0.0025 \end{pmatrix} \; ,$} \end{equation*} giving a fidelity of $F = 0.9978$. The two states are shown graphically in Fig.\ \ref{fig:QST}. \begin{figure} \centering \begin{subfigure}[t]{.45\textwidth} \centerline{$\rho$} \includegraphics[width= .9\textwidth]{Figures/rhoIdeal_rev2} \caption{} \end{subfigure} \begin{subfigure}[t]{.45\textwidth} \centerline{$\hat{\rho}$} \includegraphics[width= .9 \textwidth]{Figures/rhoQST_rev2} \caption{} \end{subfigure} \caption{Cityscape plot of the ideal quantum state (left) and that inferred from quantum state tomography (right). Only the real part of the matrix elements is shown.} \label{fig:QST} \end{figure} \section{Application of Quantum Process Tomography} \label{sec:qpt} In quantum mechanics, the evolution of a closed system is given by a unitary transformation generated by the system's Hamiltonian, in accordance with the Schr\"odinger equation. In practice, no system is ever truly isolated, and this can lead to apparent non-unitary evolution. The formalism of quantum operations gives us a framework with which to characterize the behavior of open quantum systems and, in particular, decoherence \cite{Sudarshan1961,Mike&Ike}. A quantum operation may be viewed as a superoperator on quantum states such that, if $\rho$ is the initial quantum state, then $\rho' = \mathcal{E}(\rho)$ is the state that results from some, possibly non-ideal, transformation. We will make use of an equivalent formulation of quantum operations known as the operator-sum representation \cite{Stinespring1955}. Using this formalism, the quantum operation $\mathcal{E}$ may be characterized by a discrete set of operators such that \begin{equation} \mathcal{E}(\rho) = \sum_k E_k \, \rho \, E_k^\dagger \; . \end{equation} The matrices $\{E_k\}$ are known as Kraus operators \cite{Kraus}. A further decomposition of the Kraus operators may be performed in terms of, say, the Pauli operators, as was done for QST. For our present purposes we will consider, for simplicity, only single-qubit states. In this case, each Kraus operator may be written as a linear combination of the four Pauli operators, so that \begin{equation} E_k = \sum_{i=1}^{4} e_{k,i} \, \frac{\sigma_i}{2} \; . \end{equation} Using this representation, the quantum operation may be written as \begin{equation} \mathcal{E}(\rho) = \sum_k \left(\sum_{i=1}^{4} e_{k,i}\frac{\sigma_i}{2}\right) \rho \left(\sum_{j=1}^{4} e_{k,j}\frac{\sigma_j}{2}\right)^\dagger = \sum_{i=1}^{4} \sum_{j=1}^{4} \chi_{i,j} \; \sigma_i \, \rho \, \sigma_j^\dagger \; , \end{equation} where \begin{equation} \chi_{i,j} = \sum_k e_{k,i} \, e_{k,j}^* \end{equation} are the elements of the $4\times4$ chi process matrix $\mtx{\chi}$ \cite{Choi1975,Chuang1997}. Determination of the equivalent quantum operation therefore reduces to the problem of estimating the corresponding chi matrix. This, in turn, may be accomplished using the techniques of quantum process tomography (QPT) \cite{Chuang1997,Poyatos1997,Bhandari2016}. As an example, we performed QPT on a single-qubit identity gate using the procedure outlined in Ref.\ \cite{Mike&Ike} but modified to use a maximum likelihood QPT technique \cite{AGWhite2004mlqpt,Anis2012,Yuen-Zhou2014}. We started by generating an ensemble of input states of the following form: \begin{equation} \ket{\psi_1} = \ket{0} \; , \;\; \ket{\psi_2} = \ket{1} \; , \;\; \ket{\psi_3} = \tfrac{1}{\sqrt{2}} [ \ket{0} + \ket{1} ] \; , \;\; \ket{\psi_4} = \tfrac{1}{\sqrt{2}} [ \ket{0} + i \ket{1} ] \; . \end{equation} The process, in this case an identity gate, was applied to the ensemble of states, which were then measured using this same basis. The chi matrix was parameterized using a Cholesky factorization such that $\mtx{\chi} = \Delta \Delta^\dagger$, where $\Delta$ is a lower-triangular matrix with real, positive diagonal elements. This form guarantees that the constraints of Hermiticity, trace preservation, and complete positivity are satisfied. We then optimize over $\Delta$ with respect to the following likelihood function: \begin{equation} L(\Delta) = \frac{1}{2} \sum_{\alpha} \sum_{\beta} \frac{\left[ N_{\alpha,\beta} - C \sum_{i} \sum_{j} \bra{\psi_{\beta}} \sigma_i \ket{\phi_{\alpha}} \bra{\phi_{\alpha}} \sigma_j \ket{\psi_{\beta}} (\Delta \Delta^\dagger)_{i,j} \right]^2}{ C\sum_{i} \sum_{j} \bra{\psi_{\beta}} \sigma_i \ket{\phi_{\alpha}} \bra{\phi_{\alpha}} \sigma_j \ket{\psi_{\beta}} (\Delta \Delta^\dagger)_{i,j}} \; , \end{equation} where $\ket{\phi_\alpha}$ and $\ket{\psi_\beta}$ are the input states and measurement settings, respectively, and $N_{\alpha , \beta}$ are the experimentally measured counts for the corresponding pair of input state and measurement settings. The factor $C$ is the total number of such counts. The resulting estimated chi matrix, $\hat{\mtx{\chi}}$, is illustrated graphically in Fig.\ \ref{fig:QPT}. Ideally, the matrix should be such that $\chi_{i,j} = \delta_{i,j}$, indicating that the quantum operation takes the simple form $\mathcal{E}(\rho) = \sigma_1 \, \rho \, \sigma_1 = \rho$. Empirically, we do indeed find that $\hat{\chi}_{1,1}$ is nearly 1 and is, therefore, the dominant component. A closer inspection reveals that there are other nonzero components. In particular, the $\hat{\chi}_{1,j}$ and $\hat{\chi}_{j,1}$ components appear to have non-negligible imaginary terms. \begin{figure} \centerline{\hspace{8em}$\mathrm{Re} \hat\chi$ \hfill $\mathrm{Im} \hat\chi$ \hspace{8em}} \begin{center} \includegraphics[width=\textwidth]{Figures/QPT_Identity_Rev2} \end{center} \caption{Cityscape plot of the QPT results for a single application of the identity gate. The real part of $\hat{\mtx{\chi}}$ is shown on the left, while the imaginary part of $\hat{\mtx{\chi}}$ is shown on the right. Note that the two figures are shown on very different scales to illustrate the nonzero contributions to the estimate.} \label{fig:QPT} \end{figure} The gate fidelity relative to an ideal unitary operator $U$ may be determined from the estimated quantum operation $\mathcal{E}$ according to the formula \cite{gilchrist2005distance} \begin{equation} F = \min_{\rho} \; \mathrm{Tr} \sqrt{\sqrt{\mathcal{E}(\rho)} \; U \rho U^\dagger \sqrt{\mathcal{E}(\rho)}} \; . \label{eqn:gate_fidelity} \end{equation} For the present case, $U$ is the identity and $\mathcal{E}$ is estimated by the chi matrix $\hat{\mtx{\chi}}$, for which we find that $F = 0.9933$ for a single identity gate operation. This is comparable to what was found earlier for the quantum state fidelity using QST and, so, these results appear to be consistent. Ideally, the chi matrix should give a full characterization of the quantum process (in this case, a single application of the identity gate operation). In particular, it should provide a means of forecasting the gate fidelity over multiple multiple iterations. Indeed, if the initial quantum state is determined to be $\rho_0$, then the state after $n$ iterations, denoted $\rho_n$, is given iteratively by \begin{equation} \rho_n = \mathcal{E}(\rho_{n-1}) = \mathcal{E}( \cdots \mathcal{E}(\rho_0) \cdots ) \; . \end{equation} Consider the depolarizing channel with parameter $p \in [0,1]$, for which \begin{equation} \mathcal{E}(\rho) = (1-p) U \rho U^\dagger + p \mtx{I} \; . \end{equation} The fidelity of this channel is then \begin{equation} F = \sqrt{ 1 - \frac{2p}{3} } \; . \end{equation} The depolarizing channel is closed under multiple iterations, with an effective parameter $p_n$ after $n$ iterations of \begin{equation} p_n = \frac{3}{4} \left[ 1 - \left( 1 - \frac{4p}{3} \right)^n \right] \; , \end{equation} yielding a cumulative fidelity of \begin{equation} F_n = \sqrt{1 - \frac{1}{2} \left[ 1 - \left( 1 - \frac{4p}{3} \right)^n \right] } \; . \end{equation} In addition to the gate fidelity defined in Eqn.\ (\ref{eqn:gate_fidelity}) we make use of an alternative benchmark known as the quantum process fidelity defined as \cite{gilchrist2005distance} \begin{equation} F_{\textrm{proc}}= \textrm{Tr}(\hat{\chi} uu^\dagger) \label{eqn:process_fidelity} \end{equation} where $\hat{\chi}$ is the measured chi matrix and $uu^\dagger$ is the rank-one chi matrix for the ideal unitary transformation $U$. The process fidelity has the benefit of being less computationally intensive to calculate as we need not perform the optimization step over input states. For the depolarizing channel, it may be interpreted as the probability that the ideal operation was performed. Using our device, we explored the behavior of the measured process fidelity upon performing multiple iterations of the identity gate, using QPT to estimate the fidelity for each iteration. The results are summarized in Fig.\ \ref{fig:forecasted}. Surprisingly, the process fidelity after 90 iterations drops only to about 0.874. This is well above the 0.55 cumulative process fidelity that is predicted from a simple fit to a depolarizing channel based on the fidelity of a single gate operation, which has a parameter value of $p = 0.010$. Using instead the actual chi matrix estimate, and iterating the corresponding quantum operation, yields a sequence of process fidelity values with a curious, oscillatory behavior, as shown in Fig.\ \ref{fig:forecasted}. Initially, it seems, the fidelity drops sharply. After the $35^{\rm th}$ iteration, however, it begins to climb up again, only to crest and fall once more after about the $70^{\rm th}$ iteration. This shows the folly of extrapolation based on a single QPT estimate. If one instead considers the whole sequence of iterations, a much better fit can be achieved, as illustrated in Fig.\ \ref{fig:fitted}. In this case, forecasting was done based on optimizing the channel parameterization in order to minimize the least-squares error to \emph{all} of the measured data (i.e., over all 90 iterations). Doing so, we found that a simple depolarizing channel actually did provide a good fit to data, albeit with a model parameter value of $p = 0.006$. This is just slightly lower than the value estimated from a single gate iteration, but the impact on the forecasted fidelity is quite significant. We note that the resulting depolarizing channel also agrees quite well with the forecasted chi matrix, when fitted to all 90 iterations. \begin{figure} \centering \includegraphics[width=\textwidth]{Figures/forecasted_rev2} \caption{Plot of experimentally measured process fidelity (determined via QPT) versus iteration count (red), along with forecasted results based on direct propagation of initially estimated chi matrix (black line) and depolarizing channel model (blue).} \label{fig:forecasted} \end{figure} \begin{figure} \centering \includegraphics[width=\textwidth]{Figures/fitted_rev2} \caption{Plot of experimentally measured process fidelity versus iteration count along with forecasted results based on fitting both an estimated chi matrix and a depolarizing channel model to the data using least squares fitting.} \label{fig:fitted} \end{figure} \section{Conclusions} \label{sec:end} We have considered the performance of a classical device in terms of quantum operations. Using simple electronic hardware we are able to emulate the behavior of a two-qubit quantum computer by representing the quantum state as an analog voltage signal. An arbitrary sequence of one- and two-qubit gate operations can be performed on these signals, thereby allowing for full programmability. As with any physical device, these operations are not performed perfectly but, instead, exhibit some level of degradation. In order to understand this better, we have chosen to use to the mathematical framework of quantum operations to model this classical degradation of performance in terms of quantum decoherence. In particular, we avail ourselves of the techniques from the field of quantum state and quantum process tomography to extract an empirical estimate of the equivalent quantum channel corresponding to a given gate operation. What we find is that estimate of the quantum channel obtained from performing QPT on a single gate operation provides a poor prediction of its performance upon repeated iterations. If, however, the channel is estimated over a number of iterations, then a good fit can indeed be achieved. This is particularly true when, as is the case for our device, the gate fidelity on any single iteration is quite high. In such cases, the error is quite low and the estimation process will be very sensitive to the behavior over the first few iterations. When these considerations are taken into account, we find that we are able to achieve a good fit to the measured results by assuming a simple depolarizing channel as a model for the effective quantum operation. Given that decoherence may be viewed as a classical process that can be modeled quantum mechanically, the possibility arises for the use of quantum error correction techniques to improve performance. In particular, we have shown that errors in a classical analog device can be modeled solely in terms of discrete bit-flip and phase error quantum operations, which are sufficient for modeling all forms of decoherence. This presents the exciting new prospect of using the methods of fault-tolerant quantum computing to improve the fault tolerance of classical analog devices. Whether the inclusion of additional, albeit faulty, classical resources can improve overall performance will be a subject for future investigations. \begin{acknowledgement} This work was support by the Office of Naval Research under Grant No. N00014-14-1-0323. \end{acknowledgement} \input{main.bbl} \end{document}

83c103c67da6e5a34ca0d56c3ddc19e10e13596f

\section{Introduction} A large class of gapped quantum phases are believed to be described by topological quantum field theories (TQFT) at low energy. They are broadly divided into two categories: short-ranged entangled (SRE) states, which have unique ground state on any closed (spatial) manifold and all low-lying excitations are local; long-range entangled (LRE) states, which have topology-dependent ground state degeneracy on closed manifolds and exhibit topologically nontrivial excitations. The presence of global symmetry further enriches the landscape and leads to sharp distinctions between phases protected by the symmetry. We will often loosely refer to symmetric SRE phases as symmetry-protected topological (SPT) phases, and symmetric LRE phases as symmetry-enriched topological (SET) phases. The classification of interacting SPT phases has been studied intensively in the past decade, and by now a consistent picture for SPT phases protected by internal or spatial symmetries has been largely achieved, at least in low dimensions relevant to condensed matter systems~\cite{Chen2013, kapustin2014symmetry, Kapustin_2015, freed2019reflection, Xiong_2018, Gaiotto_2019, WangGuPRX2020, ThorngrenPRX2018}. For SET phases, a systematic mathematical framework was developed for the (2+1)d case in Ref. [\onlinecite{SET}] based on tensor category, and more recently Ref. [\onlinecite{Kong_2020}] has proposed a general theory in higher dimensions. Even though SRE and LRE phases have very different physical properties, they are in fact closely related. On one hand, gauging a finite unitary symmetry in a symmetric SRE phase results in a topological gauge theory~\cite{Levin_2012}. The connection can be utilized in both ways: SPT phases may be characterized by fusion and braiding statistics of symmetry fluxes, while universal properties of the gauge theories can be traced back to the ungauged SPT phases. Another connection is provided by 't Hooft anomaly: on the boundary of an SPT phase, the symmetry must be implemented in an anomalous way~\cite{kapustin2014}, and as a consequence a symmetry-preserving gapped state must be LRE, i.e. topologically ordered~\cite{VishwanathPRX2013, Chen2014}. Therefore ground states of systems with 't Hooft anomaly are strongly constrained, in particular a symmetric SRE ground state is forbidden. Recently, a new boundary phenomenon has been discovered: there exist ``anomalous'' SPT phases which can only be realized on the boundary of a topologically trivial state in one dimension higher~\cite{WangASPT}, or in an intrinsically gapless system~\cite{thorngren2020intrinsically}. Examples discussed so far all occur in fermionic systems, so one naturally wonders whether ASPT phases also exist in bosonic systems, and how they fit into the general classification scheme. On the other hand, significant progress has been made in classifying SPT phases with crystalline symmetries~\cite{ThorngrenPRX2018, HuangPRB2017, shiozaki2018generalized, Song_2020}. In particular, the bulk-boundary correspondence of crystalline SPT phases shed new light on the celebrated Lieb-Schultz-Mattis theorems~\cite{LSM}, as manifestation of 't Hooft anomaly for spatial symmetries~\cite{ChengPRX2016, HuangPRB2017, JianPRB2018, ChoPRB2017}. This new perspective has naturally led to various generalizations of LSM-type constraints. While in most cases LSM theorems forbid SRE ground state, recent works~\cite{Lu_LSMSPT, Yang_LSMSPT, JiangLSM, ElseLSM} have revealed a new twist in which microscopic constraints allows SRE states, which nevertheless have to be in a nontrivial SPT phase. The goal of this work is to provide a unified view of the three kinds of phenomena: a) anomalies in topological gauge theories, b) anomalous SPT phases and c) LSM theorems for SPT phases. We will see that all of them can be understood within the same framework, namely a systematic theory of decorated domain wall (DW) construction of SPT phases~\cite{ChenNC2013}, which are described mathematically using Atiyah-Hirzebruch spectral sequence for generalized cohomology, as shown in Ref. [\onlinecite{Gaiotto_2019}] (see also Ref. [\onlinecite{Xiong_2018}]). We will explain the mathematical structure through explicit constructions and further expound on the relations with anomalies in this work. The basic idea is that given a global symmetry group $G$, a symmetric state can be obtained from quantum disordering a symmetry-breaking state, or proliferation of symmetry domain walls. One can imagine that domain walls and their junctions are decorated by lower-dimensional SRE phases (possibly protected by unbroken symmetries). In order to produce a fully gapped, symmetric state, it is important that all possible domain wall configurations can be smoothly deformed from one to another, so by local Hamiltonian terms that fluctuate the domain walls one can find a ground state which is the superposition of all domain wall configurations. The key observation is that consistency conditions for domain wall decorations can be organized by Atiyah-Hirzebruch spectral sequence. As will be explained below, in a general domain wall decoration one successively considers domain wall junctions with increasing codimension (codimension-$0$ is the whole system, codimension-1 is the domain wall, etc.). At each step, one needs to make sure that the decoration is consistent, which leads to an obstruction-vanishing condition. These conditions can be formulated in terms of differentials in the associated spectral sequence. Computing differentials is generally a difficult task. Recently Ref. [\onlinecite{WangGuPRX2020}] essentially derived explicit expressions for differentials for the spectral sequence that describes fermionic SPT phases. The main technical achievement of this work is to provide explicit expressions for \emph{cochain-level} differentials in the Lyndon-Hochschild-Serre (LHS) spectral sequence~\cite{Lyndon, HS}, which correspond to decorating domain walls with bosonic group-cohomology SPT phases protected by another group $A$, but $G$ is nontrivially extended by $A$. We then provide physical interpretations of the differentials. First of all, for each domain wall decoration in $D$ dimensions, differentials compute possible inconsistencies, represented by a domain wall decoration in $(D+1)$ dimensions. If the differential is nontrivial, it means that the decoration can not be realized, unless the system is actually on the boundary of a $(D+1)$-dimensional bulk. Therefore, such obstructed decorations correspond to anomalous SPT phases, which need to be excluded from the classification of SPT phases in $D$ dimensions. Using the explicit expressions for differentials, we find that the examples of SPT-LSM theorems for magnetic translation symmetry in the literature can indeed be understood as ASPTs, when the spatial symmetry is formally treated as internal. We also study a new example of LSM theorem, which forbids SRE ground states in a spin-$1/2$ model with easy-plane anisotropy and $\pi$-flux background. Using the tools developed in this work we prove that if the a fully symmetric topologically ordered ground state exists, then anyons must be permuted by some discrete symmetries of the system. Now we consider a fully consistent decoration, which should describe an actual SPT state. However, when it is the image of a differential map from another decoration in one dimension lower, the SPT state is actually trivial topologically. We will argue that the differential implies that a gapped symmetry-preserving boundary has to be an anomalous SPT phase. We give an example of a bosonic anomalous SPT phase in one-dimensional spin chain with a non-on-site $\mathbb{Z}_4$ symmetry for illustration. For another important application, it is well-known that gauging a normal subgroup of the symmetry group in an SPT phase turns it into a topological gauge theory, enriched by the remaining quotient symmetry. In this gauging procedure, the group extension structure determines how gauge charges transform under the quotient symmetry. Different decorations translate into patterns of ``fractionalization'' of the quotient symmetry on gauge flux excitations. When the differentials are nontrivial (i.e. the decoration is inconsistent), however, the gauging is also obstructed. In other words, the corresponding fractionalization class is not compactible with the prescribed symmetry action on gauge charges. A well-known special case of the obstructions is the 't Hooft anomaly, but there are other kinds of obstructions beyond 't Hooft anomaly. Using the LHS spectral sequence, we revisit the classification of symmetry-enriched Abelian gauge theory in 2D and showed that the various obstructions can be understood within the framework established in Ref. [\onlinecite{SET}]. We also consider symmetry-enriched U(1) gauge theory in 3D, recovering the classification in Ref. [\onlinecite{NingU1SL}] and furthermore identifying SPT stacking trivialization missed in previous works. \section{Domain wall decoration and spectral sequence} \label{sec:ddw} In this section we explain how the decorated domain wall construction naturally leads to a spectral sequence description of the SPT classification~\cite{Gaiotto_2019, Xiong_2018}. A brief introduction to spectral sequence can be found in Appendix \ref{sec:ss}. We will consider a finite, unitary symmetry group $G$. Here we remark that $G$ always refers to a ``bosonic'' symmetry, in the sense that a local bosonic order parameter can be defined and therefore the symmetry can be broken, either spontaneously or explicitly. As briefly explained in the introduction, an SPT wavefunction can be written as a superposition of all possible $G$ domain wall configurations. Typically, this is achieved by Hamiltonian terms that fluctuate the domain walls locally. In order for the superposition to be a SRE state, we postulate that the following conditions must be true in a symmetry-breaking state: \begin{enumerate} \item For any symmetry-breaking pattern (i.e. a configuration of domain walls), the system is fully gapped and short-range entangled, while preserving any remaining global symmetry. \item Symmetry-breaking states corresponding to different domain wall configurations can be smoothly connected. In other words, two such states can be transformed into each other by a constant-depth local unitary circuit preserving any remaining symmetry. This is necessary to ensure that a parent Hamiltonian can exist. \end{enumerate} The two conditions are clearly necessary to obtain a SRE state after the symmetry is restored by fluctuating domain walls. They are implicit in previous constructions of fixed-point wavefunctions, see for example Ref. \onlinecite{WangGu2018}. Here our focus will be on general structures, independent of any particular model construction. We will now assume that there is an unbroken symmetry group $A$ (which could be trivial). Denote the equivalence classes of SRE phases with symmetry group $A$ (and possibly with other conditions, such as fermionic/bosonic) in $D$ dimensions by $h^{D+1}(A)$. In other words, any SRE state with symmetry $A$ (obeying whatever additional conditions imposed) is associated to a unique element in $h^{D+1}(A)$. For this section we do not need more details about $h^{D+1}(A)$, besides that it is always a discrete Abelian group, with multiplication given by stacking. For brevity we will at times just write $h^{\sbullet}$ instead of $h^{\sbullet}(A)$, since the group $A$ does not play any role in the following discussions. We will denote by $\mathcal{C}^p[G, h^{\sbullet}], \mathcal{Z}^p[G, h^{\sbullet}]$ the $h^{\sbullet}$-valued $p$-cochains and $p$-cocycles of $G$, and $\mathcal{H}^p[G, h^{\sbullet}]$ the cohomology group. Below we describe the decorated domain wall construction in $D=2$ as an illustration. Our presentation is schematic and the main purpose is to demonstrate how the physical data can be naturally organized mathematically into a spectral sequence (see Appendix.\ref{sec:ss} for brief introduction). The computational details are discussed in the next section and also in Appendix \ref{sec:LHS1234}. To describe the decorated domain wall states, we start from the ``top'' level, where there is no domain wall, but the $G$ symmetry is already broken. There are $|G|$ different ways that the symmetry can be broken, related to each other by $G$ symmetry transformations. For each type of domain, the ground state wavefunction belongs to an SPT phase with the unbroken symmetry $A$, the equivalence class of which is given by an element $\omega_{0,3}\in h^{3}$ [see Fig.~\ref{fig:2D}(a)]. It is clear that in order to have a consistent decorated domain wall wavefunction, $\omega_{0,3}$ must be ``invariant'' under the natural $G$ action on $h^{3}$ (e.g. when $G$ has a nontrivial action on $A$), so that different domains belong to the same SPT phase. As the actual wavefunctions are generally different in different domains, the difference between the states before and after the symmetry action $(\delta_1 \omega_{0,3}) (\mb g)={}^{\mb g}(\omega_{0,3})/\omega_{0,3}$ should belong to the trivial class in $h^3$. So we have the first obstruction \begin{align} O_{1,3} &= \delta_1\omega_{0,3} \in \mathcal{Z}^{1}[G,h^3], \end{align} which is the $G$-invariant condition for the 2D $A$-SPT phases. Here $\delta_1$ is the usual coboundary operator acting on the $G$ elements. This obstruction condition implies that if $O_{1,3}$ is nontrivial, the 1D domain wall between two different domains has protected gapless modes, or spontaneously breaks $A$. In contrast, the vanishing of this obstruction ensures that 1D domain walls can be fully gapped out without breaking $A$. Next we consider 1D domain walls between different domains, labeled by elements of the symmetry group $G$. They can be viewed as (gapped) boundaries between two SPT states differed from each other by $G$ symmetry action [see Fig.~\ref{fig:2D}(a)]. The boundary wavefunction along the $\mb g$ DW is determined by the two 2D $A$-SPT states up to a 1D $A$-SPT phase, and will be denoted by $\omega_{1,2}(\mb g)$ [see red lines in Fig.~\ref{fig:2D}(b)]. In other words, the 1D $A$-SPT phase should be thought of as an torsor over different equivalence classes of the boundary state. Here ``equivalence'' is defined by adiabatic local deformation along the 1D domain wall. In the next step, we examine the simplest junction of domain walls, namely a tri-junction of $\mb{g,h}$ and $\mb{gh}$ domain walls depicted in Fig.~\ref{fig:2D}(b). { The tri-junction essentially provides} a way to check the consistency of group actions on the $A$-SPT states. The two sides of the junction are basically the same domain configuration, but a codimension-$1$ operator is applied to split the $\mb{gh}$ DW to $\mb{g}$ and $\mb{h}$ DWs. This codimension-$1$ operator can not create an nontrivial 1D SPT state, since such a configuration should be smoothly connected to a single domain wall by local adiabatic deformation. Equivalently, the choice of the wavefunctions of the 1D domain walls determines whether the junction harbors a 0D zero mode protected by $A$ or not. Denote this 1D SPT phase by $O_{2,2}(\mb{g,h})\in h^2$. Since the wavefuntion of a $\mb{g}$ domain wall is ambiguous up to 1D $A$-SPT $\omega_{1,2}(\mb{g})\in h^{2}$, $O_{2,2}$ is only determined up to \begin{equation} (\delta_1\omega_{1,2})(\mb{g,h})=\frac{\omega_{1,2}(\mb{g})\omega_{1,2}(\mb{h})}{\omega_{1,2}(\mb{gh})}. \label{} \end{equation} In other words, $O_{2,2}$ up to $\delta_1\omega_{1,2}$ is a function of $\omega_{0,3}$. So we have the following expression \begin{align} O_{2,2} &= (\delta_2\omega_{0,3}) (\delta_1\omega_{1,2}) \in \mathcal{Z}^{2}[G,h^2]. \label{eqn:O22} \end{align} This is the second obstruction in the decorations. Physically $O_{2,2}=\mathds{1}$ imposes the constraint that the DW tri-junction can not harbor any non-trivial 0D boundary states of 1D A-SPT in $h^2$, otherwise the configuration is not fully gapped. Here we introduce the notation $\delta_2\omega_{0,3}$, which formally represents the contribution to the obstruction on the codimension-$2$ defect junction from the decoration on the codimension-0 defect (i.e. the 2D $A$-SPT phase). The actual form of $\delta_2\omega_{0,3}$ depends on the nature of $h^{\sbullet}$ (e.g. fermionic or bosonic), and will be discussed in Sec. \ref{sec:lhs}. An important remark follows: $\omega_{0,3}$ is a 3-cocycle in $h^3$ as it is the top-dimension decoration. $\omega_{1,2}$ however is not necessarily a cocycle in ${{\mathcal{H}}}^1[G, h^2]$, since it describes the physical state on the domain wall. This is generally true for the other $\omega_{p,q}$'s with $p>0$. The only exception is that when $\omega_{0,3}$ is completely trivial, then $\delta_1\omega_{1,2}=1$ which means that $\omega_{1,2}\in \mathcal{Z}^1[G, h^2]$. On the other hand, even though $\omega_{1,2}$ itself is only a cochain in $\mathcal{C}^1[G, h^2]$, difference of two $\omega_{1,2}$'s both satisfying Eq. \eqref{eqn:O22} is a cocycle in $\mathcal{Z}^1[G, h^2]$. In other words, $\mathcal{Z}^1[G, h^2]$ (and actually, ${{\mathcal{H}}}^1[G, h^2]$ as $\mathcal{B}^1[G, h^2]$ can be deformed to a trivial decoration) forms a \emph{torsor} over different solutions of Eq. \eqref{eqn:O22}. Similar comments apply to other $\omega_{p,q}$ as well. \begin{figure}[t!] \centering \includegraphics[width=.9\columnwidth]{Fig_2D} \caption{Domain wall decorations and obstructions in 2D. Wavefunction $\omega_{i,3-i}$ is decorated on codimension-$i$ domain walls. (a) The obstruction $O_{1,3}$ indicates that the boundary between two $G$-domains should be gapped. (b) $O_{2,2}$ means that the DW tri-junction does not harbor any projective representation of $A$. (c) $O_{3,1}$ imposes the constraint of $A$-charge conservation under the $F$ move. There is a final layer of obstruction $O_{4,0}$ from the pentagon equation of DW fusions that is not shown in the figure.} \label{fig:2D} \end{figure} Once $O_{2,2}$ vanishes, we can move to the next level and consider the two configurations depicted in Fig.~\ref{fig:2D}(c). They are exactly the same away from the local patch, so they can at most differ by a $0$D SRE state, classified by $h^{1}$. To go from one configuration to the other one has to apply a ``F move'', which is a locality-preserving, $A$-symmetric unitary operator creating an additional SPT state. For now denote this SPT phase by $O_{3,1}(\mb{g,h,k})\in h^{1}$. Since the wavefuntion of the tri-junction of $\mb{g,h}$ domain wall is ambiguous up to 0D state $\omega_{2,1}(\mb{g,h})$ [see green dots in Fig.~\ref{fig:2D}(c)], $O_{3,1}$ is only determined up to \begin{equation} (\delta_1\omega_{2,1})(\mb{g,h,k})=\frac{\omega_{2,1}(\mb{h,k})\omega_{2,1}(\mb{g,hk})}{\omega_{2,1}(\mb{gh,k})\omega_{2,1}(\mb{g,h})}. \label{} \end{equation} We can split $O_{3,1}$ into terms from $\omega_{0,3}, \omega_{1,2}$ and $\omega_{2,1}$ as \begin{equation} O_{3,1} = (\delta_3\omega_{0,3}) (\delta_2\omega_{1,2}) (\delta_1\omega_{2,1}) \in \mathcal{Z}^{3}[G,h^1]. \label{eqn:O31} \end{equation} The physical meaning of this obstruction is that the F move does not change the 0D SRE state in $h^{1}(A)$. Again here we will keep the actual form of $\delta_3, \delta_2, \cdots$ implicit, except that $\delta_1$ is the usual coboundary operator. Finally, if $O_{3,1}$ vanishes, then the only known information is a phase factor in the F move (since we have gotten down to 0D quantum states) and the last consistency condition is that the F moves need to satisfy the pentagon equation. We can check whether the pentagon equation holds given all the previous decorations, and it can be violated at most by a phase factor, denoted by $O_{4,0}(\mb{g,h,k,l})$. Again we can express $O_{4,0}$ as the following form \begin{equation} O_{4,0} = (\delta_4\omega_{0,3}) (\delta_3\omega_{1,2}) (\delta_2\omega_{2,1}) \in \mathcal{Z}^{4}[G,h^0], \label{eqn:O40} \end{equation} originated from different decoration layers. Here $h^0$ is just $\mathrm{U}(1)$. The F move itself can acquire an additional phase factor $\omega_{3,0}$. If $O_{4,0} = \delta_1\omega_{3,0}$ is a coboundary of $G$, then we can absorb the F move by another phase factor $\omega_{3,0}$ such that the pentagon equation is satisfied. At this point we have exhausted all consistency conditions for the decorations. Physically, it means that the two conditions for a consistent decoration laied out in the beginning of this section are satisfied at all levels, and one can in principle define a SRE ground-state wavefunction for the corresponding SPT state using the data $\omega_{p,3-p}$ for $p=0,1,2,3$. Mathematically, the consistency conditions are the four coupled equations $O_{p,4-p}=\mathds{1}$ for $p=1,2,3,4$. For most applications, we do not need the full solutions of these equations. Instead, one may be interested the particular obstruction function arising from a certain layer of decoration, e.g. $\omega_{1,2}$. In that case, one can assume that the $\omega_{0,3}=\mathds{1}$, so $\omega_{1,2}$ is an actual cocycle in ${{\mathcal{H}}}^1[G, h^2]$. For clarify, we will rename it to $\nu_{1,2}$. Then one checks Eq. \eqref{eqn:O31} to see whether there is a $\omega_{2,1}\in \mathcal{C}^2[G, h^1]$ satisfying the condition. The solvability of Eq. \eqref{eqn:O31} is equivalent to the vanishing of $\delta_2\nu_{1,2}$ as a cohomology class in ${{\mathcal{H}}}^3[G, h^1]$. Therefore we define a \emph{cocycle-level} map, called a \emph{differential}: \begin{equation} d_2: {{\mathcal{H}}}^1[G,h^2]\rightarrow {{\mathcal{H}}}^3[G, h^1]. \label{} \end{equation} In this case, $d_2$ is basically the same as $\delta_2$, just regarded as a map between cohomology classes~\footnote{This is consistent because $\delta_2$ must commute with $\delta_1$.}, but this is not true for higher differentials. If $d_2$ of $\nu_{1,2}$ vanishes, one can find at least one solution $\omega_{2,1}^0$ for Eq. \eqref{eqn:O31}. Then one can plug $\nu_{1,2}$ and $\omega_{2,1}^0$ into Eq. \eqref{eqn:O40} to compute the obstruction function. This way one may attempt to define another differential using this procedure: \begin{equation} d_3: {{\mathcal{H}}}^1[G, h^2]\rightarrow {{\mathcal{H}}}^4[G, h^0], \label{} \end{equation} which increases the degree of the cohomology group by $3$. However, the $d_3$ map is not well-defined yet: $\omega_{2,1}^0$ is one particular solution, which is determined up to a cocycle $\nu_{2,1}$ in $\mathcal{Z}^2[G, h^1]$. If we shift $\omega_{2,1}^0\rightarrow \omega_{2,1}^0\nu_{2,1}$, the obstruction function changes by $d_2\nu_{2,1}$. In other words, $d_3$ is defined only up to $d_2\nu_{2,1}$. To resolve this ambiguity, we can restrict to a subgroup of ${{\mathcal{H}}}^2[G, h^1]$ with vanishing $d_2$. In that case, the $d_3$ map becomes well-defined. Physically the restriction makes sense, as to discuss $d_3$ we have to assume that $d_2\nu_{1,2}$ vanishes, so naturally the same assumption should be applied to other layers of decorations as well. More formally, we define $E_3^{p,q}\subset {{\mathcal{H}}}^p[G, h^{q}]$ as the kernel of $d_2$, and $d_3$ is only defined on $E_3$. In this sense, ${{\mathcal{H}}}^p[G, h^q]$ should be called $E_2^{p,q}$ as they are the kernel of $d_1\equiv\delta_1$. By the same reasoning one can define $E_n^{p,q}$ for higher $n$ iteratively. The collection of $E_n^{p,q}$ is called the $E_n$ page. The differential map $d_n$ is only defined on the $E_n$ page, and one can further show that its image automatically sits in the $E_n$ page (of one dimension higher) as well, so we have \begin{equation} d_n: E_n^{p,q}\rightarrow E_n^{p+n,q-n+1}. \label{} \end{equation} The collection of all $E_n^{p,q}$ and the differentials that map between them in fact forms an Atiyah-Hirzebruch spectral sequence. Heuristically, as the ``page number'' $n$ increases, more and more obstructions are required to vanish and one obtains better and better ``approximations'' of the actual SPT states. Eventually, when all obstructions vanish, we arrive at the ``$E_\infty$'' page, which is the collection of all physical SPT states. A concrete example of solving the consistency conditions to compute the differential maps in (2+1)d when $h^{\sbullet}$ is the group cohomology of $A=\mathbb{Z}_n$ can be found in Appendix \ref{app:example}. Now let us briefly discuss the general structure of a domain wall decoration construction in $D$ dimensions. Starting from the top dimension $p=0$, we consider topological junctions of $G$ domain walls of increasing codimension successively. The fact that the $p$-junction should be fully gapped requires that the obstruction function \begin{equation} O_{p,D+2-p} = \delta_p \omega_{0,D+1}\delta_{p-1}\omega_{1,D}\cdots \delta_1\omega_{p-1,D+2-p} \label{} \end{equation} must vanish. In other words, any gapless modes resulting from decorations on $q$-junctions with $q\leq p-1$ must cancel out on the $p$-junction. This procedure formally continues up to $p=D+2$. Note that for $p=D+2$, $\omega_{D+1,0}$ is a phase factor associated with a certain rearrangement of topological junctions, generalizing the F move. The condition follows from checking the $(D+1)$-cocycle condition for these rearranging moves. We can similarly define $d_2$ as basically the $\delta_2$ operator. To define higher differentials, first one needs to define the $E_n$ page as the subgroup of $E_{n-1}$ page whose images under $d_{n-1}$ vanish, where $E_2^{p,q}={{\mathcal{H}}}^p[G, h^q]$. $d_n\omega_{p,q}$ is then defined as the obstruction function $O_{n+p, D+2-n-p}$ arising from $\omega_{p,q}$ with all lower-codimension decorations set to $\mathds{1}$. The operation is well-defined on the $E_n$ page. So far we have focused on the construction of obstruction-free domain wall decorations. We have not yet considered the equally important question of whether the decorated domain wall state is trivial or not. This will be addressed in Sec. \ref{sec:aspt}. Below we discuss two spectral sequences relevant for the classification of SPT phases. \subsection{Lyndon-Hochschild-Serre spectral sequence} \label{sec:lhs} Suppose that $h^{D+1}(A)$ is the group-cohomology bosonic SPT phases, i.e. \begin{equation} h^{D+1}(A)={{\mathcal{H}}}^{D+2}[A, \mathbb{Z}]. \label{} \end{equation} In the following we assume the symmetry group is either finite, or a compact Lie group. In this case, we have ${{\mathcal{H}}}^*[A, \mathbb{R}]=0$ and thus ${{\mathcal{H}}}^{D+2}[A, \mathbb{Z}]={{\mathcal{H}}}^{D+1}[A, \mathrm{U}(1)]$. The same is true for $A$ a free Abelian group. The Atiyah-Hirzebruch here reduces to the so-called Lyndon-Hochschild-Serre (LHS) spectral sequence~\cite{Lyndon, HS} for group cohomology. More concretely, denote by $\tilde{G}$ the total symmetry group. Because $A$ is a normal subgroup, $\tilde{G}$ fits in the short exact sequence: \begin{equation} 1\rightarrow A\rightarrow \tilde{G}\rightarrow G\rightarrow 1. \label{eqn:extensionG} \end{equation} In general, the group extension can be specified by two pieces of data: a homomorphism $\rho$ from $G$ to the outer automorphism group $\mathrm{Out}(A)$ which satisfies a certain-obstruction vanishing condition, and a torsor $\nu$ over ${{\mathcal{H}}}^2_\rho[G, Z(A)]$ where $Z(A)$ is the center of $A$. When $A$ is Abelian, there is no distinction between $\mathrm{Out}(A)$ and $\mathrm{Aut}(A)$, and the obstruction always vanishes canonically. The multiplication rule in $\tilde G$ is given by \begin{equation} a_\mb{g}\times b_\mb{h} = [a \cdot \rho(\mb{g})(b) \cdot \nu(\mb{g,h})]_{\mb{gh}}, \end{equation} with $a,b\in A$ and $\mb{g,h}\in G$. When $\tilde{G}=A\times G$, one has the K\"unneth formula: \begin{equation} {{\mathcal{H}}}^{d}[G, \mathrm{U}(1)]=\bigotimes_{\substack{p+q=d\\ p,q\geq 0}} {{\mathcal{H}}}^{p}[G, {{\mathcal{H}}}^{q}[A, \mathrm{U}(1)]]. \label{} \end{equation} This is basically the $E_2$ page of the LHS spectral sequence, and all differentials are trivial. In general, however, explicit expressions for differentials appear to be unknown in literature except $d_2$, though partial results have been obtained in previous works~\cite{kapustin2014, Tachikawa_2020, NingU1SL}. In Appendix \ref{sec:LHS1234} we provide a complete and explicit description of the LHS spectral sequence for ${{\mathcal{H}}}^{d}[\tilde{G}, \mathrm{U}(1)]$ for $d=1,2,3,4$, at the cochain level, and present the conjectured general structure for higher $d$. We find that a general $d$-cocycle in $\mathcal{Z}^d[\tilde{G}, \mathrm{U}(1)]$ can always be constructed from cochains $F_{p,d-p}\in E_0^{p,d-p}$ for $0\le p\le d$ at different positions of the spectral sequence. These cochains satisfy the following obstruction-vanishing conditions: \begin{align} E_0^{0,d+1}&:\quad (\delta_0F_{0,d}) = 1,\\ E_0^{1,d}&:\quad (\delta_1F_{0,d})(\delta_0F_{1,d-1}) = 1,\\ E_0^{2,d-1}&:\quad (\delta_2F_{0,d})(\delta_1F_{1,d-1})(\delta_0F_{2,d-2}) = 1,\\ &\quad\quad\vdots\\ E_0^{d,1}&:\quad (\delta_dF_{0,d})(\delta_{d-1}F_{1,d-1})\cdots(\delta_0F_{d,0}) = 1,\\ E_0^{d+1,0}&:\quad (\delta_{d+1}F_{0,d})(\delta_{d}F_{1,d-1})\cdots(\delta_1F_{d,0}) = 1. \end{align} Here $\delta_i$'s are the cochain-level differentials, whose explicit expressions can be found in Eq.~\eqref{diFpq} of Appendix \ref{App:diff}. For example, the zeroth and the first differentials $\delta_0=d_A$ and $\delta_1=\pm d_G$ are just the usual differentials with respect to $A$ and $G$, respectively. Higher differentials are more complicated, so we refer the interested readers to Appendix \ref{sec:LHS1234} for details. As an example, the second differential has the following general form \begin{align} &\quad(\delta_2F_{p,q})(a_1,...,a_{q-1},\mb{g}_1,...,\mb{g}_{p+2})\\\nonumber &=\left[ \iota_{(-1)^q\nu(\mb{g_1,g_2})} \left({}^{1_{\mb{g}_1\mb{g}_2}}F_{p,q}\right)^{\cdot,...,\cdot,1_{\mb{g}_3},...,1_{\mb{g}_{p+2}}} \right]^{a_1,...,a_{q-1}}. \end{align} Here, the notation ${}^{\ag{1}{g}}\!F$ is defined in Eq.~\eqref{hF}. And $\iota$ is the slant product operation, the definition of which can be found in Eq.~\eqref{eqn:slant}. If $\delta_2$ lands on the $p$ axis (i.e., $q=1$), the above equation is simplified to \begin{align} \quad(\delta_2F_{p,1})(\mb{g}_1,...,\mb{g}_{p+2}) =\frac{1}{\left({}^{1_{\mb{g}_1\mb{g}_2}}F_{p,q}\right)^{\nu(\mb{g}_1,\mb{g}_2),1_{\mb{g}_3},...,1_{\mb{g}_{p+2}}}}, \end{align} which is basically Theorem 4 of the original LHS spectral sequence paper Ref.~\onlinecite{HS}. These equations can be thought of as the concrete mathematical representations of the physical obstruction-vanishing conditions discussed in Sec. \ref{sec:ddw} for domain wall decorations. With these cochain-level equations, one can further extract the ``cocycle-level'' differentials $d_i$'s. As a special case, the second cocycle-level differential $d_2$ is in fact equal to $\delta_2$ when viewed as operations on cocycles of the $E_2$ page. For an example of computations of higher differentials, in Appendix \ref{app:example} we derive all the differentials for $A=\mathbb{Z}_n$ and $d=3$ when the extension is central. In Appendix \ref{sec:lyndon} we provide an algorithm to extract the cochains at different positions of the LHS spectral sequence from a general cocycle, and give explicit expressions in low dimensions. An easy collorary of our results is that when $A$ is Abelian and the short exact sequence splits (i.e. $\tilde{G}=A\rtimes G$), every cochain-level differentials $\delta_n$ with $n\geq 2$ vanish and $E_\infty=E_2$. \subsection{Spectral sequence for fermionic SPT phases} Let us give another example of Atiyah-Hirzebruch spectral sequence in the classification of fermionic SPT phases~\cite{Kapustin_2015, Thorngren2018, WangGuPRX2020}. The symmetry group of a fermionic system fits into the following short exact sequence: \begin{equation} 1\rightarrow \mathbb{Z}_2^f\rightarrow G\rightarrow G_b\rightarrow 1 \label{} \end{equation} where $G_b$ is the symmetry group that act on bosonic operators, and $\mathbb{Z}_2^f$ is the conservation of the total fermion parity. A general fermionic SPT state can be constructed from decorating fermionic invertible topological phases on $G_b$ domain walls. Therefore we set $h^{D+1}$ to be the group of fermionic invertible phases, denoted by $h_F^{D+1}$. Up to spatial dimension $D=4$, they are given by \begin{center} \begin{tabular}{|c|ccccc|} \hline $D$ & $0$ & $1$ & $2$ & $3$ & $4$\\ \hline $h_F^{D+1}$ & $\mathbb{Z}_2$ & $\mathbb{Z}_2$ & $\mathbb{Z}$ & $\mathbb{Z}_1$ & $\mathbb{Z}_1$\\ \hline \end{tabular} \end{center} Here the $0$-dimensional SPT state is just a single fermion. In 1D the nontrivial phase is the Majorana chain, and in 2D the generating one is a $(p_x+ip_y)$ superconductor. Then decoration on codimension-$p$ junctions of $G_b$ is given by ${{\mathcal{H}}}^{p}[G_b, h^{D+1-p}_F]$. Explicit expressions for differentials up to $D=3$ are obtained in Ref. \onlinecite{WangGu2018} (see Ref. \onlinecite{Thorngren2018} for related results). Interestingly, in this case even when the extension is trivial, i.e. $G=G_b\times \mathbb{Z}_2^f$, the differentials are still nontrivial \cite{WangGu2018}. \section{Applications} \subsection{Anomalous SPT phases} \label{sec:aspt} As discussed in Sec. \ref{sec:ddw}, obstructions for a consistent domain wall decoration are represented by cohomology classes in ${{\mathcal{H}}}^{p}[G, h^{D+2-p}]$ for codimension-$p$ junctions. For example, we have seen that the top level decoration, i.e. $\omega_0$, may cause an obstruction represented by a class $d_2\omega_0\in {{\mathcal{H}}}^2[G, h^{D}(A)]$. $d_2\omega_0$ can be understood as boundary states of a $(D-1)$-dimensional SPT phase decorated on a codimension-2 junction in a $(D+1)$-dimensional ``bulk''. In other words, $d_2\omega_0$ gives a DW decoration of the bulk. In Fig. \ref{fig:aspt} we illustrated two examples, for $E_2^{2,2}$ and $E_2^{3,1}$, where obstructions in DW decoration on the surface are ``compensated'' by decorations in the bulk, so that the boundary can be made non-degenerate. For example, in Fig. \ref{fig:aspt}(a), the boundary decoration is obstructed on the 0D junction, which is neutralized by a compensating 1D SPT state in the bulk. Similarly, in Fig. \ref{fig:aspt}(b), the boundary decoration can be viewed as stacking the two sides of the F move together to eliminate open DWs, and the obstruction, a 0D SPT state characterized by $d_2\omega_1$, is neutralized by the corresponding bulk decoration. In other words, these ``obstructed'' decorations can only be consistently realized as boundary of a higher-dimensional SPT state. Thus they can be thought of as ``anomalous'' SPT phases. On the other hand, the bulk in this case must be a trivial SPT phase, since its boundary can form a SRE state preserving all symmetries. In summary, any nontrivial obstruction class in the decorated DW construction give rise to an anomalous SPT phase. Mathematically, these obstructions are computed using the same differential maps of the spectral sequence. One may wonder what if one adiabatically deforms the bulk to a trivial product state, which should always be possible because the bulk is topologically trivial. Naively once the bulk is completely disentangled, one is left with the boundary SRE state, contradicting the claimed anomalous nature. The resolution is that the disentangling transformation necessarily removes the boundary SRE state as well. Let us describe the required transformation for a decorated DW construction of a 1D SPT state in $D=1$, with an obstruction class given by $d_2: E_2^{0,2}\rightarrow E_2^{2,1}$. The splitting of $\mb{gh}$ DW into $\mb{g}$ and $\mb{h}$ DWs produces an additional 0D SPT state (i.e. a ``charge''), which is cancelled by the codimension-2 decoration in the 2D bulk. This bulk-boundary relation immediately suggests the following disentangling transformation: first fixing a DW configuration in the bulk. Inside the $\mb{h}$ domain, we adiabatically ``pump'' out a 1D SPT state parametrized by $^{\mb{h}}\omega_0$, and bring the 1D SPT close to the DWs. It is easy to see that along each DW the two 1D SPT states can be annihilated, which also neutralizes the 0D SPT state at the junction. On a disk, the pumping can be intuitively thought as ``pushing'' the edge SPT states inside, towards the junction. For each bulk DW configuration, this kind of pumping can be performed to remove the decoration and it should be evident that the process preserves the residual symmetry. We believe that the pumping can be realized globally by a symmetric adiabatic transformation on the whole bulk wavefunction, even though we do not know how to explicitly write it down. This discussion also makes it clear that the differential maps always have dual interpretations: consider $d_n: E_n^{p,q}\rightarrow E_n^{p+n, q-n+1}$. On one hand, it is the obstruction map for the SPT phase described by the DW decoration in $E_n^{p,q}$. Non-anomalous decorations must belong to the kernel of the differential map. On the other hand, it is also the trivialization map for SPT phases described by DW decoration in $E_n^{p+n, q-n+1}$ (in one dimension higher), and the image of the map gives trivial SPT phase. \begin{figure}[t!] \centering \includegraphics[width=\columnwidth]{aspt} \caption{Illustration of $d_2$ obstructions and anomalous SPT phase.} \label{fig:aspt} \end{figure} An alternative perspective on ASPT is to discard the bulk altogether and view the ``anomaly'' as the property of a non-on-site symmetry. To be concrete, suppose we have a $D$-dimensional bosonic lattice model, with a global symmetry $\tilde{G}$. Denote the Hilbert space of the lattice model by $\mathcal{H}$. The symmetry group is represented on the Hilbert space by locality-preserving unitary operators $\{U_\mb{g}\}_{\mb{g}\in \tilde{G}}$, which satisfy $U_\mb{g}U_\mb{g}=U_\mb{gh}$. For simplicity, let us assume that $U_\mb{g}$ is a local unitary circuit with a constant depth. Generally, it is believed that one can associate $\{U_\mb{g}\}_{\mb{g}\in \tilde{G}}$ uniquely to an element in ${{\mathcal{H}}}^{D+2}[\tilde{G}, \mathrm{U}(1)]$~\cite{ElseNayak}. When this element is nontrivial, the symmetry has a 't Hooft anomaly, forbidding the existence of any $\tilde{G}$-symmetry preserving SRE state in this Hilbert space. Now suppose $\tilde{G}$ is an extension of $G$ by $A$, where $A$ is the normal subgroup. If the anomaly class of $U_\mb{g}$'s is trivial, then a symmetry-preserving SRE state is possible. Given the group extension structure, it should be possible to locate the anomaly class in the $E_2$ page of LHS spectral sequence, i.e. the ``bulk'' can have a certain domain wall decoration pattern, trivialized by differentials. In that case, the $D$-dimensional system must realize an ASPT phase, whose differential gives the bulk domain wall decoration. Below we present a concrete example in (1+1)d, where a non-on-site internal symmetry enforces the system to realize an ASPT phase. Very similar phenomena arise for spatial symmetries as SPT-LSM theorems discussed below in Sec. \ref{sec:SPTLSM}. We consider a spin chain with $A=\mathbb{Z}_2=\{1,a\}$ and $G=\mathbb{Z}_2=\{1, g\}$ in (1+1)d. Suppose that $g^2=a$, so $G$ is extended nontrivially by $\mathbb{Z}_2$ to $\mathbb{Z}_4$. There are no nontrivial SPT phase in (1+1)d with $\mathbb{Z}_4$ symmetry, but there is an anomalous one. In terms of the LHS spectral sequence, the only nontrivial term is $E_2^{1,1}={{\mathcal{H}}}^1[G, {{\mathcal{H}}}^1[A, \mathrm{U}(1)]]=\mathbb{Z}_2$. Denote the 1-cocycle by $F^{a,g}=\pm 1$ (the other ones are normalized to 1). The $d_2$ differential gives a 3-cocycle in ${{\mathcal{H}}}^3[G, \mathrm{U}(1)]$: \begin{equation} F^{g,g,g}=F^{\nu(g,g),g}=F^{a,g}. \label{} \end{equation} Thus the nontrivial class with $F^{a,g}=-1$ is $d_2$ obstructed. Let us now turn to the lattice model. Consider a 1D chain with Ising spins $\sigma_j$ on sites and $\tau_{j+\frac12}$ on bonds. The symmetry group $A$ is generated by $U_a=\prod_{j}\tau_{j+\frac12}^x$, and the $G$ symmetry is generated by the following unitary operator: \begin{equation} U_g=\prod_j \sigma_j^x\cdot e^{\frac{i\pi}{2}\sum_j \frac{1-\sigma_j^z\sigma_{j+1}^z}{2}}\cdot e^{\frac{\pi i}{4}\sum_j(\tau^x_{j+1/2}-1)}. \label{} \end{equation} Under periodic boundary condition, it is easy to see that $U_g^2=U_a$, so we have the desired group extension structure. Besides the last factor, $U_g$ is essentially the anomalous $\mathbb{Z}_2$ symmetry transformation on the boundary of a 2D $\mathbb{Z}_2$ SPT phase~\cite{Levin_2012}. To realize a SRE state preserving both $U_g$ and $U_a$, first we define a subspace by the following constraint: \begin{equation} \tau_{j+\frac12}^x = \sigma_{j}^z\sigma_{j+1}^z. \label{eqn:Z2constraint} \end{equation} Physically, this is the $U_a$ charge decoration on $U_g$ domain wall, corresponding to a nontrivial class in ${{\mathcal{H}}}^1[\mathbb{Z}_2, {{\mathcal{H}}}^1[\mathbb{Z}_2,\mathrm{U}(1)]]$. In this subspace, the symmetry $U_g$ simplifies to $\prod_j \sigma_j^x$. Define the projector to the subspace as $P$: \begin{equation} P=\prod_j \frac{1+\tau_{j+\frac12}^x\sigma_j^z\sigma^z_{j+1}}{2}. \label{} \end{equation} We choose the Hamiltonian to be \begin{equation} H=-\sum_{j}\tau_{j-\frac12}^z \sigma_j^x\tau_{j+\frac12}^zP. \label{} \end{equation} The term is invariant under $U_a$, and under $U_g$ as well in the projected space. The ground state satisfies \begin{equation} \tau_{j+\frac12}^x\sigma_{j}^z\sigma_{j+1}^z=\tau_{j-\frac12}^z \sigma_j^x\tau_{j+\frac12}^z=1. \label{} \end{equation} We recognize that with periodic boundary condition, the ground state is the well-known 1D cluster state. It is often described as a 1D SPT state protected by the $\mathbb{Z}_2\times\mathbb{Z}_2$ symmetry group generated by $U_a$ and $U_g'=\prod_j \sigma_{j}^x$. On an open chain starting with site 1 and ending at site $L$, one finds zero mode operators $\sigma_1^z$ and $\sigma_1^x\tau_{\frac32}^z$ localized at the site $1$, which anticommute with each other and span a two-dimensional projective representation of $\mathbb{Z}_2\times\mathbb{Z}_2$. In our case, we define the restricted symmetry transformations as follows: \begin{equation} \begin{gathered} \tilde{U}_a=\sigma_1^z \prod_{j=0}^{L-1}\tau_{j+1/2}^x \sigma_L^z,\\ \tilde{U}_g=\prod_{j=1}^L \sigma_j^x\cdot e^{\frac{i\pi}{2}\sum_{j=1}^{L-1} \frac{1-\sigma_j^z\sigma_{j+1}^z}{2}}\cdot e^{\frac{\pi i}{4}\sum_{j=1}^{L-1}(\tau^x_{j+1/2}-1)}. \end{gathered} \label{} \end{equation} Notice that we introduce two additional $\sigma^z$ operators at the ends of the chain in $\tilde{U}_a$, since the restriction is ambiguous up to local operators at the ends. The restricted symmetry operators satisfy all the group relations: $\left( \tilde{U}_a \right)^2=\mathds{1}, \left( \tilde{U}_g \right)^2=\tilde{U}_a$. With these symmetry transformations, both $\sigma_1^z$ and $\sigma_1^x\tau_{3/2}^z$ are forbidden by $\tilde{U}_g$, and therefore the ground state degeneracy is protected. Intuitively, the constraint in Eq. \eqref{eqn:Z2constraint} must be enforced to get rid of the non-on-site part of $U_g$ in a suitable subspace, and therefore a non-degenerate ground state must be an anomalous SPT phase. The constraint basically binds $\tau$ charges to $\sigma$ domain walls. Without this particular decoration, the system either breaks the symmetry or forms a gapless state. For example, one may let the $\tau$ spins to form an Ising paramagnet, i.e. $\tau_{j+1/2}^x=1$ for every $j$ by adding a large on-site magnetic field $-h\sum_j \tau_{j+1/2}^x$. Within this low-energy subspace, the $U_g$ transformation effectively generates a $\mathbb{Z}_2$ group with 't Hooft anomaly, which is an example of the emergent anomaly discussed in Ref. [\onlinecite{Metlitski_2018}]. The $\sigma$ spins can form a Luttinger liquid preserving the effective $\mathbb{Z}_2$ symmetry. \subsection{Symmetry-enriched gauge theory} In a consistent decorated DW state, one can gauge the symmetry group $A$ to arrive at a $A$ gauge theory. The global symmetry becomes $G$ after $A$ is gauged, so the result is a $G$-enriched $A$ gauge theory~\cite{MesarosPRB2013, Tachikawa_2020}. This is believed to be the most general construction of a $A$ gauge theory enriched by $G$ symmetry, as long as the enrichment is compactible with the gauge structure. Let us examine the symmetry-enriched phenomena in more details. In the gauge theory, there are two types of fundamental excitations: gauge charges, which are point-like excitations labeled by irreducible representations of the gauge group, and gauge fluxes, $(D-2)$-dimensional objects labeled by conjugacy classes. The group extension determines how the gauge charges transform under the global symmetry $G$: the automorphism of $G$ on $A$ specifies how different types of gauge charges are permuted by $G$ (and at the same time determine how different kinds of fluxes are permuted), and the $2$-cocycle indicates that gauge charges transform projectively under $G$. On the other hand, each $E_2^{p,q}$ with $p\neq 0, q\neq 0$ describes a particular way that the $G$ symmetry acts on gauge flux excitations, including changing the topological type of the excitations, or symmetry transformation being ``fractionalized'' on the excitations. The latter is most familiar in 2D, and similar phenomena have also been investigated for loop excitations in 3D topological gauge theories~\cite{cheng2015symmetry, ChenPRB2016, ChenPRR2021}. Below in Sec. \ref{sec:segex} we will carefully examine the physical interpretations of these terms for Abelian gauge theories in (2+1)d and U(1) gauge theory in (3+1)d. The differentials $d_n: E_n^{p,D+1-p}\rightarrow E_n^{p+n,D+2-p-n}$ represent obstructions in gauging the symmetry group $A$. In other words, the gauged theory is ``obstructed'' (with the given action on gauge charges). More specifically, a non-vanishing obstruction typically implies that the symmetry action given by $E_2^{p,q}$ is incompactible with the underlying fusion and braiding structure of the topological order, when the group extension fixes how the symmetry acts on the charge sector. In addition, $d_n: E_n^{D+2-n, n-1}\rightarrow E_n^{D+2,0}$ leads to 't Hooft anomaly of the symmetry-enriched gauge theory. However there are a whole ladder of obstructions. We study the 2D case for general Abelian group $A$, and 3D case for $A=\mathrm{U}(1)$ below in Sec. \ref{sec:U1}, and unpack the physical meaning of the obstructions. On the other hand, if certain element in $E_n^{p,D+1-p}$ is in the image of the $d_n$ differential, according to the discussion in, the domain wall decoration after the symmetry is ungauged can be trivialized. As a result, the corresponding symmetry action on flux excitations in the gauged theory is also trivial. \section{Examples} In this section we discuss a few examples to illustrate the applications of LHS spectral sequence. \subsection{SPT-LSM theorems} \label{sec:SPTLSM} In recent years, the celebrated Lieb-Schultz-Mattis theorem has found many generalizations and refinements~\cite{Watanabe_2015, ChengPRX2016, PoPRL2017, ChengPRB2019, Lu_2020, Lu_LSMSPT, Yang_LSMSPT, ElsePRB2020}. A common theme underlying these developments is the realization that LSM theorem can be understood as the manifestation of 't Hooft anomaly of a particular class of crystalline SPT phases, the boundary of which hosts a projective representation per unit cell~\cite{ChengPRX2016, JianPRB2018, ChoPRB2017}. Here we will focus on a class of LSM theorems which allows SRE ground state~\cite{Lu_2020, Lu_LSMSPT, Yang_LSMSPT}, but nevertheless must be in a nontrivial SPT phase, thus referred to as ``SPT-LSM'' theorems. While these theorems have been understood in the theoretical framework of crystalline SPT phases~\cite{ElsePRB2019, Song_2020, shiozaki2018generalized}, here we will ``rederive'' some of the bosonic SPT-LSM theorems using the LHS spectral sequence. More specifically, we will treat the relevant crystalline symmetry formally as an internal one and identify the SPT-LSM theorems from a certain $E_2$ page corresponding to certain decoration with projective representations of the actual internal symmetry. This is motivated by the crystalline correspondence principle in Ref. [\onlinecite{ThorngrenPRX2018}], which establishes the equivalence between classification of topological phases with a crystalline symmetry group $G$ and that with an internal symmetry group $G$. Notice that however it is not clear whether the spectral sequence that classifies bosonic crystalline SPT phases~\cite{ElsePRB2019, Song_2020, shiozaki2018generalized} and the Atiyah-Hirzebruch spectral sequence studied here for internal symmetry are completely equivalent (at the level of spectrums, not just classifications). Therefore the computations performed below should be regarded as being conjectural, and suggest that the crystalline correspondence principle may actually extend to the whole spectral sequence. We will focus on SPT-LSM theorems with ``magnetic'' translation symmetry. Let us consider $G=\mathbb{Z}^D$, extended by an Abelian group $A$: \begin{align}\label{A_ZD} 1 \rightarrow A \rightarrow \tilde G \rightarrow \mathbb{Z}^D \rightarrow 1. \end{align} $G$ here represents the lattice translation symmetry in $D$-spatial dimensions, and $A$ is the internal symmetry group. When $D=2$ and $A=\mathrm{U}(1)$, $\tilde{G}$ is the usual magnetic translation symmetry group. To describe projective representation of $A$ per unit cell, we need to consider $E_2^{D,2} = {{\mathcal{H}}}^D[\mathbb{Z}^D,{{\mathcal{H}}}^2[A,\mathrm{U}(1)]]$. Roughly speaking, this is because a unit cell can be formally viewed as a particular codimension-$D$ junction of translation symmetry defects. For example, in 2D the commutator of primitive translations along the two Bravis basis vectors geometrically encloses exactly one unit cell. We will fix a $\alpha\in E_2^{D,2}$, which is uniquely determined by the $A$ projective representation. We look for SPT-LSM theorems in the strongest form, that is, nontrivial SPT ground states protected by $A$ alone are enforced when the symmetries are preserved. In order to have a SPT-LSM theorem, the ``decoration'' should be trivializable. In particular, in order to enforce a ``strong'' SPT ground state, we consider the following differential: \begin{equation} d_D: E_D^{0,D+1}\rightarrow E_D^{D,2}. \label{} \end{equation} Suppose there exists a $\omega_0\in E_2^{0,D+1}$ such that $d_D\omega_0=\alpha$. To obtain an SPT-LSM theorem, it is necessary to check that $d_n\omega_0$ vanish for $2\leq n<D$. In addition, it is also desirable that $\nu$ does not have other trivializations, to get the most stringent constraint possible. If there exists such a $\omega_0$ with all the requirements satisfied, then a SRE ground state must be a nontrivial $A$ SPT phase described by $\omega_0$. Note that higher differentials of $E_D^{0,D+1}$ vanish automatically since ${{\mathcal{H}}}^{k}[\mathbb{Z}^D, *]=\mathbb{Z}_1$ for $k>D$. For $D=2$, the above criteria are all trivially satisfied since $d_2$ is the very first obstruction. Applying the first term in formula (\ref{dFabkl}), we have \begin{align} &\quad(d_2\omega_0)(a,b,\ag{1}{k},\ag{1}{l})\\\nonumber &=\frac{{}^{\ag{1}{kl}}\!\!\left[\omega_0\!\left(\gia{kl}{a},\gia{kl}{\nugh{k,l}},\gia{kl}{b}\right)\right]}{{}^{\ag{1}{kl}}\!\!\left[\omega_0\!\left(\gia{kl}{a},\gia{kl}{b},\gia{kl}{\nugh{k,l}}\right)\right] \times {}^{\ag{1}{kl}}\!\!\left[\omega_0\!\left(\gia{kl}{\nugh{k,l}},\gia{kl}{a},\gia{kl}{b}\right)\right]}, \end{align} where $\nu\in{{\mathcal{H}}}^2[\mathbb{Z}^2,A]$ specifies the central extension of the group. For simplicity, let us consider the special case where $G=\mathbb{Z}^2$ has trivial action on $A$, and $\nu(T_x,T_y)=g\in A$, and $\nu(\mb{k,l})=0$ otherwise. It represents the magnetic translation symmetry with $T_xT_yT_x^{-1}T_y^{-1}=g$ ($g$ is a fixed element in $A$). Then $(d_2\omega_0)(a,b,T_x,T_y)$ has a simpler expression $\frac{\omega_0(a,g,b)}{\omega_0(a,b,g) \omega_0(g,a,b)}$. If it is cohomologous to a nontrivial cocycle $\alpha(a,b)$ in $E_2^{2,2}={{\mathcal{H}}}^2[\mathbb{Z}^2,{{\mathcal{H}}}^2[A,\mathrm{U}(1)]]={{\mathcal{H}}}^2[A,\mathrm{U}(1)]$ representing the projective representation in the unit cell, then we have an SPT-LSM theorem. In this way, we rederived and generalized the SPT-LSM theorem in Ref.~\onlinecite{Yang_LSMSPT}. Now we consider $D=3$, and suppose that ${{\mathcal{H}}}^3[A, \mathrm{U}(1)]=\mathbb{Z}_1$. An SPT-LSM theorem is given by the first differential term of Eq.~(\ref{dFabklm}) \begin{equation} d_3: E_3^{0,4}\rightarrow E_3^{3,2}. \label{} \end{equation} If ${{\mathcal{H}}}^3[A, \mathrm{U}(1)]$ vanishes, both differentials $d_2: E_2^{0,4}\rightarrow E_2^{2,3}$ and $d_2: E_2^{1,3}\rightarrow E_2^{3,2}$ vanish automatically. An example of 3D SPT-LSM theorem was given in Ref. \onlinecite{JiangLSM}, with $G=\mathbb{Z}^3$ and $A=\mathrm{U}(1)\times\mathbb{Z}_2^\mathsf{T}$, with translations acting as charge conjugation. \subsection{2D topological phases with magnetic translation symmetry} \label{sec:set-magnetic} We now apply the LHS spectral sequence to the problem of 2D symmetry-enriched topological (SET) phase with magnetic translation symmetry~\cite{Lu_2020, manjunath2020classification}. As in the previous section, the system has an internal symmetry group $A$ and lattice translation $G=\mathbb{Z}^2$, but the primitive translations $T_x$ and $T_y$ satisfy \begin{equation} T_xT_yT_x^{-1}T_y^{-1}=a, \label{} \end{equation} where $a\in A$ represents an internal symmetry transformation. Apparently, the internal symmetry must all commute with $a$, i.e. $a$ is a central element. For simplicity, let us assume that the internal symmetry is an Abelian group. The total symmetry group $\tilde G$ is then a central extension of $\mathbb{Z}^2$ by $A$ [see Eq.~(\ref{A_ZD})]. It is specified by $\nu\in{{\mathcal{H}}}^2[\mathbb{Z}^2,A]=A$. If a general element of the translation group is represented as $T_x^{m_x}T_y^{m_y}$ where $m_x,m_y\in\mathbb{Z}$, then we can choose \begin{equation} \nu(T_x^{m_x}T_y^{m_y}, T_x^{n_x}T_y^{n_y})=a^{m_xn_y}. \label{} \end{equation} In the topological phase, anyons can carry fractionalized quantum numbers of the global symmetry. Assuming that no anyons are permuted by the symmetry, symmetry fractionalization is classified by ${{\mathcal{H}}}^2[\tilde{G}, \cal{A}]$, where $\cal{A}$ is the group of Abelian anyons. There are three terms in the $E_2$ page: \begin{equation} \begin{split} E_2^{0,2}&={{\mathcal{H}}}^2[A, \cal{A}], \\ E_2^{1,1}&={{\mathcal{H}}}^1[\mathbb{Z}^2, {{\mathcal{H}}}^1[A, \mathcal{A}]]={({{\mathcal{H}}}^1[A, \cal{A}])}^2, \\ E_2^{2,0}&={{\mathcal{H}}}^2[\mathbb{Z}^2, \cal{A}]=\cal{A}. \end{split} \label{eqn:E2magnetic} \end{equation} Ref. \onlinecite{ChengPRX2016} gave the physical interpretations of the various terms when $a=1$: $E_2^{0,2}$ is just the fractionalization of $A$ symmetry itself on anyons. $E_2^{1,1}$ means anyons carry ``dipole moment'' of $A$, i.e. as one separates a pair of anyons apart the total $A$ charge changes. The $E_2^{2,0}$ is the background anyon charge in each unit cell. For magnetic translation the results can be understood similarly. In particular, $E_2^{2,0}$ now describes the background charge in a unit cell of the original lattice (not the magnetic unit cell). Notice that however now they are subject to obstruction and trivialization differentials. The only obstruction differential to be considered is \begin{equation} d_2: E_2^{0,2}\rightarrow E_2^{2,1}. \label{} \end{equation} Suppose $\omega_0$ is a 2-cocycle in ${{\mathcal{H}}}^2[A, \cal{A}]$. Using Eq. \eqref{dF21}, we find $d_2\omega_0\in E_2^{2,1}={{\mathcal{H}}}^2[\mathbb{Z}^2, {{\mathcal{H}}}^1[A, \cal{A}]]$ given by: \begin{equation} (d_2\omega_0)(b, \mb{g,h})=\frac{\omega_0(b, \nu(\mb{g,h}))}{\omega_0(\nu(\mb{g,h}),b)}. \label{} \end{equation} Here $b\in A$. For fixed $\mb{g,h}$, $(d_2\omega_0)(, \mb{g,h})$ gives a homomorphism from $A$ to $\cal{A}$. To check whether $d_2\omega_0$ is a nontrivial cocycle in $E_2^{2,1}$, we consider the following invariant: \begin{equation} \frac{(d_2\omega_0)(b, T_x, T_y)}{(d_2\omega_0)(b, T_y, T_x)}=\frac{\omega_0(b, a)}{\omega_0(a,b)}. \label{} \end{equation} Therefore $\omega_0$ is $d_2$-obstructed if and only if there exists $b\in A$ such that $\omega_0(a,b)\neq \omega_0(b,a)$. What this condition means heuristically is that the group element $a$ must be central (i.e. commuting with all other elements in $A$) even when acting on an individual anyon. For $A=\mathrm{U}(1)$ or a cyclic group, this obstruction always vanishes. Now let us consider the trivialization differentials. The only nontrivial one is \begin{equation} d_2: E_2^{0,1}\rightarrow E_2^{2,0}. \label{} \end{equation} Here $E_2^{0,1}={{\mathcal{H}}}^1[A, \cal{A}]$, i.e. a group homomorphism $\varphi$ from $A$ to $\cal{A}$. Under $d_2$, one finds $(d_2\varphi)(\mb{g,h})=\varphi\big(\nu(\mb{g,h})\big)$. In particular, for $G=\mathbb{Z}^2$ we only need to check the $T_x, T_y$ commutator: \begin{equation} \frac{(d_2\varphi)(T_x, T_y)}{(d_2\varphi)(T_y, T_x)} = \varphi(a). \label{} \end{equation} Namely, we exclude from $E_2^{2,0}=\cal{A}$ the elements that can be written as $\varphi(a)$ for some group homomorphism from $A$ to $\cal{A}$. Such trivializations do not occur for $A=\mathrm{U}(1)$ since ${{\mathcal{H}}}^1[\mathrm{U}(1), \cal{A}]$ is trivial. Neither the obstruction nor the trivialization affects $E_2^{1,1}$, as the physical effect only involves the translation along a certain direction. We therefore conclude that for $A=\mathrm{U}(1)$, the $E_2$ page already converges to $E_\infty$, and we obtain a complete classification of symmetry fractionalization from Eq. \eqref{eqn:E2magnetic}. \subsubsection{Magnetic LSM theorem} Now we study an example of LSM theorem with magnetic translation symmetry. Consider a spin system with $\mathrm{O}(2)=\mathrm{U}(1)_{S_z}\rtimes \mathbb{Z}_2$ symmetry. It is often sufficient to consider the (maximal) finite subgroup $\mathbb{Z}_2^2=\{1, Z, X, ZX\}$. Here $Z$ is the $\pi$ spin rotation for the $\mathrm{U}(1)_{S_z}$ subgroup of $\mathrm{O}(2)$, and $X$ is the $\pi$ rotation around $x$ axis.We will assume that on the lattice there is a $\pi$ flux of $\mathrm{U}(1)_{S_z}$ in each unit cell. In other words, the lattice translations satisfy \begin{equation} T_xT_yT_x^{-1}T_y^{-1}=Z. \label{} \end{equation} The 2-cocycle $\nu\in{{\mathcal{H}}}^2[\mathbb{Z}^2,\mathbb{Z}_2^2]=\mathbb{Z}_2^2$ characterizing the central extension takes values $\nu(T_x,T_y)=Z$ and $1$ for all other arguments. Each unit cell contains a projective representation of O(2), i.e. a spin-$1/2$. The setup is quite similar to the 2D SPT-LSM theorem discussed in Sec.~\ref{sec:SPTLSM}, but we will show that this is an actual LSM system. One way to prove this result is to apply the filling constraint in Ref. [\onlinecite{Lu_2020}]. Focusing on the $\mathrm{U}(1)_{S_z}$ subgroup, the system is half-filled with a background $\pi$ flux. Therefore, if the ground state is SRE, the Hall conductance must be an odd integer, which is not allowed for bosonic systems. Alternatively, we can use the LHS spectral sequence. In fact, it is sufficient to consider the $\mathbb{Z}_2\times\mathbb{Z}_2$ subgroup of $\mathrm O(2)$. Physically, a would-be SPT phase must have a $\mathbb{Z}_2\times\mathbb{Z}_2$ projective representation bound to a $Z$ flux. However, from the classification of $\mathbb{Z}_2\times\mathbb{Z}_2$ SPT phases in 2D, we know that no such SPT states exist. In fact, the cocycle $\omega_0\in E_2^{03}={{\mathcal{H}}}^3[\mathbb{Z}_2^2,\mathrm{U}(1)]=\mathbb{Z}_2^3$ for 2D $\mathbb{Z}_2\times\mathbb{Z}_2$ SPT can be parametrized as \begin{equation} \omega_0(a,b,c)=e^{i\pi\sum_{1\le i\le j\le 2}p_{ij}a_ib_jc_j}, \label{} \end{equation} with $p_{ij}=0,1$. From Eq.~(\ref{dFabkl}), the differential $d_2$ of $\omega_0$ is $(d_2\omega_0)(a,b,\ag{1}{k},\ag{1}{l}) = \frac{\omega_0(a,\nu(\mb{k,l}),b)}{\omega_0(a,b,\nu(\mb{k,l})) \omega_0(\nu(\mb{k,l}),a,b)}$. It is easy to check that they are all trivial in $E_2^{2,2}={{\mathcal{H}}}^2[\mathbb{Z}^2,{{\mathcal{H}}}^2[\mathbb{Z}_2^2,\mathrm{U}(1)]]=\mathbb{Z}_2$. So the anomaly does not match, and we have an LSM theorem. Now let us consider what kind of SET phases can be realized. We will show that quite surprisingly, in order to saturate the anomaly, certain symmetry transformations (among translations and $X$) must permute anyons. A key observation is that the spin-$1/2$ representation of O(2) can not be lifted to a projective representation of the full group $\tilde{G}$. If the translation symmetry is not magnetic, the spin-$1/2$ representation must be ``screened'' in the ground state by a background anyon charge which also carries the spin-$1/2$. However, such mechanism does not generalize to the present case. We now present a detailed calculation of the anomaly, assuming no anyons are permuted. For technical reasons, we focus on the $A=\mathbb{Z}_2^2$ subgroup of O(2), as the anomaly remains essentially the same. The classification for symmetry fractionalization ${{\mathcal{H}}}^2[\tilde{G}, \cal{A}]$ has already been computed from the LHS spectral sequence in Sec. \ref{sec:set-magnetic}. Here we adopt the results: \begin{itemize} \item $E_2^{0,2}$ is characterized by the three invariants \begin{equation} I_Z=\coho{w}(Z,Z), I_X=\coho{w}(X,X), I_{ZX}=\frac{\coho{w}(Z,X)}{\coho{w}(X,Z)}. \label{} \end{equation} Physically, $\coho{w}(Z,Z)$ determines which anyons carry fractional $Z$ charge, and similarly for $\coho{w}(X,X)$. $I_{ZX}$ determines which anyons carry a two-dimensional projective representations of $\mathbb{Z}_2^2$, i.e. the ``spin-$1/2$'' representation. We also notice that for projective represetations of O(2) we must have $I_Z=I_{ZX}$. It can be proven by considering the two generating classes: one of them has only nontrivial $I_X$, and $I_Z, I_{ZX}$ both trivial. The other generator can be lifted a projective representation of SO(3), which satisfies $I_Z=I_{ZX}$. According Sec. \ref{sec:set-magnetic}, $I_{ZX}$ must be trivial in order for the $d_2$ obstruction to vanish. In other words, no anyons carry the two-dimensional projective representation of $\mathbb{Z}_2\times\mathbb{Z}_2$. In conclusion, the non-obstructed class has both $I_Z$ and $I_{ZX}$ trivial. \item $E_2^{1,1}$ is characterized by four invariants, $\gamma_{Z/X}(T_{x/y})$. The invariant $\gamma_{a}(\mb{g})$ is defined as \begin{equation} \gamma_a(\mb{g})=\frac{\coho{w}(a,\mb{g})}{\coho{w}(\mb{g},a)}. \label{} \end{equation} One can show that all of them must be self-dual Abelian anyons because $Z^2=X^2=1$. In addition, since $Z$ is actually the $\pi$ rotation of $\mathrm{U}(1)_{S_z}$ and $E_2^{1,1}$ is actually completely trivial for $G=\mathrm{U}(1)$, $\gamma_{Z}(T_{x/y})$ must vanish in order to be lifted to O(2). \item $E_2^{2,0}$ is given by $\cal{A}$, but module all self-dual Abelian anyons due to a $d_2$ trivialization from $E^{0,1}$. \end{itemize} The ${{\mathcal{H}}}^4$ anomaly class, denoted by $O$ below, can be computed using the formula derived in Refs. \onlinecite{Chen2014} and \onlinecite{SET}. We can use Eqs.~(\ref{Lyn_w04}), (\ref{Lyn_w13}) and (\ref{Lyn_w22}) to extract $E_2^{0,4}, E_2^{1,3}$ and $E_2^{2,2}$. Since we are interested in those fractionalization classes that can be lifted back to O(2), we set $\gamma_Z(T_{x/y})=\mb{0}$. Recall that the obstructio-vanishing condition requires $I_Z=I_{ZX}=\mb{0}$. Let us focus on the most relevant component, $O^{2,2}$ in $E_2^{2,2}$, can be characterized by the following invariant: \begin{equation} I=\frac{O^{2,2}(Z,X,T_x, T_y)O^{2,2}(X,Z, T_y, T_x)}{O^{2,2}(Z,X, T_y, T_x)O^{2,2}(X,Z, T_x, T_y)}. \label{} \end{equation} The magnetic LSM anomaly requires $I=-1$. A tedious but straightforward calculation yields $I=1$ for the fractionalization classes in our case. Therefore, it is impossible to saturate the magnetic LSM anomaly. If we do not demand that the symmetry can be enhanced to O(2), the anomaly can be realized simply by a $\mathbb{Z}_2$ gauge theory. More specifically, denote the electric and magnetic gauge charges by $e$ and $m$, respectively. One can show that the fractionalization class \begin{equation} \gamma_Z(T_x)=e,\quad \gamma_X(T_y)=m, \label{} \end{equation} correctly produces the LSM anomaly. \subsection{Symmetry-enriched Abelian gauge theory in 2D and 3D} \label{sec:segex} \subsubsection{(2+1)d Abelian gauge theory} Let us apply the LHS spectral sequence to (2+1)d Dijkgraaf-Witten topological gauge theory with an Abelian gauge group $A$ and a symmetry group $G$. It will be assumed that $G$ is a finite group below, although the results can be easily adopted to the continuous case. Topological excitations in a $A$ gauge theory are dyons $(\hat{\chi},a)$ where $a\in A$ labels gauge flux, and gauge charges labeled by a character $\hat{\chi}\in \hat{A}$. Fusion rules and braiding statistics are determined by the cohomology twist $\alpha\in{{\mathcal{H}}}^3[A, \mathrm{U}(1)]$. The resulting topological order is denoted by $\mathcal{Z}^\alpha(A)$ (known as the twisted quantum double of $A$). We will assume that there is no type-III 3-cocycle~\cite{propitius1995}, since in that case the topological order becomes non-Abelian. We will discuss how to connect the results from the LHS spectral sequence to the classification framework established in Ref. [\onlinecite{SET}]. For simplicity, we assume that the extension is central so $G$ does not act on $A$. In the language of Ref. [\onlinecite{SET}], $G$ action does not change the types of gauge charges. The $E_2$ page has three terms, $E_2^{0,3}, E_2^{1,2}$ and $E_2^{2,1}$. The first term $E_2^{0,3}={{\mathcal{H}}}^3[A, \mathrm{U}(1)]$ is just the cohomology twist $\alpha$ in the gauge theory. Notice that it is subject to three obstruction differentials \begin{equation} \begin{split} d_2: E_2^{0,3}\rightarrow E_2^{2,2},\\ d_3: E_3^{0,3}\rightarrow E_3^{3,1},\\ d_4: E_4^{0,3}\rightarrow E_4^{4,0}. \end{split} \end{equation} In fact, similar structures already appeared in 2D SPT-LSM theorems, where the $d_2$-``obstructed'' SPT phase can only exist when the $G$ symmetry is implemented with a mixed anomaly with $A$. In the gauge theory context, the physical interpretation is that it is inconsistent to have the $G$ action on gauge charges as specified by the group extension in the twisted gauge theory. Let us illustrate this by an example of the $d_3$ obstruction, which first appeared in Ref. \onlinecite{Thorngren2015}. Let $A=\mathbb{Z}_n$, and the 3-cocycle \begin{equation} \alpha(a,b,c)=\frac{p}{n^2} a(b+c-[b+c]_n). \label{} \end{equation} Here $p=0,1,\dots, n-1$ and $a,b,c\in \mathbb{Z}/n\mathbb{Z}$. The $d_2$ map on $\alpha$ is trivial for $E_2^{2,2}=0$. The $d_3$ map to ${{\mathcal{H}}}^3[G, {{\mathcal{H}}}^1[A, \mathrm{U}(1)]]$ is (see Appendix \ref{app:example} for the derivation) \begin{equation} (d_3\alpha)(a, \mb{g,h,k})=-\frac{2p}{n}a(\beta_n\tilde{\nu})(\mb{g,h,k}), \label{} \end{equation} where $\tilde{\nu}$ is an integral lift of $\nu$, and $\beta_n\tilde{\nu}=\frac{1}{n}d\tilde{\nu}$ is the Bockstein homomorphism of $\nu$. If $-2p\nu$ is trivial as a cohomology class in ${{\mathcal{H}}}^2[G, A]$, then for the integral lift there exists a 2-cochain $\mu$ in ${\cal{C}}^2[G, \mathbb{Z}]$ and a 1-cochain $\lambda$ in ${\cal{C}}^1[G, \mathbb{Z}]$ such that $d\lambda+n\mu=-2p\tilde{\nu}$, then \begin{equation} (d_3\alpha)(a, \mb{g,h,k})=a(\beta_n\mu)(\mb{g,h,k}). \label{} \end{equation} This is actually a coboundary $d\omega \in \mathcal{B}^3[G, {{\mathcal{H}}}^1[A, \mathrm{U}(1)]]$ if we define $\omega(a, \mb{g,h})=a\frac{\mu(\mb{g,h})}{n}\in \mathcal{Z}^2[G, {{\mathcal{H}}}^1[A, \mathrm{U}(1)]]$. The converse is also true, that is if $-2p\nu$ is nontrivial in ${{\mathcal{H}}}^2[G, A]$, then $d_3\alpha$ is a nontrivial class in ${{\mathcal{H}}}^3[G, {{\mathcal{H}}}^1[A, \mathrm{U}(1)]]$. We can actually understand what goes wrong more intuitively: in this twisted $\mathbb{Z}_n$ gauge theory, the gauge flux $(0,1)$ satisfies the following fusion rules: $(0,1)^n=([2p]_n,0)$. On the other hand, the $[2p]_n$ gauge charge carries a projective representation of $G$ characterized by the 2-cocycle $2p\nu$. If $2p\nu$ is a nontrivial cohomology class, it is impossible to take a ``$n$-th'' root of the representation, so the $G$ representation on a unit flux is ill-defined. In summary, the group extension structure is inconsistent with the fusion rules. Finally, when $d_3$ vanishes, $d_4: E_4^{0,3}\rightarrow E_4^{4,0}$ contributes to the 't Hooft anomaly. An explicit expression of $d_4$ for $A=\mathbb{Z}_n$ is derived in Appendix \ref{app:example}: \begin{equation} d_4\alpha=-\frac{p}{n}\tilde{\nu}\cup_1\beta_n\tilde{\nu}+\frac{p}{n^2}\tilde{\nu}\cup\tilde{\nu}. \label{} \end{equation} Here $\tilde{\nu}$ is an integral lift of $\nu$, valued in $0,1,\cdots, n-1$. Next we have a term $E_2^{1,2}={{\mathcal{H}}}^1[G, {{\mathcal{H}}}^2[A, \mathrm{U}(1)]]$. The physical meaning is the following: consider gapped invertible domain walls in the $A$ gauge theory, obtained by embedding a 1D $A$-SPT state before gauging. After gauging it means that gauge fluxes have additional charge attached when passing through the domain wall. To see why this is the case, consider the topological response of a 1D SPT phase labeled by $\omega\in {{\mathcal{H}}}^2[A, \mathrm{U}(1)]$. When put on a ring with a flux $a\in A$ threaded, the global charge of the system under $b\in A$ changes by $\hat{i}^\omega_a(b)=\omega(b,a)/\omega(a,b)$~\cite{ZaletelJS2014}. Now if the 1D SPT state is embedded in the 2D gauge theory, the additional gauge charge must be carried away by the gauge flux to ensure charge neutrality. Therefore we have the transformation: \begin{equation} (\hat{\chi}, a)\rightarrow (\hat{\chi}\times \hat{i}^\omega_a, a) \label{eqn:permutation} \end{equation} as the particle passes through the wall. In other words, each element of ${{\mathcal{H}}}^2[A, \mathrm{U}(1)]$ corresponds to a nontrivial permutation in the MTC $\cal{Z}^\alpha(A)$. Notice that these permutations form a subgroup of the topological symmetry group $\text{Aut}\big(\cal{Z}^\alpha(A)\big)$which preserves the gauge structure (i.e. the electric charge sector remains invariant). $E_2^{1,2}$ is then a homomorphism from $G$ to this subgroup, as expected from the general classification~\cite{SET}. Let us discuss the obstruction differential $d_2: E_2^{1,2}\rightarrow E_2^{3,1}$. The $E_2^{1,2}$ cocycle is a homomorphism $\rho: G\rightarrow {{\mathcal{H}}}^2[A,\mathrm{U}(1)]$: \begin{equation} (d_2\rho)(a, \mb{h,k,l})=\frac{\rho_\mb{l}(a, \nu(\mb{h,k}))}{\rho_\mb{l}(\nu(\mb{h,k}),a))}. \label{eqn:d2E12} \end{equation} The $\mb{l}$ symmetry acts on a flux $a\in A$ by attaching a charge $\hat{i}_a^\omega$. The projective representation of this charge is characterized by the factor set $\hat{i}_a^{\rho_\mb{l}}(\nu(\mb{h,k}))$ for $\mb{h,k}\in G$, which is exactly Eq. \eqref{eqn:d2E12}. Therefore the $d_2$ represents an obstruction in assigning consistent fractionalization to fluxes, due to potential conflict between the permutations by $\rho$ and the projective represetations carried by gauge charges. As an example, let us assume $G=G_1\times G_2$. Suppose $\nu$ can be completely restricted to $G_2$, and $\rho_\mb{g}$ is nontrivial only when $\mb{g}\in G_1$. If under a transformation $\mb{g}_1\in G_1$, a certain flux is attached a charge carrying a nontrivial projective representation under $G_2$, this is physically impossible as the projective representation of the flux under $G_2$ (by assumption, it is invariant under $G_2$) should not change under symmetry transformations. This inconsistency can be detected by the $d_2$ map through slant product. The next term, $E_2^{2,1}={{\mathcal{H}}}^2[G, {{\mathcal{H}}}^1[A, \mathrm{U}(1)]]={{\mathcal{H}}}^2[G, \hat{A}]$ describes symmetry fractionalization on gauge fluxes. $d_2: E_2^{2,1}\rightarrow E_2^{4,0}$ contributes to the 't Hooft anomaly. There is also the trivialization map $d_2: E_2^{0,2}\rightarrow E_2^{2,1}$. More specifically, for $\omega\in {{\mathcal{H}}}^2[A, \mathrm{U}(1)]$, \begin{equation} (d_2\omega)(a, \mb{h,k})= \frac{\omega\big(a, \nu(\mb{h,k})\big)}{\omega\big(\nu(\mb{h,k}),a\big)}. \label{eqn:E02d2} \end{equation} As we have discussed for the obstruction map of $E_2^{1,2}$, for an $a$ flux, a charge $\hat{i}^\omega_a$ is attached, which carries a projective representation exactly given by Eq. \eqref{eqn:E02d2}. Thus the trivialization map simply indicates whether the projective representations on fluxes can be removed by relabeling. Finally, collecting all contributions to $E_2^{4,0}$ ($d_2$ from $E_2^{2,1}$, $d_3$ from $E_3^{1,2}$, and $d_4$ from $E_4^{0,3}$), one obtains a formula for the 't Hooft anomaly of the symmetry-enriched gauge theory. Together, we see that terms on the $E_2$ page describe how gauge fluxes are permuted under the symmetry group $G$, as well the fractionalization on the fluxes. The obstruction and trivialization maps can be interpreted in terms of the consistency between the symmetry actions and the underlying topological order, as well as the 't Hooft anomaly. \subsubsection{(3+1)d U(1) gauge theory} \label{sec:U1} We now consider the classification of (3+1)d U(1) gauge theory enriched by $G$. Recently the formal classification was established in Ref. [\onlinecite{NingU1SL}] (see also Ref. [\onlinecite{HsinU1SL}]), and here we will see how the results can be reproduced using the LHS spectral sequence. Suppose $G$ preserves the gauge structure (i.e. it does not mix electric and magnetic charges), then as discussed in we can use the LHS spectral sequence to compute the classification. As many details can be found in Ref. [\onlinecite{NingU1SL}], here we will be brief. The nontrivial terms on the $E_2$ page are $E_2^{1,3}, E_2^{3,1}$ and $E_2^{4,0}$. More concretely: \begin{equation} \begin{gathered} E_2^{1,3}={{\mathcal{H}}}^1[G, {{\mathcal{H}}}^3[\mathrm{U}(1), \mathrm{U}(1)]]={{\mathcal{H}}}^1[G, \mathbb{Z}]\\ E_2^{3,1}={{\mathcal{H}}}^3[G, {{\mathcal{H}}}^1[\mathrm{U}(1), \mathrm{U}(1)]]={{\mathcal{H}}}^3[G, \mathbb{Z}]\\ E_2^{4,0}={{\mathcal{H}}}^4[G, \mathrm{U}(1)]. \end{gathered} \label{} \end{equation} Physically, $E_2^{1,3}$ describes 2D bosonic integer quantum Hall states decorated on $G$ domain walls. Equivalently, $G$ implements a $T^2$ duality transformation on magnetic charges. Under $G$ magnetic charges may transform projectively, which is described by $E_2^{3,1}$. Finally ${{\mathcal{H}}}^4[G, \mathrm{U}(1)]$ describes stacking a $G$ SPT state. Now we consider the differentials. Many of them vanish automatically, and we focus on the possibly nontrivial ones. First the obstruction maps: \begin{equation} \begin{split} d_2: E_2^{3,1}&\rightarrow E_2^{5,0}\\ d_3: E_2^{1,3}&\rightarrow E_2^{4,1}\\ d_4: E_2^{1,3}&\rightarrow E_2^{5,0}\\ \end{split} \label{eqn:diffU1} \end{equation} The differentials $d_2$ and $d_4$ ending in $E_2^{5,0}$ give the 't Hooft anomaly, which was also obtained in Ref. \onlinecite{NingU1SL}. The $d_4:E_4^{1,3}\rightarrow E_4^{4,1}$ was an obstruction class to $G$ domain wall decorations (by 2D U(1) SPT phases), and it was termed ``deconfinement obstruction'' in Ref. [\onlinecite{NingU1SL}]. Explicit expressions for the differentials in Eq. \eqref{eqn:diffU1} were obtained in Ref. [\onlinecite{NingU1SL}]. Next we turn to the trivialization differentials: \begin{equation} \begin{split} d_2: E_2^{2,1}&\rightarrow E_2^{4,0}\\ d_3: E_3^{0,3}&\rightarrow E_3^{3,1}\\ d_4: E_4^{0,3}&\rightarrow E_4^{4,0}. \end{split} \label{} \end{equation} The explicit expressions for these trivialization differentials are calculated in Appendix.~\ref{app:example} (for a similar example of $A=\mathbb{Z}_n$). If $G$ is unitary and has trivial action on U(1), the differentials can be simplified: \begin{equation} \label{dU1} \begin{aligned} (d_2\omega_{2,1})(\mb{g,h,k,l})&= -\nu(\mb{g,h})n(\mb{k,l}) \\ (d_3\omega_{0,3})(a,\mb{h,k,l})&= -2 k a \cdot (\beta\nu)(\mb{h,k,l}) \\ (d_4\omega_{0,3})(\mb{g,h,k,l})&= - k [\nu(\mb{ghk,l}) \cdot (\beta\nu)(\mb{g,h,k}) \\ &\quad+ \nu(\mb{g,hkl}) \cdot (\beta\nu)(\mb{h,k,l})]\\ &\quad+ \frac{k}{2\pi}\nu(\mb{g,h})\nu(\mb{k,l}), \end{aligned} \end{equation} where $\nu(\mb{g,h})\in [0,2\pi)$ is the symmetry extension 2-cocycle from $G$ to $\mathrm{U}(1)$ in ${{\mathcal{H}}}^2[G,\mathrm{U}(1)]$. In the first equation, the cocycle $\omega_{2,1}$ in $E_2^{2,1}={{\mathcal{H}}}^2[G,{{\mathcal{H}}}^1[\mathrm{U}(1),\mathrm{U}(1)]]$ is parametrized by a $\mathbb{Z}$-valued 2-cocycle $n\in {{\mathcal{H}}}^2[G,\mathbb{Z}]$. In the last two equations, the cocycle $\omega_{0,3}$ at $E_2^{0,3}={{\mathcal{H}}}^3[\mathrm{U}(1),\mathrm{U}(1)]=\mathbb{Z}$ is chosen to be \begin{align} \omega_{0,3}(a,b,c)=\frac{k}{2\pi} a (b+c-[b+c]_{2\pi}), \end{align} with $a,b,c\in [0,2\pi) \cong \mathrm{U}(1)$ and $k\in\mathbb{Z}$. And $\beta\nu$ is the Bockstein homomorphism of $\nu$ defined as \begin{align} (\beta\nu)(\mb{g,h,k}) = \frac{\nu(\mb{h,k}) - \nu(\mb{gh,k}) + \nu(\mb{g,hk}) - \nu(\mb{g,h})}{2\pi}. \end{align} Both $d_2$ and $d_4$ lead to trivialization of SPT stacking~\cite{{WangPRX2016},{ZouPRB2018}}. The $d_2$ trivialization was discussed in Ref. [\onlinecite{HsinU1SL}]. The $d_4$ trivialization appears to be unknown before. Since the involved data is a ${{\mathcal{H}}}^3[\mathrm{U}(1),\mathrm{U}(1)]$ cocycle, we conjecture that physically it corresponds to a duality $T^2$ transformation. Finally, the $d_3$ map trivializes monopole symmetry fractionalization. The physical meaning is very clear: if the projective representation carried by the monopole is $-2k\nu$, then a duality transformation $T^{2k}$ makes the monopole completely neutral. \section{Acknowledgement} Q.-R. W. would like to thank Yang Qi for helpful discussions on diagonal approximations. M. C. and Q.-R. W acknowledge support from Alfred P. Sloan Foundation and the NSF under grant No. DMR-1846109. S.-Q. Ning acknowledges support from Research Grant Council of Hong Kong (ECS 21301018) and URC, HKU (Grant No. 201906159002).

ccbc5e108edbf98f24537516f6327d0460102e37

\section{Two-dimensional Ising models}\label{sec:2d} In order to assess the potential of the construction presented in the previous section, we have run tests on instances of the two-dimensional Ising models chosen to cover a broad range of cases ( $ L \times L $ square lattice ): \begin{itemize} \item Ferromagnetic : $J_{ij} = 1 \; \forall \; \langle i, j \rangle, \; h_i = 0 \; \forall i$. \item Antiferromagnetic : $J_{ij}=-1 \; \forall \; \langle i, j \rangle$, $h_i$ constant across the whole lattice. \item $ J' - J $ model : In this model $h_i = 0 \; \forall i$, and couplings alternate between even and odd rows or columns (see Fig.\ref{fig:fullyFrustratedCouplings}): \begin{align*} &J_{2j-1,k}=J',\; J_{2j,k}=J,\\ &J_{k, 2j-1}=J',\; J_{k,2j}=-J, \quad j=1,\,\ldots, L/2 \end{align*} The point $J = J'$, known as the fully frustrated square lattice Ising model (or Villain's model), is characterised by extensive ground state degeneracy and maximal frustration. \item Edwards-Anderson spin glass: this disordered model is such that $ h_i = 0 \; \forall i$, and $J_{ij}$ are random couplings sampled from a Gaussian distribution with zero mean and unit variance \cite{BinderYoung86}. \end{itemize} \begin{figure} \begin{center} \includegraphics[width=.45\linewidth]{JJ.pdf} \caption{Distribution of the couplings on the $J'-J$ model. Black lines indicate bonds with a coupling of $J'$, while red and blue bonds have couplings $-J$ and $J$, respectively.}\label{fig:fullyFrustratedCouplings} \end{center} \vspace{-0.75cm} \end{figure} We will be interested in the following observables: the energy per spin, $$ \varepsilon = \frac{1}{|V| Z(\beta)} \sum_{\omega \in \Omega} H(\omega) \; e^{-\beta H(\omega)}, $$ the magnetisation density, $$ m = \frac{1}{|V| Z(\beta)} \sum_{\omega \in \Omega} \left | \sum_{i \in V } \; \sigma_i \; \right | e^{-\beta H(\omega)} , $$ the staggered magnetisation density \footnote{When the external magnetic field is uniformly naught, averaging over Monte Carlo samples would result in zero (staggered) magnetisation even at temperatures where the system is known to exhibit a finite spontaneous magnetization. The absolute values appearing in our definition of the (staggered) magnetisation are meant to counter this artefact.}, $$ m_s = \frac{1}{|V| Z(\beta)} \sum_{\omega \in \Omega} \left | \sum_{i \in V } \; \text{sign}(i) \; \sigma_{i} \; \right | e^{-\beta H(\omega)}, $$ where $\text{sign}(i)$ is equal to $\pm 1$ in a checkerboard manner. Finally, we will consider also the magnetic susceptibility, defined for a system with a uniform magnetic field as $$ \chi = \partial m / \partial h. $$ \subsection*{Role of the bond dimension} The TN contractions appearing in the algorithm can be approximated in principle using any TN renormalisation strategy. Among all available methods, we have used the \emph{matrix product state renormalisation} described in \cite{Murg2005} (see also Appendix \ref{sec:tnc}). It is a choice of simplicity. In this scheme, both the accuracy and the computational effort increase with an integer parameter, the bond dimension, commonly denoted $D$. We thus expect the total variation distance between the target and the prior distribution, \begin{equation}\label{eq:tv} \| \pi^{(\beta)} - \widetilde{\pi}^{(\beta)} \|_{\text{TV}} = \frac{1}{2} \sum_{\omega \in \Omega} | \pi^{(\beta)}(\omega) - \widetilde{\pi}^{(\beta)}(\omega) |, \end{equation} to decrease with increasing values of $D$. As a result, Monte Carlo rejection rates should decrease as the bond dimension grows large. To characterize the behaviour of our method, we have explored the interplay between the bond dimension, the temperature and the rejection rate for the four different models mentioned above (Fig. \ref{fig:foobar}). In all cases, we verify that the rejection rate decreases with increasing bond dimension, and even modest values of the bond dimension may yield virtually rejection-free updates. At the same time, for a fixed $D$, rejection rates increase in the vicinity of critical points. This can be understood considering that, presumably, the distance $\| \pi^{(\beta)} - \widetilde{\pi}^{(\beta)} \|_{\text{TV}}$ for fixed $D$ will increase with the true correlation length. \begin{figure*} \begin{center} \includegraphics[width=160mm]{four-subplots.pdf} \caption{Rejection rate as a function of the temperature for four different instances of the Ising model at a fixed system size: (a) Ferromagnetic, (b) Antiferromagnetic with a constant magnetic field $h=2$, (c) Fully frustrated, (d) Gaussian spin glass. The geometry is that of a $32\times32$ square lattice with open boundary conditions in all four cases. For each model, the rejection rate was obtained by averaging over 40 independent chains, each run for a model-dependent number of steps. For the fully frustrated and spin glass cases, the points at temperatures where we believe our method starts to suffer from ill-conditioning issues are indicated by empty markers. For the ferro- and antiferromagnetic instances, the vertical lines indicate criticality in the thermodynamic limit. \footnote{ Throughout this article, we have used the PCG64 pseudo-random number generator of the NumPy module (Python) \cite{oneill-pcg2014}.}}\label{fig:foobar} \end{center} \end{figure*} Fig. \ref{fig:foobar}a demonstrates these features for the ferromagnetic case. This model exhibits, in the limit of large system sizes, a second-order phase transition at $T_c = 2/\log(1+ \sqrt{2}) \approx 2.269$ from a magnetically ordered to a paramagnetic phase \cite{cipra1987}. Still, for the system size considered, $32 \times 32$, rejection rates remain remarkably low (below $0.4$) across the whole temperature range that surrounds the critical temperature. Actually, even a bond dimension as low as $D = 2$ appears to already be sufficient to achieve our goal of producing collective updates frequently. More details regarding the vicinity of the critical point are provided on Fig. \ref{fig:scaling-fm}, where we have plotted the rejection rate as a function of the system size for different bond dimensions. Even for systems as large as $ 256 \times 256 $, acceptance rates of about $0.4$ can be obtained using only a bond dimension $D=4$. As can be appreciated from the inset of this figure, our data suggest that the bond dimension only needs to grow \emph{logarithmically} with the system size in order to maintain the acceptance rate above a threshold value. Our observations for the antiferromagnetic case (Fig. \ref{fig:foobar}b) are similar. For the fully frustrated case of the $J'-J$ model (Fig. \ref{fig:foobar}c) we obtain lower acceptance rates, as compared to the two previous cases, but still high enough that the Markov chain is usable down to $T = O(10^{-1})$. As expected, the rejection rate increases when approaching the $T=0$ critical point. Still, the minimal cost curve $D=2$ is sufficient to obtain decent acceptance rates down to at least $T=0.2$, and increasing the bond dimension again suppresses rejection events. At very low temperatures, acceptance rates drop dramatically and numerical instabilities typical of frustrated systems pointed out in \cite{zhu2019tensor} set in. Some strategies exist to mitigate these effects, but their discussion is beyond the scope of the present work, and will be the subject of a separate study \footnote{M. Frias \emph{et al.}, in preparation.}. Again, we have looked at the rejection rate as a function of the system size for different bond dimensions (see Fig. \ref{fig:scaling-ff}). Even though this instance is more challenging, the example in the figure demonstrates that at $T=0.4$ perfectly usable acceptance rates of about $0.2$ or higher can be obtained for systems of size up to $128 \times 128$ using a bond dimension, $D=6$, for which computations are not too demanding. Fig. \ref{fig:Jprime} provides additional data regarding the $J'-J$ model, beyond the fully frustrated point $J=J'$. It is remarkable that the observed maxima of rejection rates are consistent with the predicted critical lines of this model. Our findings for the Edwards-Anderson spin glass, Fig. \ref{fig:foobar}d, are qualitatively similar to those for the fully frustrated case, presumably because this spin glass is also critical at $T=0$ \cite{BinderYoung86}. Improved approximations of the contraction will generally result in a higher acceptance rate. But actually, as far as this acceptance rate does not vanish and scales well with the system size, the TNMH scheme should be applicable. \begin{figure}[h] \begin{center} \includegraphics[width=\linewidth]{scalingBondDimension.pdf} \caption{Rejection rate near the ferromagnetic Ising phase transition ($T \simeq 2.27$) as a function of the system size. Dashed lines are simply a guide to the eye. Inset: Bond dimension needed to maintain a fixed rejection rate (in this case $0.25$, although the behaviour seems to be independent of the value chosen) as a function of the system size. The fit shows that the increase in the bond dimension seems to be only logarithmic.}\label{fig:scaling-fm} \end{center} \vspace{-0.5cm} \end{figure} \begin{figure}[h] \begin{center} \includegraphics[width=\linewidth]{scalingBondDimensionT04.pdf} \caption{Rejection rate for the fully frustrated Ising model at $T = 0.4$ as a function of the system size.}\label{fig:scaling-ff} \end{center} \vspace{-0.75cm} \end{figure} \begin{figure}[h] \begin{center} \includegraphics[width=\linewidth]{rejRateFF.pdf} \caption{Rejection rate for the $J'-J$ model as a function of the temperature and the value of one of the couplings (the other has been set to unity). Computations made with a bond dimension $D=4$. Rejection rates obtained as averages over 40 independent Markov chains, each run for 200 steps. Lattice dimensions : $32\times32$. The phase separatrix predicted in \cite{Horiguchi86} is shown in black. }\label{fig:Jprime} \end{center} \end{figure} \subsection*{Equilibration and decorrelation Equilibration and auto-correlation times are two crucial time scales in Monte Carlo simulations. The former controls the number of steps needed by the Markov chain to decouple from the initial distribution (that is, the distribution from which the first configuration on the chain is sampled) and reach the desired equilibrium distribution. The latter determines the minimal time between two consecutive sample extractions in order to guarantee statistical independence. Since these times are typically difficult to bound, let alone calculate, rigorously, heuristic diagonostics are commonly used to estimate them. We have used two such heuristics to provide evidence that these two time scales are relatively short for the TNMH scheme of Section \ref{sec:asym}. A standard technique to decide that equilibration has occurred is to monitor an observable from its value at the beginning of the Markov chain until it appears to plateau at an equilibrium value around which it fluctuates \cite{Katzgraber2009}. We have used this technique for the magnetisation of a ferromagnet. Fig. \ref{fig:wolff} illustrates the evolution of the expectation value for a ferromagnetic case, and independent Markov chains evolved according to either (\ref{eq:tns-prior}) (blue), a simple spin flip Metropolis algorithm (green) or Wolff's cluster algorithm (orange) -- which is known to perform best for ferromagnetic instances. We see in this numerical experiment that the number of time steps required for TNMH to equilibrate is about $1 / 80 $ to $1 / 40 $ the number of steps required for Wolff algorithm, and about $ 1 / 10^3 $ the number of sweeps required by the single spin flip Metropolis algorithm. \begin{figure} \begin{center} \includegraphics[width=\linewidth]{magChain.pdf} \caption{Absolute value of the magnetization per site along different chains at $T=1.5$, for three algorithms, $D=4$ TNMH (blue), Wolff's algorithm (orange) and a simple spin flip (green) for a ferromagnetic $64 \times 64$ lattice. Each line represents an independent run. Time $t$ is measured in time steps for TNMH, in sweeps for the simple spin flip Metropolis, and in cluster updates for the Wolff algorithm.}\label{fig:wolff} \end{center} \end{figure} More sophisticated equilibration tests can be devised for specific problems. In particular, for the Edwards-Anderson spin glass, we have run the following test, discussed in ~\cite{PhysRevLett115077201, PhysRevB63184422}. Let $\langle X \rangle$ stand for the thermal average of an observable $X$, and $[ x ]_{\textrm{av}}$ denote the disorder average of a quantity $x$ that might depend on the coupling constants $\{ J_{ij} \}$. We consider the disorder averaged energy. \begin{equation} [ \langle H \rangle ]_{\textrm{av}} = -\int \prod_{\langle ab \rangle } \frac{dJ_{ab}}{\sqrt{2\pi}} e^{-J^2_{ab}/2} \sum_{\langle ij \rangle} J_{ij} \langle \sigma_i \sigma_j \rangle. \end{equation} At equilibrium, integrating by parts allows to prove that \begin{equation}\label{eq:Delta} \Delta \equiv \frac{1}{|V|} [ \langle H \rangle ]_{\textrm{av}} + \frac{1}{|V|} \beta \Big(|E| - \sum_{\langle ij \rangle} [ \langle \sigma_i \sigma_j \rangle^2 ]_{\textrm{av}} \Big) = 0, \end{equation} where $ \sum_{\langle ij \rangle} [ \langle \sigma_i \sigma_j \rangle^2 ]_{\textrm{av}} $ is a quantity known as link overlap. Starting configurations of Markov chains, drawn according to an easy distribution, typically have high energy and small link overlap. As a result, $\Delta$ typically has a non-zero value when the Markov chain is started. As Monte Carlo steps are taken, this value decreases in magnitude. It is a common heuristic to decide equilibration has occurred when $\Delta$ is below a given threshold value. We have implemented this heuristic test of convergence for the TNMH Markov chain. At low temperatures, we have observed the TN contraction may suffer from ill-conditionning. As a result, the approximate probability weights can be off their true value by orders of magnitude\footnote{Interestingly, it is not clear that such configurations actually correspond to local minima or maxima.}. As can be seen from Eq.(\ref{eq:mh-acc}), this mismatch affects the acceptance rates. That is, when the Markov chain hits a configuration $\omega$ such that $\widetilde{\pi}^{(\beta)}(\omega)/\pi^{(\beta)}(\omega) \ll 1$, it may remain stuck at it for a long time. The spectral gap of the transition matrix, $\{ \mathcal{T}_{\text{TNMH}}(\omega_1 \to \omega_2) \}$, which governs convergence \cite{saloff-coste-97}, should therefore be small, when compared to 'easy' models such as the ferromagnetic instance. The spectral gap of the transfer matrix of a local MC algorithm, $\mathcal{T}_{\text{loc}}$, is also expected to be relatively small for these models. But, presumably, this smaller gap has a different origin: the greater difficulty of escaping a local minimum. It could be however that alternating the two sorts of move results in a transfer matrix $\mathcal{T}_{\text{loc}} \mathcal{T}_{\text{TNMH}}$ with a larger spectral gap than each of them. We have tested this possibility. Our findings are reported on Fig. \ref{fig:hb} and on Table \ref{table:hb}, where we compare our times with state-of-the-art methods on this issue: parallel tempering (PT) and parallel tempering combined with isoenergetic cluster moves (PT+ICM)\cite{PhysRevLett115077201,PhysRevB63184422}. The comparison in Table \ref{table:hb} clearly shows that the combination of TNMH with single flip MC sweeps allows us to outperform these methods by orders of magnitude. To be fair, our simulation setup differs slightly from that of these references: whereas periodic boundary conditions were considered in these references, we have opted for open boundary conditions. The setup is identical otherwise: the number of Markov chains used for the thermal average is 30 in both cases, the number of different instances used for the disordered average is $10^4$ and the temperature is $T = 0.212$. So far, we have counted equilibration times in steps of the Markov chain, which has conceptual relevance. From a practical point of view though, it would be very interesting to know how the TNMH compares to other methods when looking at program execution times. To make such a comparison fairly is a delicate issue because we have not sought to optimise our code at all: a detailed comparison with e.g. Metropolis sweeps, which is simpler to optimise is a project in itself. We can however provide \emph{indicative} times related to the data presented on Table ~\ref{table:hb}. With our setup, we have estimated the times to get to $\Delta < 0.025$ to be $1.51 \times 10^8 \; \text{sec}$, $1.48 \times 10^6 \; \text{sec}$, and $6.75 \times 10^4 \; \text{sec}$ for the parallel tempering method, the parallel tempering method supplemented with isoenergetic cluster moves, and TNMH respectively. These figures have been obtained by multiplying an estimate for the time necessary to make a single step for each of these three methods by the number of steps necessary to reach the target value of $\Delta$. Although these estimates are rough, we believe that the order of magnitude is correct, and credibly signals the \emph{practical} potential of the TNMH Markov chain introduced here. \begin{figure} \begin{center} \includegraphics[width=\linewidth]{equilibration.pdf} \end{center} \caption{Difference between the (disordered averaged) energy per spin computed from the Hamiltonian and computed from the link overlap, Eq.~\eqref{eq:Delta}, as a function of the number of iterations. Odd times correspond to TNMH steps (bond dimension $D = 16$), even times to single spin flip Metropolis sweeps. Lattice dimensions: $32 \times 32$. Inset: Same plot, for a $16 \times 16$ lattice. The error bars, smaller than the symbols, have been computed by estimating the variance of the disorder.} \label{fig:hb} \end{figure} \begin{table} \begin{tabular}{|c|c|c|c|c|} \hline $\Delta$ & $0.25$ & $0.15$ & $0.05$ & $0.025$ \\ \hline PT & $2^{21}$ & $2^{22}$ & $2^{23}$ & $2^{24}$ \\ \hline PT + ICM & - & - & $2^{13}$ & $2^{14}$ \\ \hline TNMH + Metropolis & $3$ & $3$ & $5$ & $5$ \\ \hline \end{tabular} \caption{ First row: target value of $\Delta$, as defined by Eq.(\ref{eq:Delta}). Second and third row: each entry represents a lower bound on the number of Monte Carlo sweeps necessary to decrease $\Delta$ below the value indicated in the same column for parallel tempering (PT) and parallel tempering plus isoenergetic cluster moves (PT+ICM) (data read off Fig. 2 of \cite{PhysRevLett115077201}). Fourth row: Upper bounds on the number of TNMH iterations necessary for the same purpose. The system considered is the same as in Fig.~\ref{fig:hb}.} \label{table:hb} \end{table} We next move to autocorrelation times, that is, after equilibration is reached, the time needed between two sample extractions to guarantee (sufficient) independence. Given an observable $X$, we study the correlation function \begin{equation}\label{eq:cauto} C_{X}(t) = \frac{\langle {X}(t_0){X}(t_0+ t)\rangle - \langle {X}(t_0)\rangle\langle {X}(t_0 + t)\rangle}{ \langle {X}^2(t_0)\rangle - \langle {X}(t_0)\rangle^2}, \end{equation} where $t_0$ is assumed larger than the equilibration time \cite{Katzgraber2009}. At large $t$, we expect $C_X(t)$ to decay exponentially, with a time scale set by the decorrelation time. As the exponential tail has a fixed amount of noise, controlled by the number of samples, a useful measure to determine that time scale is the integrated correlation time $\tau_{X}^{\textrm{int}}(t) \equiv 1+2\sum_{t'=1}^t C_X(t')$. It can also be shown that it is approximately the factor that enhances the variance when averaging over samples that are not sufficiently decorrelated \cite{Katzgraber2009}. The two quantities are plotted on Fig. \ref{fig:correlationTimes} for the magnetization of a Fully Frustrated Ising model on a $32\times32$ lattice at $T=1$. The observable chosen has no clear physical meaning, but can differ greatly between different low energy configurations of this model, and thus it is a useful measure to check the ergodicity of any scheme. The data shown in Fig. \ref{fig:correlationTimes} shows that TNMH outperforms a local algorithm by almost two orders of magnitude. Furthermore, we expect that the difference in performance can only increase as the temperatures are lowered or the system size is increased. \begin{figure} \begin{center} \includegraphics[width=\linewidth]{correlationTime.pdf} \caption{Integrated correlation time and (inset) decay of the autocorrelation function of the magnetization as a function of the step, for a Fully Frustrated Ising model on a $32\times32$ lattice with open boundary conditions at $T=1$. The error bars have been computed by estimating the variance of the observables using the same samples.}\label{fig:correlationTimes} \end{center} \end{figure} To further illustrate sample-to-sample decorrelation in our algorithm, we have considered the fully frustrated Ising model and represented snapshots at different times for TNMH and for Metropolis sweeps starting from a same configuration. The difference between both techniques is striking: while the configurations appearing in our technique seem to bear no resemblance to one another from one acceptance to the next, memory of the initial configuration can still be appreciated visually after ten Metropolis sweeps. \begin{figure} \begin{center} \includegraphics[width=\linewidth]{correlationImages.pdf} \caption{Snapshots of the evolution of two different Markov chains starting from the same configuration, for the fully frustrated Ising model on a $64 \times 64$ lattice at $T = 0.5$. The top plots display the configuration obtained after three consecutive steps of the TNMH method, while those below show configurations after Metropolis sweeps at times $t=0, 1, 10$.}\label{fig: I} \end{center} \end{figure} \subsection*{Observables We now turn to the estimation of observables from the samples output by the TNMH Markov chain. We have focused on the ferromagnetic case and the antiferromagnetic one with an external field. The absolute value of the magnetisation for the ferromagnetic case is plotted on Fig. \ref{fig:ferro} (inset) and is in good agreement with the theory \cite{cipra1987}. Fig. \ref{fig:ferro} shows estimates for the fourth order Binder cumulant \cite{landau2005}, $g = (3-\langle m^4 \rangle/\langle m^2 \rangle^2)/2$. One can appreciate that the phase transition point is correctly signalled by the locus where all data sets meet, as expected. On Fig. \ref{fig:af}, we have represented the staggered susceptibility as a function of the temperature and the external magnetic field for an antiferromagnet. Our findings seem to be in good agreement with previous studies of this model \cite{Wang1997, old_antiferro}. When the external field is naught, the ferromagnetic phase transition around $T_c=2/\ln(1+ \sqrt{2})$ is recovered, as expected, since for a square lattice, a local change of variables allows a mapping between antiferro and ferromagnetic instances of the Ising model. As the field increases, the temperature at which the phase transition takes place decreases. The intuition for this fact is as follows: at $h=0$ and below the critical temperature one has antiferromagnetic order. In the large $h$ limit at the same temperature all spins would align with the external field and one would have ferromagnetic order. Thus, some phase boundary must be encountered when going from one to the other. \begin{figure} \begin{center} \includegraphics[width=\linewidth]{binderCumulant.pdf} \caption{ Binder ratio as a function of temperature for the two dimensional ferromagnetic Ising model. The data approximately cross at one point, signalling a phase transition, in great concordance with the theoretical result $T \approx 2.269$. Inset: Magnetisation of the ferromagnetic Ising model on a square $64\times64$ lattice with open boundary conditions. The error bars, computed via a jackknife analysis for the Binder cumulant and by estimating the variance for the magnetization, are smaller than the symbols. $D = 6$ was used in our algorithm. } \label{fig:ferro} \end{center} \end{figure} \begin{figure} \begin{center} \includegraphics[width=\linewidth]{sus.pdf} \caption{Susceptibility $\chi$ for an antiferromagnetic Ising model with an external field, on a $64\times64$ square lattice. $D = 6$ was used. Black: theoretical prediction of the critical line in the thermodynamical limit.} \label{fig:af} \end{center} \end{figure} \section{Three-dimensional Ising models}\label{sec:3d} Just as for planar systems, the partition function of a three-dimensional Ising model can be expressed as a tensor network. As a consequence, our TNMH construction immediately extends to three dimensions. The contraction of a three-dimensional TN is a more demanding problem than its two-dimensional analogue. Still, TN renormalisation schemes can be applied to find an approximation to the contraction \cite{Xie2012,Wang2014potts3d,Vanderstraeten2018}. We have chosen an unsophisticated renormalisation scheme involving projected entangled pair states (PEPS). Two cutoff parameters now govern the effort put in the TNMH for this implementation: a boundary PEPS bond dimension $D$, and a boundary MPS bond dimension $\chi$ (see Appendix \ref{sec:tnc} for details). We have considered two instances of the Ising model: ferromagnetic, and antiferromagnetic with an external magnetic field. The upshot is that our TNMH performs very well, even with rather low values for $D$ and $\chi$. Fig. \ref{fig:ferro3d-rejection-rate} shows the interplay between the two parameters $D$ and $\chi$, the temperature, and the rejection rate. Again, the peak in the rejection rate signals the presence of a critical point (displaced due to finite size). As this critical point is approached, rejection rates increase much faster than in two dimensions, and putting in more computational effort by increasing $D$ and $\chi$ now produces milder drops in rejection rates. We attribute this situation to an increase of correlations in the system due to a higher coordination number for each spin. Still, these preliminary results are very encouraging, since using a non-optimized contraction scheme, and modest values for the parameters $D$ and $\chi$, usable acceptance rates ($ > 0.12$ and $ > 0.05$ for the ferro- and antiferromagnetic case respectively) have been found across the whole temperature range considered, for systems as large as $ 16^3 = 4096 $ spins. Analogous to Fig.~\ref{fig:wolff}, which explored equilibration in the two dimensional case, we show the energy of a $16^3$ ferromagnetic Ising model as a function of time on Fig.~\ref{fig: wolff_3d}, both for the TNMH Markov chain and for the three-dimensional Wolff algorithm. As in two dimensions, the former appears to necessitate a lower number of steps than the latter. The magnetisation of the ferromagnetic Ising model has also been plotted in Fig.~\ref{fig: ferro_3d} and shows good agreement with previous studies \cite{Ferrenberg2018, BINDER1972508, domb1974ising}. Since observables can also be expressed as a TN, it is possible to estimate them using a direct contraction \cite{Murg2005}, and one might then wonder if the sampling procedure, which in itself requires a TN contraction, provides an advantage with respect to such a direct calculation. But, while the TNMH algorithm can succeed with a very undemanding approximate TN contraction (i.e. using very low bond dimensions), achieving a result of comparable quality by direct contraction generally requires more computational effort. To make this point more concrete, we have compared the value of the average energy in the three-dimensional ferromagnetic case, as obtained with the TNMH scheme and with direct TN contractions with different values of the bond dimensions (Fig.~\ref{fig:comparisonTRG}). We observe that, at temperatures where the direct contraction with up to $(D, \chi)=(8,16)$ was not sufficient to obtain an accurate estimate of the energy, the TNMH with $(D,\chi)=(2,2)$ was successful, since it produced decent acceptance rates, and eventually provided good samples thanks to irreducibility and reversibility. \begin{figure} \begin{center} \includegraphics[width=\linewidth]{rejectionRate3dFerro.pdf} \includegraphics[width=\linewidth]{rejectionRate3dAntiferro.pdf} \caption{ Rejection rates for the three-dimensional Ising model as a function of the temperature. The top plot corresponds to a uniform ferromagnet, the bottom plot to a uniform antiferromagnet in a field $h = 3$ (bottom). $D$ denotes the PEPS bond dimension, while $\chi$ stands for the boundary bond dimension used when compressing the PEPS associated to a plane of the lattice. A lattice of size $16\times16\times16$ was used, with open boundary conditions, and for each bond dimension and temperature, 50 chains were run for 150 steps each. The critical temperature is $T_c \approx 4.512$ \cite{Ferrenberg2018} for the ferromagnetic Ising model, and $T_c \approx 4$ \cite{domb1974ising} for the antiferromagnetic with this field (We attribute the offset with respect to this value to finite size effects.). } \label{fig:ferro3d-rejection-rate} \end{center} \end{figure} \begin{figure} \begin{center} \includegraphics[width=\linewidth]{mag_chain_3d.pdf} \caption{Single site magnetization along different Markov chains at $T=3$ for two algorithms, Wolff's (orange) and TNMH (blue) ($D = 2, \chi = 2$) for a ferromagnetic $16 \times 16 \times 16$ lattice. $t$ represents the number of cluster moves in the former case and the number of TNMH iterations in the latter.}\label{fig: wolff_3d} \end{center} \end{figure} \begin{figure} \begin{center} \includegraphics[width=\linewidth]{phaseDiagramFerro3d.pdf} \includegraphics[width=\linewidth]{phaseDiagramAntiferro3d.pdf} \caption{Magnetisation (blue) and energy (orange) per spin of the ferromagnetic Ising model (top) and staggered magnetisation and energy per spin of the antiferromagnetic Ising model in an external field (bottom) on a cubic $16\times16\times16$ lattice with open boundary conditions. The error bars are smaller than the symbols. The largest bond dimensions used to obtain the curves were $D = 4, \chi = 8$.}\label{fig: ferro_3d} \end{center} \end{figure} \begin{figure} \begin{center} \includegraphics[width=\linewidth]{comparisonContraction.pdf} \caption{ Relative error $\epsilon$ in the average energy per spin computed via different techniques for different temperatures (near the phase transition) for the three-dimensional ferromagnetic Ising model. The reference values are obtained using Wolff's algorithm, purple rhombi are obtained using samples from our algorithm with $(D, \chi) = (2, 2)$, and the other results are obtained taking derivatives of the logarithm of the approximate contraction of the TN representing the partition function. }\label{fig:comparisonTRG} \end{center} \end{figure} \section{Acknowledgements} We thank J.I. Cirac and F. Verstraete for fruitful discussions. This work was partly supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy -- EXC-2111 -- 390814868, and by the European Union through the ERC grant GAPS (Grant no. 648913), by Ministerio de Ciencia, Innovaci\'on y Universidades (Spain) (grant no. PGC2018-095862-B-C22, 'Simulaci\'on y computaci\'on cu\'anticas', grant no. MTM2017-88385-P, grant no. SEV-2015-0554, and grant no. CEX2019-000918-M, 'Mar\'ia de Maeztu'), by Generalitat de Catalunya (Spain), SGR 1761, and from the European Union Regional Development Fund within the ERDF Operational Program of Catalunya (Spain) (project QUASICAT/QuantumCat, ref. 001- P-001644) and by Comunidad de Madrid (Spain) (grant QUITEMAD-CM, ref. S2018/TCS-4342). \section{Arbitrary boundary conditions}\label{sec:bc} Although we have focused on systems with open boundary conditions, the Markov chain (\ref{eq:mh-acc}) allows us to deal with any topology that can be obtained from a rectangle by appropriate identifications. Let us show how with the simple example of a cylinder. If we make an update where we decide to leave a column of spins unchanged, \emph{e.g.} the dashed column '$L$' of Fig.\ref{fig:cylinder}, we will effectively be considering a model with open boundary conditions, where the spins in the neighbourhood of the frozen line are subjected to a local extra magnetic field. Such a model can be sampled as before. In order to make sure all spins are refreshed, the cut of frozen spins alternates between the opposite lines depicted as '$L$' and '$R$' respectively. \begin{figure}[H] \begin{center} \begin{tikzpicture}[font=\scriptsize, scale = 1.0] \draw[ thick, fill = white ] ( 0, 0 ) ellipse ( 2*0.6cm and 2*0.3cm ); \draw[ thick, fill = white ] ( 0, 3.5 ) ellipse ( 2*0.6cm and 2*0.3cm ); \draw[ thick, dashed ] ( 1.2, 0. ) --++ ( 0., 3.5 ); \draw[ thick, dashed ] ( -1.2, 0. ) --++ ( 0., 3.5 ); \node[ anchor = south ] at (1.8,1.5) { \Large{ $R$ } }; \node [anchor = south ] at (-1.8,1.5) { \Large{ $L$ } }; \end{tikzpicture} \caption{ Cylindrical boundary conditions obtained by identification from a rectangle. $L$ and $R$ denote lines of spins alternating frozen in order to be able to use a sampling scheme designed for open boundary conditions.}\label{fig:cylinder} \end{center} \end{figure} \section{Asymmetric priors and MPS}\label{sec:asym} \section{Markov chains and tensor network renormalisation}\label{sec:asym} In this section, we present a collective Monte Carlo update where, given a current configuration, tensor network renormalisation is used to propose a candidate and to decide whether it should be accepted or not. For the sake of concreteness, and with a view to the example computations that will be considered in the next two sections, we will focus on the Ising model on a square lattice. Generalisation to other nearest neighbour interactions, such as the Potts model, is immediate. Other setups, such as the \textsf{XY} model, triangular lattices, or systems of confined hard spheres will be discussed in Section \ref{sec:other-models}. We start by fixing some notation. A lattice will be denoted as $\Lambda = (V, E)$, where $V$ stands for its vertices, and $E$ for its edges. We will focus on systems made of two-state particles (classical spins) residing on the vertices. That is, the space of configurations will be $\Omega = \{-1, +1\}^{|V|}$. A spin at location $j \in V$ will be denoted by $\sigma_j$. The Ising Hamiltonian associated with a spin configuration $\omega$ is defined as \begin{equation}\label{eq:ising} H(\omega) = - \sum_{ i \in V} h_i \sigma_i - \sum_{ \langle i, j \rangle \in E} J_{i,j} \sigma_i \sigma_j. \end{equation} Our aim is to sample according to the Boltzmann distribution \begin{equation}\label{eq:boltzmann} \pi^{(\beta)}(\omega) = e^{ - \beta H( \omega ) } / Z(\beta) \; \forall \omega \in \Omega, \end{equation} where $\beta$ denotes the inverse temperature, and \begin{equation}\label{eq:pf-ising} Z(\beta) = \sum_{ \phi \in \Omega} e^{ - \beta H( \phi ) } \end{equation} is the partition function. In a Metropolis-Hastings Markov chain \cite{Hastings70}, given a current state $\omega$, a candidate configuration $\omega'$ is proposed according to some prior distribution $g^{(\beta)}( \omega' | \omega )$, from which we are able to draw. This candidate is next accepted as the new current state with probability \begin{equation}\label{eq:mh-acc} P_{\text{acc}}( \omega \to \omega' ) = \min \{1, \frac{ g^{(\beta)}( \omega | \omega' ) }{ g^{( \beta )}( \omega' | \omega ) } \times \frac{ \pi^{( \beta )}( \omega' ) }{ \pi^{( \beta )}( \omega ) } \}. \end{equation} If the prior $g^{(\beta)}$ is symmetric in its arguments, (\ref{eq:mh-acc}) reduces to the celebrated Metropolis algorithm formula. An ideal prior is one where $ g^{(\beta)}(\omega | \omega') = \pi^{(\beta)}(\omega) $. But of course, unless the instance of the Ising model under consideration is trivial (\emph{e.g.} $J_{ij} = 0 \; \forall \langle i,j \rangle $), such a prior is unavailable. Here we are going to construct $g^{(\beta)}(\omega | \omega')$ as an approximation to $\pi^{(\beta)}(\omega)$ obtained by tensor network renormalization. Let $n = |V|$, and let $\sigma_1, \sigma_2, \ldots, \sigma_n$ denote a certain sequential labelling of the vertices (see e.g. figure \ref{fig:TN} in App. \ref{sec:tnc}). Using Bayes formula, the Boltzmann distribution can be expressed as \begin{equation}\label{eq:bayes} \pi^{(\beta)}(\omega) = \pi^{(\beta)}_1( \sigma_1 ) \; \prod_{ k = 2}^{n} \pi^{(\beta)}_k( \sigma_k | \sigma_1 \ldots \sigma_{k-1} ), \end{equation} where $\pi^{(\beta)}_1$ stands for the marginal distribution of the first spin, and $\pi^{(\beta)}_k( \cdot | \sigma_1 \ldots \sigma_{k-1} )$ denotes the conditional distribution for the $k$th spin when the spins 1 through $k-1$ are fixed to values $\sigma_1, \ldots \sigma_{k-1}$. The marginal distribution for the first spin $\pi^{(\beta)}_1( \sigma_1 )$ can be expressed as the ratio of two partition functions: $\pi^{(\beta)}_1( \sigma_1 ) = Z(\beta | \sigma_1) / Z(\beta) $, where $Z(\beta | \sigma_1)$ is the partition function for a system with the same nearest neighbour Hamiltonian as in $Z(\beta)$ but where the first spin has been fixed to the value $\sigma_1$. As mentioned in the introduction, the partition function (\ref{eq:pf-ising}) of any nearest neighbour Hamiltonian can be expressed exactly as a tensor network (TN), whose bond dimension is equal to the number of states accessible by each local degree of freedom. For the Ising model, this number is equal to two. In general, neither $Z(\beta | \sigma_1)$, nor $Z(\beta) $ can be evaluated exactly, but a TN renormalisation scheme (see Appendix ~\ref{sec:tnc}) yields approximations $\widetilde{Z}(\beta | \sigma_1)$, and $\widetilde{Z}(\beta)$ for each of these quantities. With them, one can construct an approximation $\widetilde{\pi}^{(\beta)}_1( \sigma_1 ) = \widetilde{Z}(\beta | \sigma_1) / \widetilde{Z}(\beta)$ to the true marginal distribution for the first spin. Similarly, for all other sites $k > 1$, the conditional probability distribution $\pi^{(\beta)}_k( \sigma_k | \sigma_1 \ldots \sigma_{k-1} )$ can be expressed as the ratio of two TN contractions $Z(\beta | \sigma_1 \ldots \sigma_{k}) / Z(\beta| \sigma_1 \ldots \sigma_{k-1})$, and a TN renormalisation scheme provide approximations $\widetilde{Z}(\beta | \sigma_1 \ldots \sigma_{k-1})$ and $\widetilde{Z}(\beta | \sigma_1 \ldots \sigma_{k})$ to $Z(\beta | \sigma_1 \ldots \sigma_{k-1})$ and $Z(\beta | \sigma_1 \ldots \sigma_{k-1})$ repectively, which can be used to compute an approximation $\widetilde{\pi}^{(\beta)}_k( \sigma_k | \sigma_1 \ldots \sigma_{k-1} )$ to $\pi^{(\beta)}_k( \sigma_k | \sigma_1 \ldots \sigma_{k-1} )$. Fig.~\ref{fig:seq-sampling} illustrates this sequential sampling \footnote{ After completion of our work, we were made aware that a similar sampling scheme, based on Bayes' chain rule, has been proposed to directly sample an approximation to the Gibbs distribution \cite{Ueda2005}. But no study on how to use it in a Metropolis-Hastings Markov chain was made in that work. }. \begin{figure \begin{center} \includegraphics[scale=0.15]{sampling.pdf} \caption{Pictorial illustration of the first two steps of the TNMH sequential sampling. With dots refer to sites where the spin value has been fixed.}\label{fig:seq-sampling} \end{center} \end{figure} We will be interested in schemes where the Metropolis-Hastings prior, $g^{(\beta)}(\omega | \omega')$, reads: \begin{equation}\label{eq:tns-prior} \widetilde{\pi}^{(\beta)}(\omega) = \widetilde{\pi}^{(\beta)}_1( \sigma_1 ) \; \prod_{ k = 2}^{n} \widetilde{\pi}^{(\beta)}_k( \sigma_k | \sigma_1 \ldots \sigma_{k-1} ). \end{equation} As explained in Appendix ~\ref{sec:tnc}, the approximate probability $\widetilde{\pi}^{(\beta)}(\omega)$ can be evaluated for any configuration $\omega$, and the update rule (\ref{eq:mh-acc}) can be implemented. Our construction is summarized in Algorithm \ref{algorithm:tns-mhmc}. \subsection*{Properties of the TNMH Markov chain} \begin{enumerate}[label=(\roman*)] \item The construction is \emph{universal} in the sense that it is independent of the magnetic fields and couplings that define the Ising instance being considered. Yet, it is \emph{adaptive} in that the details of the Hamiltonian are taken into account when the tensors are constructed. \item The constitution of the candidate is independent of the current configuration. \item The update (\ref{eq:tns-prior}) is \emph{collective and correlated}: in principle \emph{all} spins of the system could be refreshed in a single Monte Carlo step, and the spin values proposed at different sites are conditioned by the correlations present in the tensor network. We believe this feature is the principal cause for the high acceptance rates and fast equilibration reported in the next section. Whereas a local update rule could have a hard time overcoming energy barriers, we expect our algorithm to be more capable to hop between distant regions of $\Omega$ and escape local minima in a single step. \item The transition matrix, i.e. the set of probabilities to transition from a configuration $\omega$ to a configuration $\omega'$, $\mathcal{T}( \omega \to \omega') = \widetilde{\pi}^{( \beta )}( \omega' ) \times P_{\text{acc}}( \omega \to \omega' )$, satisfies reversibility. That is, \[ \pi^{(\beta)}(\omega) \; \mathcal{T}( \omega \to \omega') = \pi^{(\beta)}(\omega') \; \mathcal{T}( \omega' \to \omega). \] Furthermore, when numerical errors are small enough that all the conditioned partition functions $\widetilde{Z}(\beta | \sigma_1 \ldots \sigma_k)$ are strictly positive (see Appendix ~\ref{sec:tnc}), the Markov chain is also irreducible: $$ \mathcal{T}( \omega \to \omega') > 0 \quad \forall \omega, \omega' \in \Omega. $$ \end{enumerate} Thus, even if the distributions $\{\pi^{(\beta)}_k : k \in V\}$ turned out to be poorly approximated by the TN renormalisation scheme used, it is still possible to guarantee that the Markov chain will eventually converge to the target probability distribution \cite{Bhanot_1988}. This last point will be illustrated with three-dimensional Ising models. \alglanguage{pseudocode} \begin{algorithm}[H] \small \begin{algorithmic}[1] \State Compute the tensors associated with the distribution (\ref{eq:boltzmann}). \State Set $t=0$, and draw some initial configuration $\omega(0)$ according to any distribution over $\Omega$. \State If $t > t_\text{max}$ go to \ref{algo-1:end}. \label{algo-1:for} \State Use the tensor network to draw a candidate configuration $\omega'$ according to Eq.(\ref{eq:tns-prior}). \State Evaluate the probabilities $\widetilde{\pi}^{(\beta)}(\omega(t))$ and $\widetilde{\pi}^{(\beta)}(\omega')$. \State Accept the change $\omega(t) \gets \omega'$ with probability $\min \{1, \frac{ \widetilde{\pi}^{(\beta)}( \omega | \omega' ) }{ \widetilde{\pi}^{( \beta )}( \omega' | \omega ) } \times \frac{ \pi^{( \beta )}( \omega' ) }{ \pi^{( \beta )}( \omega ) } \}. \State $t \gets t + 1$. Go to \ref{algo-1:for}. \State End. \label{algo-1:end} \end{algorithmic} \caption{TNMH Markov chain} \label{algorithm:tns-mhmc} \end{algorithm} \section{Discussion}\label{sec:discussion} The interplay between Monte Carlo and tensor network methods is a rich and vastly unexplored subject. While various previous works have reported on using Monte Carlo sampling for tensor network contractions, we have here presented an analysis of the converse: tensor network contractions for Monte Carlo sampling. We have introduced a new class of Markov chain Monte Carlo algorithms for many-body classical systems based on tensor network renormalisation. This class belongs in the family of Metropolis-Hastings schemes. Our construction produces collective updates. It is also irreducible and reversible; as such, asymptotic convergence towards the target probability distribution is guaranteed. We emphasize its universal nature: it works the same for any nearest neighbour Hamiltonian with finite local degrees of freedom. We have benchmarked our scheme for a variety of instances of the two-dimensional Ising model defined on a square lattice. For ferromagnets and antiferromagnets, very high acceptance rates have been observed for larger systems, even with modest values of the bond dimension. Besides, drops in acceptance rates have been shown to signal criticality. Looking at equilibration and decorrelation times, the scheme compares extremely well with single spin flip updates and Wolff algorithm. As expected, the scheme's performance is lower for frustrated and disordered instances than for the ferro- and antiferromagnets. Still, our results are very encouraging. In particular, for disordered instances, equilibration appears to be occurring orders of magnitude faster than for state-of-the-art techniques such as parallel tempering supplemented with isoenergetic cluster moves, both when time is counted in Monte Carlo steps and in seconds. We have also demonstrated the potential of the method for three dimensional systems, by testing it on ferromagnetic and an antiferromagnetic instances. Also in this case, we observe faster equilibration as compared to Wolff algorithm and, remarkably, even with a simple contraction strategy and small bond dimension, the scheme can be shown to remain usable at near critical temperatures, whereas a much more costly direct TN contraction results in considerable errors We have used simple procedures to implement tensor network renormalisation, and we have made no particular effort to write an efficient code. For these reasons, we believe the results presented here could be substantially improved. It would also be very interesting to study what can be gained by using other renormalisation schemes for approximate contractions of tensor networks \cite{Levin2007}. For example, schemes involving disentanglers would be a natural option in this regard \cite{Evenbly2015b}. Also for future work is the study of how TNMH Markov chains combine with parallel tempering \cite{hukushima96}. A major advantage of our construction is its versatility. We have seen that with little extra effort, a code valid for the Ising model on a square lattice can be used as such to construct a collective update Markov chain in other settings such as the XY model, or a triangular lattice, and that TNMH could also be used to study gases of hard spheres. In principle lattice systems with long range interactions could also be considered. For instance, given an Ising Hamiltonian $H$ where the interactions decay with the distance as a power law, one can associate an auxiliary Hamiltonian $H_\varrho$ where all interactions within some range $\varrho$ are identical to $H$, and all interactions beyond $\varrho$ have been truncated. One can next construct a tensor network prior from this Hamiltonian $H_\varrho$. Two parameters would now govern the Markov chain: the bond dimension and the range $\varrho$. We have also restricted ourselves to scalar degrees of freedom in this work. But the discussion held in Section \ref{sec:other-models} shows that TNMH sampling should also apply to matrix models, in particular lattice gauge theories. A natural variation of our work would be to depart from tensor network representations and use a quantum device to prepare Gibbs states and estimate the probability to draw a given configuration \cite{chowdhury2016quantum,Motta_2019}. Such a device would be called as an external subroutine in (classical) Metropolis-Hastings iterations. Just as our 3D computations have revealed that inaccurate contraction schemes could still be useful for sampling, it would be very interesting to investigate how much computational power such quantum devices retain when imperfect. These ideas will be studied elsewhere. Finally, it would be instructive to develop a mathematical perspective on the schemes presented here. In particular, we believe it would be meaningful to identify a non-trivial model for which the mixing time associated with our TNMH scheme could be upper bounded, \emph{e.g.} using a log-Sobolev inequality \cite{saloff-coste-97}. It would be insightful to establish the dependence of the log Sobolev constant with the bond dimension. \section{Introduction} Markov Chain Monte Carlo is central to our understanding of strongly correlated systems \cite{Metropolis1953}. When the number of degrees of freedom is too large for exact computations, and perturbative methods are ineffective, Monte Carlo sampling often emerges as the method of choice for numerical investigation. Markov chain Monte Carlo has contributed significantly to the current state-of-the-art in fields like \emph{e.g.} high temperature superconductivity \cite{Edegger2007}, ab initio quantum chemistry \cite{hammond1994monte}, or (lattice) quantum chromodynamics \cite{montvay1994quantum}. In statistical physics, Monte Carlo sampling has made it possible to chart phase diagrams of several paradigmatic spin systems \cite{krauth2006smac,landau2005}. The fundamental problem in this context is to sample according to the Boltzmann distribution. To achieve this goal, Markov chain Monte Carlo methods produce a sample by subjecting an initial configuration to a carefully designed stochastic evolution in the space of configurations. Well-known examples are the Metropolis algorithm and heat bath dynamics Markov chains where at most one spin is modified at each step \cite{landau2005}, or the Wolff algorithm, where clusters of spins are flipped at once \cite{swendsenWang,wolff}. The applications of these algorithms are countless, but there are important circumstances, such as geometric frustration or disorder, where their limitations become apparent \cite{landau2005, krauth2006smac}. Over the last two decades, a second notion has been gradually recognised as crucial to our understanding of strongly correlated systems: tensor networks states \cite{Schuch11}. In the realm of many-body quantum mechanics, the (simple) entanglement patterns, present in collections of identical particles in short range interaction, enables a description that conceptually transcends mean field approximations, but does not demand the exponential cost of exact diagonalisation \cite{Schollwoeck2011,Orus2014}. Tensor networks are also used in many-body classical physics. The first applications were proposed by Nishino in \cite{Nishino95,Nishino96,Nishino97}, and significant developments have been made possible by the advancements in tensor network algorithms. It was shown in \cite{Murg2005} that partition functions of all spin systems in nearest neighbour interaction, including inhomogeneous and finite ones, could be represented as a tensor network. While the exact contraction of the tensor network is in general computationally intractable \cite{Schuch2007complexity,Arad2010}, this idea has been used in practice to address many physical problems via an approximate contraction \cite{Nishino95,Verstraete2006,Levin2007,Xie2009,Xie2012,Evenbly2015b,yang2017loop,Ferris2015, PhysRevLett.125.060503}. Tensor network methods have been successfully applied to a variety of classical and quantum two dimensional problems (e.g \cite{Xie2009,Gu2009heis,Chen2011potts,Shimizu2014schwinger}) including continuous variables \cite{Yu2014xy,Shimizu2012boson,Unmuth-Yockey2014,Campos2019boson}, and three dimensional classical models \cite{Xie2012,Wang2014potts3d,Vanderstraeten2018}. Besides solving concrete problems to very good precision, these contributions have been insightful: we have for example learnt that the notion of bipartition Schmidt weights, ordinary in quantum information theory, is also relevant to classical statistical physics. However, unless an implausible collapse of complexity classes is found, all these methods are ultimately limited, since there exist instances of the Ising model for which the evaluation of the partition function is $\# \mathsf{P}$, even in multiplicative approximation \cite{goldberg_jerrum_2007,Galanis2016}. The downside of these fundamental obstructions is that a complete understanding of these systems will (very) likely always be out of reach. The upside is a sustained interest in developing new methods to continually push the boundary of what we can learn about these systems. In this paper, we present and explore a new connection between tensor networks and statistical physics. Instead of computing a direct estimation of the partition function, our primary goal here is \emph{to sample} configurations representative of the Boltzmann distribution of a concrete Hamiltonian at a given temperature. To achieve this goal, we introduce a Tensor Network Metropolis-Hastings (TNMH) algorithm \cite{Metropolis1953,Hastings70}, where the asymmetric prior, i.e. the distribution from which the new candidate configuration is drawn at each step, is an approximation to the target distribution, obtained via an inexpensive tensor network renormalisation contraction. In earlier works \cite{sandvikVidal, PhysRevLett.100.040501}, Monte Carlo sampling was used to evaluate tensor network contractions, whereas in \cite{Meurice2014,Yang16}, tensor network renormalisation (blocking) was compared with sampling. Our work is converse to \cite{sandvikVidal, PhysRevLett.100.040501}, in the sense that it uses tensor network contractions for Monte Carlo sampling, and instead of choosing between blocking and sampling, it aims at combining both ideas. In this way, it features concrete advantages with respect to each strategy. Some of the most remarkable ones are the following. \begin{enumerate}[label=(\roman*)] \item It is \emph{universal}, in the sense that it is identical for all instances of a given model. This is in contrast to other Monte Carlo algorithms that exploit the idea of an intelligent prior choice, but rely on a deep insight about the target distribution, and thus have limited applicability beyond the model for which they are specifically tailored. That is for instance the case of Wolff's algorithm, which performs extremely well for ferromagnetic Ising models, but rather poorly for antiferromagnets or frustrated instances. In turn, we will show that our method fares consistently well for a variety of models that are all very different from one another. \item The scheme produces collective updates. That is, the state of each of degree freedom of the considered system is susceptible to change at each Monte Carlo step. We have found that the computational effort scales mildly with increasing acceptance rates in a broad variety of instances. Presumably as a consequence, we have found that the number of Monte Carlo steps necessary to reach convergence is between $ \sim 10^1$ and $\sim 10^3$ shorter than those of other well-established Monte Carlo algorithms for several instances of two and three dimensional models of the Ising type. \item As compared to algorithms that purely rely on a tensor network renormalisation of the partition function, the shift to sampling results in the substitution of systematic errors with statistical errors, since our TNMH scheme satisfies the classical sufficient conditions for convergence (see next section). Thus, modest tensor network contraction schemes, too inaccurate for a direct evaluation of the energy, can be successfully used in our method, as they still enable collective updates with sufficiently high acceptance rates. \item Another advantage of considering sampling instead of partition function evaluation is enhanced versatility. As we shall see, a Markov chain designed for Ising models on a square lattice with open boundary conditions is useful as such to study other systems, such as the $\lambda \phi^4$ model or gases of hard spheres, other interaction graphs such as triangular lattices, and arbitrary boundary conditions. \end{enumerate} We have tested our Markov chain systematically in a variety of instances of the Ising model defined on finite square lattices: ferro- and antiferromagnetic, frustrated, disordered, in two and three spatial dimensions. One may anticipate that for systems with large (or even diverging) correlation length, our scheme will perform increasingly poorly if the bond dimension (parameter that controls the cost and accuracy of the tensor network renormalisation) is fixed. Our findings confirm this expectation, with drops in acceptance rates actually signaling phase transitions. But we have also observed that for ferro- and antiferromagnetic instances, acceptance rates remain fairly high for systems of considerable size across a phase transition, even with a bond dimension as low as $D=2$. Equilibration and decorrelation times in our TNMH scheme have been found to be systematically lower than for the Metropolis and Wolff's algorithms. As expected, frustrated and spin-glass instances have turned out to be challenging, not only for fundamental complexity-theoretic reasons, but also because their study is complicated by ill-conditioning issues \cite{zhu2019tensor}. However, even in such cases, and without any optimization of our renormalisation procedure, we have observed that acceptance rates stay high enough to be usable down to temperatures that can be considered low by nowadays state-of-the-art standards. The rest of this paper is organised as follows. The new algorithm is described in section \ref{sec:asym} in general terms. In section \ref{sec:2d} we explore its performance for two dimensional models. In particular, for a broad variety of instances of Ising models, we explore the role of the bond dimension in the acceptance rates, also in relation to the presence of critical temperatures. We further analyze equilibration and autocorrelation times, and demonstrate how the method can be used to obtain physical observables and chart phase diagrams. In section \ref{sec:3d}, we demonstrate how the algorithm is also useful for three dimensional systems, and illustrate it for ferromagnetic and antiferromagnetic instances of the Ising model in cubic lattices of up to $16^3$ sites. Section \ref{sec:other-models} is a discussion of situations where our findings could find further applications, and can be skipped on a first reading; there we discuss triangular lattices, models with continuous variables and systems of hard spheres. We provide an outlook in section \ref{sec:discussion}. \section{Other models}\label{sec:other-models} This section is a summary of ideas that will be explored elsewhere. So far, we have focused on instances of the Ising model in two and three dimensional square lattices. We here sketch variations and applications of these cases. We show how to deal with arbitrary boundary conditions in Appendix \ref{sec:bc}. \subsection*{The XY model In absence of a vector potential, the XY model describes a lattice of planar spins, interacting through the Hamiltonian \begin{equation}\label{eq:xy-model} H_{XY} = - \sum_{\langle i, j \rangle } \cos( \theta_i - \theta_j ), \end{equation} where the local variables are the angles $\{ 0 \leq \theta_i < 2 \pi : i \in V \}$. Although these variables are continuous, this model can be mapped into a system that allows to use a variation of the sampling method used for the Ising model. First, a duality transformation establishes an equivalence between (\ref{eq:xy-model}) and a system of integer variables residing on the (oriented) links of the lattice involved in four-body interactions \cite{PhysRevB161217, Xiang, PhysRevE100062136,Jha_2020}. That is, the partition function takes the form \begin{equation}\label{eq:part-func-xy-model} Z(\beta) = \lim_{N \rightarrow \infty}\prod_{l \in E} \left( \sum_{n_l = -N}^N I_{n_l}(\beta) \right) \prod_{i \in V} F^{n^{(i)}_3, n^{(i)}_4}_{n^{(i)}_1, n^{(i)}_2}, \end{equation} where $n^{(i)}_1, n^{(i)}_2, n^{(i)}_3, n^{(i)}_4$ are the values for the links meeting at site $i$. $I_{n_l}(\beta)$ are the modified Bessel functions of the first kind, and \[ F^{n_3, n_4}_{n_1, n_2} = \int_{0}^{2 \pi} \frac{d \theta}{2 \pi} e^{i \theta (n_1 + n_2 - n_3 - n_4)} = \delta_K(n_1 + n_2 - n_3 - n_4). \] At fixed $\beta$, $I_{n_l}(\beta)$ decays fast and, and truncating the sum in Eq. (\ref{eq:part-func-xy-model}) is a sensible approximation. The partition function of the XY model can thus be approximated by a tensor network where the degree of freedom at each bond takes value in a finite set. In the language of Appendix \ref{sec:tnc}, the tensor at each site $i$ would now be \[ A^{(i)}_{n_2 n_4 n_1 n_3} = \left( \prod_{k=1}^4 I_{n_k} (\beta) \right)^{1/2} F^{n_3, n_4}_{n_1, n_2}, \] and the contraction of the TN made up of these tensors would give an approximation to the partition function $\widetilde{Z}(\beta)$. Similarly, the marginal probability density of the spin at a site $i$, $\widetilde{\pi}^{(\beta)} (\theta_i)$, can be approximated by replacing the tensor at site $i$ with \[ A^{(i)}_{n_2 n_4 n_1 n_3} (\theta_i) = \left( \prod_{k=1}^4 I_{n_k} (\beta) \right)^{1/2} \frac{e^{i \theta (n_1 + n_2 - n_3 - n_4)}}{2 \pi}, \] and normalizing the contraction to the approximate partition function previously obtained. Using renormalisation to approximately contract tensor networks, and the inverse sampling method, a candidate configuration $\omega' = \{ \theta'_i : i \in V \}$ can be drawn and accepted or rejected, as we did for Ising models with Algorithm \ref{algorithm:tns-mhmc}. A vector potential could be included \cite{Xiang, PhysRevE100062136}, and other continuous variable systems admit a similar construction \cite{meurice2020tensor}. On top of the bond dimension used for the renormalisation, the number of terms kept in the series expansion of the transfer matrix in Eq. (\ref{eq:part-func-xy-model}) is another parameter that governs the accuracy of the contraction. As for the 3D Ising model discussed above, a tensor network with a low value for this parameter may be accurate enough to sample from and propose moves for a Markov chain, but not precise enough to compute the observables with a single contraction. A detailed study of the XY model is beyond the scope of this paper. But we hade made preliminary computations that show acceptance rates comparable to those of the ferromagnetic Ising model. In order to see how correlated the proposed collective moves are, we have computed the mutual information between the updates at different sites of the TNMH algorithm and compared it to that obtained from a local algorithm (left), Figure \ref{fig:corrXY}. The instance chosen for the comparison is the zero-field uniform XY model on a $16\times16$ lattice at a temperature $T = 0.5$. The difference in the results is noteworthy, and demonstrates that the TNMH algorithm is indeed capable of producing global correlated updates. \begin{figure}[h] \begin{center} \includegraphics[width=\linewidth]{correlationsXY.pdf} \caption{Mutual information in bits between the updates at different sites, that is, in the changes of the angles after a sweep of the Markov chain, $I( \theta_i(t+1) - \theta_i(t) : \theta_j(t+1) - \theta_j(t))$ for two schemes. Left: a local algorithm (a Metropolis single spin flip where the proposed local updates were chosen from $U(0, 2\pi)$). Right: TNMH. The numerical experiment was conducted on a homogeneous XY model with no external field on a $16\times16$ lattice at a temperature of $T = 0.5$. The bond dimension used in TNMH was 20, and the number of terms kept in the series of Eq. (\ref{eq:part-func-xy-model}) was $N = 4$, which gave a TN with a local dimension of 9.}\label{fig:corrXY} \end{center} \vspace{-0.75cm} \end{figure} \subsection*{Drawing configuration differences We now show that a TNMH scheme for the Ising model can be extended to deal with other nearest neighbour hamiltonians. For the sake of concreteness, we will focus on the $\lambda \phi^4$ model, defined on a two-dimensional square lattice $\Lambda = (V,E)$ by the energy function \[ H(\{ \phi_i \}) = \sum_{ \langle i, j \rangle \in E} (\phi_i - \phi_j )^2 + \sum_{i \in V} ( \frac{1}{2} m^2 \phi_i^2 + \frac{\lambda}{4!} \phi_i^4 ). \] where each local variable $\phi_i$ takes value in $\mathbb{R}$. (See also Refs.~\cite{Campos2019boson,Vanhecke_2019} for the use of tensor networks in lattice field theories.) As usual, we are interested in sampling according to the Boltzmann distribution for some fixed value $\beta$. We will use the following simple lemma. \begin{lemma}\label{lemma:2-bit-function-ising} Any real function of two binary variables, $B$, can be expressed as an Ising model energy plus some constant: \begin{equation}\label{eq:function-to-ising} B(\sigma, \sigma') = J \sigma \sigma' + h \sigma + h' \sigma' + K. \end{equation} \end{lemma} \emph{Proof} : (\ref{eq:function-to-ising}) defines a system of four linear equations for the four unknowns $J,h,h',K$, one for each assignment $(\sigma,\sigma')$. The determinant of the matrix of this system of equations does not vanish but is equal to $16$; a solution to (\ref{eq:function-to-ising}) therefore exists and is unique. Let $\omega=\{\phi_i : i \in \Omega \}$ denote the current configuration. A Markov chain with collective updates can be constructed using the TNMH presented for the Ising model in Section \ref{sec:asym} if we draw configuration changes. We proceed as follows. An integer $m$ is drawn uniformly and randomly in $\{0, 1, \ldots, m_{\text{max}}\}$, where $m_{\text{max}}$ is equal to $9$, say. $\forall i \in V$, we draw $\gamma_i$ according to a Gaussian distribution with zero mean and variance equal to $10^{-m}$. With $\Gamma = \{ \gamma_i : i \in V\}$, we construct the function: \[ H_I( \{ \sigma_i \} | \omega, \Gamma ) = \sum_{ \langle i, j \rangle \in E} (\psi_i(\sigma_i | \phi_i,\gamma_i) - \psi_j(\sigma_j | \phi_j,\gamma_j) )^2 \] \[ + \frac{1}{2} \sum_{i \in V} m^2 \psi_i(\sigma_i | \phi,\gamma_i)^2 + \frac{\lambda}{4!} \sum_{i \in V} \psi_i(\sigma_i | \phi_i,\gamma_i)^4. \] where $$ \psi_i(\sigma_i | \phi_i,\gamma_i) = \frac{1 - \sigma_i}{2} \phi_i + \frac{1 + \sigma_i}{2} ( \phi_i + \gamma_i ), $$ with $\sigma_i \in \{-1,+1 \}$ $\forall i \in V$. By lemma \ref{lemma:2-bit-function-ising}, $H_I( \{ \sigma_i \} | \omega, \Gamma )$ can be expressed as an Ising Hamiltonian for the variables $\{ \sigma_i \}$ (plus some irrelevant global constant): $$ H_I( \{ \sigma_i \} | \omega, \Gamma ) = - \sum_{ i \in V} h_i( \omega, \Gamma ) \; \sigma_i - \sum_{ \langle i, j \rangle \in \Gamma} J_{i,j}( \omega, \Gamma ) \; \sigma_i \sigma_j. $$ The Boltzmann distribution of the Ising model $H_I$, \[ \pi_I^{(\beta)}( \{ \sigma_i \} | \{ \theta_i \}, \Gamma ) = \frac{ e^{ - \beta H_I( \{ \sigma_i \} | \{ \theta_i \}, \Gamma ) } }{ \underset{\{ \sigma_j \}}{\sum} e^{ - \beta H_I( \{ \sigma_j \} | \{ \theta_i \}, \Gamma ) } } \] can generically not be sampled directly. But we can construct a tensor network approximation $\widetilde{\pi}^{(\beta)}( \cdot | \omega, \Gamma )$ for it, as described in Section \ref{sec:asym}. Given $\Gamma$ as defined above, let us define $\tau(\Gamma) = \{ - \gamma_i : i \in V \}$. The sequence of instructions listed in Algorithm \ref{algorithm:cmcu} defines an irreducible and reversible Metropolis-Hastings Markov chain that achieves collective updates for the $\lambda \phi^4$ model. \alglanguage{pseudocode} \begin{algorithm}[H] \small \begin{algorithmic}[1] \State Draw an integer $m$ u.a.r. in $\{ 0, \ldots m_{\text{max}} \}$. \State Draw $|V|$ i.i.d. Gaussians with zero mean and variance equal to $10^{-m}$: $\Gamma = \{ \gamma_i : i \in V \}$. \item Draw $\{ \sigma_i : i \in V \}$ according to $\widetilde{\pi}^{(\beta)}( \cdot | \omega, \Gamma )$. \item Accept the move $\{ \phi_i : i \in V \} \to \{ \phi_i + \frac{ 1 + \sigma_i }{ 2 } \gamma_i \}$ with probability \[ \min\{ 1, \frac{\widetilde{\pi}^{(\beta)}_I( \{ \sigma_i \} | \omega, \Gamma )}{\widetilde{\pi}^{(\beta)}_I( \{ \sigma_i \} | \omega', \tau(\Gamma) )} \times \frac{\pi^{(\beta)}(\omega')}{\pi^{(\beta)}(\omega) }\}. \] \end{algorithmic} \caption{Configuration difference collective update} \label{algorithm:cmcu} \end{algorithm} The idea of drawing configuration updates has appeared in the study of the ferromagnetic \textsf{XY} model, for which the Wolff algorithm for the ferromagnetic Ising model can be recycled \cite{landau2005}. In principle, Algorithm \ref{algorithm:cmcu} could be applied to frustrated systems. A class of systems for which we believe it could be useful to draw differences of configurations as described here are matrix models, such as $SU(d)$ lattice gauge theories \cite{RevModPhys.51.659}. The auxiliary Hamiltonian representing the possible choices for a move would no longer be two-body Ising. Still, it is not difficult to construct a tensor network representation for its Boltzmann distribution. \subsection*{Triangular lattices We now show how the construction presented in Section \ref{sec:asym}, specific to square lattices, can be used as such to deal with a triangular lattice. Let us assume that we are interested in some particular observable $X$. That is, we wish to estimate \[ \mean{X} = \frac{1}{Z(\beta)} \sum_{\omega \in \Omega} X(\omega) \; e^{- \beta H(\omega)}. \] \begin{figure*} \begin{center} \includegraphics[width=120mm]{three-triangular-lattices.pdf} \caption{Left: Interaction graph of a triangular lattice system. Centre: Same interaction graph decorated with extra degrees of freedom located on the diagonals (grey dots). Right: Square lattice on which a Hamiltonian $H_\square$ associated with the original system is defined.}\label{fig:triangular-lattice} \end{center} \end{figure*} To this end, we construct an extended model, obtained by decorating the original lattice with extra spins living on each diagonal link as shown on Fig.\ref{fig:triangular-lattice} (a) and (b). With each particle of the original model, we will associate the new extra spin located south east to it. Let $p : V \to V_{\text{new}}$ denote the function that realises this association, where $V_{\text{new}}$ denotes the set of new vertices. The Hamiltonian of the extended model reads \begin{equation}\label{eq:ext-model} H_{\text{ext}}( \omega_{\text{ext}} | \gamma ) = H(\omega) - \gamma \sum_{ j \in V } \sigma_j \sigma_{p(j)}, \end{equation} for $\gamma > 0$. $Z_{\text{ext}}(\beta | \gamma)$ will denote its partition function. \begin{proposition}\label{prop:prop-id} \begin{equation}\label{eq:prop-id} \mean{X} = \lim_{\gamma \to \infty} \frac{1}{Z_{\text{ext}}(\beta | \gamma )} \sum_{\omega_{\text{ext}} } X(\omega) \; e^{-\beta H_{\text{ext}}(\omega_{\text{ext}} | \gamma)} \end{equation} whenever $\beta$ and $|V|$ are both finite. \end{proposition} \emph{Proof:} The extended configuration space $\Omega_{\text{ext}}$ can be decomposed as $\Omega_{\text{ext}} = \Omega_{\text{ext}}^{(0)} \cup \Omega_{\text{ext}}^{(1)} \cup \ldots \cup \Omega_{\text{ext}}^{(|V|)}$, where $\Omega_{\text{ext}}^{(m)}$ denotes the subset of all configurations such that there are exactly $m$ sites $j \in \Lambda$ where $\sigma_{j} \neq \sigma_{p(j)}$. This decomposition induces another for the partition function of the extended model as $$ Z_{\text{ext}}(\beta | \gamma ) = \underset{ \omega \in \Omega}{\sum} e^{-\beta H(\omega) + \beta \gamma |V|} + \sum_{m=1}^{|V|} \zeta_m \; e^{\beta \gamma ( |V| - 2 m) }, $$ where the coefficients $\zeta_m$ are all finite and independent of $\gamma$. Similarly, the sum appearing in the r.h.s. of (\ref{eq:prop-id}) can be expressed as $$ \underset{ \omega \in \Omega}{\sum} X(\omega) e^{-\beta H(\omega) + \beta \gamma |V|} + \sum_{m=1}^{|V|} \xi_m \; e^{\beta \gamma ( |V| - 2 m) }, $$ where the coefficients $\xi_m$ are also finite and independent of $\gamma$. Finally, it is obvious that the ratio $$ \frac{ \underset{ \omega \in \Omega}{\sum} X(\omega) e^{-\beta H(\omega) + \beta \gamma |V|} + \sum_{m=1}^{|V|} \xi_m \; e^{\beta \gamma ( |V| - 2 m) }} { \underset{ \omega \in \Omega}{\sum} e^{-\beta H(\omega) + \beta \gamma |V|} + \sum_{m=1}^{|V|} \zeta_m \; e^{\beta \gamma ( |V| - 2 m) } } $$ tends to $\mean{X}$ in the limit where $\gamma$ tends to infinity. A similar argument provides the following identity between the Boltzmann weight for a configuration of the extended space $\omega_{\text{ext}}$ and the Boltzmann weight of its restriction to $\Omega$, $\omega$: \begin{equation}\label{eq:bw_equivalence} \lim_{\gamma \to \infty} \frac{ e^{-\beta H_{\text{ext}}(\omega_{\text{ext}} | \gamma)} } { Z_{ \text{ ext } }( \beta | \gamma ) } = \prod_{ j \in V} \delta_{\rm{ K } }( \sigma_j, \sigma_{ p( j ) } ) \frac{ e^{ - \beta H(\omega) } }{ Z( \beta ) }, \end{equation} where $\delta_{\rm{K}}$ denote the Kronecker delta function. The contribution of any site $j$ of the original lattice $\Lambda$ to the numerator of the r.h.s. of (\ref{eq:bw_equivalence}) reads \begin{equation}\label{eq:num} \delta_{\rm{K}}(\sigma_j, \sigma_{p(j)}) \; \text{exp}\big( \; \beta \; \big( h_j \sigma_j + \sum_{k \in N(j)} J_{jk} \sigma_j \sigma_k \big) \big). \end{equation} where $N(j)$ denotes the neighbourhood of $j$. Because of the Kronecker delta, for any bipartition of this neighbourhood $N(j) = N'(j) \cup N''(j)$, (\ref{eq:num}) remains invariant if the sum in the exponential is substituted with \begin{equation}\label{eq:sub-2} \sum_{k \in N'(j)} J_{jk} \sigma_{p(j)} \sigma_k \; + \sum_{k \in N''(j)} J_{jk} \sigma_j \sigma_k. \end{equation} Assuming w.l.o.g. the boundary conditions represented on Fig.~\ref{fig:triangular-lattice}-left, we choose, for every site $j$, $N'(j)$ to consist in the sites located east, south, and south east of $j$, $\forall j \in \Lambda$. (Edge and corner sites might require different choices of subsets $N'(j)$, depending on the boundary conditions.) This choice results in a \emph{square} lattice hamiltonian $H_{\square}$ whose couplings are shown on Fig.\, \ref{fig:extended-to-square}, and results in the interaction graph displayed on Fig.\, \ref{fig:triangular-lattice}(c). Let $\widetilde{\pi}_{\square}$ denote a probability distribution approximating the Boltzmann distribution associated with $H_\square$ through tensor network renormalisation. To deal with a triangular lattice using a TNMH code for a square lattice, a possibility is a Markov chain where, at each step, a candidate configuration $\omega'_{\text{ext}}$ is drawn according to $\widetilde{\pi}_{\square}$, and the move from the current configuration $\omega_{\text{ext}}$ to this candidate is accepted with Metropolis-Hastings probability: \[ \min \{ 1, \frac{ e^{- \beta H(\omega')} }{ e^{- \beta H(\omega)} } \times \frac{ \widetilde{\pi}_{\square}(\omega_{\text{ext}}) }{ \widetilde{\pi}_{\square}(\omega'_{\text{ext}}) }\}, \] where $\omega$ (resp. $\omega'$) denotes the restriction of $\omega_{\text{ext}}$ (resp. $\omega'_{\text{ext}}$) to $\Omega$. This mapping from a triangular lattice to a square lattice doubles the number of sites but we stress that the bond dimension of the (square) tensor network associated is unchanged and equal to that of the local degrees of freedom ($d=2$ for the Ising model). It would be very interesting to see whether the argument can be extended to three dimensions, and for example map a body centred cubic lattice model to a simple cubic lattice model. A \emph{quantum} analogue of the mapping exists: square PEPS can be used for a triangular quantum spin Hamiltonian. The extended Hamiltonian (operator) now reads $H_\text{ext} = H - \gamma \sum_{ j \in V} \sigma^z_j \sigma^z_{p(j)}$. Proposition \ref{prop:prop-id} still holds true if $\sum_{\omega_{ \text{ ext } } } X( \omega ) \; e^{ - \beta H_{\text{ext}} (\omega_{\text{ext}} | \gamma)}$ is substituted with $\textrm{Tr} ~ X e^{- \beta H_\text{ext}}$. Expressing the trace in the basis of eigenstates of $\{ \sigma^z_j \}$ operators, an analogue of the substitutions (\ref{eq:num},\ref{eq:sub-2}) holds true too. If for example, one wants a TNS approximation of the ground state, one could alternate Trotter steps with applications of the projector $\proj{00}_z+\proj{11}_z$ on each particle of the original lattice and its partner. Actually, a further reduction can be made: one readily checks that the interaction graph transformation shown on Fig.~\ref{fig:triangular-lattice} produces a hexagonal lattice when applied to a square lattice. Therefore, in principle, it should even be possible to study triangular lattices with hexagonal PEPS. \begin{figure}[h] \begin{center} \begin{tikzpicture}[scale=0.2] \draw[thick] ( 1, 1 + 8 ) -- ( 1 + 8, 1 + 8 ); \draw[thick] (1,1) -- (1,9); \draw[thick] ( 1 + 0, 9 + 0 ) -- ( 9 + 0 , 1 + 0 ); \draw[fill,] (1,1) circle [radius=0.5]; \draw[fill,] (1,9) circle [radius=0.5]; \draw[fill,] (9,1) circle [radius=0.5]; \draw[fill,] (9,9) circle [radius=0.5]; \node[anchor=south] at ( 5 + 0, 9.0 ) {\large{$J_h$}}; \node[anchor=south] at ( 0 + 0, 4.5 ) {\large{$J_v$}}; \node[anchor=south] at ( 5.3, 5 ) {\large{$J_d$}}; \draw[thick] ( 14 + 1 + 0, 9 + 0 ) -- ( 14 + 9 + 0 , 1 + 0 ); \draw[thick] ( 14 + 1 + 0, 1 + 0 ) -- ( 14 + 9 + 0 , 9 + 0 ); \draw[fill] (14+1,1) circle [radius=0.5]; \draw[fill] (14+1,9) circle [radius=0.5]; \draw[fill] (14+9,1) circle [radius=0.5]; \draw[fill] (14+9,9) circle [radius=0.5]; \draw[fill,lightgray] (14 + 5,5) circle [radius=0.5]; \node[anchor=south] at ( 14 + 5 + 0 - 2.95, -3.5 + 9 + 0.2 ) {\large{$\infty$}}; \node[anchor=south] at ( 14 + 5 + 0 + 1, -3 + 9 + 0.75 ) {\large{$J_h$}}; \node[anchor=south] at ( 14 + 2.5 + 0 , -3 + 9 - 3.0 ) {\large{$J_v$}}; \node[anchor=south] at ( 14 + 2.5 + 5.25 , -3 + 9 - 3.5 ) {\large{$J_d$}}; \end{tikzpicture} \caption{Couplings in and around a plaquette in the original and extended models (left and right respectively). The new couplings produce a square lattice rotated by a $\pi/4$ angle with respect to the original lattice.}\label{fig:extended-to-square} \end{center} \end{figure} \subsection*{Hard spheres To close this section, we show how tensor network contractions can also be used to implement collective Monte Carlo updates for systems of hard spheres (or disks in two dimensions) \cite{krauth2006smac}. We will combine three ideas for that purpose. The first is a discretisation of the domain that contains the spheres. The second is a shift of perspective where a configuration will not so much be regarded as a collection of locations for the spheres, but rather as the specification for the state of each cell of the volume that contains them (occupied or empty). The third is to use a tensor network to encode possible moves for each cell. We consider a system of $N$ hard disks in two dimensions confined in a square area discretised with a square lattice ($M$ cells). Although this is not essential, we will assume periodic boundary conditions in order to keep the presentation simple. $N$ is fixed, as well as the lattice spacing $\epsilon$. All disks have identical radius. A configuration is said to be \emph{valid} if (i) the centre of each disk is pinned on the intersection of a vertical and a horizontal line of the lattice, (ii) no cell contains bits of matter belonging to different disks. Fig. \ref{fig:example} is an example of a valid configuration. \begin{figure}[h] \begin{center} \begin{tikzpicture}[scale=0.8] \draw[lightgray,thick] (0,0.5) -- (11,0.5); \draw[lightgray,thick] (0,1) -- (11,1); \draw[lightgray,thick] (0,1.5) -- (11,1.5); \draw[lightgray,thick] (0,2) -- (11,2); \draw[lightgray,thick] (0,2.5) -- (11,2.5); \draw[lightgray,thick] (0,3) -- (11,3); \draw[lightgray,thick] (0,3.5) -- (11,3.5); \draw[lightgray,thick] (0,4) -- (11,4); \draw[lightgray,thick] (0,4.5) -- (11,4.5); \draw[lightgray,thick] (0,5) -- (11,5); \draw[lightgray,thick] (0,5.5) -- (11,5.5); \draw[lightgray,thick] (0,6) -- (11,6); \draw[lightgray,thick] (0,6.5) -- (11,6.5); \draw[lightgray,thick] (0,7) -- (11,7); \draw[lightgray,thick] (0,7.5) -- (11,7.5); \draw[lightgray,thick] (0.5,0) -- (0.5,8); \draw[lightgray,thick] (1,0) -- (1,8); \draw[lightgray,thick] (1.5,0) -- (1.5,8); \draw[lightgray,thick] (2.,0) -- (2,8); \draw[lightgray,thick] (2.5,0) -- (2.5,8); \draw[lightgray,thick] (3,0) -- (3,8); \draw[lightgray,thick] (3.5,0) -- (3.5,8); \draw[lightgray,thick] (4,0) -- (4,8); \draw[lightgray,thick] (4.5,0) -- (4.5,8); \draw[lightgray,thick] (5,0) -- (5,8); \draw[lightgray,thick] (5.5,0) -- (5.5,8); \draw[lightgray,thick] (6,0) -- (6,8); \draw[lightgray,thick] (6.5,0) -- (6.5,8); \draw[lightgray,thick] (7,0) -- (7,8); \draw[lightgray,thick] (7.5,0) -- (7.5,8); \draw[lightgray,thick] (8,0) -- (8,8); \draw[lightgray,thick] (8.5,0) -- (8.5,8); \draw[lightgray,thick] (9,0) -- (9,8); \draw[lightgray,thick] (9.5,0) -- (9.5,8); \draw[lightgray,thick] (10,0) -- (10,8); \draw[lightgray,thick] (10.5,0) -- (10.5,8); \draw[fill,lightgray] (6,4) circle [radius=0.9]; \draw[fill,cyan] (2,1) circle [radius=0.9]; \draw[fill,orange] (9,3) circle [radius=0.9]; \draw[fill,blue] (6,7) circle [radius=0.9]; \draw[fill,green] (3,5) circle [radius=0.9]; \draw[fill,red] (7,2) circle [radius=0.9]; \draw[fill,magenta] (10,6) circle [radius=0.9]; \end{tikzpicture} \caption{Example of a configuration of hard disks in a discretised volume. (Periodic boundary conditions assumed.) }\label{fig:example} \end{center} \end{figure} Our goal is to sample uniformly amongst all valid configurations. For that, we will design a Markov chain of collective updates where each disk either stands still or is moved vertically or horizontally by one lattice spacing. A configuration change must comply with the following rules: \begin{enumerate} \item A disk cannot be split. \item A disk cannot be compressed. \item Disks cannot overlap, not even completely (conservation of particle number). \end{enumerate} We will assume the disks are distinguishable and we will associate a label $\{ 1, 2, \ldots, N \}$ to each of them, which is why each disk appears with a different colour in the illustration of Fig. \ref{fig:example}. We will denote $S_0$ the set of empty cells, and $S_\alpha$ the set of cells occupied by disk $\alpha$, $1 \leq \alpha \leq N$. \emph{Given} a valid configuration $\omega$, we associate a tensor $P_j$ with each cell $j$ of the lattice, see Fig. \ref{fig:peps-plaquette}. The index $\sigma$ of this tensor encodes the move that a bit of matter located at cell $j$ would undergo: $\mathcal{M} \equiv \{ 0, -1, +1, -2, +2\}$ for \{stillness, displacement to the left, displacement to the right, downwards displacement, upwards displacement\} respectively. The role of the $w,e,n,s$ degrees of freedom of $P_j$ is to communicate the chosen move at $j$ to its neighbour cells; the indices $w',e',n',s'$ provide the information about the moves made in neighbouring cells to cell $j$. We want to assign values to these tensors $\{P_j\}$ that guarantee moves can only occur between valid configurations. \begin{figure}[h] \begin{center} \begin{tikzpicture}[scale=2] \foreach \x in { 1 } { \foreach \y in { 1 } { \pic[ pic text = $ P_{ j } $ ] at ( \x, \y ) {plaquette-tensor}; } } \end{tikzpicture} \caption{Diagrammatic representation of the PEPS tensor associated with each cell $j$ of the lattice.}\label{fig:peps-plaquette} \end{center} \end{figure} A. \textbf{Initialisation.} For each cell $j$, $P_j(\sigma)^{w',e',n',s'}_{w,e,n,s} = 1 \; \forall \sigma, w', e', n', s', w, e, n, s \in \mathcal{M}$. B. \textbf{Empty cells.} $\forall j \in S_0$, since there is no matter to be moved, we decree that $P_j(\sigma)^{w',e',n',s'}_{w,e,n,s} = 0 \; \forall w', e', n', s',w,e,n,s$ if $\sigma \neq 0$ (holes do not move). C. \textbf{Faithful move communication.} $\forall j$, $P_j(\sigma)^{w',e',n',s'}_{w,e,n,s} = 0$ unless $w=e=n=s=\sigma$. D. \textbf{Rigidity.} Let $j,k$ denote two neighbouring cells covered by a same disk $S_\alpha, \alpha \neq 0$. Let us assume, say, that $j$ is located left to $k$. We impose that $P_j(\sigma)^{w',e',n',s'}_{w,e,n,s} = 0$ if $e \neq w'$. Similar constraints are imposed on all other pairs of cells $j,k$ covered by a same disk and such that $|j-k|=1$. E. \textbf{Prevention of collisions.} By definition, a collision has occurred between two disks $\alpha$ and $\alpha'$ if and only if two bits of matter belonging to $\alpha$ and $\alpha'$ respectively are found a same cell. Therefore, it is necessary and sufficient to forbid all such events in order to prevent a collision. If there is a collision, either one disk is immobile, say $\alpha$, and $\alpha'$ moves by one cell to overlap with $\alpha$ (case A), or both $\alpha$ and $\alpha'$ move to cause the overlap (case B). Case A occurs if and only if there are pairs of adjacent cells $c$ and $c'$ in $S_\alpha$ and $S_{\alpha'}$ respectively which content will occupy a same cell. To prevent the collision, it is sufficient to impose that for each such pair $(c, c')$, the bit of matter contained in $c'$ cannot hop in $c$. There are four such moves to prohibit; they are represented by the four leftmost drawings of Fig.\ref{fig:forbidden-moves}. In case B, $\alpha$ and $\alpha'$ either move along a same direction (case BI) or along perpendicular directions (case BII). Case BI occurs if and only if there are pairs of cells $c$ and $c'$, separated by one cell, in $S_\alpha$ and $S_{\alpha'}$ respectively, which contents are moved closer to each other along a common line. It is thus enough to prevent the events represented by the rightmost drawings of Fig.\ref{fig:forbidden-moves}. Case BII is dealt with similarly, and results in the prohibition of the events represented by the four remaining diagrams of Fig.\ref{fig:forbidden-moves}. Collisions where a bit of matter contained in a cell $k \in S_{\alpha'}$ moves to its left, and lands in a cell $j \in S_{\alpha}$ already occupied by a bit of matter that does not change its position, can be prevented by imposing $P_j(0)^{-1,e',n',s'}_{0000} = 0 \; \forall e',n',s'$. The other A prohibitions admit similar translations into constraints on the tensors, and the six B prohibitions can be enforced likewise. For example the prohibition of the move depicted on the diagram located rightmost top of Fig.\ref{fig:forbidden-moves} translates into $P_j(\sigma)^{+1,-1,n',s'}_{w,e,n,s} = 0 \; \forall w,e,n,s,n',s'$ whenever the left and right neighbours of cell $j$ are occupied by different disks. \begin{figure*} \begin{center} \includegraphics[width=130mm]{ten-forbidden-moves.pdf} \caption{Forbidden moves in the discretised hard disks model. An asterisk in a cell indicate presence of matter, the rhombus symbol stands for a cell that can either be empty or filled. A dot on the side of a cell indicates no move, whereas an arrow indicates a move by one lattice spacing and its direction.}\label{fig:forbidden-moves} \end{center} \end{figure*} The exact contraction of all tensors yields a function $Q( \sigma_1, \ldots, \sigma_M | \omega )$, which value is equal to $0$ if the move $\{\sigma_1,\ldots,\sigma_M\}$ is forbidden from configuration $\omega$, and 1 otherwise. We note that for a fixed assignment $\{\sigma_1,\ldots,\sigma_M\}$, $Q( \sigma_1, \ldots, \sigma_M | \omega )$ can be evaluated exactly. Ideally, we would construct a Metropolis-Hastings Markov chain where the moves are sampled according to the prior \[ \pi_{\textrm{id}}( \sigma_1, \ldots, \sigma_M | \omega ) = \frac{ Q( \sigma_1, \ldots, \sigma_M | \omega ) }{ \sum_{ \{ \tau \} } Q( \tau_1, \ldots, \tau_M | \omega ) }. \] As we don't expect this to be possible, we propose to approximate $\pi_{\textrm{id}}$ through tensor network renormalisation, as we did for Ising models. At fixed volume $M$, the computational cost for constructing the tensors scales as $1/\epsilon^2$. The bond dimension of the tensor network is \emph{independent} of $\epsilon$. As for Ising models, acceptance rates should increase with the bond dimension used in the tensor network renormalisation. If necessary, a complementary strategy to increase acceptance rates is to select a region at each Markov step, and impose that all disks outside of it or touching its boundary remain fixed; such a region would vary from one time step to the next and may even be disconnected. In two dimensions, the hard sphere model is known to exhibit a fluid-solid phase transition for a filling fraction $\eta = \pi a^2 N / A \simeq 0.7$, where $a$ is the radius of the disks, and $A$ denotes the area of the domain that contains them \cite{krauth2006smac} ($A = M \epsilon^2$ here). It would be very interesting to see how a finite value of $\epsilon$ affects this phase transition. Actually, because of the discretisation, the model considered here is, strictly speaking, not the hard sphere model discussed in \cite{krauth2006smac}, for which the disks could in principle occupy any position in Euclidean space. It might be that the phase transition in the limit $\epsilon \to 0$ does not correspond to the transition point of the hard sphere model defined in Euclidean space. But it just might if a different lattice geometry is used. A similar phenomenon occurs in the study of fluids with cellular automata: square lattices do not relate to the Navier-Stokes equation whereas triangular lattices do \cite{Frisch-Hasslacher-Pomeau-86}. A construction similar to Fig.~\ref{fig:peps-plaquette} should hold for hard spheres in three dimensions, and we believe that an analogue also exists for dimer (and dimer-monomer) models. In this latter case, the possibility to rotate dimers by a $\pi/2$ angle produces additional constraints on the tensors. \section{Tensor network renormalisation}\label{sec:tnc} We review the relation between tensor networks and partition function \cite{Nishino96,Nishino97,Murg2005,Levin2007,Xie2012}. The setup is a slight generalization of that of Section \ref{sec:asym}. That is, we consider a nearest neighbour classical Hamiltonian \[ H( \omega ) = \sum_{\langle i,j \rangle} \varphi_{ij}(\sigma_i, \sigma_j), \] on a lattice $\Lambda = (V,E)$, where the local variables $\sigma_i$ now take value in any finite set, which size we are going to denote $d$. For the sake of simplicity, and without loss of generality, we will again only consider squares lattices, and first limit ourselves to two-dimensional systems for now. At fixed inverse temperature $\beta$, the partition function can be expressed as \begin{equation}\label{eq:pf_bond_description} Z(\beta) = \sum_{\omega \in \Omega} \prod_{\langle i, j \rangle \in E} W_{ij}( \sigma_i, \sigma_j ), \end{equation} where $W_{ij}$ is a $d \times d$ matrix, whose entries represent all possible contributions of the bond $\langle i,j \rangle$ to the Boltzmann weight of the model, i.e. $W_{ij}( \sigma, \sigma' ) = e^{-\beta \varphi_{ij}(\sigma, \sigma')}$. As an example, for the Ising model without external magnetic field, the energy associated with a given bond $\langle i,j \rangle$ reads $\varphi_{ij}( \sigma, \sigma') = - J_{ij} \sigma \sigma'$, and the $2 \times 2$ matrix $W_{ij}$ is \begin{equation}\label{eq:transfMatrixFerro} W_{ij} = \begin{pmatrix} e^{\beta J_{ij}} & e^{-\beta J_{ij}}\\ e^{-\beta J_{ij}} & e^{\beta J_{ij}} \end{pmatrix}. \end{equation} We will use the diagrammatical notation in which a tensor is represented by a vertex or a small geometric figure, with as many legs sticking out as there are indices; and where joining two lines represent a contraction of the corresponding indices. For example, a matrix $W_{ij}$ is represented as follows. \[ \includegraphics[scale = 0.7, valign=c]{transferMatrix.pdf}. \] $Z(\beta)$ can be expressed as a tensor network if we shift from a description in terms of matrices associated with the bonds of the lattice (\ref{eq:pf_bond_description}) to a description in terms of tensors associated with its vertices. Let us consider some vertex $i$ with four neighbours and let $e(i)$ denote the vertex to its right. We decompose $W_{i,e(i)}$ as: \[ W_{i,e(i)}(\sigma, \sigma') = \sum_{\mu = 1}^d L_{i}(\sigma,\mu) R_{e(i)}(\mu,\sigma'). \] Graphically, \[ \includegraphics[scale = 0.7, valign=c]{transferMatrix.pdf} = \includegraphics[scale = 0.7, valign=c]{matrixDecompo.pdf}. \] This can be achieved e.g. through a singular value decomposition (SVD) $W_{i,e(i)} = U_{i,e(i)} \Sigma_{i,e(i)} V_{i,e(i)}^\dagger$, and by setting $L_{i} = U_{i,e(i)} \sqrt{\Sigma}_{i,e(i)}$, $R_{e(i)} = \sqrt{\Sigma}_{i,e(i)} V_{i,e(i)}^\dagger$. Similarly, if $n(i), w(i), s(i)$ denote vertices located above, to the left, and below $i$ respectively, three additional SVD provide the decompositions \[ W_{w(i),i}(\sigma, \sigma') = \sum_{\nu = 1}^d L_{w(i)}(\sigma,\nu) R_{i}(\nu,\sigma'), \] \[ W_{i,n(i)}(\sigma, \sigma') = \sum_{\rho = 1}^d B_{i}(\sigma,\rho) T_{n(i)}(\rho,\sigma'), \] \[ W_{s(i),i}(\sigma, \sigma') = \sum_{\tau = 1}^d B_{s(i)}(\sigma,\tau) T_{i}(\tau,\sigma'). \] We associate a a 4-leg tensor $A^{(i)}(\sigma)$ with each site $i$ and each spin value $\sigma$, whose components are \begin{equation}\label{eq:tensor-general} A^{(i)}_{ \mu \nu \rho \tau} = \sum_{\sigma} \; L_{i}(\sigma,\mu) R_{i}(\nu,\sigma) B_{i}(\sigma,\rho) T_{i}(\tau,\sigma), \end{equation} or, in diagrammatic notation: \[ \includegraphics[scale = 0.5, valign=c]{tensorSite.pdf} = \includegraphics[scale = 0.5, valign=c]{tensorSiteV2.pdf}. \] For a system with open boundary conditions, vertices with only three or two neighbours are dealt with likewise. With these tensors, the partition function can be expressed as \begin{equation}\label{eq:pf_vertex_description} Z(\beta) = \mathcal{C}(\{ A^{(i)}\}), \end{equation} where $\mathcal{C}(\{ A^{(i)}\})$ denotes the contraction of all the tensors associated with all sites. The entire process from (\ref{eq:pf_bond_description}) to (\ref{eq:pf_vertex_description}) is illustrated on Figure \ref{fig:TN} for a $4 \times 4$ lattice. \begin{figure} \begin{center} \includegraphics[width=\linewidth ]{TN.pdf} \caption{Graphical depiction of the construction of the TN associated with the partition function of a classical Hamiltonian. (a) We start with a labelling of the vertices of the lattice in consideration. (b) Representation of the $W$ matrices(red circles) associated with each edge; their contraction yields the partition function. (c) and (d) Singular value decomposition of each $W$ matrix, and regrouping into tensors associated with each vertex of the lattice.}\label{fig:TN} \end{center} \vspace{-0.75cm} \end{figure} \begin{figure*} \begin{center} \includegraphics[width=\linewidth]{peps.pdf} \caption{Renormalisation of a PEPS used to apply the TNMH algorithm to three-dimensional systems. The bond dimension is first reduced along horizontal bonds (an index reshuffling allows to regard each row of a PEPS as an MPS), next along vertical bonds.}\label{fig:peps-cutoff} \end{center} \end{figure*} Similarly, one can construct a TN representation of the partition function with some fixed value $x$ for the degree of freedom at site $i$, $Z(\beta | \sigma_i = x)$. It is for instance sufficient that for each neighbour of $i$, $j$, we replace $W_{i,j}(\sigma,\sigma')$ with $W^{(x)}_{ij}(\sigma,\sigma') = \delta_{x,\sigma} W_{i,j}(\sigma,\sigma')$. The ratio of the two quantities, $Z(\beta | \sigma_i = x)/Z(\beta)$, would exactly be the marginal probability $\pi^{(\beta)}_1(x)$ of that spin being in state $x$, that is, a ratio of two TN contractions. Similarly, one can express any conditional probability $\pi^{(\beta)}_k( \sigma_k | \sigma_1 \ldots \sigma_{k-1} )$ as a ratio of two TN contractions. If one were able to compute the TN contractions exactly, one would have a means to sample according to the Boltzmann distribution exactly. First, the marginal probability of the first spin would be computed, and a spin value $\sigma_1$ would be sampled according to the resulting distribution. Then one would move on to the second spin, and express its probability distribution conditioned on $\sigma_1$ as the ratio of two TN contractions. A spin value $\sigma_2$ would then be sampled. And so on. Iterating across the entire system, one would obtain a sample of the underlying Boltzmann distribution. It is however well known that generically, one can only hope to approximate the TN contraction. Thus, working with the previous scheme, only an approximation to the Boltzmann distribution can be obtained. A possibility to correct the bias of the approximate probability distribution and eliminate systematic errors is to use those approximations to propose moves for a reversible Metropolis-Hastings Markov chain, as we do here. For the approximate contraction, we have used one of the simplest schemes available \cite{Murg2005, Schollwoeck2011}. We define $\ket{\mathsf{top}}$ to be the tensor resulting from contracting all the top row tensors along horizontal edges; the remaining free indices after this contraction are legs pointing downward. Similarly, we will call transfer matrix the tensor resulting from a contraction of the tensors along a horizontal bulk row; the transfer matrix resulting from contracting the tensors of row $k$ will be denoted $\mathsf{TM}_k$, $2<k<n$. Finally, in analogy to $\ket{\mathsf{top}}$, we will denote $\ket{\mathsf{bot}}$ the contraction of bottommost tensors. With these notations, the partition function can be expressed as \begin{equation}\label{eq:braket-top-tm-bot} Z(\beta) = \bra{\mathsf{bot}} \mathsf{TM}_{L-1} \ldots \mathsf{TM}_2 \ket{\mathsf{top}}. \end{equation} Both $\ket{\mathsf{top}}$ and $\bra{\mathsf{bot}}$ are matrix product states (MPS), whereas the transfer matrices $\mathsf{TM}_k$ are matrix product operators (MPO), all with a bond dimension and a 'physical' dimension equal to $d$; their length is equal to $L$. Our approximation of $Z(\beta)$ is obtained by estimating the rhs of (\ref{eq:braket-top-tm-bot}) sequentially. We initialise $\ket{\mathsf{partial}_1} \equiv \ket{\mathsf{top}}$, and for $k \in \{ 2 \ldots L - 1\}$, we define $\ket{\mathsf{partial}_k}$ to be an MPS approximation to $\mathsf{TM}_k \ket{\mathsf{partial}_{k-1}}$ obtained by tensor network renormalisation. $Z(\beta)$ is finally approximated with $\braket{\mathsf{bot}}{\mathsf{partial}_{L-1}}$. The cutoff parameter $D$ sets the accuracy of the approximation. There are many methods available for the renormalisation. Throughout this work, we have mostly used the scheme based on successive SVD \cite{Orus2014}. Two-site variational compression has been used to explore equilibration of the two dimensional Ising model with Gaussian disorder \cite{Orus2014}. The same method allows to approximate the partition function of a system where some spins have been set to definite values, $\widetilde{Z}(\beta | \sigma_1 \ldots \sigma_k )$. The only difference is that for such a site $i$ with spin value $\sigma_i$, the tensor (\ref{eq:tensor-general}) is replaced with \begin{equation} L_{i}(\sigma_i,\mu) R_{i}(\nu,\sigma_i) B_{i}(\sigma_i,\rho) T_{i}(\tau,\sigma_i). \end{equation} As claimed in section \ref{sec:asym}, an approximate Boltzmann weight $\widetilde{\pi}(\omega)$ can be evaluated since, using Bayes theorem, this probability can be expressed as \[ \frac{\widetilde{Z}(\beta | \sigma_1)}{\widetilde{Z}(\beta)} \times \ldots \times \frac{\widetilde{Z}(\beta | \sigma_1 \ldots \sigma_n )}{\widetilde{Z}(\beta | \sigma_1 \ldots \sigma_{n-1})}. \] Two remarks are in order. First, if the tensors $A^{(i)}$ are well conditioned and if $D$ is high enough, the approximations to partition functions we construct will be strictly positive. So will then be the approximated probabilities (\ref{eq:tns-prior}), and the TNMH is irreducible. Second, if the tensors $\{A^{(i)}\}$, the MPS $\ket{\mathsf{top}}$, $\ket{\mathsf{bot}}$ and the transfer matrices $\mathsf{TM_k}$ are stored, a TNMH update of the whole lattice can be performed at a computational cost that scales linearly with the lattice size Plaquette interactions can be dealt with similarly. Using singular value decomposition for the Boltzmann and regrouping all the matrices relating a given site, one obtains a ($\pi/4$ rotated) square lattice for the partition function. Bayes formula can thus again be used for sampling. We have dealt with three-dimensional models in a similar fashion. Assuming an $L \times L \times L$ lattice, the identity (\ref{eq:pf_vertex_description}) can again be obtained after sequence of SVD; $A^{(i)}$ is now a six-leg tensor. (\ref{eq:braket-top-tm-bot}) is also still valid, but $\ket{\mathsf{top}}, \ket{\mathsf{bot}}$ are now projected entangled pair states (PEPS), and the transfer matrices $\mathsf{TM}_k$ projected entangled pair operators (PEPO). As in two dimensions, without any cutoff, the bond dimension of $\mathsf{TM}_{L-1} \ldots \mathsf{TM}_2 \ket{\mathsf{top}}$ would grow exponentially with $L$, and renormalisation is in order. There exists a plethora of methods to contract three dimensional tensor networks \cite{lubasch,Xie2012}. We have not aimed at optimality and have opted for simplicity. Again, denoting $\ket{\mathsf{partial}_k}$ the approximate contraction of the first $k$ layers of the TN, the core of the renormalisation consists in constructing a PEPS approximation $\ket{\mathsf{partial}_{k+1}}$ for the contraction $\mathsf{TM}_{k+1} \ket{\mathsf{partial}_k}$. When a PEPO is superimposed on a PEPS, the resulting state is a PEPS with a larger bond dimension. The bond dimension of $\mathsf{TM}_{k+1} \ket{\mathsf{partial}_k}$ has been reduced by first cutting off along horizontal rows, then along vertical rows, see Fig. \ref{fig:peps-cutoff}. Two parameters now govern the accuracy of the approximation: the bond dimension of the PEPS $\ket{\mathsf{partial}_k}$, $D$, and the cutoff for the approximate contraction of two rows of a PEPS, $\chi$ \cite{lubasch}.

cb6b8c86fd56b29428d62a8f3a0b2cd859d880dc

\section*{Abstract} {\bf We investigate the spectral properties of one-dimensional lattices with position-dependent hopping amplitudes and on-site potentials that are smooth bounded functions of position. We find an exact integral form for the density of states (DOS) in the limit of an infinite number of sites, which we derive using a mixed Bloch-Wannier basis consisting of piecewise Wannier functions. Next, we provide an exact solution for the inverse problem of constructing the position-dependence of hopping in a lattice model yielding a given DOS. We confirm analytic results by comparing them to numerics obtained by exact diagonalization for various incarnations of position-dependent hoppings and on-site potentials. Finally, we generalize the DOS integral form to multi-orbital tight-binding models with longer-range hoppings and in higher dimensions.} \noindent\rule{\textwidth}{1pt} \tableofcontents\thispagestyle{fancy} \noindent\rule{\textwidth}{1pt} \vspace{10pt} \section{Introduction} \label{sec:intro} The density of states (DOS) is a key physical quantity in condensed matter physics -- a plethora of electronic properties of solids depend on it, such as conductivities, thermoelectric coefficients, and screening effects \cite{marder2010condensed}. In a broader context, the DOS also shows up in other areas of physics, such as optics, electronics, acoustics, and in fact for any system to which a spectrum can be assigned. Mathematically speaking, all these systems are described by Hermitian operators or matrices, which are typically large or infinite-dimensional. Their spectral features provide the most essential information about the dynamics of the systems, and in particular the energy dependence of their response functions \cite{giuliani2005}. Therefore, the ability to tailor the DOS in condensed matter systems as well as optical and photonic metamaterials is of practical interest and importance to design specific functionalities. In recent years, driven by extensive progress in fabricating photonic and acoustic metamaterials, it is more feasible than ever to manipulate and design the DOS and spectral features of these systems \cite{jacob2010engineering,lu2009phononic}. On the condensed matter side, for example, the groundbreaking fabrications of Moir\'e superlattices in van der Waals heterostructures and two-dimensional (2D) materials have provided a new class of electronic systems with extremely tunable low-energy bands which can host a variety of exotic, strongly correlated, and topological phenomena \cite{bistritzer2011moire,cao2018unconventional}. Almost all existing approaches to control or design spectral properties employ periodic structures or superlattices, for which well-established theories such as the envelope-function approximation exist \cite{bastard1981}. In such systems, including semiconductor superlattices and all kinds of existing metamaterials, the periodicity allows the use of concepts like quasi-momenta $k$, reciprocal space, and the Brillouin zone \cite{smith1990}. In particular, periodicity guarantees the presence of dispersion relations $\omega({\bf k})$ that simplify the theoretical description as compared to nonperiodic cases. Nevertheless, electronic bands also develop in the absence of spatial periodicity and translational invariance, due to the hybridization of neighboring atomic orbitals, as has been widely discussed in the literature \cite{Elliott1974,Anderson1975}. This includes amorphous materials and quasicrystalline structures which have been shown to possess well-defined electronic bands and spectra \cite{Weaire-Thorpe}. On the other hand, one may also think of general noninteracting lattice models in which the hopping integrals and on-site potentials vary with position. Such spatial variations can be a result of locally engineering the chemical structure, position-dependent doping, or the application of external fields and perturbations. The resulting Hamiltonians lack the periodicity of the underlying lattices, and calculation of the DOS even in the non-interacting limit becomes challenging since the quasimomentum $k$ is no longer a good quantum number. In this paper, we explore the spectral properties of lattice models without translation symmetries, in which the parameters of the governing Hamiltonian or dynamics (such as their hopping strengths or on-site potentials) are position-dependent and vary smoothly. We construct a general approximate scheme for calculating their DOS, which becomes exact in the limit of infinitely large lattices. The starting point in this scheme is to use a mixed representation of Bloch and Wannier functions. The basis functions in this representation are defined such that at short scales, they are delocalized similarly to Bloch functions but on longer scales they appear confined to a finite extent, covering a region over which the Hamiltonian parameters vary a little (see Fig. \ref{fig1}(a)). We may recall that Wannier and Bloch functions, in their standard definitions, are maximally localized in real and momentum space, respectively \cite{wannier1937,marzari-RevModPhys}. The essential advantage of using partial Wannier functions is that they can be tuned such that the lattice model Hamiltonian in their basis becomes almost diagonal with small corrections which can be treated in a perturbative manner. We will focus on tight-binding models, although our findings can be applied to all other position-dependent lattice models as long as they can be mapped to some non-interacting Hamiltonian. This includes photonic or acoustic metamaterials, as well as spin models which effectively map to a non-interacting problem, and even a mean-field superconducting Hamiltonian with smooth spatial variations in pairing potential. Recently, such position-dependent lattice models have also been discussed in the context of modeling curved spacetimes \cite{morice2021} to provide gravitational analogies in quantum condensed matter systems \cite{Kedem-Bergholtz-Wilczek}. In what follows, we first introduce a basic position-dependent lattice model and our main results for its DOS in Sec. \ref{sec2}. We then introduce the piecewise Wannierization approach for a 1D tight-binding (TB) model with position-dependent parameters in Sec. \ref{sec3}. By applying perturbation theory, we derive the general formula for the DOS in Sec. \ref{sec4}. Then, in Sec. \ref{sec5}, we consider the inverse problem of constructing a lattice model that yields a given DOS. Subsequently, we elaborate on specific examples and compare the results of the perturbative scheme based on Wannier functions with numerical calculations in Sec. \ref{sec6}. Generalization to higher dimensions and more general TB models are provided in \ref{sec7}, which is followed by the conclusions in Sec. \ref{sec8}. \section{Main result}\label{sec2} To illustrate our main result, we consider the basic example of a finite one-dimensional lattice with hopping amplitudes $t(x)$ between neighbors and on-site potentials $\mu(x)$, both being smooth and bounded functions of position $x$, with $0 < x \leq 1$. We take the lattice to consist of ${\cal N}$ sites at points $x_n=n/{\cal N}$ with $n=1,\cdots, {\cal N}$. This tight-binding model corresponds to the Hamiltonian \begin{eqnarray} {\cal H} = \sum_{n=1}^{\cal N} \mu\big(x_n\big) \: | n \rangle \langle n| + \, \sum_{n=1}^{{\cal N}-1} \bigg[ t\big(x_n\big) \: \bigg( | n \rangle \langle n+1|+ | n+1 \rangle \langle n| \bigg) \bigg] \label{eq:main_Hamiltonian}, \end{eqnarray} in which $| n \rangle$ denotes a real-space localized ket. This Hamiltonian corresponds to a symmetric tridiagonal $\cal N \times \cal N $ matrix with diagonal elements $\mu(x_n)$ and off-diagonal matrix elements $t(x_n)$ for which we wish to determine the density of states $D(\omega)$ in the limit $\cal N \rightarrow \infty$. We will prove that in this case \begin{eqnarray} D(\omega) &=& \Re\: \int_0^1 dx \: \bigg\{{4\,t(x)^2- \big[\omega-\mu(x)\big]^2 }\bigg\}^{-1/2}, \label{dos-approx} \end{eqnarray} where $\Re$ denotes the real part of the integral. We will also show in Sec. \ref{sec7} the generalization of the DOS relation for more general lattice models in which we have the next-nearest-neighbor or further hoppings and more than one orbital per site. Although the form of the integrand in the DOS relation change for these cases, we show that an integral form for the DOS always exists and can be evaluated at least numerically. We should mention that there is no constraint on the functions $t(x)$ and $\mu(x)$, other than boundedness and piecewise continuity and smoothness of their variations on the lattice scale. The boundedness criterion is related to the fact that for any physically reasonable lattice model, the parameters in the Hamiltonian should be finite. \section{Piecewise Wannierization }\label{sec3} In the simple case of uniform and periodic TB models, the momentum eigenstates \begin{eqnarray} | k \rangle ={\cal N} ^{-\frac{1}{2}} \: \sum_{n=1}^{{\cal N} }\: e^{ik\,n }\: | n \rangle~, \label{k-kets} \end{eqnarray} diagonalize the TB Hamiltonian. It should be noticed that for a finite lattice, $k$ takes the discrete values $k_l=2\pi l/ {\cal N} $ with $l=1,\cdots,{\cal N}$ defining the first Brillouin zone (BZ). For the nonuniform models considered here, due to the lack of translational invariance, momentum is not a good quantum number. Nevertheless, we can exploit a mixed real/momentum-space basis, which we call the basis of \emph{partial Wannier functions} (PWFs), to approximately diagonalize the Hamiltonian. There are off-diagonal correction terms that will be shown to vanish for infinitely large lattices, and therefore can be treated in a perturbative manner for finite lattices. Figure \ref{fig1}(a) schematically shows the difference between full Bloch wavefunctions, maximally localized Wannier states, and the PWFs considered here. \begin{figure}[tp!] \centering\includegraphics[width=.85 \linewidth]{scipost-lattice-models-fig1.pdf} \caption{(Color online) (a) Illustration of Bloch functions, maximally-localized Wannier functions, and partial Wannier functions for a 1D lattice model. Partial Wannier states form a mixed basis with states that lie between Bloch states and maximally localized Wannier states. In a coarse-grained picture and on long length scales, the partial Wannier state is localized, but on smaller scales, it appears extended. (b) The division of a full lattice into smaller pieces by which the Hamiltonian can be decomposed to \emph{intra}- and \emph{inter}-chain parts.} \label{fig1} \end{figure} \par We assume that the full TB chain consists of $M_c$ smaller chains each containing $N_{s}$ sites such that ${\cal N} = M_c\, N_{s}$ gives the number of total sites in the full lattice. Each site $n$ can be alternatively labelled with $m_c$ and $n_s$ corresponding to the position of the small chain and the place of the site inside that chain, respectively. This way, we have $n=N_s\, m_c +n_s$ and we can define the notation $| n \rangle \equiv | m_c,n_s \rangle$. The PWFs are now defined as \begin{eqnarray} | m_c, \vartheta \rangle ={N}_{s}^{-\frac{1}{2}} \sum_{n_s=1}^{N_s} \: e^{i\vartheta n_s} \: | m_c,n_s \rangle \end{eqnarray} in which the quasi-momentum $\vartheta$ can take $N_s$ different values $\vartheta_l=2\pi l/N_s$ with $l=1,\cdots,N_s$. As illustrated in Fig. \ref{fig1}(b), this approach fictitiously divides the full lattice into smaller pieces. The PWFs are localized to only a single small chain within the full 1D lattice but within that small chain, they have an extended Bloch form, which leads us to call this approach partial or piecewise Wannierization. The full Hamiltonian likewise is divided into two types of terms, corresponding to whether they only couple the PWFs inside each small chain or couple states between neighboring chains (These terms are labeled ``intra'' and ``inter'', respectively). It should be mentioned that if ${\cal H}$ includes hopping to the $z^{\rm th}$-nearest neighbor, we should ensure $N_s >z$, so that there is only coupling between nearest-neighbor chains. Equipped with the new basis based on PWFs, we can rewrite the Hamiltonian \eqref{eq:main_Hamiltonian} which is readily decomposed into three parts \begin{eqnarray} {\cal H}={\cal H}^{\rm intra}_0+ {\cal H}^{\rm intra}_1+ {\cal H}^{\rm inter}_1~. \label{eq:H-perturbative} \end{eqnarray} The three terms above correspond to the diagonal terms in the PWF basis, block-diagonal corrections to the intra-chain parts of the Hamiltonian due to the spatial variation of $t(x)$ and $\mu(x)$, and block-off-diagonal coupling terms originating from inter-chain hopping terms, respectively. We will see that the two correction terms ${\cal H}^{\rm intra}_1$ and ${\cal H}^{\rm inter}_1$ respectively scale as ${\cal O}\big(M_c^{-1}\big)$ and ${\cal O}\big(N_s^{-1}\big)$ which implies that by assuming both large $M_c$ and large $N_s$, they can be treated as small perturbations. On the other hand, the nonvanishing matrix elements in the PWF basis can be categorized in three different groups: \begin{itemize} \item $\langle m_c, \vartheta^\prime | {\cal H}_{\rm hop} | m_c, \vartheta \rangle$: the hopping contributions to ${\cal H}^{\rm intra}_0$ and ${\cal H}^{\rm intra}_1$; \item $\langle m_c, \vartheta^\prime | {\cal H}_{\rm onsite} | m_c, \vartheta \rangle$: the onsite potential contributions to ${\cal H}^{\rm intra}_0$ and ${\cal H}^{\rm intra}_1$; \item $\langle m_c, \vartheta^\prime | {\cal H}_{\rm hop} | m_c\pm1, \vartheta \rangle$: the inter-chain part of the Hamiltonian, ${\cal H}^{\rm inter}_1$. \end{itemize} The matrix elements $\langle m_c, \vartheta^\prime | {\cal H}_{\rm hop} | m_c, \vartheta \rangle$ are evaluated as \begin{eqnarray} &&\langle m_c, \vartheta^\prime | {\cal H}_{\rm hop} | m_c, \vartheta \rangle\nonumber\\ &&\quad=\frac{1}{{N}_{s}} \sum_{n_s^\prime=1}^{N_s} \sum_{n_s=1}^{N_s} \sum_{n=1}^{{\cal N} } \: e^{i\vartheta n_s -i\vartheta^\prime n_s^\prime} \: t\big(\frac{n}{{\cal N} }\big) \: \langle m_c,n^\prime_s| \bigg( | n \rangle \langle n+1|+ | n+1 \rangle \langle n| \bigg) |m_c,n_s \rangle \nonumber\\ &&\quad= \frac{1}{{N}_{s}} \sum_{n_s^\prime=1}^{N_s} \sum_{n_s=1}^{N_s} \: e^{i\vartheta n_s -i\vartheta^\prime n_s^\prime} \: \bigg[ t\big(\frac{ N_s\,m_c+n^\prime_s}{{\cal N} }\big) \: \delta_{n_s,n_s^\prime+1} + t\big(\frac{ N_s\,m_c+n_s}{{\cal N} }\big) \: \delta_{n_s+1,n_s^\prime} \bigg] \nonumber\\ &&\quad= \frac{\big( e^{i\vartheta } + e^{-i\vartheta' } \big)}{{N}_{s}} \sum_{n_s=1}^{N_s} \: e^{i(\vartheta -\vartheta^\prime) n_s } \: t\big(\frac{ N_s\,m_c+n_s}{{\cal N} }\big). \label{eq:intra-chain-full} \end{eqnarray} Since $n_s/{\cal N} \leq 1/M_c \ll1$, we can Taylor-expand the hopping term up to first order which results in \begin{eqnarray} \langle m_c, \vartheta^\prime | {\cal H}_{\rm hop} | m_c, \vartheta \rangle &\approx& t\big(\frac{ m_c}{{M}_c}\big) \: 2 \cos\vartheta \: \delta_{\vartheta\,\vartheta^\prime} \nonumber\\ &+& t'\big(\frac{ m_c}{{M}_c}\big) \: \frac{\big( e^{i\vartheta } + e^{-i\vartheta' } \big)}{M_c} \: \sum_{n_s=1}^{N_s} \frac{n_s }{N_s^2}\, e^{i(\vartheta -\vartheta^\prime) n_s } + {\cal O}\big(\frac{1}{M_c^{2}}\big), \label{eq:intra-chain} \end{eqnarray} with $t'(x)$ denoting the derivative of $t(x)$ evaluated at $x$. The first term in Eq. \eqref{eq:intra-chain} is diagonal and comes from approximating the Hamiltonian of this portion of the full chain with a system of uniform hopping $t( m_c/ M_c)$. The second term in the right-hand side of Eq. \eqref{eq:intra-chain} consists of both diagonal and off-diagonal terms in general, which are both of order $1/M_c$. In a similar way as above, the matrix elements of the onsite potential terms in the PWF basis can be evaluated, and read \begin{eqnarray} \langle m_c, \vartheta^\prime | {\cal H}_{\rm onsite} | m_c, \vartheta \rangle &=& \frac{1}{{N}_{s}} \sum_{n_s^\prime=1}^{N_s} \sum_{n_s=1}^{N_s} \sum_{n=1}^{{\cal N} } \: e^{i\vartheta n_s -i\vartheta^\prime n_s^\prime} \: \mu\big(\frac{n}{{\cal N} }\big) \: \langle m_c,n^\prime_s | n \rangle \langle n| m_c,n_s \rangle \nonumber\\ &=& \frac{1}{{N}_{s}} \sum_{n_s=1}^{N_s} \: e^{i\,( \vartheta - \vartheta^\prime ) n_s } \: \mu\big(\frac{ N_s m_c+n_s }{{\cal N} }\big) \nonumber\\ &=& \mu\big(\frac{ m_c }{M_c}\big)\: \delta_{\vartheta\,\vartheta^\prime} + \mu^\prime\big(\frac{ m_c }{M_c}\big) \: \frac{1}{M_c} \: \sum_{n_s=1}^{N_s} \:\frac{n_s}{N_s^2} \: e^{i(\vartheta -\vartheta^\prime) n_s } + {\cal O}\big(\frac{1}{M_c^{2}}\big).~~ \label{eq:intra-chain-onsite} \end{eqnarray} Collecting corresponding terms from Eqs. \eqref{eq:intra-chain} and \eqref{eq:intra-chain-onsite} and dropping higher-order terms which are ${\cal O}\big(M_c^{-2}\big)$, we arrive at the intra-chain parts of the Hamiltonian: \begin{eqnarray} &&{\cal H}^{\rm intra}_0= \sum_{m_c,\vartheta} \Big[ t\big(\frac{ m_c}{{M}_c}\big) \: 2 \cos\vartheta + \mu\big(\frac{ m_c }{M_c}\big) \Big] | m_c, \vartheta \rangle \langle m_c, \vartheta |, \\ && {\cal H}^{\rm intra}_1= \frac{1}{M_c}\sum_{m_c,\vartheta, \vartheta^\prime} \Big[ t'\big(\frac{ m_c}{{M}_c}\big) \: \big( e^{i\vartheta } + e^{-i\vartheta' } \big) + \mu^\prime\big(\frac{ m_c }{M_c}\big) \Big] {\cal B}(\vartheta-\vartheta^\prime) | m_c, \vartheta \rangle \langle m_c, \vartheta^\prime |, \end{eqnarray} with ${\cal B}(x)= N_s^{-2} \sum_{n_s=1}^{N_s} e^{i\,x\,n_s }$. Because there is also hopping between the sites from the neighboring small chains, we should consider the matrix elements $\langle m_c, \vartheta^\prime | {\cal H} | m_c+1, \vartheta \rangle$ and their conjugates $\langle m_c+1, \vartheta^\prime | {\cal H} | m_c, \vartheta \rangle$. They can be similarly evaluated as \begin{eqnarray} \langle m_c, \vartheta^\prime | {\cal H} | m_c+1, \vartheta \rangle &=& \frac{1}{{N}_{s}} \sum_{n_s^\prime=1}^{N_s} \sum_{n_s=1}^{N_s} \sum_{n=1}^{{\cal N} } \: e^{i\vartheta n_s -i\vartheta^\prime n_s^\prime} \: t\big(\frac{n}{{\cal N} }\big) \nonumber\\ &&\qquad \times \: \: \langle m_c,n^\prime_s| \bigg( | n \rangle \langle n+1|+ | n+1 \rangle \langle n| \bigg) |m_c+1,n_s \rangle \nonumber\\ &=& \frac{1}{{N}_{s}} \: e^{i\vartheta -i\vartheta^\prime N_s } \: t\big(\frac{ N_s\,m_c+N_s }{ {\cal N} }\big) = \frac{1}{{N}_{s}} \: e^{i\vartheta } \: t\big(\frac{ m_c+1 }{ M_c}\big) ~, \label{eq:inter-chain} \end{eqnarray} by noticing $\vartheta'\,N_s= 2\pi l'$ with $l'$ being an integer inside the range of $[0,N_s]$. Assuming a large number of sites inside each small chain, the inter-chain coupling terms can again be treated perturbatively, since they are of order $1/N_s$ . Then, the inter-chain contribution of the total Hamiltonian reads \begin{eqnarray} {\cal H}^{\rm inter}_1= \frac{1}{N_s}\sum_{m_c,\vartheta, \vartheta^\prime} t\big(\frac{ m_c+1 }{ M_c}\big) \: \Big( e^{i\vartheta } | m_c+1, \vartheta \rangle \langle m_c, \vartheta^\prime | + e^{-i\vartheta'} | m_c, \vartheta \rangle \langle m_c+1, \vartheta^\prime | \Big) . \end{eqnarray} \section{Perturbation theory based on PWFs} \label{sec4} Now that we calculated the terms of the Hamiltonian \eqref{eq:H-perturbative} in the PWF basis, we can apply perturbation theory to this Hamiltonian. Since we already know the matrix elements of all terms in the Hamiltonian, the zeroth and first-order energies can be simply obtained as \begin{eqnarray} E_{m_c,\vartheta}^{(0),\,{\rm intra}}&=& \mu\big(\frac{ m_c}{{M}_c}\big)+2\,t\big(\frac{ m_c}{{M}_c}\big) \cos\vartheta , \label{eq:0th-E} \\ E_{m_c,\vartheta}^{(1),\,{\rm intra}}&=& \frac{1}{M_c} \langle m_c, \vartheta | {\cal H}^{\rm intra}_1 | m_c, \vartheta \rangle = \frac{1}{M_c}\, \bigg[ \frac{1}{2}\,\mu'\big(\frac{ m_c}{{M}_c}\big)+ t'\big(\frac{ m_c}{{M}_c}\big) \cos\vartheta\bigg] , \label{eq:intra-1st} \\ E_{m_c,\vartheta}^{(1),\,{\rm inter}}&=& 0. \label{eq:inter-1st} \end{eqnarray} The first-order correction due to the intra-chain part of the Hamiltonian is nothing but the linear correction to the position-dependence of the hoppings inside the chain, whereas the inter-chain part which is block-off-diagonal has no first-order contribution. We can calculate the second-order correction due to ${\cal H}^{\rm inter}_1$ which reads \begin{eqnarray} E_{m_c,\vartheta}^{(2),\,{\rm inter}}&=& \frac{1}{N_s^2}\sum^{\prime}_{m_c^\prime,\vartheta^\prime}\frac{\big|\langle m_c, \vartheta | {\cal H}^{\rm inter}_1 | m_c^\prime, \vartheta^\prime \rangle\big|^2}{E_{m_c,\vartheta}^{(0)}-E_{m_c^\prime,\vartheta^\prime}^{(0)}} \nonumber\\ &=& \frac{1}{2N_s^2}\,\sum_{\vartheta^\prime}\bigg[ \frac{\big|t\big(\frac{ m_c+1}{{M}_c}\big) \big|^2}{t\big(\frac{ m_c}{{M}_c}\big)\cos\vartheta - t\big(\frac{ m_c+1}{{M}_c}\big)\cos\vartheta^\prime } \nonumber \\ && \qquad\qquad \qquad\qquad\qquad\qquad +\: \frac{\big|t\big(\frac{ m_c}{{M}_c}\big) \big|^2}{t\big(\frac{ m_c}{{M}_c}\big)\cos\vartheta - t\big(\frac{ m_c-1}{{M}_c}\big)\cos\vartheta^\prime } \bigg]. \label{eq:inter-2nd} \end{eqnarray} We note that the double sum in the first line above excludes the single term with $m_c^\prime=m_c$ and $\vartheta^\prime=\vartheta$. The single sum in the second line however, covers all possible values of $\vartheta^\prime$, including $\vartheta$, owing to the fact that we have only nonvanishing matrix elements for $m_c^\prime=m_c\pm 1$. If $\cos\vartheta=0$, we can immediately see that $E_{m_c,\vartheta}^{(2),\,{\rm inter}}$ vanishes since it is proportional to $\sum_{\vartheta^{\prime}} 1/\cos \vartheta^\prime\equiv0$. Otherwise, assuming $\cos\vartheta\neq0$, we arrive at the approximate expression \begin{eqnarray} E_{m_c,\vartheta}^{(2),\,{\rm inter}} \approx \frac{1}{N_s^2}\,\bigg\{\bigg[ t\big(\frac{ m_c}{{M}_c}\big)+\frac{1}{M_c}t'\big(\frac{ m_c}{{M}_c}\big) \bigg]\,\sum_{\vartheta^\prime\neq\vartheta} \frac{1}{\cos\vartheta-\cos\vartheta^\prime} - t\big(\frac{ m_c}{{M}_c}\big) \,\frac{1}{\cos\vartheta} \bigg\} , \label{eq:inter-2nd-final} \end{eqnarray} which is a term of the order of ${\cal O }(N_s^{-2})$. In the limit of $M_c,N_s\gg1$, the energy spectrum of the position-dependent TB model can thus be well approximated by the zeroth-order energy expectation values of the PWFs. For ${\cal N} \to \infty$ we can safely consider both $M_c$ and $N_s$ going to infinity as well, and the approximations we made become asymptotically exact. As an immediate implication of the perturbative scheme based on PWFs, the DOS per site can be written as \begin{eqnarray} D(\omega)&\approx& \frac{1}{\cal N}\sum_{m_c,\vartheta} \,\delta\big( \omega - E_{m_c,\vartheta}^{(0)} \big) = \frac{1}{M_c}\sum_{m_c=1}^{M_c}\,\frac{1}{N_s}\sum_{l=0}^{N_s-1} \,\delta\big[ \omega - \mu\big(\frac{ m_c}{{M}_c}\big) - 2\,t\big(\frac{ m_c}{{M}_c}\big) \cos\vartheta_l \big] \nonumber\\ &=& \int_0^1 dx \int_0^{2\pi} \frac{d\vartheta}{2\pi} \: \delta\big[ \omega - \mu(x) - 2\,t(x) \cos\vartheta \big] \label{dos-approx-extended} \end{eqnarray} using only the lowest-order energies $E_{m_c,\vartheta}^{(0)}$. By performing the integration over the quasi-momentum $\vartheta$, we acquire the DOS relation of Eq. \eqref{dos-approx}. Recall that $\vartheta_l=2\pi l/N_s$ for $l\in {\mathbbm Z}_{N_s}$, and that the integrals in the second line originates from taking the limits $N_s\to\infty$ and $M_c\to\infty$ of the two summations, respectively. As mentioned before, this implies that for infinitely large lattices and assuming smooth spatial variations of Hamiltonian parameters such as $t(x)$ and $\mu(x)$, the zeroth-order energy expression $E_{m_c,\vartheta}^{(0)}$ and the DOS relation of Eq. \eqref{dos-approx} become exact. In Sec. \ref{sec5}, we will compare the DOS profiles obtained from the above analytic relation with those of numerical diagonalization for various finite-sized position-dependent models, to demonstrate the versatile applicability of expression \eqref{dos-approx}. \section{Constructing lattice model to match a given DOS} \label{sec5} We now consider the inverse problem of finding a lattice model which gives a prescribed DOS. In contrast to the more common cases where we know the Hamiltonian, here the Hamiltonian is unknown and will be derived based on knowledge about the DOS. Such inverse problems have already been explored in the mathematics community, yet there are theoretic challenges about the necessary or sufficient conditions for the existence and uniqueness of solutions, and also practical questions about suitable algorithms and numerical methods \cite{chu2005inverse}. As a matter of fact, infinitely many different operators can have the same spectrum and therefore equal DOS. However, we can refine this situation by concentrating on the special family of lattice models with position-dependent hopping and/or onsite-potentials as introduced earlier. We concentrate here on the case of an infinite lattice with just nearest-neighbor hopping and in the presence of the particle-hole symmetry condition $\mu(x)=0$. We then proceed to calculate $t(x)$ such that Eq. \eqref{dos-approx} yields a desired DOS, meaning that we consider the DOS relation as an integral equation. In general, even if there exists a set of $t(x)$ that matches a given $D(\omega)$, finding them as a solution of the integral equation is only possible numerically. However, for the particle-hole symmetric cases, a formal analytical solution exists, since we can transform Eq. \eqref{dos-approx} to, \begin{align} &D(\omega) = \Re \: \int_{t(0)}^{t(1)} dt \: {\cal K}(\omega, t)\: {\cal Y}(t) =\int^{t_{\rm max}}_{|\omega|/2} dt \: {\cal K}(\omega, t)\: {\cal Y}(t ), \label{volterra}\\ &{\cal K}(\omega,t )=\big({4t^2-\omega^2 }\big)^{-1/2}, \end{align} with ${\cal Y}(t)=(dt/dx)^{-1}$ being the inverse of the derivative of the hopping parameter $t(x)$ rewritten as a function of $t$. The expression \eqref{volterra} assuming a given $D(\omega)$ defines a Volterra integral equation of the first kind. For the special kernel ${\cal K}(\omega,t )$ arising here, we show in Appendix \ref{appendix} that a formal solution \begin{align} {\cal Y}(t) = \frac{-1}{\pi} \frac{d}{dt} \int_{2t}^{\omega_{\rm max}} \omega\, d\omega \: \frac{D(\omega)}{\sqrt{\omega^2-4t^2}}\label{solution-volterra} \end{align} exists, where the upper bound $\omega_{\rm max}$ is dictated by the bandwidth above which the DOS vanishes. Note that using the particle-hole symmetry condition $D(\omega)=D(-\omega)$, we can replace the integration over all energies with twice the integral over just positive energies. Having ${\cal Y}(t)$ in hand, we define \begin{equation} {\chi}(t) = \int^t_0 dt' \: {\cal Y}(t')=\frac{1}{\pi} \bigg[ \int_{0}^{\omega_{\rm max}} d\omega \: D(\omega) - \int_{2t}^{\omega_{\rm max}} \omega\, d\omega \: \frac{D(\omega)}{\sqrt{\omega^2-4t^2}} \bigg] =x, \label{solution-t(x)} \end{equation} whose inverse function gives the spatial form of the hopping as $t(x)=\chi^{-1}(x)$. The only limitations to devising a 1D TB model with position-dependent hoppings giving a desired particle-hole symmetric DOS are the integrability criteria for Eqs. \eqref{solution-volterra} and \eqref{solution-t(x)}. Although it is not necessary, having a bounded smooth function $D(\omega)$ provides a sufficient condition for the convergence of both integrals. Therefore by varying $t(x)$ almost any smooth DOS can be obtained. A case of particular interest is when the DOS has singular behavior at particular energies, a special feature known as a van Hove singularity. Simple 1D TB models typically have van Hove singularities at the band edges as $D(\omega)\propto \sqrt{\omega^2_{\rm max}-\omega^2}$, yet it remains finite and differentiable, otherwise. Interestingly, for the position-dependent TB model, we observe that it is possible to generate a DOS with any singularity of the form $\omega^{-\beta}$ at the center of the band as long as $\beta < 1$. But higher-order singularities with $\beta \geq 1$ cannot occur in the nearest-neighbor hopping models as the integral in \eqref{solution-volterra} then becomes divergent. The significance of Van Hove singularities has been noticed long ago, as in their presence the sensitivity of the system to perturbations and interactions is substantially enhanced. The notion of higher-order van Hove singularities, with power-law form $D(\omega)\propto\omega^{-\beta}$ opposed to the logarithmic behavior at an ordinary Van Hove singularity, has also recently been put forward in the context of two-dimensional materials, and particularly Moir\'e superlattices such as twisted bilayer graphene \cite{yuan2019magic,fu-prr2010,Shtyk2017}. In general, as long as there exists a solution, one can perform the integration in \eqref{solution-t(x)} numerically to obtain $t(x)$ which yields a given DOS. But in certain situations, analytical closed forms for $t(x)$ can be obtained starting from a given form of $D(\omega)$. So, to elucidate the mathematical procedure of obtaining hopping parameters $t(x)$, we explicitly examine the two interesting examples of a constant ($D_1(\omega)$) and a semicircular ($D_2(\omega)$) DOS with bandwidth $W=2$. By evaluating the integration and derivative in Eq. \eqref{solution-volterra} we find that \begin{align} D_1(\omega)= 1 &\qquad\Longrightarrow \qquad {\cal Y}_1(t) = \frac{dx}{dt} =\frac{4 t}{\pi \sqrt{1-4 t^2}} \\ D_2(\omega) = \sqrt{1-\omega^2} &\qquad\Longrightarrow \qquad {\cal Y}_2(t) = \frac{dx}{dt} = 2t \end{align} where the solutions for $t(x)$ assuming $t(x=0)=0$ are respectively \begin{align} t_1(x) &= \sqrt{ x \,(1- x)}, \qquad 0\leq x \leq 1, \\ t_2(x) &= \sqrt{ x } , \qquad 0\leq x \leq 1 . \\ \end{align} It is interesting to note that the case of constant DOS is equivalent to having equally-distanced eigenenergies throughout a bounded spectrum and this has been studied for finite tridiagonal matrices \cite{al2009inverse,rothney2013eigenvalues,gladwell2014test}. It has been found that a tridiagonal symmetric matrix of order $n$ with off-diagonal entries \begin{equation} a_{j,j+1}= \frac{\sqrt{j(2n-j-1)}}{2}~~~ (j=1,\cdots,n-2)~, \qquad a_{n-1,n}= \sqrt{\frac{ n(n-1)}{2} }, \end{equation} has equally-distanced eigenvalues which match the form of the hopping parameter $t(x)$ we find here. The only exception is that the lowest off-diagonal entry $a_{n-1,n}$ has an extra prefactor of $\sqrt{2}$ compared to one expected from the trend of $a_{j,j+1}$. Such a difference can be regarded as a finite-size effect and that can be ignored for large $n$. We should note that, unlike the famous example of the quantum harmonic oscillator, which also has constant level spacing, here the range of the energies is bounded. Although a constant DOS for a finite range can be formally obtained assuming a linear dispersion relation $\varepsilon(k)\propto k$, such a dispersion is anomalous, meaning that it does not correspond to any 1D lattice model with uniform hoppings. This example suggests that, unlike the case of position-dependent lattice models, it is not \emph{a priori} guaranteed that any uniform-hopping models exist corresponding to a prescribed DOS. We will address this problem in the following in more detail, and show how one can construct uniform-hopping models from the DOS. As a final remark on the constant DOS, we would like to mention that it has been long-known that the nonsymmetric tridiagonal matrices with entries $a_{i,i+1}=i$ and $a_{i+1,i}=n-i$ also have equally-distanced eigenvalues \cite{sylvester1854theoreme,kac1947random}. These matrices are known as Sylvester-Kac matrices and from a physical point of view, they represent a non-Hermitian 1D position-dependent TB model. \subsection{Correspondence to periodic lattice models} We now consider the construction of lattice models giving a desired DOS, but for the uniform, position-independent case, in the presence of hopping between non-neighboring sites. The Hamiltonian of such uniform lattice models in 1D and for a single orbital reads \begin{equation} {\cal H}_{\rm uniform}=\sum_{n,m} \xi_m \: | n \rangle \langle n+m| \label{eq:uniform-hopping} \end{equation} where $\xi_m$ ($m\geq 1$) denotes the hopping between $m^{\rm th}$ nearest neighboring sites and $\xi_0$ is a constant on-site potential. We will see that, by considering the hopping between distant sites and tuning their relative strengths $\xi_m$, we can also engineer the DOS. This will serve as a kind of correspondence between the position-dependent and position-independent lattice models with short- and long-range hoppings, respectively, meaning that their DOS becomes identical in the limit of very large lattice sizes. Here, a lattice model with short-range and long-range hoppings simply refers to whether the hopping either identically vanishes when the distance between two sites becomes larger than a finite length, or the hoppings between even very distant sites remain non-zero. For a generic single-band model with position-independent hopping such as Eq. \eqref{eq:uniform-hopping}, and assuming periodic boundary conditions, there exists a dispersion relation $\varepsilon(k)$. So, the DOS can be written as \begin{eqnarray} D(\omega)=\int \frac{dk}{2\pi}\: \delta\big[\omega-\varepsilon(k)\big]= \int \frac{dk}{2\pi}\: \frac{\delta\big[k-\varepsilon^{-1}(\omega)\big]}{|d\varepsilon/dk|}=\frac{1}{|d\varepsilon/dk|_{k=\varepsilon^{-1}(\omega)}}, \label{ode-dos-dispersion} \end{eqnarray} where $\varepsilon(k)$ has been assumed to be an invertible function \footnote{This occurs for a completely non-degenerate spectrum}, although it is not a crucial constraint and can be relaxed by dividing the full range of the parameter $k$ into regions for which $\varepsilon(k)$ is monotonic and has an inverse $\varepsilon_i^{-1}(\omega)$ corresponding to the $i^{th}$ region. In the case of completely monotonic function $\varepsilon(k)$ only one $k$ exists and consequently, we can simply integrate the equation \eqref{ode-dos-dispersion} as \begin{eqnarray} k = \pm\int^\omega d\omega'\: D(\omega') \equiv \varepsilon^{-1}(\omega), \label{dos-dispersion} \end{eqnarray} which gives the inverse of the dispersion relation as an integral of the DOS. So the problem reduces to finding the lattice model parameters $\xi_m$ such that the resulting dispersion relation matches with the result of Eq. \eqref{dos-dispersion} for a given DOS. The dispersion relation for the Hamiltonian \eqref{eq:uniform-hopping} is given by \begin{eqnarray} \varepsilon(k)=2\sum_m \xi_m \cos(m k)\equiv \omega~. \label{fourier} \end{eqnarray} This is simply a Fourier series and therefore the Hamiltonian parameters $\xi_m$ can be obtained as \begin{eqnarray} \xi_m =\frac{1}{2}\,\int_{0}^{2\pi} \frac{dk}{2\pi} \: \varepsilon(k) \: \cos(m k)=\frac{1}{2}\,\int_{-\omega_{\rm max}}^{\omega_{\rm max}} d\omega \: \omega \: D(\omega)\: \cos\big[m\:\varepsilon^{-1}(\omega) \big] ~, \label{fourier-coeff} \end{eqnarray} where we have changed the integration variable from $k$ to $\omega$ in the final expression. Invoking Eq. \eqref{dos-dispersion}, we arrive at \begin{eqnarray} \xi_m = \int_{0}^{\omega_{\rm max}} d\omega \: \omega \: D(\omega)\: \cos\bigg[m \int^\omega d\omega'\: D(\omega') \bigg] ~, \label{fourier-coeff-final} \end{eqnarray} which gives $\xi_m$ in terms of the DOS. In the next section, we compare examples of position-dependent hopping models to corresponding position-independent models with the same DOS. We will see that for all periodic models we consider, with a DOS coinciding with that of a position-dependent lattice model, we must include long-range hoppings between far neighbors. \section{Examples and comparison with numerics} \label{sec6} Having derived analytical integral expressions, we explicitly examine some position-dependent TB models, by calculating their DOS using Eq. \eqref{dos-approx} and comparing them with direct numerical calculations of the spectrum. We focus on the case of a 1D chain with power-law variation of the hopping strength: $t_n= [n/({\cal N} -1)]^\gamma$ ($\gamma\geq 0$). Such a power-law form has been motivated previously by the fact that the resulting low-energy (long-wavelength continuum) physics corresponds to a 1D Dirac equation subjected to a gravitational background that possesses a horizon for $\gamma\geq 1$. In addition, as we will see in the following, the DOS for power-law variation with $\gamma\geq 1$ has a van Hove singularity at $\omega=0$. In fact, the DOS for these cases has a closed-form expression given by \begin{eqnarray} D(\omega)= \Re\: \int_0^1 dx \: \frac{1}{\sqrt{\big[2\alpha x^{\gamma}\big]^2- \omega ^2 }} = \frac{-1}{|\omega|} \Im\: \bigg[ \,_2F_1\left(\frac{1}{2},\frac{1}{2 \gamma },1+\frac{1}{2 \gamma };\,\frac{4 \alpha^2}{\omega ^2}\right)\bigg], \label{dos-approx-powerlaw} \end{eqnarray} where $_2F_1(a,b,c;\,z)$ denotes the hypergeometric function with three real parameters $a$, $b$, $c$, and the variable $z$. Using the limiting behavior of the hypergeometric function, we find that, for any $\gamma>1$, the DOS is singular as $D(\omega)\propto\omega^{1/\gamma-1}$ at zero energy ($\omega\to0$). This result is of practical interest because it introduces a model with a divergent DOS at zero energy, in the middle of the band, whereas in ordinary 1D models with uniform hopping constants, possible divergences known as Van Hove singularities occur at the edges of the band. When $\gamma<1$ we find, in contrast, a smooth behavior for the DOS without any singularity at $\omega=0$. For the special case of $\gamma=1/2$, the DOS turns out to be identical to the \emph{Wigner semicircle distribution} \begin{eqnarray} D_{\gamma=1/2}(\omega)=\frac{1}{2\alpha^2} \sqrt{4\alpha^2-\omega^2}, \label{semicircle} \end{eqnarray} as already noticed in the previous section. The Wigner semicircle distribution is known to emerge in $n\times n$ symmetric random matrices with independent and identically distributed entries in the large $n$ limit \cite{wigner1958,tao-circular-law}. Here, in contrast, we find a particular tridiagonal matrix with off-diagonal entries varying as $M_{i+1,i}=M_{i,i+1}=\sqrt{i/n}$ that yield a semicircular form for its DOS. Then, for the marginal case of $\gamma=1$, we can also find a simple form \begin{eqnarray} D_{\gamma=1}(\omega)= \frac{1}{2\alpha} {\rm log}\left( \frac{2\alpha + \sqrt{4\alpha^2-\omega^2} }{|\omega|} \right), \label{linear-hopping-dos} \end{eqnarray} which has a logarithmic divergence at $\omega=0$. As displayed in Fig. \ref{fig2}, we find good agreement between these results and those obtained from numerical calculations for large enough lattices and different exponents $\gamma$. \begin{figure}[tp!] \centering \includegraphics[width=.99 \linewidth]{scipost-lattice-models-fig2.pdf} \caption{(Color online) DOS profiles for 1D lattice models with power-law variations of the hopping parameter $t(x) \propto x^{\gamma}$ for various $\gamma$. Solid red lines are obtained from the analytical relation whereas the blue circles come from numerical calculations using a lattice with ${\cal N}=500$. The left panel shows the Wigner semicircular shape expected for $\gamma=1/2$ whereas the two other cases with $\gamma\geq 1$ show the hallmark divergence of the DOS at zero energy. } \label{fig2} \end{figure} The effects of position-dependent on-site potential assuming both uniform and position-dependent hopping parameters are shown in Fig. \ref{fig3} for various combinations of power-law forms of $\mu(x)$ and $t(x)$. Similar to the case of a power-law form of only the hopping, the power-law varying on-site potential can yield singularities in the DOS as shown in Figs. \ref{fig3}(a-d). As expected, the nonzero on-site potential breaks the electron-hole symmetry and as a result, the position and type of the singularities in the DOS can be tuned by changing $\mu(x)$. From a practical point of view, introducing a position-dependent on-site potential can be done by applying external electric fields or other perturbations. Next, we provide some examples to illustrate how we find the lattice-periodic TB models with long-range hopping corresponding to a given DOS, based on Eqs. \eqref{dos-dispersion} and \eqref{fourier-coeff}. Particularly, we consider the two DOS relations \eqref{semicircle} and \eqref{linear-hopping-dos} which arise for the position-dependent hopping models with $\gamma=1/2$ and $\gamma=1$, respectively. Plugging these forms of the DOS into \eqref{dos-dispersion} we find the inverse of the dispersion-like relations to be \footnote{Using the identity ${\rm arccosh} (x)={\rm log} \big(x+\sqrt{x^2-1} \big)$.} \begin{eqnarray} && k=\varepsilon_{\gamma=1/2}^{-1} (\omega)=2\bigg[\arcsin\bigg(\frac{\omega}{2\alpha}\bigg) +\frac{\omega}{W}\,\sqrt{1-\bigg(\frac{\omega}{2\alpha}\bigg)^2} \bigg]~, \label{disper-gamma-1}\\ && k=\varepsilon_{\gamma=1}^{-1} (\omega)= 2 \arcsin\bigg(\frac{\omega}{2\alpha}\bigg)+\frac{\omega}{\alpha}\:{\rm arccosh}\bigg(\frac{2\alpha}{|\omega|}\bigg)~. \label{disper-gamma-1/2} \end{eqnarray} Then, using \eqref{fourier-coeff}, we can obtain the $m^{\rm th}$ order hopping strength of the periodic TB model corresponding to these dispersion relations. Interestingly, the hopping between far neighbors ($m\gg 1$) for these cases approximately behaves as $\xi_m \propto - (-1)^{m} /m$ for $m\gg1$, corresponding to a long-range hopping model. The numerical results show that it is plausible that there generically exist periodic long-range hopping models and short-range position-dependent hopping models yielding the same density of states. From a mathematical point of view, such a correspondence between position-dependent short-range TB models and long-range lattice-periodic ones, indicates the existence of a similarity transformation between tridiagonal matrices and the so-called Toeplitz (or diagonal-constant) matrices \cite{gray2006toeplitz}. \begin{figure}[tp!] \centering\includegraphics[width=.99 \linewidth]{scipost-lattice-models-fig3.pdf} \caption{(Color online) The DOS profile for various forms of 1D lattice model with position-dependent hopping $t(x)\propto x^\gamma$ and on-site potential $\mu(x)\propto x^\eta$. Solid red lines are obtained from the analytical relation, whereas the blue circles come from the numerical calculations for a lattice with ${\cal N}=500$. One clear signature of on-site potential is to break electron-hole symmetry of the model as is evident from the DOS plots. } \label{fig3} \end{figure} Finally, we examine the deviations between numerically obtained energies using exact diagonalization of a finite lattice Hamiltonian and the analytical results of the perturbative scheme of Sec. \ref{sec4}. As an example, we consider the power-law varying hopping parameter $t(x)=x^2$ and show the numerical and analytical results (obtained from by \eqref{eq:0th-E}) in Fig. \ref{fig4}(a) and their difference in \ref{fig4}(b). The deviation, when we also include the first order corrections given by Eq. \eqref{eq:intra-1st}, is shown in \ref{fig4}(c). The numerical calculations are done for a lattice size ${\cal N}=1600$ and for the perturbative expression, we have divided the full chain into $M_c=40$ smaller chains each consisting of $N_s=40$ sites. The results clearly indicate that the difference between exact numerics and the perturbative framework of PWFs is very small even for a lattice size of the order of ${\cal N}\sim10^3$. Increasing the lattice size, the deviations become smaller and they vanish in the infinite-size limit. \begin{figure}[tp!] \centering\includegraphics[width=0.98 \linewidth]{scipost-lattice-models-fig4.pdf} \caption{(Color online) The profile of energies and their differences for the values obtained from numerical and analytical calculations for the hopping parameter $t(x) \propto x^{2}$ and vanishing on-site potential. The lattice size is chosen to be ${\cal N}=1600$ and we assume $M_c=N_s=40$ when dividing the lattice into smaller chains. (a) Numerically (blue circles) versus analytically (red line) obtained values of energy. (b) Absolute values of the difference between numerically and analytically calculated energies to leading order in Eq. \eqref{eq:0th-E}. (c) The difference for analytically calculated energies with the leading and the first-order perturbation terms and numerically obtained energy values. } \label{fig4} \end{figure} \section{Extension to more general lattice models} \label{sec7} So far we have shown that the DOS of single-orbital 1D TB models with a general position-dependent nearest-neighbor hopping and on-site potential can be approximated by Eq. \eqref{dos-approx}. In what follows, we explain in detail how this result can be extended for more general lattice models, such as multi-orbital (multi-band) cases, or those with farther-neighbor hoppings, and finally for higher dimensions. First, recall that a general TB model consists of $m$ different orbitals (internal degrees of freedom) and also possibly farther-neighbor hoppings. The position-space kets can be labeled by $|{\bm i}, \eta \rangle$ where ${\bm i}$ indicates the site position and $\eta$ accounts for the different orbitals. The general TB Hamiltonian, then, can be written as \begin{equation} {\cal H}_{\rm TB}=\sum_{{\bm i}, {\bm j}} \sum_{\eta,\eta^\prime}^m t^{\eta,\eta^\prime}_{{\bm i}, {\bm j}} \, |{\bm i}, \eta \rangle \langle {\bm j}, \eta^\prime | + \sum_{{\bm i}} \sum_{\eta,\eta^\prime}^m \mu^{\eta,\eta^\prime}_{{\bm i}} \, |{\bm i}, \eta \rangle \langle {\bm i}, \eta^\prime |, \end{equation} in which instead of a single hopping parameter we have a set of hoppings between potentially different orbitals at different sites. Similarly, different orbitals or sub-lattices corresponding to a single unit-cell can have all sorts of different on-site potentials which can be gathered in a $m\times m$ matrix. For lattice-periodic cases, the Hamiltonian is parameterized by a finite set of \emph{position-independent} hoppings and on-site potentials, for which we use the cumulative notations ${\bm t}$ and ${\bm \mu}$. Consequently, we can use a full Bloch basis labeled as $\psi_{{\bf k},\ell}$ to diagonalize the Hamiltonian, which yields $m$ bands $\varepsilon_{\ell,{\bf k}}\big({\bm t} ,{\bm \mu} \big)$ in which we emphasize the implicit dependence on ${\bm t}$ and ${\bm \mu}$. When hopping parameters and on-site potentials vary with position, we can use the generalized form of the PWF basis \begin{equation} | {\bm m}_c, {\bm \vartheta} \,;\, \tilde{\ell} \rangle = \sum_{{\bm n}_s,\ell} c_{\tilde{\ell},\ell} \: e^{i{\bm \vartheta}\cdot{\bm n}_s}\: | {\bm m}_c, {\bm n}_s \,;\, \ell \rangle \end{equation} assuming that the full $d$-dimensional lattice is divided into small pieces whose positions are denoted by ${\bm m}_c$, while ${\bm n}_s$ still indicates the relative location of sites in each small piece (It may help to remember that ${\bm i}\equiv N_s{\bm m}_c + {\bm n}_s$ assuming the full lattice is divided into pieces of square or cubic shape). Subsequently ${\bm \vartheta}$ is the $d$-dimensional quasi-momentum and the coefficients $c_{\tilde{\ell},\ell}$ must be determined such that the corresponding lowest-order energies of PWF $| {\bm m}_c, {\bm \vartheta} \,;\, \tilde{\ell} \rangle$ become diagonal with respect to the their orbital indices $\tilde{\ell}$. Then, following the same procedure which led to Eq. \eqref{eq:0th-E}, the lowest-order energies can be formally written as $\varepsilon^{(0)}_{\tilde{\ell},{\bm \vartheta}}\big[{\bm t}({\bm m}_c),{\bm \mu}({\bm m}_c)\big]$. The approximate lowest-order energies again become asymptotically exact for infinitely large lattices, owing to the fact that the first-order corrections scale as either $N_s^{-1}$ or $M_c^{-1}$, with $N_s$ and $M_c$ being the total number of sites in each small piece of the lattice and the total number of pieces. We can thus write the formal relation \begin{equation} D(\omega)\approx \Re\, \int d^dx \: \int d^d\vartheta \: \sum_{\tilde{\ell}} \, \delta\Big(\omega- \varepsilon^{(0)}_{\tilde{\ell},{\bm \vartheta}}\big[{\bm t}({\bm x}),{\bm \mu}({\bm x})\big] \Big), \label{dos-approx-nD} \end{equation} to find the DOS of a general position-dependent multi-band TB model in $d$-dimensions and in the infinite-size limit. In general, the expression \eqref{dos-approx-nD} cannot be analytically evaluated except for special forms of position-dependent ${\bm \mu}({\bm x})$ and ${\bm t}({\bm x})$. Nevertheless, one can always evaluate the integrals numerically to find the DOS. We should also mention that typically numerical evaluation of integrals such as those in Eq. \eqref{dos-approx-nD} is not computationally expensive when the size of the matrices becomes very large, as opposed to the direct numerical calculations based on exact diagonalization of the Hamiltonian. Furthermore, and similar to the 1D single-orbital case discussed in Sec. \ref{sec5}, we can treat the expression as an integral equation that corresponds to the inverse problem of devising the model by knowing its DOS. Therefore, employing the numerical methods for solving integral equations, in principle, we can numerically compute ${\bm \mu}({\bm x})$ and ${\bm t}({\bm x})$ as the solutions of the inverse problem. \section{Conclusions}\label{sec8} We have considered a general tight-binding model with smooth position-dependent parameters and obtained an expression for its DOS. This expression, which becomes exact in the infinite-size limit, has been derived using a mixed basis interpolating between Bloch and maximally localized Wannier functions. For the one-orbital 1D case, we found the exact solution of the inverse problem, \emph{i.e.} finding the spatial variation of hopping parameter for a given DOS. Then, we constructed a correspondence between position-dependent short-range hopping models and those having position-independent but long-range hoppings. By extension of the framework to the most general (higher dimensional multi-orbital) non-interacting lattice models, we obtained an integral form for the DOS in terms of a general hopping matrix. Our findings provide a concrete method of engineering the DOS by manipulating hopping parameters, which paves the way for a variety of applications.

60b8cf2fd054ae505f357ad780be75c5bfbdb16c

\section{Introduction} \label{sec:introduction} Human beings are the only known species that use spoken language to communicate, having a skill developed in communication that uses factors other than speech, such as, for example, body expressions and gazing~\cite{cassell2000embodied}. In this context, Embodied Conversational Agents (ECAs) are virtual agents which are able to interact and talk with humans in a natural way. In the last years, many research have been conducted to improve the quality of the communication abilities of such ECAs, both verbal and non-verbal~\cite{yalccin2020empathy,biancardi2019computational,sajjadi2019personality}. A fair amount of effort has been focused on ECAs that help people to have a healthier life~\cite{kramer2019developing,spitale2020multicriteria,das2019generation}, in clinical interviews~\cite{philip2020trust,martinez2019assessment}, training specific skills~\cite{chetty2019embodied,ayedoun2019adding}, among others. The present work aims to propose an Embodied Conversational Agent (ECA) with general purpose, called Arthur, endowed with a memory module and many other abilities. Besides a conversational module, using text and voice, this ECA is able to recognize the person he is talking to, as well to assess the user emotional state through his/her facial expressions. Also, Arthur is able to demonstrate different levels of emotion through his facial expressions. Lastly, Arthur is equipped with a memory module, which tries to replicate the behavior of human memory and, thus, allows for Arthur to learn information with and from the user while interacting. Moreover, it allows Arthur to remember such information later in the conversation or, even, in a different interaction. Section~\ref{sec:relatedWork} presents several work related with the model proposed in this work. Section~\ref{sec:method} details our methodology, while Section~\ref{sec:results} shows some results achieved when users know Arthur. Finally, Section~\ref{sec:conclusion} presents our final considerations. \section{Related Work} \label{sec:relatedWork} This Section presents some work related with methodologies used to build Arthur. Section~\ref{sec:chatbots} presents two interesting researches developed in the context of chatbots, while Section~\ref{sec:human-memory} presents papers that discuss the behaviour of human memory and its use for virtual agents. Finally, Section~\ref{sec:will-related} discusses the importance of appearance and expressiveness for virtual agents, also presenting work related with Embodied Conversational Agents (ECAs). \subsection{Chatbots} \label{sec:chatbots} Zhou et al.~\cite{zhou2018emotional} focus their work on the proposal of the Emotional Chat Machine (ECM), a chatbot that can generate relevant responses to content that are also emotionally consistent. As discussed by the authors, there are three main challenges: obtaining high-quality data marked by emotions in a large-scale corpus, considering emotions in a natural and coherent way and incorporating that information of emotions into a neural model. The results achieved by their work show that ECM can generate appropriate responses, if the emotion category of the response and the emotion of the post both belong to one of the frequent Emotion Interaction Patterns (EIP). EIP is defined as the pair: categories of emotion of the publication and its response. Zhang et al.~\cite{zhang2019consistent} propose to solve the problem of consistency on chatbot responses, concerning both context and personas (casual speaker). In their work, they present a self-supervised approach that uses the natural structure of conversational data to learn and leverage both topic and persona features. For the tests, two data sets are used: Twitter FireHose, collected from 2012 until 2016; and Maluuba dataset. The results achieved indicate that the proposed model is able to capture meaningful topics and personas' features. Also, the incorporation of the learned features helps to significantly improve the quality of generated responses on both data sets, even when comparing with models which explicit persona information. In our method, we mix a pre-defined dialog with an existent technology to provide chat and voice, in order to build the communication module of Arthur. More details are given in Section~\ref{sec:chat-voice}. \subsection{Human and Agent Memory} \label{sec:human-memory} One of the most accepted models concerning human memory cited in literature is known as Autobiographical Memory. As defined by Bluck et al.~\cite{bluck1998reminiscence}, autobiographical memory is "a system that encodes, stores and guides retrieval of all episodic information related to our person experiences". Also, according to Conway et al.~\cite{conway2000construction}, autobiographical memory can be grouped in three levels: lifetime periods, general events and event-specific knowledge. So, such memories can be directly accessed if the cues are specific and relevant to the person. Otherwise, if the cues are too general, a generative retrieval process must be used to produce more specific cues for the retrieval of relevant memories. Following this definition of autobiographical memory, Wang et al.~\cite{wang2016modeling} build a model to mimic such behavior. Their model, known as Autobiographical Memory-Adaptive Resonance Theory (AM-ART), is a three-layer neural network that encode lifetime periods, general events and event-specific knowledge, respectively, being, therefore, consistent with the model presented by Conway et al.~\cite{conway2000construction}. Also, it encodes the 5W1H schema, which represents \textit{when} an event occurred, \textit{where} it happened, \textit{who} was involved, \textit{what} happened, \textit{which} pictorial memory was associated with the event and \textit{how} was the person feeling during the event. The results achieved by their work show that AM-ART was able to perform better that the keyword-based query method, since the last can not deal with noisy cues in many existing photo or memory repositories. The main difference between the work presented in this Section and ours is that we do not train a neural network. We modeled our memory in a procedural way, similarly to~\cite{kope2013modeling}. Also, our memory retrieval can be guided by the emotional state and, even, change the mood of the agent. More details are presented in Section~\ref{sec:agent-memory}. \subsection{Appearance and Expressiveness} \label{sec:will-related} When developing an interactive agent, there is a concern about how it presents itself. It is important that the interaction occurs so that the subject in contact with this agent does not feel uncomfortable and/or embarrassed. As shown in Dill et al.~\cite{dill2012evaluation}, when trying to present a character who pretends to be human, there is a certain strange feeling (Uncanny Valley) when it seems human, but not as close as expected. It should be noted that, considering the face of a character, the elements that can cause more strangeness are the eyes and the mouth. Real human eyes tend to not stay static for too much time, having some involuntary or voluntary movement at some point. Such phenomena is known as saccade behavior, being defined by Leigh and Zee~\cite{leigh2015neurology} as rapid movements of both eyes from one gaze position to another. Lee et al.~\cite{lee2002eyes} propose an algorithm to simulate such behavior. The authors obtained a statistical model through the analysis of eye-tracking images from a subject during a conversation with a software. In combination with literature data, they achieved a model capable of generate movements considering various aspects, like magnitude and duration, of a human saccade. As result, the model obtained a natural and friendly perception from subject into author's survey. In this work, we chose to model our virtual agent in a cartoon manner to avoid the strange feeling discussed by Dill et al.~\cite{dill2012evaluation}. Also, we included saccade eyes movement, following the algorithm proposed by Lee et al.~\cite{lee2002eyes}. More details are presented in Section~\ref{sec:facial-express}. \section{Proposed Model} \label{sec:method} This work aims to build an interactive empathetic ECA which can store and recover memories about previous interactions: Arthur. One of his features is being able to recognize the person he is talking to. Also, Arthur has a memory to store knowledge about people he interacts, such as their face, emotion and interactions. Finally, in order to keep the interactions as natural as possible, the agent is able to properly react to interactions with people, delivering meaningful responses and facial cues. \subsection{Overview} \label{sec:overview} The overview of Arthur is illustrated in Figure~\ref{fig:overview-model}. This work was mostly developed using Unity3D~\cite{Unity}. Some modules were developed using Python, as, for example, the Face Recognition module. \begin{figure}[!htb] \centering \includegraphics[width=\linewidth]{images/OverviewMario.png} \caption{Overview of the proposed model. Behavior Control is responsible to define the appropriate behavior of the virtual agent, while the Memory Control is responsible of store and retrieve memories. The dolls inside the boxes represent the modules where an user input exist.} \label{fig:overview-model} \end{figure} Behavior Control is responsible for defining Arthur's appropriate behavior. This module uses all the information available to Arthur, that is, the person recognized by Arthur and with whom he is interacting, the user's facial expressions, facts and information stored in Arthur's memory, the user's emotion detected and the dialogue through text/voice. The Chat and Voice Detection modules are responsible by the interaction between the user and the agent, in the scope of verbal and textual behavior, where the Chat Module allows for the user to type some sentence and the Voice Detector Module is able to transform the voice of the user into words, which are used as input to the agent. The Face Recognition Module allows Arthur to recognize the person he is talking to, while the Emotion Detection Module allows him to identify the person's perceived emotion. The Facial Expressions Module is responsible to animate the facial expressions of Arthur, such as emotions and eyes movement. Finally, Memory Control is responsible for managing the memory of the virtual agent and is linked to all memory resources (i.e., memory learning, memory retrieval, memory consolidation, general events and ESK). These aspects will be explained in Section~\ref{sec:agent-memory}. We use a common method to define artificial memories that is to use Short and Long term memories, where information is transferred from one to the other (STM and LTM, respectively)~\cite{loftus2019human}. So, during the interaction between Arthur and the user, the information is stored in STM. Then, a phase called Memory Consolidation Module deals with the consolidation of the Long-Term Memory (LTM) of the agent. Therefore, saved memories can be forgotten by Arthur if their importance is too low or if they are not used for too long. Such process is detailed in Section~\ref{sec:memoryConsolidation}. Next sections give more details about each of the modules. \subsection{Chat and Voice Detector} \label{sec:chat-voice} The Chat module allows for the user to communicate with the virtual agent, by speaking or writing words and sentences. The user can use its own voice to interact with the virtual agent or write sentences. The Voice Detection module is able to transform the voice input into words, which can be understood by the agent. To do so, the DictationRecognizer~\footnote{https://docs.unity3d.com/ScriptReference/Windows.Speech. DictationRecognizer.html} class was used. It is available for C\# and, thus, for Unity3D, allowing to access the system scripts which deal with voice translation to text. Such tool was used for the ease of use, once it is already integrated with Windows itself. \subsubsection{Arthur Conversation} \label{sec:conversation} At the beginning of the conversation, we modeled some initial questions that our agent can ask the person it is talking to. These questions serve both as an ice-breaker and to know some basic information about that person, that are going to be stored in the memory. To the extent of this work, the modeled questions are: \begin{itemize} \item How old are you? \item Do you work? \item Do you study? \item Do you have children? \end{itemize} Each question can, also, have another question(s) linked with them. For example, if the person answers that it has children, the virtual agent can ask how many children, or the names of them. All this information is stored in the memory of the agent, as further explained in Section~\ref{sec:agent-memory}. If, at any time, the user asks something about a certain subject that Arthur already knows, he can answer. For example, suppose that John tells Arthur that he is 42 years old. Then, in another conversation, someone asks the virtual agent "how old is John?". In this case, Arthur is able to answer that John is 42 years old. Finally, if Arthur is unsure about how to handle a question, the user's written/spoken sentence is sent to a chatbot API~\footnote{https://rapidapi.com} that processes the question and returns an answer, which is presented to the user. Since we are using an API for chatbot responses, it is quite easy to use another chatbot model, if necessary. \subsection{Face Recognition} \label{sec:face-recog} The face recognition module allows Arthur to recognize the person he is talking to. To do this, we base our work on the Python library for face recognition~\footnote{Available at https://github.com/ageitgey/face\_recognition}. In summary, this library is a trained neural network that allows to recognize people's faces by comparing a particular image with other images saved in a data directory. If the person photographed has a low chance of being in the database (returned by the method), it is a new person, so it is stored in the dataset for further consultation. For more information, see the link in the footnote. \subsection{Emotion Detection} \label{sec:emotion-detect} The Emotion Detection module provides to Arthur the skill to detect the emotion of the user. We use the Affectiva~\footnote{https://developer.affectiva.com/} plugin, available for Unity. Affectiva is a well accepted framework which uses a trained neural network to identify many points on the person's face. Depending on the arrangement of such points, the network is able to detect the emotion that the person is expressing. For more information, see the link in the footnote. \subsection{Agent Memory} \label{sec:agent-memory} Arthur's memory is used to record data collected during interactions with people. For this, we adopted a model known as Autobiographical Memory. In short, as defined by Bluck et al.~\cite{bluck1998reminiscence}, autobiographical memory is "a system that encodes, stores and guides retrieval of all episodic information related to our person experiences". Therefore, it is capable of storing different types of information (for example, text, episodes, images, voice) and retrieving it, if a certain suggestion (or suggestions) is informed. For example, when someone meets a new person, they usually greet each other and say their names. If the memory serves that person well, he/she will be able to remember the other person's name when they meet again. According to Conway et al.~\cite{conway2000construction}, autobiographical memory can be grouped in three levels: Lifetime Periods, General Events and Event-Specific Knowledge. Lifetime Periods serve as an index to cue General Events and can be linked to different periods across the lifespan of a person. For example, someone's first job can be seen as a lifetime period. General Events are events that occur within a certain Lifetime Period. Using the first example of work given earlier, some general events within this can be the first day of work, meeting a new colleague or a happy hour with people in the office. Finally, the Event Specific Knowledge (ESK) includes a set of resources that form the memory of a given event. This information can be stored in different types, such as images, grammatical or audio. As it is possible to see in Figure~\ref{fig:overview-model}, we model our memory using General Events and ESK. As the intent of our virtual agent is to interact with people, we chose not to model Lifetime Periods. General Events represent the events which occur during the interaction between Arthur and the user. For example, when Arthur meets someone new, a new General Event is generated (i.e. Meet new person). More details about General Events are given in Section~\ref{sec:generalEvents}. Moreover, following the autobiographical memory model, General Events are comprised of a set of resources known as ESK. In this work, we store such resources in two levels: Short Term Memory (STM) and Long Term Memory (LTM). More details are given in Section~\ref{sec:stmAndLtm}. \subsubsection{General Events} \label{sec:generalEvents} In our model, General Events are comprised of: \begin{itemize} \item Timestamp: the moment this event was created or updated. \item ID: an unique ID which refers to this event. \item Type: refers to the type of the event and there are currently three possible types: \textit{i)} Meet new person: when the agent meets someone new; \textit{ii)} Learn thing: when the agent learn something new; and \textit{iii)} Interaction: when it does not fit in any of the other types). \item Emotion: the emotion associated with the event, recognized from the Emotion Detection module. For example, if a person is talking about the death of its beloved pet, it is expected that such person assumes a sad face. So, this general event should be associated with the Sadness emotion. \item Polarity [-1,1]: the polarity of the sentence, it means, if it is a positive or negative sentence. For example, the sentence "I am feeling good" would be a positive sentence, while the sentence "I am not feeling good" would be a negative one. When we divide the sentence into tokens, we also use the sentiment library\footnote{https://www.nltk.org/howto/sentiment.html}, provided by NLTK, to calculate the polarity of the given sentence. To do so, we chose to work with Vader method. It returns a float value lying between -1 and 1, where negative values reflect a negative polarity and positive values reflect a positive polarity. \item Resources: a list with the resources (i.e. ESK) associated with this General Event. Using the previous example, some resources associated with this event could be a picture of the deceased pet and some grammatical information like the pet name, death and when it passed away. \end{itemize} \subsubsection{ESK: STM and LTM} \label{sec:stmAndLtm} As commented before, we store the resources (i.e. ESK) in two levels: STM and LTM. According to Loftus et al.~\cite{loftus2019human}, STM is used to store important information for a short period of time, while LTM is an information storage with virtually unlimited capacity that each human being has. In our work, we model STM and LTM separately. The STM is comprised of a list of resources. This list of resources can have, at most, seven items, as defined by Miller's Law~\cite{miller1956magical}. If a new resource should enter the STM, the less important information is forgotten according its weight (i.e. the resource with the lower weight is removed). Each resource has the following information: \begin{itemize} \item Timestamp: the moment this resource was created or updated. \item ID: an unique ID which refers to this resource. \item Type: refers to the type of the resource (e.g. grammatical, image, audio). \item Information: refers to the information itself that the resource should reflect. For example, if the resource is an image, the information is going to contain the path to such image. \item Activation [0,1]: As described in Loftus et al.~\cite{loftus2019human}, an information present in the short-term memory is rapidly forgotten (around 15 seconds), unless it is rehearsed, it means, repeated over and over. The activation represents this rehearsal process. When a new resource enters the STM of Arthur, activation is set at its maximum value (i.e. 1). As time passes by, a logarithmic-based decay function is applied for each resource, diminishing their activation value. It is defined as follows: $A_{ID_{STM}}^* = Log(A_{ID_{STM}} + 1)$, where $A_{ID_{STM}}^*$ is the new activation value of STM for resource $ID$, $A_{ID_{STM}}$ is the current activation value and $Log()$ is a logarithmic function. If any resource in the memory is rehearsed (e.g. remembered, seen again), its activation value is set back to 1. This attribute exists only for STM. \item Weight [0,1]: represents the importance of the resource. For example, meeting a new person can be considered more important than talking about the weather. We empirically defined the initial importance of each resource (at STM) based on the Type of the General Event it belongs to, as follows: Meet new person: 0.9; Learn thing: 0.9; Interaction: 0.1. At LTM this attribute is not initialized but indeed affected by the values of $W_{ID_{LTM}}^* = f(A_{ID_{STM}}^*,W_{ID_{STM}}^*)$ at STM, in the consolidation of the memory, as explained in Section~\ref{sec:memoryConsolidation}. \end{itemize} In order to store grammatical resources, we divide a given sentence in significant tokens and keep each of them separately. To do so, we use the Natural Language ToolKit (NLTK)\footnote{https://www.nltk.org/}. since it is a well accepted tool for natural language processing. It is a platform for building Python programs to work with human language data. Such tool is able to "tokenize" a text, it means, split it in sentence or word tokens. It is also able to remove "stop words" from the text, it means, words which have low or no meaningfulness for the context of the sentence. We built a script which is able to remove the "stop words" from the text and "tokenize" it word by word. For example, assuming the user told to the virtual agent "I am going on vacation with my dad to Glasgow", the tokens returned by the NLTK script would be "vacation", "dad" and "Glasgow". The same type of resources can be stored in STM and LTM, but the LTM is theoretically unlimited, it means, can maintain an unlimited number of resources. To keep track of the content of the Long-term memory, we save all its information in a database, so we can retrieve it at any moment. \subsubsection{Memory Retrieval} \label{sec:memoryRetrieval} Assuming the information is stored, how can one retrieve such information if required? The Memory Retrieval module deals with the retrieval of the information from the storage. According to Conway et al.~\cite{conway2000construction}, there are two types of memory construction: Generative Retrieval and Direct Retrieval. Generative retrieval is a method guided by cues, it means, the retrieval depends on some "hint" in order to find some information. For example, the word "dog" can make some people remember of its beloved pet. Thus, the word "dog" acted like a cue to the generative retrieval method, constructing a memory of a special dog. In its turn, Direct Retrieval method refers to spontaneously recalled memories, it means, memories that are recalled automatically, with no apparent cue. According to Berntsen~\cite{berntsen1996involuntary}, such process occurs between two to three times each day. In our work, the Memory Retrieval is developed following the Generative Retrieval method. When the user interacts with the virtual agent, the information provided can be used as cue(s) to the Retrieval method. For example, if the user tells "I went fishing with my dad", the words "fish" and "dad" are used as cues for the retrieval method, so we proceed with a search of those words in the database. \subsubsection{Memory Consolidation: from STM to LTM} \label{sec:memoryConsolidation} This module is responsible for consolidating the Long-term Memory of the virtual agent based on data available at STM. According Klinzing et al.~\cite{klinzing2019mechanisms}, the formation of such LTM is a major function of sleep. As their work affirms, the authors "consider the formation of long-term memory during sleep as an active systems consolidation process that is embedded in a process of global synaptic downscaling". In short, the consolidation process prioritizes important memories over mundane ones. For example, emotional memories are more important and have more impact than neutral memories. Therefore, less important information can be placed in a second plan or, even, forgotten. In our work, we developed a module to simulate such process, which uses two information from the STM's resources: Activation and Weight. If the Activation value $A_{ID}$ of a given resource $ID$ is below an empirically defined threshold (i.e. $A_{ID}<0.2$), the Weight attribute of this resource is reduced at LTM, as follows: $W_{ID_{LTM}}^* = Log(W_{ID_{STM}} + 1)$, where $W_{ID_{LTM}}^*$ is the new Weight this resource will have at LTM, $W_{ID_{STM}}$ is the current weight at STM, and $Log()$ is a logarithmic function. Such process is repeated for all stored resources at STM. Then, resources which have low importance are wiped out from the LTM. We empirically define that a resource has low importance if its weight drops below 0.2. In its turn, General Events are not forgotten by the virtual agent, unless all the resources belonging to a given event are also forgotten. We do so to make possible for the agent to forget just parts of the information, just like it happens with a real person. In addition, please notice that during the memory consolidation, all data at STM is erased. \subsection{Facial Expressions} \label{sec:facial-express} In order to model our agent facial characteristics, we proceeded the following way: the various parts of agent's face (such as eyes, mouth, etc) where created using Unity3D sprites. Then, after placing each of them into their position, a set of constraints were defined in order to avoid further distortions, e.g., when agent's head moves, the other parts follow that movement in a proper way. Finally, the different expression where modeled using a plug-in called Anima2D~\footnote{\label{foots}https://assetstore.unity.com/packages/essentials/unity-anima2d-79840}. The processes of animation consists in the usage of skeletal animation. For a given mesh in the face that should suffer some kind of deformation, such as the eyebrows and mouth, a set of bones is defined and linked to this mesh. After that, it is possible for us to animate the agent through the time just by applying transformations into the bones, which is done through Anima2D\footnote{See footnote 7}. Therefore, after animating a certain facial expression (which can by either animated or a static one), we can save it for further usage. The expressions of our agent are: anger, disgust, doubt, fear, joy, sadness, surprise and worry. Also, there is a neutral expression and one for sleeping (in order to indicate that the memory consolidation process was triggered). Finally, in order to give some expression to the eyes of the virtual agent, it was implemented saccade movements on them, according to Lee et al~\cite{lee2002eyes}. As commented in Section~\ref{sec:will-related}, human eyes are not static for much time, tending to have involuntary movements, even when the person is focusing its vision on something. Such phenomena is known as saccade, from the french word \textit{saccade}, which means \textit{jerk}. As defined by Leigh and Zee~\cite{leigh2015neurology}, saccades are rapid movements of both eyes from one gaze position to another. \section{Results} \label{sec:results} With all features presented in Section~\ref{sec:method} working altogether, we developed an Embodied Conversational Agent (ECA), named Arthur, able to recognize the person he is talking to, as well the emotions hinted by this person's facial expressions.Arthur is also capable of demonstrate some level of emotions, depending the situation. Also, it can learn things from the user and remember such information when needed. Figure~\ref{fig:arthur} shows Arthur and its interface. We chose to model Arthur in a cartoon manner to avoid the Uncanny Valley, as commented in Section~\ref{sec:will-related}. At the bottom, the image captured by the camera is shown, with the face of the person talking with Arthur. The green square is where the name of the recognized person is presented. The blue square is where the person can chat with Arthur. The larger text area is where Arthur speaks, while the smaller input text below is where the user can write his/her sentences. The red square is where the person's emotion is shown. Finally, in the yellow square, two buttons: Voice and Sleep. The Voice button toggles the voice management, allowing for the user to talk with Arthur using its own voice, while the Sleep button puts Arthur to sleep, triggering the memory consolidation process, as explained in Section~\ref{sec:memoryConsolidation}. \begin{figure}[!htb] \centering \includegraphics[width=\linewidth]{images/arthur.png} \caption{Arthur and its interface. The green square is where the name of the recognized person is presented. The blue square is where the person can chat with Arthur. The red square is where the person's emotion is shown. Finally, in the yellow square we have the Voice and Sleep buttons, which allow for the user to talk naturally with Arthur and puts it to sleep, respectively.} \label{fig:arthur} \end{figure} In order to test our method, we developed some experiments exploring the various features of Arthur. Section~\ref{sec:result-agent-memory} shows some emergent behaviors regarding Arthur's memory while Section~\ref{sec:results-user-study} presents research carried out with subjects on the functioning and behavior of Arthur. \subsection{Agent Memory} \label{sec:result-agent-memory} In order to evaluate the Arthur's memory, we test some possible scenarios: \begin{itemize} \item Introduction Scenario: Arthur meets a new person and starts to asking questions about him/her. After that, we check if Arthur can remember all information. \item Learning Scenario: we feed Arthur with information about some objects. Later, we check if it recognizes such objects. \end{itemize} \subsubsection{Introduction Scenario} \label{sec:result-intro-scenario} In this scenario, Arthur begins with no knowledge about the person it is seeing on the webcam, thus, its first action should be to ask who it is. Figure~\ref{fig:meet-new-person} shows how Arthur deals with this situation. In Figure~\ref{fig:meet-new-person} (a), Arthur is meeting a new person and, therefore, asks for this person's name with the statement "Hello stranger! May I know your name?". After the proper introduction (i.e. the person in question answers that its name is "Knob"), Arthur stores the name of this person on its memory. When they meet again (Figure~\ref{fig:meet-new-person} (b)), Arthur is able to remember both the face and the name of this person. Then, it proceeds to a cordial greeting ("Greetings Knob!"). Plus, Arthur tries to maintain the conversation asking questions about the person. In Figure~\ref{fig:meet-new-person} (b), it is possible to see Arthur asking the age of the person ("How old are you?"). When the answer is given, such information is also stored into the memory, increasing the information Arthur knows about this user. \begin{figure}[!htb] \centering \subfigure[Arthur does not know this user.]{\includegraphics[width=0.49\linewidth]{images/stranger2.png}}\hfill \subfigure[After introduction, Arthur is able to remember this person.]{\includegraphics[width=0.49\linewidth]{images/greeting2.png}} \caption{Arthur is meeting a new person. In (a), he asks for this user's name. After a proper introduction, Arthur stores that person's name in his memory and then he can remember that person's face and name, greeting him/her when they see each other again (b).} \label{fig:meet-new-person} \end{figure} \subsubsection{Learning Scenario} \label{sec:result-learning-scenario} As commented in Section~\ref{sec:agent-memory}, Arthur is able to learn information from the person he is talking to. Figure~\ref{fig:learning-things} shows Arthur learning 2 types of information. As commented in Section~\ref{sec:result-intro-scenario}, Arthur tries to starts or maintain a conversation with the person he is interacting, e.g., asking questions about him/her. Figure~\ref{fig:learning-things} (a) shows Arthur asking if the person studies. All possible questions are presented in Section~\ref{sec:chat-voice}. In Figure~\ref{fig:learning-things} (b), Arthur answers the question about the age of Knob, saying he is 31 years old. It is able to do so because the person (i.e. Knob) already informed Arthur about his age in a past interaction. Besides learning information about the person Arthur is talking to, he can also learn about new things. For example, in Figure~\ref{fig:learning-things} (c), it is asked if Arthur knows what a cellphone is. "No, it does not!", is the Arthur answer. Then, he asks the person if he/she desires to show him a picture of a cellphone. If the person agrees, a new term is stored in Arthur's memory, next to the image provided. Then, when Arthur is asked again about a cellphone (Figure~\ref{fig:learning-things} (d)), he is able to remember what it is, showing the image which was provided before. \begin{figure}[!htb] \centering \subfigure[Arthur asking a question to the person.]{\includegraphics[width=0.49\linewidth]{images/question2.png}} \subfigure[After being asked about the age of Knob, Arthur gives the answer.]{\includegraphics[width=0.49\linewidth]{images/old2.png}} \subfigure[After being asked about if it knows a cellphone, Arthur says it does not.]{\includegraphics[width=0.49\linewidth]{images/dunno2.png}} \subfigure[After it learns what a cellphone is, Arthur is able to identify it.]{\includegraphics[width=0.49\linewidth]{images/cellphone2.png}} \caption{Arthur learning new things. In (a) and (b), it is shown the process where Arthur learns new information about a given person. In (c) and (d), Arthur learns what a cellphone is.} \label{fig:learning-things} \end{figure} \subsection{Study with Subjects} \label{sec:results-user-study} In order to test our hypothesis with a qualitative evaluation, we conducted an experiment with subjects. Each person was presented to three video sequences. In the first video sequence, Arthur meets a new person and asks questions about he/she, as explained in Section~\ref{sec:chat-voice}, but the Memory Module is deactivated, thus, it does not retain information. In the second video sequence, the same interaction is conducted, but the Memory Module is activated, thus, Arthur can retain information and remember it further. Finally, in the third video sequence is specifically presented the process where Arthur learns a new information and remembers it further. Arthur is asked if he knows an object (i.e. cellphone), which he does not. Then, we show to him what a cellphone is and ask him again if he knows a cellphone. Since the Memory Module is activated, Arthur remembers what a cellphone is and answer the person. After each video sequence, each subject was asked to answer some questions: \begin{itemize} \item Q1: How do you evaluate your comfort level regarding agent's appearance? \item Q2: How do you evaluate the agent's facial expressions realism level? \item Q3: How do you evaluate agent's information comprehension abilities? \item Q4: After watching the interaction, how do you evaluate agent's memory? \end{itemize} Subjects answered each question following a Likert scale, where: $1$: Very low, $2$: Low, $3$: Moderate, $4$: Good and $5$: Very good. Regarding our expectations, we formulate the following hypothesis: \begin{itemize} \item H1: Since the facial expressions are the same for all three videos, we expect little or no difference of how people will evaluate both their comfort level and the realism level of the agent's facial expressions (Q1 and Q2), throughout the video sequences. \item H2: We believe that the memory of Arthur is going to help him to understand the context of the interaction better. Therefore, in the videos where the Memory Module is activated, we believe the participants are going to rate Arthur's comprehension abilities better (Q3). \item H3: We expect that people rate Arthur's memory higher in Videos 2 and 3, when compared with Video 1 (Q4). \end{itemize} Concerning the subjects, we had 51 answers, where 39 were male and 12 were female. Also, we had participants ranging from 17 to 61 years old. The average age was 31.49 with a standard deviation of 12.49. All participants agreed with the terms of the research and conducted the experiment until the end. \begin{table}[htb] \centering \caption{Quantity of answers acquired by the experiment, separated by each video sequence. The head numbers (i.e. 1, 2, 3, 4 and 5) represent each value of the Likert scale.} \label{tbl:answers-experiment} \begin{tabular}{|c||c||c||c||c||c|} \hline \textbf{Video 1} & \textbf{1} & \textbf{2} & \textbf{3} & \textbf{4} & \textbf{5} \\ \hline \textbf{Q1} & 4 & 7 & \textbf{19} & \textbf{19} & 2 \\ \textbf{Q2} & 8 & 12 & \textbf{19} & 10 & 2 \\ \textbf{Q3} & 3 & 14 & \textbf{15} & \textbf{15} & 4 \\ \textbf{Q4} & 7 & 8 & \textbf{17} & 15 & 4 \\ \hline \end{tabular} \begin{tabular}{|c||c||c||c||c||c|} \hline \textbf{Video 2} & \textbf{1} & \textbf{2} & \textbf{3} & \textbf{4} & \textbf{5} \\ \hline \textbf{Q1} & 6 & 9 & 15 & \textbf{19} & 2 \\ \textbf{Q2} & 5 & 12 & \textbf{20} & 12 & 2 \\ \textbf{Q3} & 4 & 4 & 14 & \textbf{23} & 6 \\ \textbf{Q4} & 5 & 4 & 11 & \textbf{22} & 9 \\ \hline \end{tabular} \begin{tabular}{|c||c||c||c||c||c|} \hline \textbf{Video 3} & \textbf{1} & \textbf{2} & \textbf{3} & \textbf{4} & \textbf{5} \\ \hline \textbf{Q1} & 4 & 7 & 16 & \textbf{19} & 5 \\ \textbf{Q2} & 5 & 12 & \textbf{19} & 10 & 5 \\ \textbf{Q3} & 2 & 6 & 12 & \textbf{19} & 12 \\ \textbf{Q4} & 1 & 5 & 13 & \textbf{19} & 13 \\ \hline \end{tabular} \end{table} Table~\ref{tbl:answers-experiment} shows the quantity of answers obtained per question, for each video sequence. In order to investigate the hypothesis variation of response of the users, we rely on ANOVA test. The results reveal an uniformity in the participants answers concerning Q1 (F(3.05)=0.63,p=.53)) and Q2 (F(3.05)=0.64,p=.52)), thus, there is little variation on the answers about the appearance and realism level of the agent (i.e. Q1 and Q2), which confirms H1. For Q1, most answers were between Moderate (3) and Good (4). For Q2, the most answers were Moderate (3). With this, we can also conclude that the appearance of Arthur has some space to improve, specially concerning its facial expressions. An unexpected behavior was observed on the answers about the comprehension abilities and memory of the agent (i.e. Q3 and Q4). Both questions had a significant amount of answers between Moderate (3) and Good (4) for Video 1. For example, if we take Q4, it was expected to have most answers lying between Very Low (1) and Low (2), since the agent has the Memory Module deactivated. However, we had 17 answers for Moderate (3) and 15 for Good (4), while having a total of 15 for the ones we were expecting. One possible explanation could be the order of the sentences that Arthur speaks. Since they are presented to the user in an orderly way, it may cause the impression that Arthur is indeed understanding and processing the information given by the user. Further investigation would be necessary to confirm or disprove it. Even with such unexpected behavior, it seems that H2 and H3 were also confirmed. In Video 1, we had a total of 19 answers for Good (4) and Very Good (5), for both Q3 and Q4. In Video 2, with the Memory Module activated, we had a total of 29 answers, in the same range, for Q3 and 31 answers for Q4. In Video 3, we had 31 for Q3 and 32 for Q4. Also, if we look at the amount of answers Very Low (1) and Low (2), we can clearly perceive that it diminishes in Videos 2 and 3, when compared with Video 1. Finally, the results of the ANOVA test show a change of perception when comparing Video 1 and Video 3, for both Q3 (F(3.93)=7.57,p=.007)) and Q4 (F(3.93)=11.31,p=.001)). \section{Final Considerations} \label{sec:conclusion} This work presented a model of an Embodied Conversational Agent (ECA) endowed with a memory module and many other abilities, like face recognition, emotion detection and expressiveness. We conducted some experiments in order to test our model and collect both quantitative and qualitative information. The results achieved seem to confirm that Arthur presented the expected behavior. Concerning the appearance of the virtual agent, despite having a good result in the user study, we conclude that it has room to improve. A few changes could be made regarding the geometric model and the animations, pursuing a better comfort level and higher realism. Also, concerning the memory model, a satisfactory result was achieved. In both quantitative and qualitative tests, the results achieved showed that the memory worked as intended and its influence on the interaction could be perceived. As for future work, there are a few paths that can be followed. As have already been commented, we plan to improve the appearance of the agent. Also, according the literature, as presented in Sections~\ref{sec:human-memory} and~\ref{sec:agent-memory}, the human memory is able to store information in different types, like text, images and voice. This work presents a memory model which deals with text and images, thus, the next logical step would be to make it able to deal with audio also. Moreover, besides being able to recognize the person it is talking with, the virtual agent could be able to recognize objects present in the image, both to learn what such object is and identify it later. \section*{Acknowledgments} The authors would like to thank to Brazilian agencies CNPq e CAPES for partially funding this project. \bibliographystyle{IEEEtran}

cb6a9f1dda0587dde048f87d1d833cd24344bdad

\section{Introduction} Massive multiple-input multiple-output (MIMO) has been demonstrated, in both theory and practice, as one of the major performance boosters for the next generation (5G and beyond) wireless communication systems \cite{larsson2014massive,bjornson2017massive}. Operating massive MIMO in time division duplex (TDD) mode is technically favorable, because the inherent uplink-downlink channel reciprocity makes it convenient to reconstruct the downlink channel vectors directly from the uplink pilot observations without requiring downlink training. Nevertheless, from the mobile network operators’ standpoint, the frequency division duplex (FDD) mode seems more preferable, as the current wireless systems are mainly operating in FDD mode, for which a lot of resource has been invested (e.g., in acquiring the spectrum), and FDD systems show a much better performance in scenarios with symmetric data traffic and delay-sensitive applications. For FDD massive MIMO, however, the uplink-downlink channel reciprocity does not hold in general, due to the large uplink/downlink frequency separation that exceeds the fading coherence bandwidth. The base stations (BSs) have to probe the downlink channels via pilot training and ask for channel information feedback from the users. The high-dimensional channel vectors (due to the large number of antennas) incur prohibitively expensive feedback overhead and therefore result in inevitable performance degradation provided the limited channel coherence time and bandwidth. Recently, a fast-growing number of techniques have been proposed to make FDD as competitive as TDD systems by reducing downlink training and uplink feedback overhead, e.g., joint spatial division and multiplexing (JSDM) \cite{JSDM, JSDM-UG}. It has been observed that the high-dimensional channel vectors admit a sparse representation in the angular/beam domain, such that they could be efficiently represented by low-dimensional ones \cite{JSDM, JSDM-UG,rao2014distributed, gao2015spatially}. As such, the pilot dimension of downlink training and the feedback overhead can be substantially reduced. By exploiting the sparse representation in the beam domain, a number of techniques have been proposed. First, the compressed sensing (CS)-inspired methods (e.g., \cite{rao2014distributed, gao2015spatially, ding2018dictionary}) exploit this sparse representation to reconstruct the channel vectors at the BS using the compressed downlink pilot signals fed back from users. In particular, it allows each user to obtain the compressive measurements of the probing signals locally, and feed back to the BS, so that the BS can jointly recover the channel vectors using CS techniques \cite{rao2014distributed}. {Although effective to some extent, these CS-based techniques rely highly on the knowledge of channel vector sparsity order (i.e., the number of significant elements), and probably fail to reconstruct downlink channels reliably when the devoted pilot dimension is less than the sparsity order. } Second, channel reconstruction by exploiting the second-order statistics has attracted more and more attention \cite{xie2018channel, miretti2018fddmassive, haghighatshoar2018multi}. Instead of feeding back compressed measurements of pilot signals, these approaches leverage the angular domain reciprocity to reconstruct downlink channel covariance matrices from the uplink training. Third, there is a new trend of using deep learning to predict downlink channel from the observations of uplink training \cite{wen2018deep, wang2018deep, jang2019deep, arnold2019enabling}. The basic idea is to build up a mapping from uplink channel vectors to downlink ones by using an over-parameterized deep neural network. In principle, the deep neural networks with a sufficiently large number of parameters are able to approximate any complicated functions, as long as the training dataset is large enough. Nevertheless, there are still many challenges to design an efficient deep neural network for channel reconstruction. More recently, channel reconstruction methodologies using second-order channel statistics have been advanced by e.g, \cite{B.K, khalilsarai2020dual, khalilsarai2020structured, liu2020statistical}, which aim to reconstruct downlink channel for FDD massive MIMO by exploiting the angular scattering function (ASF) reciprocity. This technique relies on the key assumption of the reciprocity of the ASF -- it assumes that the ASF is frequency-invariant over both uplink and downlink frequency bands \cite{haghighatshoar2018multi}. It consists of two major components: (1) Acquiring downlink channel support information in the angular domain from uplink channel training by exploiting uplink/downlink angular domain reciprocity; and (2) exploiting structural properties of such support information to design efficient downlink probing and uplink feedback schemes. As the channel angular support (i.e., non-zero elements) information is contained in the covariance matrix, its acquisition can be done by estimating downlink covariance matrices from the uplink ones, followed by the angular supports extraction. The channel support information of all users establishes a beam-user association (that can be modeled by a bipartite graph), in which the support of a user’s channel vector indicates the corresponding beams that can be utilized to serve this user. Such a beam-user association can be exploited for intelligent beam-user assignment that leads to artificially sparsified users’ channels. The active channel sparsification (ACS) will finally help reduce the pilot dimension for downlink probing, while allowing for simultaneous multiple-access of a large number of users using spatial multiplexing. However, the ACS methodology is still facing some challenges in the potential deployment in the practical massive MIMO systems. On one hand, dual-polarized uniform planar array (DP-UPA) is commonly used in the practical systems, although attempts have been made to extend from ULA to DP-ULA \cite{khalilsarai2020dual}. On the other hand, although the current ACS implementation using the mixed integer linear program (MILP) formulation is elegant in theory, its computational complexity scales as the number of antennas and users. To address these issues, in this paper, we consider to extend the ACS formulation from ULA/DP-ULA to DP-UPA by leveraging a matrix-weight bipartite graph representation for users' channels. By relaxing the original MILP formulation, we come up with a new nonlinear integer program (NIP) formulation. Hence, we propose a greedy algorithm to solve the NIP problem in an efficient way. Specifically, our contributions are summarized as follows. \begin{itemize} \item The channel covariance matrix of massive MIMO with DP-UPA antennas can be recognized as a doubly block Toeplitz matrix. By leveraging Toeplitz matrix theory, we characterize the spectral properties of channel covariance matrices by investigating their matrix-valued spectral density function, which is also referred to as angular scattering function \cite{B.T}. There exhibits some sparsity in the spectral density function when the angular spread is narrow under the context of DP-UPA massive MIMO scenarios. \item Inspired by these properties, we extend channel representation of ACS using bipartite graph from the original scale-weight to the matrix-weight counterpart. The matrix-weight bipartite graph establishes the association between block beams (correspond to dual-polarized antenna) and users according to the asymptotic block diagonalization of the channel covariance matrices. Building upon the matrix-weight bipartite graph representation, we propose a multi-dimensional ACS (MD-ACS) method, which is a generalized version of original ACS formulation and is more suitable for DP-UPA antenna configurations. The MD-ACS can be formulated as a generalized multi-assignment problem, which includes the original ACS formulation (i.e., assignment problem) as a special case. \item By taking into account the sum rate maximization and multiuser interference control, we reformulate the MD-ACS approach as a nonlinear integer program, for which we propose a simple yet efficient greedy algorithm to solve it. The extensive simulation results using QuaDRiGa channel models demonstrate the superiority of the proposed MD-ACS with greedy algorithm to the state-of-the-art methods, including the recently advanced ACS method concerning DP-ULA antenna configurations. \end{itemize} The rest of this paper is organized as follows. In the next section, we describe the channel and system model of the DP-UPA FDD massive MIMO system with downlink training and precoding. In Section III, we study channel covariance matrices through Toeplitz theory, and characterize the spectral properties of the spectral density functions. The proposed MD-ACS is detailed in Section IV, including the review of the original ACS, the matrix-weight graph representation, and the NIP formulation with a greedy algorithm. The numerical results can be found in Section V, followed by the Conclusion in Section VI. {\bf Notation:} We use $x$, $\vect{x}$, and $\mat{x}$ to represent scalar, vector, and matrix, respectively. For any scalar $x$, we denote $\{x_n\}_{n=1}^N \triangleq \{x_1,x_2,\dots, x_N\}$. For the integer $N$, we denote $[N]\triangleq \{1,2,\dots,N\}$. A matrix $\mat{x}$ is Hermitian if and only if $\mat{x}=\mat{x}^{\H}$, where $\mat{x}^\H$ is the conjugate transpose of $\mat{x}$. ${\hbox{tr}}{(\mat{x})}$ denotes the trace of a matrix $\mat{x}$. $\mbox{\bb E}\{\cdot\}$ denotes the expectation. The Kronecker and Hadamard products of two matrices $\mat{x}$ and $\mat{y}$ are denoted by $\mat{x} \otimes \mat{y}$ and $\mat{x} \odot \mat{y}$, respectively. {$\CN(\alpha,\beta)$ denotes the complex normal distribution, where $\alpha$ and $\beta$ are mean (vector) and variance (matrix), respectively. } $\mat{i}_M$ is the $M\times M$ identity matrix, and $\mat{f}_M$ is the discrete Fourier transform (DFT) matrices with $[\mat{f}_{M}]_{p,q}=\frac{1}{\sqrt{M}}e^{-\jmath\frac{2\pi (p-1)(q-1)}{M}}$ for all $p\in [M],q\in[M]$. \section{Channel and Signal Model} \subsection{DP-UPA Channel Model} We consider a single-cell massive MIMO system where the base station is equipped with an $M_x\times M_y\times 2$ dual-polarized uniform planar array (DP-UPA) serving $N_U$ single-polarized single-antenna users. {The DP-UPA consists of in total $M=2 M_x M_y$ antenna elements with $M_x$ ports in each column and $M_y$ ports in each row, and for each port there are two polarized antenna elements.} According to 3GPP {TR-36.873} \cite{3GPP}, which is also referred by e.g., \cite{C.Q} and \cite{L.M}, the channel vector $\vect{h}$ of DP-UPA can be represented as \begin{align}\label{eq:channelV_channelH} \vect{h}=\begin{bmatrix}\vect{h}_V\\ \vect{h}_H \end{bmatrix} \in \mbox{\bb C}^{M \times 1} \end{align} where $\vect{h}_V \in \mbox{\bb C}^{\frac{M}{2} \times 1}$ and $\vect{h}_H \in \mbox{\bb C}^{\frac{M}{2} \times 1}$ correspond to the channel between the vertical ($V$)/horizontal ($H$) antenna and the user, respectively. For notational simplicity, let $q\in\{V,H\}$. Given the angle intervals of azimuth ${\mathcal A}$ and elevation ${\mathcal B}$, according to the channel model of 3GPP \cite{3GPP}, the $q$-th sub-channel vector can be written as \begin{align} \vect{h}_q=\int_{\mathcal B}\int_{\mathcal A} \beta_q(\theta, \phi)\gamma_q\vect{a}(\theta,\phi)d\theta d\phi \end{align} where ${\mathcal A}=[\theta_{\min},\theta_{\max}],{\mathcal B}=[\phi_{\min},\phi_{\max}]$ and $|{\mathcal A}|=2\delta_\theta$ and $|{\mathcal B}|=2\delta_\phi$, in which $\delta_\theta$ and $\delta_\phi$ are the angular spread (AS) of azimuth and elevation, respectively; $\beta_q(\theta,\phi)\sim\CN(0,\beta_q)$ denotes the complex gain that is independent and identically distributed (i.i.d.) across paths; $\gamma_q$ is the polarization factor of the $q$-th sub-channel; and $\vect{a}(\theta,\phi)$ is the steering vector of DP-UPA antenna that possesses the same structure as that of UPA, and it can be written as \cite{3GPP}\cite{C.Q}\cite{lu2020omnidirectional} \begin{align} \label{eq:steering-vector} &\vect{a}(\theta,\phi)=\vect{a}_y(\theta,\phi)\otimes \vect{a}_x(\theta,\phi) = \begin{bmatrix} 1\\e^{\jmath\frac{2\pi d_y}{\lambda_w}\sin(\phi)\sin(\theta)}\\ \vdots \\e^{\jmath\frac{2\pi d_y(M_y-1)}{\lambda_w}\sin(\phi)\sin(\theta)} \end{bmatrix}\otimes \begin{bmatrix} 1\\e^{\jmath\frac{2\pi d_x }{\lambda_w}\sin(\phi)\cos(\theta)}\\ \vdots \\e^{\jmath\frac{2\pi d_x (M_x-1)}{\lambda_w}\sin(\phi)\cos(\theta)} \end{bmatrix} \end{align} where $d_x$ and $d_y$ are antenna spacing of column and row array respectively, and $\lambda_w$ is the carrier wavelength. \subsection{Downlink Training and Precoding} \label{sec:signal-model} In this paper, we follow the comprehensive framework proposed in \cite[Figure 4]{khalilsarai2020dual}, which consists of (1) uplink pilot transmission, (2) uplink covariance estimation, (3) uplink-downlink covariance transformation, (4) downlink pilot transmission, (5) feeding back pilot measurements, (6) downlink channel estimation, and (7) downlink beamforming. As our focus in this paper is on the downlink precoding/beamforming, we assume the availability of downlink covariance matrix at the base station via the above steps (1)-(3). In what follows, we briefly reiterate the procedure of (4)-(7) to maintain certain level of self-containedness. \subsubsection{Downlink Pilot Transmission} As in \cite{khalilsarai2020dual}, the base station sends a space-time pilot matrix $\mat{s}\in\mbox{\bb C}^{T\times M'}$ to all users through a sparsifying precoder $\mat{V}_h \in \mbox{\bb C}^{M \times M'}$, where $T$ is the number of time slots used for pilot trainsmission, {$M' \le M$ is the dimension after the active sparsification, and the columns of $\mat{V}_h$ are chosen from an orthogonal matrix that will be specified later.} As such, the received pilot signal $\vect{y}_i^{\mathrm{p}}$ of $i$-th user can be written as \begin{align}\label{eq:estimated channel} &\vect{y}^{\mathrm{p}}_i = \mat{s} \mat{V}_h^\H \vect{h}_i +\vect{n}, \end{align} where $\vect{h}_i\in\mbox{\bb C}^{M\times 1}$ is the downlink channel vector of $i$-th user, and $\vect{n} \sim \CN(\vect{0}, \sigma^2 \mat{i}_M)$ is the additive white Gaussian noise (AWGN). The pilot matrix $\mat{s}$ is up to design, subject to a total power constraint ${\hbox{tr}}(\mat{s}\mat{V}_h^\H\mat{V}_h\mat{s}^\H)\leq \rho^{\mathrm{p}}T$, where {$\rho^{\mathrm{p}}$ is the pilot signal power in each time slot.} \subsubsection{Feeding Back Pilot Measurements} For simplicity, we assume the users feed back their pilot signals $\vect{y}^{\mathrm{p}}_i \in \mbox{\bb C}^{T \times 1}$ to the base station in an analog form. The digital feedback with quantization can be implemented according to well-developed techniques (see \cite{love2003equal} and references therein). Due to possible user selection, only the selected users are required to send the pilot signals back to the base station. In doing so, the base station could successfully acquires the perfect pilot signals $\{\vect{y}^{\mathrm{p}}_i \}_{i \in {\mathcal S}}$ with {${\mathcal S}$ being the subset of selected users, which will be specified later.} \subsubsection{Downlink Channel Estimation} Given the $T \times 1$ pilot signal $\vect{y}^{\mathrm{p}}_i$, we aim to recover the $M \times 1$ channel vector $\vect{h}_i$ with $M > T$, relying on the sparsity of $\vect{h}_i$ in the angular domain. Following the footstep in \cite{khalilsarai2020dual}, we obtain the estimated channel vector via MMSE estmators as \begin{align}\label{est-channel} \hat{\vect{h}}_i = \mat{r}_{h,i}\mat{r}_{y,i}^{-1}\vect{y}_i^{\mathrm{p}}, \end{align} where $\mat{r}_{h,i} = \mbox{\bb E}\{\vect{h}_i(\vect{y}_i^\mathrm{p})^\H\}= \mat{r}_i \mat{V}_h\mat{s}^\H$, $\mat{r}_{y,i} = \mbox{\bb E}\{\vect{y}_i^\mathrm{p}(\vect{y}_i^\mathrm{p})^\H\}=\mat{s}\mat{V}_h^\H\mat{r}_i\mat{V}_h\mat{s}^\H + \sigma^2 \mat{i}$ with $\mat{r}_i \triangleq \mbox{\bb E} [\vect{h}_i \vect{h}_i^{\H}]$ being the downlink channel covariance matrix of user-$i$. \subsubsection{Downlink Precoding} With channel estimates, the base station transmit users' data $\{d_i\}_{i \in {\mathcal S}}$ through sparsifying precoders $\vect{p}_i \in \mbox{\bb C}^{M \times 1}$ for each selected user $i \in {\mathcal S}$. Thus, the received signal under FDD DP-UPA downlink data phase $y_i^\d$ of $i$-th user can be written as \begin{align} y_i^{\d}=\vect{h}_i^\H\vect{p}_id_i+\sum_{j \in {\mathcal S} \backslash i}\vect{h}_i^\H\vect{p}_jd_j+n_i \end{align} where $n_i\sim\CN(0,\sigma_i^2)$ is the AWGN, and the sparsifying precoder $\vect{p}_i$ will be specified later. As the downlink covariance matrix estimation has been extensively investigated in the literature (e.g, \cite{L.M,khalilsarai2020dual,miretti2018fddmassive}), we place our focus instead on designing the downlink precoder assuming that the downlink channel covariance matrix $\{\mat{r}_i\}_{i=1}^{N_U}$ is perfectly known at the base station. \section{Spectral Properties of Covariance Matrix}\label{sec-3} While some existing works have mentioned the Toeplitz structure of covariance matrices for ULA/UPA massive MIMO (e.g., \cite{JSDM, Yu:ACS}), the extension to DP-UPA has not been fully understood. In what follows, we will inspect the structural properties of downlink channel covariance matrices $\{\mat{r}_i\}_{i=1}^{N_U}$ for DP-UPA massive MIMO through the lens of Toeplitz matrix theory. \subsection{Toeplitz Matrix Theory} Before proceeding further, we first introduce the definitions related to Toeplitz matrix \cite{gray2006toeplitz} and its extension to block Toeplitz matrix \cite{B.T} and doubly Toeplitz matrix \cite{pa1996atheorem,oudin2008asymptotic}. Given a sequence of scalars $\{t_{-n+1}, \dots, t_{-1}, t_0, t_1 \dots, t_{n-1}\}$, an $n\times n$ matrix $\mat{t}_n$ is a Toeplitz matrix if $[\mat{t}_n]_{i,j}=t_{i-j}$ for all $i,j\in[n]$. Similarly, given a sequence of $M_1 \times M_2$ matrices $\{\mat{t}_{-n+1}, \dots, \mat{t}_{-1}, \mat{t}_0, \mat{t}_1 \dots, \mat{t}_{n-1}\}$ , an $nM_1\times nM_2$ matrix $\mat{b}_n$ is a block Toeplitz matrix if the $(i,j)$-th $M_1 \times M_2$ submatrix $[\mat{b}_n]_{i,j}=\mat{t}_{i-j}$ for all $i,j\in[n]$. In particular, if $\mat{t}_{m}$ is an $N \times N$ Toeplitz matrix for all $-n+1 \le m \le n-1$, then $\mat{b}_n$ is an $nN \times nN$ doubly Toeplitz matrix, also known as Toeplitz-block-Toeplitz (TBT) matrix (i.e., block Toeplitz matrix with Toeplitz blocks). Further, if $\mat{t}_{m}$ is an $NM_1 \times NM_2$ block Toeplitz matrix for all $-n+1 \le m \le n-1$, then $\mat{b}_n$ becomes an $nNM_1 \times nNM_2$ doubly block Toeplitz matrix. It can be viewed as a Toeplitz matrix with each element being block Toeplitz matrices, or a doubly Toeplitz matrix with each element being a general matrix. Throughout this paper, we consider Hermitian matrix, that is $t_{-i}=t_{i}^*$ for Toeplitz matrix and $\mat{t}_{-i} = \mat{t}_{i}^\H$ for block Toeplitz matrix. The circulant, block circulant, doubly circulant, doubly block circulant matrices can be similarly defined as their Teoplitz counterparts, where the only difference is the circular operation using {the modulo operator}$\!\mod\!\!$, i.e., for the scalar sequence $[\mat{c}_n]_{i,j}=c_{(i-j) \!\mod n}$ for all $i,j\in[n]$ and for the matrix sequence $[\mat{b}_n]_{i,j}=\mat{c}_{(i-j) \!\mod n}$ for all $i,j\in[n]$. Apparently, the (doubly block) circulant matrix is the special case of (doubly block) Toeplitz matrix. Given an $n \times n$ circulant matrix $\mat{c}_n$, it can be diagonalized by DFT matrix, i.e., $\mat{c}_n = \mat{f}_n \hbox{\boldmath$\Lambda$} \mat{f}_n^\H$ with $\hbox{\boldmath$\Lambda$}$ being a diagonal matrix. For an $nN \times nN$ doubly circulant matrix, it can be diagonalized by 2D-DFT matrix $\mat{f}_n \otimes \mat{f}_N$. For an $nM_1 \times nM_2$ block circulant matrix $\mat{b}_n$, it can be block-diagonalized by $\mat{b}_n = (\mat{f}_n \otimes \mat{i}_{M_1}) \hbox{\boldmath$\Sigma$} (\mat{f}_n \otimes \mat{i}_{M_2})^\H$ where $\hbox{\boldmath$\Sigma$}$ is an $nM_1 \times nM_2$ block diagonal matrix, with each block being an $M_1 \times M_2$ matrix. When $n$ tends to infinity, each Toeplitz matrix can be associated with a generating function, which is a continuous and periodic function \cite{gray2006toeplitz,B.T,pa1996atheorem,oudin2008asymptotic}. For instance, the Hermitian Toeplitz matrix $\mat{t}_n$ can be generated by a real function $F: [-1/2, 1/2] \mapsto \mbox{\bb R}$, i.e., $F(\omega)=\sum_{k=-\infty}^{\infty}t_ke^{\jmath 2\pi k \omega}$. Similarly, the Hermitian block Toeplitz matrix $\mat{b}_n$ can be generated by a matrix-valued real function $F: [-1/2,1/2] \mapsto \mbox{\bb R}^{M_1 \times M_2}$, i.e., $F(\omega)=\sum_{k_1=-\infty}^{\infty}\mat{t}_{k_1}e^{\jmath 2 \pi k_1\omega}$. Further, for the Hermitian doubly block Toeplitz matrix, it can be generated by $F: [-1/2,1/2]^2 \mapsto \mbox{\bb R}^{M_1 \times M_2}$, i.e., \begin{align}\label{eq:TBT_G} &F(\omega_1,\omega_2)=\sum_{k_1=-\infty}^{\infty}\sum_{k_2=-\infty}^{\infty}\mat{t}_{k_1,k_2}e^{\jmath 2 \pi k_1\omega_1}e^{\jmath 2 \pi k_2\omega_2}. \end{align} The circulant conterparts share the same generating functions. \subsection{Spectral Properties} By leveraging the Toeplitz matrix theory, we inspect the spectral properties of channel covariance matrix through a function analysis perspective. In particular, instead of looking into the channel covariance matrix, we investigate its spectral density in the angular domain. This is underpinned by the following lemma. \begin{lemma} \label{lemma:structure_cov} The channel covariance matrix $\mat{r}$ of DP-UPA massive MIMO can be represented, subject to row/column permutation, as a Hermitian doubly block Toeplitz matrix $\hat{\mat{r}}$, which can be asymptotically block diagonalized by an orthongal matrix \begin{align} \mat{V} = \mat{f}_{M_x} \otimes \mat{f}_{M_y} \otimes \mat{i}_2, \end{align} as $M_x,M_y\to \infty$, where $\mat{f}_{n}$ is an $n \times n$ DFT matrix, and the block-diagonal submatrices are uniformly sampled from the matrix-valued spectral density function, i.e., \begin{align} &\hbox{\boldmath$\Sigma$}(\omega_1,\omega_2)=\sum_{m_1=-M_y+1}^{M_y-1}\sum_{m_2=-M_x+1}^{M_x-1}[\hat{\mat{r}}]_{m_1,m_2}e^{ \jmath 2 \pi (m_1\omega_1+m_2\omega_2)} \end{align} with $[\hat{\mat{r}}]_{m_1,m_2}$ being a $2 \times 2$ submatrix of $\hat{\mat{r}}$. \end{lemma} \begin{proof} See Appendix \ref{proof:lemma-toeplitz}. \end{proof} \begin{remark} The $2 \times 2$ matrix-valued spectral density function $\hbox{\boldmath$\Sigma$}(\omega_1,\omega_2)$ over $(\omega_1,\omega_2) \in [-1/2,1/2]^2$ is the generating function of the doubly block Toeplitz matrix $\hat{\mat{r}}$. As row/column permutation does not change spectral properties, $\hbox{\boldmath$\Sigma$}(\omega_1,\omega_2)$ is the spectral density function of channel covariance matrix $\mat{r}$ over the two-dimensional angular domain $ [-1/2,1/2]^2$. Similar to the ULA massive MIMO, we can also transform the signals from spatial to angular domain to exploit possible (block) sparsity of the spectral density. The columns of the DFT-type orthogonal matrices have been widely used as the common basis for Toeplitz, block Toeplitz and TBT matrix in massive MIMO such as precoder design for ULA\cite{X.R,khalilsarai2020dual} and UPA\cite{Yu:ACS} array, and pilot decontamination\cite{Haifan,you2015pilot,Z.C}. \end{remark} Equipped with Lemma \ref{lemma:structure_cov}, we are able to inspect the spectra of channel covariance matrix $\mat{r}$ through its spectral density function $\hbox{\boldmath$\Sigma$}(\omega_1,\omega_2)$. As such, we come up with the sparsity properties of DP-UPA antennas in the angular domain, as in Theorem \ref{theorem:sparsity}. \begin{theorem} \label{theorem:sparsity} The spectral density function $\hbox{\boldmath$\Sigma$}(\omega_1,\omega_2)$ has a compact support over the two-dimensional frequencies $(\omega_1,\omega_2) \in [-1/2,1/2]^2$, i.e., \begin{align} \hbox{\boldmath$\Sigma$}(\omega_1,\omega_2) = \mathbf{0}, \quad \mathrm{ if~} (\omega_1, \omega_2) \notin \left[-\frac{d}{\lambda_w}z_1^{\max},\frac{d}{\lambda_w}z_1^{\min}\right] \times \left[-\frac{d}{\lambda_w}z_2^{\max},\frac{d}{\lambda_w}z_2^{\min}\right] \end{align} where $z_i^{\min}$ and $z_i^{\max}$ depend on a fixed AOA $\theta_c,\phi_c$ and AS $\Delta_1, \Delta_2$. \end{theorem} \begin{proof} See Appendix \ref{proof:theorem-sparsity}. \end{proof} \begin{remark} Theorem \ref{theorem:sparsity} is a generalization of the compact properties of spectral density function from ULA reported in \cite{JSDM} to DP-UPA antenna configurations. In contrast with the ULA, the spectral density function of DP-UPA is $2\times 2$ matrix-valued because of the dual-polarization. In addition, for UPA and DP-UPA antennas, the compact supports of $\hbox{\boldmath$\Sigma$}(\omega_1,\omega_2)$ could be more dispersed, thanks to the two-dimensional array. This enables UPA-type antennas to server more users without causing severe pilot contamination or multiuser interference. Thanks to the high resolution of large-scale antenna arrays, the azimuth and elevation AoAs are usually limited within a narrow range \cite{jaeckel2014quadriga}, so that $z_i^{\max}$ and $z_i^{\min}$ are confined within small intervals in $[-1,1]$. As such, the compact support only covers a limited range of frequency range, and thus the spectral density exhibits sparsity properties in the angular domain. To illustrate the above points, we plot the spectral density of covariance matrices of ULA, UPA, and DP-UPA with the same number of antennas, using channels generated by QuaDRiGa \cite{jaeckel2014quadriga} (See Section \ref{sec:numerical} for the configurations). In particular, Figure \ref{Com_set_1} shows the normalized diagonal elements of DFT-diagonalized covariance matrices for $128\times 1$ ULA, $16\times 8$ UPA, and $8\times 8\times 2$ DP-UPA, respectively. It can be observed that ULA has one single yet wide support, and UPA and DP-UPA have multiple narrow supports. Additionally, for DP-UPA, it exhibits the block support where the supports appear in pair, which agrees with the $2 \times 2$ matrix-value spectral density function. \end{remark} \begin{figure}[t] \centering \includegraphics[width= 3.5in,angle=0]{Com_set_2} \vspace{-10pt} \caption{The normalized spectra of covariance matrices for $128 $ ULA, $16\times 8$ UPA, and $8\times 8\times2$ DP-UPA antennas.} \label{Com_set_1} \vspace{-20pt} \end{figure} \section{Multi-Dimensional Active Channel Sparsification} \label{sectionIV} Inspired by the matrix-valued spectral density function, we extend the active channel sparsification to multi-dimensional scenarios, with a generalized optimization problem formulation. In what follow, we first summarize the merit of active channel sparsification proposed in \cite{B.K}, followed by a matrix-weight graph representation, and finally we bridge the general optimization problem formulation to an existing problem in combinatorial optimization. \subsection{Preliminary: Active Channel Sparsification} For the sake of self-containedness, we briefly introduce the main idea of active channel sparsification proposed in \cite{B.K}, with the focus on ULA antenna geometry. \subsubsection{Channel Representation} Each user's channel is represented as a weighted sum of a set of vectors chosen from common bases. These common basis vectors are also referred as to virtual beams in the angular/beam domain. Usually, for ULA antenna setting, the columns of discrete Fourier transform (DFT) matrix are adopted as common basis vectors. This is underpinned by the fact that the channel covariance matrix of ULA antenna is a Toeplitz matrix, which asymptotically approximates the circulant matrix that can be diagonalized by DFT matrix. Such a representation ensures that all users are represented in the same vector space, such that beam selection can be alternatively done by switching on/off the basis vectors. If we use $x_m$ to denote the status of the beam-$m$, we have \begin{align} x_m = \left\{ \Pmatrix{1, \quad \text{if beam-$m$ is selected,}\\ 0, \quad \text{otherwise.}} \right. \end{align} In a similar way, we can also impose a binary variable $y_i$ to denote user selection, i.e., \begin{align} y_i = \left\{ \Pmatrix{1, \quad \text{if user-$i$ is selected,}\\ 0, \quad \text{otherwise.}} \right. \end{align} Hence, after adopting the user and beam selection strategy, the estimate of $i$-th user's channel $\vect{h}_i$, as in \eqref{est-channel} in Section \ref{sec:signal-model}, can be asymptotically written as \begin{align} \hat{\vect{h}}_i \approx \sum_{m=1}^M x_m {\iota_{i,m}} \sqrt{w_{i,m}}\vect{f}_m \end{align} if user-$i$ is selected, i.e., $ y_i=1$, where $\vect{f}_m$ is the $m$-th common basis vectors used for channel representation, $w_{i,m}$ is the corresponding coefficient that can be estimated by downlink training, {and $\iota_{i,m}$ is a random variable}. Usually, in the ULA setting, $\vect{f}_m$ comes from the columns of the DFT matrix $\mat{f}_M$ (e.g., \cite{B.K, T.M}), and in the $M_x \times M_y \times 2$ DP-UPA setting, $\vect{f}_m$ is usually from the columns of the common basis $\mat{i}_2\otimes\mat{f}_{M_x}\otimes \mat{f}_{M_y}$ (e.g., \cite{khalilsarai2020dual}). Accordingly, the sparsifying precoder $\mat{V}_h$ in \eqref{eq:estimated channel} can be specified as the collection of basis vectors $\{\vect{f}_m: x_m=1\}$. \subsubsection{Graph Representation} To describe the interaction between beam and user selection, we can construct a weighted bipartite graph, where beams are on one side and users are on the other side, and a beam and a user is connected if the beam contributes to channel representation of such user. By such bipartite graph representation, we establish the user-beam association with respect to the weighted combinations of channel representation. For the readers' reference, we introduce some graph definitions. Consider an undirected bipartite graph ${\mathcal G}=({\mathcal U},{\mathcal V},{\mathcal E})$ with two vertex sets ${\mathcal U}$ and ${\mathcal V}$, and an edge set ${\mathcal E}$. For any $u \in {\mathcal U}$ and $v \in {\mathcal V}$, $e=(u,v) \in {\mathcal E}$ if and only if $u$ and $v$ are connected with an edge $e$. The neighborhood of a vertex $v$ is the set of nodes $u \in {\mathcal U}$ such that $(u,v) \in {\mathcal E}$, i.e., ${\mathcal N}_{{\mathcal G}}(v) \triangleq \{u \in {\mathcal U}: (u,v) \in {\mathcal E} \}$. The degree of a vertex $v$ is the number of nodes in the neighborhood of $v$, i.e., $\text{deg}_{{\mathcal G}}(v) \triangleq \abs{{\mathcal N}_{{\mathcal G}}(v)}$ where $\abs{{\mathcal N}}$ is the cardinality of the set ${\mathcal N}$. The adjacency matrix $\mat{a}$ of the bipartite graph ${\mathcal G}=({\mathcal U},{\mathcal V},{\mathcal E})$ is a binary matrix, where $\mat{a}_{i,j}=1$ if $(i,j) \in {\mathcal E}$ and 0 otherwise. The {Bipartite matching} of the bipartite graph ${\mathcal G}=({\mathcal U},{\mathcal V},{\mathcal E})$ is a subset of edges ${\mathcal M}_{{\mathcal G}} \subset {\mathcal E}$ such that there are not edges in ${\mathcal M}_{{\mathcal G}}$ sharing the same vertex. \subsubsection{Beam/user Selection} The aim of beam/user selection is to switch on/off beams and users to avoid beam overlapping among selected users, in order to achieve the maximum multiplexing gain (i.e., prelog of the sum rate expression). The optimization problem was given in \cite{B.K} as \begin{subequations}\label{*} \begin{align} ({\mathcal P}_1): \quad \max & \quad \left| {\mathcal M}_{{\mathcal G}'}\right|\\ \label{eq:acs-degree_constraint} \mathrm{s.t.} & \quad \text{deg}_{{\mathcal G}'}(u_i) \leq T, \quad \forall u_i\in{\mathcal U}' \\ \label{eq:acs-power_constraint} & \quad \sum_{b_m \in {\mathcal N}_{{\mathcal G}'}(u_i)} w_{i,m} \ge P, \quad \forall u_i\in{\mathcal U}' \end{align} \end{subequations} where $\left| {\mathcal M}_{{\mathcal G}}\right|$ is the maximum cardinality bipartite matching number of the selected subgraph ${\mathcal G}'=({\mathcal B}',{\mathcal U}',{\mathcal E}')$, and the degree constraint \eqref{eq:acs-degree_constraint} guarantees that for each the selected user $u_i \in {\mathcal U}'$ the number of beams to represent this user's channel vector is no more than $T$, and the power constraint \eqref{eq:acs-power_constraint} is to ensure that for each selected user $u_i \in {\mathcal U}'$ the sum power of representing beams is no less than $P$. The degree and power constraints ensure that each selected user should have a sufficient number of (but not too many) representing beams selected, so that those beams with little contribution to a user's channel representation can be switched off. As usually there are much more beams than users, by intuition, the maximum cardinality bipartite matching tends to select all users and only the users have severe conflicting representing beams will be unselected. As such, there is only implicit user selection through beam selection. \subsubsection{Casting as an MILP} By establishing the equivalence between the multiplexing gain of the users' effective channel and the maximum cardinality bipartite matching of the graph representation, the objective of ACS can be solved by finding the solutions to an MILP \cite{B.K} involving two sets of binary variables $\{x_m\}_{m=1}^{M}$ and $\{y_i\}_{i=1}^{N_U}$, and a set of continuous ones, $\{z_{i,m}\}_{i=1,m=1}^{N_U,M}$ i.e., \begin{subequations}\label{eq:ACS-MILP} \begin{align} ({\mathcal P}_1'): \max_{x_m,y_i, z_{i,m}} & \sum_{b_m\in{\mathcal B}}\sum_{u_i\in{\mathcal U}} z_{i,m} \label{eq:MILP-0}\\ \mathrm{s.t.} \quad & z_{i,m} \le [\mat{a}]_{i,m}, \quad \forall b_m\in {\mathcal B},\ u_i \in {\mathcal U} \label{eq:MILP-1}\\ & \sum_{u_i \in {\mathcal U}} z_{i,m} \leq x_m, \quad \forall b_m\in {\mathcal B} \label{eq:MILP-2}\\ & \sum_{b_m \in {\mathcal B}} z_{i,m}\leq y_i, \quad u_i\in{\mathcal U}\label{eq:MILP-3}\\ & \sum_{b_m \in {\mathcal B}} [\mat{a}]_{i,m}z_{i,m} \leq T y_i + M(1-y_i), \quad \forall u_i\in {\mathcal U} \label{eq:MILP-4}\\ & P y_i \le \sum_{b_m \in {\mathcal B}} [\mat{w}]_{i,m} x_{m}, \quad \forall u_i \in {\mathcal U} \label{eq:MILP-5}\\ & x_m \le \sum_{u_i \in {\mathcal U}} [\mat{a}]_{i,m} y_i, \quad \forall b_m \in {\mathcal B} \label{eq:MILP-6}\\ & x_m,y_i\in\{0,1\}\quad \forall u_i\in{\mathcal U},\ b_m\in {\mathcal B} \label{eq:MILP-7}\\ & z_{i,m}\in[0,1]\quad \forall u_i\in {\mathcal U}, \ b_m\in{\mathcal B} \end{align} \end{subequations} where binary matrix $\mat{a}$ is the adjacency matrix of graph ${\mathcal G}$, and $[\mat{w}]_{i,m}=w_{i,m}$ indicates the contribution of the block beam $m$ to the $i$-th user's channel representation. By such an MILP formulation, we can adopt off-the-shelf solvers to find a {feasible} solution $\{x^*_m\}_{m=1}^{M}$, $\{y^*_i\}_{i=1}^{N_U}$ and $\{z^*_{i,m}\}_{i=1,m=1}^{N_U,M}$ efficiently, where the selected beams and users are indicated by $\{m: x^*_m=1\}$ and $\{i: y^*_i=1\}$ respectively in the optimal solution yielded by the MILP. \subsection{Matrix-weight Bipartite Graph Representation}\label{sec-GAP} From Section \ref{sec-3}, the covariance matrix $\hat{\mat{r}}_i$ can be asymptotically block-diagonalized by \begin{align} \label{eq:blk-diag} \lim_{M_x,M_y \to \infty}\hat{\mat{r}}_i &= (\mat{f}_{M_y}\otimes \mat{f}_{M_x}\otimes \mat{i}_2) \hbox{\boldmath$\Sigma$}_i (\mat{f}_{M_y}\otimes \mat{f}_{M_x}\otimes \mat{i}_2)^\H\\ &= \sum_{m_1=1}^{M_y} \sum_{m_2=1}^{M_x} (\vect{f}_{v,m_1} \otimes \vect{f}_{h,m_2} \otimes \mat{i}_2) \hbox{\boldmath$\Sigma$}_i(m_1,m_2) (\vect{f}_{v,m_1} \otimes \vect{f}_{h,m_2} \otimes \mat{i}_2)^\H \end{align} where $\vect{f}_{v,m}$ and $\vect{f}_{h,m}$ are the $m$-th column of DFT matrices $\mat{f}_{M_y}$ and $\mat{f}_{M_x}$, respectively, and $\hbox{\boldmath$\Sigma$}_i(m_1,m_2)$ is the $(M_y(m_1-1)+m_2)$-th diagonal block matrix of $\hbox{\boldmath$\Sigma$}_i$. Instead of using a vector to represent a virtual beam in the ULA and UPA settings, here we use a $M \times 2$ submatrix $\mat{V}_{m_1,m_2} \triangleq \vect{f}_{v,m_1} \otimes \vect{f}_{h,m_2} \otimes \mat{i}_2$ to represent a virtual \emph{cross-polarized} block beam. Similarly, we can represent all users' channels by a bipartite graph with matrix-valued weights, where the cross-polarized block beams $\{\mat{V}_{m_1,m_2}, m_1 \in [M_y], m_2 \in [M_x]\}$ on one side and the users on the other side, and the beams and users are connected with edges of matrix-valued weights $\{\hbox{\boldmath$\Sigma$}_i(m_1,m_2), m_1 \in [M_y], m_2 \in [M_x]\}$. For notational simplicity, we use $[\hbox{\boldmath$\Sigma$}_i]_m$ to denote the matrix-valued weight for $m \in [M/2]$ corresponding to some $(m_1,m_2)$. \begin{figure}[t] \centering \includegraphics[width= 6in,angle=0]{DP-UPA-ONOFF-H} \caption{Matrix-weight bipartite graph for channel representations, where the virtual block beams are denoted by a square with crossed lines (cf. cross-polarized antenna elements), the users are denoted by triangles, and the weights between beams and users $\mat{e}_{i,m}$ are $2 \times 2$ matrices. (a) Channel representations from different users are overlapping in the sense that they share some common block beams (indicated by red edges) to represent their respective channels. (b) After active channel sparsification applied, some block beams (marked in gray) and users (marked in black) are switched off to avoid channel overlapping, so that the remaining users are not overlapping on active block beams.} \label{fig_2} \vspace{-15pt} \end{figure} We refer to the scalar-weight graph representation of previous ACS formulation as single-dimension, and the matrix-weight one as multi-dimension bipartite graph representation. In particular, we represent the users' channel covariance matrices in respect of the block beams in a matrix weighted bipartite graph ${\mathcal G}$ as in Fig. \ref{fig_2}, where a block beam corresponds to a pair of cross-polarized antennas. For notational simplicity, we index the block beams as $m \in [M/2]$. We define the matrix weighted bipartite graph ${\mathcal G}=({\mathcal B},{\mathcal U},{\mathcal E})$, in which the block beams $b\in{\mathcal B}$ is on one side and users $u\in{\mathcal U}$ on the other side. Therefore, a beam $b_m$ and a user $u_i$ are connected with an edge $(b_m,u_i)\in{\mathcal E}$ if $\mat{a}_{i,m}=1$. It is worth noting that, the weight of edges $(b_m,u_i) \in {\mathcal E}$, i.e., $\mat{e}_{i,m}=[\hbox{\boldmath$\Sigma$}_i]_m$ with $m$ corresponding to some $(m_1,m_2)$, is a $2 \times 2$ matrix rather than a scalar. With the block beam and user selection parameters $x_m$ and $y_i$, the estimated channel of $i$-th user, as in \eqref{est-channel} in Section \ref{sec:signal-model}, can be approximately written as \begin{align} \hat{\vect{h}}_i \approx \sum_{m=1}^{M/2}x_m \mat{V}_m \left([\hbox{\boldmath$\Sigma$}_i]_m\right)^{\frac{1}{2}} \hbox{\boldmath$\iota$}_{i,m} \end{align} where $\mat{V}_m \in \mbox{\bb C}^{M \times 2}$ corresponds to the block basis vectors $\mat{V}_{m_1,m_2}$ with $m$ corresponding to some $(m_1,m_2)$, and $\hbox{\boldmath$\iota$}_{i,m} \in \mbox{\bb C}^{2 \times 1}$ is a random vector. Similarly, the sparsifying precoder $\mat{V}_h$ in \eqref{eq:estimated channel} can be specified as the collection of $\{\mat{V}_m: x_m=1\}$. Let us explain the physical meaning of the matrix weighted bipartite graph. Each block beam is illustrated as a pair of crossed lines, in which the cross with red and black lines corresponds to the vertical and horizontal polarization antennas in the DP-UPA array, respectively. As a matter of fact, such a correspondence is resulted from the block-diagonalization of the channel covariance matrix, where it combines the cross-polarized antennas at the same position. The diagonal elements of $\mat{e}_{i,m}\in\mbox{\bb C}^{2\times 2}$ represent the channel characteristics of the corresponding antenna, while the off-diagonal elements indicate the channel correlation between the vertical and horizontal antennas due to their cross-polarization. In Fig. \ref{fig_2}(a), the edges in red indicate the inter-user spectral correlation between users' channels in the angular domain, which results in potential inter-user interference for multi-user transmission. Fig. \ref{fig_2}(b) presents a simple beam and user selection to reduce the possible beam overlapping for activated users. When actively switching off some block beams and users, the partial channel correlation of the remaining users is reduced, for which there is not any overlap on the activated beams anymore. Note here that the original channels covariance matrices are partially represented by the active block beams only. While this may result in partial channel estimation and exploitation, it is expected not to degrade the overall multi-user performance as long as a proper beam and user selection strategy is designed. For the scalar case, it has been evidenced in \cite{B.K} that the beam and user selection by ACS could improve overall performance over the ones without channel sparsification. In what follows, before proceeding with the matrix-weight bipartite graph representation, we take a step back to propose a more general formulation of ACS from the lens of combinatorial optimization. \subsection{Generalized Multi-dimensional Active Channel Sparsification (MD-ACS)} Given the above matrix-weight graph representation, we reformulate the original ACS \cite{B.K} in a more general way. The generalization lies in two aspects: one is to extend one-to-one matching to many-to-many matching, the other one is to {generalize scale-weight (i.e., single-dimensional) to matrix-weight (i.e., multi-dimensional) matching with rate maximization and interference mitigation embedded instead of maximizing multiplexing gain.} \subsubsection{From One-to-one to Many-to-Many Matching} In the original formulation in \eqref{*} of ACS, a subgraph ${\mathcal G}'$ is selected with active beams ${\mathcal B}'$ and users ${\mathcal U}'$, and the maximal bipartite matching is constructed in the induced subgraph ${\mathcal G}'=({\mathcal B}',{\mathcal U}',{\mathcal E}')$. It has been shown in \cite{B.K} that the cardinality of the maximal bipartite matching is equal to the multiplexing gain of multi-user transmission. If we take a step back, instead of working on the maximal (one-to-one) bipartite matching in the selected graph, we could consider the many-to-many matching on the original bipartite graph ${\mathcal G}$, where a number of beams can be associated to one user, and likewise each beam can serve multiple users as long as inter-user interference is properly controlled. As such, a more general formulation can be given by \begin{subequations}\label{GACS} \begin{align} ({\mathcal P}_2): \quad \max & \quad w({\mathcal M}^*_{{\mathcal G}})\\ \label{eq:degree_constraint} \mathrm{s.t.} & \quad \text{deg}_{{\mathcal G}}(u_i) \leq \kappa_{b,i}, \quad \forall u_i\in{\mathcal U} \\ \label{eq:power_constraint} & \quad \text{deg}_{{\mathcal G}}(b_m) \le \kappa_{u,m}, \quad \forall b_m\in{\mathcal B} \end{align} \end{subequations} where ${\mathcal M}^*_{{\mathcal G}}$ is the set of many-to-many matching, which is a generalization of one-to-one matching, {$\kappa_{b,i} \le T/2$ is the maximum beams can be assigned to user $i$ to guarantee that channel estimation is feasible \cite{B.K}}, and $\kappa_{u,m}$ the maximum users that can reuse the same beam $m$ so that not much interference is caused one another. In contrast to the one-to-one matching, many-to-may matching allows each vertex on one side to be matched with multiple vertices on the other side. The above many-to-many weighted matching is equivalent to the generalized multi-assignment problem (GMAP) \cite{park1998lagrangian}, which is a generalized version of the assignment problem corresponding to one-to-one matching. The GMAP considers to assign a set of tasks to a set of agents. When a task is assigned to an agent, it produces profit and incurs cost. The aim of GMAP is to assign each task to multiple agents, where one agent can conduct multiple tasks, so that the total cost of all tasks is minimized and/or the total profit is maximized. Under the context of the multiple beam-user assignment, the above generalized ACS formulation can be reformulated as a GMAP with an integer programming as follows \begin{subequations}\label{eq:GAP_1} \begin{align} ( {\mathcal P}_2'): \max_{z_{i,m}} \quad & \sum_{m=1}^{M/2} \sum_{i=1}^{N_U} w_{i,m} z_{i,m} \label{eq:gap-obj}\\ {\rm s.t.} \quad &\sum_{i=1}^{N_U}{z_{i,m}} \leq \kappa_u,\ \forall b_m\in{\mathcal B} \label{eq:gap-beam} \\ &\sum_{m=1}^{M/2} {z_{i,m}} \leq \kappa_b,\ \forall u_i\in{\mathcal U} \label{eq:gap-user}\\ &z_{i,m}\in \{0,1\}, \ \forall b_m\in{\mathcal B}, \forall u_i\in{\mathcal U} \end{align} \end{subequations} where $z_{i,m}$ is a binary decision variable such that $z_{i,m}=1$ indicates the $m$-th block beam is assigned to $i$-th user, and 0 otherwise; $w_{i,m}$ is the corresponding profit for such an assignment, and it is a function of the matrix-weight $\mat{e}_{i,m}$; $\kappa_u$ and $\kappa_b$ are the maximum number of users and block beams to match each beam and user, respectively. For simplicity, we assume each user (resp. beam) is associated to the same number of beams (resp. users). \subsubsection{From Single-dimensional to Multi-dimensional Matching} The generalized formulation in \eqref{eq:GAP_1} reduces the size of the integer program compared to \eqref{*} to a great extent, thanks to the GMAP formulation and the matrix-weight bipartite representation of beam-user association. However, the merits in the original formulation, e.g., multiplexing gain maximization in \eqref{eq:MILP-0} and interference control in \eqref{eq:MILP-5}, are totally lost. To remedy the above reformulation, in what follows, we integrate the consideration of sum rate maximization into the objective function, especially into the parameters $\{w_{i,m}\}$, and relegate the interference control to a constraint. Such a remedy results in a nonlinear formulation, which motivates us to propose a greedy algorithm to solve it in an efficient way. \paragraph{Embedding Sum Rate Maximization and Interference Control} Given the subset of selected users ${\mathcal S}=\{i:y^*_i=1\}$, the achievable rate of $i$-th user with downlink precoder $\vect{p}_i$ can be written by \begin{align} R_i=\log\left(1+\frac{\Abs{\vect{h}_i^\H \vect{p}_i}}{\sigma^2+\sum_{j \in {\mathcal S} \backslash i} \Abs{\vect{h}_i^\H \vect{p}_j}} \right). \end{align} For the sake of tractability of optimization, we consider an asymptotic version of sum rate when $M_x,M_y \to \infty$ so that the asymptotic zero-forcing precoder of $i$-th user can be simply written by the column vectors of common basis $\mat{V}$. In particular, we have \begin{align} \vect{p}_i & \in{\mathcal R}\{\vect{h}_i \} \cap {\mathcal N} \{\vect{h}_j, j\in {\mathcal S} \backslash i\}\\ &={\{\mat{V}_m: x_m {\hbox{tr}}([\hbox{\boldmath$\Sigma$}_{i}]_m) \ge \delta, x_my_j {\hbox{tr}}([\hbox{\boldmath$\Sigma$}_{j}]_m) \le \delta, \forall j\in[N_U] \backslash i,m\in [M/2]\}} \end{align} where ${\mathcal R}\{\cdot\}$ and $ {\mathcal N}\{\cdot\}$ are the range and null spaces of the subspace spanned by the vectors, and $\delta$ is a threshold to determine if the block beam is strong enough to be considered. Hence, with such asymptotic precoder, the asymptotic rate\footnote{We point out that the asymptotic rate here is with respect to the number of antennas, which is different from those at high SNR in the literature.} of $i$-th user can be written as \begin{align} R_i^\infty&=\log{\left(1+\frac{{\hbox{tr}}\left(\sum_{m=1}^{M/2}x_m[\hbox{\boldmath$\Sigma$}_i]_m[\hbox{\boldmath$\Sigma$}_i]_m^\H\right)}{\sigma^2+{\hbox{tr}}\left(\sum_{m=1}^{M/2}\sum_{j=1,j\neq i}^{N_U}y_j x_m[\hbox{\boldmath$\Sigma$}_i]_m[\hbox{\boldmath$\Sigma$}_j]_m^\H\right)}\right)}\\ &=\log{\Big(\sigma^2+{\hbox{tr}}\big(\sum_{m=1}^{M/2}\sum_{j=1}^{N_U}y_j x_m[\hbox{\boldmath$\Sigma$}_i]_m[\hbox{\boldmath$\Sigma$}_j]_m^\H\big)\Big)}-\log{\Big(\sigma^2+{\hbox{tr}}\big(\sum_{m=1}^{M/2}\sum_{j=1,j\neq i}^{N_U}y_j x_m[\hbox{\boldmath$\Sigma$}_i]_m[\hbox{\boldmath$\Sigma$}_j]_m^\H\big)\Big)} \notag \\ &\ge \sum_{m=1}^{M/2} \left(\log{\Big(\frac{2\sigma^2}{M}+{\hbox{tr}}\big(\sum_{j=1}^{N_U}y_j x_m[\hbox{\boldmath$\Sigma$}_i]_m[\hbox{\boldmath$\Sigma$}_j]_m^\H\big)\Big)} - \eta_{i,m}\right) \end{align} where the first term is due to Jensen's inequality with $\log(\cdot)$ being a concave function, and the second term is due to an artificially introduced constraint \begin{align} \log{\Big(\sigma^2+{\hbox{tr}}\big(\sum_{m=1}^{M/2}\sum_{j=1,j\neq i}^{N_U}y_j x_m[\hbox{\boldmath$\Sigma$}_i]_m[\hbox{\boldmath$\Sigma$}_j]_m^\H\big)\Big)} \le \sum_{m=1}^{M/2} {\eta_{i,m}} \end{align} With Jensen's inequality, the above constraint can be relaxed to \begin{align} \label{eq:rate-constraint} \sum_{m=1}^{M/2} \log{\Big(\frac{2\sigma^2}{M}+{\hbox{tr}}\big(\sum_{j=1,j\neq i}^{N_U}y_j x_m[\hbox{\boldmath$\Sigma$}_i]_m[\hbox{\boldmath$\Sigma$}_j]_m^\H\big)\Big)} \le \sum_{m=1}^{M/2} \eta_{i,m}. \end{align} Let us introduce two matrices $\mat{p} \in \mbox{\bb C}^{N_U \times \frac{M}{2}}$ and $\mat{c} \in \mbox{\bb C}^{N_U \times \frac{M}{2}}$ such that \begin{align} [\mat{p}]_{i,m}&=\log {\hbox{tr}}\Big(\sum_{j=1}^{N_U}y_j[\hbox{\boldmath$\Sigma$}_i]_m[\hbox{\boldmath$\Sigma$}_j]_m^\H\Big),\label{profit}\\ [\mat{c}]_{i,m}&=\log {\hbox{tr}}\Big(\sum_{j=1,j\neq i}^{N_U}y_j[\hbox{\boldmath$\Sigma$}_i]_m[\hbox{\boldmath$\Sigma$}_j]_m^\H\Big).\label{cost} \end{align} The maximization of the asymptotic sum rate with joint user and beam selection can be approximately formulated in the following way \begin{subequations}\label{eq:GAP} \begin{align} ({\mathcal P}_3): \max_{z_{i,m}} \quad & \sum_{m=1}^{M/2} \sum_{i=1}^{N_U} z_{i,m} [\mat{p}]_{i,m} \label{eq:gap-obj}\\ {\rm s.t.} \quad & \eqref{eq:gap-beam}, \eqref{eq:gap-user},\\ &[\mat{c}]_{i,m} z_{i,m} \le \eta_{i,m}, \ \forall b_m\in{\mathcal B}, \forall u_i\in{\mathcal U} \label{eq:gap-cost} \\ &z_{i,m}\in \{0,1\}, \ \forall b_m\in{\mathcal B}, \forall u_i\in{\mathcal U} \end{align} \end{subequations} where the objective function \eqref{eq:gap-obj} comes from the lower bound of the asymptotic rate, with the constant parts dropped for simplicity, and the final constraint \eqref{eq:gap-cost} due to the constraint \eqref{eq:rate-constraint} to control interference, and $z_{i,m}=x_my_i$ is binary-valued. The above optimization formulation can be recognized as a GMAP with an additional constraint \eqref{eq:gap-cost}. The lower bound of the asymptotic rate can be regraded as the profits, and the constrained term in \eqref{eq:rate-constraint} can be treated as costs. As such, we refer to $\mat{p}$ and $\mat{c}$ as the profits and costs matrices, respectively. While the optimization problem \eqref{eq:GAP} is linear for the parameters $\{z_{i,m}\}$, the profits and costs matrices $\mat{p}$ and $\mat{c}$ are dependent of user selection $\{y_j\}$, which is entangled with $\{z_{i,m}\}$ as $z_{i,m}=x_my_i$. This makes the problem a nonlinear integer program with respect to $\{x_m\}$ and $\{y_i\}$, which is challenging to solve. To overcome this, we propose a low-complexity greedy algorithm, avoiding overlaps between any two users in the matrix-weight bipartite graph. \paragraph{Greedy Algorithm} As detailed earlier, given the channel covariance matrix $\hat{\mat{r}}_i$ with permuted rows and columns from the original one ${\mat{r}}_i$, we can construct a matrix-weight bipartite graph representation where the matrix weights $[{\hbox{\boldmath$\Sigma$}}_i]_m$ come from the block diagonalization of $\hat{\mat{r}}_i$. {However, when it comes to the practical scenarios with a finite number of antennas, $\hat{\mat{r}}_i$ is not perfectly block-diagonalizable with the DFT matrix as in \eqref{eq:blk-diag}. To overcome this, a possible way is to approximate the matrix-weight $[{\hbox{\boldmath$\Sigma$}}_i]_m$ by \begin{align} \label{eq:practical-decompose} [\hat{\hbox{\boldmath$\Sigma$}}_i]_m &= [(\mat{f}_{M_y}\otimes \mat{f}_{M_x}\otimes \mat{i}_2)^\H \hat{\mat{r}}_i (\mat{f}_{M_y}\otimes \mat{f}_{M_x}\otimes \mat{i}_2)]_{m,m} \end{align} where $[\cdot]_{m,m}$ is the $m$-th $2\times 2$ block diagonal submatrix with $m \in [M/2]$. It is readily verified that $\lim_{M_x,M_y \to \infty} [\hat{\hbox{\boldmath$\Sigma$}}_i]_m = [{\hbox{\boldmath$\Sigma$}}_i]_m$ for all $i,m$. Thus, in what follows, we use $[\hat{\hbox{\boldmath$\Sigma$}}_i]_m$ instead of $[{\hbox{\boldmath$\Sigma$}}_i]_m$ for algorithm design in the practical scenarios. } For ease of presentation, we introduce a $N_U \times \frac{M}{2} $ matrix $\hbox{\boldmath$\Psi$}$ with $[\hbox{\boldmath$\Psi$}]_{i,m}={\hbox{tr}}([\hat{\hbox{\boldmath$\Sigma$}}_i]_m)$ to indicate the contribution of the $m$-th block-beam to the $i$-th user. Let us define a binary matrix $\mat{a}'$ for the greedy algorithm with elements specified as \begin{align} \label{eq:adj-matrix} [\mat{a}']_{i,m}=\left\{ \Pmatrix{1, \quad \text{if } m \in \text{max}^{n_p} \big\{[\hbox{\boldmath$\Psi$}]_{i,m'},\forall m' \in [M/2]\big\},\\0, \quad \text{Otherwise,}} \right. \end{align} where $n_p\in[M/2]$ is a tunable integer parameter, and $\text{max}^{n_p}\{{\mathcal A}\}$ returns the indices of the largest $n_p$ values in the set ${\mathcal A}$. Here $\mat{a}'$ serve as a mask to filter out the insignificant weight matrices $\{[\hat{\hbox{\boldmath$\Sigma$}}_i]_m, \forall m\}$ and only keep $n_p$ largest ones. In particular, if we set $n_p=\kappa_b$, then after masking with $\mat{a}'$, there are at most $\kappa_b$ block beams left that are connected to each user, so that the constraint \eqref{eq:gap-user} is automatically satisfied. For the greedy algorithm, according to the asymptotic analysis of sum rate, we define a specific evaluation function as \begin{align}\label{eq:final-profit} \Phi(\mat{p},\mat{c}) = \sum_{i} \sum_m x_my_i([\mat{p}]_{i,m} - [\mat{c}]_{i,m}), \end{align} where $\mat{p}$ and $\mat{c}$ are profit and cost matrix as shown in \eqref{profit} and \eqref{cost}, for which $[\hat{\hbox{\boldmath$\Sigma$}}_i]_m$ is used. Given the bipartite graph representation ${\mathcal G}=({\mathcal B},{\mathcal U},{\mathcal E})$ and the matrix-weight $[\hat{\hbox{\boldmath$\Sigma$}}_i]_m$ on the edge $(b_m,u_i)$, we propose a greedy algorithm to solve the optimization problem $({\mathcal P}_3)$. The detailed procedure is outlined in Algorithm \ref{greedy_algorithm}. \begin{algorithm}[t] \caption{Greedy Algorithm for Generalized MD-ACS} \begin{algorithmic}[1] \State {\bf Input}: $\{\hat{\mat{r}}_i, \forall i\}$, $\kappa_u, \kappa_b$ \State {\bf Initialization}: $x_m=y_i=1$ for all $i \in [N_U],m \in [M/2]$ \State Produce $2 \times 2$ diagonal submatrices $\{[\hat{\hbox{\boldmath$\Sigma$}}_i]_m\}_{i=1,m=1}^{N_U, M/2}$ from $\{\hat{\mat{r}}_i\}_{i=1}^{N_U}$ according to \eqref{eq:practical-decompose} \State Construct the bipartite graph representation ${\mathcal G}=({\mathcal B},{\mathcal U},{\mathcal E})$ with matrix-weights $\{[\hat{\hbox{\boldmath$\Sigma$}}_i]_m\}$ for the edge $(b_m,u_i) \in {\mathcal E}$, and construct the weight matrix $\hbox{\boldmath$\Psi$}$ with $[\hbox{\boldmath$\Psi$}]_{i,m}={\hbox{tr}}([\hat{\hbox{\boldmath$\Sigma$}}_i]_m)$ \State Compute profit and cost matrix $\mat{p}, \mat{c}$ as \eqref{profit}-\eqref{cost} with $[\hat{\hbox{\boldmath$\Sigma$}}_i]_m$ \State Construct a binary matrix $\mat{a}'$ as in \eqref{eq:adj-matrix} with $n_p=\kappa_b$, such that \eqref{eq:gap-user} is satisfied \State Update $\hbox{\boldmath$\Psi$}$ as $\hbox{\boldmath$\Psi$}\gets\mat{a}'\odot \hbox{\boldmath$\Psi$}$, and set ${\mathcal M} = \{m: \sum_{i=1}^{N_U} x_my_i [\mat{a}']_{i,m}> \kappa_u, \forall m \in [M/2]\}$ \While{${\mathcal M} \neq \emptyset$} \State Select the beam $m \in {\mathcal M}$ \State Compute \eqref{eq:final-profit} as $\Phi_b$ if the beam is switched off, i.e., $x_m=0$ \State Compute \eqref{eq:final-profit} as $\Phi_u$ if only $\kappa_u$ users with the largest $[\hbox{\boldmath$\Psi$}]_{i,m}$ are selected, i.e., $y_i=0$ for all $i \notin \max^{\kappa_u}\{[\hbox{\boldmath$\Psi$}]_{i',m}, \forall i'\}$ \If{$\Phi_b > \Phi_u$} \State $x_m = 0$, and $[\mat{a}']_{i,m}=0$, $\forall i \in [N_U]$ \Else \State $y_i=0$, and $[\mat{a}']_{i,m}=0$, $\forall m \in [M/2]$, $i \notin \max^{\kappa_u}\{[\hbox{\boldmath$\Psi$}]_{i',m},\forall i'\}$ \EndIf \State Update $\hbox{\boldmath$\Psi$}$ as $\hbox{\boldmath$\Psi$} \gets [\mat{a}']\odot \hbox{\boldmath$\Psi$}$ \State Update ${\mathcal M} \gets \{m: \sum_{i=1}^{N_U} x_m y_i [\mat{a}']_{i,m}> \kappa_u, \forall m \in [M/2]\}$ \EndWhile \State {\bf Output}: $\{x_m\}_{m=1}^{M/2}$, $\{y_i\}_{i=1}^{N_U}$ \end{algorithmic} \label{greedy_algorithm} \end{algorithm} Let us explain the greedy algorithm in detail. Instead of maximizing the profit with the cost as the constraint in \eqref{eq:GAP}, we define a new profit function as in \eqref{eq:final-profit} which takes both original profits and costs into account. At the beginning, each user $i$ selects $\kappa_b$ block beams with the largest weights as specified by $\mat{a}'$ in \eqref{eq:adj-matrix}. This is to make the constraint \eqref{eq:gap-user} automatically satisfied. Then, each block beam $m$ determines whether the number of served users exceeds its capability $\kappa_u$ to satisfy the constraint \eqref{eq:gap-beam}. For those beams with more than $\kappa_u$ users served that violate the constraint \eqref{eq:gap-beam}, we need to determine if it is better to switch off this beam $m$, or some users so that the constraint \eqref{eq:gap-beam} is satisfied. To make such a decision, we compute and compare two quantities $\Phi_b$ and $\Phi_u$ when either option is applied with respect to the newly defined profit in \eqref{eq:final-profit}. This operation repeats till the constraint \eqref{eq:gap-beam} is satisfied for all active block beams. After each iteration, the weight matrix $\hbox{\boldmath$\Phi$}$ and the binary matrix $\mat{a}'$ will be updated, so that the deactivated users or beams will not be considered in the future. As such, the greedy algorithm results in a feasible solution after at most $\frac{M}{2}$ updates. To summarize, compared with the original single-dimension ACS formulation in \cite{B.K,khalilsarai2020dual}, our proposed MD-ACS with greedy algorithm has the following advantages. \begin{itemize} \item While the original ACS is dedicated to the maximization of multiplexing gain, our proposed MD-ACS takes both sum rate maximization and interference control into account, which leads to better performance at finite SNR, as will be shown in Section \ref{sec:numerical}. \item In the original ACS, the same threshold is applied for all users and beams to construct the bipartite graph representation, and the resulting graph is sensitive to such threshold. In addition, there are quite a few tunable parameters in \eqref{eq:ACS-MILP}, which are challenging to fine-tune to arrive at the sweet spot for the optimal solution, so that an improper choice probably results in severe performance degradation. For our proposed MD-ACS, the integer-valued parameters $\kappa_u$ and $\kappa_b$ are used to construct the bipartite graph, and the resulting graph is more flexible and suitable for greedy search. \item Due to the pre-determined bipartite graph representation and the implicit user selection, the original ACS is suitable to the homogeneous scenarios, whereas our proposed greedy algorithm is suitable for both homogeneous and heterogeneous scenarios (e.g., with both indoor and outdoor users), thanks to the adaptive bipartite graph construction and the explicit user selection, as will be demonstrated in Section \ref{sec:numerical}. \end{itemize} \section{Numerical Results} \label{sec:numerical} In this section, we provide the numerical results of our proposed method --- generalized multi-dimensional active channel sparsification (MD-ACS) --- compared with the state-of-the-art ones in the practical DP-UPA FDD massive MIMO scenarios. The following baseline methods are considered for comparison. \begin{itemize} \item \textbf{No Selection}: All users and beams are activated. \item \textbf{JSDM}: A clustering algorithm that divides users into groups according to the similarity of their channel covariance matrices \cite{JSDM}. {In each group, a user is randomly selected on behalf of the corresponding cluster}. In JSDM, the number of clusters $K$ is essential, and is set to $K=\sum_i y_i^*$, where $\vect{y}^*$ is the user selection vector obtained by our greedy algorithm. \item \textbf{ACS}: The original ACS on scale-weight bipartite graph representation, which was first proposed in \cite{B.K} for ULA, and later on extended to DP-ULA in \cite{khalilsarai2020dual}; \item \textbf{ACS-Matrix}: The conventional ACS with the MILP formulation on the matrix-weight bipartite graph representation, where the constraint \eqref{eq:MILP-5} is replaced by \begin{align} P y_i \le \sum_{b_m \in {\mathcal B}} {\hbox{tr}}\left([\mat{w}]_{i,m}\right) x_{m}, \quad \forall u_i \in {\mathcal U}. \end{align} \item \textbf{Greedy Algorithm}: The proposed MD-ACS with greedy algorithm implementation as specified in Algorithm \ref{greedy_algorithm}. \end{itemize} The downlink channel training and precoding follow the procedure in Section \ref{sec:signal-model}, where the pilot matrix $\mat{s}$ is a $T\times M'$ orthogonal matrix, with {$M'=2\sum_{m=1}^{M/2}x_m \le M$ being the number of activated virtual beams}, and the average pilot signal power is set to $\rho^{\text{p}}=1$. \begin{figure} \begin{minipage}{0.48\linewidth} \centering \includegraphics[width= 3.2in,angle=0]{SNR_32_15_1} \caption{Sum rate versus SNR with $4\times 4\times 2$ DP-UPA, $N_U=15$ users and $T=16$ timeslots.} \label{SNR_32_15} \end{minipage} \quad \begin{minipage}{0.48\linewidth} \centering \includegraphics[width= 3.2in,angle=0]{T_32_15_1} \caption{Sum rate versus pilot dimension with $4\times 4\times 2$ DP-UPA, $N_U=15$ users and $\mathrm{SNR}=20$ dB.} \label{T_32_15} \end{minipage} \begin{minipage}{0.48\linewidth} \centering \includegraphics[width= 3.2in,angle=0]{SNR_32_30_1} \caption{Sum rate versus SNR with $4\times 4\times 2$ DP-UPA, $N_U=30$ users and $T=16$ timeslots.} \label{SNR_32_30} \end{minipage} \quad \begin{minipage}{0.48\linewidth} \centering \includegraphics[width= 3.2in,angle=0]{T_32_30_1} \caption{Sum rate versus pilot dimension with $4\times 4\times 2$ DP-UPA, $N_U=30$ users and $\mathrm{SNR}=20$ dB.} \label{T_32_30} \end{minipage} \vspace{-20pt} \end{figure} For the simulation scenarios, we consider FDD downlink transmission in a single-cell massive MIMO network, where the base station is equipped with $M=M_x \times M_y \times 2$ DP-UPA antenna and serves $N_U$ single-antenna users. In order to evaluate the algorithms comprehensively and fairly, we adopt the QuaDRiGa channel model \cite{jaeckel2014quadriga} to generate downlink channel vectors $\vect{h}$. According to the 3GPP and the QuaDRiGa specifications \cite{QuadrigaD}, the Inter-Site Distance is set to be 500m and the `3GPP-3D-UMA' scenario is considered. For the users' located rules, the minimum distance from users to the base station is 10m. We choose 50\% indoor and 50\% outdoor users for downlink transmission, where the height of all the outdoor users is set to 1.5m. In all simulation scenarios, we assume the downlink channel covariance matrix is somehow available, which can be simply obtained by $\mat{r}=\frac{1}{N}\sum_{t=1}^{N} \vect{h}_t \vect{h}_t^\H$ using $N=1000$ downlink channel vector realizations $\vect{h}_t$ generated from QuaDRiGa, or obtained from uplink channel covariance matrix by leveraging uplink-downlink reciprocity (e.g., \cite{B.K}). Unless otherwise explicitly specified, for all the simulation scenarios, we choose the following parameters: $\kappa_b=3$; $\kappa_u=20$ for 64 antenna configuration and $\kappa_u=12$ for 32 antenna configuration; pilot dimensions of training phase $T=16$; for 32 antennas (Fig. \ref{SNR_32_15}-\ref{T_32_30}), the antenna array is $M_x=M_y=4$ and the FDD frame length is $T_c = 64$, and for 64 antennas (Fig. \ref{SNR_64_30}-\ref{User_64_20T}), $M_x=4$, $M_y=8$ and $T_c = 72$. Figures \ref{SNR_32_15} and \ref{T_32_15} illustrate the sum rate of downlink transmission with $N_U=15$ users in total versus SNR and the pilot dimension of the training phase, respectively. We can observe in Fig. \ref{SNR_32_15} that, the proposed MD-ACS with greedy algorithm outperforms all other methods, and the gap is increasing as SNR goes. The ACS-like methods (i.e., ACS and ACS-Matrix) perform poorly at high SNR compared with No Selection, which is probably due to the fact that the number of users for selection is quite limited so that activating all users may not be a bad idea. In Fig. \ref{T_32_15}, it appears the sum rate first increases as the pilot dimension does, because higher pilot dimension yields higher estimation accuracy of downlink channel, and therefore more accuracy downlink precoding. The sum rate is decreasing when pilot dimension increases further, because the more resource the training phase occupies, the less the transmission phase could use. For the ACS-like methods (i.e., ACS and ACS-Matrix), it looks too many users have been switched off, which results in severe performance degradation when $T$ is large. In Figures \ref{SNR_32_30} and \ref{T_32_30}, $N_U=30$ users are considered. It is observed that our proposed MD-ACS with greedy algorithm consistently outperform other methods. In this scenario, with a sufficiently large number of users, both ACS-Matrix and ACS perform better than JSDM and No Selection. The No Selection method confronts severe performance degradation -- it is because there are too many users in the network, and user selection is crucial. In these simulations, ACS-Matrix outperforms the conventional ACS approach, which demonstrates the effectiveness of using matrix-weight bipartite graph representation. Notably, from Fig. \ref{T_32_15} and Fig. \ref{T_32_30}, the optimal pilot dimensions that maximize the sum rate are different across algorithms. The optimal pilot dimension of ACS is around $T=12$ while others are around $T = 16$. This suggests that ACS seems more dedicated to beam selection, while others (including the greedy algorithm) prefer user selection. \begin{figure} \begin{minipage}{0.48\linewidth} \centering \includegraphics[width= 3.2in,angle=0]{SNR_64_30_1} \caption{Sum Rate versus SNR with $4\times 8\times 2$ DP-UPA, $N_U=30$ users and $T=16$ timeslots.} \label{SNR_64_30} \end{minipage} \quad \begin{minipage}{0.48\linewidth} \centering \includegraphics[width= 3.2in,angle=0]{T_64_30_1} \caption{Sum Rate versus Timeslots with $4\times 8\times 2$ DP-UPA, $N_U=30$ users and $\mathrm{SNR}=20$ dB.} \label{T_64_30} \end{minipage} \begin{minipage}{0.48\linewidth} \centering \includegraphics[width= 3.2in,angle=0]{SNR_64_60_1} \caption{Sum Rate versus SNR with $4\times 8\times 2$ DP-UPA, $N_U=60$ users and $T=16$ timeslots.} \label{SNR_64_60} \end{minipage} \quad \begin{minipage}{0.48\linewidth} \centering \includegraphics[width= 3.2in,angle=0]{T_64_60_1} \caption{Sum Rate versus Timeslots with $4\times 8\times 2$ DP-UPA, $N_U=60$ users and $\mathrm{SNR}=20$ dB.} \label{T_64_60} \end{minipage} \end{figure} Further, in Fig. \ref{SNR_64_30}-\ref{T_64_60}, we increase the number of antennas from 32 to 64, and consider two scenarios with $N_U=30$ and $N_U=60$ users. For the 64 DP-UPA antenna scenario with 30 users, compared with the 32-antenna case, the sum rate improvement of our greedy algorithm over other methods is diminishing, because the capability of serving more users is enhanced with more antennas, and thus user selection is not crucial. When the number of users increases from $N_U=30$ to $N_U=60$, we observe the same phenomenon as that in Figures \ref{SNR_32_30} and \ref{T_32_30}. Interestingly, from Figures \ref{T_64_30} and \ref{T_64_60}, we find that even if the number of antennas is increased, the optimal pilot dimension is still around $T=16$ timeslots. It is worth noting that, ACS-Matrix with matrix-weight graph representation outperforms the conventional scalar-weight ACS method. It suggests that the matrix-weight formulation is more suitable than scalar-weight ACS for the DP-UPA scenario. As such, the improvement of our proposed greedy algorithm comes from two aspects: the matrix-weight MILP formulation and the search-based user/beam selection strategy. \begin{figure} \centering \includegraphics[width= 3.2in,angle=0]{User_64_16T} \caption{Sum Rate versus the number of users with $4\times 8\times 2$ DP-UPA, $T=16$ timeslots and $\mathrm{SNR}=20$ dB.} \label{User_64_20T} \vspace{-20pt} \end{figure} Fig. \ref{User_64_20T} shows the sum rate versus the number of users in the cell when the pilot dimension is set to $T=16$. We can observe that: (1) When the number of users is small, e.g., $N_U \le 30$, user selection is unnecessary, because the sum rates are nearly the same for the proposed greedy algorithm, compared with No Selection. (2) As the number of users increases, the benefit of user selection emerges, and it becomes crucial when the number of users is large, e.g., $N_U \ge 45$. (3) Our proposed MD-ACS with greedy algorithm always outperforms the conventional ACS thanks to the matrix-weight graph representation and the search-based greedy user/beam selection. \section{Conclusion} In this work, we have investigated downlink sparsifying precoder design and user selection in DP-UPA FDD massive MIMO systems using active channel sparsification (ACS). By extending the original scalar-weight bipartite graph representation of user-beam association to a matrix-weight bipartite graph, we proposed a generalized multi-dimensional ACS (MD-ACS) for DP-UPA antenna configurations with a nonlinear integer program formulation. Inspired by the generalized multi-assignment problem, we proposed an efficient greedy algorithm to solve the nonlinear integer problem, and observed its superiority in extensive simulation results using QuaDriGa channel models. We believe such an improvement of the ACS methodology could pave the way for the potential deployment of ACS to the practical FDD massive MIMO systems.

522b68fcf5496f12d3407d0008598a8065daa7fe

\section{#1}} \renewcommand{\thesection.\arabic{equation}}{\arabic{section}.\arabic{equation}} \def\thesection.\arabic{equation}{\thesection.\arabic{equation}} \newcommand{\ell}{\ell} \newcommand{\nonumber\\}{\nonumber\\} \newcommand\para{\paragraph{}} \newcommand{\ft}[2]{{\textstyle\frac{#1}{#2}}} \newcommand{\eqn}[1]{(\ref{#1})} \newcommand\mf{\mathcal{F}} \newcommand\ma{\mathcal{A}} \newcommand\Q{\mathcal{Q}} \newcommand\R{\mathbb{R}} \newcommand\Z{\mathbb{Z}} \newcommand\F{\mathbb{F}} \newcommand\ml{\mathcal{L}} \newcommand\Pf{\mathrm{Pf}} \newcommand\D{\mathcal{D}} \newcommand\cp{\mathbb{CP}} \newcommand\C{\mathbb{C}} \newcommand\T{\mathbb{T}} \newcommand\diff{\mathrm{d}} \newcommand{\partial}{\partial} \newcommand{\mathbb{E}}{\mathbb{E}} \newcommand{\mathrm{d}}{\mathrm{d}} \newcommand{\mathrm{e}}{\mathrm{e}} \newcommand{\mathrm{i}}{\mathrm{i}} \newcommand{\overline\nabla}{\overline\nabla} \newcommand{\mathrm{e}}{\mathrm{e}} \newcommand{\mathrm{vol}}{\mathrm{vol}} \newcommand{\mathrm{Vol}}{\mathrm{Vol}} \newcommand{\mathscr{Y}^{p,q}}{\mathscr{Y}^{p,q}} \newcommand{p}{p} \newcommand{q}{q} \newcommand{\mathscr{L}^{\mathtt{a},\mathtt{b},\mathtt{c}}}{\mathscr{L}^{\mathtt{a},\mathtt{b},\mathtt{c}}} \newcommand{\gamma}{\gamma} \newcommand{B}{B} \newcommand{\Bsca}{B} \newcommand{\Btsca}{\tilde B} \newcommand{\mathcal J}{\mathcal J} \newcommand{\mathtt{a}}{\mathtt{a}} \newcommand{\mathtt{b}}{\mathtt{b}} \newcommand{\mathtt{c}}{\mathtt{c}} \newcommand{\mathtt{d}}{\mathtt{d}} \newcommand{{p}}{{p}} \newcommand{{q}}{{q}} \newcommand{\mathscr{F}}{\mathscr{F}} \newcommand{\mathscr{Z}}{\mathscr{Z}} \newcommand{c_{\mathrm{sugra}}}{c_{\mathrm{sugra}}} \newcommand{{\varphi}}{{\varphi}} \newcommand{B}{B} \newcommand{\mathtt{c}}{c} \newcommand{S_{\mathrm{SUSY}}}{S_{\mathrm{SUSY}}} \newcommand{z}{z} \newcommand{\mathtt{m}}{\mathtt{m}} \newcommand{\mathrm{Re}\, }{\mathrm{Re}\, } \newcommand{\mathrm{Im}\, }{\mathrm{Im}\, } \newcommand{\hat{g}}{\hat{g}} \newcommand{\hat{J}}{\hat{J}} \newcommand{\hat{\Omega}}{\hat{\Omega}} \newcommand{\hat \omega}{\hat \omega} \newcommand{\mathrm{E}}{\mathrm{E}} \newcommand{\mathtt{1}}{\mathtt{1}} \newcommand\pamo{{\cal P}_\mathrm{A}(p)} \newcommand\pb{{\cal P}_\mathrm{B}(p)} \newcommand\qamo{{\cal Q}_\mathrm{A}(q)} \newcommand\qb{{\cal Q}_\mathrm{B}(q)} \newlength{\sswidth} \newcommand{\sla}[1]{ \settowidth{\sswidth}{$#1$} \mbox{$\not{\hspace*{-0.3\sswidth}#1}$}} \DeclareMathOperator{\image}{\mathrm{Im}\,} \numberwithin{equation}{section} \newcommand{\JFS}[1]{{\bf [JFS: #1]}} \newcommand{\remove}[1]{\textcolor{red}{#1}} \newcommand{\jmpi}[1]{{\bf [jmpi: #1]}} \newcommand{\ab}[1]{{\bf [ab: #1]}} \newcommand{\kappa}{\kappa} \newcommand{U(1)_{\mathcal{J}}}{U(1)_{\mathcal{J}}} \newcommand{\mathcal{A}_{\mathcal{J}}}{\mathcal{A}_{\mathcal{J}}} \newcommand{\mathcal{F}_{\mathcal{J}}}{\mathcal{F}_{\mathcal{J}}} \newcommand{\mathcal{J}}{\mathcal{J}} \begin{document} \begin{titlepage} \vskip 1cm \begin{center} {\Large \bf Twisted D3-brane and M5-brane compactifications} \vskip 0.5cm {\Large \bf from multi-charge spindles} \vskip 2.0cm {Andrea Boido, Juan Manuel P\'erez Ipi\~na and James Sparks} \vskip 0.5cm \textit{Mathematical Institute, University of Oxford,\\ Andrew Wiles Building, Radcliffe Observatory Quarter,\\ Woodstock Road, Oxford, OX2 6GG, U.K.\\} \end{center} \vskip 2.0 cm \begin{abstract} \noindent We construct families of supersymmetric AdS$_3\times Y_7$ and AdS$_3\times Y_8$ solutions to type IIB string theory and M-theory, respectively. Here $Y_7$ is an $S^5$ fibration over $\Sigma$, while $Y_8$ is an $S^4$ fibration over $\Sigma_g\times \Sigma$, where $\Sigma_g$ is a Riemann surface of genus $g>1$ and $\Sigma$ is a two-dimensional orbifold known as a spindle. We interpret the solutions as near-horizon limits of $N$ D3-branes wrapped on $\Sigma$ and $N$ M5-branes wrapped on $\Sigma_g\times \Sigma$, respectively. These are holographically dual to $d=2$, $(0,2)$ SCFTs, and we show that the central charge and superconformal R-symmetry of the gravity solutions agree with dual field theory calculations. \end{abstract} \end{titlepage} \pagestyle{plain} \setcounter{page}{1} \newcounter{bean} \baselineskip18pt \tableofcontents \newpage \section{Introduction and summary}\label{sec:intro} Interesting classes of superconformal field theories (SCFTs) may be obtained in string theory and M-theory by wrapping branes on a compact space $\Sigma$, and flowing to the infrared (IR). Supergravity solutions describing the near-horizon limits of such wrapped branes were first constructed in the pioneering work of \cite{Maldacena:2000mw}, describing D3-branes and M5-branes wrapping a constant curvature Riemann surface $\Sigma_g$ of genus $g>1$. This work has subsequently been generalized in many directions, leading to a large literature on the subject. Until recently such constructions have realized supersymmetry on the wrapped brane worldvolume via a so-called topological twist \cite{Witten:1988ze}. Here the spin connection on the wrapped directions $\Sigma$ is effectively cancelled by the coupling of the spinor to external R-symmetry gauge fields, which geometrically are connections on the normal bundle to $\Sigma$ in spacetime. The upshot is that the preserved Killing spinors are sections of trivial bundles over $\Sigma$, and in fact constant. A different realization of supersymmetry has recently been presented in \cite{Ferrero:2020laf}. Here a family of supersymmetric AdS$_3\times \Sigma$ solutions was constructed in $d=5$ minimal gauged supergravity, where $\Sigma$ is a two-dimensional orbifold surface known as a spindle. This is topologically a two-sphere, but with conical deficit angles $2\pi (1-\frac{1}{n_\pm})$ at the two poles, specified by coprime positive integers $n_+\neq n_-$. Equivalently, $\Sigma=\mathbb{WCP}^1_{[n_-,n_+]}$ is a weighted projective space. When $n_--n_+$ is divisible by 3 these may be uplifted on an $S^5$ internal space, leading to AdS$_3\times Y_7$ solutions of type IIB supergravity in which $Y_7$ takes the fibred form \begin{align}\label{introfibre} S^5 \ \hookrightarrow\ Y_7 \ \rightarrow\ \Sigma \, = \, \mathbb{WCP}^1_{[n_-,n_+]}\, . \end{align} Since these ten-dimensional solutions are sourced only by five-form flux $F_5$, with $N$ units of flux through the $S^5$ fibre of \eqref{introfibre}, these are naturally interpreted as near-horizon limits of $N$ D3-branes wrapped on the spindle $\Sigma$. However, the Killing spinors on $\Sigma$ are not constant, and moreover are sections of non-trivial spinor bundles \cite{Ferrero:2020twa}. The flux through $\Sigma$ of the $d=5$ Abelian R-symmetry gauge field in gauged supergravity is correspondingly not simply proportional to the Euler number $\chi(\Sigma)$, given by \begin{align}\label{introEuler} \chi(\Sigma)\, = \, \frac{n_-+n_+}{n_-n_+}\, , \end{align} as is characteristic of a topological twist. In the first part of this paper we generalize the construction of \cite{Ferrero:2020laf} by allowing for more general fibrations in \eqref{introfibre}. More precisely, the $d=5$ minimal gauged supergravity solution of \cite{Ferrero:2020laf} is a special case of a more general class of (local) AdS$_3$ solutions to $d=5$, $U(1)^3$ gauged supergravity with two additional parameters, first constructed in \cite{Kunduri:2007qy}. These also uplift on $S^5$, with the three gauge fields of $U(1)^3$ determining the fibration of this internal space over the $d=5$ spacetime. The resulting supersymmetric AdS$_3\times Y_7$ solutions were in fact constructed before reference \cite{Kunduri:2007qy}, in \cite{Gauntlett:2006ns}. The approach we take to the global analysis of these supergravity solutions is different to that in \cite{Kunduri:2007qy}, \cite{Gauntlett:2006ns}, and instead follows \cite{Ferrero:2020laf}. Specifically, we first construct AdS$_3\times \Sigma$ solutions of the $d=5$, $U(1)^3$ gauged supergravity theory, where $\Sigma=\mathbb{WCP}^1_{[n_-,n_+]}$ is a spindle specified by coprime positive integers $n_->n_+$. We then appropriately quantize the three Abelian R-symmetry gauge field fluxes through $\Sigma$, in such a way that the resulting fibration \eqref{introfibre} is well-defined, with a smooth total space $Y_7$. More precisely, defining the fluxes \begin{align}\label{introfluxes} Q_I\, \equiv\, \frac{1}{2\pi}\int_\Sigma F^{(I)}\, , \end{align} where $F^{(I)}$, $I=1,2,3$, are the three Abelian R-symmetry gauge field curvatures, we construct solutions with \begin{align}\label{introcharges} Q_1\, = \, \frac{q}{n_-n_+}\, , \qquad Q_2\, = \, \frac{p}{n_-n_+}\, , \qquad Q_3\, = \, \frac{n_--n_+-q-p}{n_-n_+}\, . \end{align} Here $q,p>0$ are positive integers, constrained so that $q+p<n_--n_+$, and where $q$ and $p$ have no common factor with either of $n_-,n_+$. Notice that the charges \eqref{introcharges} satisfy \begin{align}\label{introchargesum} Q_1+Q_2+Q_3\, =\, \frac{n_--n_+}{n_-n_+}\, . \end{align} This should be contrasted with a topological twist, where instead on the right hand side of \eqref{introchargesum} one would have the Euler number of the spindle \eqref{introEuler}. The solutions of \cite{Ferrero:2020laf} are recovered by setting $Q_1=Q_2=Q_3$, or equivalently $p=q=\frac{1}{3}(n_--n_+)$, which then requires $n_--n_+$ to be divisible by 3. The central charge of these solutions is given by \begin{equation} \label{introc} c \, = \, \frac{3 p q (n_- - n_+ - q - p)}{n_- n_+ [(n_--n_+) (p+q) + n_-n_+-p^2-p q-q^2]} N^2 \,. \end{equation} The solution is thus specified by five positive integers: $n_-, n_+$ determine the topology of the wrapped spindle $\Sigma=\mathbb{WCP}^1_{[n_-,n_+]}$, $q,p$ determine the R-symmetry fluxes \eqref{introcharges} and hence twisting in the fibration \eqref{introfibre}, and finally $N$ is the number of wrapped D3-branes. As in \cite{Ferrero:2020laf}, the description of these solutions suggests that the dual $d=2$, $(0,2)$ SCFTs arise from a twisted compactification of $\mathcal{N}=4$ SYM on $\Sigma$, where there are background R-symmetry fluxes given by \eqref{introcharges}. We make a check on this conjecture by computing both the central charge and superconformal R-symmetry of the $d=2$ field theories, making use of the anomaly polynomial of $\mathcal{N}=4$ SYM together with $c$-extremization \cite{Benini:2012cz}. We find precise agreement with the corresponding supergravity quantities. When $Q_1=Q_2$, or equivalently $p=q$, the solutions to $d=5$, $U(1)^3$ supergravity are also solutions to Romans $\mathcal{N}=4$, $SU(2)\times U(1)$ supergravity \cite{Romans:1985ps}. The latter is a consistent truncation of $d=11$ supergravity on the Maldacena-N\'u\~nez solution referred to at the start of this introduction \cite{Gauntlett:2007sm}. We thus also obtain AdS$_3\times Y_8$ solutions of $d=11$ supergravity, where $Y_8$ takes the fibred form \begin{align}\label{introfibreM5} S^4 \ \hookrightarrow\ Y_8 \ \rightarrow\ \Sigma_g\times \Sigma\, . \end{align} Here $\Sigma_g$ is a constant curvature Riemann surface of genus $g>1$. The original Maldacena-N\'u\~nez solution corresponds to the AdS$_5$ vacuum of the Romans theory, and is dual to the $d=4$, $\mathcal{N}=2$ SCFT living on $N$ M5-branes wrapped on the Riemann surface $\Sigma_g$, which is holomorphically embedded inside a Calabi-Yau two-fold. The uplift to $d=11$ of our $Q_1=Q_2$ solution, with parameter $q$, then describes the further twisted compactification of this four-dimensional theory on the spindle $\Sigma$, with $0<q<\frac{1}{2}(n_--n_+)$ determining the twisting. The central charge of these gravity solutions is \begin{align}\label{cM5intro} c \, = \, \frac{4q^2(n_- - n_+ - 2 q)}{ n_- n_+ [n_- (n_+ + 2 q) - q (2 n_+ + 3 q)]} (g-1) N^3 \, . \end{align} We again reproduce this result from a dual field theory calculation, this time utilizing the anomaly polynomial of the theory on $N$ M5-branes, doubly wrapped on $\Sigma_g\times \Sigma$. The outline of the paper is as follows. In section \ref{sec:sugra} we construct the AdS$_3\times Y_7$ solutions of interest by first constructing $d=5$ AdS$_3\times \Sigma$ solutions in $U(1)^3$ gauged supergravity, and then imposing appropriate quantization conditions on the gauge field fluxes through the spindle $\Sigma=\mathbb{WCP}^1_{[n_-,n_+]}$ so that the uplift on $S^5$ leads to a smooth seven-manifold $Y_7$. We also compute the central charge and R-symmetry gauge field for these gravity solutions. In section \ref{sec:M5} we similarly construct AdS$_3\times Y_8$ solutions to M-theory by uplifting the same $d=5$ solutions, but now necessarily with $Q_1=Q_2$, and compute the central charge. In section \ref{sec:fieldtheory} we compute the $d=2$ anomaly polynomial for the twisted compactification of $\mathcal{N}=4$ SYM on $\Sigma$, and using $c$-extremization then compute the exact central charge and superconformal R-symmetry in field theory, finding agreement with the gravity duals. We perform a similar computation for the theory on $N$ M5-branes wrapped on $\Sigma_g\times \Sigma$, again finding agreement with the central charge computed from the supergravity solution. We conclude in section \ref{sec:discuss} with a brief discussion. \ \noindent {\bf Note added}: As this paper was being completed we learned of the work \cite{Hosseini:2021fge}, which has overlap with the D3-brane results presented here. \section{D3-brane supergravity solutions}\label{sec:sugra} In this section we construct a family of supersymmetric AdS$_3$ solutions to $d=5$, $U(1)^3$ gauged supergravity. These uplift on an $S^5$ internal space to corresponding AdS$_3\times Y_7$ solutions of type IIB supergravity. The local form of the solutions was first presented in \cite{Gauntlett:2006ns}, where they were obtained from multi-charge superstar solutions in \cite{Cvetic:1999xp}. Shortly afterwards the same solutions were rediscovered in \cite{Kunduri:2007qy}, where they arise as the uplifts of near-horizon limits of a family $d=5$ unbalanced black ring solutions. The novelty of this section is the global analysis that we perform for the allowed values of the parameters in the solution, which is very different to the approach taken in \cite{Kunduri:2007qy} and \cite{Gauntlett:2006ns}. Instead we follow \cite{Ferrero:2020laf}. The resulting description will be crucial to match the dual field theory computation in section \ref{sec:fieldtheory}. \subsection{Local form of the solutions}\label{sec:local} The action of $d=5$, $U(1)^3$ gauged supergravity is \begin{equation} \begin{split} S & \, = \, \frac{1}{16\pi G_{(5)}}\int \Big\{ \diff^5 x \sqrt{-\det g}\, \Big[ R + 4\sum_{I=1}^3 (X^{(I)})^{-1} - \frac{1}{2}\sum_{I=1}^3 (X^{(I)})^{-2}(\partial X^{(I)})^2 \\ & \qquad \qquad \qquad \qquad \qquad \qquad - \frac{1}{4}\sum_{I=1}^3(X^{(I)})^{-2}(F^{(I)})^2\Big] - F^{(1)}\wedge F^{(2)} \wedge A^{(3)}\,\Big\} . \end{split} \end{equation} Here $A^{(I)}$, $I=1,2,3$, are the three $U(1)$ gauge fields, with field strengths $F^{(I)}=\diff A^{(I)}$, and the three scalar fields $X^{(I)}$ satisfy the constraint $X^{(1)}X^{(2)}X^{(3)}=1$. We have set the gauge coupling constant to $1$. A consistent truncation to minimal gauged supergravity may be obtained by setting the scalars $X^{(I)}=1$ and the three gauge fields equal $A^{(I)}=A$, $I=1,2,3$, where to compare to the conventions in \cite{Ferrero:2020laf} one should also rescale the gauge field $A$ by factor of $\frac{3}{2}$. Our starting point is the following supersymmetric AdS$_3$ solution to this theory \cite{Kunduri:2007qy}: \begin{equation} \begin{split} \label{5dmetric} \diff s^2_5 & \, = \, H(x)^{1/3} \left[\diff s^2_{\mathrm{AdS}_3} + \frac{1}{4P(x)} \diff x^2 +\frac{P(x)}{H(x)}\diff\phi^2\right]\, ,\\ A^{(I)}&\, = \, \frac{x-\alpha}{x+3K_I} \diff \phi\, ,\\ X^{(I)} &\, =\, \frac{H(x)^{1/3}}{x+3K_I}\, . \end{split} \end{equation} Here $\diff s^2_{\mathrm{AdS}_3}$ a unit radius metric on AdS$_3$, the metric functions are the polynomials \begin{equation}\label{metricfunctions} \begin{split} H(x) & \, \equiv \, (x + 3K_1)(x + 3K_2)(x + 3K_3) \, = \, x^3 + 3 c_1 x + c_2\, ,\\ P(x) &\, \equiv \, H(x) - (x-\alpha)^2\, , \\ \end{split} \end{equation} and $\alpha$, $K_I$ are constants, with the latter satisfying the constraint $K_1+K_2+K_3=0$. In the first line of \eqref{metricfunctions} we have introduced the elementary symmetric polynomials for the parameters $K_I$, namely $c_1 \equiv 3(K_1 K_2 + K_2 K_3 + K_1 K_3)$, $c_2 \equiv 27 K_1 K_2 K_3$. We note that the solution to minimal gauged supergravity studied in \cite{Ferrero:2020laf} is obtained by setting $K_I=0$ for all $I=1,2,3$, which is then parametrized by the single parameter~$\alpha$. Any solution of $d=5$, $U(1)^3$ gauged supergravity may be uplifted (locally) on an $S^5$ internal space to a solution of type IIB supergravity \cite{Cvetic:1999xp}. The ten-dimensional metric~is \begin{align}\label{lift} L^{-2}\diff s^2_{10}& \, = \, W^{1/2}\diff s^2_5 +W^{-1/2}\sum_{I=1}^3 (X^I)^{-1} \left[ \diff \mu_I^2 + \mu_I^2 (\diff\phi_I + A^{(I)})^2 \right]\, , \end{align} while the self-dual five-form flux is \begin{align} \label{fiveform} L^{-4} F_5 \, = \, (1 + \star_{10}) \Big\{ &2 \sum_{I=1}^3 [(X^{(I)})^2 \mu_I^2 - W X^{(I)}] \mathrm{vol}_5 + \frac{1}{2} \sum_{I=1}^3 (X^{(I)})^{-1} \star_5 \diff X^{(I)} \wedge \diff (\mu_I^2) \nonumber \\ & + \frac{1}{2 } \sum_{I=1}^3 (X^{(I)})^{-2} \diff (\mu_I^2) \wedge (\diff \phi_I + A^{(I)}) \wedge \star_5 F^{(I)} \Big\}\, . \end{align} Here we have introduced an arbitrary length scale parameter $L>0$ into the form of the solution, which will be fixed via flux quantization later. We have denoted the volume form on $\diff s^2_5$ by $\mathrm{vol}_5$, and $\star$ denotes the Hodge duality operator in the appropriate dimension. As in \cite{Gauntlett:2006ns} we have corrected the sign of the second factor in \eqref{fiveform}, compared with the expression given in \cite{Cvetic:1999xp}. We have also introduced the warp factor function \begin{align}\label{warp} W\, = \, \sum_{I=1}^3 \mu_I^2 X^{(I)} \, > \, 0 \, . \end{align} Finally, $\{\mu_I, \phi_I\}$ form a system of polar coordinates on $S^5\subset \R^2\oplus\R^2\oplus\R^2$, where correspondingly $\sum_{I=1}^3\mu_I^2=1$ and the angular coordinates $\phi_I$ each have period $2\pi$. Notice that the warp factor function \eqref{warp} depends on both the internal $S^5$ coordinates $\mu_I$ and {\it a priori} on the $d=5$ coordinates, via the scalar fields $X^{(I)}$. Applying the uplifting formula \eqref{lift} to the AdS$_3$ solution \eqref{5dmetric} gives rise to the following warped AdS$_3\times Y_7$ metric \begin{align}\label{3+7} \diff s^2_{10}& \, = \, L^2W^{1/2}H(x)^{1/3} \left(\diff s^2_{\mathrm{AdS}_3} + \diff s^2_{Y_7}\right)\, , \end{align} where the seven-dimensional metric on $Y_7$ is \begin{equation} \begin{split}\label{7dmetric} \diff s^2_{Y_7} & \, = \, \frac{\diff x^2}{4P(x)} + \frac{P(x)}{H(x)} \diff\phi^2\\ & \quad + \frac{1}{W H(x)^{2/3}}\sum_{I=1}^3 (x + 3K_I) \left[\diff \mu_I^2+\mu_I^2 \left(\diff \phi_I + \frac{(x-\alpha)}{x+3K_I}\diff \phi \right)^2 \right]\, . \end{split} \end{equation} One can verify that this is the same metric as that given in section 5.2 of \cite{Gauntlett:2006ns}, where we identify their coordinates and parameters, in terms of those presented here, as $w=x-\alpha$, $z=-2\phi$, $q_I=\alpha+3K_I$. In particular, reference \cite{Gauntlett:2006ns} also shows that the solution is supersymmetric, with the dual $d=2$ CFT having $(0,2)$ supersymmetry. \subsection{Global analysis and general solution}\label{sec:global} Given a local AdS$_3\times Y_7$ solution of the form \eqref{7dmetric}, one would like to choose the parameters $\alpha, K_I$ and ranges of coordinates so as to obtain a smooth metric on a compact internal space $Y_7$. An analysis of this was carried out in \cite{Gauntlett:2006ns} (see also \cite{Kunduri:2007qy}), where one looks at the degeneration loci of Killing vector fields on $Y_7$, which are constant linear combinations of the four vector fields $\partial_\phi$ and $\partial_{\phi_I}$, $I=1,2,3$. We instead follow the approach of \cite{Ferrero:2020laf}, first obtaining a suitably regular $d=5$ AdS$_3\times \Sigma$ solution, where $\Sigma=\mathbb{WCP}^1_{[n_-,n_+]}$ is a two-dimensional orbifold known as a spindle. We then appropriately quantize the $U(1)^3$ gauge field fluxes so as to obtain a regular fibration of $S^5$ over this five-dimensional solution. We begin then with the $d=5$ metric in \eqref{5dmetric}, which we rewrite as \begin{align} \diff s^2_5 & \, = \, H(x)^{1/3} \left(\diff s^2_{\mathrm{AdS}_3} + \diff s^2_\Sigma\right)\, , \end{align} with the two-dimensional surface $\Sigma$ having coordinates $x$, $\phi$. In order for the metric to have the correct signature and be non-singular we would like to choose the parameters so that $H(x)>0$ is strictly positive, and similarly $P(x)\geq 0$. In order that $\Sigma$ forms a compact surface without boundary, we furthermore assume that $x\in [x_1,x_2]$ lies between two roots of the polynomial $P(x)$ in \eqref{metricfunctions}. In a certain regime of parameters, that we make explicit below, we find that the three roots $x_i$, $i=1,2,3$, of $P(x)$ are real and positive, and we then take $0<x_1<x_2<x_3$ so that $P(x)\geq 0$ for $x\in [x_1,x_2]$. Defining $\varrho_i = 2 (x-x_i)^{1/2}$ for $i=1,2$ we find that as $x \rightarrow x_i$ the metric on the two-dimensional surface $\Sigma$ approaches \begin{align} \diff s^2_{\Sigma}\, \simeq\, \frac{1}{4 P'(x_i)} \left(\diff\varrho_i^2+\kappa_i^2 \varrho_i^2 \, \diff \phi^2\right)\,, \qquad \mbox{where}\quad \kappa_i \, \equiv \, \left|\frac{P'(x_i)}{x_i - \alpha}\right|\,, \quad i\, = \, 1,2\, . \end{align} As for the solution in \cite{Ferrero:2020laf}, it is not possible to remove both conical singularities at the roots $x=x_i$ by making a single choice $\Delta\phi$ of the period for $\phi$. Instead we impose \begin{align}\label{delwcond} \Delta \phi\, = \, \frac{2\pi}{\kappa_1 n_+}\, = \, \frac{2\pi}{\kappa_2 n_-}\, , \end{align} where $n_\pm\in \mathbb{N}$. The resulting surface $\Sigma=\mathbb{WCP}^1_{[n_-,n_+]}$ is then an orbifold known as a spindle. This is topologically a two-sphere, but with conical deficit angles $2\pi(1-\frac{1}{n_\pm})$ at the poles $x=x_1$, $x_2$. After a little computation we find that the second equality in \eqref{delwcond} may be written as \begin{equation} \label{quotweights2} \frac{n_-}{n_+} \, = \, \left| \frac{x_2-\alpha}{x_1-\alpha} \right| \left| \frac{-1+2x_1+x_2}{-1+x_1+2x_2} \right| \, , \end{equation} where the roots $x_1, x_2$ satisfy the equations \begin{equation} \begin{split} \label{ceq} 2\alpha + 3c_1 & \, = \, x_1+x_2 -x_1^2-x_2^2-x_1 x_2\,,\\ c_2 & \, = \, \alpha^2+x_1^2x_2+x_1x_2^2-x_1x_2\, , \end{split} \end{equation} and we have eliminated $x_3$ via the relation $x_1+x_2+x_3=1$. We will also need the fluxes of the gauge field strengths $F^{(I)}=\diff A^{(I)}$ through the surface $\Sigma$. Using the expression for $A^{(I)}$ in \eqref{5dmetric} we compute \begin{equation} \label{fluxes} \begin{split} Q_I &\, \equiv \, \frac{1}{2\pi} \int_\Sigma F^{(I)} \, = \, \frac{(x_2 - x_1) (\alpha + 3K_I)}{(x_1 + 3K_I) (x_2 + 3K_I)}\frac{\Delta\phi}{2\pi} \, . \end{split} \end{equation} As explained in \cite{Ferrero:2020laf}, \cite{Ferrero:2020twa}, for $\Sigma=\mathbb{WCP}^1_{[n_-,n_+]}$ the appropriate quantization condition on these fluxes is that \begin{align}\label{quantize} Q_I \, \equiv \, \frac{p_I}{n_-n_+}\, , \qquad \mbox{where}\quad p_I\in \Z\, . \end{align} Specifically, the circles parametrized by $\phi_I$ inside the internal space $S^5\subset \C^3$ then give a well-defined orbifold circle fibration over $\Sigma$. Moreover, provided the integers $p_I$ are coprime to both of $n_\pm$, the total space of this fibration is smooth.\footnote{See appendix A of \cite{Ferrero:2020twa} for a detailed discussion of this.} This leads to a fibration \begin{align}\label{fibre} S^5\ \hookrightarrow \ Y_7 \ \rightarrow \ \Sigma \, = \, \mathbb{WCP}_{[n_-,n_+]}\, , \end{align} where the total space $Y_7$ is a compact seven-manifold, and the twisting is determined by the integers $p_I$. At this stage we have a three-parameter family of solutions, determined by the constants $\alpha, K_1, K_2$, where the roots $x_1,x_2$ obey \eqref{ceq}. We would like to express the solution in terms of the more physical flux parameters $p_I$ and $n_\pm$, the ratio of the latter determined via \eqref{quotweights2}. It turns out to be convenient to express quantities in terms of two of the three fluxes, which we take to be \begin{align} p_1 \, \equiv \, q\, , \qquad p_2 \, \equiv \, p\, . \end{align} After some work, it is possible to solve the three equations \eqref{quotweights2}, and \eqref{ceq} for $\alpha$, $K_1$, $K_2$, $x_1,$ and $x_2$ in terms of $n_\pm$, together with $q$ and $p$. We find the rather unwieldy expressions \begin{align}\label{soln} x_1 &\, = \, \frac{1}{\mathcal{D}} \Big\{(p^2+p q+q^2)^2+3 n_+ p q (p+q+n_+)+2 (n_+-n_-) (p+q) (p^2+q^2) \nonumber\\ & \qquad +(n_-^2+n_+^2) (p^2+q^2)-n_- n_+ (3 p^2+4 p q+3 q^2)+n_- n_+ (n_--n_+) (p+q)\Big\} \:, \nonumber\\[10pt] x_2 &\, = \, \frac{1}{\mathcal{D}} \Big\{(p^2+p q+q^2)^2-3 n_- p q (p+q-n_-)+2 (n_+-n_-) (p+q) (p^2+q^2) \nonumber\\ & \qquad +(n_-^2+n_+^2) (p^2+q^2)-n_- n_+ (3 p^2+4 p q+3 q^2)+n_- n_+ (n_--n_+) (p+q)\Big\} \:,\nonumber\\[2pt] &\\[-7pt] K_1 & \, =\, \frac{1}{3\mathcal{D}} \Big\{3 p^2 (p+q)^2-(p^2+p q+q^2)^2+2 (n_--n_+) (q^2-2 p^2) (p+q)-2n_- n_+ p q\nonumber\\ &\qquad\qquad -(n_-^2+n_+^2) (q^2-2 p^2)+3n_- n_+ (q^2-2p^2)-n_- n_+ (n_--n_+) (q-2 p)\Big\} \:,\nonumber\\[10pt] K_2 &\, =\, \frac{1}{3\mathcal{D}} \Big\{3 q^2 (p+q)^2-(p^2+p q+q^2)^2+2 (n_--n_+) (p^2-2 q^2) (p+q)-2n_- n_+ p q\nonumber\\ &\qquad\qquad -(n_-^2+n_+^2) (p^2-2 q^2)+3n_- n_+ (p^2-2q^2)-n_- n_+ (n_--n_+) (p-2 q)\Big\} \:,\nonumber \end{align} where we have defined the denominator term \begin{align} \mathcal{D}\, \equiv\, 3\, [(n_- - n_+)(p+q)+n_-n_+-p^2-p q-q^2]^2\, . \end{align} The parameter $\alpha$ may then be determined from the first equation in \eqref{ceq}. The third flux is computed to be \begin{equation} Q_3\, = \, \frac{n_- - n_+ - q - p}{n_- n_+}\, , \end{equation} which notice also satisfies the quantization condition \eqref{quantize}, with \begin{align} p_1 \, = \, q\, , \qquad p_2\, = \, p\, , \qquad p_3\, = \, n_--n_+-q-p\, . \end{align} We then also have \begin{align}\label{chargesum} Q_1 + Q_2 + Q_3 \, = \, \frac{n_--n_+}{n_-n_+}\, . \end{align} We find that the roots obey the assumed inequalities $0 < x_1 < x_2 < x_3$ provided \begin{align}\label{inequalities} q, \, p \, > \, 0\, , \qquad q+p\, < \, n_--n_+\, . \end{align} We now have a family of regular supersymmetric AdS$_3\times Y_7$ solutions, with $Y_7$ having the fibration structure \eqref{fibre}, parametrized by the integers $n_-,n_+,p$ and $q$. We address quantization of the five-form flux $F_5$ in the next subsection. \subsection{Central charge and R-symmetry}\label{sec:central} The AdS$_3$ solutions are, via the AdS/CFT correspondence, expected to be dual to $d=2$, $(0,2)$ SCFTs. The central charge of these CFTs is computed in gravity using the standard formula \cite{Brown:1986nw} \begin{equation} \label{ccharge} c \, = \, \frac{3L}{2G_{(3)}}\, , \end{equation} where $G_{(3)}$ is the effective Newton constant in three dimensions. In turn the latter is computed via dimensional reduction on $Y_7$, and is given by \cite{Couzens:2018wnk} \begin{equation} \frac{1}{G_{(3)}} \, = \, \frac{L^7}{G_{(10)}} \int_{Y_7} W^2 H(x)^{4/3} \, \text{vol}_7 \, = \, \frac{\pi^3 L^7 \Delta\phi}{2 G_{(10)}} \, (x_2-x_1) \: , \end{equation} where we have explicitly evaluated the integrals over $S^5$ and the spindle $\Sigma$. The ten-dimensional Newton constant is \begin{equation} G_{(10)} \, = \, \frac{(2\pi)^7 g_s^2 \, \ell_s^8}{16\pi} \:, \end{equation} where $\ell_s$ and $g_s$ are the constant string length and string coupling constant, respectively. The central charge \eqref{ccharge} is then \begin{equation} \label{cchargeagain} c \, = \, \frac{3 L^8 \Delta\phi}{32\pi^3 g_s^2 \, \ell_s^8} \, (x_2-x_1) \:. \end{equation} In order to obtain a good string theory solution we must also quantize the flux of the closed five-form \eqref{fiveform} through five-cycles in the ten-dimensional spacetime. In particular, integrating $F_5$ through a copy of the $S^5$ fibre, at any point in the $d=5$ spacetime, we find the total flux is \begin{equation} N \, \equiv \, \frac{1}{(2\pi \ell_s)^4 g_s} \int_{S^5} F_5 \, = \, \frac{L^4}{4\pi \ell_s^4 \, g_s}\, . \end{equation} Eliminating $L$ in the central charge \eqref{cchargeagain} using this last equation then gives \begin{equation}\label{cchargeneat} c \, = \, \frac{3 }{2 \pi} \Delta \phi (x_2-x_1)N^2\,. \end{equation} Finally, substituting the solution \eqref{soln} into \eqref{cchargeneat}, we find the remarkably simple expression \begin{equation} \label{cgrav3charge} c \, = \, \frac{3 p q (n_- - n_+ - q - p)}{n_- n_+ [(n_--n_+) (p+q) + n_-n_+-p^2-p q-q^2]} N^2 \,. \end{equation} This is our final formula for the central charge, parametrized in terms of the integers $n_-, n_+, p,q$ and $N$. It will be useful to note that \eqref{delwcond} gives the period $\Delta\phi$ of $\phi$ to be \begin{align}\label{Deltaphi} \frac{\Delta\phi}{2\pi}\, = \, \frac{ (n_- - n_+)(p+q)+n_-n_+-p^2-p q-q^2}{ n_-n_+( n_-+ n_+)}\, , \end{align} an expression we will need momentarily. The $U(1)$ R-symmetry of the dual $d=2$, $(0,2)$ field theory is realized in the gravity solution as a Killing vector field $R_{\mathrm{2d}}$ of the AdS$_3\times Y_7$ gravity solution. This is defined as a certain bilinear in the Killing spinors on $Y_7$ \cite{Gauntlett:2007ts}, and combining the general results of the latter reference with the form of the solution given in \cite{Gauntlett:2006ns} we can read off that this vector field is simply \begin{align}\label{Rvec} R_{\mathrm{2d}}\, = \, \partial_\phi\, . \end{align} Here we have normalized the R-symmetry so that the Killing spinor on $Y_7$ has unit charge under the Lie derivative along this vector field, giving an explicit phase dependence of $\mathrm{e}^{\mathrm{i} \phi}$. On the other hand, this Killing spinor arises as a tensor product of the $d=5$ Killing spinor with a spinor on the internal $S^5$. This ansatz preserves an $\mathcal{N}=1$ spinor, out of the full $\mathcal{N}=4$ supersymmetry of $S^5$, with the spinor on $S^5$ having charge $\frac{1}{2}$ under each of the three vector fields $\partial_{\phi_I}$ generating the $U(1)^3$ isometry of $S^5\subset \C^3$. As discussed in \cite{Ferrero:2020laf}, we would like to choose a gauge for the $d=5$ gauge fields such that the Killing spinor is uncharged under the $U(1)$ isometry that rotates the spindle. Since a gauge transformation of the $A^{(I)}$ precisely leads to a phase rotation in the corresponding Killing spinor, which has unit charge under each $A^{(I)}$, this may be achieved via the gauge/coordinate transformation \begin{align}\label{coordtrans} \tilde{\phi}_I \, =\, \phi_I + \tfrac{2}{3} \, \phi\, , \qquad \tilde{A}^{(I)} \, = \, A^{(I)} - \tfrac{2}{3} \, \diff \phi\, , \end{align} which ensures that $\diff\phi_I+A^{(I)}=\diff\tilde{\phi}_I+\tilde{A}^{(I)}$. Defining also \begin{align} \varphi \, \equiv \, \frac{2\pi}{\Delta\phi} \, \phi\, , \end{align} so that $\Delta\varphi = 2\pi$, and completing the coordinate transformation \eqref{coordtrans} by defining \begin{align}\label{tildephi} \tilde{\varphi}\, = \ \varphi\, , \end{align} we see that the R-symmetry vector \eqref{Rvec} is \begin{equation}\label{premixing} \begin{split} R_{\mathrm{2d}} &\, = \, \partial_\phi \, = \, \frac{2\pi}{\Delta\phi} \, \partial_\varphi \, = \, \frac{2\pi}{\Delta\phi} \left(\sum_{I=1}^3 \frac{\partial \tilde{\phi}_I}{\partial \varphi} \, \partial_{\tilde{\phi}_I} + \frac{\partial \tilde{\varphi}}{\partial \varphi} \, \partial_{\tilde{\varphi}}\right) \, = \, \frac{2}{3} \sum_{I=1}^3 \partial_{\tilde{\phi}_I} + \frac{2\pi}{\Delta\phi} \, \partial_{\tilde{\varphi}}\, . \end{split} \end{equation} Note that $\partial_{\tilde{\varphi}}$ generates the $U(1)$ isometry of the spindle, that we refer to as $U(1)_{\mathcal{J}}$ in the subsequent discussion. The gauge transformation \eqref{coordtrans} leads to a phase $\mathrm{e}^{\mathrm{i} (\tilde{\phi}_1+\tilde{\phi}_2+\tilde{\phi}_3)/2}$ in the Killing spinor on $Y_7$, but this is now invariant under $U(1)_{\mathcal{J}}$. We may then identify the Killing vector field on $S^5$ \begin{equation} R_{\mathrm{4d}}\, = \, \frac{2}{3} \sum_{I=1}^3 \partial_{\tilde{\phi}_I}\, , \end{equation} with the superconformal $\mathcal{N}=1$ R-symmetry before compactification of the theory on $\Sigma$, and from \eqref{premixing} hence write \begin{equation}\label{mixing} R_{\mathrm{2d}} \, = \, R_{\mathrm{4d}} + \frac{n_-n_+( n_-+ n_+)}{ (n_- - n_+)(p+q)+n_-n_+-p^2-p q-q^2}\, \partial_{\tilde{\varphi}}\, . \end{equation} Here in the second term we have substituted for $\Delta\phi$ using \eqref{Deltaphi}. Equation \eqref{mixing} states that the $d=4$ $U(1)$ R-symmetry mixes with $U(1)_{\mathcal{J}}$ in flowing to the $d=2$ $U(1)$ R-symmetry in the IR. We shall recover this formula from a dual field theory calculation in section~\ref{sec:fieldtheory}, along with the central charge \eqref{cgrav3charge}. \subsection{Special cases}\label{sec:cases} In this section we briefly analyse some interesting special cases of the general solutions of section \ref{sec:global}, in particular making contact with \cite{Ferrero:2020laf}. Setting $p=q$ implies that the fluxes $Q_1=Q_2$ are equal, and hence also $A^{(1)}=A^{(2)}$, and $X^{(1)}=X^{(2)}$. In this case \eqref{soln} simplifies~to \begin{equation}\label{twosoln} \begin{split} x_1 &\, = \, \frac{q\,(n_+ + q) [2 n_-^2 - 2 n_- (n_+ + 4 q) + q\, (5 n_+ + 9 q)]}{3\, [n_- (n_+ + 2 q) - q\,(2 n_+ + 3 q) ]^2} \,,\\[6pt] x_2 &\, = \, -\frac{q\,(n_- - q)[2 n_+^2 - 2 n_+ (n_- - 4 q) - q\, (5 n_- - 9 q)]}{3\, [n_- (n_+ + 2 q) - q\,(2 n_+ + 3 q) ]^2} \,,\\[6pt] K_1 &\, = \, K_2\, = \, \frac{q\,(n_- - n_+ - 3 q) (n_+ + q) (n_- - q)}{9\, [n_- (n_+ + 2 q) - q\,(2 n_+ + 3 q) ]^2} \,, \end{split} \end{equation} with corresponding central charge \begin{equation} \label{cgrav} c \, = \, \frac{3 q^2 (n_- - n_+ - 2q)}{n_- n_+ [n_- (n_+ + 2 q) - q\,(2 n_+ + 3 q) ]}N^2 \,. \end{equation} This sub-family of solutions also uplift to solutions of $d=11$ supergravity, as discussed in section \ref{sec:M5}. Finally, setting all three charges equal gives, from \eqref{chargesum}, \begin{align}\label{equals} p \, = \, q\, = \, \frac{1}{3}(n_--n_+)\, . \end{align} This is precisely the solution to minimal gauged supergravity studied in \cite{Ferrero:2020laf}, where we note that $p,q\in \Z$ then requires $n_--n_+$ to be divisible by 3, as discussed in \cite{Ferrero:2020laf}.\footnote{In the notation of that reference, we take the K\"ahler-Einstein four-manifold to be $\mathrm{KE}_4=\mathbb{CP}^2$, and $k=1$ then gives $S^5$ as the internal space. But this then requires $n_--n_+$ to be divisible by the Fano index of $\mathbb{CP}^2$, which is $I=3$. The parameters $p$ and $q$ in this paper should not be confused with those in \cite{Ferrero:2020laf}, and also recall that one should rescale our gauge field by a factor of $\frac{3}{2}$ to match to the conventions in \cite{Ferrero:2020laf}.} Imposing \eqref{equals}, our central charge \eqref{cgrav3charge} and R-symmetry gauge field \eqref{mixing} reduce to the expressions given in \cite{Ferrero:2020laf}. \section{M5-brane supergravity solutions}\label{sec:M5} In this section we construct a family of supersymmetric AdS$_3\times Y_8$ solutions to $d=11$ supergravity. These arise by uplifting the $Q_1=Q_2$ solutions of $d=5$, $U(1)^3$ gauged supergravity in section \ref{sec:cases} on the Maldacena-N\'u\~nez solution \cite{Maldacena:2000mw}, using the consistent truncation of \cite{Gauntlett:2007sm}. We interpret these M-theory solutions as the near-horizon limit of $N$ M5-branes wrapped on $\Sigma_g\times \Sigma$. \subsection{Romans supergravity and uplift}\label{sec:upliftM5} There are a number of different consistent truncations of type IIB supergravity on $S^5$, among them being both the $U(1)^3$ gauged supergravity of section \ref{sec:local}, but also the Romans $\mathcal{N} = 4$, $SU(2) \times U(1)$ supergravity \cite{Romans:1985ps}. The latter preserves a different subgroup of the full $SO(6)$ R-symmetry of the internal $S^5$. The bosonic sector of the Romans theory contains a scalar field $X$, a triplet of $SU(2)$ gauge fields $B_\mu^i$, $i=1,2,3$, a $U(1)$ gauge field $A_\mu$, and two two-forms. This theory can be further truncated by setting to zero the two-forms and truncating $SU(2)$ to its Abelian subgroup. It was shown in \cite{Lu:1999bw} that the resulting theory is the same as the $U(1)^3$ theory with two gauge fields $A^{(1)}=A^{(2)}$ and two scalar fields $X^{(1)}=X^{(2)}$ set equal, and in particular then \begin{equation}\label{trun} X \, = \, X^{(1)} \:, \qquad B_\mu^1 \, = \, B_\mu^2 \, = \, 0 \:, \qquad B_\mu^3 \, = \, A^{(1)}_\mu \:, \qquad A_\mu \, = \, A_\mu^{(3)} \:. \end{equation} On the other hand, this Romans supergravity theory is also a consistent truncation of $d=11$ supergravity, as shown in \cite{Gauntlett:2007sm}. The vacuum AdS$_5$ solution uplifts to a warped AdS$_5\times N_6$ solution, which for a given internal space $N_6$ is dual to an $\mathcal{N}=2$ SCFT in four dimensions. A particular example, studied in \cite{Gauntlett:2007sm}, is the Maldacena-N\'u\~nez solution, which is dual to the $d=4$, $\mathcal{N}=2$ SCFT living on M5-branes wrapping a Riemann surface, holomorphically embedded in a Calabi-Yau two-fold. More generally there are the AdS$_5\times N_6$ solutions of \cite{Gaiotto:2009gz}, corresponding to M5-branes wrapping a Riemann surface with punctures, and holographically dual to $d=4$, $\mathcal{N}=2$ SCFTs of class $\mathcal{S}$. Our five-dimensional solutions with $Q_1=Q_2$, given by \eqref{twosoln}, may thus also be uplifted to interesting classes of M5-brane solutions. Specifically, the M5-brane is wrapped on a Riemann surface (in general with punctures), which is then further wrapped on the spindle $\Sigma=\mathbb{WCP}^1_{[n_-,n_+]}$ to obtain a two-dimensional theory. In this section for simplicity we focus on the uplift on the Maldacena-N\'u\~nez solution. A solution to the $d=5$ Romans theory with scalar field $X$, and Abelian gauge fields $B^3=A^{(1)}$, $A=A^{(3)}$, with $B^1=B^2=0$ as in \eqref{trun}, uplifts to a $d=11$ solution with metric \cite{Gauntlett:2007sm}\footnote{We note that in \cite{Gauntlett:2007sm} the uplifting formula is stated for $d=5$ gauge coupling set to $m_{\mathrm{there}}=\frac{1}{2}$, while our $d=5$ solutions have this quantity set to 1. We thus need to rescale our fields, and in the notation of \cite{Gauntlett:2007sm} correspondingly $A^3_{\mathrm{there}}=2\sqrt{2} A^{(1)}$, $B_{\mathrm{there}}=2A^{(3)}$.} \begin{equation}\label{11d} \begin{split} L^{-2} \diff s_{11}^2 &= 2^{-2/3} \Omega^{1/3} \diff s_5^2 +2^{1/3} X \Omega^{1/3} \left(\diff \theta^2 + \diff s_{\Sigma_g}^2\right) \\ & + 2^{1/3} X \Omega^{-2/3} \sin^2\theta \left(\diff \psi + V + A^{(3)}\right)^2 + \frac{2^{-2/3}}{X^2} \, \Omega^{-2/3} \cos^2\theta D\mu_i D\mu_i \:. \end{split} \end{equation} Here $\diff s^2_5$ denotes the five-dimensional gauged supergravity metric, $\diff s^2_{\Sigma_g}$ is the metric on a unit radius hyperbolic plane, quotiented by a discrete group of isometries to obtain a compact Riemann surface of genus $g>1$, and we have introduced the warp factor function \begin{align} \Omega\, \equiv \, \cos^2\theta + \frac{1}{2 X^3} \,\sin^2\theta\, . \end{align} The coordinates $\theta$, $\psi$, and $\mu_i$, $i=1,2,3$, are (constrained) coordinates on $S^4 \subset \mathbb{R}^2 \oplus \mathbb{R}^3$, where $\theta\in [0,\frac{\pi}{2}]$ describes the polar direction in the latter splitting. Smoothness of the metric \eqref{11d} at $\theta=0$ fixes the period $\Delta\psi=2\pi$, and the local one-form $V$ on $\Sigma_g$ is such that $\diff V = - \mathrm{vol}_{\Sigma_g}$. We note that \begin{align}\label{VolSigmag} \mathrm{Vol}(\Sigma_g)\, = \, \int_{\Sigma_g} \mathrm{vol}_{\Sigma_g}\, = \, 4\pi (g-1) \, = \ -2\pi\, \chi(\Sigma_g)\, , \end{align} with $\chi(\Sigma_g)$ the Euler number of $\Sigma_g$. This identifies the $\R^2$ bundle over $\Sigma_g$, with unit circle in $\R^2=\C$ having coordinate $\psi$, as $T^*\Sigma_g$. This is a local Calabi-Yau two-fold, with the $d=11$ solution describing the near-horizon limit of a stack of M5-branes wrapped on the zero-section $\Sigma_g$. The M5-branes have corresponding normal space $\R^5=\C\oplus\R^3$, and denoting the first factor by the complex line bundle $\mathcal{N}_1$, we see that \begin{align} \mathcal{N}_1 \, = \, T^*\Sigma_g\otimes L_3\, , \end{align} where $A^{(3)}$ is a connection on a complex line bundle $L_3$ over the $d=5$ spacetime. For our solution with $Q_1=Q_2$ in \eqref{twosoln}, this is a line bundle over the spindle $\Sigma=\mathbb{WCP}^1_{[n_-,n_+]}$ with charge $Q_3$. Finally, $\mu_i$, $i=1,2,3$, are constrained coordinates describing a round $S^2\subset \R^3$, with $\sum_{i=1}^3\mu_i^2 = 1$, and the connection terms are given by \begin{align} D\mu_1 &\, = \, \diff \mu_1 +2 A^{(1)} \mu_2 \:,\qquad D\mu_2 \, =\, \diff \mu_2 -2 A^{(1)} \mu_1 \:, \qquad D\mu_3 \, = \, \diff \mu_3 \:. \end{align} Writing $\mu_1=\sin\vartheta\cos\nu$, $\mu_2=\sin\vartheta\sin\nu$, $\mu_3=\cos\vartheta$ in spherical polar coordinates, the twisted metric on $S^2$ is \begin{align} D\mu_i D\mu_i \, = \, \diff\vartheta^2 + \sin^2\vartheta(\diff\nu - 2A^{(1)})^2\, . \end{align} Since $\nu$ has $\Delta\nu=2\pi$, we see that writing $\R^3=\C\oplus\R$, the complex line $\C$ is twisted via $L_1^2$, with connection $2A^{(1)}$, so we have \begin{align} \mathcal{N}_2 \, = \, L_1^2\, . \end{align} The total normal bundle of the M5-branes is then $\mathcal{N}_1\oplus\mathcal{N}_2\oplus \R$, with $\R$ the $z$-axis direction in $\R^3$. The M-theory four-form $G_4$ may also be read off from the expression in \cite{Gauntlett:2007sm}, although we won't need its explicit form in what follows. \subsection{Central charge}\label{sec:M5central} We begin by writing the metric \eqref{11d} as \begin{equation} \diff s_{11}^2 \, = \, \left(16\,\Omega H(x)\right)^{1/3} \left(\diff s^2_{\mathrm{AdS}_3} + \diff s^2_{Y_8}\right) \:, \end{equation} where we have uplifted the $d=5$ solution of section \ref{sec:cases} with $Q_1=Q_2$. The eight-dimensional metric on $Y_8$ is then \begin{equation} \begin{split} \diff s^2_{Y_8} &= \frac{1}{4P(x)} \, \diff x^2 + \frac{P(x)}{H(x)} \, \diff\phi^2 + \frac{ X}{2H(x)^{1/3}}\left(\diff \theta^2 + \diff s_{\Sigma_g}^2\right) \\ &\quad + \frac{ X}{2\Omega H(x)^{1/3}} \sin^2\theta \left(\diff \psi + V + A^{(3)}\right)^2 + \frac{1}{4X^2\Omega H(x)^{1/3}} \cos^2\theta D\mu_i D\mu_i \:. \end{split} \end{equation} The central charge of the dual $d=2$, $(0,\, 2)$ field theories is again given by formula \eqref{ccharge}, where $G_{(3)}$ is computed via dimensional reduction on $Y_8$. We find \begin{equation} \frac{1}{G_{(3)}} \, = \, \frac{L^8}{G_{(11)}} \int_{Y_8} \left(16\,\Omega H(x)\right)^{3/2} \mathrm{vol}_8 \, = \, \frac{16\pi^2 L^8 \Delta\phi}{3 G_{(11)}} \,(x_2-x_1) \mathrm{Vol}(\Sigma_g)\:, \end{equation} where in our conventions the eleven-dimensional Newton constant is \begin{equation} G_{(11)} \, = \, \frac{(2\pi)^8\, \ell_p^9}{16\pi} \:. \end{equation} The central charge is then given by \begin{equation}\label{cchargeM} c \, =\, \frac{L^9 \Delta\phi}{2\pi^5 \ell_p^9}\,(x_2-x_1)\mathrm{Vol}(\Sigma_g)\:. \end{equation} We also need to impose flux quantization on the four-form in order to have a consistent M-theory solution. Integrating the expression for $G_4$ in \cite{Gauntlett:2007sm} over a copy of the $S^4$ fibre we get the number of M5 branes \begin{equation} N \, \equiv \, \frac{1}{(2\pi \ell_p)^3}\int_{S^4} G_4 \, = \, \frac{L^3}{\pi \ell_p^3} \:. \end{equation} Substituting for the length scale $L$ in \eqref{cchargeM}, $\mathrm{Vol}(\Sigma_g)$ given by \eqref{VolSigmag}, and also for $x_1$, $x_2$ in \eqref{twosoln} and $\Delta\phi$ in \eqref{Deltaphi} with $p=q$, we find the final central charge \begin{align}\label{cM5gravity} c \, = \, \frac{4q^2(n_- - n_+ - 2 q) }{ n_- n_+ [n_- (n_+ + 2 q) - q (2 n_+ + 3 q)]}(g-1) N^3 \, . \end{align} This is the central charge of a family of $d=2$, $(0,2)$ theories obtained by wrapping $N$ M5-branes on $\Sigma_g\times \Sigma$, with $\Sigma=\mathbb{WCP}^1_{[n_-,n_+]}$ a spindle, and the integer $q$ with $0<q<\frac{1}{2}(n_--n_+)$ determines the twisting of the normal bundle of the M5-branes. We shall recover this from a dual theory calculation in section \ref{sec:fieldtheory}. \section{Field theory}\label{sec:fieldtheory} Although our AdS$_3\times Y_7$ solutions were first constructed in \cite{Gauntlett:2006ns}, the description in the latter reference meant that the dual $d=2$, $(0,2)$ SCFTs were left unidentified. However, as in \cite{Ferrero:2020laf} our alternative $d=5$ construction of the solutions leads to a natural conjecture. Specifically, one begins with $\mathcal{N}=4$ SYM theory, which is holographically dual to the AdS$_5\times S^5$ vacuum of the $d=5$, $U(1)^3$ gauged supergravity theory. One then compactifies this theory on $\Sigma=\mathbb{WCP}^1_{[n_-,n_+]}$, together with background fluxes for the $U(1)^3$ Abelian symmetry given by \eqref{quantize}. Recall here that $p_1=q$, $p_2=p$, $p_3=n_--n_+-q-p$. The solution we have described suggests that this compactification of $\mathcal{N}=4$ SYM flows to a $d=2$, $(0,2)$ SCFT in the IR. We give evidence for this in sections~\ref{sec:anom} and \ref{sec:extreme} by computing the central charge and superconformal $U(1)$ R-symmetry using purely field theory methods, finding precise agreement with the supergravity results. There is a similar interpretation of the M5-brane solutions in section \ref{sec:M5}. In section \ref{sec:anomM5} we likewise compactify the theory on $N$ M5-branes on $\Sigma_g\times \Sigma$, with appropriate background fluxes, and compute the central charge of the $d=2$, $(0,2)$ SCFTs, again finding agreement with supergravity. \subsection{D3-brane anomaly polynomial}\label{sec:anom} Generalizing the approach of \cite{Ferrero:2020laf}, our starting point is the anomaly polynomial for $\mathcal{N}=4$ SYM, with background gauge field fluxes $F^{(I)}$ for the $U(1)^3\subset SO(6)$ Abelian global symmetry group of this theory. In the large $N$ limit this anomaly polynomial reads (see, for example, \cite{Hosseini:2020vgl}) \begin{equation} \label{anomaly4d} \mathcal{A}_{\mathrm{4d}} \, = \, c_1(F^{(1)}) c_1(F^{(2)}) c_1(F^{(3)}) \frac{N^2}{2} \, . \end{equation} Here $c_1(F^{(I)})$ denote the first Chern classes of the $U(1)$ bundles with gauge field curvatures $F^{(I)}$, respectively. We want to compactify this theory on $\Sigma =\mathbb{WCP}^1_{[n_-,n_+]}$, with fluxes given by \eqref{quantize}, and compute the anomaly polynomial of the resulting $d=2$ theory. For this we also need to take into account the $U(1)_{\mathcal{J}}$ global symmetry in $d=2$, coming from the $U(1)_{\mathcal{J}}$ isometry of $\Sigma$. Geometrically, this means that we want to compute the anomaly polynomial \eqref{anomaly4d}, where the six-manifold $Z_6$ on which it is defined is the total space of a $\Sigma$ fibration over $Z_4$. We may achieve this by introducing a corresponding connection form $\mathcal{A}_{\mathcal{J}}$ for $U(1)_{\mathcal{J}}$, and replacing $\diff \varphi \rightarrow \diff \varphi + \mathcal{A}_{\mathcal{J}}$. In doing so it is important that the $d=5$ Killing spinor is independent of $\tilde\varphi=\varphi$, which is true in the tilded gauge defined by \eqref{coordtrans}, \eqref{tildephi}. We thus introduce the connection one-forms on $Z_6$: \begin{equation}\label{AIscript} \mathscr{A}^{(I)} \,\equiv\, \left( \frac{x - \alpha}{x + 3K_I} - \frac{2}{3} \right) \frac{\Delta \phi}{2\pi} (\diff \varphi + \mathcal{A}_{\mathcal{J}}) \, \equiv \, \rho_I(x) (\diff \varphi + \mathcal{A}_{\mathcal{J}})\,, \quad I\, = \, 1,2,3\, , \end{equation} which by construction restricts to the supergravity gauge field $\tilde{A}^{(I)}$ on each $\Sigma$ fibre. We then compute the curvature \begin{equation} \mathscr{F}^{(I)}\, \equiv \, \diff \mathscr{A}^{(I)}\, = \, \rho_I'(x) \diff x \wedge (\diff \varphi + \mathcal{A}_{\mathcal{J}}) + \rho_I(x) \mathcal{F}_{\mathcal{J}}\,, \end{equation} where $\mathcal{F}_{\mathcal{J}} = \diff \mathcal{A}_{\mathcal{J}}$. By construction the integral of $\mathscr{F}^{(I)}$ over a fibre $\Sigma$ gives the flux $Q_I$, as in \eqref{fluxes}. The curvature form $\mathscr{F}^{(I)}$ defines a $U(1)$ bundle $\mathcal{L}_I$ over $Z_6$ by taking $c_1 (\mathcal{L}_I) = [\mathscr{F}^{(I)}/ 2\pi] \in H^2(Z_6,\R)$. We also define $c_1(\mathcal{J}) = [\mathcal{F}_{\mathcal{J}} / 2\pi] \in H^2(Z_4,\Z)$. In the anomaly polynomial we then write \begin{equation} c_1 (F^{(I)}) \, = \, \Delta_I\, c_1(R_{\mathrm{2d}}) + c_1 (\mathcal{L}_I)\, , \end{equation} where $R_{\mathrm{2d}}$ is the pull-back of a $U(1)$ bundle over $Z_4$. The trial R-charges $\Delta_I$ satisfy the constraint $\Delta_1+\Delta_2+\Delta_3=2$, which is equivalent to the superpotential (in $\mathcal{N}=1$ language) having R-charge 2. The $d=2$ anomaly polynomial is then obtained by integrating $\mathcal{A}_{\mathrm{4d}}$ in \eqref{anomaly4d} over $\Sigma$: \begin{equation} \begin{split} \label{anomalyintegral} \mathcal{A}_{\mathrm{2d}} & \, = \, \int_{\Sigma} \mathcal{A}_{\mathrm{4d}} \, = \, \frac{N^2}{2} \int_{\Sigma} c_1(F^{(1)}) c_1(F^{(2)}) c_1(F^{(3)})\, , \end{split} \end{equation} which reads \begin{equation}\label{A2dfirst} \begin{split} \mathcal{A}_{\mathrm{2d}} & \, = \, \frac{N^2}{2} \int_{\Sigma} \Big\{ \Delta_1 \Delta_2 \Delta_3 c_1(R_{\mathrm{2d}})^3 + c_1(R_{\mathrm{2d}})^2 \big[ \Delta_2 \Delta_3 c_1(\mathcal{L}_1) + \Delta_1 \Delta_3 c_1(\mathcal{L}_2) \\ & \qquad \qquad \quad + \Delta_1 \Delta_2 c_1(\mathcal{L}_3) \big] + c_1(R_{\mathrm{2d}}) \big[\Delta_3 c_1(\mathcal{L}_1) c_1(\mathcal{L}_2) + \Delta_2 c_1(\mathcal{L}_1) c_1(\mathcal{L}_3) \\ & \qquad \qquad \quad + \Delta_1 c_1(\mathcal{L}_2) c_1(\mathcal{L}_3) \big]+ c_1(\mathcal{L}_1) c_1(\mathcal{L}_2) c_1(\mathcal{L}_3) \Big\}\,. \end{split} \end{equation} After a computation using the explicit functions $\rho_I(x)$ defined in \eqref{AIscript}, we find the integral in \eqref{A2dfirst} gives \begin{align}\label{anomaly2dfinalfinal} \mathcal{A}_{\mathrm{2d}}& \, = \, \frac{N^2}{2} \left\{ \left[ \frac{q \Delta_2 \Delta_3}{n_- n_+} + \frac{p \Delta_1 \Delta_3}{n_- n_+} + \frac{n_- - n_+-p-q}{n_- n_+} \, \Delta_1\Delta_2\right] c_1(R_{\mathrm{2d}})^2 \right. \nonumber\\ [4pt] & \qquad \left.+\frac{\Delta_1 f_1(n_-,n_+,p,q)+\Delta_2 f_2(n_-,n_+,p,q)+\Delta_3 f_3(n_-,n_+,p,q)}{3 n_-^2 n_+^2 (n_- + n_+)}\, c_1(R_{\mathrm{2d}}) c_1(\mathcal{J}) \right. \nonumber\\ [4pt] & \qquad \left. + \frac{g_1(n_-,n_+,p,q) \, g_2(n_-,n_+,p,q)}{9 n_-^3 n_+^3 (n_- + n_+)^2}\, c_1(\mathcal{J})^2 \right\}\, , \end{align} where we have defined \begin{align} \begin{split} f_1(n_-,n_+,p,q) & \, \equiv \, -2 (n_-^2+n_+^2) (p+q)+(n_--n_+) (2 p^2+7 p q+4 q^2)\nonumber\\ &\qquad +2 n_- n_+ (-n_-+n_++2 p+3 q)-q (5 p^2+5 p q+2 q^2)\, , \end{split}\\[10pt] \begin{split} f_2(n_-,n_+,p,q) & \, \equiv \, -2 (n_-^2+n_+^2) (p+q)+(n_--n_+) (2 q^2+7 p q+4 p^2)\nonumber\\ &\qquad +2 n_- n_+ (-n_-+n_++2 q+3 p)-p (5 q^2+5 p q+2 p^2)\, , \end{split}\\[10pt] \begin{split} f_3(n_-,n_+,p,q) & \, \equiv \, -(n_--n_+) (2 p^2+p q+2 q^2)-2 n_- n_+ (p+q)\\ &\qquad +(2 p^2-p q+2 q^2) (p+q)\, , \end{split}\\[10pt] \begin{split} g_1(n_-,n_+,p,q) & \, \equiv \, -(n_--n_+) (4 p^2+13 p q+4 q^2)+4 (n_--n_+)^2 (p+q) \nonumber\\ & \qquad +4 n_- n_+ (n_--n_+)+9 p q (p+q)\, , \end{split}\\[10pt] g_2(n_-,n_+,p,q) & \, \equiv \, (n_- - n_+)(p+q) + n_- n_+ - (p^2 + p q + q^2)\, .\nonumber \end{align} \subsection{$c$-extremization and superconformal R-symmetry}\label{sec:extreme} The coefficient of $\frac{1}{2} c_1(L_i)c_1(L_j)$ in the anomaly polynomial $\mathcal{A}_{\mathrm{2d}}$ is precisely $\text{Tr}\, \gamma^3 \mathcal{Q}_i \mathcal{Q}_j$, where the global symmetry $\mathcal{Q}_i$ is associated to the $U(1)$ bundle $L_i$ over $Z_4$, and $\gamma^3$ is the $d=2$ chirality operator. We know from $c$-extremization \cite{Benini:2012cz} that the $d=2$ superconformal $U(1)$ R-symmetry extremizes the trial function \begin{equation}\label{ct} c_{\mathrm{trial}} \, =\, 3\, \text{Tr}\, \gamma^3 R_{\mathrm{trial}}^2\,, \end{equation} over the space of possible R-symmetries. We accordingly set \begin{equation}\label{Rtrial} R_{\mathrm{trial}} \, = \, R_{\mathrm{2d}} + \epsilon\, \mathcal{J}\,, \end{equation} so that \begin{equation} \label{ctriale} c_{\mathrm{trial}}\, = \, 3 \left( \text{Tr}\, \gamma^3 R_{\mathrm{2d}}^2 +2 \epsilon\, \text{Tr}\, \gamma^3 R_{\mathrm{2d}} \mathcal{J} + \epsilon^2\, \text{Tr}\, \gamma^3 \mathcal{J}^2 \right)\,. \end{equation} Next we can substitute for the traces using \eqref{anomaly2dfinalfinal}, and extremize over the trial R-charges $\epsilon$, $\Delta_I$, subject to the constraint $\Delta_1+\Delta_2+\Delta_3=2$. We find the extremal values of these parameters to be \begin{equation}\label{epsilonstar} \epsilon_* \, = \, \frac{ n_- n_+ (n_- + n_+)}{(n_--n_+) (p+q) + n_-n_+-p^2-p q-q^2}\, , \ \quad \Delta_1^* \, = \, \Delta_2^* \, = \, \Delta_3^* \, = \, \frac{2}{3} \,. \end{equation} The right-moving central charge, which is equal to the central charge $c$ in the large $N$ limit, is then given by evaluating \eqref{ct} on the superconformal R-symmetry with $\epsilon=\epsilon_*$ and $\Delta_I=\Delta_I^*$. We find \begin{equation} \label{cfield3charge} c \, = \, \frac{3 p q(n_- - n_+ - p - q)}{n_- n_+ [(n_--n_+) (p+q) + n_-n_+-p^2-p q-q^2]} N^2\,. \end{equation} This precisely matches the gravity computation \eqref{cgrav3charge}, and moreover the R-symmetry \eqref{Rtrial} with $\epsilon=\epsilon_*$ and $\Delta_I=\Delta_I^*$ precisely matches \eqref{mixing}! \subsection{M5-brane anomaly polynomial}\label{sec:anomM5} The large $N$ limit of anomaly polynomial for $N$ M5-branes is given by (see {\it e.g.} \cite{Hosseini:2020vgl}) \begin{align}\label{A6d} \mathcal{A}_{\mathrm{6d}}\, = \, \frac{1}{24}p_2(R)N^3\, . \end{align} Here $R$ denotes the $SO(5)_R$ symmetry of the M5-brane theory, which geometrically is identified with the normal bundle to the M5-brane in spacetime, and $p_2$ denotes the second Pontryagin class. The anomaly polynomial \eqref{A6d} is only valid to leading order in the large $N$ limit, and more generally receives $O(N)$ corrections involving also Pontryagin classes of the tangent bundle of the M5-brane. The supergravity solution in section \ref{sec:M5} has normal bundle twisted via the Cartan $U(1)\times U(1)\subset SO(5)_R$. Denoting these as two complex line bundles $\mathcal{N}_1$, $\mathcal{N}_2$, as in section \ref{sec:upliftM5}, we may write the Pontryagin class in terms of Chern classes so that \begin{align}\label{A6dagain} \mathcal{A}_{\mathrm{6d}}\, = \, \frac{1}{24}c_1(\mathcal{N}_1)^2c_1(\mathcal{N}_2)^2 N^3\, . \end{align} As described in section \ref{sec:upliftM5}, the M5-branes are wrapped on $\Sigma_g\times \Sigma$, with $\Sigma=\mathbb{WCP}^1_{[n_-,n_+]}$ a spindle. From the metric we saw that the normal bundles are \begin{align} \mathcal{N}_1\, = \, T^*\Sigma_g\otimes L_3\, , \qquad \mathcal{N}_2\, = \, {L}_1^2\, , \end{align} where ${L}_1\cong{L}_2$ because we have set the first two charges equal, $Q_1=Q_2$, and $T^*\Sigma_g$ is the cotangent bundle to the Riemann surface $\Sigma_g$, with total space being a Calabi-Yau two-fold. The corresponding first Chern class gives minus the Euler number of this Riemann surface: \begin{align} \int_{\Sigma_g}c_1(T^*\Sigma_g)\, = \, 2(g-1)\, . \end{align} Since $c_1(\mathcal{N}_1) = c_1(T^*\Sigma_g) + c_1(F^{(3)})$, we may first integrate the anomaly polynomial \eqref{A6dagain} over $\Sigma_g$ to obtain \begin{equation} \begin{split} \mathcal{A}_{\mathrm{4d}}& \, = \, \int_{\Sigma_g}\mathcal{A}_{\mathrm{6d}} \, = \, \frac{N^3}{24}\cdot 2 \left(\int_{\Sigma_g}c_1(T^*\Sigma_g)\right) c_1(F^{(3)})\left[2c_1(F^{(1)})\right]^2\, \\ & \, = \, c_1(F^{(3)})c_1(F^{(1)})^2 \cdot \frac{2}{3}(g-1) N^3\, . \end{split} \end{equation} We have thus effectively reduced to the four-dimensional anomaly polynomial \eqref{anomaly4d}, where we must take $F^{(1)}=F^{(2)}$ and the overall factor of $N^2/2$ should be replaced by $2(g-1)N^3/3$. The rest of the computation of the $d=2$ anomaly polynomial, obtained by integrating $\mathcal{A}_{\mathrm{4d}}$ over the spindle, and extracting the central charge, then goes through \emph{mutatis mutandis}. The final central charge is thus given by \eqref{cfield3charge} with $p=q$, and replacing an overall factor of $N^2/2$ by $2(g-1)N^3/3$. We obtain \begin{align} c \, = \, \frac{4(n_--n_+-2q)q^2}{n_-n_+[n_-(n_++2q) -q (2n_++3q)]} (g-1) N^3\, . \end{align} This perfectly matches the supergravity result \eqref{cM5gravity}. \section{Discussion}\label{sec:discuss} In this paper we have constructed a five-parameter family of AdS$_3\times Y_7$ solutions of type IIB supergravity, in which $Y_7$ is the total space of an $S^5$ fibration over a spindle $\Sigma=\mathbb{WCP}^1_{[n_-,n_+]}$. In addition to the coprime positive integers $n_->n_+$ specifying $\Sigma$, the solution is also characterized by positive integers $q,p$ that determine the twisting of the $S^5$ over $\Sigma$, together with the quantized five-form flux $N$ through the $S^5$. Setting $q=p=\frac{1}{3}(n_--n_+)$ recovers the solutions of \cite{Ferrero:2020laf}, and we have thus generalized those solutions via the addition of the twisting parameters $q, p$. As in \cite{Ferrero:2020laf} we interpret these as the near-horizon limit of $N$ D3-branes wrapping the spindle $\Sigma$, with $q$ and $p$ now determining the background $U(1)^3$ Abelian R-symmetry fluxes via \eqref{introfluxes}, \eqref{introcharges}. As a consistency check on this interpretation, we have reproduced the gravity formulas for the central charge and superconformal R-symmetry from a dual field theory computation, starting with the anomaly polynomial of $\mathcal{N}=4$ SYM and performing an appropriate twisted compactification on $\Sigma$. Similarly, we have constructed a four-parameter family of AdS$_3\times Y_8$ solutions of $d=11$ supergravity, which we interpret as the near-horizon limit of $N$ M5-branes wrapping $\Sigma_g\times \Sigma$, and we have reproduced the gravitational formula for the central charge from a dual field theory computation. One of the most interesting features of the solutions of \cite{Ferrero:2020laf}, that our solutions inherit, is that the Killing spinors on $\Sigma$ are not simply given by a topological twist. This is exemplified by the formula \eqref{introchargesum}, where the right hand side is not the Euler number $\chi(\Sigma)$ given in \eqref{introEuler}. An analogous family of $d=4$ AdS$_2\times \Sigma$ solutions was studied in more detail in \cite{Ferrero:2020twa}. In fact these are near-horizon limits of full $d=4$ accelerating extremal black hole solutions, where it is the acceleration parameter that effectively leads to the conical deficit singularities on the black hole horizon $\Sigma$. In this case we also have a UV geometry, on the conformal boundary of the black holes in AdS$_4$, and remarkably one finds that there is a topological twist on the copy of $\Sigma$ in the UV, but that it has effectively been cut in half along an equator by the acceleration horizon of the black hole! The spinor is a \emph{different} constant on the two halves of the spindle in the UV. The physical interpretation of this, in terms of wrapped branes, is still somewhat obscure, and it would be interesting to find an analogous class of black string solutions in $d=5$, $U(1)^3$ gauged supergravity that have our solutions as a near-horizon limit. Alternatively, one could directly attempt to study $\mathcal{N}=4$ SYM on a rigid supersymmetric background $\R^2\times \Sigma$, with background fluxes \eqref{introfluxes}, \eqref{introcharges}, making more explicit the twisting of the fields and their boundary conditions at the conical singularities of $\Sigma$. In particular, one might envisage that additional data needs to be specified at the conical singularities, in describing the behaviour of the fields. We note that the anomaly polynomial method gives the correct supergravity result, despite the presence of the conical deficit singularities on $\Sigma$, and it would be interesting to justify this more carefully in such an analysis. Finally, in section \ref{sec:M5} we uplifted the $d=5$, $Q_1=Q_2$ solutions on the Maldacena-N\'u\~nez solution, but we also commented that one can uplift on other internal six-manifolds $N_6$ to obtain solutions to $d=11$ supergravity. Such solutions \cite{Gaiotto:2009gz} describe M5-branes wrapped on a Riemann surface, in general with punctures, which are then further wrapped on the spindle $\Sigma$ to obtain $d=2$, $(0,2)$ SCFTs. It would be interesting to investigate these solutions in more detail, in particular computing the central charge in gravity and via the anomaly polynomial for the theory on $N$ M5-branes. We leave this, together with other interesting generalizations, for future work. \subsection*{Acknowledgments} We thank Dario Martelli for helpful discussions. The work of JFS was supported in part by STFC grant ST/T000864/1. \bibliographystyle{utphys}

9e7e8d6b632191ae8167f8680c1a1df87217f8ec

\section{Proofs} \subsection{Proof of \Cref{prop:cade}} \subsubsection{Any $\xi$-name of $\vec x\in\mathbb R^n$ is a Cauchy-name of $\vec x$} \begin{lemma}\label{lem:dist-cade} For any $\vec x\in\mathbb R^n$ and $\xi$-name $\varphi$ of $\vec x$, $\forall k:|\vec x-\varphi_k|\leq 2^{-k}$ holds. \end{lemma} \begin{proof} For simplicity we assume a dimension of $n=1$. The general case can be proved similarly. Let $x\in\mathbb R$ and $\varphi$ be a $\xi$-name of $x$ and let $k\in\omega$. By construction $x=\lim\varphi$, hence there is $n_0\in\omega$ such that for every $n\geq n_0$ the bound $|\varphi_n-x|<2^{-k-1}$ holds. If $n_0\leq k$, the previous bound already gives the required property. Otherwise $n_0>k$, then $|\varphi_k-x|\leq|\varphi_k-\varphi_{n_0}|+|\varphi_{n_0}-x|$ holds. Since $\varphi\in\dom\xi$, the first summand is bounded by $2^{-\min(k,n_0)-1}=2^{-(k+1)}$. By the property above, so is the second. Ergo $|\varphi_k-x|\leq2^{-k}$. \end{proof} The property that $\xi$-names are Cauchy-names follows directly from \Cref{lem:dist-cade}. \subsubsection{$\xi$ is computably equivalent to the Cauchy representation} \begin{proof} For simplicity we assume a dimension of $1$. The general case can be proved similarly. \begin{itemize} \item[$\Rightarrow$)] Let $\psi$ be a $\xi$-name of $x\in\mathbb R$. By \Cref{it:xi2rho-cade} of \Cref{prop:cade}, $\psi$ is a name of $x$. \item[$\Leftarrow$)] Given $\varphi\in\dom\rho$ and $n\in\omega$. Compute $\psi_n\coloneqq\round{\varphi_{n+4}\cdot2^{n+1}}/2^{n+1}\in\mathbb D_n$ where $\round\cdot:\mathbb Q\to\mathbb Z$ is a computable rounding operation with $|\round q-q|\leq1/2$. Then with $x\coloneqq\lim\varphi$: \begin{align*} |\psi_n-x|&=|\round{\varphi_{n+4}\cdot2^{n+1}}/2^{n+1}-x| \\ &\leq|\round{\varphi_{n+4}\cdot2^{n+1}}-\varphi_{n+4}\cdot2^{n+1}|/2^{n+1}+ |\varphi_{n+4}-x| \\ &\leq2^{-(n+2)}+2^{-(n+4)} \end{align*} We show $\psi\coloneqq(\psi_n)_n$ is a $\xi$-name of $x$. Let $n,k\in\omega$ with $k>0$. \begin{align*} |\psi_n-\psi_{n+k}| &\leq|\psi_n-x|+|\psi_{n+k}-x| \\ &\leq2^{-(n+2)}+2^{-(n+4)}+2^{-(n+k+2)}+2^{-(n+k+4)} \\ &\leq2^{-(n+2)}+2^{-(n+4)}+2^{-(n+3)}+2^{-(n+5)} \\ &\leq2^{-(n+1)} \end{align*} Thus, $\psi\in\dom\xi$ and therefore $\psi$ is a $\xi$-name of $\lim\psi=x$. \end{itemize} \end{proof} \subsubsection{$\xi$ is continuous} \begin{proof} Computable equivalence between two representations implies there are continuous maps between them. Since the Cauchy representation is continuous itself~\cite{DBLP:journals/tcs/BrattkaH02}, so is $\xi$. \end{proof} \subsection{Proof of \Cref{lem:local-lin''}}\label{prf:local-lin''} \paragraph{\upshape\textbf{\Cref{lem:local-lin''}.}} \emph{% Let $P:f(\vec x)\diamond0$ be a non-linear constraint in $\mathcal N$ and $\alpha$ be an assignment of $\vec x$ to rationals in $\dom f$. Whenever $C={}$\Call{LineariseLocal$_\delta$}{$f,\vec x,\diamond,\alpha$} and $C\neq\None$, $C$ is an \epsfull linearisation of $P$ at $\alpha$, with $\epsilon$ corresponding to the effective local modulus of continuity induced by $M_f^?$ at a $\xi$-name of $\interp{\vec x}^\alpha$. } \begin{proof} Let $P:f(\vec x)\diamond 0$, $\alpha$ and $C\neq\None$ be as in the premise and let $p,\tilde y$ and $\varphi$ as in \Cref{alg:2}. Since $C\neq\None$ by construction $C=(\vec x\notin B(\interp{\vec x}^\alpha,2^{-k}))$ where $k=\gamma_{M^?_f,\varphi}(p)$ is the maximum query $M_f^\varphi(p+2)$ poses to its oracle. Thus, by definition, $C$ is \epsfull with $\epsilon=2^{-k}$. In order to show that $C$ indeed is a linearisation of $P$ at $\alpha$, let $\vec z\in B(\interp{\vec x}^\alpha,2^{-k})\cap\widebar{D_P}\subseteq\dom f$. As $\gamma_{M^?_f,\varphi}$ is a local modulus of continuity of $f$ at $\interp{\vec x}^\alpha$~\cite[Theorem~2.13]{DBLP:books/daglib/0067010}, $f(\vec z)$ is within distance $2^{-p}\leq\delta/4$ of $f(\interp{\vec x}^\alpha)$, which, by definition of $M_f^?$, is at most $2^{-(p+2)}<\delta/4$ away from $\tilde y$. By construction of $C$ in \Cref{alg:2} the property $\neg(\tilde y\diamond-\delta/2)$ holds. As in case~\ref{th:int:lin:unsat} in the proof of \Cref{th:int:lin-correct-cade}, this property implies $\neg(f(\vec z)\diamond 0)$. Therefore, according to \Cref{def:eps-full}, $C$ is an \epsfull linearisation of $P$ at $\alpha$. \end{proof} \subsection{Proof of \Cref{prop:names-compact-cade}}\label{prf:names-compact-cade} The proof of the following lemma follows that of~\cite[Theorem 7.2.5.2]{DBLP:series/txtcs/Weihrauch00}. \begin{lemma}\label{prop:names-closed-cade} Let $X\subset\mathbb R^n$ be closed. Then the set $\xi^{-1}(X)\subset\mathbb D_\omega^n$ of $\xi$-names of elements in $X$ is closed as well. \end{lemma} \begin{proof} Again, for sake of simplicity we assume $n=1$ while the general case be proved in a similar manner. We first introduce the notation $\mathbb D_*$ for the set of finite prefixes of elements in $\mathbb D_\omega$ and for any prefix $u\in\mathbb D_*$ let $\mathbb D_\omega[u]\coloneqq\{u\varphi\mid \varphi\in\omega^\omega,u\varphi\in\mathbb D_\omega\}$ denote `the ball' around $u$ in $\mathbb D_\omega$. Observe that $\mathbb D_\omega[u]$ is a basic open set in $\mathbb D_\omega$ for any prefix $u\in\mathbb D_*$. In order to show $\xi^{-1}(X)$ is closed, we prove that its complement is open. Since $X$ is closed there is a \KK{$\mathcal B$ not needed \fb y}% collection $\mathcal B$ of open subsets of $\mathbb R$ such that $\bigcup\mathcal B=\mathbb R\setminus X$. Then \[ C\coloneqq\{u\in\omega^*\mid\exists B\in\mathcal B~\text{s.t.}~ \xi(\mathbb D_\omega[u])\subseteq B\} \] is a subset of $\omega^*$. Define \begin{align*} U_1&\coloneqq\bigcup_{u\in C}\mathbb D_\omega[u] \\ U_2&\coloneqq\{(p_n)_n\in\mathbb D_\omega\mid \exists i,j~\text{s.t.}~|p_i-p_{i+j}|>2^{-(i+1)}\} \end{align*} and $U\coloneqq U_1\cup U_2$. By definition, $U_1$ is open in $\mathbb D_\omega$. To see that $U_2$ is open in $\mathbb D_\omega$ as well, observe that \[ U_2=\mathbb D_\omega\setminus\dom\xi=\bigcup_{w\in C'}\mathbb D_\omega[w] \] where $w\in C'\subseteq\mathbb D_*$ iff there are $i,j\in\omega$ such that $|w_i-w_{i+j}|>2^{-(i+1)}$. We show $\xi^{-1}(X)=\mathbb D_\omega\setminus U$. \begin{itemize} \item Assume $p\in\xi^{-1}(X)$, that is, $\xi(p)\in X$. \begin{itemize} \item Then $p=(p_k)_k$ with $p_k\cdot2^{k+1}\in\mathbb Z$ for every $k$ such that $\forall i,j:|p_i-p_{i+j}|\leq2^{-(i+1)}$, thus, $p\notin U_2$. \item Suppose $p\in U_1$. Then there is $u\in C$ and $B\in\mathcal B$ such that $p\in\mathbb D_\omega[u]$ and $\xi(\mathbb D_\omega[u])\subseteq B$ or $\mathbb D_\omega[u]\cap\dom\xi=\varnothing$. In both cases $p\notin\xi^{-1}(X)$, a contradiction. \end{itemize} From $p\notin U_1$ and $p\notin U_2$ we obtain $p\notin U$. \item Assume $p\notin\xi^{-1}(X)$. \begin{itemize} \item First, consider $p\notin\dom\xi$. Then $p\in U_2$. \item Finally, consider $p\in\dom\xi$. Since $\xi(p)\notin X$, there are \KK{elaborate: whole prefix will be outside \fb y}% some prefix $u$ of $p$ and some $B\in\mathcal B$ such that $\xi(\mathbb D_\omega[u])\subseteq B$, and so $p\in U_1$. \end{itemize} From $p\in U_1$ or $p\in U_2$ we obtain $p\in U$. \end{itemize} Therefore, $U$ is the open complement of $\xi^{-1}(X)$, which is a closed subset of $\mathbb D_\omega$. \end{proof} Now, the proof of \Cref{prop:names-compact-cade} follows from Tychonoff's theorem, which states that arbitrary products of non-empty compact spaces again are compact~\cite{Tychonoff1930}. \paragraph{\upshape\textbf{\Cref{prop:names-compact-cade}.}} \emph{% Let $X\subset\mathbb R^n$ be compact. Then the set $\xi^{-1}(X)\subset\mathbb D_\omega^n$ of $\xi$-names of elements in $X$ is compact as well. } \begin{proof} By \Cref{lem:dist-cade}, $\xi^{-1}(X)$ is a subset of the product of the finite and therefore compact spaces \[ \{\vec y\in\mathbb D_k^n:\Vert\vec x-\vec y\Vert\leq2^{-k},\vec x\in X\} \] over $k\in\omega$. As a closed (\Cref{prop:names-closed-cade}) subset of a compact space, $\xi^{-1}(X)$ is compact as well. \end{proof} \section{Introduction}\label{sec:intro} \strut\KK{Spell check final version!}% \todo{All references to appendices need to be replaced. a) Upcoming journal version, b) TR on arXiv, c) just drop \fb b \kk b \mk b Remember to eliminate all words ``appendix''.}% Solving non-linear constraints is important in many applications, including verification of cyber-physical system , software verificatio , proof assistants for mathematic ~\cite{DBLP:books/sp/Platzer18,DBLP:conf/formats/KuratkoR14,https://doi.org/10.1002/rnc.2914,DBLP:conf/fm/BardBD19,DBLP:journals/corr/HalesABDHHKMMNNNOPRSTTTUVZ15,DBLP:conf/fmcad/BrausseKK20}. Hence there has been a number of approaches for solving non-linear constraints, involving symbolic methods~\cite{DBLP:journals/cca/JovanovicM12,DBLP:conf/cade/MouraP13,DBLP:journals/fmsd/TiwariL16,DBLP:conf/casc/KorovinKS14} as well as numerically inspired ones, in particular for dealing with transcendental functions~\cite{DBLP:conf/cade/GaoAC12,DBLP:conf/cade/TungKO16}, and combinations of symbolic and numeric methods~\cite{DBLP:conf/frocos/BrausseKKM19,DBLP:journals/tocl/CimattiGIRS18,DBLP:conf/frocos/FontaineOSV17}. In \cite{DBLP:conf/frocos/BrausseKKM19} we introduced the \ksmt calculus for solving non-linear constraints over a large class of functions including polynomial, exponential and trigonometric functions. The \ksmt calculus% \footnote{Implementation is available at \url{http://informatik.uni-trier.de/~brausse/ksmt/}} combines CDCL-style reasoning~\cite{DBLP:conf/iccad/SilvaS96,DBLP:conf/vmcai/MouraJ13,DBLP:journals/jar/BonacinaGS20} over reals based on conflict resolution~\cite{CR:KTV09} with incremental linearisations of non-linear functions using methods from computable analysis~\cite{DBLP:series/txtcs/Weihrauch00,DBLP:conf/cca/Muller00}. Our approach is based on computable analysis and exact real arithmetic which avoids limitations of double precision computations caused by rounding errors and instabilities in numerical methods. In particular, satisfiable and unsatisfiable results returned by \ksmt are exact as required in many applications. This approach also supports implicit representations of functions as solutions of ODEs and PDEs~\cite{DBLP:books/daglib/0087495}. It is well known that in the presence of transcendental functions the constraint satisfiability problem is undecidable~\cite{DBLP:journals/jsyml/Richardson68}. However if we only require solutions up to some specified precision $\delta$, then the problem can be solved algorithmically on bounded instances and that is the motivation behind $\delta$-completeness, which was introduced in~\cite{DBLP:conf/cade/GaoAC12}. In essence a $\delta$-complete procedure decides if a formula is unsatisfiable or a $\delta$ weakening of the formula is satisfiable. In this paper we investigate theoretical properties of the \ksmt calculus, and its extension $\delta$-\ksmt for the $\delta$-SMT setting. Our main results are as follows: \begin{enumerate} \item We introduced a notion of \emph{\epsfull{} linearisations} and prove that all \epsfull{} runs of \ksmt are terminating on bounded instances. \item We extended the $\ksmt$ calculus to the $\delta$-satisfiability setting and proved that $\delta$-\ksmt is a \emph{$\delta$-complete decision procedure} for bounded instances. \item We introduced an algorithm for computing \epsfull{} \emph{local linearisations} and integrated it into $\delta$-\ksmt. Local linearisations can be used to considerably narrow the search space by taking into account local behaviour of non-linear functions avoiding computationally expensive global analysis. \end{enumerate} In \Cref{sec:ksmt-overview}, we give an overview about the \ksmt calculus and introduce the notion of \epsfull linearisation used throughout the rest of the paper. We also present a completeness theorem. \Cref{sec:delta-sat} introduces the notion of $\delta$-completeness and related concepts. In \Cref{sec:delta-ksmt-cade} we introduce the $\delta$-\ksmt adaptation, prove it is correct and $\delta$-complete, and give concrete effective linearisations based on a uniform modulus of continuity. Finally in \Cref{sec:eps-from-M-cade}, we introduce local linearisations and show that termination is independent of computing uniform moduli of continuity, before we conclude in \Cref{sec:concl}. \section{Preliminaries}\label{sec:prelim} \irrelcade{ \todo[inline]{NM, KK: Reframe notions to intervals which might be easier to digest for target audience. \fb n} } The following conventions are used throughout this paper. By $\Vert\cdot\Vert$ we denote the maximum-norm $\Vert (x_1,x_2,\ldots,x_n)\Vert=\max\{|x_i|:1\leq i\leq n\}$. When it helps clarity, we write finite and infinite sequences $\vec x=(x_1,\ldots,x_n)$ and $\vec y=(y_i)_i$ in bold typeface. We are going to use open balls $B(\vec c,\epsilon)=\{\vec x:\Vert\vec x-\vec c\Vert<\epsilon\}\subseteq\mathbb R^n$ for $\vec c\in\mathbb R^n$ and $\eps>0$ and $\widebar A$ to denote the closure of the set $A\subseteq\mathbb R^n$ in the standard topology induced by the norm. By $\mathbb Q_{>0}$ we denote the set $\{q\in\mathbb Q:q>0\}$. For sets $X,Y$, a (possibly partial) function from $X$ to $Y$ is written as $X\to Y$. We use the notion of compactness: a set $A$ is compact iff every open cover of $A$ has a finite subcover. \todofb{Note: We use this for $\widebar{D_P}$ in the proof of \Cref{th:int:lin:alt-cade}.}% In Euclidean spaces this is equivalent to $A$ being bounded and closed~\cite{Willard70}. \subsection*{Basic notions of Computable Analysis} \label{sec:ca} Let us recall the notion of computability of functions over real numbers used throughout this paper. A rational number $q$ is an {\it $n$-approximation} of a real number $x$ if $\Vert q-x\Vert\leq2^{-n}$. Informally, a function $f$ is {\it computed} by a function-oracle Turing machine $M_f^?$, where $^?$ is a placeholder for the oracle representing the argument of the function, in the following way. The real argument $x$ is represented by an oracle function $\varphi:\mathbb N\to\mathbb Q$, for each $n$ returning an $n$-approximation $\varphi_n$ of $x$. For simplicity, we refer to $\varphi$ by the sequence $(\varphi_n)_n$. When run with argument $p\in\mathbb N$, $M_f^\varphi(p)$ computes a rational $p$-approximation of $f(x)$ by querying its oracle $\varphi$ for approximations of $x$. Let us note that the definition of the oracle machine does not depend on the concrete oracle, i.e., the oracle can be seen as a parameter. In case only the machine without a concrete oracle is of interest, we write $M_f^?$. We refer to~\cite{DBLP:books/daglib/0067010} for a precise definition of the model of computation by function-oracle Turing machines which is standard in computable analysis. \begin{definition}[\cite{DBLP:books/daglib/0067010}]\label{def:computable} Consider $\vec x \in \mathbb R^n$. A \emph{name} for $\vec x$ is a rational sequence $\vec\varphi=(\vec\varphi_k)_k$ such that $\forall k:\Vert\vec\varphi_k-\vec x\Vert\leq2^{-k}$. A function $f:\mathbb R^n\to\mathbb R$ is \emph{computable} iff there is a function-oracle Turing machine $M_f^?$ such that for all $\vec x\in\dom f$ and names $\vec\varphi$ for $\vec x$, $|M_f^{\vec\varphi}(p)-f(\vec x)|\leq2^{-p}$ holds for all $p\in\mathbb N$. \end{definition} This definition is closely related to interval arithmetic with unrestricted precision, but enhanced with the guarantee of convergence and it is equivalent to the notion of computability used in~\cite{DBLP:series/txtcs/Weihrauch00}. The class of computable functions contains polynomials and transcendental functions like $\sin$, $\cos$, $\exp$, among others. It is well known \cite{DBLP:books/daglib/0067010,DBLP:series/txtcs/Weihrauch00} that this class is closed under composition and that computable functions are continuous. By continuity, a computable function $f:\mathbb R^n\to\mathbb R$ total on a compact $D\subset\mathbb R^n$ has a computable \emph{uniform modulus of continuity} $\mu_f:\mathbb N\to\mathbb N$ on $D$~\cite[Theorem 6.2.7]{DBLP:series/txtcs/Weihrauch00}, that is, \begin{align} \forall k\in\mathbb N\,\forall\vec y,\vec z\in D: \Vert\vec y-\vec z\Vert\leq2^{-\mu(k)}\implies|f(\vec y)-f(\vec z)|\leq2^{-k} \text. \label{eq:mu} \end{align} A uniform modulus of continuity of $f$ expresses how changes in the value of $f$ depend on changes of the arguments in a uniform way. \section{The \ksmt calculus}\label{sec:ksmt-overview} We first describe the \ksmt calculus for solving non-linear constraints~\cite{DBLP:conf/frocos/BrausseKKM19} informally, and subsequently recall the main definitions which we use in this paper. The \ksmt calculus consists of transition rules, which, for any formula in linear separated form, allow deriving lemmas consistent with the formula and, in case of termination, produce a satisfying assignment for the formula or show that it is unsatisfiable. A quantifier-free formula is in separated linear form $\mathcal L\cup\mathcal N$ if $\mathcal L$ is a set of clauses over linear constraints and $\mathcal N$ is a set of non-linear atomic constraints; this notion is rigorously defined below. In the \ksmt calculus there are four transition rules applied to its states: Assignment refinement $(A)$, Conflict resolution $(R)$, Backjumping $(B)$ and Linearisation $(L)$. The final \ksmt states are \SAT and \UNSAT. A non-final \ksmt state is a triple $(\alpha,\mathcal L,\mathcal N)$ where $\alpha$ is a (partial) assignment of variables to rationals. A \ksmt derivation starts with an initial state where $\alpha$ is empty and tries to extend this assignment to a solution of $\mathcal L \cup \mathcal N$ by repeatedly applying the Assignment refinement rule. When such assignment extension is not possible we either obtain a linear conflict which is resolved using the conflict resolution rule, or a non-linear conflict which is resolved using the linearisation rule. The main idea behind the linearisation rule is to approximate the non-linear constraints around the conflict using linear constraints in such a way that the conflict will be shifted into the linear part where it will be resolved using conflict resolution. Applying either of these two rules results in a state containing a clause evaluating to \False under the current assignment. They either result in application of the backjumping rule, which undoes assignments or in termination in case the formula is \UNSAT. In this procedure, only the assignment and linear part of the state change and the non-linear part stays fixed. \begin{figure}[tp] \centering\begin{tikzpicture}[node distance=6em,every node/.style={scale=0.74},align=center] \matrix[row sep=1em,column sep=2em] { & \node (input) [input] {\clap{separated linear form}}; \\ \\ & \node (lin-check) [draw,thick,star,star points=9,star point ratio=0.87,inner sep=.5ex,fill=yellow!40] {lin. \\ check}; \\ \node (choice) [draw,thick,diamond,inner sep=.5ex,fill=green!30] {choice}; & & \node (A) [process,thick] {A}; & & \node (B) [process,very thick,draw=red] {B}; \\ & & \node (ex-z) [draw,red,text=black,very thick,star,star points=9,star point ratio=0.87,inner sep=.5ex,fill=yellow!40] {$\exists z$}; & \node (R) [process,thick] {R}; \\ & \node (nlin-check) [draw,star,thick,star points=9,star point ratio=0.87,inner sep=.5ex,fill=yellow!40] {nlin. \\ check}; & & \node (L) [process,thick] {L}; \\ }; \draw[->] (input) -- node [right] {$\alpha=\nil$} (lin-check); \draw[->] (lin-check) edge [in=90,out=180] node [above left] {p.lin.cons.} (choice); \draw[->] (lin-check) edge [in=90,out=0] node [above right] {p.lin.incons.} (B); \draw[->] (choice) edge [out=-45,in=180] (ex-z); \draw[->] (choice) edge [out=270,in=180] (nlin-check); \draw[->] (A) edge [bend right=10] (choice.east); \draw[->] (B) edge [bend right=20] (choice); \draw[->] (ex-z) -- node [left,midway,yshift=-1ex] {$\exists q$} (A); \draw[->] (ex-z) -- node [above,near start] {$\neg\exists q$} (R); \draw[->] (R) edge [bend right=20] (B); \draw[->] (nlin-check) edge [bend right=15] node [right,near start,yshift=-.5ex] {p.nlin.cons.} (ex-z); \draw[->] (nlin-check) edge [bend right=15] node [below] {p.nlin.incons.} (L); \draw[->] (L) edge [bend right] (B); \end{tikzpicture} \caption{Core of \ksmt calculus. Derivations terminate in red nodes.} \label{fig:white-box} \end{figure} \paragraph{Notations.} Let $\mathcal F_{\mathrm{lin}}$ consist of rational constants, addition and multiplication by rational constants; $\mathcal F_{\mathrm{nl}}$ denotes an arbitrary collection of non-linear computable functions including transcendental functions and polynomials over the reals. We consider the structure $(\mathbb R,\langle\mathcal F_{\mathrm{lin}}\cup\mathcal F_{\mathrm{nl}},\mathcal P\rangle)$ where $\mathcal P=\{{<},{\leq},{>},{\geq},{=},{\neq}\}$ and a set of variables $V=\{x_1,x_2,\ldots,x_n,\ldots\}$. We will use, possibly with indices, $x$ to denote variables and $q,c,e$ for rational constants. Define terms, predicates and formulas over $V$ in the standard way. An \emph{atomic linear constraint} is a formula of the form: $q+c_1x_1+\ldots+c_nx_n \diamond 0$ where $q,c_1,\ldots,c_n\in\mathbb Q$ and $\diamond\in\mathcal P$. Negations of atomic formulas can be eliminated by rewriting the predicate symbol $\diamond$ in the standard way, hence we assume that all literals are positive. A \emph{linear constraint} is a disjunction of atomic linear constraints, also called \emph{(linear) clause}. An \emph{atomic non-linear} constraint is a formula of the form $f(\vec x)\diamond 0$, where $\diamond\in\mathcal P$ and $f$ is a composition of computable non-linear functions from $\mathcal F_{\mathrm{nl}}$ over variables $\vec x$. Throughout this paper for every computable real function $f$ we use $M_f^?$ to denote a function-oracle Turing machine computing $f$. We assume quantifier-free formulas in \emph{separated linear form}~\cite[Definition~1]{DBLP:conf/frocos/BrausseKKM19}, that is, $\mathcal L\cup\mathcal N$ where $\mathcal L$ is a set of linear constraints and $\mathcal N$ is a set of non-linear atomic constraints. Arbitrary quantifier-free formulas can be transformed equi-satisfiably into separated linear form in polynomial time~\cite[Lemma~1]{DBLP:conf/frocos/BrausseKKM19}. Since in separated linear form all non-linear constraints are atomic we will call them just \emph{non-linear constraints}. Let $\alpha:V\to\mathbb Q$ be a partial variable assignment. The interpretation $\interp{\vec x}^\alpha$ of a vector of variables $\vec x$ under $\alpha$ is defined in a standard way as component-wise application of $\alpha$. Define the notation $\interp{t}^\alpha$ as evaluation of term $t$ under assignment $\alpha$, that can be partial, in which case $\interp{t}^\alpha$ is treated symbolically. We extend $\interp\cdot^\alpha$ to predicates, clauses and CNF in the usual way and $\True,\False$ denote the constants of the Boolean domain. The evaluation $\interp{t\diamond 0}^\alpha$ for a predicate $\diamond$ and a term $t$ results in $\True$ or $\False$ only if all variables in $t$ are assigned by $\alpha$. In order to formally restate the calculus, the notions of linear resolvent and linearisation are essential. A resolvent $R_{\alpha,\mathcal L,z}$ on a variable $z$ is a set of linear constraints that do not contain $z$, are implied by the formula $\mathcal L$ and which evaluate to $\False$ under the current partial assignment $\alpha$; for more details see~\cite{CR:KTV09,DBLP:conf/frocos/BrausseKKM19}. \begin{definition}\label{def:linearisation} Let $P$ be a non-linear constraint and let $\alpha$ be an assignment with $\interp P^\alpha=\False$. A \emph{linearisation of $P$ at $\alpha$} is a linear clause $C$ with the properties: \begin{enumerate} \item\label{def:linearisation:it1} $\forall\beta:\interp P^\beta=\True\implies\interp C^\beta=\True$, and \item\label{def:linearisation:it2} $\interp C^\alpha=\False$. \end{enumerate} \end{definition} Wlog.\ we can assume that the variables of $C$ are a subset of the variables of $P$. Let us note that any linear clause $C$ represents the complement of a rational polytope $R$ and we will use both interchangeably. Thus for a rational polytope $R$, $\vec{x}\not \in R$ also stands for a linear clause. In particular, any linearisation excludes a rational polytope containing the conflicting assignment from the search space. \paragraph{Transition rules.} For a formula $\mathcal L_0\cup\mathcal N$ in separated linear form, the initial \ksmt state is $(\nil,\mathcal L_0,\mathcal N)$. The calculus consists of the following transition rules from a state $S=(\alpha,\mathcal L,\mathcal N)$ to $S'$: \begin{description} \item[$(A)$]\label{rule:A} \emph{Assignment.} $S'=(\alpha::z\mapsto q,\mathcal L,\mathcal N)$ iff $\interp{\mathcal L}^\alpha\neq\False$ and there is a variable $z$ unassigned in $\alpha$ and $q\in\mathbb Q$ with $\interp{\mathcal L}^{\alpha::z\mapsto q}\neq\False$. \item[$(R)$]\label{rule:R'} \emph{Resolution.} $S'=(\alpha,\mathcal L\cup R_{\alpha,\mathcal L,z},\mathcal N)$ iff $\interp{\mathcal L}^\alpha\neq\False$ and there is a variable $z$ unassigned in $\alpha$ with $\forall q\in\mathbb Q:\interp{\mathcal L}^{\alpha::z\mapsto q}=\False$ and $R_{\alpha,\mathcal L,z}$ is a resolvent. \item[$(B)$]\label{rule:B} \emph{Backjump.} $S'=(\gamma,\mathcal L,\mathcal N)$ iff $\interp{\mathcal L}^\alpha=\False$ and there is a maximal prefix $\gamma$ of $\alpha$ such that $\interp{\mathcal L}^\gamma\neq\False$. \item[$(L)$]\label{rule:L} \emph{Linearisation.} $S'=(\alpha,\mathcal L\cup \{L_{\alpha, P}\},\mathcal N)$ iff $\interp{\mathcal L}^\alpha\neq\False$, there is $P$ in $\mathcal N$ with $\interp P^\alpha=\False$ and there is a linearisation $L_{\alpha,P}$ of $P$ at $\alpha$. \item[\Fsat] \emph{Final \SAT.} \todo{prefer $\SAT_\alpha$ \kk y \mk y}% $S'=\SAT$ if all variables are assigned in $\alpha$, $\interp{\mathcal L}^\alpha=\True$ and none of the rules $(A),(R),(B),(L)$ is applicable. \item[\Funsat] \emph{Final \UNSAT.} $S'=\UNSAT$ if $\interp{\mathcal L}^\nil=\False$. In other words a trivial contradiction, e.g., $0>1$ is in $\mathcal L$. \end{description} A path (or a run) is a derivation in a \ksmt. A procedure is an effective (possibly non-deterministic) way to construct a path. \paragraph{Termination.}\label{par:termination} If no transition rule is applicable, the derivation terminates. For clarity, we added the explicit rules $\Fsat$ and $\Funsat$ which lead to the final states. This calculus is sound \cite[Lemma~2]{DBLP:conf/frocos/BrausseKKM19}: if the final transition is $\Fsat$, then $\alpha$ is a solution to the original formula, or $\Funsat$, then a trivial contradiction $0>1$ was derived and the original formula is unsatisfiable. The calculus also makes progress by reducing the search space~\cite[Lemma~3]{DBLP:conf/frocos/BrausseKKM19}. \begin{figure \centering \small \begin{minipage}{.47\linewidth} \[ \mathcal C= \underbrace{\textcolor{red}{(y\leq 1/x)}}_P{} \begin{aligned}[t] &{}\land{}\textcolor{green!40!black}{(x/4+1\leq y)} \\ &{}\land\textcolor{green!40!black}{(y\leq 4\cdot(x-1))} \\ &{{}\land\big((x\leq\tfrac{12}{19})\lor(y\leq\tfrac{19}{12})\big)} \\ &{{}\land\textcolor{blue}{\big((x\leq\tfrac{220}{223})\lor(y\leq\tfrac{223}{220})\big)}} \\ &{{}\land\textcolor{purple}{(\tfrac43\leq x)\land(x\leq\tfrac{220}{223})}} \\ &{{}\land\textcolor{cyan!50!black}{(\tfrac43\leq\tfrac{220}{223})}} \end{aligned} \] \centering \includegraphics[width=.8\linewidth]{pictures/unsat1-4.pdf} \end{minipage}% \begin{minipage}{.53\linewidth}\small Linearisation of $P$ on conflicts $(x,y)$ at $\alpha$ here: \begin{itemize} \vspace{-.5\baselineskip} \item choose $d\coloneqq(1/\interp x^\alpha+\interp y^\alpha)/2$, \item $C=\big(x\leq 1/d\;\lor\; y\leq d\big)$ \end{itemize} \centering \begin{tabular}{c|l|c}\removelastskip rule & $\alpha$ & note \\ \hline $(A)$ & $x\mapsto 2$ & \\ $(A)$ & $x\mapsto 2$, $y\mapsto \tfrac83$ & (3a) \\ $(L)$ & $x\mapsto 2$, $y\mapsto \tfrac83$ & (3b) \\ $(B)$ & $x\mapsto 2$ & \\ $(A)$ & $x\mapsto 2$, $y\mapsto\tfrac{84}{55}$ & (4a) \\ $(L)$ & $x\mapsto 2$, $y\mapsto\tfrac{84}{55}$ & (4b) \\ $(B)$ & $x\mapsto 2$ & \\ $(R)$ & $x\mapsto 2$ & \textcolor{green!40!black}{on} \textcolor{blue}{$y$} \\ $(B)$ & & \\ $(R)$ & & \textcolor{purple}{on $x$} \\ \Funsat && \texttt{unsat} \end{tabular} \end{minipage} \caption{\texttt{unsat} example run of \ksmt using interval linearisation~\cite{DBLP:conf/frocos/BrausseKKM19}.} \label{fig:lin-exampe-unsat-run} \end{figure} An example run of the \ksmt calculus is presented in \Cref{fig:lin-exampe-unsat-run}. We start in a state with a non-linear part $\mathcal N=\{y\leq 1/x\}$, which defines the pink area and the linear part $\mathcal L=\{ (x/4+1\leq y), (y\leq4\cdot(x-1))\}$, shaded in green. Then we successively apply \ksmt rules excluding regions around candidate solutions by linearisations, until we derive linearisations which separates the pink area from the green area thus deriving a contradiction. \begin{remark}\label{rem:lin-infty-often} In general a derivation may not terminate. The only cause of non-termination is the linearisation rule which adds new linear constraints and can be applied infinitely many times. To see this, observe that \ksmt with only the rules $(A),(R),(B)$ corresponds to the conflict resolution calculus which is known to be terminating~\cite{CR:KTV09,DBLP:conf/cade/KorovinV11}. Thus, in infinite \ksmt runs the linearisation rule $(L)$ is applied infinitely often. This argument is used in the proof of \Cref{th:bounded} below. Let us note that during a run the \ksmt calculus neither conflicts nor lemmas can be generated more than once. In fact, any generated linearisation is not implied by the linear part, prior to adding this linearisation. \end{remark} \subsection{Sufficient termination conditions} In this section we will assume that $(\alpha,\mathcal L,\mathcal N)$ is a \ksmt state obtained by applying \ksmt inference rules to an initial state. As in \cite{DBLP:conf/cade/GaoAC12} we only consider bounded instances. In many applications this is a natural assumption as variables usually range within some (possibly large) bounds. \irrelcade{\todofb{remove sentence, implied by `bounded set'}}% We can assume that these bounds are made explicit as linear constraints in the system. \begin{definition} Let $F$ be the formula $\mathcal L_0\land\mathcal N$ in separated linear form over variables $x_1,\ldots,x_n$ and let $B_i$ be the set defined by the conjunction of all clauses in $\mathcal L_0$ univariate in $x_i$, for $i=1,\ldots,n$; in particular, if there are no univariate linear constraints over $x_i$ then $B_i=\mathbb R$. We call $F$ a \emph{bounded instance} if: \begin{itemize} \item $D_F\coloneqq\bigtimes_{i=1}^n B_i$ is bounded, and \item for each non-linear constraint $P:f(x_{i_1},\ldots,x_{i_k})\diamond 0$ in $\mathcal N$ with $i_j\in\{1,\ldots,n\}$ for $j\in\{1,\ldots,k\}$ it holds that $\widebar{D_P}\subseteq\dom f$ where $D_P\coloneqq\bigtimes_{j=1}^k B_{i_j}$. \end{itemize} \end{definition} By this definition, already the linear part of bounded instances explicitly defines a bounded set by univariate constraints. Consequently, the set of solutions of $F$ is bounded as well. In \Cref{th:bounded} we show that when we consider bounded instances and restrict linearisations to so-called $\epsilon$-full linearisations, then the procedure terminates. We use this to show that the \ksmt-based decision procedure we introduce in \Cref{sec:delta-ksmt-cade} is $\delta$-complete. \begin{definition}\label{def:eps-full} Let $\eps>0$, $P$ be a non-linear constraint over variables $\vec x$ and let $\alpha$ be an assignment of $\vec x$. A linearisation $C$ of $P$ at $\alpha$ is called \emph{\epsfull} iff for all assignments $\beta$ of $\vec x$ with $\interp{\vec x}^\beta\in B(\interp{\vec x}^\alpha,\eps)$, $\interp{C}^\beta=\False$. A \ksmt \todofb{`almost all' weakens \Cref{th:int:lin-epsfull-cade,th:int:lin:alt-cade}.}% run is called \epsfull for some $\eps>0$, if all but finitely many linearisations in this run are $\epsilon$-full. \end{definition} \begin{comment} As in \cite{DBLP:conf/cade/GaoAC12} we only consider bounded instances. In many applications this is a natural assumption as variables usually range within some (possibly large) bounds. \todofb{remove sentence, implied by `bounded set'}% We can assume that these bounds are made explicit as linear constraints in the system. \end{comment} The next theorem provides a basis for termination of \ksmt-based decision procedures for satisfiability. \begin{theorem}\label{th:bounded} Let $\eps>0$. On bounded instances, \epsfull \ksmt runs are terminating. \end{theorem} \begin{proof} Let $F:\mathcal L_0 \wedge \mathcal N$ be a bounded instance and $\eps>0$. Towards a contradiction assume there is an infinite \epsfull derivation $(\alpha_0,\mathcal L_0,\mathcal N),\dots, (\alpha_n,\mathcal L_n,\mathcal N), \dots $ in the \ksmt calculus. Then, by definition of the transition rules, $\mathcal L_k\subseteq\mathcal L_l$ for all $k,l$ with $0\leq k\leq l$. According to \Cref{rem:lin-infty-often} in any infinite derivation the linearisation rule must be applied infinitely many times. During any run of \ksmt the set of non-linear constraints $\mathcal N$ is fixed and therefore there is a non-linear constraint $P$ in $\mathcal N$ over variables $\vec x$ to which linearisation is applied infinitely often. Let $(\alpha_{i_1},\mathcal L_{i_1},\mathcal N),\dots, (\alpha_{i_n},\mathcal L_{i_n},\mathcal N), \dots$ be a corresponding subsequence in the derivation such that $C_{i_1}\in \mathcal L_{i_1+1},\ldots,C_{i_n}\in \mathcal L_{i_n+1},\ldots$ are $\epsilon$-full linearisations of $P$. Consider two different linearisation steps $k,\ell\in\{i_j:j\in\mathbb N\}$ in the derivation where $k < \ell$. By the precondition $\interp{\mathcal L_\ell}^{\alpha_\ell}\neq\False$ of rule $(L)$ applied in step $\ell$, in particular the linearisation $C_k\in\mathcal L_{k+1}\subseteq\mathcal L_\ell$ of $P$ constructed in step $k$ does not evaluate to \False under $\alpha_\ell$. Since the set of variables in $C_k$ is a subset of those in $P$, $\interp{C_k}^{\alpha_\ell}\neq\False$ implies $\interp{C_k}^{\alpha_\ell}=\True$. By assumption, the linearisation $C_k$ is \epsfull, thus from Definition~\ref{def:eps-full} it follows that $\interp{\vec x}^{\alpha_\ell}\notin B(\interp{\vec x}^{\alpha_k},\epsilon)$. Therefore the distance between $\interp{\vec x}^{\alpha_k}$ and $\interp{\vec x}^{\alpha_\ell}$ is at least $\epsilon$. However, every conflict satisfies the variable bounds defining $D_F$, so there could be only finitely many conflicts with pairwise distance at least $\epsilon$. This contradicts the above. \end{proof} Concrete algorithms to compute \epsfull linearisations are presented in \Cref{sec:delta-ksmt-cade,sec:eps-from-M-cade}. \section{$\delta$-decidability}\label{sec:delta-sat} In the last section, we proved termination of the \ksmt calculus on bounded instances when linearisations are \epsfull. Let us now investigate how \epsfull linearisations of constraints involving non-linear computable functions can be constructed. To that end, we assume that all non-linear functions are defined on the closure of the bounded space $D_F$ defined by the bounded instance $F$. So far we described an approach which gives exact results but at the same time is necessarily incomplete due to undecidability of non-linear constraints in general. On the other hand, non-linear constraints usually can be approximated using numerical methods allowing to obtain approximate solutions to the problem. This gives rise to the bounded $\delta$-SMT problem~\cite{DBLP:conf/cade/GaoAC12} which allows an overlap between the properties $\delta$-\SAT and \UNSAT of formulas as illustrated by \Cref{fig:lin-mv-cade}. It is precisely this overlap that enables $\delta$-decidability of bounded instances. \irrelcade{ The notion of computability for functions over the reals is \todofb{cite}% closely related to effective continuity. That way, the problem of deciding satisfiability of a formula can be lifted to a continuous domain where functions are no longer handled on a purely algebraic basis only as symbols, but also gain the quality of approximability. In particular, computation of these functions is performed on continuous spaces of sequences of approximations of real numbers. } Let us recall the notion of $\delta$-decidability, adapted from~\cite{DBLP:conf/cade/GaoAC12}. \begin{definition}\label{def:pred-approx}% Let $F$ be a formula in separated linear form and let $\delta\in\mathbb Q_{>0}$. We inductively define the $\delta$-weakening $F_\delta$ of $F$. \begin{itemize} \item If $F$ is linear, let $F_\delta\coloneqq F$. \item If $F$ is a non-linear constraint $f(\vec x)\diamond 0$, let \[ F_\delta\coloneqq\begin{cases} f(\vec x)-\delta\diamond 0,&\text{if}~\diamond\in\{<,\leq\} \\ f(\vec x)+\delta\diamond 0,&\text{if}~\diamond\in\{>,\geq\} \\ |f(\vec x)|-\delta\leq 0,&\text{if}~\diamond\in\{=\} \\ (f(\vec x)<0\lor f(\vec x)>0)_\delta,&\text{if}~\diamond\in\{\neq\}\text. \end{cases} \] \item Otherwise, $F$ is $A\circ B$ with $\circ\in\{\land,\lor\}$. Let $F_\delta\coloneqq(A_\delta\circ B_\delta)$. \end{itemize} \noindent\emph{$\delta$-deciding} $F$ designates computing \[ \begin{cases} \text{\UNSAT},&\text{if}~\interp F^\alpha=\False~\text{for all}~\alpha \\ \text{$\delta$-\SAT},&\text{if}~\interp{F_\delta}^\alpha=\True~\text{for some}~\alpha\text. \end{cases} \] In case both answers are valid, the algorithm may output any. An assignment $\alpha$ with $\interp{F_\delta}^\alpha=\True$ we call a \emph{$\delta$-satisfying assignment} for $F$. \end{definition} \begin{figure}[tp] \centering \begin{tikzpicture} \draw[->] (-1,0) -- (2,0) node [right] {$f(x)$}; \draw (0,-.15) node [below] {$0$} -- (0,.15); \draw (1,-.15) node [below] {$\delta$} -- (1,.15); \draw (-1,1) node [left] {$\delta$-\SAT} -- (1,1) node {$]$}; \draw (0,.5) node {$($} -- (2,.5) node [right] {\UNSAT}; \end{tikzpicture} \caption{% The overlapping cases in the $\delta$-SMT problem $f(x)\leq 0$. } \label{fig:lin-mv-cade} \end{figure} For non-linear constraints $P$ this definition of the $\delta$-weakening $P_\delta$ corresponds exactly to the notion of $\delta$-weakening $P^{-\delta}$ used in the introduction of $\delta$-decidability \cite[Definition 4.1]{DBLP:conf/lics/GaoAC12}. \begin{remark}\label{rem:delta:eq-le-ge} The $\delta$-weakening of a non-linear constraint $f(\vec x)\neq 0$ is a tautology. \end{remark} We now consider the problem of $\delta$-deciding quantifier-free formulas in separated linear form. The notion of $\delta$-decidability is slightly stronger than in \cite{DBLP:conf/cade/GaoAC12} in the sense that we do not weaken linear constraints. Consider a formula $F$ in separated linear form. As before, we assume variables $\vec x$ to be bounded by linear constraints $\vec x\in D_F$. We additionally assume that for all non-linear constraints $P:f(\vec x)\diamond0$ in $\mathcal N$, $f$ is defined on $\widebar{D_P}$ and, in order to simplify the presentation, throughout the rest of paper we will assume only the predicates $\diamond\in\{{>},{\geq}\}$ are part of formulas, since the remaining ones ${<},{\leq},{=}$ can easily be expressed by the former using simple arithmetic transformations, and by \Cref{rem:delta:eq-le-ge} predicates $\neq$ are irrelevant for $\delta$-deciding formulas. An algorithm is \emph{$\delta$-complete}, if it $\delta$-decides bounded instances% ~\cite{DBLP:conf/cade/GaoAC12}. \section{$\delta$-\ksmt}\label{sec:delta-ksmt-cade} Since $\delta$-decidability as introduced above adapts the condition when a formula is considered to be satisfied to $\delta$-\SAT, this condition has to be reflected in the calculus, which we show solves the bounded $\delta$-SMT problem in this section. Adding the following rule $\Fsat[\delta]$ together with the new final state $\delta$-\SAT to \ksmt relaxes the termination conditions and turns it into the extended calculus we call $\delta$-\ksmt. \begin{description} \item[\ensuremath{\Fsat[\delta]}] \emph{Final $\delta$-\SAT.} If $(\alpha,\mathcal L,\mathcal N)$ is a $\delta$-\ksmt state where $\alpha$ is a total assignment and $\interp{\mathcal L\land\mathcal N_\delta}^\alpha=\True$, transition to the $\delta$-\SAT state. \end{description} The applicability conditions on the rules $(L)$ and $\Fsat[\delta]$ individually are not decidable~\cite{DBLP:journals/jsyml/Richardson68,Brattka2008}, however, when we compute them simultaneously, we can effectively apply one of these rules, as we will show in \Cref{th:int:lin-correct-cade}. In combination with \epsfull{}ness of the computed linearisations (\Cref{th:int:lin-epsfull-cade}), this leads to \Cref{thm:delta-ksmt-complete-cade}, showing that $\delta$-\ksmt is a $\delta$-complete decision procedure. Let us note that if we assume $\delta=0$ then $\delta$-\ksmt would just reduce to \ksmt as $\Fsat$ and $\Fsat[\delta]$ become indistinguishable, but in the following we always assume $\delta>0$. In the following sub-section, we prove that terminating derivations of the $\delta$-\ksmt calculus lead to correct results. Then, in \Cref{sec:pf-snd-thm-cade}, we present a concrete algorithm for applying rules $(L)$ and $\Fsat[\delta]$ and show its linearisations to be \epsfull, which is sufficient to ensure termination, as shown in \Cref{th:bounded}. These properties lead to a $\delta$-complete decision procedure. In \Cref{sec:eps-from-M-cade} we develop a more practical algorithm for $\epsilon$-full linearisations that does not require computing a uniform modulus of continuity. \subsection{Soundness} In this section we show soundness of the $\delta$-\ksmt calculus, that is, validity of its derivations. In particular, this implies that derivability of the final states $\UNSAT$, $\delta$-\SAT and \SAT directly corresponds to unsatisfiability, $\delta$-satisfiability and satisfiability of the original formula, respectively. \begin{lemma}\label{lem:preserve_assign-cade} For all $\delta$-\ksmt derivations of $S'=(\alpha',\mathcal L',\mathcal N)$ from a state $S=(\alpha,\mathcal L,\mathcal N)$ and for all total assignments $\beta$, $\interp{\mathcal L\land\mathcal N}^\beta= \interp{\mathcal L'\land\mathcal N}^\beta$. \end{lemma} \begin{proof} Let $\beta$ be a total assignment of the variables in $\mathcal L\land\mathcal N$. Since the set of variables remains unchanged by $\delta$-\ksmt derivations, $\beta$ is a total assignment for $\mathcal L'\land\mathcal N$ as well. Let $S'=(\alpha',\mathcal L',\mathcal N)$ be derived from $S=(\alpha,\mathcal L,\mathcal N)$ by a single application of one of $\delta$-\ksmt rules. By the structure of $S'$, its derivation was not caused by neither $\Funsat,\Fsat$ or $\Fsat[\delta]$. For rules $(A)$ and $(B)$ there is nothing to show since $\mathcal L=\mathcal L'$. If $(R)$ caused $S\mapsto S'$, the claim holds by soundness of arithmetical resolution. Otherwise $(L)$ caused $S\mapsto S'$ in which case the direction $\Rightarrow$ follows from the definition of a linearisation (condition~\ref{def:linearisation:it1} in \Cref{def:linearisation}) while the other direction trivially holds since $\mathcal L\subseteq\mathcal L'$. The condition on derivations of arbitrary lengths then follows by induction. \end{proof} \begin{lemma}\label{thm:delta_sound-cade''} Let $\delta\in\mathbb Q_{>0}$. Consider a formula $G=\mathcal L_0\land\mathcal N$ in separated linear form and let $S=(\alpha,\mathcal L,\mathcal N)$ be a $\delta$-\ksmt state derivable from the initial state $S_0=(\nil,\mathcal L_0,\mathcal N)$. The following hold. \begin{itemize} \item If rule $\Funsat$ is applicable to $S$ then $G$ is unsatisfiable. \item If rule $\Fsat[\delta]$ is applicable to $S$ then $\alpha$ is a $\delta$-satisfying assignment for $G$, hence $G$ is $\delta$-satisfiable. \item If rule $\Fsat$ is applicable to $S$ then $\alpha$ is a satisfying assignment for $G$, hence $G$ is satisfiable. \end{itemize} \end{lemma} \begin{proof} Let formula $G$ and states $S_0,S$ be as in the premise. As $S$ is not final in $\delta$-\ksmt, only \ksmt rules have been applied in deriving it. The statements for rules $\Funsat$ and $\Fsat$ thus hold by soundness of \ksmt~\cite[Lemma~2]{DBLP:conf/frocos/BrausseKKM19}. Assume $\Fsat[\delta]$ is applicable to $S$, that is, $\interp{\mathcal L\land\mathcal N_\delta}^\alpha$ is \True. Then, since $\mathcal L_0\subseteq\mathcal L$, we conclude that $\alpha$ satisfies $\mathcal L_0\land\mathcal N_\delta$ which, according to \Cref{def:pred-approx}, equals $G_\delta$. Therefore $\alpha$ is a $\delta$-satisfying assignment for $G$. \end{proof} Since the only way to derive one of the final states \UNSAT, $\delta$-\SAT and \SAT from the initial state in $\delta$-\ksmt is by application of the rule $\Funsat,\Fsat[\delta]$ and $\Fsat$, respectively, as corollary of \Cref{lem:preserve_assign-cade,thm:delta_sound-cade''} we obtain soundness. \begin{theorem}[Soundness]\label{thm:delta_sound-cade} Let $\delta\in\mathbb Q_{>0}$. The $\delta$-\ksmt calculus is sound. \end{theorem} \subsection{$\delta$-completeness}\label{sec:pf-snd-thm-cade} We proceed by introducing \Cref{alg:box-linearisation'-cade} computing linearisations and deciding which of the rules $\Fsat[\delta]$ and $(L)$ to apply. These linearisations are then shown to be \epsfull for some $\eps>0$ depending on the bounded instance. By \Cref{th:bounded}, this property implies termination, showing that $\delta$-\ksmt is a $\delta$-complete decision procedure. Given a non-final $\delta$-\ksmt state, the function \Call{nlinStep$_\delta$}{} in \Cref{alg:box-linearisation'-cade} computes a $\delta$-\ksmt state derivable from it by application of $\Fsat[\delta]$ or $(L)$. This is done by evaluating the non-linear functions and adding a linearisation $\ell$ based on their uniform moduli of continuity as needed. To simplify the algorithm, it assumes total assignments as input. It is possible to relax this requirement, e.g., by invoking rules $(A)$ or $(R)$ instead of returning $\delta$-\SAT for partial assignments. \begin{algorithm}[tp] \setlength{\columnsep}{0cm} \begin{multicols}{2} \begin{algorithmic} \Function{linearise$_\delta$}{$f,\vec x,\diamond,\alpha$} \State compute $p\geq-\lfloor\log_2(\min\{1,\delta/4\})\rfloor$ \State $\varphi\gets(n\mapsto\interp{\vec x}^\alpha)$ \State $\eps\gets2^{-\mu_f(p)}$ \State $\tilde y\gets M_f^\varphi(p)$ \If{$\tilde y\diamond-\delta/2$} \State\Return $\None$ \EndIf \State\Return $(\vec x\notin B(\interp{\vec x}^\alpha, \eps))$ \EndFunction \columnbrea \Function{nlinStep$_\delta$}{$\alpha,\mathcal L,\mathcal N$} \For{$P:(f(\vec x)\diamond 0)$ \textbf{in} $\mathcal N$} \State $\ell\gets{}$\Call{linearise$_\delta$}{$f,\vec x,\diamond,\alpha$} \If{$\ell\neq\Non $} \State\Return $(\alpha,\mathcal L\cup\{\ell\},\mathcal N)$ \Comment $(L)$ \EndIf \EndFor \State \Return $\delta$-\SAT \Comment $\Fsat[\delta]$ \EndFunction \end{algorithmic} \end{multicols} \caption{% (\textsc{nlinStep$_\delta$}) Algorithm computing a $\delta$-\ksmt derivation according to either rule $(L)$ or $\Fsat[\delta]$ from a state $(\alpha,\mathcal L,\mathcal N)$ where $\alpha$ is total. The functions $f$ are assumed to be computed by machines $M_f^?$ and $\mu_f$ to be a computable uniform modulus of continuity of $f$. } \label{alg:box-linearisation'-cade} \end{algorithm} \begin{lemma}\label{th:int:lin-correct-cade} Let $\delta\in\mathbb Q_{>0}$ and let $S=(\alpha,\mathcal L,\mathcal N)$ be a $\delta$-\ksmt state where $\alpha$ is total and $\interp{\mathcal L}^\alpha=\True$. Then \Call{nlinStep$_\delta$}{$\alpha,\mathcal L,\mathcal N$} computes a state derivable by application of either $(L)$ or $\Fsat[\delta]$ to $S$. \end{lemma} \begin{proof} In the proof we will use notions from computable analysis, as defined in \Cref{sec:ca}. Let $(\alpha,\mathcal L,\mathcal N)$ be a state as in the premise and let $P:f(\vec x)\diamond0$ be a non-linear constraint in $\mathcal N$. Let $M_f^?$ compute $f$ as in \Cref{alg:box-linearisation'-cade}. The algorithm computes a rational approximation $\tilde y=M_f^{(\interp{\vec x}^\alpha)_i}(p)$ of $f(\interp{\vec x}^\alpha)$ where $p\geq-\lfloor\log_2(\min\{1,\delta/4\})\rfloor\in\mathbb N$. $\interp{\mathcal L}^\alpha=\True$ implies $\interp{\vec x}^\alpha\in D_P\subseteq\dom f$, thus the computation of $\tilde y$ terminates. Since $M_f^?$ computes $f$, $\tilde y$ is accurate up to $2^{-p}\leq\delta/4$, that is, $\tilde y\in[f(\interp{\vec x}^\alpha)\pm\delta/4]$. By assumption $\diamond\in\{{>},{\geq}\}$, thus \begin{enumerate} \item\label{th:int:lin:delta-sat} $\tilde y\mathrel\diamond-\delta/2$ implies $f(\interp{\vec x}^\alpha)\mathrel\diamond-\delta$, which is equivalent to $\interp{P_\delta}^\alpha=\True$, and \item\label{th:int:lin:unsat} $\neg(\tilde y\mathrel\diamond-\delta/2)$ implies $\neg(f(\interp{\vec x}^\alpha)\mathrel\diamond-\delta/2+\delta/4)$, which in turn implies $\interp P^\alpha=\False$ and the applicability of rule $(L)$. \end{enumerate} For \cref{th:int:lin:delta-sat} no linearisation is necessary and indeed the algorithm does not linearise $P$. Otherwise (\Cref{th:int:lin:unsat}), it adds the linearisation $(\vec x\notin B(\interp{\vec x}^\alpha,\eps_P))$ to the linear clauses. Since $\interp{\vec x}^\alpha\in D_P$ by \cref{eq:mu} we obtain that $0\notin B(f(\vec z),\delta/4)$ holds, implying $\neg(f(\vec z)\diamond 0)$, for all $\vec z\in B(\interp{\vec x}^\alpha,\eps_P)\cap\widebar{D_P}$. Hence, $(\vec x\notin B(\interp{\vec x}^\alpha,\eps_P))$ is a linearisation of $P$ at $\alpha$. In case \Call{nlinStep$_\delta$}{$\alpha,\mathcal L,\mathcal N$} returns $\delta$-\SAT, the premise of \Cref{th:int:lin:delta-sat} holds for every non-linear constraint in $\mathcal N$, that is, $\interp{\mathcal N_\delta}^\alpha=\True$. By assumption $\interp{\mathcal L}^\alpha=\True$, hence the application of the $\Fsat[\delta]$ rule deriving $\delta$-\SAT is possible in $\delta$-\ksmt. \end{proof} \begin{lemma}\label{th:int:lin-epsfull-cade} For any bounded instance $\mathcal L_0\land\mathcal N$ there is a computable $\eps\in\mathbb Q_{>0}$ such that any $\delta$-\ksmt run starting in $(\nil,\mathcal L_0,\mathcal N)$, where applications of $(L)$ and $\Fsat[\delta]$ are performed by \Call{nlinStep$_\delta$}{}, is \epsfull. \end{lemma} \begin{proof} Let $P:f(\vec x)\diamond 0$ be a non-linear constraint in $\mathcal N$. Since $\mathcal L_0\land\mathcal N$ is a bounded instance, $D_P\subseteq\mathbb R^n$ is also bounded. Let $\eps_P\coloneqq2^{-\mu_f(p)}$ where $p\geq-\lfloor\log_2(\min\{1,\delta/4\})\rfloor\in\mathbb N$ as in \Cref{alg:box-linearisation'-cade}. As $\mu_f$ is a uniform modulus of continuity, the inequalities in the following construction hold on the whole domain $\widebar{D_P}$ of $f$ and do not depend on the concrete assignment $\alpha$ where the linearisation is performed. Since $\log_2$ and $\mu_f$ are computable, so are $p$ and $\eps_P$. There are finitely many non-linear constraints $P$ in $\mathcal N$, therefore the linearisations the algorithm \Call{nlinStep$_\delta$}{} computes are \epsfull with $\epsilon=\min\{\eps_P:P~\text{in}~\mathcal N\}>0$. \end{proof} We call $\delta$-\ksmt derivations when linearisation are computed using Algorithm~\ref{alg:box-linearisation'-cade} $\delta$-\ksmt with full-box linearisations, or \emph{$\delta$-\ksmt-fb} for short. As the runs computed by it are \epsfull for $\eps>0$, by \Cref{th:bounded} they terminate. \begin{theorem}\label{thm:delta-ksmt-complete-cade} $\delta$-\ksmt-fb is a $\delta$-complete decision procedure. \end{theorem} \begin{proof} $\delta$-\ksmt-fb is sound (\Cref{thm:delta_sound-cade}) and terminates on bounded instances (\Cref{th:bounded,th:int:lin-epsfull-cade}). \end{proof} \section{Local \epsfull linearisations} \label{sec:eps-from-M-cade} In practice, when the algorithm computing \epsfull linearisations described in the previous section is going to be implemented, the question arises of how to get a good uniform modulus of continuity $\mu_f$ for a computable function $f$. Depending on how $f$ is given, there may be several ways of computing it. Implementations of exact real arithmetic, e.g., iRRAM \cite{DBLP:conf/cca/Muller00} and Ariadne \cite{https://doi.org/10.1002/rnc.2914}, are usually based on the formalism of function-oracle Turing machines (see \Cref{def:computable}) which allow to compute with representations of computable functions~\cite{DBLP:journals/corr/BrausseS17} including implicit representations of functions as solutions of ODEs/PDEs~\cite{DBLP:books/daglib/0087495,DBLP:conf/ershov/BrausseKM15}. If $f$ is only available as a function-oracle Turing machine $M_f^?$ computing it, a modulus $\mu_f$ valid on a compact domain can be computed, however, in general this is not possible without exploring the behaviour of the function on the whole domain, which in many cases is computationally expensive. Moreover, since $\mu_f$ is uniform, $\mu_f(n)$ is constant throughout $D_F$, independent of the actual assignment $\alpha$ determining where $f$ is evaluated. Yet, computable functions admit \emph{local} moduli of continuity that additionally depend on the concrete point in their domain. In most cases these would provide linearisations with $\eps$ larger than that determined by $\mu_f$ leading to larger regions being excluded, ultimately resulting in fewer linearisation steps and general speed-up. Indeed, machines producing finite approximations of $f(x)$ from finite approximations of $x$ internally have to compute some form of local modulus to guarantee correctness. In this section, we explore this approach of obtaining linearisations covering a larger part of the function's domain. In order to guarantee a positive bound on the local modulus of continuity extracted directly from the run of the machine $M_f^?$ computing $f$, it is necessary to employ a restriction on the names of real numbers $M_f^?$ computes on. The set of names should in a very precise sense be ``small'', i.e., it has to be compact. The very general notion of names used in \Cref{def:computable} is too broad to satisfy this criterion since the space of rational approximations is not even locally compact. Here, we present an approach using practical names of real numbers as sequences of dyadic rationals of lengths restricted by accuracy. For that purpose, we introduce another representation~\cite{DBLP:series/txtcs/Weihrauch00} of $\mathbb R$, that is, the surjective mapping $\xi:\mathbb D_\omega\to\mathbb R$. Here, $\mathbb D_\omega$ denotes the set of infinite sequences $\varphi$ of dyadic rationals with bounded length. If $\varphi$ has a limit (in $\mathbb R$), we write $\lim\varphi$. \begin{definition}\label{def:Domega-xi-Cauchy} \begin{itemize} \item For $k\in\omega$ let $\mathbb D_k\coloneqq\mathbb Z\cdot2^{-(k+1)}=\{m/2^{k+1}:m\in\mathbb Z\}\subset\mathbb Q$ and let $\mathbb D_\omega\coloneqq\bigtimes_{k\in\omega}\mathbb D_k$ be the set of all sequences $(\varphi_k)_k$ with $\varphi_k\in\mathbb D_k$ for all $k\in\omega$. By default, $\mathbb D_\omega$ is endowed with the Baire space topology, which corresponds to that induced by the metric \[ d:(\varphi,\psi)\mapsto\begin{cases} 0&\text{if}~\varphi=\psi \\ 1/{\min\{1+n:n\in\omega,\varphi_n\neq\psi_n\}}&\text{otherwise.} \end{cases} \] \item Define $\xi:\mathbb D_\omega\to\mathbb R$ as the partial function mapping $\varphi\in\mathbb D_\omega$ to $\lim\varphi$ iff $\forall i,j:|\varphi_i-\varphi_{i+j}|\leq2^{-(i+1)}$. Any $\varphi\in\xi^{-1}(x)$ is called a \emph{$\xi$-name} of $x\in\mathbb R$. \item The representation $\rho:(x_k)_k\mapsto x$ mapping names $(x_k)_k$ of $x\in\mathbb R$ to $x$ as per \Cref{def:computable} is called \emph{Cauchy representation}. \end{itemize} \end{definition} Using a standard product construction we can easily generalise the notion of $\xi$-names to $\xi^n$-names of $\mathbb R^n$. When clear from the context, we will drop $n$ and just write $\xi$ to denote the corresponding generalised representation $\mathbb D_\omega^n\to\mathbb R^n$. Computable equivalence between two representations not only implies that there are continuous maps between them but also that names can computably be transformed~\cite{DBLP:series/txtcs/Weihrauch00}. Since the Cauchy representation itself is continuous~\cite{DBLP:journals/tcs/BrattkaH02} we derive continuity of $\xi$, which is used below to show compactness of preimages $\xi^{-1}(X)$ of compact sets $X\subseteq\mathbb R$ under $\xi$. \cade{% All proofs can be found in~\todo[inlinepar]{\cite{}}. }{% All proofs can be found in the appendix. } \begin{lemma}\label{prop:cade} The following properties hold for $\xi$. \begin{enumerate} \item $\xi$ is a representation of $\mathbb R^n$: it is well-defined and surjective. \item\label{it:xi2rho-cade} Any $\xi$-name of $\vec x\in\mathbb R^n$ is a Cauchy-name of $\vec x$. \item\label{it:xi-equiv-rho-cade} $\xi$ is computably equivalent to the Cauchy representation. \item $\xi$ is continuous. \end{enumerate} \end{lemma}% The converse of \cref{it:xi2rho-cade} does not hold. An example for a Cauchy-name of $0\in\mathbb R$ is the sequence $(x_n)_n$ with $x_n=(-2)^{-n}$ for all $n\in\omega$, which does not satisfy $\forall i,j:|x_i-x_{i+j}|\leq2^{-(i+1)}$. However, given a name of a real number, we can compute a corresponding $\xi$-name, this is one direction of the property in \cref{it:xi-equiv-rho-cade}. As a consequence of \cref{it:xi2rho-cade} a function-oracle machine $M^?$ computing $f:\mathbb R^n\to\mathbb R$ according to \Cref{def:computable} can be run on $\xi$-names of $\vec x\in\mathbb R^n$ leading to valid Cauchy-names of $f(\vec x)$. Note that this proposition does not require $M_f^?$ to compute a $\xi$-name of $f(\vec x)$. Any rational sequence rapidly converging to $f(\vec x)$ is a valid output. This means, that the model of computation remains unchanged with respect to the earlier parts of this paper. It is the set of names the machines are operated on, which is restricted. This is reflected in \Cref{alg:2} by computing dyadic rational approximations $\tilde{\vec x}_k$ of $\interp{\vec x}^\alpha$ such that $\tilde{\vec x}_k\in\mathbb D_k^n$ instead of keeping the name of $\interp{\vec x}^\alpha$ constant as has been done in \Cref{alg:box-linearisation'-cade}. \begin{algorithm}[tp] \begin{algorithmic} \Function{LineariseLocal$_\delta$}{$f,\vec x,\diamond,\alpha$} \State $\varphi\gets(m\mapsto \Approx(\interp{\vec x}^\alpha,m))$ \Comment then $\varphi$ is a $\xi$-name of $\interp{\vec x}^\alpha$ \State compute $p\geq-\lfloor\log_2(\min\{1,\delta/4\})\rfloor$ \State run $M_f^\varphi(p+2)$, record its output $\tilde y$ and its maximum query $k\in\omega$ to $\varphi$ \If{$\tilde y\diamond-\delta/2$} \State \Return $\None$ \Else \State \Return $(\vec x\notin B(\interp{\vec x}^\alpha,2^{-k}))$ \EndIf \EndFunction \end{algorithmic} \caption{\textbf{(Local linearisation)} Algorithm $\delta$-deciding $P:f(\vec x)\diamond 0$ and -- in case \UNSAT{} -- computing a linearisation at $\alpha$ or returning ``\None'' and in this case $\alpha$ satisfies $P_\delta$. The function $f$ is computed by machine $M_f^?$. } \label{alg:2} \end{algorithm} In particular, in \Cref{th:int:lin:alt-cade} we show that linearisations for the $(L_\delta)$ rule can be computed by \Cref{alg:2}, which -- in contrast to \Call{linearise$_\delta$}{} in \Cref{alg:box-linearisation'-cade} -- does not require access to a procedure computing an upper bound $\mu_f$ on the uniform modulus of continuity of the non-linear function $f\in\mathcal F_{\mathrm{nl}}$ valid on the entire bounded domain. It not just runs the machine $M_f^?$, but also observes the queries $M_f^\varphi$ poses to its oracle in order to obtain a local modulus of continuity of $f$ at the point of evaluation. The function $\Approx(\vec x,m)\coloneqq\round{\vec x\cdot 2^{m+1}}/2^{m+1}$ used to define \Cref{alg:2} computes a dyadic approximation of $\vec x$, with $\round\cdot:\mathbb Q^n\to\mathbb Z^n$ denoting a rounding operation, that is, it satisfies $\forall \vec q:\Vert\round{\vec q}-\vec q\Vert\leq\frac12$. On rationals (our use-case), $\round\cdot$ is computable by a classical Turing machine. \begin{definition}[{\cite[Definition 6.2.6]{DBLP:series/txtcs/Weihrauch00}}]\label{def:local-mod} Let $f:\mathbb R^n\to\mathbb R$ and $\vec x\in\dom f$. A function $\gamma:\mathbb N\to\mathbb N$ is called \emph{a (local) modulus of continuity of $f$ at $\vec x$} if for all $p\in\mathbb N$ and $\vec y\in\dom f$, $\Vert\vec x-\vec y\Vert\leq 2^{-\gamma(p)}\implies\vert f(\vec x)-f(\vec y)\vert\leq2^{-p}$ holds. \end{definition} We note that in most cases a local modulus of continuity of $f$ at $\vec x$ is smaller than the best uniform modulus of $f$ on its domain, since it only depends on the local behaviour of $f$ around $x$. One way of computing a local modulus of $f$ at $\vec x$ is using the function-oracle machine $M_f^?$ as defined next. \begin{definition}\label{def:local-mod-M} Let $M^?_f$ compute $f:\mathbb R^n\to\mathbb R$ and let $\vec x\in\dom f$ have Cauchy-name $\varphi$. The function $\gamma_{M_f^?,\varphi}:p\mapsto\max\{0,k:k~\text{is queried by}~M_f^\varphi(p+2)\}$ is called \emph{the effective local modulus of continuity induced by $M_f^?$ at $\varphi$}. \end{definition} The effective local modulus of continuity of $f$ at a name $\varphi$ of $\vec x\in\dom f$ indeed is a local modulus of continuity of $f$ at $\vec x$~\cite[Theorem~2.13]{DBLP:books/daglib/0067010}. We prove that \Cref{alg:2} indeed computes linearisations in \Cref{prf:local-lin''}. \begin{lemma}\label{lem:local-lin''} Let $P:f(\vec x)\diamond0$ be a non-linear constraint in $\mathcal N$ and $\alpha$ be an assignment of $\vec x$ to rationals in $\dom f$. Whenever $C={}$\Call{LineariseLocal$_\delta$}{$f,\vec x,\diamond,\alpha$} and $C\neq\None$, $C$ is an \epsfull linearisation of $P$ at $\alpha$, with $\epsilon$ corresponding to the effective local modulus of continuity induced by $M_f^?$ at a $\xi$-name of $\interp{\vec x}^\alpha$. \end{lemma} Thus, the function \textsc{lineariseLocal$_\delta$}{} in \Cref{alg:2} is a drop-in replacement for \textsc{linearise$_\delta$}{} in \Cref{alg:box-linearisation'-cade} since the condition on returning a linearisation of $P$ versus accepting $P_\delta$ is identical. The linearisations however differ in the radius $\epsilon$, which now, according to \Cref{lem:local-lin''}, corresponds to the effective local modulus of continuity. The resulting procedure we call \Call{nlinStepLocal$_\delta$}{}. One of its advantages over \Call{nlinStep$_\delta$}{} is running $M_f^?$ on $\xi$-names instead of Cauchy-names, is that they form a compact set for bounded instances, unlike the latter. This allows us to bound $\eps>0$ for the computed \epsfull local linearisations of otherwise arbitrary $\delta$-\ksmt runs. A proof of the following Lemma showing compactness of preimages $\xi^{-1}(X)$ of compact sets $X\subseteq\mathbb R$ under $\xi$ is given in \Cref{prf:names-compact-cade}. \begin{lemma}\label{prop:names-compact-cade} Let $X\subset\mathbb R^n$ be compact. Then the set $\xi^{-1}(X)\subset\mathbb D_\omega^n$ of $\xi$-names of elements in $X$ is compact as well. \end{lemma} The proof involves showing $\xi^{-1}(X)$ to be closed and uses the fact that for each component $\varphi_k$ of names $(\varphi_k)_k$ of $\vec x\in X$ there are just finitely many choices from $\mathbb D_k$ due to the restriction of the length of the dyadics. This is not the case for the Cauchy representation used in \Cref{def:computable} and \todorev[2]{`sceptical'. \emph{`We will [...] add clarifications'}}% it is the key for deriving existence of a strictly positive lower bound $\epsilon$ on the \epsfull{}ness of linearisations. \begin{theorem}\label{th:int:lin:alt-cade} Let $\delta\in\mathbb Q_{>0}$. For any bounded instance $\mathcal L_0\land\mathcal N$ there is $\epsilon>0$ such that any $\delta$-\ksmt run starting in $(\nil,\mathcal L_0,\mathcal N)$, where applications of $(L)$ and $\Fsat[\delta]$ are performed according to \Call{nlinStepLocal$_\delta$}{}, is \epsfull. \end{theorem} \begin{proof} Assume $\mathcal L_0\land\mathcal N$ is a bounded instance. Set $\eps\coloneqq\min\{\eps_P:P\in\mathcal N\}$, where $\eps_P$ is defined as follows. Let $P:f(\vec x)\diamond 0$ in $\mathcal N$. Then the closure $\widebar{D_P}$ of the bounded set $D_P$ is compact. Let $E$ be the set of $\xi$-names of elements of $\widebar{D_P}\subseteq\dom f$ (see \Cref{def:Domega-xi-Cauchy}) and for any $\varphi\in E$ let $k_\varphi$ be the maximum index queried by $M_f^\varphi(p)$ where $p$ is computed from $\delta$ as in \Cref{alg:2}. Therefore $\varphi\mapsto k_\varphi$ is continuous. By \Cref{prop:names-compact-cade} $E$ is compact, thus, there \todofb{could also state $k_\psi=\max\{k_\varphi:\varphi\in E\}$ instead; which is clearer? \fb{keep} \mk{keep}}% is $\psi\in E$ such that $2^{-k_\psi}= \inf\{2^{-k_\varphi}:\varphi\in E\}$. Set $\eps_P\coloneqq2^{-k_\psi}$. The claim then follows by \Cref{lem:local-lin''}. \end{proof} Thus we can conclude. \begin{corollary} $\delta$-\ksmt with local linearisations is a $\delta$-complete decision procedure. \end{corollary} \section{Conclusion}\label{sec:concl} In this paper we extended the the \ksmt calculus to the $\delta$-satisfiability setting and proved that the resulting $\delta$-\ksmt calculus is a $\delta$-complete decision procedure for solving non-linear constraints over computable functions which include polynomials, exponentials, logarithms, trigonometric and many other functions used in applications. We presented algorithms for constructing $\epsilon$-full linearisations ensuring termination of $\delta$-\ksmt. Based on methods from computable analysis we presented an algorithm for constructing local linearisations. Local linearisations exclude larger regions from the search space and can be used to avoid computationally expensive global analysis of non-linear functions. \bibliographystyle{plain}

aebdb3814179fe98b5c5c1ee55b66b55d15c60dc

\section{Introduction} \IEEEPARstart{A}{utomatic} speech recognition (ASR) is a well established field with diverse applications including captioning voiced speech and recognizing voice commands. Deep learning has revolutionised this task in the past years, to the point where state of the art models are able to achieve very low word error rates (WER) \cite{DBLP:conf/interspeech/LiLGLKCNG19}. Although these models are reliable for clean audio, they struggle under noisy conditions \cite{DBLP:conf/interspeech/MaasLOVNN12,DBLP:journals/tcyb/StewartSPM14}, and they are not effective when gaps are found in the audio stream \cite{DBLP:conf/iccv/Zhou0X0019}. The recurrence of these edge cases has driven researchers towards Visual Speech Recognition (VSR), also known as lipreading, which performs speech recognition based on video only. Although the translation from video-to-text can now be achieved with remarkable consistency, there are various applications that would benefit from a video-to-audio model, such as videoconferencing in noisy conditions; speech inpainting \cite{DBLP:conf/iccv/Zhou0X0019}, i.\,e., filling in audio gaps from video in an audiovisual stream; or generating an artificial voice for people suffering from aphonia (i.\,e., people who are unable to produce voiced sound). One approach for this task would be to simply combine a lipreading model (which outputs text) with a text-to-speech (TTS) model (which outputs audio). This approach is especially attractive since state-of-the-art TTS models can now produce realistic speech with considerable efficacy \cite{DBLP:conf/icassp/ShenPWSJYCZWRSA18,DBLP:conf/icml/OordLBSVKDLCSCG18}. Combining video-to-text and text-to-speech models to perform video-to-speech has, however, some disadvantages. Firstly, these models require large transcribed datasets, since they are trained with text supervision. This is a sizeable constraint given that generating transcripts is a time consuming and expensive process. Secondly, generation can only happen as each word is recognized, which imposes a delay on the throughput of the model, jeopardizing the viability of real-time synthesis. Lastly, using text as an intermediate representation removes any intonation and emotion from the spoken statement, which are fundamental for natural sounding speech. Given these constraints, some authors have developed end-to-end video-to-speech models which circumvent these issues. The first of these models \cite{DBLP:conf/interspeech/CornuM15} used visual features based on discrete cosine transform (DCT) and active appearance models (AAM) to predict linear predictive coding (LPC) coefficients and mel-filterbank amplitudes. Following works have mostly focused on predicting spectrograms \cite{DBLP:conf/iccvw/EphratHP17,DBLP:conf/icassp/AkbariACM18,DBLP:journals/corr/abs-2005-08209}, which is also a common practice in text-to-speech works \cite{DBLP:conf/icassp/ShenPWSJYCZWRSA18}. These models achieve intelligible results, but are only applied to seen speakers, i.\,e., there is exact correspondence between the speakers in the training, validation and test sets, or choose to focus on single speaker speech reconstruction \cite{DBLP:journals/corr/abs-2005-08209}. Recently, \cite{DBLP:journals/corr/abs-2004-02541} has proposed an alternative approach based on predicting WORLD vocoder parameters \cite{DBLP:journals/ieicet/MoriseYO16} which generates clear speech for unseen speakers as well. However, the reconstructed speech is still not realistic. It is clear that previous works have avoided synthesising raw audio, likely due to the lack of a suitable loss function, and have focused on generating intermediate representations which are then used for reconstructing speech. To the best of our knowledge, the only work which directly synthesises the raw audio waveform from video is \cite{DBLP:journals/corr/abs-1906-06301}. This work introduces the use of GANs \cite{DBLP:journals/corr/GoodfellowPMXWOCB14,DBLP:journals/corr/ArjovskyCB17}, and thanks to the adversarial loss, it is able to directly reconstruct the audio waveform. This approach also produces realistic utterances for seen speakers, and is the first to produce intelligible speech for unseen speakers. Our work builds upon the model presented in \cite{DBLP:journals/corr/abs-1906-06301} by proposing architectural changes to the model, and to the training procedure. Firstly, we replace the original encoder composed of five stacked convolutional layers with a ResNet-18 \cite{DBLP:conf/cvpr/HeZRS16} composed of a front end 3D convolutional layer (followed by a max pooling layer), four blocks containing four convolutional layers each and an average pooling layer. Additionally, we replace the GRU (Gated Recurrent Unit) layer following the encoder with two bidirectional GRU layers, increasing the capacity of our temporal model. The adversarial methodology was a major factor towards generating intelligible waveform audio in \cite{DBLP:journals/corr/abs-1906-06301}. Hence, our approach is also based on the Wasserstein GAN \cite{DBLP:journals/corr/ArjovskyCB17}, but we propose a new critic adapted from \cite{DBLP:conf/nips/KumarKBGTSBBC19}. We also propose an additional critic which discriminates real from synthesized spectrograms. Furthermore, we revise the loss configuration presented in \cite{DBLP:journals/corr/abs-1906-06301}. Firstly, we decide to forego the use of the total variation loss and the L1 loss, as their benefit was minimal. Secondly, we use the recently proposed PASE (Problem Agnostic Speech Encoder) \cite{DBLP:journals/corr/abs-1904-03416} as a perceptual feature extractor. Finally, we propose two additional losses, the power loss and the MFCC loss. The power loss is an L1 loss between the (log-scaled) spectrograms of the real and generated waveforms. The MFCC loss is an L1 Loss between the MFCCs (Mel Frequency Cepstral Coefficients) of the real and generated waveforms. Our contributions for this work are described as follows: \textbf{1)} We propose a new approach for reconstructing waveform speech directly from video based on GANs without using any intermediate representations. We use two separate critics to discriminate real from synthesized waveforms and spectrograms respectively, and apply three comparative losses to improve the quality of outputs. \textbf{2)} We include a detailed ablation study where we measure the effect of each component on the final model. We also investigate how the type of visual input, size of training set and range of vocabulary affect the performance. \textbf{3)} We show results on two different datasets (GRID \cite{grid} and TCD-TIMIT \cite{7050271}) for seen speakers. We find that our model substantially outperforms the state-of-the-art for GRID and adapts well to a larger pool of speakers. \textbf{4)} We also include results for unseen speakers on two datasets (GRID and LRW \cite{Chung16}). We show that our model achieves intelligible results, even when applied to utterances recorded `in the wild', and outperforms the state-of-the-art for the corpora we present. \textbf{5)} Finally, we study our model's ability to generalize for videos of silent speakers, and discuss our findings. \section{Related Work} \label{section2} Video-driven speech reconstruction is effectively the combination of two tasks: lipreading and speech synthesis. As such, we begin by briefly describing the main works in each field, and then go on to describe existing approaches for video-to-speech. \subsection{Lipreading} Traditional lipreading approaches relied on HMMs (Hidden Markov Models) \cite{DBLP:journals/tsp/GurbanT09} or SVMs (Support Vector Machines) \cite{DBLP:journals/tmm/ZhaoBP09} to transcribe videos from manually extracted features such as DCT \cite{DBLP:journals/tsp/GurbanT09} or mouth geometry \cite{DBLP:conf/icassp/KumarCS07}. Recently, end-to-end models have attracted attention due to their superior performance over traditional approaches. One of the first end-to-end architectures for lipreading was \cite{DBLP:journals/corr/AssaelSWF16}. This model featured a convolutional encoder as the visual feature extractor and a two-layer BGRU-RNN (Bidirectional GRU recurrent neural network) followed by a linear layer as the classifier, and it achieved state of the art performance for the GRID corpus. This work was followed by \cite{DBLP:conf/accv/ChungZ16}, whose model relied entirely on CNNs (Convolutional Neural Networks) and RNNs, and was successfully applied to spoken utterances recorded in the wild. Various works have followed which apply end-to-end deep learning models to achieve competitive lipreading performance. \cite{DBLP:conf/icassp/PetridisLP17,DBLP:journals/prl/PetridisWMLP20} propose an encoder composed of fully connected layers and performs classification using LSTMs (Long-short Term Memory RNNs). Other works choose to use convolutional encoders \cite{DBLP:conf/interspeech/ShillingfordAHP19}, often featuring residual connections \cite{DBLP:conf/interspeech/StafylakisT17}, and then apply RNNs to perform classification. Furthermore, these end-to-end architectures have been extended for multi-view lipreading \cite{DBLP:conf/bmvc/PetridisWLP17} and audiovisual \cite{DBLP:conf/icassp/PetridisSMCTP18} speech recognition. \subsection{Speech Synthesis} \label{wss} One of the most popular speech synthesis models in recent years has been WaveNet \cite{DBLP:conf/ssw/OordDZSVGKSK16}, which proposed dilated causal convolutions to compose waveform audio sample by sample, taking advantage of the large receptive field achieved by stacking these layers. This model achieved far more realistic results than any artificial synthesizer proposed before then. Another work \cite{DBLP:conf/interspeech/WangSSWWJYXCBLA17} introduced a vastly different sequence-to-sequence model that predicted linear-scale spectrograms from text, which were then converted into waveform using the Griffin-Lim Algorithm (GLA) \cite{DBLP:journals/taslp/ZhuBW07}. This process produced very clear and intelligible audio. In the following years, \cite{DBLP:conf/icassp/ShenPWSJYCZWRSA18} combined these two methodologies to push the state-of-the-art once more, and \cite{DBLP:conf/icml/OordLBSVKDLCSCG18} accelerated and improved the original WaveNet. The first model to apply GANs for end-to-end speech synthesis was \cite{DBLP:conf/iclr/DonahueMP19}, which used simple convolutional networks with large kernels as the generator and discriminator and applied the improved Wasserstein loss \cite{DBLP:conf/nips/GulrajaniAADC17}. In a later work \cite{DBLP:journals/corr/abs-1910-11480}, the original WaveNet vocoder \cite{DBLP:conf/ssw/OordDZSVGKSK16} has been combined with the adversarial methodology introduced in \cite{DBLP:conf/iclr/DonahueMP19}. This results in a network which has far less parameters than the original WaveNet, but remains on par with the latest WaveNet-based models. Recently, the first end-to-end adversarial Text-To-Speech model \cite{DBLP:journals/corr/abs-2006-03575} was also proposed, whose performance is comparable to the state-of-the-art. \subsection{Reconstructing audio from visual speech} To the best of the authors' knowledge, the first work to attempt the task of video-to-speech synthesis directly was \cite{DBLP:conf/interspeech/CornuM15}. The proposed model aims to predict the spectral envelope (LPC or mel-filterbanks) from manually extracted visual features (DCT or AAM) using Gaussian Mixture Models (GMMs) or deep neural networks. These acoustic parameters are then fed into an HMM-based vocoder, together with an estimate of the voicing parameters. Through multiple user studies, the speech reconstructed by this model is shown to have fairly low intelligibility (WER $\approx 50\,\%$), but shows that this task is indeed achievable. This work was extended in \cite{DBLP:journals/taslp/CornuM17}, which introduced additional temporal information in the visual features and in the model itself. These improvements yielded an impressive 15\,\% WER for GRID (single speaker), based on user studies. The next development in this field comes with \cite{DBLP:conf/icassp/EphratP17}, which uses a deep CNN architecture to predict acoustic features -- LPC analysis followed by LSP (Line Spectral Pairs) decomposition, frame by frame -- from gray-scale video frames. These are combined with white noise (excitation signal) and fed into a source-filter speech synthesizer which produces unvoiced speech. This model produces intelligible results (WER $< 20\,\%$) when trained and tested on a single speaker from GRID, and constitutes a step forward given that it no longer relies on handcrafted visual features as input. An improved version of this model was presented in \cite{DBLP:conf/iccvw/EphratHP17}, which predicts spectrograms that are then translated into waveform using the Griffin-Lim algorithm. This extension also proposes a new encoder composed of two ResNet-18s followed by a post-processing network which increases temporal resolution. This work is the first to experiment with multiple speakers and achieves much more realistic speech than any previous work for this task. Lip2Audspec \cite{DBLP:conf/icassp/AkbariACM18} proposes a similar CNN+RNN encoder to predict spectrograms directly from the gray-scale frames of the video. As in \cite{DBLP:conf/iccvw/EphratHP17}, the spectrograms are converted to waveform using a phase estimation method. The resulting spectrograms are very close to the original samples, but the reconstructed waveforms sound noticeably robotic. Another recent work \cite{DBLP:journals/corr/abs-2004-02541} uses CNNs+RNNs to predict vocoder parameters (aperiodicity and spectral envelope), rather than spectrograms. Additionally, the model is trained to predict the transcription of the speech, in other words performing speech reconstruction and recognition simultaneously in a multi-task fashion. This approach achieves results which are very impressive when measured with objective speech quality metrics (PESQ, STOI), but yields samples which still sound noticeably robotic. Finally, a recent work \cite{DBLP:journals/corr/abs-2005-08209} proposes an approach based on the Tacotron 2 architecture \cite{DBLP:conf/icassp/ShenPWSJYCZWRSA18}, predicting mel-frequency spectrograms from video rather than text. To perform this task, it applies a stack of residual 3D convolutional layers as a spatio-temporal encoder for the video, and combines it with an attention-based decoder adapted from \cite{DBLP:conf/icassp/ShenPWSJYCZWRSA18}, which generates the spectrograms. Unlike Tacotron, these spectrograms are decoded into waveform audio using the Griffin-Lim algorithm \cite{DBLP:journals/taslp/ZhuBW07} rather than WaveNet, as the authors claim the generated spectrograms are not as accurate as modern TTS works, and therefore do not perform well with neural vocoders. This work is able to generate remarkably intelligible audio from visual speech and achieves state-of-the-art performance in all presented metrics. However, it focuses on speaker specific speech reconstruction, i.\,e., it is trained and tested on the same speaker. An aspect which is worth highlighting is that none of these models attempt to generate the waveform end-to-end from video, instead predicting spectrograms or other features which can be translated into waveform. This is likely due to the notoriously arduous task of generating realistic waveforms, which can be attributed to the lack of suitable loss functions. The only model to perform video-to-waveform speech reconstruction without the use of intermediate representations is \cite{DBLP:journals/corr/abs-1906-06301}. This work proposes a generative adversarial network based on a convolutional encoder-decoder model (combined with a GRU) which encodes video into visual features and decodes them directly into waveform audio. The generator is trained with an adversarial loss based on a convolutional waveform critic, as well as three other comparative losses. This procedure achieves competitive results for speech reconstruction on both seen and unseen speaker datasets (GRID). \subsection{Reconstructing audio from multi-view visual speech} The majority of works in video-driven speech reconstruction use frontal views of the face. In this section, we briefly describe a set of works which use multiple-views in order to improve the quality of reconstructed speech. The first work to use multi-view video for this task was \cite{DBLP:conf/mm/KumarANSSZ18}. This model is very similar to \cite{DBLP:conf/icassp/EphratP17} in the sense that it applies a CNN to extract visual features directly from video, which then predict vocoder parameters (LPC followed by LSP). This work, however, uses video taken from two different angles for every speaker (Oulu VS2 dataset \cite{DBLP:conf/fgr/AninaZZP15}). The results presented in this paper show that the use of multiple views can substantially improve speech reconstruction performance. This model has been improved in \cite{DBLP:conf/ism/KumarJSSZY18} by replacing the LSTM with a BGRU and using more than two views as input. It is shown that the use of three views can yield improvements of 20\,\% in the quality of reconstructed outputs (measured with PESQ). This has been extended in \cite{DBLP:conf/aaai/KumarJSSYZ19} by including a view classifier to attribute view labels to the input videos and by also generating text transcriptions. The latest work in this field \cite{DBLP:conf/interspeech/UttamKSASMS19} follows the trend seen in single-view speech reconstruction research \cite{DBLP:conf/icassp/EphratP17,DBLP:conf/iccvw/EphratHP17} and speech synthesis in general \cite{5931414,DBLP:conf/interspeech/WangSSWWJYXCBLA17} by switching from LPC coefficients to spectrograms as the predicted audio representation. \subsection{Audio reconstruction from video in other applications } Finally, a set of past works has approached the application of Video-to-Audio models to domains outside speech \cite{DBLP:conf/cvpr/OwensIMTAF16,DBLP:conf/cvpr/ZhouWFBB18,DBLP:conf/mm/ChenSDX17}. Namely, these papers have focused on a diverse range of datasets which feature a set of generic sounds such as fireworks and drums \cite{DBLP:conf/cvpr/ZhouWFBB18}; different instruments being played \cite{DBLP:conf/mm/ChenSDX17}; or even objects composed of different materials being hit with a drumstick \cite{DBLP:conf/cvpr/OwensIMTAF16}. The methodology applied to reconstruct audio from video is similar to what is seen for video-to-speech systems. CNNs are applied to encode the video frames, followed by RNNs or fully connected layers to produce acoustic features which are decoded into audio using vocoders. While some of these works struggle to reproduce the corresponding audio, \cite{DBLP:conf/cvpr/ZhouWFBB18} is able to produce remarkably realistic audio (as proven by its user studies) by combining the extraction of optical flow with a neural network-based vocoder. \section{Video-Driven Speech Reconstruction} \label{section3} \begin{figure} \centering \includegraphics[width = 0.86\linewidth]{./figures/architecture.pdf} \caption{Architecture of the generator (encoder, bidirectional GRU, decoder) and critics (waveform critic, power critic) used in this work, as well as the losses that are used for training.} \label{architecture} \end{figure} Our model is composed of a video encoder based on a ResNet-18 combined with a Bidirectional GRU, as well as a convolutional decoder which transforms the visual features into waveform audio. This generator is trained using two separate critics, to ensure the realism of the outputs, as well as three L1 losses to minimize the difference between real and synthesized audio for each video. \subsection{Generator} Given that we aim to synthesize speech directly from video, our generator accomplishes two sequential tasks: encode temporal visual features and decode them into an audio waveform. Firstly, we encode the frames of the video using a Resnet-18 preceded by a spatio-temporal 3D convolutional layer (combined with a max pooling layer). This initial layer has a receptive field of 5 frames centered on the frame it will encode meaning that the encoding for each frame will depend on the previous two frames and on the following two frames. We experimented with different numbers of frames as input to this layer (3 and 7), but found that this did not considerably affect results. The ResNet-18 is composed of 4 blocks of 4 convolutional layers, each followed by batch normalization and ReLU (Rectified Linear Unit) activation, and an adaptive average pooling layer. The features extracted from the ResNet encoder are then fed into a 2-layer bidirectional GRU which temporally correlates the features produced from each set of frames. This architecture is described in detail in Figure \ref{encoder}. \begin{figure*} \centering \includegraphics[width = 0.9\linewidth]{./figures/encoder_rotated.pdf} \caption{Description of the layers in the encoder (generator).} \label{encoder} \end{figure*} After this, the decoder upsamples the features from each video frame into a waveform segment of $N$ audio samples. The length of each segment is given by: \begin{align} N = \frac{audio\ sampling\ rate}{video\ frame\ rate}. \end{align} Since we use a sampling rate of 16\,kHz and a frame rate of 25 frames per second, $N$ is equal to 640 (corresponding to 40\,ms of audio). The decoder is composed of six stacked transposed convolutional layers, each followed by batch normalization and ReLU activation except for the last layer which uses a hyperbolic tangent activation function. In an attempt to alleviate the issue of abrupt frame transitions, we use an overlap of 50\,\% between the generated waveform frames, as proposed in \cite{DBLP:conf/icassp/EphratP17}. The overlapped segments are linearly averaged sample by sample in order to maintain the original waveform scale. The detailed architecture of the decoder is shown in Figure \ref{decoder}. \subsection{Critics} As demonstrated in recent works \cite{DBLP:conf/iclr/DonahueMP19,DBLP:conf/nips/KumarKBGTSBBC19,DBLP:journals/corr/abs-1910-11480}, the use of a waveform critic can dramatically increase the realism and clarity of synthesized speech. To discriminate the real from the synthesized waveforms, we adapt the critic from \cite{DBLP:conf/nips/KumarKBGTSBBC19}. After experimenting extensively with and without weight normalization for this module, as well as for the generator, we find that weight normalization increases the stability of adversarial training but overall leads to worse results. Therefore, we remove weight normalization from this critic but otherwise keep the original architecture: 7 convolutional layers, each followed by Leaky ReLU activation, as shown in Figure \ref{wave_critic}. \begin{figure} \centering \includegraphics[width = 0.9\linewidth]{./figures/decoder_rotated.pdf} \caption{Description of the layers in the decoder (generator).} \label{decoder} \end{figure} We did not attempt batch normalization, which worked well for the generator, since this interferes with the gradient penalty for our adversarial loss \cite{DBLP:conf/nips/GulrajaniAADC17}. We compared this architecture to other convolutional critics similar to the one proposed in \cite{DBLP:conf/iclr/DonahueMP19} as well as a one-dimensional ResNet 18, and found that this critic produced the best results. Remarkably, this critic has a far smaller receptive field than any of the critics we experimented with. This may indicate that waveform critics work best when focusing on the small scale. Inspired by the SpecGAN model \cite{DBLP:conf/iclr/DonahueMP19}, we propose to combine the waveform critic, which judges the audio in the temporal domain, with a power critic, which judges the audio in the spectral domain. This module discriminates the spectrograms computed from real and generated audio. We first compute the spectrogram from both the real and generated samples using the short-time Fourier transform (STFT) with a window size of 25\,ms, a hop size of 10\,ms and frequency bins of size 512. We then compute the natural logarithm of the spectrogram magnitudes, normalize these values to mean 0 and variance 1, clip values outside [-3,3] and normalize them to [-1,1], similarly to \cite{DBLP:conf/iclr/DonahueMP19}. In this case, we use a ResNet18 identical to the one presented in our generator, except with a two-dimensional front end convolutional layer in the beginning, since our input is a single image. As with the waveform critic, we cannot use batch normalization in this module due to the gradient penalty, and found that weight normalization did not improve results. The architecture for the power critic is shown in Figure \ref{power_critic}. \subsection{Losses} To train our network, we apply the Wasserstein GAN loss \cite{DBLP:journals/corr/ArjovskyCB17}, which aims to minimize the Wasserstein Distance between the distributions of real and synthesized data. We also add the gradient penalty \cite{DBLP:conf/nips/GulrajaniAADC17} in order to satisfy the Lipschitz constraint in the Wasserstein GAN objective. The losses for the generator and respective critic(s) are defined as: \begin{align}\label{eq:1} &L_{G} = -\underset{\widetilde{x} \sim \mathbb{P}_{G}}{\mathbb{E}}[D(\widetilde{x})] + \lambda \underset{\hat{x}\sim \mathbb{P}_{\hat{x}}}{\mathbb{E}} [(\|\nabla_{\hat{x}} D(\hat{x})\| - 1)^2] \\ &L_{D} = \underset{\widetilde{x} \sim \mathbb{P}_{G}}{\mathbb{E}}[D(\widetilde{x})] - \underset{x \sim \mathbb{P}_{R}}{\mathbb{E}}[D(x)], \end{align} where $G$ is the generator, $D$ is the critic, $x \sim \mathbb{P}_{R}$ are samples from the real distribution, $\widetilde{x} \sim \mathbb{P}_{G}$ are samples from the estimated distribution (produced by the generator) and $\hat{x}\sim \mathbb{P}_{\hat{x}}$ are sampled uniformly between two points from $P_G$ and $P_R$ respectively. In this work, we apply two critics: the waveform critic and the power critic. Each critic is trained with their own losses $L_{D_{wave}}$ and $L_{D_{power}}$, whereas the generator combines the losses from the two critics such that: \begin{align} &L_{G_{adv}} = L_{G_{wave}} + L_{G_{power}}, \end{align} where $L_{G_{wave}}$ and $L_{G_{power}}$ are calculated as mentioned in Eq. \ref{eq:1}. The coefficient for the gradient penalty $\lambda$ is kept at the value of 10 for both critics, as proposed in \cite{DBLP:conf/nips/GulrajaniAADC17}. In addition to this adversarial loss, we also apply three other losses to train the generator. The first is a perceptual loss: \begin{align} L_{PASE} = \| \delta(x) - \delta(\widetilde{x}) \|, \end{align} where $x$ is the real waveform, $\widetilde{x}$ is the synthesized waveform from the same video and $\delta$ is our perceptual feature extractor. In this work, we use the pre-trained PASE model \cite{DBLP:journals/corr/abs-1904-03416} to extract perceptual features $\delta(x)$. PASE has been trained in a self-supervised manner to produce meaningful speech representations. We have also tried using PASE+ \cite{DBLP:journals/corr/abs-2001-09239}, which is an improved version of PASE, however, no improvement in the speech reconstruction quality was observed. Furthermore, we experimented with multiple ASR models as feature extractors, but we found that they also did not improve results. The second loss we apply is the Power Loss. This function aims to improve the accuracy of the reconstructed audio by attempting to match it with the real audio in the frequency domain. For this purpose, we use the L1 loss between the STFT magnitudes of the real and synthesized audio as follows: \begin{align} L_{power} = \| log \|STFT(x)\|^2 - log \|STFT(\widetilde{x})\|^2 \|, \end{align} where $x$ is the real waveform, $\widetilde{x}$ is the synthesized waveform from the same video and $STFT$ is the Short Time Fourier Transform with a window size of 25\,ms, a hop size of 10\,ms and frequency bins of size 512 (same parameters used for the power critic). We found that scaling the magnitudes using the natural logarithm and using an L1 Loss rather than the L2 Loss chosen in \cite{DBLP:conf/icml/OordLBSVKDLCSCG18} greatly improve training stability and performance. The third loss we apply is the MFCC Loss: \begin{align} L_{MFCC} = \| MFCC(x) - MFCC(\widetilde{x}) \|, \end{align} where $x$ is the real waveform, $\widetilde{x}$ is the synthesized waveform from the same video and $MFCC$ is the MFCC function which extracts 25 mel-frequency cepstral coefficients from the corresponding waveform. The objective of this loss lies in increasing the accuracy and intelligibility of the synthesized speech, given that MFCCs are known to be effective in ASR \cite{DBLP:conf/iscas/HanCCP06} and emotion recognition \cite{6514336}. We adapt the function provided on an open-source repository\footnote{https://github.com/skaws2003/pytorch-mfcc}. Finally, the loss for the generator is described based on the losses mentioned above as: \begin{align} L_{G} = \alpha_1 L_{G_{adv}} + \alpha_2 L_{PASE} + \alpha_3 L_{power} + \alpha_4 L_{MFCC}. \end{align} We tune the coefficients $\alpha_{1,2,3,4}$ by sequentially training multiple models on GRID (4 speakers, seen speaker split) and incrementally finding the coefficients that yield the best WER on the validation set. Through our search, we find that $\alpha_1 = 1$, $\alpha_2 = 140$, $\alpha_3 = 50$, $\alpha_4 = 0.4$ yield the best results. \begin{figure} \centering \includegraphics[width = 0.9\linewidth]{./figures/wave_critic_rotated.pdf} \caption{Description of the layers in the waveform critic used to train our model.} \label{wave_critic} \end{figure} \begin{figure} \centering \includegraphics[width = 0.9\linewidth]{./figures/power_critic_rotated.pdf} \caption{Description of the layers in the power critic used to train our model, and the process used to extract spectrograms from waveform samples.} \label{power_critic} \end{figure} \subsection{Training details} We use the Adam optimizer with a learning rate of 0.0001 and $\beta_1 = 0.5$, $\beta_2 = 0.99 $ to train our generator and critics end-to-end. Given that the critics should be trained to completion before every generator training step, we perform 6 training steps on the critics before every training step of the generator. It should also be noted that we feed a one second clip randomly sampled from the real and synthesized audio to each of the critics, rather than the entire utterance. The other losses are computed using the entire real and synthesized utterances. Additionally, we employ two data augmentation methods during training. Firstly, we apply random cropping on the input frame, producing a frame with roughly 90\,\% of the original size. Furthermore, we apply horizontal flipping to each frame with a probability of 50\,\%. These procedures help make our model more robust and provide regularization. During test time, the same cropping is performed on the center of the frame and no horizontal flipping is performed. Training our model for each of the experiments generally takes aprroximately one week on an Nvidia RTX 2080 Ti GPU. Synthesizing a 3 second audio clip sampled at 16\,kHz from 75 frames of video takes approximately 32\,ms on the same high-end GPU, excluding pre-processing. \section{Datasets} \label{section4} \begin{table} \centering \def1.5{1.5} \begin{tabular}{ m{2.2cm} | >{\centering}m{1.9cm} >{\centering}m{1.6cm} >{\centering\arraybackslash}m{1.6cm} } \hline Corpus & Training set (clips / hours) & Validation set (clips / hours) & Test set \break (clips / hours) \\ \hline \hline GRID (4 speakers, seen speakers) & 3576 / 2.98 & 210 / 0.18 & 210 / 0.18 \\ \hline GRID (33 speakers, seen speakers) & 29584 / 24.65 & 1642 / 1.37 & 1641 / 1.37 \\ \hline GRID (33 speakers, unseen speakers) & 15888 / 13.24 & 7000 / 5.83 & 9982 / 8.32 \\ \hline TCD-TIMIT \hfill \break (3 lipspeakers) & 1014 / 1.64 & 57 / 0.09 & 60 / 0.09 \\ \hline LRW (full) & 488763 / 157.49 & 25000 / 8.06 & 25000 / 8.06 \\ \hline FLRW 500 Words & 112811 / 36.35 & 5878 / 1.89 & 5987 / 1.93 \\ \hline FLRW 100 Words & 22055 / 7.11 & 1151 / 0.37 & 1144 / 0.37 \\ \hline FLRW 20 Words & 4347 / 1.40 & 266 / 0.09 & 248 / 0.08 \\ \hline \end{tabular} \caption{Number of speech clips and total number of hours of speech for each dataset used in our study.} \label{datasets} \end{table} For the purpose of this work, we use three separate audiovisual datasets to train and evaluate our model: GRID, TCD-TIMIT and LRW. GRID contains 33 speakers, each uttering 1000 short sentences composed of 6 simple words from a constrained vocabulary of 51 words. It is the most commonly used dataset for video-driven speech reconstruction \cite{DBLP:conf/icassp/EphratP17,DBLP:conf/icassp/AkbariACM18,DBLP:journals/corr/abs-2004-02541,DBLP:conf/icassp/AkbariACM18,DBLP:journals/corr/abs-2005-08209} due to the clean recording conditions and the limited vocabulary. TCD-TIMIT is another audiovisual dataset composed of 62 speakers, three of which are trained lipspeakers. In order to compare with previous works \cite{DBLP:journals/corr/abs-2005-08209}, we only use the audiovisual data uttered by the three lipspeakers. Each lipspeaker utters 375 unique phonetically rich sentences, as well as two additional sentences which are uttered by all three speakers. This results in a total of 1\,131 clips. The video/audio for this data is recorded in studio conditions with exceptional clarity given the particular speaking ability of the professional lipspeakers. Finally, LRW contains roughly 500\,000 speech samples (500 words, up to 1\,000 clips per word) uttered by hundreds of different speakers, taken from television broadcasts. Due to the fact that these utterances are recorded `in the wild' from a large variety of speakers, LRW presents a far more substantial challenge for speech reconstruction than the datasets mentioned above. Additionally, we use a subset of this corpus which keeps only the videos that are approximately frontal, i.\,e., videos with yaw, pitch and roll below 10 degrees. This leads to a corpus containing 124\,676 samples in total and will be referred to as \textit{F(rontal)LRW}. We also randomly select 20/100 words from this subset to experiment with different ranges of vocabulary during training/testing. These smaller sets will be referred to as \textit{FLRW20} and \textit{FLRW100}, respectively. Further statistics for each dataset are presented in Table \ref{datasets}. Rather than using the full face as input to our network, as is standard in other speech reconstruction works \cite{DBLP:conf/iccvw/EphratHP17,DBLP:conf/icassp/AkbariACM18,DBLP:journals/corr/abs-2005-08209}, we crop the mouth of the speaker, and use it as the input for every frame. We do this by performing face detection and alignment using dlib's 68 landmark model \cite{dlib09}, aligning each face to a reference mean face shape and extracting a mouth ROI (Region of Interest) from each frame. The mouth ROI is of size 128x74 for GRID and 96x96 for TCD-TIMIT and LRW. \section{Evaluation Metrics} \label{section5} Although many metrics have been proposed for evaluating the quality of speech \cite{DBLP:series/sci/Loizou11}, it is widely acknowledged that none of the existing metrics are highly correlated with human perception. For this reason, we evaluate our speech reconstruction model using 4 objective metrics which capture different properties of the audio: PESQ, STOI, MCD and WER. PESQ (Perceptual Evaluation of Speech Quality) \cite{DBLP:conf/icassp/RixBHH01} is an objective speech quality metric originally proposed for telephony quality assessment. It consists of a complex series of filters and transforms which result in a speech quality score. For the purposes of our work, we use this metric to measure how clean a speech signal is. STOI (Short-Time Objective Intelligibility measure) \cite{DBLP:journals/taslp/TaalHHJ11} aims to measure how intelligible a speech signal is through a comparative DFT-based (Discrete Fourier Transform) approach. It has been found that it achieves close correlation to human intelligibility scores. In our experiments, we use this metric to measure the intelligibility of the reconstructed samples. MCD (Mel-Cepstral Distance) \cite{407206} is designed to evaluate speech quality based on the cepstrum distance on the mel-scale. In practice, this is calculated as the distance between the MFCCs extracted from two signals. We find that it works quite reliably in measuring perceptual quality in our synthesized outputs, when compared to the original signal. WER (Word Error Rate) measures the accuracy of a speech recognition system. It is calculated as: \begin{align} WER = \frac{S+D+I}{N}, \end{align} where $S$ is the number of substitutions, $D$ is the number of deletions, $I$ is the number of insertions and $N$ is the total number of words in an utterance. For our work, we apply pre-trained ASR models to measure WER, which serves as an objective intelligibility metric for the reconstructed speech. \section{Results on Seen Speakers} \label{section6a} In this section, we present our experiments for seen speakers. For direct comparison with other works we use the same 4 speakers from GRID (1, 2, 4 and 29) as in \cite{DBLP:conf/icassp/AkbariACM18,DBLP:journals/corr/abs-1906-06301,DBLP:journals/corr/abs-2005-08209,DBLP:journals/corr/abs-2004-02541} and the 3 lipspeakers from TCD-TIMIT as in \cite{DBLP:journals/corr/abs-2005-08209}. In order to investigate the impact of the number of speakers and the amount of training data, we also present results for all 33 speakers from the GRID dataset. We split the utterances in each of these datasets using a 90-5-5\,\% ratio for training, validation and testing respectively similarly to \cite{DBLP:conf/icassp/AkbariACM18,DBLP:journals/corr/abs-1906-06301,DBLP:journals/corr/abs-2005-08209,DBLP:journals/corr/abs-2004-02541}, such that the speakers in the validation and test sets are identical to the speakers seen in the training set (but the utterances are different). To measure the Word Error Rate (WER) for our GRID samples, we use a pre-trained ASR model (based on \cite{ma2019investigating}) which was trained and tested on the full GRID dataset (using the split mentioned in Section \ref{section6b}), achieving a baseline of 4.23\,\% WER on the test set. Audio samples, as well as spectrogram and waveform figures are presented on our website\footnote{\label{note1}\url{https://sites.google.com/view/video-to-speech/home}} for the experiments presented in sections \ref{section6a}, \ref{section6b} and \ref{section6c}. Additionally, we present a publicly available repository\footnote{\url{https://github.com/miraodasilva/evalaudio}} which can be used to reproduce each of the evaluation metrics presented in this work. We are also available to provide generated test samples for researchers hoping to reproduce or compare with our work. \subsection{Ablation Study} Results for the ablation study are shown in Table \ref{ablation_seen}. For this study, we only consider the 4 subjects from GRID presented above (1,2,4 and 29). Firstly, we observe that each of the three comparative losses $L_{PASE}$, $L_{power}$ and $L_{MFCC}$ yield considerable improvements in the verbal accuracy of samples (as shown by the WER), even when only one is removed. We can also observe that $L_{MFCC}$ and $L_{power}$ are particularly impactful on the MCD of the reported samples, which is unsurprising since this is an MFCC-based metric. On the other hand, it is clear that $L_{PASE}$ is essential towards achieving high intelligibility, given its particular impact on STOI. Finally, all three losses also seem to positively impact the PESQ score, indicating an increase in overall audio clarity. We can see that the simultaneous removal of $L_{PASE}$ and $L_{power}$ greatly decreases PESQ and STOI, indicating that these losses are particularly important towards the clarity of generated samples. We also show that the absence of $L_{MFCC}$ and $L_{power}$ sharply increases MCD, indicating that these two losses greatly increase the similarity between real and synthesized audio. On the other hand, this model maintains a WER below 10\,\%, which means that $L_{PASE}$ alone (together with the adversarial losses) can achieve intelligible audio. Finally, the removal of all three L1 losses results in realistic yet unintelligible audio. This is because the adversarial losses are the only objective used for training, and therefore there is no incentive for the network to learn the exact words corresponding to the input video. We observe that the use of the waveform critic yields noticeable improvements through our metrics, particularly in WER and STOI, suggesting that its inclusion substantially increases intelligibility. Additionally, the power critic also yields moderate improvements in PESQ, STOI and WER. Finally, we observe that the removal of both critics results in substantially lower MCD and WER, but mantains PESQ and STOI at a similar value. This again indicates that our model can generate intelligible and accurate words without the adversarial losses. However, these synthesized samples lack realism, which drastically improves when the critics are used. To demonstrate this effect, readers are encouraged to listen to examples on our website$^{\ref{note1}}$. We also experiment with using the full face as input, as this is commonly used in previous studies. Through this ablation, we show that using a cropped mouth region instead of the full face improves our results substantially regarding WER, effectively improving intelligibility. We also prove that the use of overlap improves all metrics slightly, suggesting that its purpose of minimizing the issue of frame transitions is benefitial towards output quality. A qualitative comparison with other works can be seen in Figure \ref{ablation_qualitative}. Compared to the real audio, our spectrogram is similar overall, but is slightly blurrier and fails to model some of the fine details in the frequency bins, especially in the higher frequencies. The model trained without adversarial critics features a much blurrier spectrogram than the full model, failing to reproduce even the lower frequency bands during voiced speech, highlighting the importance of adversarial training. \begin{table} \centering \def1.5{1.5} \begin{tabular}{ m{3.5cm} | >{\centering}m{0.8cm} >{\centering}m{0.8cm} >{\centering}m{0.8cm} >{\centering\arraybackslash}m{0.8cm} } \hline Model & PESQ & STOI & MCD & WER \\ \hline \hline w/o $L_{PASE}$ & 2.06 & \textbf{0.597} & \textbf{26.44} & 8.97\,\% \\ \hline w/o $L_{power}$ & 2.05 & 0.575 & 28.64 & 9.54\,\% \\ \hline w/o $L_{MFCC}$ & 2.08 & 0.591 & 28.09 & 9.09\,\% \\ \hline w/o $L_{PASE}$, w/o $L_{power}$ & 1.86 & 0.545 & 27.47 & 13.44\,\% \\ \hline w/o $L_{PASE}$, w/o $L_{MFCC}$ & 2.02 & 0.589 & 28.82 & 13.33\,\% \\ \hline w/o $L_{MFCC}$, w/o $L_{power}$ & 2.00 & 0.569 & 31.43 & 9.71\,\% \\ \hline w/o $L_{PASE}$, w/o $L_{power}$, w/o $L_{MFCC}$ & 1.14 & 0.311 & 53.63 & 89.12\,\% \\ \hline w/o waveform critic & 2.07 & 0.583 & 26.66 & 8.47\,\% \\ \hline w/o power critic & 2.08 & 0.594 & 26.73 & 7.30\,\% \\ \hline w/o waveform critic, w/o power critic & 2.07 & 0.584 & 27.45 & 9.01\,\% \\ \hline w/o overlap & 2.06 & 0.590 & 26.73 & 7.40\,\% \\ \hline w/ full face & 2.07 & 0.596 & 26.46 & 9.94\,\% \\ \hline full model & \textbf{2.10} & 0.595 & 26.78 & \textbf{7.03}\,\% \\ \hline \end{tabular} \caption{Ablation study performed on GRID for seen speaker speech reconstruction.} \label{ablation_seen} \end{table} \begin{figure*} \centering \begin{subfigure}{0.3\linewidth} \includegraphics[width=\linewidth]{./figures/nocritic_spectrogram.pdf} \caption{w/o Wave Critic, w/o Power Critic} \end{subfigure} \begin{subfigure}{0.3\linewidth} \includegraphics[width=\linewidth]{figures/baseline_spectrogram.pdf} \caption{Full Model} \end{subfigure} \begin{subfigure}{0.3\linewidth} \includegraphics[width=\linewidth]{figures/real_spectrogram.pdf} \caption{Real Audio} \end{subfigure} \caption{Mel-frequency spectrograms taken from the audio reconstructed with our seen speaker ablation models. The clip we present is from GRID, speaker 1, utterance 'Bin blue at L 9 again'.} \label{ablation_qualitative} \end{figure*} \subsection{Comparison with Other Works} We compare our proposed model with previous works on the commonly used 4 GRID speakers as shown in Table \ref{grid_seen_comparison}. We note that the metrics reported on Lip2Wav \cite{DBLP:journals/corr/abs-2005-08209} are taken directly from their paper due to test samples not being publicly available, and that their WER was calculated using the Google Speech-to-Text (STT) API rather than our ASR model. Regarding PESQ, it is clear that our model is superior to the previous approaches by a sizeable margin. This suggests that the quality of our synthesized speech is somewhat higher than past models. Our model also outperforms previous works on STOI, excluding Lip2Wav. This shows that our samples are more intelligible than most other approaches, but are outperformed by the robustness and consistency of the speech produced by Lip2Wav. Furthermore, our generated samples achieve a better MCD than previous works, indicating that our reconstructed audio is more accurate than previous approaches on the frequency domain. Finally, our work achieves the best WER out of all methods, which shows that our model is more accurate than any of the previous approaches by a large factor, outperforming our previous model by more than 10\,\%. A qualitative comparison is shown in Figure \ref{qualitative_comparison}, which displays waveforms, mel-frequency spectrogram, and mel-frequency spectrogram differences, i.\,e., the element-wise absolute difference between the real and synthesized spectrograms. This difference is calculated as: \begin{align} \| MelSpec(x) - MelSpec(\widetilde{x}) \|, \end{align} where $x$ is the real waveform and $\widetilde{x}$ is the synthesized waveform. Through the spectrograms, it is clear that Lip2Audspec is the least accurate in the frequency domain, failing to model many frequencies, particularly in the higher bands. The other three approaches are clearly more accurate, but all feature some inaccuracies during voiced speech and also noise in unvoiced segments. While \cite{DBLP:journals/corr/abs-1906-06301} and \cite{DBLP:journals/corr/abs-2004-02541} feature an excessive amount of low frequency noise, our model seems to accurately emulate the low amount of noise in the real audio and therefore achieves the least substantial spectrogram difference. We also compare our model to Lip2Wav on TCD-TIMIT (3 lipspeakers) in Table \ref{timit_comparison}. Once more, it is clear that our model outperforms Lip2Wav \cite{DBLP:journals/corr/abs-2005-08209} on PESQ, but achieves lower performance on STOI, which indicates that our model produces clearer, yet somewhat less intelligible audio. Additionally, our samples achieve a reasonably low MCD, indicating moderate similarity in the frequency domain. \begin{figure*} \centering \begin{subfigure}{\linewidth} \centering \begin{minipage}[c]{0.01\linewidth} \centering \caption{} \end{minipage} \begin{minipage}[c]{0.95\linewidth} \centering \begin{subfigure}{0.3\linewidth} \includegraphics[width=\linewidth]{./figures/l2as_spec.pdf} \end{subfigure} \begin{subfigure}{0.3\linewidth} \includegraphics[width=\linewidth]{./figures/l2as_spec_diff.pdf} \end{subfigure} \begin{subfigure}{0.3\linewidth} \includegraphics[width=\linewidth]{./figures/l2as_wav.pdf} \end{subfigure} \end{minipage} \end{subfigure} \begin{subfigure}{\linewidth} \centering \begin{minipage}[c]{0.01\linewidth} \centering \caption{} \end{minipage} \begin{minipage}[c]{0.95\linewidth} \centering \begin{subfigure}{0.3\linewidth} \includegraphics[width=\linewidth]{./figures/dino_spec.pdf} \end{subfigure} \begin{subfigure}{0.3\linewidth} \includegraphics[width=\linewidth]{./figures/dino_spec_diff.pdf} \end{subfigure} \begin{subfigure}{0.3\linewidth} \includegraphics[width=\linewidth]{./figures/dino_wav.pdf} \end{subfigure} \end{minipage} \end{subfigure} \begin{subfigure}{\linewidth} \centering \begin{minipage}[c]{0.01\linewidth} \centering \caption{} \end{minipage} \begin{minipage}[c]{0.95\linewidth} \centering \begin{subfigure}{0.3\linewidth} \includegraphics[width=\linewidth]{./figures/michelsanti_spec.pdf} \end{subfigure} \begin{subfigure}{0.3\linewidth} \includegraphics[width=\linewidth]{./figures/michelsanti_spec_diff.pdf} \end{subfigure} \begin{subfigure}{0.3\linewidth} \includegraphics[width=\linewidth]{./figures/michelsanti_wav.pdf} \end{subfigure} \end{minipage} \end{subfigure} \begin{subfigure}{\linewidth} \centering \begin{minipage}[c]{0.01\linewidth} \centering \caption{} \end{minipage} \begin{minipage}[c]{0.95\linewidth} \centering \begin{subfigure}{0.3\linewidth} \includegraphics[width=\linewidth]{./figures/ours_spec.pdf} \end{subfigure} \begin{subfigure}{0.3\linewidth} \includegraphics[width=\linewidth]{./figures/ours_spec_diff.pdf} \end{subfigure} \begin{subfigure}{0.3\linewidth} \includegraphics[width=\linewidth]{./figures/ours_wav.pdf} \end{subfigure} \end{minipage} \end{subfigure} \begin{subfigure}{\linewidth} \hspace{0.5em} \begin{minipage}[c]{0.01\linewidth} \centering \caption{} \end{minipage} \begin{minipage}[c]{0.95\linewidth} \centering \begin{subfigure}{0.3\linewidth} \includegraphics[width=\linewidth]{./figures/real_spec.pdf} \end{subfigure} \begin{subfigure}{0.3\linewidth} \includegraphics[width=\linewidth]{./figures/real_spec_diff.pdf} \end{subfigure} \begin{subfigure}{0.3\linewidth} \includegraphics[width=\linewidth]{./figures/real_wav.pdf} \end{subfigure} \end{minipage} \end{subfigure} \caption{Mel-frequency spectrograms (left), Mel-frequency spectrogram differences (middle) and waveforms (right) taken from the audio reconstructed with Lip2AudSpec \cite{DBLP:conf/icassp/AkbariACM18} (a), our previous work \cite{DBLP:journals/corr/abs-1906-06301} (b), a previous vocoder-based model \cite{DBLP:journals/corr/abs-2004-02541} (c) and our model (d), as well as the real audio (e) -- GRID, Speaker 1, utterance 'Bin white at T 3 soon'. All models were trained on the same split of GRID (4 speakers, seen speaker split), as presented in our comparison.} \label{qualitative_comparison} \end{figure*} \begin{table} \centering \def1.5{1.5} \begin{minipage}{\linewidth} \begin{tabular}{ m{2.5cm} | >{\centering}m{1cm} >{\centering}m{1cm} >{\centering}m{1cm} >{\centering\arraybackslash}m{1.1cm} } \hline Method & PESQ & STOI & MCD & WER \\ \hline \hline Lip2Audspec \cite{DBLP:conf/icassp/AkbariACM18} & 1.81 & 0.425 & 63.88 & 46.36\,\% \\ \hline GAN-based \cite{DBLP:journals/corr/abs-1906-06301} & 1.70 & 0.539 & 45.37 & 21.11\,\% \\ \hline Vocoder-based \cite{DBLP:journals/corr/abs-2004-02541} & 1.90 & 0.553 & 46.64 & 22.14\,\% \\ \hline Lip2Wav \cite{DBLP:journals/corr/abs-2005-08209} & 1.77 & \textbf{0.731} & - & 14.08\footnote{Reported using Google STT API.}\,\% \\ \hline Ours & \textbf{2.10} & 0.595 & \textbf{26.78} & \textbf{7.03\,\%} \\ \hline \end{tabular} \end{minipage} \caption{Comparison between our model and previous works, using the GRID subset (4 speakers) with a seen speaker split. } \label{grid_seen_comparison} \end{table} \begin{table} \centering \def1.5{1.5} \begin{tabular}{ m{2.5cm} | >{\centering}m{1cm} >{\centering}m{1cm} >{\centering\arraybackslash}m{1.1cm} } \hline Method & PESQ & STOI & MCD \\ \hline \hline Lip2Wav \cite{DBLP:journals/corr/abs-2005-08209} & 1.35 & \textbf{0.558} & - \\ \hline Ours & \textbf{1.61} & 0.295 & \textbf{32.12} \\ \hline \end{tabular} \caption{Comparison between our model and Lip2Wav, using TCD-Timit (3 lipspeakers) with a seen speaker split.} \label{timit_comparison} \end{table} \subsection{Performance as a Function of Training Set Size} For the purposes of this study, we use all 33 subjects from GRID and we report results as we vary the size of the training set from 20\,\% to 100\,\% in steps of 20\,\%. Results are shown in Table \ref{varying_training_set}. When compared to the results reported for GRID (4 speakers, seen split), we observe comparable performance for 33 speakers when using the full training set. This shows that our network adapts well to larger datasets and is able to model a large amount of speakers with no substantial drop in performance. Regarding the models which are trained using a smaller subset of the training set, it is clear that the performance drops as the amount of training data is gradually reduced. However, it is worth highlighting that the overall performance remains moderately consistent, even when we use only 20\,\% of the training data. This shows that our model adapts well to smaller datasets. We note that all 5 models were trained for the same amount of total training steps to avoid any bias in our comparative results. \begin{table} \centering \def1.5{1.5} \begin{tabular}{ m{2cm} | >{\centering}m{1cm} >{\centering}m{1cm} >{\centering}m{1cm} >{\centering\arraybackslash}m{1.1cm} } \hline \% of Training Set & PESQ & STOI & MCD & WER \\ \hline \hline 20\,\% & 1.96 & 0.583 & 29.22 & 11.78\,\% \\ \hline 40\,\% & 2.00 & 0.594 & 28.49 & 10.10\,\% \\ \hline 60\,\% & 2.02 & 0.595 & 27.94 & 9.06\,\% \\ \hline 80\,\% & 2.02 & 0.596 & 27.68 & 8.36\,\% \\ \hline 100\,\% & 2.02 & 0.601 & 27.78 & 8.03\,\% \\ \hline \end{tabular} \caption{Study on the performance of our speech reconstruction model using varying training set sizes, using the full GRID seen speaker split mentioned in Section \ref{section4}.} \label{varying_training_set} \end{table} \section{Results on Unseen Speakers} \label{section6b} In this section, we investigate the performance of the proposed approach on unseen speakers. For the purposes of this study, we use all speakers from the GRID dataset, using a 50-20-30\,\% split ratio similarly to \cite{DBLP:journals/corr/abs-1906-06301,DBLP:journals/corr/abs-2004-02541}, such that there is no overlap between the speakers featured in the training, validation and test sets. To measure WER, we use the GRID pre-trained model mentioned in the previous section. \subsection{Ablation Study} In this study, we use all 33 GRID speakers. The results for the ablation study are shown in Table \ref{ablation_unseen}. For this, task, we find that $L_{power}$ provides the greatest impact on the quality of results, providing a substantial improvement in all metrics. On the other hand, $L_{PASE}$ and $L_{MFCC}$ show noticeable improvements in PESQ and STOI, indicating that these losses contribute to the clarity and intelligibility of the generated samples. Furthermore, we once more find that $L_{MFCC}$ and $L_{power}$ are particularly important towards achieving a low MCD, meaning that these losses are essential towards achieving accurate MFCCs in our synthesized samples. Regarding the adversarial loss, we can see that, as reported in the seen speaker ablation, PESQ, STOI and MCD improve with the addition of the waveform and power critics. This suggests that these critics have a positive effect on the clarity and intelligiblity of samples, and that the accuracy on the frequency domain is improved as well. However, we observe that the WER remains at a similar value with the removal of both critics, indicating that the network is generally capable of reproducing the correct words from the corresponding video samples while relying only on the three proposed L1 losses. \begin{table} \centering \def1.5{1.5} \begin{tabular}{ m{2.5cm} | >{\centering}m{1cm} >{\centering}m{1cm} >{\centering}m{1cm} >{\centering\arraybackslash}m{1.1cm} } \hline Model & PESQ & STOI & MCD & WER \\ \hline \hline w/o $L_{PASE}$ & 1.44 & 0.520 & 38.19 & 22.66\,\% \\ \hline w/o $L_{power}$ & 1.37 & 0.503 & 39.59 & 24.32\,\% \\ \hline w/o $L_{MFCC}$ & 1.44 & 0.518 & 39.03 & \textbf{21.70}\,\% \\ \hline w/o Waveform Critic, w/o Power Critic & 1.43 & 0.516 & 38.48 & 22.82\,\% \\ \hline Full Model & \textbf{1.47} & \textbf{0.523} & \textbf{37.91} & 23.13\,\%\\ \hline \end{tabular} \caption{Ablation study performed on GRID for unseen speaker speech reconstruction.} \label{ablation_unseen} \end{table} \subsection{Comparison with Other Works} We present our comparison with other works \cite{DBLP:journals/corr/abs-1906-06301,DBLP:journals/corr/abs-2004-02541} on the subject-independent split of GRID in Table \ref{grid_unseen_comparison}. It is clear that our model outperforms previous works in all performance measures. Although, the improvement in PESQ and STOI compared to these works is not as emphatic as the gains reported for seen speakers, WER sees a very substantial reduction. This improvement in WER can easily be observed in our synthesized speech, and clearly shows that our model is far more consistent for this task than previous approaches. Furthermore, the observed MCD is substantially lower in our work, indicating that our synthesized speech yields more accurate spectrograms, which suggests a greater similarity between the content of real and synthesized samples. \begin{table} \centering \def1.5{1.5} \begin{tabular}{ m{2.5cm} | >{\centering}m{1cm} >{\centering}m{1cm} >{\centering}m{1cm} >{\centering\arraybackslash}m{1.1cm} } \hline Method & PESQ & STOI & MCD & WER \\ \hline \hline GAN-based \cite{DBLP:journals/corr/abs-1906-06301} & 1.24 & 0.470 & 51.28 & 37.10\,\% \\ \hline Vocoder-based \cite{DBLP:journals/corr/abs-2004-02541} & 1.23 & 0.477 & 55.02 & 55.23\,\% \\ \hline Ours & \textbf{1.47} & \textbf{0.523} & \textbf{37.91} & \textbf{23.13}\,\% \\ \hline \end{tabular} \caption{Comparison between our current and previous model, using full GRID (33 speakers) with an unseen speaker split.} \label{grid_unseen_comparison} \end{table} \subsection{Additional Experiments} Additionally, we present a study on silent speakers. For this experiment, we artificially produce a video of a speaker from the GRID corpus being silent for five seconds by feeding Brownian noise into the facial animation model proposed in \cite{DBLP:journals/ijcv/VougioukasPP20}. We then use this video as input for our model trained on the full GRID dataset (33 speakers, unseen speaker split). This aims to measure two distinct factors: firstly, our model's ability to recognize a silent speaker and not produce any voiced speech; and secondly, the baseline noise that is present in the audio we synthesize with our network, which is clear to observe when the speaker is silent. As discussed in Figure \ref{silent_experiment}, our model performs well in this scenario and produces minimal noise for this silent example. \begin{figure*} \centering \begin{subfigure}{0.3\linewidth} \includegraphics[width=\linewidth]{figures/silent_video.pdf} \caption{Silent Video} \end{subfigure} \begin{subfigure}{0.3\linewidth} \includegraphics[width=\linewidth]{figures/silent_waveform.pdf} \caption{Waveform} \end{subfigure} \begin{subfigure}{0.3\linewidth} \includegraphics[width=\linewidth]{figures/silent_spectrogram.pdf} \caption{Mel-frequency spectrogram} \end{subfigure} \caption{The spectrogram and waveform for the audio produced by our model for a video of a silent speaker (Speaker 2 from GRID) are portrayed in (a). As displayed in the waveform (b), the audio is almost completely silent, disregarding some low frequency noise which is higlighted in the spectrogram (c). This shows that our model is robust to the scenario of silent speakers and produces minimal baseline noise under these circumstances. This audio sample is also available on our website$^{\ref{note1}}$.} \label{silent_experiment} \end{figure*} \section{Results in the Wild} \label{section6c} In this section, we investigate the performance of the proposed approach on utterances recorded `in the wild'. For this purpose, we use the full LRW dataset, and its subsets FLRW 500 Words, FLRW 100 Words and FLRW 20 Words, which are introduced in Section \ref{section4}. We split the utterances using the default split for LRW (90-5-5\,\% ratio), such that there is no overlap between the utterances in the training, validation and test sets. To measure the Word Error Rate (WER) for our samples, we use a pre-trained model (based on \cite{DBLP:conf/icassp/PetridisSMCTP18}) which was trained and tested on full LRW using the same split, and achieve a baseline WER of 1.68\,\% on the test set. \subsection{Comparison with Other Works} Our comparison with Lip2Wav \cite{DBLP:journals/corr/abs-2005-08209} on LRW (500 Words) is presented in Table \ref{lrw_comparison}. We compare our model to Lip2Wav on LRW (500 Words), in order to compare our model's performance ``in the wild'' to this recent work. Our work shows a great improvement in PESQ compared to Lip2Wav, which suggests that our samples are able to achieve a superior clarity in this regard. On the other hand, our STOI is very similar to the one reported in Lip2Wav, achieving a slight edge which could indicate a minor improvement in intelligibility. \begin{table} \centering \def1.5{1.5} \begin{minipage}{\linewidth} \begin{tabular}{ m{2.5cm} | >{\centering}m{1cm} >{\centering}m{1cm} >{\centering}m{1cm} >{\centering\arraybackslash}m{1.1cm} } \hline Method & PESQ & STOI & MCD & WER \\ \hline \hline Lip2Wav \cite{DBLP:journals/corr/abs-2005-08209} & 1.20 & 0.543 & - & \textbf{34.20}\footnote{Reported using Google STT API.}\,\% \\ \hline Ours & \textbf{1.45} & \textbf{0.556} & \textbf{39.32} & 42.51\,\% \\ \hline \end{tabular} \end{minipage} \caption{Comparison between our model and Lip2Wav, using the full LRW dataset.} \label{lrw_comparison} \end{table} \subsection{Performance for Different Subsets} In order to demonstrate our model's ability to reconstruct speech in less constrained conditions, we experiment with the LRW dataset, as well as some of its subsets. These subsets present increasing degrees of challenge, culminating with the full LRW dataset which presents the greatest challenge given its large vocabulary and large variance in video perspective. Regarding the experiments with frontal LRW, we observe that our model maintains a similar overall quality of outputs for larger vocabularies, as demonstrated by the consistency in PESQ, STOI and MCD. However, it is clear that the more difficult task presented by larger vocabularies yields a decrease in the average accuracy of samples, shown by the increasing WER. This implies that our model scales well with larger datasets, but has difficulties in adapting to larger vocabularies in very unconstrained and inconsistent environments. Even still, the word error rate reported for FLRW 20 Words is noticeably low, implying that our model can realistically reconstruct speech for hundreds of different speakers, even under such `wild' conditions. Finally, we found that the full LRW dataset yields a better performance than our full frontal subset (FLRW 500 Words). Although we expected the frontal data to provide an easier task for the network during training and testing, this result shows that the network benefits strongly from a larger training set, even if the visual data is less consistent. \begin{table} \centering \def1.5{1.5} \begin{tabular}{ m{2.5cm} | >{\centering}m{1cm} >{\centering}m{1cm} >{\centering}m{1cm} >{\centering\arraybackslash}m{1.1cm} } \hline Corpus & PESQ & STOI & MCD & WER \\ \hline \hline FLRW 20 Words & 1.43 & 0.523 & 43.87 & 25.00\,\% \\ \hline FLRW 100 Words & 1.40 & 0.528 & 41.56 & 36.54\,\% \\ \hline FLRW 500 Words & 1.44 & 0.555 & 39.72 & 44.28\,\% \\ \hline LRW 500 Words & 1.45 & 0.556 & 39.32 & 42.51\,\% \\ \hline \end{tabular} \caption{Study on the performance of our speech reconstruction model for the three subsets of LRW mentioned in Section \ref{section4}, as well as the full LRW dataset.} \label{lrw_results} \end{table} \section{Conclusion} \label{section7} In this work, we have presented our end-to-end video-to-waveform synthesis model using a generative adversarial network with two critics on waveform and spectrogram. First, we showed through an ablation study on GRID that the use of our losses, adversarial critics and other choices in training methodology provide a positive impact on the quality of our results for both seen and unseen speaker video-to-speech. Furthermore, we demonstrated through our experiments on LRW that our model is able to generate intelligible speech for videos recorded entirely in the wild by hundreds of different speakers. Finally, we compared our model to previous video-to-speech models and found that it produces the best results on most metrics for GRID and LRW and achieves state-of-the-art performance on PESQ for TCD-TIMIT. We observed that the choice of good critics as well as adequate comparative losses is fundamental towards obtaining realistic results. Therefore, we believe that the pursuit of alternative loss functions (including different adversarial losses) is a promising option for future work. Additionally, we believe that there would be substantial benefit in experimenting with a speaker embedding as input to the generator, in addition to the video, in order to generalize to unseen speakers with a more accurate voice profile, as proposed in \cite{DBLP:conf/icassp/ShenPWSJYCZWRSA18,DBLP:journals/corr/abs-2005-08209}. Finally, extending our model towards other practical applications such as speech inpainting i.\,e., reconstructing missing audio segments in an audiovisual stream, would be a promising research pursuit in order to show the empirical value of video-to-speech synthesis. \section*{Acknowledgments} All datasets used in the experiments and all training, testing and ablation studies have been contacted at Imperial College. Rodrigo Mira would like to thank Samsung for their continued support of his work on this project. Additionally, the authors would like to than AWS for providing cloud computation resources for the experiments discussed in this paper. \ifCLASSOPTIONcaptionsoff \newpage \fi \section*{References} \AtNextBibliography{\footnotesize} \printbibliography[heading=none] \iffalse \begin{IEEEbiography}[{\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{portraits/rodrigo.jpg}}]{Rodrigo Mira} is a PhD at the IBUG research group at Imperial College London under the supervision of Prof. Björn Schuller and Prof. Maja Pantic. He received his BSc in Computer Science Engineering from Instituto Superior Técnico in Portugal in 2017 and his MSc in Advanced Computing from Imperial College London in 2018. For his MSc thesis, he worked in Music Composition using Deep Learning under the supervision of Prof. Björn Schuller and Dr. Eduardo Coutinho. His research interests currently include speech synthesis as well as other generative tasks, as well as self-supervised learning in multimodal contexts. \end{IEEEbiography} \begin{IEEEbiography}[{\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{portraits/dino_portrait.png}}]{Konstantinos Vougioukas} received his Diploma in Electrical and Computer Engineering from the University of Patras in 2011 and his M.Sc. in Artificial Intelligence from the University of Edinburgh in 2012. He has worked for 5 years as a software engineer in the field of mobile underwater robotics. He has worked as a Research Assistant at the iBUG group and is now doing his Ph.D. under the supervision of Prof. Maja Pantic. \end{IEEEbiography} \begin{IEEEbiography}[{\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{portraits/pingchuan_portrait.jpeg}}]{Pingchuan Ma} received his BSc degree from Beihang University in 2015 and his MSc degree in Machine Learning from Imperial College London in 2017. Currently, he is a PhD student at the iBUG group under the supervision of Prof. Maja Pantic and Dr. Stavros Petridis. \end{IEEEbiography} \begin{IEEEbiography}[{\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{portraits/stavros.jpg}}]{Stavros Petridis} is a research associate at the Department of Computing at Imperial Collegel London.He received his BSc degree in electrical and computer engineering from the Aristotle University of Thessaloniki, Greece in 2004 and his MSc degree in Advanced Computing and Ph.D. from Imperial College London in 2005 and 2012, respectively. He has been a research intern in the Image Processing Group at University College London and the Field Robotics Centre, Robotics Institute, Carnegie Mellon University and a visiting researcher at the affect analysis group at University of Pittsburgh. His research interests lie in the areas of pattern recognition and machine learning and their application to multimodal recognition of human non-verbal behaviour and non-linguistic vocalisations. He is currently working on deep learning approaches for audiovisual fusion. He is a member of the IEEE. \end{IEEEbiography} \begin{IEEEbiography}[{\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{portraits/bjorn_portrait.jpg}}]{Björn W. Schuller} received his diploma, doctoral degree, habilitation, and Adjunct Teaching Professor in Machine Intelligence and Signal Processing all in EE/IT from TUM in Munich/Germany. He is Full Professor of Artificial Intelligence and the Head of GLAM at Imperial College London/UK, Full Professor and Chair of Embedded Intelligence for Health Care and Wellbeing at the University of Augsburg/Germany, co-founding CEO and current CSO of audEERING – an Audio Intelligence company based near Munich and in Berlin/Germany, and permanent Visiting Professor at HIT/China amongst other Professorships and Affiliations. Previous stays include Full Professor at the University of Passau/Germany, and Researcher at Joanneum Research in Graz/Austria, and the CNRS-LIMSI in Orsay/France. He is a Fellow of the IEEE and Golden Core Awardee of the IEEE Computer Society, Fellow of the ISCA, President-Emeritus of the AAAC, and Senior Member of the ACM. He (co-)authored 900+ publications (30k+ citations, h-index=83), is Field Chief Editor of Frontiers in Digital Health and was Editor in Chief of the IEEE Transactions on Affective Computing amongst manifold further commitments and service to the community. His 30+ awards include having been honoured as one of 40 extraordinary scientists under the age of 40 by the WEF in 2015. He served as Coordinator/PI in 15+ European Projects, is an ERC Starting Grantee, and consultant of companies such as Barclays, GN, Huawei, or Samsung. \end{IEEEbiography} \begin{IEEEbiography}[{\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{portraits/maja.jpg}}]{Maja Pantic} is a professor in affective and behavioral computing in the Department of Computing at Imperial College London, UK. She was the Research Director of Samsung AI Centre, Cambridge, UK from 2018 to 2020 and is currently an AI Scientific Research Lead at Facebook London. She currently serves as an associate editor for both the IEEE Transactions on Pattern Analysis and Machine Intelligence and the IEEE Transactions on Affective Computing. She has received various awards for her work on automatic analysis of human behavior,including the Roger Needham Award 2011. She is a fellow of the UK’s Royal Academy of Engineering, the IEEE, and the IAPR \end{IEEEbiography} \fi \end{document}

703f4ab69ae5a11f35c9020950069e20044f83f8

\section{Introduction} With the rapid development of internet technology (IT) and internet of things (IoT) technology, cybersecurity issues are attached more and more importance these years due to people's increasing reliance on internet. About 2 million to 5 million computers worldwide suffered from malware (are infected with malware) every day, estimating by the domestic network security company Jinshan in 2016. Cybersecurity, as a new subject for research, has received extensive attention from the academic community. The emerging research field 'cybersecurity dynamics' \cite{XuCybersecurityDynamicsHotSoS2014,XuBookChapterCD2019} is an interdisciplinary field, conceived from the methodology of several early studies in {\em biological epidemiology} (e.g., \cite{McKendrick1926,Kermack1927,Bailey1975,Anderson1991,HethcoteSIAMRew00}) and its variants in {\em cyber epidemiology} (e.g., \cite{KephartOkland91,KephartOkland93,Pastor2001,Moreno2002}) {\em interacting particle systems} \cite{Liggett1985}, and {\em microfoundation in economics} \cite{Hoover2010}. Different from the classical researches oriented to specific tools, such as Cryptography and Database Security, cybersecurity dynamics studies the offensive and defensive models under various circumstances from a whole-network perspective. \subsection{Our Contributions} \label{sec:our-contributions} In this paper, we investigate how to control the evolution of cybersecurity dynamics more efficiently and effectively in the context of {\em preventive and reactive cyber defense dynamics}, and guarantee its globally convergence to a safe state. When it comes to the control problem of the highly nonlinear network dynamics system, the traditional method is to use a single control strategy, that is, to adjust the dynamic parameters and maintain them, so that the security dynamics of the network space converges to a safe state globally. However, maintaining a high level of dynamic defense strategy may result in high control costs. Such excessive prevention and control will lead to a waste of control resources to some extent, and may even have negative effects on the stable operation of the network systems. In order to solve this problem, this paper proposes an event-based parameter switching approach to save control resources and control the evolution of cybersecurity dynamics in a decentralized manner. It is also proved that this approach is Zeno-free, that is, it will not fall victim to the Zeno behavior. Numerical examples show that the maintenance hours of high-cost control strategies can be reduced by more than $40\%$ with our parameter switching approach. In addition, this paper provides an estimation method to bridge the gap between the probability state and the sampling state when using the method in practice. Different from the equilibrium state estimation problem\cite{liu2020using}, the event-based control method requires considering the timeliness of probability estimation. A new adaptive estimation method is proposed to verify the effectiveness of the event-based parameter switching method in the control process through numerical examples. \ \subsection{Related Work} \label{sec:related-work} Similar event-based methods have been employed in many other application settings before (see, for example, \cite{Astrom2002,Tabuada2007,6425820}). In practical application, one of the essential problems is that this method should not fall victim to the Zeno behavior, which can lead to infinitely many events within a finite period of time, thus invalidate the method \cite{Johansson1999}. The importance of assuring Zeno-freeness in event-triggered control is witnessed by a body of literature, including \cite{Tabuada2007,Johannesson2007,WangLemmon2011,Dimarogonas2012}. The first cybersecurity dynamics model was proposed in \cite{XuAINA07}. Ten years later, \cite{ZhengSub} demonstrated that a certain class of cybersecurity dynamics is globally convergent in the entire parameter universe, which laid a foundation for our further research. The notion of cybersecurity dynamics, as discussed in \cite{XuCybersecurityDynamicsHotSoS2014,XuBookChapterCD2019}, has opened the door to a new research field. The preventive and reactive cyber defense dynamics is a particular kind of cybersecurity dynamics. Several other kinds of cybersecurity dynamics have been demonstrated in early studies, such as the models aiming to accommodate adaptive defenses \cite{XuTAAS2014}, active defenses \cite{XuHotSoS15,XuInternetMath2015ACD,XuGameSec13}, and proactive defenses \cite{XuHotSOS14-MTD}. It is worth mentioning that the event-based parameter switching method may be extended and applied to the various kinds of dynamics. \subsection{Paper Outline} In Section \ref{sec:model}, we briefly review the preventive and reactive cyber defense dynamics model and its global convergence in the entire parameter universe \cite{ZhengSub}. And then we state the problems we addressed in this paper. In Sections \ref{sec:proof-of-problem}, we present an event-based parameter switching method for controlling the cybersecurity dynamics, and prove the effectiveness of the control method (without Zeno behavior). In Section \ref{sec:simulation}, we present numerical examples on the theoretical model. Then in Section \ref{sec:gap}, we show how to apply this event-based parameter switching method in practice by bridging the gap between the probability-state in the theoretical model and the sample-state in practice, using an stochastic process method with adaptive time windows. In Section \ref{sec:conclusion}, we conclude the paper with open problems. \section{Problem Statement} \label{sec:model} \subsection{Review of Preventive and Reactive Defense Dynamics} As a particular kind of cybersecurity dynamics, the preventive and reactive defense dynamics model is first introduced in \cite{XuAINA07} and the convergence properties of the dynamics is studied in \cite{XuTAAS2012} . Later \cite{ZhengSub} fully analyzed the convergence issues, not only considered the common situation with node homogeneity (i.e., the parameters are node-independent), but also a more general situation with node heterogeneity (i.e., the parameters are nodes-dependent). The paper proved that this dynamics model is globally convergent in the entire parameter universe, that is, there is always a unique equilibrium, whose exact value (or position) depends on the specific parameter values instead of the initial state of the dynamics. In the preventive and reactive defense dynamics model, we consider two classes of defenses: {\em preventive defenses} and {\em reactive defenses}, and two classes of attacks: {\em push-based attacks} and {\em pull-based attacks}. Suppose that the attack-defense interaction occurs over an {\em attack-defense} graph structure $G=(V,E)$, where $V$ is the vertex set representing computers and $(u,v)\in E$ means computer $u$ can directly attack computer $v$ using push-based attack strategy (i.e., the communication from $u$ to $v$ is allowed by the security policy). $G$ can be derived from the security policy of a networked system and the physical network in question. Without loss of generality, we do {\em not} make any restrictions on the structure of $G$ (e.g., $G$ may be directed or undirected). Denote the adjacency matrix of $G$ by $A=[a_{vu}]_{n\times n}$, where $a_{vu}=1$ if and only if $(u,v) \in E$. Since the model aims to describe the attacks between computers, we set $a_{vv}=0$. Let $N_{v}=\{u\in V:~(u,v)\in E\}$. In this paper, we consider the continuous-time model described in \cite{ZhengSub}. At any time point, a node $v \in V$ is in one of two states: ``0'' means {\em secure}\ but vulnerable, or ``1'' means {\em compromised}. Let $s_v(t)$ and $i_v(t)$ be the probability that $v$ is {\em secure} and {\em compromised} at time $t$ respectively. It is obvious that $s_v(t)+i_v(t)=1$, $s_v(t)$ and $i_v(t)$ explain the term {\em probability-state}. For a node $v\in V$ at time $t$, let $\theta_{v,1\to0}(t)$ abstract the effectiveness of the reactive defenses and $\theta_{v,0\to1}(t)$ abstract the capability of attacks against the preventive defenses. $\beta_v\in (0,1]$ represents the probability that the {\em compromised}\ computer $v$ changes to the {\em secure}\ state because the attacks are detected and cleaned up by the reactive defenses. Then, $\theta_{v,1\to0}(t)=\beta_v$. Let $\alpha_v\in [0,1]$ denote the probability that the {\em secure}\ computer $v$ becomes {\em compromised}\ despite the presence of the preventive defenses (i.e., the preventive defenses are penetrated by the pull-based attacks). And let $\gamma_{uv}\in (0,1]$ denote the probability that a {\em compromised}\ computer $u$ wages a successful attack against a {\em secure}\ computer $v$ despite the preventive defenses (i.e., the preventive defenses are penetrated by push-based attacks), where $(u,v)\in E$. Under the assumption that the attacks are waged independent of each other, it holds that \begin{equation} \label{eq:nonlinear-term} \theta_{v,0\to1}(t)=1-(1-\alpha_v)\prod_{u\in N_{v}}\big(1-\gamma_{uv} i_{u}(t)\big). \end{equation} The dynamics can be rewritten as a system of $n$ nonlinear equations for $v\in V$ \cite{ZhengSub}: \begin{equation} \begin{aligned} \label{Main} \dfrac{{\rm d}i_{v}(t)}{{\rm d}t}=f_{v}(i)=-\beta_v i_{v}(t) +\bigg[1-(1-\alpha_v)\prod_{u\in N_{v}}\big(1-\gamma_{uv} i_{u}(t)\big)\bigg]\big(1-i_{v}(t)\big). \end{aligned} \end{equation} Notice that system \eqref{Main} is globally stable (i.e., there exists a unique equilibrium $i^{*}\in [0,1]^n$ such that every trajectory of system \eqref{Main} converges to $i^{*}$) no matter whether the parameters are nodes-dependent or nodes-independent \cite{ZhengSub}. If the parameters of the network system are nodes-independent (i.e., $\alpha_v=\alpha$, $\beta_v=\beta$ for any $v\in V$ and $\gamma_{uv}=\gamma$ for any $u,v\in V$, $(u,v)\in E$), the global convergence of system \eqref{Main} can be summarized as follows: \begin{itemize} \item If the attacker wages both push-based and pull-based attacks on some nodes $v\in V$ (i.e., $\alpha_v>0$ for some nodes $v\in V$), system \eqref{Main} is globally convergent in the entire parameter universe and the dynamics converges to a unique {\em nonzero} equilibrium exponentially. \item If the attacker only wages push-based attacks (i.e., $\alpha_v=0$ for any nodes $v\in V$), system \eqref{Main} is still globally convergent in the parameter universe but the convergence speed depends on all the model parameters $(\beta_v,\gamma_{uv})$ and the largest eigenvalue $\lambda_{A,1}$ of adjacency matrix $A$. \end{itemize} In this paper, we need to control the dynamics with nodes-dependent parameters converging to equilibrium zero. Despite the complexity of nodes heterogeneity, we can still take advantage of the convergence properties of the dynamics with nodes-independent parameters. \subsection{Problem Statement: Controlling Cybersecurity Dynamics} In this paper, we focus on controlling the convergence process of the cybersecurity dynamics model in a decentralized control manner, which means that we only need to observe the state of the target node $v$ during the control process of $v$, with no need to observe the states of its neighbors within the network. For every node $v$ of the network, we need to control $i_v(t)$ converging to zero at our target convergence speed with relatively low control cost by switching its parameter $\beta_v$ according to our control rule. Before presenting the control method, we need to finish two pre-control steps to assure the effectiveness of our method. As introduced before, we need to control the preventive and reactive defense dynamics converging to equilibrium zero globally. So firstly, as discussed above, we need to force the parameter $\alpha_v=0$ for each node $v$, which means that the threats of pull-based attacks are eliminated after the first step of the control process (e.g., connections between some compromised websites and the network system in which the pull-based epidemic spreading takes place are all cut off). This is the first step of the pre-control process. Then the corresponding push-based dynamics model we focus on can be rewritten as: \begin{equation} \begin{aligned} \label{Main-push} \dfrac{{\rm d}i_{v}(t)}{{\rm d}t}=f_{v}(i)=-\beta_v i_{v}(t) +\bigg[1-\prod_{u\in N_{v}}\big(1-\gamma_{uv} i_{u}(t)\big)\bigg]\big(1-i_{v}(t)\big). \end{aligned} \end{equation} As for $\beta_v$, we select two reactive defense strategy with different control cost for our parameter switching method, including one relatively strict defense strategy with higher control cost, denoted by $\beta_{+}$, and one relatively relaxed defense strategy with lower control cost, denoted by $\beta_{-}$. Then we apparently have $\beta_{+}>\beta_{-}$. These two strategies should satisfy the conditions that they are both able to make the dynamics converge to equilibrium zero. The key difference between them is, comparing to our target convergence speed with regard to the dynamics to be controlled, the dynamics with $\beta_v=\beta_{+}$ should converge faster than the target speed for all nodes in $V$, while the dynamics with $\beta_v=\beta_{-}$ may converge more slowly than the target speed. The classical approach to control the dynamics through adjusting the reactive defense strategy is the trivial method that forcing $\beta_{v}=\beta_{+}$ during the entire control process, which is inefficient because it may cost too many defense resources to maintain the relatively strict reactive defense strategy. Besides, the convergence speed of the dynamics under high-cost control may be faster than what we actually need, causing redundancies and wastes of defense resources to some extent. As discussed above, both the relatively strict defense strategy and the relatively relaxed defense strategy need to be able to make the dynamics converge to equilibrium zero. So a safe method to get equilibrium zero is to let $\beta_v/\gamma_{max}\ge\lambda_{A,1}$ for both $\beta_v=\beta_{+}$ and $\beta_v=\beta_{-}$, where $\gamma_{max}$ denotes the maximum value of probability $\gamma_{uv}$ for all neighbor nodes pair $(u,v)\in E$. But due to the variety of parameter $\gamma_{uv}$, $\gamma_{max}$ can be relatively large in practice, leaving little choice for parameter $\beta_v$. Therefore, the second step of the pre-control process is to force $\gamma_{max}$ to a relatively small value, which means we need to permanently reinforce the preventive defense strategy for the nodes which are more vulnerable to push-based attacks launched by the attacker (e.g., a stronger network firewall or filter is deployed). After the two pre-control steps, in which all the pull-based attacks have been eliminated and $\gamma_{max}$ is relatively small, we now employ an parameter switching method to control the convergence speed with relatively low cost of defense resources. We use an event-based mechanism to define the cost-saving parameter switching rule. \subsection{Notations} Table \ref{table:model-variables-and-parameters} summarizes the major notations used in the paper. \begin{table}[!htbp] \centering \caption{Notations used throughout the paper.\label{table:model-variables-and-parameters}} \begin{tabular}{|c|p{0.8\textwidth}|} \hline $I_n$ & the $n*n$ identity matrix\\ \hline $\mathbb{R}$ & the set of real numbers\\ \hline $\mathbb{N}$ & the set of positive integers and zero\\ \hline $\|i\|_{1}$ & $\|i\|_{1}=\sum_{v=1}^{n}\|i_{v}\|$ is the $l_{1}$-norm for an $n$-dimensional vector $i=[i_{1},\ldots,i_{n}]\in\mathbb{R}^n$. Note that the result equally holds with respect to other norms.\\ \hline $G=(V,E), A$ & the attack-defense graph structure $G$ with adjacency matrix $A=[a_{vu}]_{n\times n}$ where $a_{vu}=1$ if and only if $(u,v)\in E$\\ \hline $N_{v}$ & $N_{v}=\{u\in V:(u,v)\in E\}$\\ \hline $\alpha_v \in [0,1]$ & the probability that {\em secure}\ node $v$ becomes {\em compromised}\ because pull-based attack penetrates preventive defense \\ \hline $\beta_v \in (0,1]$ & the probability that {\em compromised}\ node $v$ becomes {\em secure}\ because reactive defense detects and cleans compromise \\ \hline $\gamma_{uv} \in (0,1]$ & the probability that a {\em compromised}\ neighbor node $u$ wages a successful push-based attack against {\em secure}\ node $v$ \\ \hline $\gamma_{max} \in (0,1]$ & the maximum value of probability $\gamma_{uv}$ for all neighbor nodes pair $(u,v)\in E$ \\ \hline $\beta_{-}$, $\beta_{+}$ & the value of parameter $\beta_v$ for all nodes $v\in V$ in low-cost (high-cost) reactive defense setting\\ \hline $i_v(t)$, $i(t)$ & the probability $v$ is in {\em compromised}\ state at time $t$; $i(t)=[i_{1}(t),\cdots,i_{n}(t)]$ \\ \hline $\varphi_{up}(s)$, $\varphi_{up}(s)$ & the decision functions that trigger high-cost (low-cost) control events\\ \hline $t_{k}^{v}$, $\tau_{k}^{v}$ & the time for the $k$-th high-cost (low-cost) control event at $v\in V$ in the event-based control method; $t_{1}^{v}=\tau_{0}^{v}=0$ \\ \hline $T_{-}^v$, $T_{+}^v$ & the total time of maintaining the parameter $\beta_v$ of the target node in low-cost (high-cost) reactive defense setting during control process\\ \hline $\mathcal{S}(t)$, $\mathcal{S}_{v}(t)$ & the exponential speed index of the convergence speed of $i(t)$ ($i_v(t)$) when the dynamics converge exponentially\\ \hline $\chi_v(t)$ & the sample-state of node $v$ at time $t$; 0 means {\em secure} and 1 means {\em compromised} \\ \hline $\widehat{i_{v}(t)}$, $\widehat{s_{v}(t)}$ & the probability $v$ is in the {\em compromised}\ ({\em secure}) state at time $t$ as estimated from the sample-states \\ \hline $\mathcal{W}$, $\mathcal{W}^{'}(t)$ & the time window with fixed (adaptive) time length for estimating the probability-states from sample-states\\ \hline \end{tabular} \end{table} \section{An Event-based Parameter Switching Method} \label{sec:proof-of-problem} In this section, we initiate an event-based parameter switching method to control the convergence speed to our target speed, with relatively low cost of defense resources. That is to say, we switch the reactive defense strategy between two predefined settings in practice, according to the event-based trigger rule we proposed. We first show that the global dynamics under control will converge to equilibrium zero at our target convergence speed. Then we prove that there is no Zeno behavior during the entire control process. \subsection{Designing Event-based Parameter Switching Rule} \label{subsec:Design-Rule} We apply the control method on all nodes of the network system. But for the purpose of clarification, we focus on one target node $v$ to explain the control process corresponding to the decentralized control manner as discussed above. We switch the parameter $\beta_v$ of the target node $v$ between two different groups of parameters alternately, which are the low-cost parameter $\beta_v=\beta_{-}$ and the high-cost parameter $\beta_v=\beta_{+}$. Here the footnotes 'lc' stands for 'low cost' and 'hc' stands for 'high cost'. We use $T_{-}^v$ and $T_{+}^v$ respectively to denote the total time of maintaining the two parameters $\beta_{-}$ and $\beta_{+}$ during the entire control process for node $v\in V$. We want the dynamics to converge to zero at target convergence speed with a relatively low total cost, which means to make the mean of time ratio $\frac{1}{n}\sum_{v\in V}\frac{T_{+}^v}{T_{+}^v+T_{-}^v}$ relatively small. Before defining the event-trigger rule, let us review these definition and lemma. For all nodes $v\in V$, let $n*n$ real matrix $K=\{\gamma_{uv}\}_{u,v=1}^n$, where $\gamma_{vv}=0$ for all nodes $v\in V$, and let non-singular $n*n$ real diagonal matrix $B=diag(\{\beta_v\}_{v=1}^n)$. Notice that $\beta_v$ takes values in $\{\beta_{-},\beta_{+}\}$ for all nodes $v\in V$. Let $B_{+}=\beta_{+}I_n$ and $B_{-}=\beta_{-}I_n$, where $I_n$ is $n*n$ identity matrix. Then we have following lemma. \begin{lemma} \label{lemma-get-m-matrix} If there exists some positive constant number $\iota$, so that $J=B_{+}-K-\iota I_n$ is a monotone matrix (M-matrix), then there exists a positive diagonal matrix $P=diag(\{p_v\}_{v=1}^n)$ such that $[(-J)P+P^T(-J)]$ is negative definite. \end{lemma} Apparently, our target convergence speed should be faster than the dynamics with $\beta_v=\beta_{-}$ for all nodes in $V$, and slower than the dynamics with $\beta_v=\beta_{+}$ for all nodes in $V$. We can set our target convergence speed as $C{\rm e}^{-\iota t}$ as long as $\iota>0$ satisfies Lemma \ref{lemma-get-m-matrix}, where $C$ is a positive constant number. We will prove the effectiveness later. In our event-based parameter switching method, we switch the parameter $\beta_v$ when an event is triggered (e.g. certain conditions are satisfied). Between two consecutive events, the parameter $\beta_v$ holds. We define two criterion functions according to our target speed of the convergence process: \begin{eqnarray*} \begin{cases} \displaystyle \varphi_{up}(t)&=~{\rm e}^{-\iota t},{\forall}t\ge0,\\[8pt] \displaystyle \varphi_{low}(t)&=~L*{\rm e}^{-\iota t},{\forall}t\ge0. \end{cases} \end{eqnarray*} where $L$ is an positive constant number satisfying $0<L<1$. Notice that $\varphi_{up}(t)$ represents the ideal convergence process at our target convergence speed. Theoretically, $\varphi_{up}(t)$ can be a polynomial function if the original dynamics converge polynomially. But from the Sard's Lemma in \cite{sard1942measure}, the parameter regime that causing polynomial convergence speed is in a zero measure set and cannot be chosen in practice. So without loss of generality, we let $\varphi_{up}(t)$ be an exponential function for the simplification of narrative. It is also worth noting that $\varphi_{low}(t)$ can be defined in other functional form (only need to satisfy the conditions of convergence speed). With the matrix $P$ defined in Lemma \ref{lemma-get-m-matrix}, we now define rigger rule as follows: \begin{definition}[event-based trigger rule] \label{trigger-rule-control} $P=diag(\{p_v\}_{v=1}^n)$ is as defined in Lemma \ref{lemma-get-m-matrix}. Let $m_v(t)=p_v^{-1}i_v(t)$, then for $k=1,2,\ldots$, the trigger rule is defined as: \begin{itemize} \item if $m_v(0)\ge 1$, then let $t_{1}^{v}=0$, and \begin{eqnarray*} \begin{cases} \displaystyle \tau_{k}^{v}=&\inf\bigg\{s\ge t_{k}^{v}:m_v(s)\le \varphi_{low}(s)\bigg\}\\[8pt] \displaystyle t_{k+1}^{v}=&\inf\bigg\{s\ge \tau_{k}^{v}:m_v(s)\ge \varphi_{up}(s)\bigg\} \end{cases} \end{eqnarray*} \item if $m_v(0)<1$, then let $\tau_{0}^{v}=0$, and \begin{eqnarray*} \begin{cases} \displaystyle t_{k}^{v}=&\inf\bigg\{s\ge \tau_{k-1}^{v}:m_v(s)\ge \varphi_{up}(s)\bigg\}\\[8pt] \displaystyle \tau_{k}^{v}=&\inf\bigg\{s\ge t_{k}^{v}:m_v(s)\le \varphi_{low}(s)\bigg\} \end{cases} \end{eqnarray*} \end{itemize} which specifies two sequences of parameter switching events: \begin{itemize} \item High-cost control event: At time $t_{k}^{v}$, the value of parameter $\beta_v$ switches to $\beta_{+}$ which generates relatively high control cost. \item Low-cost control event: At time $\tau_{k}^{v}$, the value of parameter $\beta_v$ switches to $\beta_{-}$ which generates relatively low control cost. \end{itemize} \end{definition} As discussed above, both $\beta_{-}$ and $\beta_{+}$ need to be able to make the dynamics converge to equilibrium zero in the parameter regime after pre-control. So system \eqref{Main-push} can be written as: for target node $v\in V$, \begin{itemize} \item If $t\in[t_{k}^{v},\tau_{k}^{v})$, \begin{equation} \begin{aligned} \label{high-cost-dynamics} \frac{{\rm d}i_{v}(t)}{{\rm d}t =-\beta_{+} i_{v}(t) +\bigg[1- \prod_{u\in N_{v}}\Big(1-\gamma_{uv} i_{u}(t)\Big)\bigg]\big(1-i_{v}(t)\big), \end{aligned} \end{equation} \item If $t\in[\tau_{k}^{v},t_{k+1}^{v})$, \begin{equation} \begin{aligned} \label{low-cost-dynamics} \frac{{\rm d}i_{v}(t)}{{\rm d}t =-\beta_{-} i_{v}(t) +\bigg[1- \prod_{u\in N_{v}}\Big(1-\gamma_{uv} i_{u}(t)\Big)\bigg]\big(1-i_{v}(t)\big), \end{aligned} \end{equation} \end{itemize} For $k=1,2,\ldots$, we regard the two control steps \eqref{high-cost-dynamics} and \eqref{low-cost-dynamics} during $t\in[t_{k}^{v},t_{k+1}^{v})$ as the $k$-th {\em control cycle}. \subsection{Analyzing the Event-based Parameter Switching Method} Under the control procedure proposed above, we will prove the effectiveness of the event-based parameter switching method, that is, the new dynamics of the target node under control will converge to zero at our target convergence speed, and what's more important, with no Zeno behavior. \begin{theorem} \label{theorem-main} For any node $v\in V$, system \eqref{high-cost-dynamics} \eqref{low-cost-dynamics} generated by the event-based parameter switching control strategy (trigger rule Definition \ref{trigger-rule-control}) will converge to zero at the convergence speed same as $\varphi_{up}(t)$, with no Zeno behavior. \end{theorem} \begin{proof} We first prove that $m_v(t)=p_v^{-1}i_v(t)$ under parameter switching control strategy (trigger rule Definition \ref{trigger-rule-control}) could never exceed $\varphi_{up}(t)$ after time $\tau_{1}^{v}$ for all nodes $v\in V$ (i.e., it always holds that $p_v^{-1}i_v(t)\le\varphi_{up}(t)$ after time $\tau_{1}^{v}$). For any node $v\in V$, with regard to its original dynamic \eqref{Main-push}, let $u_{-}$ be the smallest index in $N_{v}$. Notice that \begin{align*} &~1-\prod_{u\in N_{v}}\Big(1-\gamma_{uv} i_{u}(t)\Big)\\ =&~\bigg[\Big(1-\gamma_{u_{-}v} i_{u_{-}}(t)\Big)+\gamma_{u_{-}v} i_{u_{-}}(t)\bigg] -\prod_{u\in N_{v}}\Big(1-\gamma_{uv} i_{u}(t)\Big)\\ =&~\gamma_{u_{-}v} i_{u_{-}}(t)+\Big(1-\gamma_{u_{-}v} i_{u_{-}}(t)\Big) \bigg[1-\prod_{u>u_{-},u\in N_{v}}\Big(1-\gamma_{uv} i_{u}(t)\Big)\bigg]. \end{align*} This recurrent process will lead to \begin{align*} \frac{{\rm d}}{{\rm d}t}i_{v}(t) =-\beta_v i_{v}(t)+\Big(1-i_{v}(t)\Big)\sum_{\omega\in N_{v}}\gamma_{\omega v}i_{\omega}(t) \prod_{u<\omega,u\in N_{v}}\Big(1-\gamma_{uv} i_{u}(t)\Big) \end{align*} So we have the following inequality \begin{align} \label{inequality-node} \frac{{\rm d}}{{\mathrm d}t}i_{v}(t)\le-\beta_v i_{v}(t)+\sum_{\omega\in N_{v}}\gamma_{\omega v}i_{\omega}(t). \end{align} At time $\tau_1^{v}$, we have $m_v(\tau_1^{v})=\varphi_{low}(\tau_1^{v})<\varphi_{up}(\tau_1^{v})$ for all nodes $v\in V$. Let $M(t)=\max_{v\in V}m_v(t)$. Assume there is some time point $t^{*}$ so that $M(t^{*})=\varphi_{up}(t^{*})$, then for each $v^{*}\in V$ so that $m_{v^{*}}(t^{*})=M(t^{*})$, by inequality \eqref{inequality-node} we have \begin{align*} \frac{{\mathrm d}}{{\mathrm d}t}\bigl[i_{v^{*}}(t){\mathit e}^{\iota t}\bigr]\biggl|_{t=t^{*}} &\le& -\beta_v^{*} i_{v^{*}}(t^{*}){\mathit e}^{\iota t^{*}}+\sum_{\omega\in N_{v^{*}}}\gamma_{\omega v^{*}}i_{\omega}(t^{*}){\mathit e}^{\iota t^{*}}+\iota i_{v^{*}}(t^{*}){\mathit e}^{\iota t^{*}}\\ &=& -\beta_v^{*} p_{v^{*}}m_{v^{*}}(t^{*}){\mathit e}^{\iota t^{*}}+\sum_{\omega\in N_{v^{*}}}\gamma_{\omega v^{*}}p_{\omega^{*}}m_{\omega}(t^{*}){\mathit e}^{\iota t^{*}} +\iota p_{v^{*}}m_{v^{*}}(t^{*}){\mathit e}^{\iota t^{*}}\\ \end{align*} Notice that $m_{v^{*}}(t^{*})\ge m_{\omega}(t^{*})$, so we have \begin{align*} \frac{{\mathrm d}}{{\mathrm d}t}\big[i_{v^{*}}(t){\mathit e}^{\iota t}\big]\bigg|_{t=t^{*}} \le (-\beta_v^{*} p_{v^{*}}+\sum_{\omega\in N_{v^{*}}}\gamma_{\omega v^{*}}p_{\omega^{*}}+\iota p_{v^{*}})m_{v^{*}}(t^{*}){\mathit e}^{\iota t^{*}} \end{align*} According to the event-based trigger rule Definition \ref{trigger-rule-control}, the reactive defense strategy switches to high-cost setting $\beta_+$ right after time point $t^{*}$. Recall Lemma \ref{lemma-get-m-matrix}, $J=B_{+}-K-\iota I_n$ is a monotone matrix (M-matrix) and $[(-J)P+P^T(-J)]$ is negative definite, which leads to $$-\beta_v^{*} p_{v^{*}}+\sum_{\omega\in N_{v^{*}}}\gamma_{\omega v^{*}}p_{\omega^{*}}+\iota p_{v^{*}}<0$$ So we have $$\frac{{\mathrm d}}{{\mathrm d}t}\big[i_{v^{*}}(t){\mathit e}^{\iota t}\big]\bigg|_{t=t^{*}}<0$$ This implies that $m_{v^{*}}(t){\mathit e}^{\iota t}=p_{v}^{-1}i_{v^{*}}(t){\mathit e}^{\iota t}$ is strictly decreasing after time point $t^{*}$. In the mean time, notice that $\varphi_{up}(t){\mathit e}^{\iota t}=1$, and $m_{v^{*}}(t^{*})=\varphi_{up}(t^{*})$. Finally we have $m_{v}(t)\le\varphi_{up}(t)$ for all nodes $v\in V$ at any time point $t$. Next we prove that the parameter switching events will continue to exist till infinite time. That is to say, $t\to+\infty$ implies $k\to+\infty$ for both $\{t_{k}^{v}\}$ and $\{\tau_{k}^{v}\}$ (i.e., there are infinitely many parameter switching events). Besides, we prove that there is no Zeno behavior during the entire control process. For the convenience of clarification, let ${\mathit e}^{-t\mathcal{S}_{+}^{v}(t)}$ be the convergence speed of dynamics \eqref{high-cost-dynamics} when $\beta_v=\beta_+$ (i.e., the high-cost control), and ${\mathit e}^{-t\mathcal{S}_{-}^{v}(t)}$ be the convergence speed of dynamics \eqref{low-cost-dynamics} when $\beta_v=\beta_-$ (i.e., the low-cost control). Notice that $\mathcal{S}_{+}^{v}(t)$ and $\mathcal{S}_{-}^{v}(t)$ may not be constant numbers, but they always satisfy $\mathcal{S}_{+}^{v}(t)>\iota$ and $\mathcal{S}_{-}^{v}(t)<\iota$. For any $k=1,2,3\cdots$, with respect to the $k$-th {\em control cycle} during $t\in[t_{k}^{v},t_{k+1}^{v})$, We analyze the two control steps respectively. i)\ {\em Step one}: With regard to the high-cost control step \eqref{high-cost-dynamics} during $t\in[t_{k}^{v},\tau_{k}^{v})$, notice that the parameter $\beta_v$ switches to high-cost setting $\beta_+$ since $m_v(t_{k}^{v})=\varphi_{up}(t_{k}^{v})$. Due to $\mathcal{S}_{+}^{v}(t)>\iota$, $i_v(t)$ converges to zero faster than $\varphi_{up}(t)$ and $\varphi_{low}(t)$. This leads to the existence of the next event triggered at $\tau_{k}^{v}$ so that $m_v(\tau_{k}^{v})=\varphi_{low}(\tau_{k}^{v})$, which is a low-cost control event. In the following, we prove that there is no Zeno behavior in the high-cost control step \eqref{high-cost-dynamics} during $t\in[t_{k}^{v},\tau_{k}^{v})$. According to the dynamic, we have \begin{align*} \bigg|\int_{t_{k}^{v}}^{\tau_k^v}\frac{{\rm d}}{{\rm d}t}\Big[m_v(t)\Big]{\rm d}t\bigg|=\bigg|\varphi_{up}(t_{k}^{v})-\varphi_{low}(\tau_{k}^{v})\bigg| \end{align*} For the left side, we have \begin{align*} \bigg|\int_{t_{k}^{v}}^{\tau_k^v}\frac{{\rm d}}{{\rm d}t}\Big[m_v(t)\Big]{\rm d}t\bigg| \le~&p_v^{-1}\int_{t_{k}^{v}}^{\tau_k^v}\bigg|\frac{{\rm d}}{{\rm d}t}\Big[i_v(t)\Big]\bigg|{\rm d}t\\ \le~&M\int_{t_{k}^{v}}^{\tau_k^v}\bigg|{\rm e}^{-t\mathcal{S}_{+}^{v}(t)}\bigg|{\rm d}t\\ \le~& M {\rm e}^{-\iota t_k^v}(\tau_{k}^{v}-t_{k}^{v}), \end{align*} where $M$ is a positive constant. For the right side, we have \begin{align*} \bigg|\varphi_{up}(t_{k}^{v})-\varphi_{low}(\tau_{k}^{v})\bigg| =~&i_v(0){\rm e}^{-\iota t_{k}^{v}}-L*i_v(0){\rm e}^{-\iota \tau_{k}^{v}}\\ \ge~&(1-L)i_v(0){\rm e}^{-\iota \tau_{k}^{v}} \end{align*} Then for both sides, we have \begin{align*} (1-L)i_v(0){\rm e}^{-\iota \tau_{k}^{v}} \le M {\rm e}^{-\iota t_k^v}(\tau_{k}^{v}-t_{k}^{v}) \end{align*} Then we have \begin{align*} (1-L)i_v(0){\rm e}^{-\iota (\tau_{k}^{v}-t_{k}^{v})} \le M (\tau_{k}^{v}-t_{k}^{v}) \end{align*} It shows the existence of a positive number $\eta_v$, which is the root of the transcendental equation $(1-L)i_v(0){\rm e}^{-\iota \eta_v}\le M \eta_v$ and satisfies $\tau_{k}^{v}-t_{k}^{v}\ge\eta_{v}$, which essentially means that for every $v\in V$, $\inf\{\tau_{k}^{v}-t_{k}^{v}\}>0$. That is to say, there is no Zeno behavior in the high-cost control step \eqref{high-cost-dynamics} during $t\in[t_{k}^{v},\tau_{k}^{v})$. ii)\ {\em Step two}: With regard to the low-cost control step \eqref{low-cost-dynamics} during $t\in[\tau_{k}^{v},t_{k+1}^{v})$, notice that the parameter $\beta_v$ switches to low-cost setting $\beta_-$ since $m_v(\tau_{k}^{v})=\varphi_{low}(\tau_{k}^{v})$. Due to $\mathcal{S}_{-}^{v}(t)<\iota$, $\varphi_{up}(t)$ and $\varphi_{low}(t)$ converge to zero faster than $i_v(t)$. This leads to the existence of the next event triggered at $t_{k+1}^{v}$ so that $m_v(t_{k+1}^{v})=\varphi_{up}(t_{k+1}^{v})$, which is a high-cost control event. Now we prove that there is no Zeno behavior in the low-cost control step \eqref{low-cost-dynamics} during $t\in[\tau_{k}^{v},t_{k+1}^{v})$. This proof is much simpler than that of the high-cost control step. Notice that $i_v(t)$ converges to zero, so we have \begin{align*} \varphi_{up}(t_{k+1}^{v})\le\varphi_{low}(\tau_{k}^{v}) \end{align*} So we have \begin{align*} {\rm e}^{-\iota (t_{k+1}^{v}-\tau_{k}^{v})}\le L \end{align*} which essentially means that for every $v\in V$, $\inf\{t_{k+1}^{v}-\tau_{k}^{v}\}\ge -\ln{L}/\iota>0$. That is to say, there is no Zeno behavior in the low-cost control step \eqref{low-cost-dynamics} during $t\in[\tau_{k}^{v},t_{k+1}^{v})$. From the proof above, we have shown that $m_v(t)=p_v^{-1}i_v(t)$ continues to touch $\varphi_{up}(t)$ but could never exceed $\varphi_{up}(t)$ till infinite time, so $i_v$ converges to zero at the convergence speed same as $\varphi_{up}(t)$ (the convergence speed here refers to the average speed through time). Note that the proof under periodic reference setting (see also \cite{liu2020using}) is similar. \end{proof} \subsection{Translating Trigger Rule in Definition \ref{trigger-rule-control} to Algorithm} In order to employ the parameter switching control method presented above, we need to translate the event-based trigger rule in Definition \ref{trigger-rule-control} into a control algorithm. For this purpose, we need to observe the states of nodes first. Due to the node heterogeneity (i.e., the parameters are nodes-dependent) of the network system in this paper, the event-based observing method proposed in \cite{liu2020using} is no longer applicable, so we simply use the classical periodic observing method to handle this issue. Besides, with respect to the decentralized control manner of the parameter switching method, we illustrate the algorithm by focusing on one target node $v\in V$. \begin{algorithm}[!htbp] \caption{Event-based parameter switching control process according to the trigger rule in Definition \ref{trigger-rule-control}}\label{algorithm-trigger-control} \textbf{input:} ~$G=(V,E)$, $i_{v}(0)$, $\varphi_{up}$, $\varphi_{low}$, $\beta_{+}$, $\beta_{-}$, $h$\\ \textbf{output:} $\{t_{k}^{v}\}_{k=1}^{+\infty}$ and $\{\tau_{k}^{v}\}_{k=0}^{+\infty}$ for $v\in V$\\ \textbf{initialize:} $t_{1}^{v}\leftarrow 0$ and $\tau_{0}^{v}\leftarrow 0$; $k\leftarrow 1$;\\ Get $p_v>0$ for $v$ as specified in Lemma \ref{lemma-get-m-matrix}\\ $Cycle\leftarrow 0$\\ \While{{\tt true}} $t\leftarrow t_{k}^{v}$\\ \While{$Cycle = 0$}{ \If{$p_v^{-1}i_v(t)\le \varphi_{low}(t)$}{ switch reactive defense strategy of $v$ to $\beta_{-}$\\ $Cycle \leftarrow 1$\\ $\tau_{k}^{v}\leftarrow t$\\ } $t\leftarrow t+h$\\ } \While{$Cycle = 1$}{ \If{$p_v^{-1}i_v(t)\ge \varphi_{up}(t)$}{ switch reactive defense strategy of $v$ to $\beta_{+}$\\ $Cycle \leftarrow 0$\\ $t_{k+1}^{v}\leftarrow t$\\ } $t\leftarrow t+h$\\ } $k \leftarrow k+1$\\ } \end{algorithm} There are four groups of inputs in Algorithm \ref{algorithm-trigger-control}: attack-defense graph $G=(V,E)$; initial values $i_v(0)$ for node $v\in V$; two criterion functions $\varphi_{up}(t)$ and $\varphi_{low}(t)$; two reactive defense strategies with their corresponding parameter values $\beta_{+}$ and $\beta_{-}$ and a step length parameter $h$ (i.e., the constant time interval of the periodic observing method, see also \cite{liu2020using}). \subsection{Numerical Examples} \label{sec:simulation} We use numerical examples to illustrate the convergence process of the dynamics under control. The numerical examples exhibit the the effectiveness of the proposed event-based parameter switching method. The settings of the examples are defined as follows. For graph $G$ in the dynamics model, we conduct experiments on both undirected graph and directed graph to put the method into practice. The following network structures are obtained from \url{http://snap.stanford.edu/data/} Note that the extraction of $G$ in practice demands access to the enterprise's physical network topologies and security policies, which are usually confidential data unavailable to academic researchers. \begin{itemize} \item Enron email network: This is an undirected graph with $|V|=5242$ nodes, $|E|=28980$ edges, maximal node degree $81$ and $\lambda_{A,1}=45.6167$. \item Gnutella peer-to-peer network: This is a directed graph with $|V|=8,717$ nodes, $|E|=31,525$ links, maximal node in-degree $64$ and $\lambda_{A,1}=4.7395$. \end{itemize} We set $\beta_{+}=0.8$, $\beta_{-}=0.1$ with respect to $\beta_v$ for all nodes $v\in V$. As for $\gamma_v$, we randomly select the values for all nodes $v\in V$ with an upper bound $\gamma_{max}=0.002$ for undirected Enron email network and $\gamma_{max}=0.013$ for directed Gnutella peer-to-peer network. With respect to the criterion functions $\{\varphi_{up},\varphi_{low}\}$ of the event-based trigger rule Definition \ref{trigger-rule-control}, we set $\iota=0.5$ and $L=0.5$ for both undirected Enron email network and directed Gnutella peer-to-peer network. Thus, the conditions of the parameter switching method proposed above are satisfied. We calculate the matrix $P=diag(\{p_v\}_{v=1}^n)$ in Lemma \ref{lemma-get-m-matrix} for each graph respectively. As for the initial values, each node $v\in V$ is assigned with an initial compromise probability $i_v(0)\in_R [0,1]$ where $\in_R$ means sampling uniformly at random. Besides, we consider $t\in[0,500]$ with a fixed step-length $h=0.025$. The convergence processes of dynamics are shown in Figure \ref{fig:prob-state-control}. Notice that the grey curve (i.e., the dynamics under control) refers to $i_v(t)$, while the the blue curve (i.e., the adjusted control target) refers to $m_v(t)=p_v^{-1}i_v(t)$. \begin{figure}[!htbp] \centering \subfigure[Node 1855 of undirected Enron email network]{\includegraphics[width=0.46\textwidth]{UD_Dynam1855.png}\label{Undirected1}}\hspace{2ex} \subfigure[Node 2923 of undirected Enron email network]{\includegraphics[width=0.46\textwidth]{UD_Dynam2923.png}\label{Undirected2}}\hspace{2ex} \subfigure[Node 1187 of directed Gnutella peer-to-peer network]{\includegraphics[width=0.46\textwidth]{D_Dynam1186.png}\label{Directed1}}\hspace{2ex} \subfigure[Node 6992 of directed Gnutella peer-to-peer network]{\includegraphics[width=0.46\textwidth]{D_Dynam6992.png}\label{Directed2}} \caption{The convergence processes of dynamics under parameter switching control for both undirected and directed graph. \label{fig:prob-state-control}} \end{figure} Figure \ref{fig:prob-state-control} exhibits the control process of the proposed event-based parameter switching method and is consistent with the proof of Theorem \ref{theorem-main}. Then we verify that the convergence speed (i.e., the average convergence speed through time) of the dynamics under control is close to the target speed given by the criterion function $\varphi_{up}$. In order to confirm the result, we define the following indicator of convergence speed and name it {\em exponential speed index}:$$\mathcal{S}(t)=-\frac{1}{\Delta t}\ln\frac{i(t+\Delta t)}{i(t)}.$$ Notice that the exponential speed index of the criterion function is equal to $\iota=0.5$ in our settings, which also represents the target speed index. The exponential convergence speed indexes of the dynamics $i(t)=[i_{1}(t),\cdots,i_{n}(t)]$ under parameter switching control for both undirected and directed graph are shown in Figure \ref{fig:convergence-speed-prob}. Notice that the value of the blue line $\mathcal{S}$ refers to the whole time average of the green curve for $t\in[0,500]$ and should be close to the red line (i.e., the target speed index). \begin{figure}[!htbp] \centering \subfigure[The convergence speed index of undirected Enron email network]{\includegraphics[width=0.46\textwidth]{UD_Speed.png}\label{Undirected1}}\hspace{2ex} \subfigure[The convergence speed index of directed Gnutella peer-to-peer network]{\includegraphics[width=0.46\textwidth]{D_Speed.png}\label{Directed2}} \caption{The exponential convergence speed indexes of the dynamics under parameter switching control for both undirected and directed graph. \label{fig:convergence-speed-prob}} \end{figure} For the presented convergence speed experiments, the threshold of effectiveness is defined as $\frac{|\mathcal{S}-\iota|}{\iota}$, which should be less than $10\%$. For the undirected Enron email network, we have $\frac{|\mathcal{S}-\iota|}{\iota}=3.72\%$, and for the directed Gnutella peer-to-peer network, we have $\frac{|\mathcal{S}-\iota|}{\iota}=2.60\%$, which shows the effectiveness. Next, it comes to the control cost, which is indicates by the mean of time ratio $\frac{1}{n}\sum_{v\in V}\frac{T_{+}^v}{T_{+}^v+T_{-}^v}$. For the classical control approach without parameter switching, it holds that $\beta_v=\beta_{+}$ for $t\in[0,500]$ for all nodes $v\in V$. Suppose the control cost is equal to $1$. From the experiments, by using the event-based parameter switching method, the control cost is equal to $0.50$ in the case of undirected Enron email network and $0.53$ in the case of directed Gnutella peer-to-peer network respectively. That is to say, the new control method should save at least $40\%$ of the cost incurred by the classical approach. So we conclude that the event-based parameter switching method can reduce more than $40\%$ of the control cost compared with the classical approach, which shows the efficiency. \section{Putting the Event-based Method into Practice} \label{sec:gap} Similar to \cite{liu2020using}, we need to bridge the gap between the following two kinds of states for utilizing the event-based parameter switching control method in practice. In the aforementioned model, the state of node $v\in V$ at time $t$ is represented by $i_v(t)$, namely the probability that $v$ is in {\em compromised}\ state at time $t$. In practice, this state is often measured as a Boolean value, with ``0'' indicating $v$ is {\em secure}\ but vulnerable and ``1'' indicating $v$ is {\em compromised}. In other words, the {\em sample-state} of node $v\in V$ at time $t$ can be denoted by \begin{align} \label{0-1-State} \chi_{v}(t)= \begin{cases} 0 & v \text{ is in the {\em secure}\ state at time $t$}\\ 1 & v \text{ is in the {\em compromised}\ state at time $t$}. \end{cases} \end{align} This difference underlines the gap between the probability-states in the model and the sample-states in practice. It is worth noticing that, the algorithm proposed in \cite{liu2020using}, which estimates the probability-states via 0-1 state ergodic process, can not be simply transferred and applied to the current control problem. Unlike the equilibrium estimation task in which the accuracy of estimation is of vital importance, the timeliness is the main focus in our event-based parameter switching dynamics control. Without the timeliness, the event triggered by rule may suffer from time lag , which could invalidate the event-based control method. So despite the 0-1 state sequences over the whole time (i.e., $[0,t]$) may provide higher accuracy of probability estimation, yet we should avoid using them in the present task. Instead, we propose a new algorithm which takes advantage of 0-1 state sequences within a time window. \subsection{Estimation via 0-1 State Sequences within a Time Window} Motivated by the theorem of two-valued processes introduced in \cite[Chapter~1]{parzen1999stochastic}, we propose a new method to bridge the aforementioned gap by obtaining an estimation $\widehat{i_{v}(t)}$ of probabilities $i_{v}(t)$ and an estimation $\widehat{s_{v}(t)}$ of probabilities $s_{v}(t)$ from a 0-1 state sequence within a time window, as indicated by \eqref{0-1-State}. We design a new algorithm for the event-based parameter switching method from the theorem. Firstly, let us review the theorem of two-valued processes. We use the Lebesgue measure $\mathcal{M}$ to define \begin{align} \label{Random-Variables-T} \begin{cases} \displaystyle \mathcal{T}_{v0}(t)=\mathcal{M}\big(\big\{\tau\le t:\chi_{v}(\tau)=0\big\}\big)\\[8pt] \displaystyle \mathcal{T}_{v1}(t)=\mathcal{M}\big(\big\{\tau\le t:\chi_{v}(\tau)=1\big\}\big). \end{cases} \end{align} Theorem \ref{Theorem:0-1} below shows how to generate probabilities $\widehat{i_{v}(t)}$ and $\widehat{s_{v}(t)}$ from a 0-1 state ergodic process over time. \begin{theorem}[\cite{parzen1999stochastic}] \label{Theorem:0-1} Let $\{\chi_{v}(t),t>0\}$ for $v\in V$ be a 0-1 state ergodic process. Let \begin{eqnarray*} \begin{cases} \displaystyle \widehat{s_{v}(t)} \frac{\mathcal{T}_{v0}(t)}{t}\\[8pt] \displaystyle \widehat{i_{v}(t)} \frac{\mathcal{T}_{v1}(t)}{t}, \end{cases} \end{eqnarray*} then it is obvious that $\lim_{t\to+\infty}\big[\mathbb{P}\big(\chi_{v}(t)=0\big)-\widehat{s_{v}(t)}\big] = 0$, $\lim_{t\to+\infty}\big[\mathbb{P}\big(\chi_{v}(t)=1\big)-\widehat{i_{v}(t)}\big] = 0$. \end{theorem} In \cite{liu2020using}, $\widehat{i_{v}(t)}$ and $\widehat{s_{v}(t)}$ is used to estimate $i_{v}(t)$ and $s_{v}(t)$ at sufficiently large time $t$ respectively. However, as the time-averaged estimations of probability state, $\widehat{i_{v}(t)}$ and $\widehat{s_{v}(t)}$ suffer from an increasing time lag as $t\to\infty$. Due to the convergence property of the preventive and reactive cyber defense dynamics, the time-averaged estimation still converges to the original equilibrium, which makes it effective in the scenario of \cite{liu2020using}. But in our present scenario, we need a new estimation method with less time lag and faster reaction. For simplicity, sometimes $\mathcal{W}$ also represents the time window which covers the $\mathcal{W}$ most recent time points before $t$. Let \begin{align} \label{Random-Variables-T} \begin{cases} \displaystyle \mathcal{T}_{v0}^{\mathcal{W}}(t)=\mathcal{M}\big(\big\{t-\mathcal{W}<\tau\le t:\chi_{v}(\tau)=0\big\}\big)\\[8pt] \displaystyle \mathcal{T}_{v1}^{\mathcal{W}}(t)=\mathcal{M}\big(\big\{t-\mathcal{W}<\tau\le t:\chi_{v}(\tau)=1\big\}\big)\\[8pt] \end{cases} \end{align} Notice that the smaller $\mathcal{W}$ implies the less time lag and the lower accuracy the probability estimation exhibits. So we face the immediacy-accuracy trade-off dilemma when selecting the proper time window $\mathcal{W}$. It is also worth noting that a simple way to enhance the estimation accuracy is to increase the sampling frequency $h$ (i.e., the constant time interval of the periodic observing method, see also \cite{liu2020using}). But there is always a limitation on the sampling frequency in practice. In order to cope with this trade-off dilemma, we later propose the design of adaptive time windows, which shows great performance. In order to simulate a 0-1 ergodic process for $\forall v\in V$, the paper samples node $v$ at time $t$ by its compromise probability $i_v(t)$: \begin{align} \label{RandomChoice} \chi_{v}(t)=H\big[i_v(t)-Rand(0,1)\big] \end{align} where $Rand(0,1)$ means drawing a random real number uniformly from $[0,1]$, and $H$ is the Discrete Heaviside step function: \begin{align} \label{Heaviside} H(x)= \begin{cases} 0 & x<0\\ 1 & x\ge0. \end{cases} \end{align} \subsection{Using the Event-based Control Method in Practice} With the aid of the newly proposed estimation method with adaptive time windows, which exhibits less time lag and faster response, the gap between probability-states and sample-states has been properly bridged. We now use the aforementioned event-based parameter switching method in practice to control the preventive and reactive cyber defense dynamics. Since undirected networks are a special case of directed networks, only experiments on the directed Gnutella peer-to-peer Network are performed here. Let the time window $\mathcal{W}=30$. As proposed above, we use an adaptive time window $$\mathcal{W}^{'}(t)=\max(\mathcal{W},\frac{t}{C_0})$$ to replace the original time window $\mathcal{W}$ with a fixed time length, where $C_0$ is a positive constant number. In the settings of our numerical examples, let $C_0=3$ for $\mathcal{W}^{'}(t)$. Figure \ref{fig:sample-state-control-adaptive} shows the convergence process of the dynamics under control, using a time window with adaptive time length to estimate the probability $m_v(t)$. We can find that there are still some events triggered in the advanced stage $t\in [300,500]$ when probability $i_v(t)$ has fully converged to equilibrium zero, which means the event-based parameter switching method becomes effective by using the adaptive time window. With respect to the convergence speed indexes, notice that the threshold for it defined in Section \ref{sec:simulation} is $10\%$. Figure \ref{speed-adaptive-window} shows the effectiveness of the time window with adaptive time length, with $\frac{|\mathcal{S}-\iota|}{\iota}=6.79\%<10\%$, while $\frac{|\mathcal{S}-\iota|}{\iota}=27.27\%>10\%$ for the time window with fixed time length exhibited in Figure \ref{speed-fixed-window}. Besides, with respect to the control cost, which is indicated by the mean of time ratio $\frac{1}{n}\sum_{v\in V}\frac{T_{+}^v}{T_{+}^v+T_{-}^v}$, recall the statement in Section \ref{sec:simulation}, the threshold of efficiency is defined as $40\%$ of the cost. The control cost in the current numerical example is $0.60$, which means the reactive defense setting is under low-cost setting for $40\%$ of the time.(the low-cost reactive defense setting takes up $40\%$ of the time) \begin{figure}[!htbp] \centering \subfigure[Node 2688 using a time window with adaptive time length]{\includegraphics[width=0.46\textwidth]{Discrete_Adaptive_Dynam2688.png}\label{Discrete1}}\hspace{2ex} \subfigure[Node 4011 using a time window with adaptive time length]{\includegraphics[width=0.46\textwidth]{Discrete_Adaptive_Dynam4011.png}\label{Discrete2}} \caption{The convergence processes of dynamics under parameter switching control aiming at adjusted sample-states estimation, using a time window with adaptive time length. \label{fig:sample-state-control-adaptive}} \end{figure} \begin{figure}[!htbp] \centering \subfigure[The convergence speed index, using a time window with fixed time length]{\includegraphics[width=0.46\textwidth]{Speed_Fixed_Window.png}\label{speed-fixed-window}}\hspace{2ex} \subfigure[The convergence speed index, using a time window with adaptive time length]{\includegraphics[width=0.46\textwidth]{Speed_Adaptive_Window.png}\label{speed-adaptive-window}} \caption{The exponential convergence speed indexes of the dynamics under parameter switching control, using a time window with fixed or adaptive time length. \label{fig:speed-fixed-window-or-not}} \end{figure} \section{Conclusion} \label{sec:conclusion} In this paper, an event-based parameter switching method is proposed for the control tasks of cybersecurity, which helps avoid excessive control costs as well as guarantees the dynamics to converge as our desired speed. The Zeno-free property is proved, implying the feasibility of the method. Meanwhile, we designed a new estimation method with adaptive time windows in order to bridge the gap between the probability state and the sampling state with less time lags. Both theoretical and practical experiments are given to illustrate the parameter switching method, which show the effectiveness and efficiency of our new method. There are many open problems for future research: Do there exist some better event-based parameter switching trigger rule so as to save more defense resources and guarantee the convergence speed? Can this parameter switching approach be applied to other dynamics control? Furthermore, similar parameter switching approaches regarding the parameter $\gamma$ for push-based attacks are worth further studying. \section{Introduction} With the rapid development of internet technology (IT) and internet of things (IoT) technology, cybersecurity issues are attached more and more importance these years due to people's increasing reliance on internet. About 2 million to 5 million computers worldwide suffered from malware (are infected with malware) every day, estimating by the domestic network security company Jinshan in 2016. Cybersecurity, as a new subject for research, has received extensive attention from the academic community. The emerging research field 'cybersecurity dynamics' \cite{XuCybersecurityDynamicsHotSoS2014,XuBookChapterCD2019} is an interdisciplinary field, conceived from the methodology of several early studies in {\em biological epidemiology} (e.g., \cite{McKendrick1926,Kermack1927,Bailey1975,Anderson1991,HethcoteSIAMRew00}) and its variants in {\em cyber epidemiology} (e.g., \cite{KephartOkland91,KephartOkland93,Pastor2001,Moreno2002}) {\em interacting particle systems} \cite{Liggett1985}, and {\em microfoundation in economics} \cite{Hoover2010}. Different from the classical researches oriented to specific tools, such as Cryptography and Database Security, cybersecurity dynamics studies the offensive and defensive models under various circumstances from a whole-network perspective. \subsection{Our Contributions} \label{sec:our-contributions} In this paper, we investigate how to control the evolution of cybersecurity dynamics more efficiently and effectively in the context of {\em preventive and reactive cyber defense dynamics}, and guarantee its globally convergence to a safe state. When it comes to the control problem of the highly nonlinear network dynamics system, the traditional method is to use a single control strategy, that is, to adjust the dynamic parameters and maintain them, so that the security dynamics of the network space converges to a safe state globally. However, maintaining a high level of dynamic defense strategy may result in high control costs. Such excessive prevention and control will lead to a waste of control resources to some extent, and may even have negative effects on the stable operation of the network systems. In order to solve this problem, this paper proposes an event-based parameter switching approach to save control resources and control the evolution of cybersecurity dynamics in a decentralized manner. It is also proved that this approach is Zeno-free, that is, it will not fall victim to the Zeno behavior. Numerical examples show that the maintenance hours of high-cost control strategies can be reduced by more than $40\%$ with our parameter switching approach. In addition, this paper provides an estimation method to bridge the gap between the probability state and the sampling state when using the method in practice. Different from the equilibrium state estimation problem\cite{liu2020using}, the event-based control method requires considering the timeliness of probability estimation. A new adaptive estimation method is proposed to verify the effectiveness of the event-based parameter switching method in the control process through numerical examples. \ \subsection{Related Work} \label{sec:related-work} Similar event-based methods have been employed in many other application settings before (see, for example, \cite{Astrom2002,Tabuada2007,6425820}). In practical application, one of the essential problems is that this method should not fall victim to the Zeno behavior, which can lead to infinitely many events within a finite period of time, thus invalidate the method \cite{Johansson1999}. The importance of assuring Zeno-freeness in event-triggered control is witnessed by a body of literature, including \cite{Tabuada2007,Johannesson2007,WangLemmon2011,Dimarogonas2012}. The first cybersecurity dynamics model was proposed in \cite{XuAINA07}. Ten years later, \cite{ZhengSub} demonstrated that a certain class of cybersecurity dynamics is globally convergent in the entire parameter universe, which laid a foundation for our further research. The notion of cybersecurity dynamics, as discussed in \cite{XuCybersecurityDynamicsHotSoS2014,XuBookChapterCD2019}, has opened the door to a new research field. The preventive and reactive cyber defense dynamics is a particular kind of cybersecurity dynamics. Several other kinds of cybersecurity dynamics have been demonstrated in early studies, such as the models aiming to accommodate adaptive defenses \cite{XuTAAS2014}, active defenses \cite{XuHotSoS15,XuInternetMath2015ACD,XuGameSec13}, and proactive defenses \cite{XuHotSOS14-MTD}. It is worth mentioning that the event-based parameter switching method may be extended and applied to the various kinds of dynamics. \subsection{Paper Outline} In Section \ref{sec:model}, we briefly review the preventive and reactive cyber defense dynamics model and its global convergence in the entire parameter universe \cite{ZhengSub}. And then we state the problems we addressed in this paper. In Sections \ref{sec:proof-of-problem}, we present an event-based parameter switching method for controlling the cybersecurity dynamics, and prove the effectiveness of the control method (without Zeno behavior). In Section \ref{sec:simulation}, we present numerical examples on the theoretical model. Then in Section \ref{sec:gap}, we show how to apply this event-based parameter switching method in practice by bridging the gap between the probability-state in the theoretical model and the sample-state in practice, using an stochastic process method with adaptive time windows. In Section \ref{sec:conclusion}, we conclude the paper with open problems. \section{Problem Statement} \label{sec:model} \subsection{Review of Preventive and Reactive Defense Dynamics} As a particular kind of cybersecurity dynamics, the preventive and reactive defense dynamics model is first introduced in \cite{XuAINA07} and the convergence properties of the dynamics is studied in \cite{XuTAAS2012} . Later \cite{ZhengSub} fully analyzed the convergence issues, not only considered the common situation with node homogeneity (i.e., the parameters are node-independent), but also a more general situation with node heterogeneity (i.e., the parameters are nodes-dependent). The paper proved that this dynamics model is globally convergent in the entire parameter universe, that is, there is always a unique equilibrium, whose exact value (or position) depends on the specific parameter values instead of the initial state of the dynamics. In the preventive and reactive defense dynamics model, we consider two classes of defenses: {\em preventive defenses} and {\em reactive defenses}, and two classes of attacks: {\em push-based attacks} and {\em pull-based attacks}. Suppose that the attack-defense interaction occurs over an {\em attack-defense} graph structure $G=(V,E)$, where $V$ is the vertex set representing computers and $(u,v)\in E$ means computer $u$ can directly attack computer $v$ using push-based attack strategy (i.e., the communication from $u$ to $v$ is allowed by the security policy). $G$ can be derived from the security policy of a networked system and the physical network in question. Without loss of generality, we do {\em not} make any restrictions on the structure of $G$ (e.g., $G$ may be directed or undirected). Denote the adjacency matrix of $G$ by $A=[a_{vu}]_{n\times n}$, where $a_{vu}=1$ if and only if $(u,v) \in E$. Since the model aims to describe the attacks between computers, we set $a_{vv}=0$. Let $N_{v}=\{u\in V:~(u,v)\in E\}$. In this paper, we consider the continuous-time model described in \cite{ZhengSub}. At any time point, a node $v \in V$ is in one of two states: ``0'' means {\em secure}\ but vulnerable, or ``1'' means {\em compromised}. Let $s_v(t)$ and $i_v(t)$ be the probability that $v$ is {\em secure} and {\em compromised} at time $t$ respectively. It is obvious that $s_v(t)+i_v(t)=1$, $s_v(t)$ and $i_v(t)$ explain the term {\em probability-state}. For a node $v\in V$ at time $t$, let $\theta_{v,1\to0}(t)$ abstract the effectiveness of the reactive defenses and $\theta_{v,0\to1}(t)$ abstract the capability of attacks against the preventive defenses. $\beta_v\in (0,1]$ represents the probability that the {\em compromised}\ computer $v$ changes to the {\em secure}\ state because the attacks are detected and cleaned up by the reactive defenses. Then, $\theta_{v,1\to0}(t)=\beta_v$. Let $\alpha_v\in [0,1]$ denote the probability that the {\em secure}\ computer $v$ becomes {\em compromised}\ despite the presence of the preventive defenses (i.e., the preventive defenses are penetrated by the pull-based attacks). And let $\gamma_{uv}\in (0,1]$ denote the probability that a {\em compromised}\ computer $u$ wages a successful attack against a {\em secure}\ computer $v$ despite the preventive defenses (i.e., the preventive defenses are penetrated by push-based attacks), where $(u,v)\in E$. Under the assumption that the attacks are waged independent of each other, it holds that \begin{equation} \label{eq:nonlinear-term} \theta_{v,0\to1}(t)=1-(1-\alpha_v)\prod_{u\in N_{v}}\big(1-\gamma_{uv} i_{u}(t)\big). \end{equation} The dynamics can be rewritten as a system of $n$ nonlinear equations for $v\in V$ \cite{ZhengSub}: \begin{equation} \begin{aligned} \label{Main} \dfrac{{\rm d}i_{v}(t)}{{\rm d}t}=f_{v}(i)=-\beta_v i_{v}(t) +\bigg[1-(1-\alpha_v)\prod_{u\in N_{v}}\big(1-\gamma_{uv} i_{u}(t)\big)\bigg]\big(1-i_{v}(t)\big). \end{aligned} \end{equation} Notice that system \eqref{Main} is globally stable (i.e., there exists a unique equilibrium $i^{*}\in [0,1]^n$ such that every trajectory of system \eqref{Main} converges to $i^{*}$) no matter whether the parameters are nodes-dependent or nodes-independent \cite{ZhengSub}. If the parameters of the network system are nodes-independent (i.e., $\alpha_v=\alpha$, $\beta_v=\beta$ for any $v\in V$ and $\gamma_{uv}=\gamma$ for any $u,v\in V$, $(u,v)\in E$), the global convergence of system \eqref{Main} can be summarized as follows: \begin{itemize} \item If the attacker wages both push-based and pull-based attacks on some nodes $v\in V$ (i.e., $\alpha_v>0$ for some nodes $v\in V$), system \eqref{Main} is globally convergent in the entire parameter universe and the dynamics converges to a unique {\em nonzero} equilibrium exponentially. \item If the attacker only wages push-based attacks (i.e., $\alpha_v=0$ for any nodes $v\in V$), system \eqref{Main} is still globally convergent in the parameter universe but the convergence speed depends on all the model parameters $(\beta_v,\gamma_{uv})$ and the largest eigenvalue $\lambda_{A,1}$ of adjacency matrix $A$. \end{itemize} In this paper, we need to control the dynamics with nodes-dependent parameters converging to equilibrium zero. Despite the complexity of nodes heterogeneity, we can still take advantage of the convergence properties of the dynamics with nodes-independent parameters. \subsection{Problem Statement: Controlling Cybersecurity Dynamics} In this paper, we focus on controlling the convergence process of the cybersecurity dynamics model in a decentralized control manner, which means that we only need to observe the state of the target node $v$ during the control process of $v$, with no need to observe the states of its neighbors within the network. For every node $v$ of the network, we need to control $i_v(t)$ converging to zero at our target convergence speed with relatively low control cost by switching its parameter $\beta_v$ according to our control rule. Before presenting the control method, we need to finish two pre-control steps to assure the effectiveness of our method. As introduced before, we need to control the preventive and reactive defense dynamics converging to equilibrium zero globally. So firstly, as discussed above, we need to force the parameter $\alpha_v=0$ for each node $v$, which means that the threats of pull-based attacks are eliminated after the first step of the control process (e.g., connections between some compromised websites and the network system in which the pull-based epidemic spreading takes place are all cut off). This is the first step of the pre-control process. Then the corresponding push-based dynamics model we focus on can be rewritten as: \begin{equation} \begin{aligned} \label{Main-push} \dfrac{{\rm d}i_{v}(t)}{{\rm d}t}=f_{v}(i)=-\beta_v i_{v}(t) +\bigg[1-\prod_{u\in N_{v}}\big(1-\gamma_{uv} i_{u}(t)\big)\bigg]\big(1-i_{v}(t)\big). \end{aligned} \end{equation} As for $\beta_v$, we select two reactive defense strategy with different control cost for our parameter switching method, including one relatively strict defense strategy with higher control cost, denoted by $\beta_{+}$, and one relatively relaxed defense strategy with lower control cost, denoted by $\beta_{-}$. Then we apparently have $\beta_{+}>\beta_{-}$. These two strategies should satisfy the conditions that they are both able to make the dynamics converge to equilibrium zero. The key difference between them is, comparing to our target convergence speed with regard to the dynamics to be controlled, the dynamics with $\beta_v=\beta_{+}$ should converge faster than the target speed for all nodes in $V$, while the dynamics with $\beta_v=\beta_{-}$ may converge more slowly than the target speed. The classical approach to control the dynamics through adjusting the reactive defense strategy is the trivial method that forcing $\beta_{v}=\beta_{+}$ during the entire control process, which is inefficient because it may cost too many defense resources to maintain the relatively strict reactive defense strategy. Besides, the convergence speed of the dynamics under high-cost control may be faster than what we actually need, causing redundancies and wastes of defense resources to some extent. As discussed above, both the relatively strict defense strategy and the relatively relaxed defense strategy need to be able to make the dynamics converge to equilibrium zero. So a safe method to get equilibrium zero is to let $\beta_v/\gamma_{max}\ge\lambda_{A,1}$ for both $\beta_v=\beta_{+}$ and $\beta_v=\beta_{-}$, where $\gamma_{max}$ denotes the maximum value of probability $\gamma_{uv}$ for all neighbor nodes pair $(u,v)\in E$. But due to the variety of parameter $\gamma_{uv}$, $\gamma_{max}$ can be relatively large in practice, leaving little choice for parameter $\beta_v$. Therefore, the second step of the pre-control process is to force $\gamma_{max}$ to a relatively small value, which means we need to permanently reinforce the preventive defense strategy for the nodes which are more vulnerable to push-based attacks launched by the attacker (e.g., a stronger network firewall or filter is deployed). After the two pre-control steps, in which all the pull-based attacks have been eliminated and $\gamma_{max}$ is relatively small, we now employ an parameter switching method to control the convergence speed with relatively low cost of defense resources. We use an event-based mechanism to define the cost-saving parameter switching rule. \subsection{Notations} Table \ref{table:model-variables-and-parameters} summarizes the major notations used in the paper. \begin{table}[!htbp] \centering \caption{Notations used throughout the paper.\label{table:model-variables-and-parameters}} \begin{tabular}{|c|p{0.8\textwidth}|} \hline $I_n$ & the $n*n$ identity matrix\\ \hline $\mathbb{R}$ & the set of real numbers\\ \hline $\mathbb{N}$ & the set of positive integers and zero\\ \hline $\|i\|_{1}$ & $\|i\|_{1}=\sum_{v=1}^{n}\|i_{v}\|$ is the $l_{1}$-norm for an $n$-dimensional vector $i=[i_{1},\ldots,i_{n}]\in\mathbb{R}^n$. Note that the result equally holds with respect to other norms.\\ \hline $G=(V,E), A$ & the attack-defense graph structure $G$ with adjacency matrix $A=[a_{vu}]_{n\times n}$ where $a_{vu}=1$ if and only if $(u,v)\in E$\\ \hline $N_{v}$ & $N_{v}=\{u\in V:(u,v)\in E\}$\\ \hline $\alpha_v \in [0,1]$ & the probability that {\em secure}\ node $v$ becomes {\em compromised}\ because pull-based attack penetrates preventive defense \\ \hline $\beta_v \in (0,1]$ & the probability that {\em compromised}\ node $v$ becomes {\em secure}\ because reactive defense detects and cleans compromise \\ \hline $\gamma_{uv} \in (0,1]$ & the probability that a {\em compromised}\ neighbor node $u$ wages a successful push-based attack against {\em secure}\ node $v$ \\ \hline $\gamma_{max} \in (0,1]$ & the maximum value of probability $\gamma_{uv}$ for all neighbor nodes pair $(u,v)\in E$ \\ \hline $\beta_{-}$, $\beta_{+}$ & the value of parameter $\beta_v$ for all nodes $v\in V$ in low-cost (high-cost) reactive defense setting\\ \hline $i_v(t)$, $i(t)$ & the probability $v$ is in {\em compromised}\ state at time $t$; $i(t)=[i_{1}(t),\cdots,i_{n}(t)]$ \\ \hline $\varphi_{up}(s)$, $\varphi_{up}(s)$ & the decision functions that trigger high-cost (low-cost) control events\\ \hline $t_{k}^{v}$, $\tau_{k}^{v}$ & the time for the $k$-th high-cost (low-cost) control event at $v\in V$ in the event-based control method; $t_{1}^{v}=\tau_{0}^{v}=0$ \\ \hline $T_{-}^v$, $T_{+}^v$ & the total time of maintaining the parameter $\beta_v$ of the target node in low-cost (high-cost) reactive defense setting during control process\\ \hline $\mathcal{S}(t)$, $\mathcal{S}_{v}(t)$ & the exponential speed index of the convergence speed of $i(t)$ ($i_v(t)$) when the dynamics converge exponentially\\ \hline $\chi_v(t)$ & the sample-state of node $v$ at time $t$; 0 means {\em secure} and 1 means {\em compromised} \\ \hline $\widehat{i_{v}(t)}$, $\widehat{s_{v}(t)}$ & the probability $v$ is in the {\em compromised}\ ({\em secure}) state at time $t$ as estimated from the sample-states \\ \hline $\mathcal{W}$, $\mathcal{W}^{'}(t)$ & the time window with fixed (adaptive) time length for estimating the probability-states from sample-states\\ \hline \end{tabular} \end{table} \section{An Event-based Parameter Switching Method} \label{sec:proof-of-problem} In this section, we initiate an event-based parameter switching method to control the convergence speed to our target speed, with relatively low cost of defense resources. That is to say, we switch the reactive defense strategy between two predefined settings in practice, according to the event-based trigger rule we proposed. We first show that the global dynamics under control will converge to equilibrium zero at our target convergence speed. Then we prove that there is no Zeno behavior during the entire control process. \subsection{Designing Event-based Parameter Switching Rule} \label{subsec:Design-Rule} We apply the control method on all nodes of the network system. But for the purpose of clarification, we focus on one target node $v$ to explain the control process corresponding to the decentralized control manner as discussed above. We switch the parameter $\beta_v$ of the target node $v$ between two different groups of parameters alternately, which are the low-cost parameter $\beta_v=\beta_{-}$ and the high-cost parameter $\beta_v=\beta_{+}$. Here the footnotes 'lc' stands for 'low cost' and 'hc' stands for 'high cost'. We use $T_{-}^v$ and $T_{+}^v$ respectively to denote the total time of maintaining the two parameters $\beta_{-}$ and $\beta_{+}$ during the entire control process for node $v\in V$. We want the dynamics to converge to zero at target convergence speed with a relatively low total cost, which means to make the mean of time ratio $\frac{1}{n}\sum_{v\in V}\frac{T_{+}^v}{T_{+}^v+T_{-}^v}$ relatively small. Before defining the event-trigger rule, let us review these definition and lemma. For all nodes $v\in V$, let $n*n$ real matrix $K=\{\gamma_{uv}\}_{u,v=1}^n$, where $\gamma_{vv}=0$ for all nodes $v\in V$, and let non-singular $n*n$ real diagonal matrix $B=diag(\{\beta_v\}_{v=1}^n)$. Notice that $\beta_v$ takes values in $\{\beta_{-},\beta_{+}\}$ for all nodes $v\in V$. Let $B_{+}=\beta_{+}I_n$ and $B_{-}=\beta_{-}I_n$, where $I_n$ is $n*n$ identity matrix. Then we have following lemma. \begin{lemma} \label{lemma-get-m-matrix} If there exists some positive constant number $\iota$, so that $J=B_{+}-K-\iota I_n$ is a monotone matrix (M-matrix), then there exists a positive diagonal matrix $P=diag(\{p_v\}_{v=1}^n)$ such that $[(-J)P+P^T(-J)]$ is negative definite. \end{lemma} Apparently, our target convergence speed should be faster than the dynamics with $\beta_v=\beta_{-}$ for all nodes in $V$, and slower than the dynamics with $\beta_v=\beta_{+}$ for all nodes in $V$. We can set our target convergence speed as $C{\rm e}^{-\iota t}$ as long as $\iota>0$ satisfies Lemma \ref{lemma-get-m-matrix}, where $C$ is a positive constant number. We will prove the effectiveness later. In our event-based parameter switching method, we switch the parameter $\beta_v$ when an event is triggered (e.g. certain conditions are satisfied). Between two consecutive events, the parameter $\beta_v$ holds. We define two criterion functions according to our target speed of the convergence process: \begin{eqnarray*} \begin{cases} \displaystyle \varphi_{up}(t)&=~{\rm e}^{-\iota t},{\forall}t\ge0,\\[8pt] \displaystyle \varphi_{low}(t)&=~L*{\rm e}^{-\iota t},{\forall}t\ge0. \end{cases} \end{eqnarray*} where $L$ is an positive constant number satisfying $0<L<1$. Notice that $\varphi_{up}(t)$ represents the ideal convergence process at our target convergence speed. Theoretically, $\varphi_{up}(t)$ can be a polynomial function if the original dynamics converge polynomially. But from the Sard's Lemma in \cite{sard1942measure}, the parameter regime that causing polynomial convergence speed is in a zero measure set and cannot be chosen in practice. So without loss of generality, we let $\varphi_{up}(t)$ be an exponential function for the simplification of narrative. It is also worth noting that $\varphi_{low}(t)$ can be defined in other functional form (only need to satisfy the conditions of convergence speed). With the matrix $P$ defined in Lemma \ref{lemma-get-m-matrix}, we now define rigger rule as follows: \begin{definition}[event-based trigger rule] \label{trigger-rule-control} $P=diag(\{p_v\}_{v=1}^n)$ is as defined in Lemma \ref{lemma-get-m-matrix}. Let $m_v(t)=p_v^{-1}i_v(t)$, then for $k=1,2,\ldots$, the trigger rule is defined as: \begin{itemize} \item if $m_v(0)\ge 1$, then let $t_{1}^{v}=0$, and \begin{eqnarray*} \begin{cases} \displaystyle \tau_{k}^{v}=&\inf\bigg\{s\ge t_{k}^{v}:m_v(s)\le \varphi_{low}(s)\bigg\}\\[8pt] \displaystyle t_{k+1}^{v}=&\inf\bigg\{s\ge \tau_{k}^{v}:m_v(s)\ge \varphi_{up}(s)\bigg\} \end{cases} \end{eqnarray*} \item if $m_v(0)<1$, then let $\tau_{0}^{v}=0$, and \begin{eqnarray*} \begin{cases} \displaystyle t_{k}^{v}=&\inf\bigg\{s\ge \tau_{k-1}^{v}:m_v(s)\ge \varphi_{up}(s)\bigg\}\\[8pt] \displaystyle \tau_{k}^{v}=&\inf\bigg\{s\ge t_{k}^{v}:m_v(s)\le \varphi_{low}(s)\bigg\} \end{cases} \end{eqnarray*} \end{itemize} which specifies two sequences of parameter switching events: \begin{itemize} \item High-cost control event: At time $t_{k}^{v}$, the value of parameter $\beta_v$ switches to $\beta_{+}$ which generates relatively high control cost. \item Low-cost control event: At time $\tau_{k}^{v}$, the value of parameter $\beta_v$ switches to $\beta_{-}$ which generates relatively low control cost. \end{itemize} \end{definition} As discussed above, both $\beta_{-}$ and $\beta_{+}$ need to be able to make the dynamics converge to equilibrium zero in the parameter regime after pre-control. So system \eqref{Main-push} can be written as: for target node $v\in V$, \begin{itemize} \item If $t\in[t_{k}^{v},\tau_{k}^{v})$, \begin{equation} \begin{aligned} \label{high-cost-dynamics} \frac{{\rm d}i_{v}(t)}{{\rm d}t =-\beta_{+} i_{v}(t) +\bigg[1- \prod_{u\in N_{v}}\Big(1-\gamma_{uv} i_{u}(t)\Big)\bigg]\big(1-i_{v}(t)\big), \end{aligned} \end{equation} \item If $t\in[\tau_{k}^{v},t_{k+1}^{v})$, \begin{equation} \begin{aligned} \label{low-cost-dynamics} \frac{{\rm d}i_{v}(t)}{{\rm d}t =-\beta_{-} i_{v}(t) +\bigg[1- \prod_{u\in N_{v}}\Big(1-\gamma_{uv} i_{u}(t)\Big)\bigg]\big(1-i_{v}(t)\big), \end{aligned} \end{equation} \end{itemize} For $k=1,2,\ldots$, we regard the two control steps \eqref{high-cost-dynamics} and \eqref{low-cost-dynamics} during $t\in[t_{k}^{v},t_{k+1}^{v})$ as the $k$-th {\em control cycle}. \subsection{Analyzing the Event-based Parameter Switching Method} Under the control procedure proposed above, we will prove the effectiveness of the event-based parameter switching method, that is, the new dynamics of the target node under control will converge to zero at our target convergence speed, and what's more important, with no Zeno behavior. \begin{theorem} \label{theorem-main} For any node $v\in V$, system \eqref{high-cost-dynamics} \eqref{low-cost-dynamics} generated by the event-based parameter switching control strategy (trigger rule Definition \ref{trigger-rule-control}) will converge to zero at the convergence speed same as $\varphi_{up}(t)$, with no Zeno behavior. \end{theorem} \begin{proof} We first prove that $m_v(t)=p_v^{-1}i_v(t)$ under parameter switching control strategy (trigger rule Definition \ref{trigger-rule-control}) could never exceed $\varphi_{up}(t)$ after time $\tau_{1}^{v}$ for all nodes $v\in V$ (i.e., it always holds that $p_v^{-1}i_v(t)\le\varphi_{up}(t)$ after time $\tau_{1}^{v}$). For any node $v\in V$, with regard to its original dynamic \eqref{Main-push}, let $u_{-}$ be the smallest index in $N_{v}$. Notice that \begin{align*} &~1-\prod_{u\in N_{v}}\Big(1-\gamma_{uv} i_{u}(t)\Big)\\ =&~\bigg[\Big(1-\gamma_{u_{-}v} i_{u_{-}}(t)\Big)+\gamma_{u_{-}v} i_{u_{-}}(t)\bigg] -\prod_{u\in N_{v}}\Big(1-\gamma_{uv} i_{u}(t)\Big)\\ =&~\gamma_{u_{-}v} i_{u_{-}}(t)+\Big(1-\gamma_{u_{-}v} i_{u_{-}}(t)\Big) \bigg[1-\prod_{u>u_{-},u\in N_{v}}\Big(1-\gamma_{uv} i_{u}(t)\Big)\bigg]. \end{align*} This recurrent process will lead to \begin{align*} \frac{{\rm d}}{{\rm d}t}i_{v}(t) =-\beta_v i_{v}(t)+\Big(1-i_{v}(t)\Big)\sum_{\omega\in N_{v}}\gamma_{\omega v}i_{\omega}(t) \prod_{u<\omega,u\in N_{v}}\Big(1-\gamma_{uv} i_{u}(t)\Big) \end{align*} So we have the following inequality \begin{align} \label{inequality-node} \frac{{\rm d}}{{\mathrm d}t}i_{v}(t)\le-\beta_v i_{v}(t)+\sum_{\omega\in N_{v}}\gamma_{\omega v}i_{\omega}(t). \end{align} At time $\tau_1^{v}$, we have $m_v(\tau_1^{v})=\varphi_{low}(\tau_1^{v})<\varphi_{up}(\tau_1^{v})$ for all nodes $v\in V$. Let $M(t)=\max_{v\in V}m_v(t)$. Assume there is some time point $t^{*}$ so that $M(t^{*})=\varphi_{up}(t^{*})$, then for each $v^{*}\in V$ so that $m_{v^{*}}(t^{*})=M(t^{*})$, by inequality \eqref{inequality-node} we have \begin{align*} \frac{{\mathrm d}}{{\mathrm d}t}\bigl[i_{v^{*}}(t){\mathit e}^{\iota t}\bigr]\biggl|_{t=t^{*}} &\le& -\beta_v^{*} i_{v^{*}}(t^{*}){\mathit e}^{\iota t^{*}}+\sum_{\omega\in N_{v^{*}}}\gamma_{\omega v^{*}}i_{\omega}(t^{*}){\mathit e}^{\iota t^{*}}+\iota i_{v^{*}}(t^{*}){\mathit e}^{\iota t^{*}}\\ &=& -\beta_v^{*} p_{v^{*}}m_{v^{*}}(t^{*}){\mathit e}^{\iota t^{*}}+\sum_{\omega\in N_{v^{*}}}\gamma_{\omega v^{*}}p_{\omega^{*}}m_{\omega}(t^{*}){\mathit e}^{\iota t^{*}} +\iota p_{v^{*}}m_{v^{*}}(t^{*}){\mathit e}^{\iota t^{*}}\\ \end{align*} Notice that $m_{v^{*}}(t^{*})\ge m_{\omega}(t^{*})$, so we have \begin{align*} \frac{{\mathrm d}}{{\mathrm d}t}\big[i_{v^{*}}(t){\mathit e}^{\iota t}\big]\bigg|_{t=t^{*}} \le (-\beta_v^{*} p_{v^{*}}+\sum_{\omega\in N_{v^{*}}}\gamma_{\omega v^{*}}p_{\omega^{*}}+\iota p_{v^{*}})m_{v^{*}}(t^{*}){\mathit e}^{\iota t^{*}} \end{align*} According to the event-based trigger rule Definition \ref{trigger-rule-control}, the reactive defense strategy switches to high-cost setting $\beta_+$ right after time point $t^{*}$. Recall Lemma \ref{lemma-get-m-matrix}, $J=B_{+}-K-\iota I_n$ is a monotone matrix (M-matrix) and $[(-J)P+P^T(-J)]$ is negative definite, which leads to $$-\beta_v^{*} p_{v^{*}}+\sum_{\omega\in N_{v^{*}}}\gamma_{\omega v^{*}}p_{\omega^{*}}+\iota p_{v^{*}}<0$$ So we have $$\frac{{\mathrm d}}{{\mathrm d}t}\big[i_{v^{*}}(t){\mathit e}^{\iota t}\big]\bigg|_{t=t^{*}}<0$$ This implies that $m_{v^{*}}(t){\mathit e}^{\iota t}=p_{v}^{-1}i_{v^{*}}(t){\mathit e}^{\iota t}$ is strictly decreasing after time point $t^{*}$. In the mean time, notice that $\varphi_{up}(t){\mathit e}^{\iota t}=1$, and $m_{v^{*}}(t^{*})=\varphi_{up}(t^{*})$. Finally we have $m_{v}(t)\le\varphi_{up}(t)$ for all nodes $v\in V$ at any time point $t$. Next we prove that the parameter switching events will continue to exist till infinite time. That is to say, $t\to+\infty$ implies $k\to+\infty$ for both $\{t_{k}^{v}\}$ and $\{\tau_{k}^{v}\}$ (i.e., there are infinitely many parameter switching events). Besides, we prove that there is no Zeno behavior during the entire control process. For the convenience of clarification, let ${\mathit e}^{-t\mathcal{S}_{+}^{v}(t)}$ be the convergence speed of dynamics \eqref{high-cost-dynamics} when $\beta_v=\beta_+$ (i.e., the high-cost control), and ${\mathit e}^{-t\mathcal{S}_{-}^{v}(t)}$ be the convergence speed of dynamics \eqref{low-cost-dynamics} when $\beta_v=\beta_-$ (i.e., the low-cost control). Notice that $\mathcal{S}_{+}^{v}(t)$ and $\mathcal{S}_{-}^{v}(t)$ may not be constant numbers, but they always satisfy $\mathcal{S}_{+}^{v}(t)>\iota$ and $\mathcal{S}_{-}^{v}(t)<\iota$. For any $k=1,2,3\cdots$, with respect to the $k$-th {\em control cycle} during $t\in[t_{k}^{v},t_{k+1}^{v})$, We analyze the two control steps respectively. i)\ {\em Step one}: With regard to the high-cost control step \eqref{high-cost-dynamics} during $t\in[t_{k}^{v},\tau_{k}^{v})$, notice that the parameter $\beta_v$ switches to high-cost setting $\beta_+$ since $m_v(t_{k}^{v})=\varphi_{up}(t_{k}^{v})$. Due to $\mathcal{S}_{+}^{v}(t)>\iota$, $i_v(t)$ converges to zero faster than $\varphi_{up}(t)$ and $\varphi_{low}(t)$. This leads to the existence of the next event triggered at $\tau_{k}^{v}$ so that $m_v(\tau_{k}^{v})=\varphi_{low}(\tau_{k}^{v})$, which is a low-cost control event. In the following, we prove that there is no Zeno behavior in the high-cost control step \eqref{high-cost-dynamics} during $t\in[t_{k}^{v},\tau_{k}^{v})$. According to the dynamic, we have \begin{align*} \bigg|\int_{t_{k}^{v}}^{\tau_k^v}\frac{{\rm d}}{{\rm d}t}\Big[m_v(t)\Big]{\rm d}t\bigg|=\bigg|\varphi_{up}(t_{k}^{v})-\varphi_{low}(\tau_{k}^{v})\bigg| \end{align*} For the left side, we have \begin{align*} \bigg|\int_{t_{k}^{v}}^{\tau_k^v}\frac{{\rm d}}{{\rm d}t}\Big[m_v(t)\Big]{\rm d}t\bigg| \le~&p_v^{-1}\int_{t_{k}^{v}}^{\tau_k^v}\bigg|\frac{{\rm d}}{{\rm d}t}\Big[i_v(t)\Big]\bigg|{\rm d}t\\ \le~&M\int_{t_{k}^{v}}^{\tau_k^v}\bigg|{\rm e}^{-t\mathcal{S}_{+}^{v}(t)}\bigg|{\rm d}t\\ \le~& M {\rm e}^{-\iota t_k^v}(\tau_{k}^{v}-t_{k}^{v}), \end{align*} where $M$ is a positive constant. For the right side, we have \begin{align*} \bigg|\varphi_{up}(t_{k}^{v})-\varphi_{low}(\tau_{k}^{v})\bigg| =~&i_v(0){\rm e}^{-\iota t_{k}^{v}}-L*i_v(0){\rm e}^{-\iota \tau_{k}^{v}}\\ \ge~&(1-L)i_v(0){\rm e}^{-\iota \tau_{k}^{v}} \end{align*} Then for both sides, we have \begin{align*} (1-L)i_v(0){\rm e}^{-\iota \tau_{k}^{v}} \le M {\rm e}^{-\iota t_k^v}(\tau_{k}^{v}-t_{k}^{v}) \end{align*} Then we have \begin{align*} (1-L)i_v(0){\rm e}^{-\iota (\tau_{k}^{v}-t_{k}^{v})} \le M (\tau_{k}^{v}-t_{k}^{v}) \end{align*} It shows the existence of a positive number $\eta_v$, which is the root of the transcendental equation $(1-L)i_v(0){\rm e}^{-\iota \eta_v}\le M \eta_v$ and satisfies $\tau_{k}^{v}-t_{k}^{v}\ge\eta_{v}$, which essentially means that for every $v\in V$, $\inf\{\tau_{k}^{v}-t_{k}^{v}\}>0$. That is to say, there is no Zeno behavior in the high-cost control step \eqref{high-cost-dynamics} during $t\in[t_{k}^{v},\tau_{k}^{v})$. ii)\ {\em Step two}: With regard to the low-cost control step \eqref{low-cost-dynamics} during $t\in[\tau_{k}^{v},t_{k+1}^{v})$, notice that the parameter $\beta_v$ switches to low-cost setting $\beta_-$ since $m_v(\tau_{k}^{v})=\varphi_{low}(\tau_{k}^{v})$. Due to $\mathcal{S}_{-}^{v}(t)<\iota$, $\varphi_{up}(t)$ and $\varphi_{low}(t)$ converge to zero faster than $i_v(t)$. This leads to the existence of the next event triggered at $t_{k+1}^{v}$ so that $m_v(t_{k+1}^{v})=\varphi_{up}(t_{k+1}^{v})$, which is a high-cost control event. Now we prove that there is no Zeno behavior in the low-cost control step \eqref{low-cost-dynamics} during $t\in[\tau_{k}^{v},t_{k+1}^{v})$. This proof is much simpler than that of the high-cost control step. Notice that $i_v(t)$ converges to zero, so we have \begin{align*} \varphi_{up}(t_{k+1}^{v})\le\varphi_{low}(\tau_{k}^{v}) \end{align*} So we have \begin{align*} {\rm e}^{-\iota (t_{k+1}^{v}-\tau_{k}^{v})}\le L \end{align*} which essentially means that for every $v\in V$, $\inf\{t_{k+1}^{v}-\tau_{k}^{v}\}\ge -\ln{L}/\iota>0$. That is to say, there is no Zeno behavior in the low-cost control step \eqref{low-cost-dynamics} during $t\in[\tau_{k}^{v},t_{k+1}^{v})$. From the proof above, we have shown that $m_v(t)=p_v^{-1}i_v(t)$ continues to touch $\varphi_{up}(t)$ but could never exceed $\varphi_{up}(t)$ till infinite time, so $i_v$ converges to zero at the convergence speed same as $\varphi_{up}(t)$ (the convergence speed here refers to the average speed through time). Note that the proof under periodic reference setting (see also \cite{liu2020using}) is similar. \end{proof} \subsection{Translating Trigger Rule in Definition \ref{trigger-rule-control} to Algorithm} In order to employ the parameter switching control method presented above, we need to translate the event-based trigger rule in Definition \ref{trigger-rule-control} into a control algorithm. For this purpose, we need to observe the states of nodes first. Due to the node heterogeneity (i.e., the parameters are nodes-dependent) of the network system in this paper, the event-based observing method proposed in \cite{liu2020using} is no longer applicable, so we simply use the classical periodic observing method to handle this issue. Besides, with respect to the decentralized control manner of the parameter switching method, we illustrate the algorithm by focusing on one target node $v\in V$. \begin{algorithm}[!htbp] \caption{Event-based parameter switching control process according to the trigger rule in Definition \ref{trigger-rule-control}}\label{algorithm-trigger-control} \textbf{input:} ~$G=(V,E)$, $i_{v}(0)$, $\varphi_{up}$, $\varphi_{low}$, $\beta_{+}$, $\beta_{-}$, $h$\\ \textbf{output:} $\{t_{k}^{v}\}_{k=1}^{+\infty}$ and $\{\tau_{k}^{v}\}_{k=0}^{+\infty}$ for $v\in V$\\ \textbf{initialize:} $t_{1}^{v}\leftarrow 0$ and $\tau_{0}^{v}\leftarrow 0$; $k\leftarrow 1$;\\ Get $p_v>0$ for $v$ as specified in Lemma \ref{lemma-get-m-matrix}\\ $Cycle\leftarrow 0$\\ \While{{\tt true}} $t\leftarrow t_{k}^{v}$\\ \While{$Cycle = 0$}{ \If{$p_v^{-1}i_v(t)\le \varphi_{low}(t)$}{ switch reactive defense strategy of $v$ to $\beta_{-}$\\ $Cycle \leftarrow 1$\\ $\tau_{k}^{v}\leftarrow t$\\ } $t\leftarrow t+h$\\ } \While{$Cycle = 1$}{ \If{$p_v^{-1}i_v(t)\ge \varphi_{up}(t)$}{ switch reactive defense strategy of $v$ to $\beta_{+}$\\ $Cycle \leftarrow 0$\\ $t_{k+1}^{v}\leftarrow t$\\ } $t\leftarrow t+h$\\ } $k \leftarrow k+1$\\ } \end{algorithm} There are four groups of inputs in Algorithm \ref{algorithm-trigger-control}: attack-defense graph $G=(V,E)$; initial values $i_v(0)$ for node $v\in V$; two criterion functions $\varphi_{up}(t)$ and $\varphi_{low}(t)$; two reactive defense strategies with their corresponding parameter values $\beta_{+}$ and $\beta_{-}$ and a step length parameter $h$ (i.e., the constant time interval of the periodic observing method, see also \cite{liu2020using}). \subsection{Numerical Examples} \label{sec:simulation} We use numerical examples to illustrate the convergence process of the dynamics under control. The numerical examples exhibit the the effectiveness of the proposed event-based parameter switching method. The settings of the examples are defined as follows. For graph $G$ in the dynamics model, we conduct experiments on both undirected graph and directed graph to put the method into practice. The following network structures are obtained from \url{http://snap.stanford.edu/data/} Note that the extraction of $G$ in practice demands access to the enterprise's physical network topologies and security policies, which are usually confidential data unavailable to academic researchers. \begin{itemize} \item Enron email network: This is an undirected graph with $|V|=5242$ nodes, $|E|=28980$ edges, maximal node degree $81$ and $\lambda_{A,1}=45.6167$. \item Gnutella peer-to-peer network: This is a directed graph with $|V|=8,717$ nodes, $|E|=31,525$ links, maximal node in-degree $64$ and $\lambda_{A,1}=4.7395$. \end{itemize} We set $\beta_{+}=0.8$, $\beta_{-}=0.1$ with respect to $\beta_v$ for all nodes $v\in V$. As for $\gamma_v$, we randomly select the values for all nodes $v\in V$ with an upper bound $\gamma_{max}=0.002$ for undirected Enron email network and $\gamma_{max}=0.013$ for directed Gnutella peer-to-peer network. With respect to the criterion functions $\{\varphi_{up},\varphi_{low}\}$ of the event-based trigger rule Definition \ref{trigger-rule-control}, we set $\iota=0.5$ and $L=0.5$ for both undirected Enron email network and directed Gnutella peer-to-peer network. Thus, the conditions of the parameter switching method proposed above are satisfied. We calculate the matrix $P=diag(\{p_v\}_{v=1}^n)$ in Lemma \ref{lemma-get-m-matrix} for each graph respectively. As for the initial values, each node $v\in V$ is assigned with an initial compromise probability $i_v(0)\in_R [0,1]$ where $\in_R$ means sampling uniformly at random. Besides, we consider $t\in[0,500]$ with a fixed step-length $h=0.025$. The convergence processes of dynamics are shown in Figure \ref{fig:prob-state-control}. Notice that the grey curve (i.e., the dynamics under control) refers to $i_v(t)$, while the the blue curve (i.e., the adjusted control target) refers to $m_v(t)=p_v^{-1}i_v(t)$. \begin{figure}[!htbp] \centering \subfigure[Node 1855 of undirected Enron email network]{\includegraphics[width=0.46\textwidth]{UD_Dynam1855.png}\label{Undirected1}}\hspace{2ex} \subfigure[Node 2923 of undirected Enron email network]{\includegraphics[width=0.46\textwidth]{UD_Dynam2923.png}\label{Undirected2}}\hspace{2ex} \subfigure[Node 1187 of directed Gnutella peer-to-peer network]{\includegraphics[width=0.46\textwidth]{D_Dynam1186.png}\label{Directed1}}\hspace{2ex} \subfigure[Node 6992 of directed Gnutella peer-to-peer network]{\includegraphics[width=0.46\textwidth]{D_Dynam6992.png}\label{Directed2}} \caption{The convergence processes of dynamics under parameter switching control for both undirected and directed graph. \label{fig:prob-state-control}} \end{figure} Figure \ref{fig:prob-state-control} exhibits the control process of the proposed event-based parameter switching method and is consistent with the proof of Theorem \ref{theorem-main}. Then we verify that the convergence speed (i.e., the average convergence speed through time) of the dynamics under control is close to the target speed given by the criterion function $\varphi_{up}$. In order to confirm the result, we define the following indicator of convergence speed and name it {\em exponential speed index}:$$\mathcal{S}(t)=-\frac{1}{\Delta t}\ln\frac{i(t+\Delta t)}{i(t)}.$$ Notice that the exponential speed index of the criterion function is equal to $\iota=0.5$ in our settings, which also represents the target speed index. The exponential convergence speed indexes of the dynamics $i(t)=[i_{1}(t),\cdots,i_{n}(t)]$ under parameter switching control for both undirected and directed graph are shown in Figure \ref{fig:convergence-speed-prob}. Notice that the value of the blue line $\mathcal{S}$ refers to the whole time average of the green curve for $t\in[0,500]$ and should be close to the red line (i.e., the target speed index). \begin{figure}[!htbp] \centering \subfigure[The convergence speed index of undirected Enron email network]{\includegraphics[width=0.46\textwidth]{UD_Speed.png}\label{Undirected1}}\hspace{2ex} \subfigure[The convergence speed index of directed Gnutella peer-to-peer network]{\includegraphics[width=0.46\textwidth]{D_Speed.png}\label{Directed2}} \caption{The exponential convergence speed indexes of the dynamics under parameter switching control for both undirected and directed graph. \label{fig:convergence-speed-prob}} \end{figure} For the presented convergence speed experiments, the threshold of effectiveness is defined as $\frac{|\mathcal{S}-\iota|}{\iota}$, which should be less than $10\%$. For the undirected Enron email network, we have $\frac{|\mathcal{S}-\iota|}{\iota}=3.72\%$, and for the directed Gnutella peer-to-peer network, we have $\frac{|\mathcal{S}-\iota|}{\iota}=2.60\%$, which shows the effectiveness. Next, it comes to the control cost, which is indicates by the mean of time ratio $\frac{1}{n}\sum_{v\in V}\frac{T_{+}^v}{T_{+}^v+T_{-}^v}$. For the classical control approach without parameter switching, it holds that $\beta_v=\beta_{+}$ for $t\in[0,500]$ for all nodes $v\in V$. Suppose the control cost is equal to $1$. From the experiments, by using the event-based parameter switching method, the control cost is equal to $0.50$ in the case of undirected Enron email network and $0.53$ in the case of directed Gnutella peer-to-peer network respectively. That is to say, the new control method should save at least $40\%$ of the cost incurred by the classical approach. So we conclude that the event-based parameter switching method can reduce more than $40\%$ of the control cost compared with the classical approach, which shows the efficiency. \section{Putting the Event-based Method into Practice} \label{sec:gap} Similar to \cite{liu2020using}, we need to bridge the gap between the following two kinds of states for utilizing the event-based parameter switching control method in practice. In the aforementioned model, the state of node $v\in V$ at time $t$ is represented by $i_v(t)$, namely the probability that $v$ is in {\em compromised}\ state at time $t$. In practice, this state is often measured as a Boolean value, with ``0'' indicating $v$ is {\em secure}\ but vulnerable and ``1'' indicating $v$ is {\em compromised}. In other words, the {\em sample-state} of node $v\in V$ at time $t$ can be denoted by \begin{align} \label{0-1-State} \chi_{v}(t)= \begin{cases} 0 & v \text{ is in the {\em secure}\ state at time $t$}\\ 1 & v \text{ is in the {\em compromised}\ state at time $t$}. \end{cases} \end{align} This difference underlines the gap between the probability-states in the model and the sample-states in practice. It is worth noticing that, the algorithm proposed in \cite{liu2020using}, which estimates the probability-states via 0-1 state ergodic process, can not be simply transferred and applied to the current control problem. Unlike the equilibrium estimation task in which the accuracy of estimation is of vital importance, the timeliness is the main focus in our event-based parameter switching dynamics control. Without the timeliness, the event triggered by rule may suffer from time lag , which could invalidate the event-based control method. So despite the 0-1 state sequences over the whole time (i.e., $[0,t]$) may provide higher accuracy of probability estimation, yet we should avoid using them in the present task. Instead, we propose a new algorithm which takes advantage of 0-1 state sequences within a time window. \subsection{Estimation via 0-1 State Sequences within a Time Window} Motivated by the theorem of two-valued processes introduced in \cite[Chapter~1]{parzen1999stochastic}, we propose a new method to bridge the aforementioned gap by obtaining an estimation $\widehat{i_{v}(t)}$ of probabilities $i_{v}(t)$ and an estimation $\widehat{s_{v}(t)}$ of probabilities $s_{v}(t)$ from a 0-1 state sequence within a time window, as indicated by \eqref{0-1-State}. We design a new algorithm for the event-based parameter switching method from the theorem. Firstly, let us review the theorem of two-valued processes. We use the Lebesgue measure $\mathcal{M}$ to define \begin{align} \label{Random-Variables-T} \begin{cases} \displaystyle \mathcal{T}_{v0}(t)=\mathcal{M}\big(\big\{\tau\le t:\chi_{v}(\tau)=0\big\}\big)\\[8pt] \displaystyle \mathcal{T}_{v1}(t)=\mathcal{M}\big(\big\{\tau\le t:\chi_{v}(\tau)=1\big\}\big). \end{cases} \end{align} Theorem \ref{Theorem:0-1} below shows how to generate probabilities $\widehat{i_{v}(t)}$ and $\widehat{s_{v}(t)}$ from a 0-1 state ergodic process over time. \begin{theorem}[\cite{parzen1999stochastic}] \label{Theorem:0-1} Let $\{\chi_{v}(t),t>0\}$ for $v\in V$ be a 0-1 state ergodic process. Let \begin{eqnarray*} \begin{cases} \displaystyle \widehat{s_{v}(t)} \frac{\mathcal{T}_{v0}(t)}{t}\\[8pt] \displaystyle \widehat{i_{v}(t)} \frac{\mathcal{T}_{v1}(t)}{t}, \end{cases} \end{eqnarray*} then it is obvious that $\lim_{t\to+\infty}\big[\mathbb{P}\big(\chi_{v}(t)=0\big)-\widehat{s_{v}(t)}\big] = 0$, $\lim_{t\to+\infty}\big[\mathbb{P}\big(\chi_{v}(t)=1\big)-\widehat{i_{v}(t)}\big] = 0$. \end{theorem} In \cite{liu2020using}, $\widehat{i_{v}(t)}$ and $\widehat{s_{v}(t)}$ is used to estimate $i_{v}(t)$ and $s_{v}(t)$ at sufficiently large time $t$ respectively. However, as the time-averaged estimations of probability state, $\widehat{i_{v}(t)}$ and $\widehat{s_{v}(t)}$ suffer from an increasing time lag as $t\to\infty$. Due to the convergence property of the preventive and reactive cyber defense dynamics, the time-averaged estimation still converges to the original equilibrium, which makes it effective in the scenario of \cite{liu2020using}. But in our present scenario, we need a new estimation method with less time lag and faster reaction. For simplicity, sometimes $\mathcal{W}$ also represents the time window which covers the $\mathcal{W}$ most recent time points before $t$. Let \begin{align} \label{Random-Variables-T} \begin{cases} \displaystyle \mathcal{T}_{v0}^{\mathcal{W}}(t)=\mathcal{M}\big(\big\{t-\mathcal{W}<\tau\le t:\chi_{v}(\tau)=0\big\}\big)\\[8pt] \displaystyle \mathcal{T}_{v1}^{\mathcal{W}}(t)=\mathcal{M}\big(\big\{t-\mathcal{W}<\tau\le t:\chi_{v}(\tau)=1\big\}\big)\\[8pt] \end{cases} \end{align} Notice that the smaller $\mathcal{W}$ implies the less time lag and the lower accuracy the probability estimation exhibits. So we face the immediacy-accuracy trade-off dilemma when selecting the proper time window $\mathcal{W}$. It is also worth noting that a simple way to enhance the estimation accuracy is to increase the sampling frequency $h$ (i.e., the constant time interval of the periodic observing method, see also \cite{liu2020using}). But there is always a limitation on the sampling frequency in practice. In order to cope with this trade-off dilemma, we later propose the design of adaptive time windows, which shows great performance. In order to simulate a 0-1 ergodic process for $\forall v\in V$, the paper samples node $v$ at time $t$ by its compromise probability $i_v(t)$: \begin{align} \label{RandomChoice} \chi_{v}(t)=H\big[i_v(t)-Rand(0,1)\big] \end{align} where $Rand(0,1)$ means drawing a random real number uniformly from $[0,1]$, and $H$ is the Discrete Heaviside step function: \begin{align} \label{Heaviside} H(x)= \begin{cases} 0 & x<0\\ 1 & x\ge0. \end{cases} \end{align} \subsection{Using the Event-based Control Method in Practice} With the aid of the newly proposed estimation method with adaptive time windows, which exhibits less time lag and faster response, the gap between probability-states and sample-states has been properly bridged. We now use the aforementioned event-based parameter switching method in practice to control the preventive and reactive cyber defense dynamics. Since undirected networks are a special case of directed networks, only experiments on the directed Gnutella peer-to-peer Network are performed here. Let the time window $\mathcal{W}=30$. As proposed above, we use an adaptive time window $$\mathcal{W}^{'}(t)=\max(\mathcal{W},\frac{t}{C_0})$$ to replace the original time window $\mathcal{W}$ with a fixed time length, where $C_0$ is a positive constant number. In the settings of our numerical examples, let $C_0=3$ for $\mathcal{W}^{'}(t)$. Figure \ref{fig:sample-state-control-adaptive} shows the convergence process of the dynamics under control, using a time window with adaptive time length to estimate the probability $m_v(t)$. We can find that there are still some events triggered in the advanced stage $t\in [300,500]$ when probability $i_v(t)$ has fully converged to equilibrium zero, which means the event-based parameter switching method becomes effective by using the adaptive time window. With respect to the convergence speed indexes, notice that the threshold for it defined in Section \ref{sec:simulation} is $10\%$. Figure \ref{speed-adaptive-window} shows the effectiveness of the time window with adaptive time length, with $\frac{|\mathcal{S}-\iota|}{\iota}=6.79\%<10\%$, while $\frac{|\mathcal{S}-\iota|}{\iota}=27.27\%>10\%$ for the time window with fixed time length exhibited in Figure \ref{speed-fixed-window}. Besides, with respect to the control cost, which is indicated by the mean of time ratio $\frac{1}{n}\sum_{v\in V}\frac{T_{+}^v}{T_{+}^v+T_{-}^v}$, recall the statement in Section \ref{sec:simulation}, the threshold of efficiency is defined as $40\%$ of the cost. The control cost in the current numerical example is $0.60$, which means the reactive defense setting is under low-cost setting for $40\%$ of the time.(the low-cost reactive defense setting takes up $40\%$ of the time) \begin{figure}[!htbp] \centering \subfigure[Node 2688 using a time window with adaptive time length]{\includegraphics[width=0.46\textwidth]{Discrete_Adaptive_Dynam2688.png}\label{Discrete1}}\hspace{2ex} \subfigure[Node 4011 using a time window with adaptive time length]{\includegraphics[width=0.46\textwidth]{Discrete_Adaptive_Dynam4011.png}\label{Discrete2}} \caption{The convergence processes of dynamics under parameter switching control aiming at adjusted sample-states estimation, using a time window with adaptive time length. \label{fig:sample-state-control-adaptive}} \end{figure} \begin{figure}[!htbp] \centering \subfigure[The convergence speed index, using a time window with fixed time length]{\includegraphics[width=0.46\textwidth]{Speed_Fixed_Window.png}\label{speed-fixed-window}}\hspace{2ex} \subfigure[The convergence speed index, using a time window with adaptive time length]{\includegraphics[width=0.46\textwidth]{Speed_Adaptive_Window.png}\label{speed-adaptive-window}} \caption{The exponential convergence speed indexes of the dynamics under parameter switching control, using a time window with fixed or adaptive time length. \label{fig:speed-fixed-window-or-not}} \end{figure} \section{Conclusion} \label{sec:conclusion} In this paper, an event-based parameter switching method is proposed for the control tasks of cybersecurity, which helps avoid excessive control costs as well as guarantees the dynamics to converge as our desired speed. The Zeno-free property is proved, implying the feasibility of the method. Meanwhile, we designed a new estimation method with adaptive time windows in order to bridge the gap between the probability state and the sampling state with less time lags. Both theoretical and practical experiments are given to illustrate the parameter switching method, which show the effectiveness and efficiency of our new method. There are many open problems for future research: Do there exist some better event-based parameter switching trigger rule so as to save more defense resources and guarantee the convergence speed? Can this parameter switching approach be applied to other dynamics control? Furthermore, similar parameter switching approaches regarding the parameter $\gamma$ for push-based attacks are worth further studying.

7d9d2c5740049633a952dda116962ec4c680360a

\section{Introduction} There are many interesting connections between representation categories of vertex operator algebras and quantum groups associated to the same simple Lie algebra $\mathfrak{g}$. The most prominent example is the work of Kazhdan and Lusztig \cite{KL1,KL2,KL3,KL4}, where they proved a braided equivalence between representation categories of affine Lie algebras at generic level and quantum groups at an appropriate corresponding parameter $q$. A vertex algebra fulfilling sufficient finiteness conditions gives rise to a braided tensor category of representations \cite{Hu1, HLZ}. Over the last two decades logarithmic theories received increased attention. The word logarithmic comes from conformal field theory and refers to logarithmic singularities in correlation functions, see \cite{CR} for an introduction. This happens if the representation category of the underlying vertex operator algebra is non-semisimple. It is in general a difficult problem to completely understand the braided tensor category of a given vertex algebra. Our approach is to use structure of the vertex algebra and its tensor category that is fairly accessible in order to characterize a quasi Hopf algebra with braided equivalent tensor category. First examples of vertex algebras with non semi-simple representation categories were the triplet algebras $\mathcal{W}(p)$ and our main focus are correspondences of triplet algebras and related vertex algebras and quasi Hopf modifications of quantum groups. Connections between the triplet vertex operator algebra and restricted quantum groups of $\mathfrak{sl}_2$ at roots of unity first appeared in the work of Feigin, Gainutdinov, Semikhatov, and Tipunin \cite{FGST1,FGST2,FGST3,FGST4}. They conjectured therein the ribbon equivalence of particular representation categories of $\mathcal{W}(p)$ and the small quantum group $u_q(\mathfrak{sl}_2)$ at $2p$-th root of unity $q$, and an abelian equivalence between these categories was later claimed to be shown by Nagatomo and Tsuchiya \cite{NT}. Proofs were however lacking and details have finally and very recently been filled in by McRae and Yang \cite{MY}, which used the work of Adamovic, Milas and others \cite{TW, A, AM1, AM2, AM3, AM4, CJORY,CMY}. For a general Lie algebra $\mathfrak{g}$, the {\it Logarithmic Kazhdan Lusztig Conjecture} is a conjectured ribbon equivalence between the representation categories of higher rank analogs of the triplet algebras and a quantum group $u_q(\mathfrak{g})$ and a $2p$-th root of unity $q$ \cite{FT, AM5,L, S}. However, in this form the conjecture cannot be true, because at even roots of unity the category of representations of a quantum group is typically not a (non-semisimple) modular tensor category. For example for $\mathfrak{sl}_2,p>2$ this was demonstrated by Kondo and Saito \cite{KS}. It was then shown by Gainutdinov, Runkel and the first author that there exists a quasi Hopf algebra $\tilde{u}_q(\mathfrak{sl}_2)$ whose underlying algebra is $u_q(\mathfrak{sl}_2)$ \cite{FGR2,CGR}, and whose representation category is a non-semisimple modular tensor category. This conjecture is a consequence of the conjectural correspondence between the singlet algebra $\mathcal{M}(p)$ and the unrolled restricted quantum group $u_q^H(\mathfrak{sl}_2)$ at $2p$-th root of unity $q$. This conjecture first appeared in \cite{CGP}, and has been motivated in \cite{CM,CMR,CGR,Ru}. To be precise, it is expected that the category of finite-dimensional weight $u_q^H(\mathfrak{sl}_2)$-modules $\mathcal{C}^H$ should be ribbon equivalent to the smallest subcategory of singlet modules generated by irreducibles with respect to finite sums, tensor products, and quotients. In fact, it was this conjecture that motivated the work of \cite{CGR}. The triplet can be realized as a simple current extension of the singlet, and its representation category is therefore expected to correspond to the category of local modules $\mathrm{Rep}^0 \mathcal{A}_p$ for some commutative algebra object in $\mathcal{A}_p \in \mathcal{C}^H$. The quasi Hopf algebra constructed in \cite{CGR} is precisely the one which realizes $\mathrm{Rep}^0 \mathcal{A}_p$ as its own category of modules. The modified Kazhdan Lusztig conjecture was then that the representation category of this quasi Hopf modification is braided tensor equivalent to the category of the triplet algebra. \\ For arbitrary Lie algebras $\mathfrak{g}$ and even order of $q$ such a modified quasi Hopf algebra $\tilde{u}_q(\mathfrak{g})$ that gives rise to a modular tensor category was constructed by Gainutdinov, Ohrmann and the second author \cite{GLO} and from the perspective of de-equivariantization by Negron \cite{N}. Our construction proceeds in the following way, which is inspired by the Andruskiewitsch Schneider program \cite{AS} for classifying pointed Hopf algebras: It starts with the correct modular tensor category of vector spaces graded by the Cartan part, an abelian group of here even order, which is a quasi Hopf algebra. On the vertex algebra side, this is the semisimple modular tensor category of representations of the lattice vertex algebra underlying the free field construction. Then one extends this category by Nichols algebras \cite{H} inside the category, which correspond the the Borel parts. On the vertex algebra side, these are the algebras of screening operators \cite{L}. As for Lie algebras, between these two categories there is an adjunctions of induction to Verma modules and restriction to weight spaces. On the vertex algebra side, these is reversed an adjunction of restriction from the lattice vertex algebra and induction. \\ We aim to develop a general technology to prove braided tensor equivalences between categories of modules of vertex algebras and quasi Hopf algebras. Indeed, there are many more interesting vertex algebras that are expected to be related to quasi Hopf modifications of quantum groups. The most obvious examples are those which are closely related to the triplet such as the singlet algebra $\mathcal{M}(p)$, and $\mathcal B_p$ vertex algebras \cite{CRW, ACKR}. The singlet is the $U(1)$-orbifold of the triplet and the $\mathcal B_p$-algebras are quantum Hamiltonian reductions of $\mathfrak{sl_{p-1}}$ at level $-(p-1)^2/p$ for the subregular nilpotent element \cite{ACGY}. Their Heisenberg coset is $\mathcal{M}(p)$ \cite{A2, CRW}. We mention here especially the $\mathcal B_p$-algebras as they have relaxed-highest weight modules and spectral flow twists thereof, i.e. modules that neither have finite dimensional conformal weight spaces nor is the conformal weight necessarily lower bounded. Affine vertex (super)algebras have similar representations (if the level is not a positive integer) and understanding their representation theory is of major importance. \subsection{From vertex operator algebras to quasi Hopf algebras} We aim to develop a general formalism to characterize the quasi Hopf algebra structure of representation categories of vertex operator algebras. \subsubsection{The vertex algebra set-up} We start by listing assumptions that should hold for general classes of vertex algebras with non semi-simple representation category. \begin{enumerate} \item Let $\mathcal{W}$ be a vertex operator algebra and $\mathcal{C}= {\rm Rep}(\mathcal{W})$ a vertex tensor category of $\mathcal{W}$-modules. In particulary it is a braided tensor category. We also assume that every object in $\mathcal{C}$ is rigid, has integral Frobenius-Perron dimension and that this category is locally finite. Let $U$ be an algebra and $\Psi_\mathcal{W} : {\rm Rep}(U) \rightarrow \mathcal{C}$ an equivalence of abelian categories that preserves the Frobenius Perron dimensions. \item There is an abelian group $L$ and non-degenerate quadratic form $Q$ on $L$, such that the quasi-fiber functor $ \mathcal{C} \rightarrow \mathrm{Vect}$ factors through $\mathrm{Vect}_L^Q$. \item There is a family of embeddings $\iota_a: \mathcal{W} \rightarrow \mathcal{V}$ of vertex operator algebras of finite index for $a= 1, \dots, n$, such that ${\rm Rep}(\mathcal{V}) \cong \mathrm{Vect}_L^Q$. In particular there are induction functors $\mathcal{F}_a: {\rm Rep}(\mathcal{W}) \rightarrow {\rm Rep}(\mathcal{V})^{\text{tw}}$ to a category ${\rm Rep}(\mathcal{V})^{\text{tw}}$ that contains ${\rm Rep}(\mathcal{V})$ as subcategory. The right adjoint $\mathcal{G}_a$ is just the restriction functor that forgets the structure of the larger algebra $\mathcal{V}$. The right adjoint of a monoidal functor is oplax and so in particular it provides an lax tensor functor $\mathcal{G}_a : {\rm Rep}(\mathcal{V})^{\text{tw}} \rightarrow {\rm Rep}(\mathcal{W})$. Especially the restriction to ${\rm Rep}(\mathcal{V})$ of $\mathcal{G}$ is oplax. We require that \[ \bigotimes_{a=1}^n \mathcal{G}_a( {\rm Rep}(\mathcal{V})) \] is a projective generator of ${\rm Rep}(\mathcal{W})$. \item The semi-simple part of the ribbon twist on objects in ${\rm Rep}(\mathcal{W})$ is known. \end{enumerate} We now comment on these assumptions and explain that they are satisfied in the main example of current interest, the triplet $\mathcal{W}(p)$. The theory of vertex tensor categories is due to Huang, Lepowsky and Zhang \cite{HLZ} and there are a few general Theorems ensuring the existence of vertex tensor categories. \begin{enumerate} \item $C_2$-cofinite vertex operator algebras of positive energy \cite{Hu1}. \item Vertex operator algebras with the properties that all irreducible ordinary modules are $C_1$-cofininte and all generalized Verma modules are of finite length \cite{CY}. \item Let $V \subset W$ be vertex operator algebras, such that their conformal vectors coincide. If $W$ is an object in a suitable completion of a vertex tensor category $\mathcal{C}_V$ of $V$-modules, then the category of $W$-modules that lie in $\mathcal{C}_V$ is a vertex tensor category as well \cite{CKM, CMY2}. \end{enumerate} In particular the triplet algebra $\mathcal{W}(p)$ is $C_2$-cofinite \cite{AM2} and obviously of positive energy. It is believed but unproven that the category of ordinary modules of the singlet algebra $\mathcal{M}(p)$ falls into the second type, see section 6 of \cite{CMR}. Currently only vertex tensor category is known for a subcategory \cite{CMY} by using the last approach and vertex tensor category results of ordinary modules of the Virasoro algebra \cite{CJORY}. There is no general rigidity Theorem for non-rational VOAs, however it has been explicitly verified for $\mathcal{W}(p)$ \cite{TW} and for the just mentioned subcategory of $\mathcal{M}(p)$ \cite{CMY}. Integrality of Frobenius-Perron dimensions follows from fusion rules that are determined in both cases \cite{TW, CMY, MY}. The abelian equivalence between ${\rm Rep}(\mathcal{W}(p))$ and ${\rm Rep}(u_q(\mathfrak{sl}_2))$ is settled as mentioned above. The Cartan subalgebra $C \subset u_q(\mathfrak{sl}_2)$ has as representation category $\mathrm{Vect}_L^Q$. Verifying that the quasi-fiber functor factors through $\mathrm{Vect}_L^Q$ is just a fusion rule computation that in our case follows from the knowledge that fusion rules in ${\rm Rep}(\mathcal{W}(p))$ and ${\rm Rep}(u_q(\mathfrak{sl}_2))$ coincide for the monoidal structure on ${\rm Rep}(u_q(\mathfrak{sl}_2))$ determinend in \cite{CGR} and since the fiber functor for ${\rm Rep}(u_q(\mathfrak{sl}_2))$ clearly factors through $\mathrm{Vect}_{\mathbb Z_{2p}}^Q$ the same must be true for ${\rm Rep}(\mathcal{W}(p))$. Finally, the ribbon twist on VOA-modules is determined by conformal weights, which are among the first quantities one is usually able to determine. It remains the family of oplax tensor functors. We prove \begin{theoremX}[\ref{thm:oplax}] Let $\mathcal{C}_{\mathcal{W}(p)}$ be the category of modules of the triplet $\mathcal{W}(p)$ vertex operator algebra for integer $p>1$ and let $\mathcal{C}_{V_L}$ be the category of modules of the lattice vertex algebra of the lattice $L= \sqrt{2p}\mathbb Z$. Then there exist (not additive) subcategories $\mathcal{V}, \overline{\mathcal{V}}$ of $\mathcal{C}_{\mathcal{W}(p)}$ with the properties that $\mathcal{V} \boxtimes \overline{\mathcal{V}}$ is a projective generator of $\mathcal{C}_{\mathcal{W}(p)}$ and there are surjective oplax tensor functors $\mathcal{G} : \mathcal{C}_{V_L} \rightarrow \mathcal{V}$ and $\mathcal{G}^\sigma : \mathcal{C}_{V_L} \rightarrow \overline{\mathcal{V}}$. \end{theoremX} We prove this by twisting the usual embedding of the triplet in $V_L$ by an automorphism and then compute relevant fusion rules. We used the results of \cite{CMY}. We find, more explicitly spoken, the fusion product of Verma module and twisted (opposite) Verma module is projective. \subsubsection{The characterization Theorem} With the above assumptions on the vertex algebra $\mathcal{W}$ and its associated algebra $U$ in place, we consider by Tannakian reconstruction a quasi bialgebra structure on $U$, such that $\mathcal{C} = {\rm Rep}(\mathcal{W})$ is tensor equivalent ${\rm Rep}(U)$. First we prove a slightly enhanced ``relative" version of Tannakian reconstruction relevant to our situation:\\ \begin{propositionX}[\ref{prop2}] Let $\mathcal{C}, \mathcal{D}$ be a finite tensor categories such that there exists a quasi-fiber functor $F_{\mathcal{D}}:\mathcal{D} \to \mathrm{Vec}$ and an essentially surjective quasi-tensor functor $G:\mathcal{C} \to \mathcal{D}$. Then we have an inclusion $C \subset U$ of the associated quasi bialgebra $U=\mathrm{End}(F_{\mathcal{D}} \circ G)$, $C=\mathrm{End}(F_{\mathcal{D}})$ preserving the coproduct, counit, and algebra structure. \end{propositionX} Then we prove that if the category can be ``covered" with images of oplax tensor functors, then the previous inclusion is an inclusion of quasi bialgebras resp. $G$ is a true tensor functor. In essence, the proof reflects the fact that the regular representation of a bialgebras is a coalgebra in the respective tensor category, but for a true quasi bialgebra this fails. \\ \begin{theoremX}[\ref{Catthm}] Let $\mathcal{C}, \mathcal{D}$ be a finite tensor categories such that there exists a quasi-fiber functor $F_{\mathcal{D}}:\mathcal{D} \to \mathrm{Vec}$ and an essentially surjective quasi-tensor functor $G:\mathcal{C} \to \mathcal{D}$ as in Proposition \ref{prop2}. Suppose we have oplax monoidal functors $\mathbb{V}_k:\mathrm{Rep}(C) \to \mathrm{Rep}(U)$, $k=1,...,n$ such that $\bigotimes_k \mathbb{V}_k(C_{reg})$ is a projective generator of $\mathrm{Rep}(U)$, such that in addition the coproduct $\delta$ induced by the coproduct of the regular $C$-$C$-bimodule $C_{reg}$ sends the image of a generator $1$ to a $U$-$U$-bimodule generator. Then there exist a twist $U^J$ of $U$, such that $\Phi_{U^J}=\Phi_C$ and thus $C=\mathrm{End}(F_{\mathcal{D}})$ is a sub quasi bialgebra of $U^J$. Moreover there is a unique choice of such a twist such that $\delta(1)=1\otimes 1$. \end{theoremX} The technical use of this theorem is that it fixes a twist representative of $U$ and that it provides us a-priori with $U$-$C$-comodule coalgebras $\mathbb{V}_k(C_{reg})$ with a known associator $\Phi_U=\Phi_C$. Eventually, we can then determine these coalgebras explicitly and thus our category in question. \subsubsection{The correspondence of $\mathcal{W}(2)$ and $\tilde{u}_i(\mathfrak{sl}_2)$} Now we consider the vertex algebra $\mathcal{W}=\mathcal{W}(2)$, the algebra $U=u_i(\mathfrak{sl}_2)=\tilde{u}_i(\mathfrak{sl}_2)$, its Cartan subalgebra $C=\mathbb{C}[\mathbb{Z}_4]$, and the lattice $L=\sqrt 2 A_1$. Denote by $\mathbb{V},\overline{\mathbb{V}}:\mathrm{Rep}(C) \to {\rm Rep}(U)$ the oplax tensor functors which send an irreducible $C$-module $\mathbb{C}_k$ of weight $i^k$ to the corresponding Verma and opposite Verma modules $M_k$ and $\overline{M}_k$ of highest weight $i^k$. Then we have the following \begin{theorem}\label{classif} The abelian category ${\rm Rep}(U)$ admits up to equivalence a unique structure of a braided tensor category such that $\mathbb{V},\overline{\mathbb{V}}:{\rm Rep}(C) \to {\rm Rep}(U)$ are braided oplax tensor functors with the braiding on $C$ given by the quadratic form $Q(k)=\beta^{(k^2)}$, for fixed $\beta^4=-1$. \end{theorem} \noindent It follows from Theorem \ref{thm:oplax} that the triplet algebra $\mathcal{W}(2)$ admits two such oplax tensor functors $\mathbb{V},\overline{\mathbb{V}}:{\rm Rep} V_L \to {\rm Rep} \mathcal{W}(2)$ where the $Q$ is determined by the quadratic form of the lattice $L=\sqrt{2} A_1$, so $\beta=e^{2\pi i(1/8)}$. It then follows that the abelian category $\mathrm{Rep}(U)$ of the Kazhdan-Lusztig dual $U$ of $\mathcal{W}(2)$ must admit the functors described in Theorem \ref{classif}. We therefore obtain as a corollary a proof of the Kazhdan-Lusztig conjecture for $\mathcal{W}(2)$: \begin{corollary} There is a braided tensor equivalence of ${\rm Rep}(\mathcal{W}(2))$ to the representation category of the previously constructed quasi-Hopf algebra for this value of $\beta$. The quasi Hopf algebra coincides with the one constructed in \cite{FGR2}. \end{corollary} The proof has the following steps: \begin{itemize} \item We have already established that $C$ is a sub quasi bialgebra of $U$, identified with the Cartan part $\mathbb{C}[K]/(K^ 4-1)$. We have four possibilities parametrized by $\beta^4=-1$, two of which are equivalent. \item In Proposition \ref{prop_Vreg} we determine the possible coproducts on the $U$-$C$-coalgebras $\mathbb{V}(C_{reg}),\delta$ and $\overline{\mathbb{V}}(C_{reg}),\bar{\delta}$ fulfilling in addition $\delta(1)=\bar{\delta}(1)=1\otimes 1$. What we find matches the expectation of a dual Nichols algebra of type $\mathfrak{sl}_2$ (a truncated polynomial ring) and a family of twists of the Cartan part parametrized by $c^a,\bar{c}^a$. \item From this we can directly compute in Proposition \ref{prop_EF} all coproducts $\Delta(E),\Delta(F)$. Some general arguments show that there are no mixed $E$-$F$-terms, so the coproducts are determined by the previous steps and their compatibility implies a relation between $c^a,\bar{c}^a$. We also use the existence of an antipode for a nonzero condition. \item In Proposition \ref{prop_automorphism} we apply algebra automorphisms to reduce $\Delta(F)$ to a standard form and find a $1$-parameter family of coproducts $\Delta(E)$ parametrized by a scalar $d$ (and a sign $\epsilon$ that disappears after switching $E,F$). These are precisely all quasi Hopf algebras resp. tensor categories with the desired properties. For $d=\pm i$ we recover the quasi Hopf algebra in \cite{FGR2} (for the negative sign after switching $K,K^ {-1}$) \item In Proposition \ref{braiding} we compute all extensions of the $R$-matrix of $C$ to $U$. We find that our previous quasi Hopf algebras are only braidable if $d=\pm i$, which concludes our result. \end{itemize} \subsection{Outlook} We have seen that our characterization of the category of the $\mathcal{W}(2)$-algebra already uniquely up to twist equivalence determines the quasi Hopf algebra structure on ${\rm Rep}(\tilde{u}_i(\mathfrak{sl}_2))$. It is work in progress to show that the same happens for general $\mathcal{W}(p)$. We hope to report on this in the near future. Next, one would like to settle the equivalence between the singlet algebras and the unrolled quantum groups. For this the vertex tensor category results of \cite{CMY} need to be extended to the complete category of ordinary modules. We remark that none of our results require finiteness assumptions on the category, as long as we start with a known equivalence of abelian categories to representations over an algebra, finite or infinite dimensional. More generally, we would like to apply our techniques to the logarithmic Kazhdan Lusztig conjecture and beyond: Many vertex algebras characterized as intersections of kernels of sets of screening charges acting on lattice vertex algebras, Heisenberg vertex algebras, or some more complicated vertex algebras. The logarithmic Kazhdan Lusztig conjecture is concerned with the Feigin-Tipunin algebras \cite{FT, S}. These are the natural higher rank generalizations of the triplet $\mathcal{W}(p)$, and their representation category is conjectured to be equivalent to the representation category of a corresponding quasi quantum group. Higher rank generalizations of the singlet and $\mathcal B_p$-algebras also exist \cite{CM2, C}, and they are equally nicely characterized as kernels of screenings on free field vertex algebras. Before applying our technology to these cases much of the representation theory of these algebras has to first be studied. There is however already one very interesting example whose representation theory is well enough understood, namely the affine vertex superalgebra of $\mathfrak{gl}_{1|1}$ \cite{CMY}. Unlike the Lie algebra case, its category of ordinary modules has uncountable infinitely many inequivalent simple objects. We note that nonetheless we will try to employ our technology to determine the quasi Hopf superalgebra whose representation category is braided equivalent to the category of ordinary modules of the affine vertex superalgebra of $\mathfrak{gl}_{1|1}$. The underlying algebra must of course be the quantum supergroup of $\mathfrak{gl}_{1|1}$, i.e. we expect the linear equivalence to be easy to see. This example would thus provide a first Kazhdan-Lusztig correspondence for an affine vertex superalgebra and a quantum supergroup. Again it is clear from \cite{L} that the respective screening algebras are the Nichols algebras and it is natural to expect that they give precisely the coalgebra structures on the image of the oplax tensor functors. Other more exotic Nichols algebras over abelian groups or in more complicated braided tensor categories point to more exotic examples beyond super Lie algebras. \\ \begin{remark} Terry Gannon and Cris Negron informed us that they proved the equivalence between $\mathcal{W}(p)$-mod and the quasi Hopf modification of $\tilde{u}_q(\mathfrak{sl}_2)$-mod at $2p$-th root of unity \cite{GN}. Their approach uses a universal property of the Temperly-Lieb category \cite{O} and de-equivariantization \cite{N} and is very different from ours. \end{remark} \section{The Singlet and Triplet Vertex Operator Algebras} Let $\mathcal{C}_{\mathcal{W}(p)}$ be the category of modules of the triplet $\mathcal{W}(p)$ vertex operator algebras for integer $p>1$ and let $\mathcal{C}_{V_L}$ be the category of modules of the lattice vertex algebra of the lattice $L= \sqrt{2p}\mathbb Z$. The aim of this section is to prove the following Theorem, which summarizes Corollary \ref{cor:projective} and Proposition \ref{prop:oplax}. \begin{theorem}\label{thm:oplax} There exist subcategories $\mathcal{V}, \overline{\mathcal{V}}$ of $\mathcal{C}_{\mathcal{W}(p)}$ with the properties that $\mathcal{V} \boxtimes \overline{\mathcal{V}}$ is a projective generator of $\mathcal{C}_{\mathcal{W}(p)}$ and there are surjective oplax tensor functors $\mathcal{G} : \mathcal{C}_{V_L} \rightarrow \mathcal{V}$ and $\mathcal{G}^\sigma : \mathcal{C}_{V_L} \rightarrow \overline{\mathcal{V}}$. \end{theorem} We review the singlet $\mathcal{M}(p)$ and triplet $\mathcal{W}(p)$ vertex operator algebras for integer $p>1$. We restrict to the category $\mathcal C^0_{\mathcal{M}(p)}$ of \cite{CMY}, since the category $\mathcal{C}_{\mathcal{W}(p)}$ of the triplet algebra is obtained from this one via vertex algebra extension. Moreover the category $\mathcal C^0_{\mathcal{M}(p)}$ is braided monoidal and rigid \cite{CMY} and most fusion rules are known. We will determine missing ones in a moment. Let $\alpha_+ = \sqrt{2p}, \alpha_- = -\sqrt{2/p}$, $\alpha_0 = \alpha_+ + \alpha_-$ and \[ \alpha_{r, s} = \frac{1-r}{2} \alpha_+ + \frac{1-s}{2}\alpha_- \] Then the simple modules of the singlet algebra are denoted by $M_{r, s}$ for integers $r, s$ and in addition $1\leq s \leq p$. The top level of these modules has conformal weight \[ h_{r, s} = \frac{\alpha_{r, s}(\alpha_{r, s}- \alpha_0)}{2} \] and so especially the ribbon twist on these modules is $e^{2\pi i h_{r, s}}\text{Id}_{M_{r, s}}$. The projective cover of $M_{r, s}$ is denoted by $P_{r, s}$. The modules $M_{r, p}$ are projective, $M_{r, p} = P_{r, p}$. There are then further modules $F_{r, s}$ and $\overline{F}_{r, s}$ uniquely characterized by fitting in the non-split exact sequence \begin{equation}\label{ses1} 0 \rightarrow M_{r, s} \rightarrow F_{r, s} \rightarrow M_{r+1, p-s} \rightarrow 0, \qquad 0 \rightarrow M_{r, s} \rightarrow \overline{F}_{r, s} \rightarrow M_{r-1, p-s} \rightarrow 0 \end{equation} for $s \neq p$ and $F_{r, p} \cong \overline{F}_{r, p} \cong M_{r, p} \cong P_{r, p}$. The $F_{r, s}$ and $\overline{F}_{r, s}$ are submodules and quotients of $P_{r, s}$ and fit in the non-split exact sequences \begin{equation}\label{ses2} 0 \rightarrow F_{r, s} \rightarrow P_{r, s} \rightarrow F_{r-1, p-s} \rightarrow 0, \qquad 0 \rightarrow \overline{F}_{r, s} \rightarrow P_{r, s} \rightarrow \overline{F}_{r+1, p-s} \rightarrow 0 \end{equation} for $s\neq p$. The $M_{n+1, 1}$ are simple currents and satisfy the fusion rules \begin{equation} \begin{split} M_{n+1, 1} \boxtimes M_{r, s} &\cong M_{r+n, s} \\ M_{n+1, 1} \boxtimes P_{r, s} &\cong P_{r+n, s} \\ M_{n+1, 1} \boxtimes F_{r, s} &\cong F_{r+n, s} \\ M_{n+1, 1} \boxtimes \overline{F}_{r, s} &\cong \overline{F}_{r+n, s}. \end{split} \end{equation} These fusion rules are proven in \cite{CMY}, see especially section 3.2. The only one not stated there is the last one, but it follows immediately from simple currents preserving non-split short exact sequences, see Proposition 2.5 of \cite{CKLR}. The triplet algebra $\mathcal{W}(p)$ has simple modules $W_{r, s}$ with $r \in \{1, 2\}$ and $s \in \{ 1, \dots, p\}$. The ribbon twist on these modules is $e^{2\pi i h_{r, s}}\text{Id}_{W_{r, s}}$. The module $W_{r, p}$ is projective and otherwise one has the extensions characterized by the non-split exact sequences \begin{equation}\label{Wses1} 0 \rightarrow W_{r, s} \rightarrow V_{r, s} \rightarrow W_{3-r, p-s} \rightarrow 0, \qquad 0 \rightarrow W_{r, s} \rightarrow \overline{V}_{r, s} \rightarrow W_{3-r, p-s} \rightarrow 0. \end{equation} The modules $V_{r, s}$ and $\overline{V}_{r, s}$ are constructed as follows. The triplet algebra is \[ \mathcal{W}(p) \cong \bigoplus_{n \in \mathbb Z} M_{1+2n, 1} \] as a singlet module. In particular there is an induction functor $\mathcal{F} : \mathcal{C}^0_{\mathcal{M}(p)} \rightarrow \mathcal{C}_{\mathcal{W}(p)}$ that satisfies \cite{CMY} \begin{equation} \begin{split} \mathcal{F}(M_{r, s}) &= W_{\bar r, s} \\ \mathcal{F}(P_{r, s}) &= R_{\bar r, s} \\ \mathcal{F}(F_{r, s}) &= V_{\bar r, s} \\ \mathcal{F}(\overline{F}_{r, s}) &= \overline{V}_{\bar r, s} \\ \end{split} \end{equation} Here $\bar r = 2$ if $r$ is even and $\bar r =1$ if $r$ is odd. The modules $R_{\bar r, s}$ are all projective and in particular the induction of a projective module is always projective. The last two can be taken as the definiton of the corresponding triplet modules and the short-exact sequences follow from exactness of $\mathcal{F}$, see \cite{CKM}. The right adjoint of $\mathcal{F}$ is the forgetful functor $\mathcal{G}$. It satisfies \[ \mathcal{G}(\mathcal{F}(M)) \cong \bigoplus_{n \in \mathbb Z} M_{2n+1, 1} \boxtimes M \] for an object $M$ in $\mathcal{C}^0_{\mathcal{M}(p)}$. The important properties are that $\mathcal{F}$ is monoidal and that Frobenius reciprocity holds, that is \begin{equation} \begin{split} \mathcal{F}(M \boxtimes M') &\cong \mathcal{F}(M) \boxtimes \mathcal{F}(M')\quad \text{and} \quad \\ \text{Hom}_{\mathcal{C}_{\mathcal{W}(p)}}(\mathcal{F}(M), X) &\cong \text{Hom}_{\text{Ind}(\mathcal{C}^0_{\mathcal{M}(p)})}(M, \mathcal{G}(X)) \end{split} \end{equation} for objects $M, M'$ in $\mathcal{C}^0_{\mathcal{M}(p)}$ and $X$ in $\mathcal{W}(p)$. Here $\text{Ind}(\mathcal{C}^0_{\mathcal{M}(p)})$ is the direct limit completion of $ \mathcal{C}^0_{\mathcal{M}(p)}$ \cite{CMY2} (alternatively one can also take the direct sum completion of \cite{AR}). Frobenius reciprocity tells us that $V_{r, s}$ and $\overline{V}_{r, s}$ are not isomorphic since \[ \text{Hom}_{\mathcal{C}_{\mathcal{W}(p)}}(V_{r, s} , \overline{V}_{r, s}) \cong \text{Hom}_{\mathcal{C}_{\mathcal{W}(p)}}(\mathcal{F}(F_{r, s}) , \mathcal{F}(\overline{V}_{r, s})) = \text{Hom}_{\text{Ind}(\mathcal{C}^0_{\mathcal{M}(p)})}(F_{r, s}, \bigoplus_{n \in \mathbb Z} \overline{F}_{r+2n, s} ) = 0. \] \subsection{Fusion Rules} We need the fusion rules for $F_{r, s}$ with $\overline{F}_{r', s'}$. Define \[ P_{r', s', r, s} := \bigoplus_{\substack{ \ell =2p+1-s-s' \\ \ell +s + s' \ \text{odd}}}^{p} P_{r+r'-1, \ell}. \] \begin{theorem} The fusion rules \begin{equation} \begin{split} M_{r', s'} \boxtimes M_{r, s} &=P_{r', s', r, s} \oplus \bigoplus_{\substack{ \ell = |s-s'| +1 \\ \ell +s + s' \ \text{odd}}}^{\text{min}\{s+s'-1, 2p-1-s-s' \}} M_{r+r'-1, \ell} \\ M_{r', s'} \boxtimes F_{r, s} &=P_{r', s', r, s} \oplus P_{r', s', r+1, p-s } \oplus \bigoplus_{\substack{ \ell = |s-s'| +1 \\ \ell +s + s' \ \text{odd}}}^{\text{min}\{s+s'-1, 2p-1-s-s' \}} F_{r+r'-1, \ell} \\ M_{r', s'} \boxtimes \overline{F}_{r, s} &=P_{r', s', r, s} \oplus P_{r', s', r-1, p-s } \oplus \bigoplus_{\substack{ \ell = |s-s'| +1 \\ \ell +s + s' \ \text{odd}}}^{\text{min}\{s+s'-1, 2p-1-s-s' \}} \overline{F}_{r+r'-1, \ell} \\ \end{split} \end{equation} hold. \end{theorem} \begin{proof} The first fusion rule of the Theorem is part of the main Theorem of \cite{CMY}. The other rules follow from these together with rigidity and the structure of the projective modules. We start with the fusion rules $M_{n+1,1} \boxtimes M_{r, s} = M_{r+n, 1}$, i.e. the $M_{n_1, 1}$ are simple currents. These preserve non-split exact sequences \cite[Prop. 2.5 ]{CKLR} and thus $M_{n+1,1} \boxtimes F_{r, s} = F_{r+n, 1}$. Next, we use \begin{equation} \begin{split} M_{1, 2} \boxtimes M_{r, s} &= \begin{cases} M_{r, 2} & \quad s=1 \\ M_{r, s-1} \oplus M_{r, s+1} &\quad 1<s<p \\ P_{r, p-1} &\quad s=p \end{cases} \\ M_{1, 2} \boxtimes P_{r, s} &= \begin{cases} P_{r, 2} \oplus P_{r+1, p} \oplus P_{r-1, p} & \quad s=1 \\ P_{r, s-1} \oplus P_{r, s+1} &\quad 1<s<p-1 \\ P_{r, p-2} \oplus 2P_{r, p} &\quad s=p-1 \end{cases} \end{split} \end{equation} Consider the case $s=1$, We tensor \eqref{ses1} and \eqref{ses2} with $M_{1, 2}$ and since $\mathcal C^0_{\mathcal{M}(p)}$ is rigid we obtain the short exact sequences \begin{equation}\label{ses3} \begin{split} &0 \rightarrow M_{r, 2} \rightarrow M_{1, 2} \boxtimes F_{r, 1} \rightarrow M_{r+1, p-2} \oplus M_{r+1, p} \rightarrow 0, \qquad \\ &0 \rightarrow M_{r-1, p-2} \oplus M_{r-1, p} \rightarrow M_{1, 2} \boxtimes F_{r-1, p-1} \rightarrow M_{r, 2} \rightarrow 0, \qquad \\ &0 \rightarrow M_{1, 2} \boxtimes F_{r, 1} \rightarrow P_{r, 2} \oplus P_{r+1, p} \oplus P_{r-1, p} \rightarrow M_{1, 2} \boxtimes F_{r-1 , p-1} \rightarrow 0 \end{split} \end{equation} Recall that $M_{r \pm 1, p} \cong P_{r \pm1, p}$ and hence we get the fusion rule \[ M_{1, 2} \boxtimes F_{r, 1} = F_{r, 2} \oplus P_{r+1,p}, \qquad M_{1, 2} \boxtimes F_{r, p-1} = F_{r, p-2} \oplus P_{r,p} \] Let now $1 < s < p-1$. Tensoring \eqref{ses1} and \eqref{ses2} with $M_{1, 2}$ we obtain the non-split short exact sequences \begin{equation}\label{ses3} \begin{split} &0 \rightarrow M_{r, s-1} \oplus M_{r, s+1} \rightarrow M_{1, 2} \boxtimes F_{r, s} \rightarrow M_{r+1, p-s-1} \oplus M_{r+1, p-s+1} \rightarrow 0, \qquad \\ &0 \rightarrow M_{r-1, p-s-1} \oplus M_{r-1, p-s+1} \rightarrow M_{1, 2} \boxtimes F_{r-1, p-s} \rightarrow M_{r, s-1} \oplus M_{r, s+1} \rightarrow 0, \qquad \\ &0 \rightarrow M_{1, 2} \boxtimes F_{r, s} \rightarrow P_{r, s-1} \oplus P_{r, s+1} \rightarrow M_{1, 2} \boxtimes F_{r-1 , p-s} \rightarrow 0 \end{split} \end{equation} From which we get \[ M_{1, 2} \boxtimes F_{r, s} = F_{r, s-1} \oplus F_{r, s+1}. \] In a rigid tensor category projective modules form a tensor ideal. Denote by $\mathcal{H}$ the functor to the quotient category. It is monoidal. We have \begin{equation}\label{ses3} \begin{split} \mathcal{H}(M_{1, 2}) \boxtimes \mathcal{H}(M_{r, s}) &= \begin{cases} \mathcal{H}(M_{r, 2}) & \quad s=1 \\ \mathcal{H}(M_{r, s-1}) \oplus \mathcal{H}(M_{r, s+1}) &\quad 1<s<p \end{cases}, \\ \mathcal{H}(M_{1, 2}) \boxtimes \mathcal{H}(F_{r, s}) &= \begin{cases} \mathcal{H}(F_{r, 2}) & \quad s=1 \\ \mathcal{H}(F_{r, s-1}) \oplus \mathcal{H}(F_{r, s+1}) &\quad 1<s<p \end{cases}. \end{split} \end{equation} Since the fusion rules $M_{r', s'} \boxtimes M_{r, s}$ are uniquely determined by associativity, commutativity, the simple currents $M_{n+1, 1}$, and the fusion rules of $M_{1, 2} \boxtimes M_{r, s}$, the same must be true for the fusion rules in the quotient by projective modules and hence the same must be true for the fusion rules $M_{r,' s'} \boxtimes F_{r, s}$, i.e. we get \[ M_{r', s'} \boxtimes F_{r, s} =P^+_{r', s', r, s} \oplus \bigoplus_{\substack{ \ell = |s-s'| +1 \\ \ell +s + s' \ \text{odd}}}^{\text{min}\{s+s'-1, 2p-1-s-s' \}} F_{r+r'-1, \ell} \] for some projective module $P^+_{r', s', r, s}$. Comparing with $M_{r', s'} \boxtimes ( M_{r, s} \oplus M_{r+1, p-s})$ then gives the claim. The argument for the fusion rule $M_{r', s'} \boxtimes \overline{F}_{r, s}$ is essentially the same and we don't repeat it. \end{proof} A few special cases are \begin{equation} \begin{split} M_{r',1} \boxtimes F_{r, 1} &= F_{r+r'-1, 1} \\ M_{r'-1,p-1} \boxtimes F_{r, 1} &= F_{r+r'-2, p-1} \oplus \bigoplus_{\substack{\ell = 3 \\ \ell \ \text{odd}}}^p P_{r+r'-1, \ell}\\ M_{r,1} \boxtimes \overline{F}_{r', 1} &= \overline{F}_{r+r'-1, 1} \\ M_{r+1,p-1} \boxtimes \overline{F}_{r, 1} &= \overline{F}_{r+r', p-1} \oplus \bigoplus_{\substack{\ell = 3 \\ \ell \ \text{odd}}}^p P_{r+r'-1, \ell} \end{split} \end{equation} Using the short exact sequences characterizing $F_{r, 1}$ and $\overline{F}_{r', 1}$ we get the two short exact sequences for $F_{r, 1} \boxtimes \overline{F}_{r', 1}$ \begin{equation} \begin{split} E_1 &= 0 \rightarrow M_{r', 1} \boxtimes F_{r, 1} \rightarrow \overline{F}_{r', 1} \boxtimes F_{r, 1} \rightarrow M_{r'-1, p-1} \boxtimes F_{r, 1} \rightarrow 0 \\ &= 0 \rightarrow F_{r+r'-1, 1} \rightarrow \overline{F}_{r', 1} \boxtimes F_{r, 1} \rightarrow F_{r+r'-2, p-1} \oplus \bigoplus_{\substack{\ell = 3 \\ \ell \ \text{odd}}}^p P_{r+r'-1, \ell} \rightarrow 0 \\ E_2 &= 0 \rightarrow \overline{F}_{r', 1} \boxtimes M_{r, 1} \rightarrow \overline{F}_{r', 1} \boxtimes F_{r, 1} \rightarrow \overline{F}_{r', 1} \boxtimes M_{r+1, p-1} \rightarrow 0 \\ &= 0 \rightarrow \overline{F}_{r+r'-1, 1} \rightarrow \overline{F}_{r', 1} \boxtimes F_{r, 1} \rightarrow\overline{F}_{r+r', p-1} \oplus \bigoplus_{\substack{\ell = 3 \\ \ell \ \text{odd}}}^p P_{r+r'-1, \ell} \rightarrow 0 \end{split} \end{equation} It follows that $\overline{F}_{r', 1} \boxtimes F_{r, 1}$ equals $\bigoplus\limits_{\substack{\ell = 3 \\ \ell \ \text{odd}}}^p P_{r+r'-1, \ell}$ plus a summand that has the same composition factors as $F_{r+r'-1, 1} \oplus F_{r+r'-2, p-1}$, that is $M_{r+r'-1, 1}$ with multiplicity two and $M_{r+r'-2, p-1}$ and $M_{r+r', p-1}$ once as composition factors. Moreover this module has to contain both $F_{r+r'-1, 1}$ and $\overline{F}_{r+r'-1, 1}$ as submodules and both $F_{r+r'-2, p-1}$ and $\overline{F}_{r+r', p-1}$ as quotients. Clearly the only possibility for this is $P_{r+r'-1, 1}$, i.e. \begin{equation}\label{fus1} \overline{F}_{r', 1} \boxtimes F_{r, 1} = \bigoplus_{\substack{\ell = 1 \\ \ell \ \text{odd}}}^p P_{r+r'-1, \ell} . \end{equation} Since $F_{r, s}$ is a summand of $M_{1, s} \boxtimes F_{r, 1}$ and $\overline{F}_{r', s'}$ is a summand of $M_{1, s'} \boxtimes \overline{F}_{r,' 1}$ and since projective modules form a tensor ideal due to rigidity we have \begin{corollary} The modules $F_{r, s} \boxtimes \overline{F}_{r', s'}$ are projective for all $r, s, r', s'$. \end{corollary} As an example we have \begin{equation}\label{fus2} \begin{split} \overline{F}_{r', 1} \boxtimes F_{r, 2} &=\overline{F}_{r', 1} \boxtimes (F_{r, 1} \boxtimes M_{1, 2}) = (\overline{F}_{r', 1} \boxtimes F_{r, 1}) \boxtimes M_{1, 2} \\ &= \bigoplus_{\substack{\ell = 1 \\ \ell \ \text{odd}}}^p P_{r+r'-1, \ell} \boxtimes M_{1, 2} \\ &= P_{r+r', p} \oplus P_{r+r'-2, p} \oplus \bigoplus_{\substack{\ell = 2 \\ \ell \ \text{even}}}^p 2 P_{r+r'-1, \ell} \end{split} \end{equation} \begin{corollary} The modules $V_{r, s} \boxtimes \overline{V}_{r', s'}$ are projective for all $r, s, r', s'$. \end{corollary} \begin{proof} The induction functor $\mathcal{F}$ is monoidal and maps projectives to projectives. \end{proof} In particular the fusion rules \begin{equation} \begin{split} \overline{V}_{r, 1} \boxtimes V_{1, 1} = \bigoplus_{\substack{\ell = 1 \\ \ell \ \text{odd}}}^p R_{r, \ell} \qquad \text{and} \qquad \overline{V}_{r, 1} \boxtimes V_{1, 2} = 2R_{r+1, p} \oplus \bigoplus_{\substack{\ell = 2 \\ \ell \ \text{even}}}^p 2 R_{r, \ell} \end{split} \end{equation} hold via induction from \eqref{fus1} and \eqref{fus2}. We thus see that every indecomposable projective object appears in the fusion product of some $V_{r, s} \boxtimes \overline{V}_{r', s'}$. Let $\mathcal{V}$ be the subcategory of $\mathcal{C}_{\mathcal{W}(p)}$ whose objects are direct sums of the modules of type $V_{r, s}$ and $\overline{\mathcal{V}}$ the subcategory whose modules are direct sums of the $\overline{V}_{r, s}$. Then this observation can be rephrased as \begin{corollary}\label{cor:projective} $\mathcal{V} \boxtimes \overline{\mathcal{V}}$ is a projective generator of $\mathcal{C}_{\mathcal{W}(p)}$. \end{corollary} \subsection{Free field realizations} Let $\pi$ be the rank one Heisenberg algebra with generator $\alpha(z)$ satisfying \[ \alpha(z)\alpha(w) = (z-w)^{-2}. \] Let $\pi_\lambda$ be the Fock module of highest-weight $\lambda$ and let $L = \alpha_+ \mathbb Z$. Then the singlet and triplet algebras are subalgebras of $\pi$ and $V_L$ characterized as \[ \mathcal{W}(p) = \text{ker}( e_0^{\alpha_-\alpha}: V_L \rightarrow V_{L+\alpha_-}), \qquad \mathcal{M}(p) = \text{ker}(e_0^{\alpha_-\alpha}: \pi \rightarrow \pi_{\alpha_-}). \] Let us denote these embeddings by $\iota$, \[ \iota: \mathcal{W}(p) \hookrightarrow V_L, \qquad \iota: \mathcal{M}(P) \hookrightarrow \pi. \] Especially the triplet is an extension of the singlet, namely \begin{equation} \begin{split} \mathcal{W}(p) &= \bigoplus_{\lambda \in L } \text{ker}(e_0^{\alpha_-\alpha}: \pi_\lambda \rightarrow \pi_{\lambda+\alpha_-}) \\ &= \bigoplus_{n \in \mathbb Z} \text{ker}(e_0^{\alpha_-\alpha}: \pi_{\alpha_{2n+1, 1}} \rightarrow \pi_{\alpha_{2n+2, p-1}}) \\ &= \bigoplus_{n \in \mathbb Z} M_{2n+1, 1}. \end{split} \end{equation} Here we used that \[ M_{r, s} = \text{ker}(e_0^{\alpha_-\alpha}: \pi_{\alpha_{r, s}} \rightarrow \pi_{\alpha_{r+1, p-s}}). \] In fact $\pi_{\alpha_{r, s}}$ satisfies the non-split exact sequence \[ 0 \rightarrow M_{r, s} \rightarrow \pi_{\alpha_{r, s}} \rightarrow M_{r+1, p-s} \rightarrow 0, \] that is $\pi_{\alpha_{r, s}} \cong F_{r, s}$ as $\mathcal{M}(p)$-modules. The triplet vertex algebra is strongly generated by fields $W^\pm, W^0$ together with a Virasoro field. While the singlet is strongly generated by $W^0$ and the Virasoro field. The field $W^+$ is the field associated to the top level vector of $M_{3, 1}$ in $\pi_{\alpha_{3, 1}}$ and the field associated to $W^-$ is the top level vector of $M_{-1, 1}$ in $\pi_{\alpha_{-1, 1}}$. These vertex algebras have automorphisms and we can twist these embeddings by an automorphism. The full automorphism group of the triplet is $PSL(2, \mathbb C)$ \cite{ALM} and especially there is an involution $\sigma$ corresponding to the non-trivial Weyl reflection. The fields $W^\pm, W^0$ carry the adjoint representation of $PSL(2, \mathbb C)$ and the superscript indicates the weight. In particular we can normalize these strong generators such that $\sigma(W^\pm) = W^\mp$ and $\sigma(W^0) = -W^0$. Note that the Virasoro field is invariant under $\sigma$. We see that $\sigma$ restricts to an automorphism of the singlet algebra as well. We denote by $M^\sigma$ the $\sigma$-twisted singlet module corresponding to $M$. Since $\sigma(W^\pm) = W^\mp$ and $W^+$ corresponds to the top level of $M_{3, 1}$ and $W^-$ to the top level of $M_{-1, 1}$ we have $M^\sigma_{3, 1} \cong M_{-1, 1}$. Let $\iota^\sigma := \sigma\circ\iota$. We denote by $\pi^\sigma_\lambda$ the Fock-module $\pi_\lambda$ viewed as an $\mathcal{M}(p)$-module via the embedding $\iota^\sigma$. We have \[ 0 \rightarrow M^\sigma_{r, s} \rightarrow \pi^\sigma_{\alpha_{r, s}} \rightarrow M^\sigma_{r+1, p-s} \rightarrow 0, \] and since $\sigma$ leaves the Virasoro subalgebra invariant one necessarily has $M^\sigma_{r, s} \cong M_{r, s}$ as modules for the Virasoro algebra. This fixes $M_{r, s}^\sigma \in \{ M_{r, s}, M_{2-r, s}\}$. We claim that the second case happens. \begin{proposition} $M_{r, s}^\sigma \cong M_{2-r, s}$ for all $r, s$. Especially $\pi^\sigma_{\alpha_{r, s}} \cong \overline{F}_{r, s}$. \end{proposition} \begin{proof} For this we note that some fusion rules of the singlet algebra have been computed using only the free field realization in section 3.2 of \cite{CMY}. The exact same argument as the one of the proof of Proposition 3.2.4 of \cite{CMY} applies to get \[ M_{3, 1} \boxtimes \pi^\sigma_{\alpha_{r, s}} \cong M_{-1, 1}^\sigma \boxtimes \pi^\sigma_{\alpha_{r, s}} \cong \pi^\sigma_{\alpha_{r-2, s}}. \] Using rigidity we thus get by tensoring the non-split short exact sequence for $\pi^\sigma_{\alpha_{r, s}}$ with $M_{3, 1}$ the non-split exact sequence \[ 0 \rightarrow M_{3, 1} \boxtimes M^\sigma_{r, s} \rightarrow \pi^\sigma_{\alpha_{r-2, s}} \rightarrow M_{3, 1} \boxtimes M^\sigma_{r+1, p-s} \rightarrow 0, \] and hence $M_{3, 1} \boxtimes M^\sigma_{r, s} \cong M_{r-2 s}^\sigma \in \{ M_{r-2, s}, M_{4-r, s}\}$ but $M_{3, 1} \boxtimes M_{r, s} \cong M_{r+2, s}$ and $M_{3, 1} \boxtimes M_{2-r, s} \cong M_{4-r, s}$ so that the only possibility is the claim $M_{r-2, s}^\sigma \cong M_{4-r, s}$. \end{proof} Let $V_{\alpha_{r, s}+L}$ be the $V_L$-module corresponding to the coset $L+\alpha_{r, s}$ of $L$. It becomes an $\mathcal{W}(p)$-module via the embeddings $\iota$ and $\iota^\sigma$ and we denote these by $V_{\alpha_{r, s}+L}$ and $V^\sigma_{\alpha_{r, s}+L}$ respectively. By Frobenius reciprocity we have \[ \text{Hom}_{\mathcal{C}_{\mathcal{W}(p)}}(V_{r, s} , V_{\alpha_{r, s}+L}) \cong \text{Hom}_{\mathcal{C}_{\mathcal{W}(p)}}(\mathcal{F}(F_{r, s}) , V_{\alpha_{r, s}+L}) = \text{Hom}_{\text{Ind}(\mathcal{C}^0_{\mathcal{M}(p)})}(F_{r, s}, \bigoplus_{n \in \mathbb Z} F_{r+2n, s} ) = \mathbb C \] and hence $V_{r, s} \cong V^\sigma_{\alpha_{r, s}+L}$. The same argument for $\overline{V}_{r, s}$ and $V^\sigma_{\alpha_{r, s}+L}$ gives $\overline{V}_{r, s} \cong V^\sigma_{\alpha_{r, s}+L}$. The embeddings $\iota$ and $\iota^\sigma$ provide $V_L$ with the structure of a commuative algebra in $\mathcal{C}_{\mathcal{W}(p)}$ and hence there are induction functors from $\mathcal{C}_{\mathcal{W}(p)}$ to the category of $V_L$-modules that lie in $\mathcal{C}_{\mathcal{W}(p)}$. This category is larger than the category $\mathcal{C}_{V_L}$ of vertex algebra modules for $V_L$, but it contains $\mathcal{C}_{V_L}$ as tensor catgory. Induction is monoidal and so we have oplax tensor functors from $\mathcal{C}_{V_L}$ that we denote by $\mathcal{G}$ and $\mathcal{G}^\sigma$ that map the lattice vertex algebra module $V_{L+\alpha_{r, s}}$ to $V_{r, s}$ respectively $\overline{V}_{r, s}$, i.e. we have \begin{proposition}\label{prop:oplax} There are surjective oplax tensor functors $\mathcal{G} : \mathcal{C}_{V_L} \rightarrow \mathcal{V}$ and $\mathcal{G}^\sigma : \mathcal{C}_{V_L} \rightarrow \overline{\mathcal{V}}$. \end{proposition} \section{Categorical Setup} \subsection{Module Coalgebras} We include here the fundamental definitions and results on module coalgebras required in the following subsection. The content of this subsection can be found in greater detail in \cite[Subsection 4.2]{BCPO}. Throughout, we denote the coassociator, coproduct, and unit of a quasi-bialgebra $H$ by $\Phi_H$, $\Delta_H$, and $\epsilon_H$. \begin{definition} Let $H$ be a quasi-bialgebra and $C$ a left $H$-module. We call $C$ a left $H$-module coalgebra if it is equipped with morphisms \[ \delta:C \to C \otimes C, \qquad \varepsilon:C \to \mathbb{C} \] subject to the following conditions: \begin{align} \Phi_H \cdot (\delta \otimes \mathrm{Id}_C) \circ \delta(c)&=(\mathrm{Id}_C \otimes \delta) \circ \delta(c) \\ \label{Hcou}\varepsilon(c_1)c_2&=\varepsilon(c_2)c_1=c \\ \label{Hcop}\delta(h \cdot c)&=\Delta_H(h) \cdot \delta(c),\\ \varepsilon(h \cdot c)&=\epsilon_H(h)\varepsilon(c), \end{align} for all $c \in C$ and $h \in H$, where we are using Sweedler's notation $\delta(c)=c_1 \otimes c_2$. Right $H$-module coalgebras are defined similarly. \end{definition} We also consider coalgebra objects in categories of bimodules over two distinct Hopf algebras. \begin{definition} Let $H,G$ be Hopf algebras and $C$ a $H$-$G$ bimodule. We call $C$ a $H$-$G$ bimodule coalgebra if is equipped with morphisms \[ \delta:C \to C \otimes C, \qquad \varepsilon:C \to \mathbb{C} \] satisfying \begin{align} \label{HGcoass}\Phi_H \cdot (( \delta \otimes \mathrm{Id}_C) \circ \delta(c) ) \cdot \Phi_G^{-1}&=(\mathrm{Id}_C \otimes \delta) \circ \delta(c)\\ \delta(h \cdot c)=\Delta_H(h) \cdot \delta(c), \; & \; \delta(c \cdot g)=\delta(c) \cdot \Delta_G(g)\\ \varepsilon(c_1)c_2&=\varepsilon(c_2)c_1=c\\ \varepsilon(h \cdot c)=\epsilon_H(h)\varepsilon(c), \; & \; \varepsilon(c \cdot g)=\varepsilon(c)\epsilon_G(g) \end{align} \end{definition} We have the following useful result (\cite[Proposition 4.17]{BCPO}): \begin{proposition}\label{Gaugeprop} Let $H$ be a quasi-bialgebra, $F \in H \otimes H$ a gauge transformation on $H$, and $(C,\delta,\varepsilon)$ a left $H$-module coalgebra. If we define $\delta_F(c)=F\delta(c)$ for all $c \in C$, then $(C,\delta_F,\varepsilon)$ is an $H^F$-module coalgebra, where $H^F$ is the $F$-twist of $H$. \end{proposition} \subsection{Categorical Tools} \label{catsec} We proceed first by deriving some general results on quasi Hopf algebras which will be necessary for the following sections. \begin{proposition}\label{prop1} Let $\mathcal{C}$ be a finite tensor category with integral Frobenius-Perron dimensions and $U$ an algebra such that we have an abelian equivalence $\mathcal{C} \cong \mathrm{Rep}(U)$ identifying Frobenius-Perron dimensions in $\mathcal{C}$ with vector space dimensions in $\mathrm{Rep}(U)$. Then, $U$ can be given the structure of a quasi bialgebra such that $\mathcal{C}$ and $\mathrm{Rep}(U)$ are tensor equivalent. \end{proposition} \begin{proof} Since $\mathcal{C}$ is a finite tensor category with integral Frobenius-Perron dimensions, there exists a quasi-fiber functor $F:\mathcal{C} \to \mathrm{Vec}$ and a quasi bialgebra $\tilde{U}:=\mathrm{End}(F)$ such that $\mathcal{C} \cong \mathrm{Rep}(\tilde{U})$ as tensor categories. It remains only to show that $\tilde{U} \cong U$ as algebras. However, we have an abelian isomorphism $\mathrm{Rep}(U) \cong \mathrm{Rep}(\tilde{U})$ identifying vector space dimensions with Frobenius-Perron dimensions and it follows from the proof of the reconstruction theorem for finite-dimensional quasi bialgebras \cite[Theorem 5.3.12]{EGNO} that the algebra structure of $\tilde{U}$ is uniquely determined by the abelian structure of $\mathrm{Rep}(\tilde{U})$ and Frobenius-Perron dimensions (i.e. the associated quasi-fiber functor). \end{proof} \begin{proposition}\label{prop2} Let $\mathcal{C}, \mathcal{D}$ be a finite tensor categories such that there exists a quasi-fiber functor $F_{\mathcal{D}}:\mathcal{D} \to \mathrm{Vec}$ and an essentially surjective quasi-tensor functor $G:\mathcal{C} \to \mathcal{D}$. Then we have an inclusion $C \subset U$ of the associated quasi bialgebra $U=\mathrm{End}(F_{\mathcal{D}} \circ G)$, $C=\mathrm{End}(F_{\mathcal{D}})$ preserving the coproduct, counit, and algebra structure. \end{proposition} \begin{proof} We have \cite[Subsection 1.10]{EGNO} \begin{align*} \mathrm{Coend}(F_{\mathcal{D}})&= \bigoplus\limits_{Y \in \mathcal{D}} F_{\mathcal{D}}(Y)^* \otimes F_{\mathcal{D}}(Y)/E_{\mathcal{D}}\\ \mathrm{Coend}(F_{\mathcal{D}} \circ G)&= \bigoplus\limits_{X \in \mathcal{C}} F_{\mathcal{D}}(G(X))^* \otimes F_{\mathcal{D}}(G(X))/E_{\mathcal{C}} \end{align*} where $E_{\mathcal{D}}$ is spanned by the elements of the form $y_* \otimes F_{\mathcal{D}}(f)x-F_{\mathcal{D}}(f)^*y_*\otimes x$ for $x \in F_{\mathcal{D}}(X)$, $y_* \in F_{\mathcal{D}}(Y)^*$, $f \in \mathrm{Hom}_{\mathcal{D}}(X,Y)$ and $E_{\mathcal{C}}$ is spanned by elements of the form $y_* \otimes F_{\mathcal{D}}(G(g))^*y_* \otimes x$ for $x \in F_{\mathcal{D}}(G(X))$, $y_* \in F_{\mathcal{D}}(G(Y))^*$, $g \in \mathrm{Hom}_{\mathcal{C}}(X,Y)$. It is clear that $E_{\mathcal{C}} \subset E_{\mathcal{D}}$ and $G$ is essentially surjective, so we have a canonical surjection \[ \mathrm{Coend}(F_{\mathcal{D}} \circ G) \twoheadrightarrow \mathrm{Coend}(F_{\mathcal{D}})\] which yields an inclusion \[ \mathrm{End}(F_{\mathcal{D}}) \hookrightarrow \mathrm{End}(F_{\mathcal{D}} \circ G)\] For a finite tensor $\mathcal{B}$ and quasi-fiber functor $F:\mathcal{B} \to \mathrm{Vec}$, the coproduct and counit on an element $a$ of $\mathrm{End}(F)$ is defined uniquely by the action of $a$ on $F(X \otimes Y)$ and $F(1)$ viewed as modules for $\mathrm{End}(F)$ for any $X,Y \in \mathcal{B}$ \cite[Theorem 5.2.1]{EGNO} so the inclusion preserves the coproduct and counit. \end{proof} \begin{proposition} \label{bimod}~ \begin{itemize} \item[a)] Let $U$ be a quasi bialgebra. If there exists a module coalgebra structure $(U_{reg},\delta)$ on the regular representation $U_{reg}$ of $U$ such that $\delta(1)$ is invertible in $U \otimes U$, then there is a twist equivalent quasi bialgebra $U^J$ such that $\Phi^{J}=1 \otimes 1 \otimes 1$ and $\delta^J$ is the coproduct on $U^J$. \item[b)] Let $U$ be a quasi bialgebra and $C \subset U$ a subalgebra. Let $(U,\delta)$ be a $U$-$C$-bimodule coalgebra such that its structure as a $U$-$C$ bimodule is that of the regular representation of $U$. If $\delta(1)$ is invertible, then there exists a twist $J$ of $U$ such that $\Phi_{U^J}=\Phi_C$ as operators on $U$. \end{itemize} \end{proposition} \begin{proof} a) Since $\delta(1)$ is invertible in $U \otimes U$, it follows from Equation \eqref{Hcou} that $\delta(1)$ is a gauge transformation on $U$ so we can define the twisted quasi bialgebra $U^J$ where $J:=\delta(1)^{-1}$. It then follows from Proposition \ref{Gaugeprop} that $(U_{reg},\delta^J)$ is a module-coalgebra for $U^J$ where $\delta^J(x)=J\delta(x)$ for all $x \in U^J$, so $\delta^J(1)=1 \otimes 1$. Since $\delta^J$ is coassociative, we have $\Phi^J(\delta^J \otimes \mathrm{Id}) \circ \delta^J = (\mathrm{Id} \otimes \delta^J) \circ \delta^J$ so $\Phi^J(1 \otimes 1) \otimes 1=1 \otimes (1 \otimes 1)$ and because $U_{reg}$ is a faithful representation, we have $\Phi^J=1 \otimes 1 \otimes 1$. It follows from Equation \eqref{Hcop} that $\delta^J$ is the coproduct on $U^J$.\\ b) As in a), we have a gauge transformation $J:=\delta(1)^{-1}$ such that $\delta^J(1)=1 \otimes 1$ and $(U_{reg},\delta^J)$ is a $U^J$-$C$ bimodule coalgebra. Coassociativity gives \[ \Phi_{U^J} \cdot (\delta^J \otimes \mathrm{Id}) \circ \delta(x) \cdot \Phi_C^{-1}=(\mathrm{Id} \otimes \delta^J) \circ \delta^J(x)\] Taking $x=1$ gives \[ \Phi_{U^J} \cdot (1 \otimes 1) \otimes 1 \cdot \Phi_C^{-1}=(1 \otimes 1) \otimes 1 \] so $\Phi_{U^J}=\Phi_C$ since we are in the regular representation. \end{proof} \begin{proposition}\label{oplax} Suppose we have oplax monoidal functors $\mathbb{V}_k:\mathrm{Rep}(C) \to \mathrm{Rep}(U)$, $k=1,...,n$ such that $\bigotimes\limits_{k=1}^n \mathbb{V}_k(C_{reg})$ is a projective generator of $\mathrm{Rep}(U)$, such that in addition the coproduct $\delta$ induced by the coproduct of the regular $C$-$C$-bimodule $C_{reg}$ sends the image of a generator $1$ to a $U$-$U$-bimodule generator. Then there exists a twist $J$ of $U$ such that $\Phi_{U^J}=\Phi_{C}$. \end{proposition} \begin{proof} We can view $C$ as a $C$-$C$-bimodule coalgebra with the regular representation. The category of $C$-bimodules is equivalent to the relative Deligne tensor category $\mathrm{Rep}(C) \boxtimes_{\mathrm{Rep}(C)} \mathrm{Rep}(C)$. Since the image of a coalgebra object under an oplax functor is a coalgebra object, the image of $C$ under $\mathbb{V}_K \boxtimes \mathrm{Id}$ is a $U$-$C$ bimodule coalgebra for each $k=1,...,n$ with $\delta(1)$ invertible. Therefore, it follows from Proposition \ref{bimod} that the gauge transformation $J:=\delta(1)$ yields a twist $U^J$ such that $\Phi_{U^J}$ and $\Phi_C$ agree as operators on $\mathbb{V}_k(C)$. However, \[ \bigotimes_{k=1}^n \mathbb{V}_k(C)\] is a projective generator for $\mathrm{Rep}(U)$, so $\Phi_{U^J}$ and $\Phi_C$ agree as operators on a projective generator and therefore agree. \end{proof} \begin{theorem}\label{Catthm} Let $\mathcal{C}, \mathcal{D}$ be a finite tensor categories such that there exists a quasi-fiber functor $F_{\mathcal{D}}:\mathcal{D} \to \mathrm{Vec}$ and an essentially surjective quasi-tensor functor $G:\mathcal{C} \to \mathcal{D}$ as in Proposition \ref{prop2}. Suppose we have oplax monoidal functors $\mathbb{V}_k:\mathrm{Rep}(C) \to \mathrm{Rep}(U)$, $k=1,...,n$ such that $\bigotimes_k \mathbb{V}_k(C_{reg})$ is a projective generator of $\mathrm{Rep}(U)$, such that in addition the coproduct $\delta$ induced by the coproduct of the regular $C$-$C$-bimodule $C_{reg}$ sends the image of a generator $1$ to a $U$-$U$-bimodule generator, as in Proposition \ref{oplax}. Then there exist a twist $U^J$ of $U$, such that $\Phi_{U^J}=\Phi_C$ and thus $C=\mathrm{End}(F_{\mathcal{D}})$ is a sub quasi bisubalgebra of $U^J$. Moreover there is a unique choice of such a twist such that $\delta$ \end{theorem} \begin{proof} We have seen in Proposition \ref{prop2} that there is an inclusion $C \subset U$ of the associated quasi bialgebra $U=\mathrm{End}(F_{\mathcal{D}} \circ G)$, $C=\mathrm{End}(F_{\mathcal{D}})$ preserving the coproduct, counit, and algebra structure. Since the conditions of Propositiion \ref{oplax} are satisfied, we also have $\Phi_{U^J}=\Phi_C$ for some twist $J$ of $U$. \end{proof} \section{Characterization} Our goal is to classify all quasitriangular quasi bialgebras whose representation categories coincide with that of the triplet vertex operator algebra $\mathcal{W}(2)$. We proceed in steps, first classifying the possible coproducts and antipodes. That is, we first classify those quasi bialgebras whose representation categories are tensor equivalent to that of the triplet. We then classify the quasitriangular quasi bialgebras whose representation categories are braided tensor equivalent to that of the triplet. \\ \begin{definition} As an algebra, the small quantum group $\tilde{u}_i(\mathfrak{sl}_2)=u_i(\mathfrak{sl}_2)$ at fourth root of unity $i=\sqrt{-1}$ is the $\mathbb{C}$-algebra with generators $E,F,K^{\pm 1}$, and relations \begin{align} \label{Uq1} KE&=-EK, & KF&=-FK, & [E,F]&=\frac{1}{2}(K-K^{-1})\\ \label{Uq2} \quad & & F^2&=E^2=0, & KK^{-1}&=K^{-1}K=1 \end{align} \end{definition} \noindent $\tilde{u}_i(\mathfrak{sl}_2)$ has four Verma modules $M_k$ with $k\in \{0,1,2,3\}$. Each $M_k$ is two dimensional spanned by $\{v_k,Fv_k\}$ where \[ Kv_k=i^kv_k, \quad KFv_k=-i^kFv_k, \quad Ev_k=0, \quad EFv_k=\delta_{2|k}i^{k-1} v_k \] where $\delta_{2|k}$ is $1$ if $2|k$ and $0$ otherwise. It is clear then that $M_k$ is irreducible if $k$ is odd and reducible if $k$ is even. Let $X_k$, $k\in \{0,...,3\}$ denote the irreducible modules of $\tilde{u}_i(\mathfrak{sl}_2)$ where $X_k$ is the irreducible quotient of $M_k$. We therefore have $X_k=M_k$ if $k$ is odd and \[ K \cdot X_k = i^k \cdot X_k\] \noindent The following semisimple quasi-triangular quasi-Hopf algebra $C$ appears below as the Cartan subalgebra of $u_i(\mathfrak{sl}_2)$: \begin{definition} Let $C=\mathbb{C}[\mathbb{Z}_4]=\mathrm{Span}_{\mathbb{C}}\{K\}$ with $K^4=1$ be a group algebra. This algebra is isomorphic to the dual group algebra $\mathbb{C}^{\mathbb{Z}_4}=\bigoplus\limits_{k=0}^3 \mathbb{C}_k $ via \begin{align*} e_0&:=\frac{1}{4}(1+K+K^2+K^3) & e_1&:=\frac{1}{4}(1-iK-K^2+iK^3)\\ e_2&:=\frac{1}{4}(1-K+K^2-K^3) & e_3&:=\frac{1}{4}(1+iK-K^2-iK^3) \end{align*} which are orthogonal idempotents $ \label{eij} e_i \cdot e_j= \delta_{i,j} e_i$.\\ \noindent The algebra $C$ has accordingly four simple representations $\mathbb{C}_i$ with the action \begin{equation}\label{Kaction} K \cdot \mathbb{C}_k = i^k \mathbb{C}_k. \end{equation} The algebra $C$ can be given the structure of a quasi-triangular quasi-Hopf algebra with coproduct, counit, antipode \begin{align} \label{Ccop}\Delta_C(e_k)&=\sum\limits_{i+j=k} e_i \otimes e_j\\ \label{Ccou}\epsilon_C(e_k)&=\delta_{k,0}\\ \label{Cant}S_C(e_k)&=e_{-k}\\ \intertext{and with coassociator and $R$-matrix given by a quadratic form $Q(k)=\beta^{(k^2)}$ on $\mathbb{Z}_4$ for any fixed $\beta$ with $\beta^4=-1$. We take the following representative associator and $R$-matrix in accordance with \cite{FGR1}}\\ \label{Ccoass}\Phi_C^{\pm1}&= \sum\limits_{a,b,c=1}^3 \phi_{abc}^\pm (e_a \otimes e_b \otimes e_c) \\ \phi_{abc}^\pm&:=\begin{cases} \mp \beta^2,\quad &2\nmid a,b\text{ and } c=1\\ -1,\quad &2\nmid a,b\text{ and } c=2\\ \pm \beta^2,\quad &2\nmid a,b\text{ and } c=3\\ 1, \quad &\mathrm{otherwise} \end{cases}\\ R&=\sum\limits_{a,b=1}^3 R^{a,b}(e_a\otimes e_b)\\ R^{a,b}&:=\begin{pmatrix} 1 & 1 & 1 & 1 \\ 1 & \beta & 1 & \beta \\ 1 & -1 & -1 & 1 \\ 1 & -\beta & -1 & \beta \end{pmatrix}, R^{a,b}R^{b,a}=\begin{pmatrix} 1 & 1 & 1 & 1 \\ 1 & \beta^2 & -1 & -\beta^2 \\ 1 & -1 & 1 & -1 \\ 1 & -\beta^2 & -1 & \beta^2 \end{pmatrix} \end{align} \end{definition} \begin{definition} The oplax tensor functors $\mathbb{V},\overline{\mathbb{V}}$ from ${\rm Rep}(C)$ to ${\rm Rep}(U)$ are defined by induction \begin{align*} \mathbb{V}(X)&:=X\otimes_{U^{\leq 0}} U \\ \overline{\mathbb{V}}(X)&:=X \otimes_{U^{\geq 0}} U\\ \end{align*} We have $\mathbb{V}(\mathbb{C}_k)=\mathrm{Span}\{e_k,Fe_k\}$ where $Ee_k=0$, and it follows from Equations \eqref{Kaction} and \eqref{Uq1} that \begin{align*} EFe_0&=0, & EFe_1&=e_1, & EFe_2&=0, & EFe_3&=-e_3. \end{align*} Similarly, $\overline{\mathbb{V}}(\mathbb{C}_k^{\pm})=\mathrm{Span}\{e_k^{\pm}, Ee_k^{\pm}\}$ where $Fe_k^{\pm}=0$ and \begin{align*} FEe_0^+&=0, & FEe_1&=-e_1, & FEe_2&=0, & FEe_3&=e_3. \end{align*} We therefore see that $\mathbb{V}(e_k) \cong M_k$, $\overline{\mathbb{V}}(e_k) \cong \overline{M}_k$ as modules over $u_i(\mathfrak{sl}_2)$ where $\overline{M}_k$ is the opposite Verma module. Then the images of the regular representation of $C$ under $\mathbb{V},\overline{\mathbb{V}}$ are as $u_i(\mathfrak{sl}_2)$-modules \[ \mathbb{V}(C_{reg}) \cong \bigoplus\limits_{k=0}^3 M_k, \qquad \overline{\mathbb{V}}(C_{reg}) \cong \bigoplus\limits_{k=0}^3 \overline{M}_k\] We now consider the regular $C$-$C$-bimodule $C_{\mathrm{reg}}=\bigoplus_{k=0}^3 \mathbb{C}_{k}$ and the induced module $\mathbb{V}(C_\mathrm{reg})=\bigoplus \mathbb{V}(\mathbb{C}_k)$ with basis $\{e_k,Fe_k\}$, $k\in \{0,1,2,3\}$. The left action of $C=\mathbb{C}^{\mathbb{Z}_4}$ is by $C\subset U$ and the right action of $C$ conveniently coincides in this notation with the right-multiplication by $C\subset U$. \end{definition} \subsection{Classification of braided tensor structure on ${\rm Rep}(U)$} The image of a coalgebra under an oplax tensor functor is a coalgebra. Similarly, the image of a $C$-$C$-bimodule coalgebra under $\mathbb{V},\overline{\mathbb{V}}$ is a $U$-$C$-bimodule coalgebra. Note that the regular representation $C_{reg}$ is a $C$-$C$-bimodule coalgebra but not a $C$-module coalgebra due to the nontrivial associator. We now classify all possible coalgebra structures on $\mathbb{V}(C_\mathrm{reg}),\overline{\mathbb{V}}(C_\mathrm{reg})$. In this explicit case it follows for $1=\sum_ae_a$ from the grading and the coassociativity that $\delta(1)$ is again a generator. Hence we may invoke Theorem \ref{Catthm} and thus assume that after going to a twist equivalent $U$ we have $\delta(1)=1\otimes 1$ and $\Phi_U=\Phi_C$. \begin{proposition}\label{prop_Vreg} The following are all $\mathbb{Z}_4$-$\mathbb{Z}_4$-graded coalgebra structures on $\mathbb{V}(C_\mathrm{reg})$ such that $\delta(1)=1\otimes 1$. \[ \delta(F \cdot1) := (F\otimes 1)\sum_{a,b} c_L^{ab}(e_a\otimes e_b) + (1\otimes F)\sum_{a,b} c_R^{ab}(e_a\otimes e_b)\] with $c_L^{ab},c_R^{a,b} \in \mathbb{C}$. If $c^{a,b}_L, c^{a,b}_R\neq 0$, then \begin{align} \label{cL}c_L^{ab}&=\frac{c_{a+b}}{c_a}\\ \label{cR}c_R^{ab}&=\epsilon^a \frac{c_{a+b}}{c_b}(-1)^{a(a-1)/2} \end{align} for any choice of scalars $c_0=1,c_1,c_2,c_3$ and any global choice $\epsilon^4=1$. By the same argument, $\delta(E \cdot1)$ is determined on $\overline{V}(C_\mathrm{reg})$ by the same equation, swapping $F$ and $E$ with coefficients $\bar{c}_0=1,\bar{c}_1,\bar{c}_2,\bar{c}_3$ and $\bar{\epsilon}^4=1$. \end{proposition} \begin{proof} We have assumed $\delta(1)=1\otimes 1$ where $1=\sum\limits_{k=0}^3 e_k$ which gives \[ \sum\limits_{k=0}^3 \delta(e_k)=\delta(1)=1 \otimes 1 =\sum\limits_{a,b=0}^3 e_a \otimes e_b. \] Since $\mathbb{V}(C_{\mathrm{reg}})$ is $\mathbb{Z}_4$-graded as a coalgebra with $\mathbb{V}(C_{\mathrm{reg}})_{\bar{k}}=\mathrm{Span}\{e_k,Fe_k\}$, we have $$ \delta(e_k)=\sum_{a+b=k} e_{a}\otimes e_{b}$$ Since the coproduct must respect the $\mathbb{Z}_4$-grading, an arbitrary expression for $\delta(F \cdot e_k)$ is \[ \delta(F \cdot e_k)=\sum\limits_{\substack{ u,v \in \{0,1\} \\ a+b=k}} F^ue_a \otimes F^v e_b, \] and $K \otimes K \delta(F \cdot e_k)= -i^k \delta( F \cdot e_k)$, so we have $(-1)^{u+v}i^k=-i^k$ and therefore $u+v$ is odd. Hence, an arbitrary expression for $\delta(F \cdot 1)=\sum\limits_{k=0}^3\delta(F \cdot e_k)$ is \begin{align*} \delta(F \cdot 1)&=: (F\otimes 1)\sum_{a,b=0}^3 c_L^{ab}(e_a\otimes e_b) + (1\otimes F)\sum_{a,b=0}^3 c_R^{ab}(e_a\otimes e_b)\\ \intertext{and thus by the right $C$-module structure} \delta(F \cdot e_k)&= (F \otimes 1) \sum\limits_{a+b=k} c_L^{ab} (e_a \otimes e_b)+(1 \otimes F) \sum\limits_{a+b=k}c_R^{ab}(e_a \otimes e_b) \end{align*} By applying counitality to each $\delta(F \cdot e_k)$ we have $c_L^{a,0}=c_R^{0,b}=1$. We now compute constraints on the coefficients using coassociativity \begin{align*} &(\delta \otimes \mathrm{Id}) \circ \delta(F \cdot 1)=\sum\limits_{a,b=0}^3 c_L^{ab}\delta(F \cdot e_a) \otimes e_b+\sum\limits_{a,b=0}^3 c_R^{ab} \delta(e_a) \otimes F \cdot e_b \\ &= \sum\limits_{a,b=0}^3 \sum\limits_{u+v=a} c_L^{ab} c_L^{uv} F e_u \otimes e_v \otimes e_b\\ &+\sum\limits_{a,b=0}^3 \sum\limits_{u+v=a} c_L^{ab}c_R^{uv}e_u \otimes Fe_v \otimes e_b+\sum\limits_{a,b=0}^3\sum\limits_{u+v=a}c_R^{ab} e_u \otimes e_v \otimes Fe_b\\ \qquad \\ &(\mathrm{Id} \otimes \delta) \circ \delta(F \cdot 1)=\sum\limits_{a,b=0}^3 c_L^{ab}Fe_a \otimes \delta(e_b)+ \sum\limits_{a,b=0}^3 c_R^{ab}e_a \otimes \delta(Fe_b)\\ &=\sum\limits_{a,b=0}^3 \sum\limits_{u+v=b} c_L^{ab}Fe_a \otimes e_u \otimes e_v \\ &+\sum\limits_{a,b=0}^3 \sum\limits_{u+v=b} c_R^{ab}c_L^{uv} e_a \otimes Fe_u \otimes e_v+\sum\limits_{a,b=0}^3 \sum\limits_{u+v=b} c_R^{ab}c_R^{uv}e_a \otimes e_u \otimes Fe_v\\ \end{align*} Here, we have to use the coassociator of $U$-$C$-bimodules, which we have proven in Theorem \ref{Catthm} coincides with the coassociator of $C$ from Equation \eqref{Ccoass}. By Equation \eqref{HGcoass}, the coassociator acts on $U$-$C$-bimodules by $X \mapsto \Phi \cdot X \cdot \Phi^{-1}$, which acts trivially on $(e_a\otimes e_b\otimes e_c)$, $(Fe_a\otimes e_b\otimes e_c)$, and $(e_a\otimes Fe_b \otimes e_c)$, while it acts as $(-1)^{ab}$ on $(e_a\otimes e_b\otimes Fe_c)$. Now coassociativity reads for all $a,b,c$: \begin{align*} c_L^{ab}c_L^{a+b,c} &= c_L^{a,b+c}\\ c_R^{ab}c_L^{a+b,c} &=c_L^{bc}c_R^{a,b+c}\\ (-1)^{ab}c_R^{a+b,c} &=c_R^{bc}c_R^{a,b+c} \end{align*} \noindent The first and third equations with $a=0$ and $c=0$ respectively give \begin{align*} c_L^{bc} &=\frac{c_L^{0,b+c}}{c_L^{0,b}}, \\ c_R^{ab} &=\frac{c_R^{a+b,0}}{c_R^{b,0}}(-1)^{ab}. \end{align*} This conversely solves the equations for any values $c_a:=c_L^{0,a},\;d_a:=c_R^{a,0}$, where $c_0=d_0=1$. If we plug this into the second equation we get $$(-1)^{ab}\frac{d_{a+b+c}d_b}{d_{a+b}d_{b+c}} =\frac{c_{a+b+c}c_b}{c_{a+b}c_{b+c}}.$$ In particular $b=0$ is a $2$-cocycle condition $$(-1)^{ac}\frac{d_{a+c}}{d_{a}d_{c}} =\frac{c_{a+c}}{c_{a}c_{c}}$$ which gives a necessary expression for $d_n$ in terms of $c_n$ and $\epsilon:=d_1/c_1$ \begin{equation} \label{bn} d_n=\epsilon^n c_n(-1)^{n(n-1)/2} \end{equation} and the necessary condition $\epsilon^4=1$. Rewriting the coefficients in terms of $c_n$ then gives $c_L^{ab}=\frac{c_{a+b}}{c_a}$ and \begin{align*} c_R^{ab}&=\frac{(-1)^{ab}d_{a+b}}{d_b}\\ &=(-1)^{ab}\frac{\epsilon^{a+b}c_{a+b}(-1)^{(a+b)(a+b-1)/2}}{\epsilon^b c_b (-1)^{b(b-1)/2}}\\ &=\epsilon^a \frac{c_{a+b}}{c_b}(-1)^{a(a-1)/2} \end{align*} \end{proof} We now use our knowledge of the coalgebras $\mathbb{V}(C_{reg}),\overline{\mathbb{V}}(C_{reg})$ to determine the coproduct on $U$. This proceeds in several steps. Some of the following arguments are more general, but for clarity we spell them out on the explicit algebra in question: \begin{enumerate}[a)] \item We write down an arbitrary expression for $\Delta(E)$ and $\Delta(F)$, keeping in mind that $K$ acts adjointly on $E,F$ by $-1$. Therefore, since $\Delta(KX)=-\Delta(XK)$ for $X\in \{E,F\}$ and $\Delta(K)=K \otimes K$, the summands in $\Delta(E)$ and $\Delta(F)$ have to contain $E,F$ an odd number of times \begin{align*} \Delta(F) &=\sum\limits_{a,b =0}^3\big( (F \otimes 1) c_{L}^{ ab}+(1 \otimes F) c_{R}^{ ab}\\ &+(E \otimes 1) r_{L}^{ab}+(1 \otimes E) r_{R}^{ ab}\\ &+(E \otimes EF) t_{L}^{ab}+(EF \otimes E) t_{R}^{ab}\\ &+(F \otimes EF) s_{L}^{ab}+(EF \otimes F) s_{R}^{ab}\big) (e_a \otimes e_b )\\ \quad\\ \Delta(E) &=\sum\limits_{a,b =0}^3\big( (E \otimes 1) \bar{c}_{L}^{ ab}+(1 \otimes E) \bar{c}_{R}^{ab}\\ &+(F \otimes 1) \bar{r}_{L}^{ab}+(1 \otimes F) \bar{r}_{R}^{ ab}\\ &+(F \otimes EF) \bar{t}_{L}^{ab}+(EF \otimes F) \bar{t}_{R}^{ab}\\ &+(E \otimes EF) \bar{s}_{L}^{ab}+(EF \otimes E) \bar{s}_{R}^{ab}\big) (e_a \otimes e_b ) \end{align*} \item It follows from $\Delta(X) \cdot (1 \otimes 1) = \Delta(X) \cdot \delta(1)=\delta(X \cdot 1)$ that the coefficients $c_L,c_R,\bar{c}_L,\bar{c}_R$ coincide with the corresponding coefficients of $\delta(F \cdot 1)$ appearing in Proposition \ref{prop_Vreg}. The existence of a right quasi-antipode requires that $c_L^{a,-a}$ is invertible, and it follows from the coassociativity relations in the proof of Proposition \ref{prop_Vreg} that $c^{a,-a}c^{0,a}=c^{a,0}=1$, so the additional nonzero condition in Proposition \ref{prop_Vreg} is fulfilled, that is, $c_L^{ab}$ and $c_R^{ab}$ are non-zero. On the other hand, since we have a zero action $E \cdot 1= F \cdot \bar{1} =0$, we have $r^{ab}_L=r^{ab}_R=\bar{r}^{ab}_L=\bar{r}^{ab}_R=0$.\\ \item The coefficient of $E e_a \otimes E e_0 \otimes F e_b$ in $(\Delta \otimes \mathrm{Id}) \circ \Delta(F)$ is zero, but the coefficient of the same term in $(\mathrm{Id} \otimes \Delta) \circ \Delta(F)$ is $t_L^{ab}\bar{c}_L^{0,b}c_R^{0,b+2}$. We therefore see that $t_L^{a,b}=0$ since $c_L^{a,b}$ and $c_R^{a,b}$ are non-zero. Similarly, inspecting the coefficients of $E e_a \otimes Fe_0 \otimes Ee_b$, $Ee_a \otimes Fe_0 \otimes Fe_b$, and $Fe_a \otimes Ee_0 \otimes Fe_b$ respectively show that $t_R^{a,b}$, $s_R^{a,b}$, and $s_L^{a,b}$ all vanish. Applying the same argument to $\Delta(E)$ shows that the corresponding coefficients for $\Delta(E)$ also vanish. Therefore, we have \[ t_L^{a,b}=t_R^{a,b}=s_L^{a,b}=s_R^{a,b}=\bar{t}_L^{a,b}=\bar{t}_R^{a,b}=\bar{s}_L^{a,b}=\bar{s}_R^{a,b}=0 \\ \] \item We now use the nilpotency of $F$ to further restrict the coefficients: \begin{align*} 0=\Delta(F)^2 &= \big((F\otimes 1)\sum_{a,b=0}^3 c_L^{ab}(e_a\otimes e_b) + (1\otimes F)\sum_{a,b} c_R^{ab}(e_a\otimes e_b)\big)^2\\ &=\sum_{a,b=0}^3 (c_L^{a,b+2}c_R^{a,b}+c_R^{a+2,b}c_L^{a,b})(F\otimes F)(e_a\otimes e_b)\\ \end{align*} Now, since the coefficients are non-zero and have the form described in Proposition \ref{prop_Vreg}, we have: \begin{align*} \quad &=\frac{c_{a+b+2}}{c_{a}}\cdot \epsilon^a (-1)^{a(a-1)/2}\frac{c_{a+b}}{c_{b}} +\epsilon^{a+2} (-1)^{(a+2)(a+1)/2}\frac{c_{a+b+2}}{c_{b}}\cdot \frac{c_{a+b}}{c_a} \end{align*} so $\epsilon^2=1$ in Proposition \ref{prop_Vreg}. Similarly $\Delta(E)^2=0$ implies $\bar{\epsilon}^2=1$. \item We now check the commutation relations for $E$ and $F$, $[\Delta(E),\Delta(F)]=K(e_1+e_3)$. \begin{align*} &\Delta(E)\Delta(F)-\Delta(F)\Delta(E)\\ &=\left(\sum_{a',b'}\big((E \otimes 1) \bar{c}_{L}^{ a'b'}+(1 \otimes E) \bar{c}_{R}^{a'b'}\big)(e^{a'} \otimes e^{b'} )\right) \cdot\left(\sum_{a,b}\big((F \otimes 1) c_{L}^{ ab}+(1 \otimes F) c_{R}^{ ab}\big)(e^{a} \otimes e^{b} )\right)\\ &-\left(\sum_{a',b'}\big((F \otimes 1) c_{L}^{ a'b'}+(1 \otimes F) c_{R}^{ a'b'}\big)(e^{a'} \otimes e^{b'} )\right) \cdot \left(\sum_{a,b}\big((E \otimes 1) \bar{c}_{L}^{ ab}+(1 \otimes E) \bar{c}_{R}^{ab}\big)(e^{a} \otimes e^{b} )\right)\\ &=\sum_{a,b}\big((EF\otimes 1)\bar{c}_L^{a+2,b}c_L^{a,b} +(F\otimes E)\bar{c}_R^{a+2,b}c_L^{a,b} +(E\otimes F)\bar{c}_L^{a,b+2}c_R^{a,b} +(1\otimes EF)\bar{c}_R^{a,b+2}c_R^{a,b}\\ &-(FE\otimes 1)c_L^{a+2,b}\bar{c}_L^{a,b} -(E\otimes F)c_R^{a+2,b}\bar{c}_L^{a,b} -(F\otimes E)c_L^{a,b+2}\bar{c}_R^{a,b} -(1\otimes FE)c_R^{a,b+2}\bar{c}_R^{a,b}\big)(e^{a} \otimes e^{b} )\\ \intertext{and on the other hand} &\Delta \left(K(e_1+e_3)\right) =\sum\limits_{a+b=1} Ke_a \otimes Ke_b + \sum\limits_{a+b=3} Ke_a \otimes Ke_b\\ &= \sum\limits_{a,b=0}^3 i^{a+b}(\delta_{2\mid a} \delta_{2 \nmid b}+\delta_{2 \nmid a} \delta_{2 \mid b})e_a \otimes e_b \end{align*} The vanishing of the mixed term $F\otimes E$ reads \begin{align*} \bar{c}_R^{a+2,b}c_L^{a,b}&=c_L^{a,b+2}\bar{c}_R^{a,b}\\ \bar{\epsilon}^{a+2} (-1)^{(a+2)(a+1)/2}\frac{\bar{c}_{a+b+2}}{\bar{c}_{b}} \cdot \frac{c_{a+b}}{c_{a}} &=\frac{c_{a+b+2}}{c_{a}} \cdot \bar{\epsilon}^a (-1)^{a(a-1)/2}\frac{\bar{c}_{a+b}}{\bar{c}_{b}}\\ \Rightarrow\quad -\bar{c}_{t+2}/\bar{c}_{t}&=c_{t+2}/c_t \end{align*} so in particular $-\bar{c}_2=c_2$). The vanishing of the mixed term $E\otimes F$ gives the same quasiperiodicity relation. \\ \begin{align*} \bar{c}_L^{a,b+2}c_R^{a,b}&=c_R^{a+2,b}\bar{c}_L^{a,b} \\ \frac{\bar{c}_{a+b+2}}{\bar{c}_{a}} \cdot \epsilon^a (-1)^{a(a-1)/2}\frac{c_{a+b}}{c_{b}} &=\epsilon^{a+2} (-1)^{(a+2)(a+1)/2}\frac{c_{a+b+2}}{c_{b}} \cdot \frac{\bar{c}_{a+b}}{\bar{c}_{a}} \\ \end{align*} \noindent The term $(EF-FE)\otimes 1 $ requires first that the two coefficients are equal \begin{align*} \bar{c}_L^{a+2,b}c_L^{a,b} &=c_L^{a+2,b}\bar{c}_L^{a,b}\\ \frac{\bar{c}_{a+b+2}}{\bar{c}_{a+2}}\frac{c_{a+b}}{c_{a}} &=\frac{c_{a+b+2}}{c_{a+2}}\frac{\bar{c}_{a+b}}{\bar{c}_{a}} \end{align*} which is always true by the quasiperiodicity relation. \\ \noindent The term $1\otimes (EF-FE)$ requires first that the two coefficients are equal \begin{align*} \bar{c}_R^{a,b+2}c_R^{a,b} &=c_R^{a,b+2}\bar{c}_R^{a,b}\\ \bar{\epsilon}^a (-1)^{a(a-1)/2}\frac{\bar{c}_{a+b+2}}{\bar{c}_{b+2}} \cdot \epsilon^a (-1)^{a(a-1)/2}\frac{c_{a+b}}{c_{b}} &=\epsilon^a (-1)^{a(a-1)/2}\frac{c_{a+b+2}}{c_{b+2}} \cdot \bar{\epsilon}^a (-1)^{a(a-1)/2}\frac{\bar{c}_{a+b}}{\bar{c}_{b}} \end{align*} which is always true by the quasiperiodicity relation. \\ Now we apply $[E,F]=K(e_1+e_3)$ which acts on $e_a$ by $i^a\delta_{2\nmid a}$ and compute the joint result of $(EF-FE)\otimes 1$ and $1\otimes (EF-FE)$, then the final equation to be fulfilled is: \begin{align*} i^a\delta_{2\nmid a} \cdot \frac{\bar{c}_{a+b+2}}{\bar{c}_{a+2}}\frac{c_{a+b}}{c_{a}} +i^b\delta_{2\nmid b} \cdot \bar{\epsilon}^a\frac{\bar{c}_{a+b+2}}{\bar{c}_{b+2}} \cdot \epsilon^a\frac{c_{a+b}}{c_{b}} &=i^a\delta_{2\mid a}i^b\delta_{2\nmid b}+i^a\delta_{2\nmid a}i^b\delta_{2\mid b} \end{align*} For $2\mid a,b$ both sides are zero. For $2\nmid a,b$ \begin{align*} i^a \cdot \frac{\bar{c}_{a+b+2}}{\bar{c}_{a+2}}\frac{c_{a+b}}{c_{a}} +i^b (\epsilon\bar{\epsilon})^a \cdot\frac{\bar{c}_{a+b+2}}{\bar{c}_{b+2}} \frac{c_{a+b}}{c_{b}} &=0 \\ i^a \cdot \frac{1}{\bar{c}_{a+2}c_{a}} +i^b(\epsilon\bar{\epsilon})^a \cdot\frac{1}{\bar{c}_{b+2}c_{b}} &=0 \end{align*} which for $a=b$ (odd) means $\epsilon\bar{\epsilon}$ and which for $b=a+2$ means $\bar{c}_{1}c_3=-\bar{c}_{3}c_1$ so it holds again by the quasiperiodicity relation.\\ \noindent For $2\nmid a,2\mid b$ resp. $2\mid a,2\nmid b$ \begin{align*} \frac{\bar{c}_{a+b+2}}{\bar{c}_{a+2}}\frac{c_{a+b}}{c_{a}} &=i^b\\ \frac{\bar{c}_{a+b+2}}{\bar{c}_{b+2}} \frac{c_{a+b}}{c_{b}} &=i^a \end{align*} which are equivalent after switching $a,b$, so we continue with the first. The first relation always holds for $b=0$, and the case $b=2$ always holds by the quasiperiodicity relation. \end{enumerate} \noindent The previous steps thus yield: \begin{proposition}\label{prop_EF} The coproduct is necessarily \begin{align*} \Delta(F) &=\sum\limits_{a,b \in \{\pm \}}\big( (F \otimes 1) c_{L}^{ ab}+(1 \otimes F) c_{R}^{ ab}\big)(e_a\otimes e_b)\\ \Delta(E) &=\sum\limits_{a,b \in \{\pm \}}\big( (E \otimes 1) \bar{c}_{L}^{ ab}+(1 \otimes E) \bar{c}_{R}^{ab}\big)(e_a\otimes e_b) \end{align*} where $c_{L}^{ ab}, c_{R}^{ ab}$ and $\bar{c}_{L}^{ ab},\bar{c}_{R}^{ ab}$ are defined in terms of $c_a,\epsilon,\bar{c}_a,\bar{\epsilon}$ as in Equations \eqref{cL} and \eqref{cR} under the following additional assumptions: $$\epsilon^2=\bar{\epsilon}^2=1,\quad \epsilon\bar{\epsilon}=-1,\quad c_{t+2}/c_t=-\bar{c}_{t+2}/\bar{c}_{t}$$ Conversely, every such coproduct defines a quasi-bialgebra structure on $U$ with respect to the associator $\Phi=\Phi_C$. \end{proposition} \noindent We now apply an algebra automorphism to remove the parameters $c_a,\bar{c}_a$:\\ \begin{proposition}~\label{prop_automorphism} \begin{enumerate}[a)] \item For any $x_0=1,x_1,x_2,x_3\in\mathbb{C}^\times$ the element $F':=F(\sum_a x_ae_a)$ also fulfills the coproduct formula in Proposition \ref{prop_Vreg}, with modified parameters $c'_a=c_ax_a$. \item The assignment \begin{align*} K&\mapsto K\\ E&\mapsto E':=E(\sum_a \bar{x}_a e_a)\\ F&\mapsto F':=F(\sum_a x_a e_a) \end{align*} extends to an algebra automorphism of $U$ iff $$\bar{x}_1=x_3^{-1},\quad \bar{x}_2=x_2,\quad \bar{x}_3=x_1^{-1}$$ \item Applying the algebra automorphism for $x_i=c_i$ yields the coproducts \begin{align*} \Delta(F)&=F\otimes 1+\omega_{-\epsilon}\otimes F\\ \Delta(E)&=(E\otimes 1)(\sum_{a,b}(\bar{c}')_L^{ab}e_a\otimes e_b)+(\omega_{\epsilon}\otimes E) (\sum_{a,b}(\bar{c}')_L^{ba}e_a\otimes e_b)\\ \omega_{\epsilon}&:=\sum_a(-\epsilon)^a(-1)^{a(a-1)/2}e_a=((e_0+e_2)+i\epsilon (e_1+e_3))K \\ (\bar{c}')_L^{ab}&= \begin{pmatrix} 1 & d & -1 & -d \\ 1 & -d^{-1} & -1 & d^{-1} \\ 1 & d & -1 & -d \\ 1 & -d^{-1} & -1 & d^{-1} \\ \end{pmatrix} \end{align*} \end{enumerate} \end{proposition} \begin{example} In particular for $d=\pm i, \epsilon=1$ the quasi-bialgebra above is the quasi-Hopf algebra in \cite{FGR2} Section 3.1 \begin{align*} \Delta(F)&=F\otimes 1+\omega_{-}\otimes F\\ \Delta(E)&=E\otimes K^{\pm 1}+\omega_{+}K^{\pm 1}\otimes E\\ \omega_{\pm}&:=\sum_a(-\epsilon)^a(-1)^{a(a-1)/2}=((e_0+e_2)\pm i\epsilon (e_1+e_3))K \end{align*} with the conventions $F=\mathsf{f^-},E=K\mathsf{f^+}$ and central idempotents $e_0+e_2=\mathsf{e_0},e_1+e_3=\mathsf{e_2}$. For $\epsilon=-1$ we get an isomorphic quasi-Hopf algebra with $\mathsf{f^-},\mathsf{f^+}$ switched, and for $d=-i$ we get an isomorphic quasi-Hopf algebra with switched $K,K^{-1}$ and a different initial coassociator $\Phi$ for $\beta'=i\beta$. \end{example} \begin{proof}[Proof of Proposition \ref{prop_automorphism}] \begin{enumerate}[a)] \item By multiplicativity of the coproduct \begin{align*} \Delta(F(\sum_a x_ae_a)) &=(F\otimes 1)\sum_{a,b} c_L^{ab}(e_a\otimes e_b)(\sum_{a',a''} x_{a'+a''} e_{a'}\otimes e_{a''}) \\ &+ (1\otimes F)\sum_{a,b} c_R^{ab}(e_a\otimes e_b)(\sum_{a',a''} x_{a'+a''} e_{a'}\otimes e_{a''})\\ &=(F(\sum_a x_ae_a)\otimes 1)\sum_{a,b} c_L^{ab}\frac{x_{a+b}}{x_a}(e_a\otimes e_b)\\ &+ (1\otimes F(\sum_b x_ae_b))\sum_{a,b} c_R^{ab}\frac{x_{a+b}}{x_b}(e_a\otimes e_b) \end{align*} and $c_L^{ab},c_R^{ab}$ are proportional to $\frac{c_{a+b}}{c_a},\frac{c_{a+b}}{c_b}$ respectively.\\ \item Nilpotency and commutation relations with $K$ surely hold for the new $E',F'$, it remains to check $[E',F']=Ke_1$: \begin{align*} [E',F'] &=E(\sum_a \bar{x}_a e_a)F(\sum_a {x}_a e_a) -F(\sum_a {x}_a e_a)E(\sum_a \bar{x}_a e_a)\\ &=EF(\sum_a \bar{x}_{a+2}x_a e_a) -FE(\sum_a {x}_{a+2}\bar{x}_a e_a) \end{align*} The two factors are equal iff $\bar{x}_{a+2}/\bar{x}_{a}={x}_{a+2}/{x}_{a}$, so iff $\bar{x}_2=x_2$ and $\bar{x}_3=\bar{x}_1(x_3/x_1)$, and they are equal to $1$ for $2\nmid a$ (where $[E,F]\neq0$) iff $\bar{x}_{a+2}x_a=1$, so iff $\bar{x}_{3}x_1=1$ and $\bar{x}_{1}x_3=1$, which then holds automatically.\\ \item If we set $x_i=c_i^{-1}$ the coproduct of $F'$ is as asserted $$\Delta(F')=\sum_{a,b}\big((F\otimes 1)+(1\otimes F)\epsilon^a(-1)^{a(a-1)/2}\big)(e_a\otimes e_b)$$ On the other hand by the choice of $\bar{x}_i$ in b): $$\bar{c}_1'=\bar{c}_1c_3,\; \bar{c}_2'=\bar{c}_2/c_2,\; \bar{c}_3'=\bar{c}_3c_1$$ If we introduce $d=\bar{c}_1c_3$ then by the quasiperiodicity relation $-d=\bar{c}_3c_1$, on the other hand $\bar{c}_2/c_2=-1$. Going through all $a,b$ gives the matrix $$(\bar{c}')_L^{ab}=\frac{\bar{c}'_{a+b}}{\bar{c}'_a} = \begin{pmatrix} 1 & d & -1 & -d \\ 1 & -d^{-1} & -1 & d^{-1} \\ 1 & d & -1 & -d \\ 1 & -d^{-1} & -1 & d^{-1} \\ \end{pmatrix}$$ \end{enumerate} \end{proof} \begin{proposition}\label{braiding} The previous bialgebra for $\epsilon=1$ has an $R$-matrix iff $d=\pm i$, and it is of the form \begin{align*} R &=\sum\limits_{a,b =0}^3\bigg( R^{a,b}(1 \otimes 1)+R^{a,b}_{F,E}(F \otimes E)\bigg)(e^a \otimes e^b) R^{a,b} &=\begin{pmatrix} 1 & 1 & 1 & 1 \\ 1 & \beta & 1 & \beta \\ 1 & -1 & -1 & 1 \\ 1 & -\beta & -1 & \beta \end{pmatrix},\\ R_{F,E}^{a,b} &=2\begin{pmatrix} \mp 1 & -i & \mp 1 & -i \\ \mp 1 & -i\beta & \mp 1 & -i\beta \\ \mp 1 & i & \pm & -i \\ \mp 1 & i\beta & \pm 1 & -i\beta \\ \end{pmatrix} \end{align*} \end{proposition} \begin{proof} An arbitrary element in $u_q \otimes u_q$ is a $u^0_q$-combination of $E^iF^j \otimes E^kF^l$ with $i,j,k,l \in \{0,1\}$. We restrict the form in the following steps (the order of which chosen to minimize terms in computations) \begin{enumerate}[a)] \item Compatibility of the $R$-matrix with the coproduct ($R\Delta_{U(e_0)}(K)=\Delta^{op}(K)R$) means, since $XK=-KX$ for $X \in \{E,F\}$, that for each term in $R$, we must have $2|(i+j+k+l)$. Therefore, \begin{align*} R&=\sum\limits_{a,b =0}^3\bigg( R^{a,b}(1 \otimes 1)+R^{a,b}_{E,F}(E \otimes F)+R_{F,E}^{a,b}(F \otimes E)+R^{a,b}_{E,E}(E \otimes E)+R^{a,b}_{F,F}(F \otimes F)\\ &+R_{EF,1}^{a,b}(EF \otimes 1)+R_{1,EF}^{a,b}(1 \otimes EF)+R_{EF,EF}^{a,b}(EF \otimes EF) \bigg)(e^a \otimes e^b) \end{align*} \item Counitality $(\epsilon\otimes 1)(R)=(1\otimes\epsilon)(R)=1$ implies $$R^{0b}=R^{a0}=1,\quad R_{EF,1}^{a0}=R_{1,EF}^{0b}=0$$ \item The functors $\mathbb{V},\overline{\mathbb{V}}$ are braided, and the tensor product $\mathbb{V}(\mathbb{C}_a)\otimes \mathbb{V}(\mathbb{C}_b)$ decomposes into the images, so $R^{a,b}=\sigma(a,b)$ is the braiding of $C$: \item Writing out the left hexagon identity $(\Delta\otimes 1)(R)=R_{13}R_{23}$ implies $$R^{a,b}_{EF,1}=R^{a,b}_{1,EF}=0$$ because the terms $E\otimes F\otimes 1$ on the left hand side cannot appear on the right hand side, and similarly for the right hexagon. We later write out the full hexagon identity. \item Compatibility of the $R$ matrix with the coproduct of $F$ reads, using also $FE=EF-e_1K$: \begin{align*} R\Delta(F)&=\sum\limits_{a,b}\bigg( R^{a+2,b}c_L^{ab}(F \otimes 1)+R_{E,F}^{a+2,b}c_L^{ab}(EF \otimes F)\\ &+R^{a,b+2}c_R^{ab}(1 \otimes F)+R_{F,E}^{a,b+2}c_R^{ab}(F \otimes EF) \bigg) (e^a \otimes e^b)\\ \Delta(F)^{\tau}R &=\sum\limits_{a,b} \bigg( c_L^{ba}R^{a,b}(1 \otimes F)+c_L^{b+2,a+2}R_{F,E}^{a,b}(F \otimes (EF-e_1K))\\ &+c_R^{ba}R^{a,b}(F \otimes 1)+c_R^{b+2,a+2}R_{E,F}^{a,b}((EF-e_1K) \otimes F) \bigg) (e^a \otimes e^b) \end{align*} Comparing coefficients then gives the following relations \begin{align*} R^{a+2,b}c_L^{ab}&=c_R^{ba}R^{a,b}+c_L^{b+2,a+2}R_{F,E}^{a,b}(-\delta_{2\nmid b}i^b)\\ R^{a,b+2}c_R^{ab} &=c_L^{ba}R^{a,b}+c_R^{b+2,a+2}R_{E,F}^{a,b}(-\delta_{2\nmid a}i^a)\\ R_{E,F}^{a+2,b}c_L^{ab} &= c_R^{b+2,a+2}R_{E,F}^{a,b} \\ R_{F,E}^{a,b+2}c_R^{ab} &=c_L^{b+2,a+2}R_{F,E}^{a,b} \end{align*} We now assume the standard form of the coproduct achieved in Proposition \ref{prop_automorphism} $$c_L^{ab}=1, c_R^{ab}=\epsilon^a(-1)^{a(a-1)/2}, \epsilon=1$$ then the previous relations read explicitly \begin{align*} R_{F,E}^{a,b}(-\delta_{2\nmid b}i^b) &=R^{a+2,b}-R^{a,b}(-1)^{b(b-1)/2}=2\begin{pmatrix} 0 & -1 & 0 & 1 \\ 0 & -\beta & 0 & \beta \\ 0 & 1 & 0 & 1 \\ 0 & \beta & 0 & \beta \\ \end{pmatrix}\\ R_{F,E}^{a,b} &=2\begin{pmatrix} * & -i & * & -i \\ * & -i\beta & * & -i\beta \\ * & i& * & -i \\ * & i\beta & * & -i\beta \\ \end{pmatrix}\\ R_{E,F}^{a,b}(\delta_{2\nmid a}i^a) &=R^{a,b}-R^{a,b+2}(-1)^{a(a-1)/2} =\begin{pmatrix} 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ \end{pmatrix}\\ R_{E,F}^{a,b} &=\begin{pmatrix} * & * & * & * \\ 0 & 0 & 0 & 0 \\ * & * & * & * \\ 0 & 0 & 0 & 0 \\ \end{pmatrix}\\ \intertext{The other two equations are fulfilled on the visible part of $R_{E,F}^{a,b},R_{F,E}^{a,b}$ and fix the unknown third row/column relative to the unknown first row/column} R_{E,F}^{a+2,b} &= (-1)^{b(b-1)/2}R_{E,F}^{a,b} \\ R_{F,E}^{a,b+2} &=(-1)^{a(a-1)/2}R_{F,E}^{a,b} \end{align*} \item Compatibility of the $R$ matrix with the coproduct of $E$ reads \begin{align*} R\Delta(E)&=\sum\limits_{a,b}\bigg( R^{a+2,b}\bar{c}_L^{ab}(E \otimes 1)+R_{F,E}^{a+2,b}\bar{c}_L^{ab}((EF-e_1K) \otimes E)\\ &+R^{a,b+2}\bar{c}_R^{ab}(1 \otimes E)+R_{E,F}^{a,b+2}\bar{c}_R^{ab}(E \otimes (EF-e_1K)) \bigg) (e^a \otimes e^b)\\ \Delta(E)^{\tau}R &=\sum\limits_{a,b} \bigg( \bar{c}_L^{ba}R^{a,b}(1 \otimes E)+\bar{c}_L^{b+2,a+2}R_{E,F}^{a,b}(E \otimes EF)\\ &+\bar{c}_R^{ba}R^{a,b}(E \otimes 1)+\bar{c}_R^{b+2,a+2}R_{F,E}^{a,b}(EF \otimes E) \bigg) (e^a \otimes e^b) \end{align*} Comparing coefficients then gives the following relations \begin{align*} R^{a+2,b}\bar{c}_L^{ab}+R_{E,F}^{a,b+2}\bar{c}_R^{ab}(-\delta_{2\nmid b}i^b)&=\bar{c}_R^{ba}R^{a,b}\\ R^{a,b+2}\bar{c}_R^{ab}+R_{F,E}^{a+2,b}\bar{c}_L^{ab}(-\delta_{2\nmid a}i^a) &=\bar{c}_L^{ba}R^{a,b}\\ R_{F,E}^{a+2,b}\bar{c}_L^{ab} &= \bar{c}_R^{b+2,a+2}R_{F,E}^{a,b} \\ R_{E,F}^{a,b+2}\bar{c}_R^{ab} &=\bar{c}_L^{b+2,a+2}R_{E,F}^{a,b} \end{align*} We now assume the standard form of the coproduct achieved in Proposition \ref{prop_automorphism} $$\bar{c}_L^{ab}= \begin{pmatrix} 1 & d & -1 & -d \\ 1 & -d^{-1} & -1 & d^{-1} \\ 1 & d & -1 & -d \\ 1 & -d^{-1} & -1 & d^{-1} \\ \end{pmatrix},\quad \bar{c}_R^{ab}=\epsilon^a(-1)^{a(a-1)/2}\bar{c}_L^{ba}, \quad \epsilon=-1$$ then the previous relations read explicitly \begin{align*} R_{E,F}^{a,b+2}\epsilon^a(-1)^{a(a-1)/2}\bar{c}_L^{ab}(\delta_{2\nmid b}i^b) &=(R^{a+2,b}-R^{a,b}(-1)^b(-1)^{b(b-1)/2})\bar{c}_L^{ab} =\begin{pmatrix} 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ \end{pmatrix}\\ % R_{E,F}^{a,b} &=\begin{pmatrix} * & 0 & * & 0 \\ * & 0 & * & 0 \\ * & 0 & * & 0 \\ * & 0 & * & 0 \\ \end{pmatrix}\\ R_{F,E}^{a+2,b}\bar{c}_L^{ab}(-\delta_{2\nmid a}i^a) =(R^{a,b}-&R^{a,b+2}(-1)^a(-1)^{a(a-1)/2})\bar{c}_L^{ba} =2\begin{pmatrix} 0 & 0 & 0 & 0 \\ d & -d^{-1}\beta & d & -d^{-1}\beta \\ 0 & 0 & 0 & 0 \\ -d & -d^{-1}\beta & d & d^{-1}\beta \\ \end{pmatrix}\\ R_{F,E}^{a,b} &= 2\begin{pmatrix} * & * & * & * \\ di & -i\beta & di & -i\beta \\ * & * & * & * \\ di & i\beta & -di & -i\beta \\ \end{pmatrix}\\ \intertext{The other two equations are fulfilled on the visible part of $R_{E,F}^{a,b},R_{F,E}^{a,b}$ and fix the unknown third row/column relative to the unknown first row/column (we use the 2-periodicity behaviour $\pm1$ of $\bar{c}_L^{ab}$)} R_{F,E}^{a+2,b} &= -(-1)^{b+2}(-1)^{(b+2)(b+1)/2}R_{F,E}^{a,b} \\ R_{E,F}^{a,b+2}&=-(-1)^a(-1)^{a(a-1)/2}R_{E,F}^{a,b} \end{align*} \item Combining the partial matrices and the periodicity relations in the two previous bullets we get \begin{align*} R_{E,F}^{a,b} &=2\begin{pmatrix} X & 0 & -X & 0 \\ 0 & 0 & 0 & 0 \\ X & 0 & X & 0 \\ 0 & 0 & 0 & 0 \\ \end{pmatrix}\\ R_{F,E}^{a,b} &=2\begin{pmatrix} Y & -i & Y & -i \\ di & -i\beta & di & -i\beta \\ Y & i & -Y & -i \\ di & i\beta & -di & -i\beta \\ \end{pmatrix} \end{align*} \item We now turn to the full hexagon identity $$R_1^{(1)}\otimes R_1^{(2)} \otimes R_2 =(\Phi_{2}\otimes\Phi_{3}\otimes \Phi_{1})^{-1} (R_1\otimes 1\otimes R_2) (\Phi_{1}\otimes\Phi_{3}\otimes \Phi_{2}) (1\otimes R_1\otimes R_2) (\Phi_{1}\otimes\Phi_{2}\otimes \Phi_{3})^{-1} $$ We denote $\phi_{abc}^\pm=1,\mp \beta^2,-1,\pm \beta^2$ for $a,b$ odd and $c=0,1,2,3$ and else $\phi_{abc}=1$, and we denote by $\odot$ the component-wise multiplication of matrices.\\ For the term $(1\otimes 1 \otimes 1)$ the hexagon equation holds, because $R^{ab}$ is the R-matrix of $c$. We now consider the term $(F\otimes 1\otimes E)(e_a\otimes e_b\otimes e_c)$, and \begin{align*} c_L^{ab}R_{F,E}^{a+b,c} &=\phi_{c+2,a+2,b}^- R_{F,E}^{a,c} \phi^+_{acb} R^{b,c} \phi^-_{abc} \\ \intertext{$b=0$ holds trivially. We spell this equation out for $b=1$ and use $\phi_{c+2,a+2,b}^-\phi_{acb}^+=1$} \begin{pmatrix} di & -i\beta & di & -i\beta \\ Y & i & -Y & -i \\ di & i\beta & -di & -i\beta \\ Y & -i & Y & -i \\ \end{pmatrix} &= \begin{pmatrix} Y & -i & Y & -i \\ di & -i\beta & di & -i\beta \\ Y & i & -Y & -i \\ di & i\beta & -di & -i\beta \\ \end{pmatrix} \odot \begin{pmatrix} 1 & \beta & 1 & \beta \\ 1 & \beta & 1 & \beta \\ 1 & \beta & 1 & \beta \\ 1 & \beta & 1 & \beta \ \end{pmatrix} \odot \begin{pmatrix} 1 & 1 & 1 & 1 \\ 1 & \beta^2 & -1 & -\beta^2 \\ 1 & 1 & 1 & 1 \\ 1 & \beta^2 & -1 & -\beta^2 \\ \end{pmatrix}\\ \intertext{This equation is fulfilled for $di=Y$.} \intertext{The same calcuation for $E\otimes 1\otimes F$ shows clearly $X=0$ and thus $R_{E,F}^{ab}=0$.\newline The calculation for $E\otimes F\otimes EF$ relates $R_{EF,EF}^{a+b,c}$ to the product of $R_{E,F}$ and $R_{F,E}$, so we have $R_{EF,EF}^{a,b}=0$. Hence $(F\otimes 1\otimes E)$ and $(1\otimes F\otimes E)$ are the only terms appearing. } \end{align*} \item We now turn to the other hexagon identity $$R_1\otimes R_2^{(1)}\otimes R_2^{(2)} =(\Phi_{3}\otimes\Phi_{1}\otimes \Phi_{2}) (R_1\otimes 1\otimes R_2) (\Phi_{2}\otimes\Phi_{1}\otimes \Phi_{3})^{-1} (R_1\otimes R_2\otimes 1) (\Phi_{1}\otimes\Phi_{2}\otimes \Phi_{3}) $$ We consider the term $(F\otimes E\otimes 1)$ for $c=1$ using $\phi^-_{b+2,a+2,c}\phi^+_{abc}=1$ $$ \bar{c}_L^{bc}R_{F,E}^{a,b+c} =\phi_{b+2,c,a+2}^+ R^{a+2,c} \phi^-_{b+2,a+2,c} R_{F,E}^{a,b} \phi^+_{abc}$$ \begin{align*} &\begin{pmatrix} d & d & d & d \\ -d^{-1} & -d^{-1} & -d^{-1} & -d^{-1} \\ d & d & d & d \\ -d^{-1} & -d^{-1} & -d^{-1} & -d^{-1} \\ \end{pmatrix} \odot \begin{pmatrix} -i & Y & -i & Y \\ -i\beta & di & -i\beta & di\\ i & -Y & -i & Y \\ i\beta & -di & -i\beta & di\\ \end{pmatrix}\\ &= \begin{pmatrix} 1 & -1 & 1 & -1 \\ 1 & \beta^2 & 1 & \beta^2 \\ 1 & 1 & 1 & 1 \\ 1 & -\beta^2 & 1 & -\beta^2 \\ \end{pmatrix} \odot \begin{pmatrix} -1 & -1 & -1 & -1 \\ -\beta & -\beta & -\beta & -\beta \\ 1 & 1 & 1 & 1 \\ \beta & \beta & \beta & \beta \\ \end{pmatrix} \odot \begin{pmatrix} Y & -i & Y & -i \\ di & -i\beta & di & -i\beta \\ Y & i & -Y & -i \\ di & i\beta & -di & -i\beta \\ \end{pmatrix} \intertext{These equations hold iff $d^2=-1$ (all equations with odd $a+b$).\qedhere} \end{align*} \end{enumerate} \end{proof} \noindent We formulate the results of this section: \begin{theorem} The abelian category ${\rm Rep}(U)$ admits up to equivalence a unique structure of a braided tensor category such that $\mathbb{V},\overline{\mathbb{V}}:{\rm Rep}(C) \to {\rm Rep}(U)$ are braided oplax tensor functors with the braiding on $C$ given by the quadratic form $Q(k)=\beta^{(k^2)}$, for fixed $\beta^4=-1$. \end{theorem} \noindent It follows from Theorem \ref{thm:oplax} that the triplet algebra $\mathcal{W}(2)$ admits two such oplax tensor functors $\mathbb{V},\overline{\mathbb{V}}:{\rm Rep} V_L \to {\rm Rep} \mathcal{W}(2)$ where the $Q$ is determined by the quadratic form of the lattice $L=\sqrt{2} A_1$, so $\beta=e^{2\pi i(1/8)}$. It then follows that the abelian category $\mathrm{Rep}(U)$ of the Kazhdan-Lusztig dual $U$ of $\mathcal{W}(2)$ must admit the functors described in Theorem \ref{classif}. We therefore obtain as a corollary a proof of the Kazhdan-Lusztig conjecture for $\mathcal{W}(2)$: \begin{corollary} There is a braided tensor equivalence of ${\rm Rep}(\mathcal{W}(2))$ to the representation category of the previously constructed quasi-Hopf algebra for this value of $\beta$. The quasi Hopf algebra coincides with the one constructed in \cite{FGR2}. \end{corollary} \newcommand\arxiv[2] {\href{http://arXiv.org/abs/#1}{#2}}

e29c8f63ce748375884257fe93e0108c4c4bd732

\section{Introduction} Entanglement, first introduced in the EPR paper \cite{einstein}, is a quantum mechanical feature that can be used as a resource for computational and communicational purposes \cite{ekert1,horodecki4}. It plays a central role in many information processing protocols such as quantum cryptography \cite{ekert}, quantum superdense coding \cite{bennett2} and quantum teleportation \cite{bennett1,bennett3}. The potential offered by quantum entanglement to computing, security and communication makes it a topic of vital interest to researchers all across the globe.\\ One of the important problem in quantum information theory is the detection of entanglement in a quantum mechanical system. A pure two-qubit entangled system always violated Bell-CHSH inequality and thus detected by Bell-CHSH operator \cite{bennett4,deutsch}. On the other hand, the Bell-CHSH inequality fails to detect the several mixed bipartite entangled state. This loophole can be fixed using Peres-Horodecki (PH) positive partial transpose (PPT) criteria, which is necessary and sufficient for the detection of entanglement in $2 \otimes 2$ and $2 \otimes 3$ systems \cite{peres,horo1}. In higher dimensional systems, all states with negative partial transpose (NPT) are entangled but the states with positive partial transpose may or may not be entangled \cite{phorodecki}. The entangled states which are described by a density matrix that remains positive under partial transposition are known as bound entangled states. Thus, the separability problem can also be framed as analysing whether states with positive partial transposition are entangled or not.\\ The separability problem can be tackled to certain extent by witness operator \cite{lewenstein,ganguly}. Witness operators are hermitian operators with at least one negative eigenvalues. The witness operators are more powerful than Bell inequalities in the sense that it can detect multipartite entanglement in different cuts, if some prior information about the state under investigation is provided. They not only detect multipartite entanglement in different cuts but also detect genuine entanglement and classify entanglement in multipartite system. They are observables and thus experimentally realizable also.\\ A map $\Lambda:M_{d_{1}}(C)\rightarrow M_{d_{2}}(C)$ is said to be positive map if $\Lambda(A) \in M_{d_{2}}(C)$ is positive, for any positive $A \in M_{d_{1}}(C)$. Unfortunately, the structure of positive map is not completely understood and still it is under extensive research \cite{stormer1,stormer2,woronowicz}. Indecomposable maps plays an important role among all positive linear maps due to the fact that it can detect positive partial transpose entangled states. Let $\Lambda$ be a positive map and let $I_{d}:M_{d}(C)\rightarrow M_{d}(C)$ denote an identity map. Then we say that $\Lambda$ is completely positive if for all d, the extended map $I_{d} \otimes \Lambda$ is positive. Trace map is an example of a completely positive map. There also exist a map such as transposition map which is positive but not completely positive map. These positive but not completely positive maps are important in detecting the entanglement in a composite quantum system.\\ The separability problem can be reformulated in terms of positive maps \cite{horo1} as follows: Let us suppose that $H_{d_{1}}$ and $H_{d_{2}}$ represent two Hilbert spaces with dimensions $d_{1}$ and $d_{2}$ respectively. A bipartite quantum state described by the density operator $\sigma \in H_{d_{1}}\otimes H_{d_{2}}$ is separable if and only if $(I_{d_{1}} \otimes \Lambda)\sigma$ is positive for any positive map $\Lambda$. Thus there is a deep relation between the theory of detection of entanglement and operator theory. This linkage has been established by Choi-Jamiolkowski isomorphism \cite{choi,jamiolkowski}. According to Choi-Jamiolkowski isomorphism, there is a one to one correspondence between entanglement witnesses and a positive but not completely positive map.\\ The motivation of this work is as follows: The construction and studying the structure of new positive map may give useful insight in the understanding of positive map, which gives us the first motivation of this work. Secondly, we find that the problem of constructing the positive but not completely positive map and its relation in the detection of entanglement may take one step further in the development of not only operator theory but also quantum information theory.\\ The work is organized as follows: In section-II, we review few earlier results that is needed in our work. In section-III, we have constructed a family of map and then characterize it for when the map is (i) positive or (ii) completely positive or (iii) positive but not completely positive. In section-IV, we have chosen a particular map from the class of positive but not completely positive map and then shown that it can detect bound entangled entangled state. In section-V, we end up with the conclusion. \section{Preliminaries} \textbf{Definition-1:} A matrix $V$ is said to be a contraction if \begin{equation} \|V\|\leq 1 \label{contraction} \end{equation} where $\|.\|$ denote the operator norm.\\ \textbf{Result-1 \cite{stormer}:} Let $\phi: M_{m}(C)\rightarrow M_{n}(C)$ be a linear map. Then the following statements are equivalent:\\ (i) $\phi$ is completely positive.\\ (ii) $C_{\phi}$ is positive semidefinite, where $C_{\phi}$ denotes the Choi matrix of $\phi$.\\ \textbf{Result-2 \cite{zhang}:} Let A be a partitioned block matrix of the form \begin{eqnarray} A= \begin{pmatrix} X & Y \\ Y^{*} & Z \end{pmatrix} \label{block} \end{eqnarray} Then A is positive semidefinite if and only if X and Z are positive semidefinite and there exists a contraction $V$ such that \begin{equation} Y = X^{\frac{1}{2}}VZ^{\frac{1}{2}} \label{psdcond1} \end{equation} \textbf{Result-3 \cite{sharma}:} For positive definite blocks $X$ and $Z$, the matrix $A$ given in (\ref{block}) is positive semidefinite iff $X\geq YZ^{-1}Y^{\dagger}$. \section{Construction of a family of map} In this section, we will construct a map and derive the condition for which the map is positive. Further, we will probe that whether the constructed map is completely positive. Moreover, we will provide the explicit matrix form of the map, which is positive but not completely positive.\\ Let us take a positive integer $n~~(n\geq 2)$ and then we define a general family of map $\Phi: M_{n}(C)\rightarrow M_{n}(C) \otimes M_{n}(C)$ as \begin{equation} \Phi_{\alpha,\beta}(A) = \alpha((A + A^{T}) \otimes I_{n}) + \beta ( |\psi_{+}\rangle\langle \psi_{+}| )^{\Gamma} \label{map} \end{equation} where $A$ denote $n\times n$ matrix, $\alpha,\beta \in R$, $\Gamma$ represent the partial transposition, $I_{n}$ denote the identity matrix of order $n$ and $|\psi_{+}\rangle=\frac{1}{\sqrt{n}}\sum_{i=1}^{n}|ii\rangle$.\\ To discuss our result, we will fix $n=2$ and re-define the map $\Phi: M_{2}(C)\rightarrow M_{2}(C) \otimes M_{2}(C)$ as \begin{equation} \Phi_{\alpha,\beta}(A) = \alpha((A + A^{T}) \otimes I_{2}) + \beta ( |\psi_{+}\rangle\langle \psi_{+}| )^{\Gamma} \label{map} \end{equation} where $I_{2}$ denote the identity matrix of order $2$ and $|\psi_{+}\rangle=\frac{1}{\sqrt{2}}(|00\rangle + |11\rangle)$.\\ For any $a,d \geq 0$ and $b,c \in \textrm{R}$, we can take the input matrix $A \in M_{2}(R)$ of the form \begin{eqnarray} A= \begin{pmatrix} a & b \\ c & d \end{pmatrix} \label{input} \end{eqnarray} In matrix notation, the output of the map $\Phi_{\alpha,\beta}$ can be expressed as \begin{eqnarray} \Phi_{\alpha,\beta}= \begin{pmatrix} 2a\alpha + \frac{\beta}{2}& 0 & \alpha(b+c) & 0 \\ 0 & 2a\alpha & \frac{\beta}{2} & \alpha(b+c) \\ \alpha(b+c) & \frac{\beta}{2} & 2d\alpha & 0 \\ 0 & \alpha(b+c) & 0 & 2d\alpha + \frac{\beta}{2} \end{pmatrix} \label{mapmat} \end{eqnarray} \subsection{Conditions for which a map $\phi$ will be positive} We will derive here the conditions for which $\Phi$ represent a positive map. The map $\Phi$ will be positive if the matrix represented by $\Phi_{\alpha,\beta}$ given in (\ref{mapmat}) is a positive semi-definite matrix. To accomplish this task, we re-express $\Phi_{\alpha,\beta}$ in a block matrix form as \begin{eqnarray} \Phi_{\alpha,\beta}= \begin{pmatrix} X & Y \\ Y^{*} & Z \end{pmatrix} \label{block1} \end{eqnarray} where \begin{eqnarray} &&X= \begin{pmatrix} 2a\alpha + \frac{\beta}{2}& 0 \\ 0 & 2a\alpha \end{pmatrix}, Y=\begin{pmatrix} \alpha(b+c) & 0 \\ \frac{\beta}{2} & \alpha(b+c) \end{pmatrix},\nonumber\\&&Z=\begin{pmatrix} 2d\alpha & 0 \\ 0 & 2d\alpha + \frac{\beta}{2} \end{pmatrix} \label{block1} \end{eqnarray} Applying Result-2 on $\Phi_{\alpha,\beta}$, we can state that the matrix $\Phi_{\alpha,\beta}$ will be positive semidefinite if the following conditions hold:\\ \begin{eqnarray} (i) X\geq 0\Rightarrow 2a\alpha \geq 0 ~~\textrm{and}~~ 4a\alpha + \beta \geq 0 \label{cond1} \end{eqnarray} \begin{eqnarray} (ii) Z\geq 0\Rightarrow 2d\alpha \geq 0 ~~\textrm{and}~~ 4d\alpha + \beta \geq 0 \label{cond2} \end{eqnarray} \begin{eqnarray} (iii)&& \|V\|=\|X^{\frac{-1}{2}}YZ^{\frac{-1}{2}}\|\leq 1 \nonumber\\&& \Rightarrow \|\begin{pmatrix} \frac{\alpha(b+c)}{\sqrt{d\alpha(4a\alpha+\beta)}} & 0\\ \frac{\beta}{4\alpha\sqrt{ad}} & \frac{\alpha(b+c)}{\sqrt{a\alpha(4d\alpha+\beta)}} \end{pmatrix}\|\leq 1 \label{cond3} \end{eqnarray} where $\|V\|$ denote the operator norm of $V$.\\ Conditions $(i)$ and $(ii)$ given by (\ref{cond1}) and (\ref{cond2}) are collectively given by \begin{eqnarray} 2\alpha(a+d) +\beta \geq 0,~~\alpha \geq 0 \label{cond100} \end{eqnarray} Now our task is to take into account condition (iii) in which we need to calculate the operator norm of the matrix $V$. Operator norm of the matrix $V$ is defined as the maximum eigenvalue of $V^{T}V$. The eigenvalue of $V^{T}V$ can be calculated from the characteristic equation of $V^{T}V$. The characteristic equation of $V^{T}V$ is given by \begin{eqnarray} &&\lambda^2 - k_{1}\lambda + \frac{k_{2}}{4}=0 \label{char} \end{eqnarray} where $k_{1}=\frac{\alpha(b+c)^2}{(4a\alpha+\beta)d} + \frac{\beta}{4a\alpha} + \frac{\beta^2}{16ad\alpha^2} + \frac{d}{a}$ and $k_{2}=\frac{(b+c)^2 (\beta+4d\alpha)}{ad(4a\alpha+\beta)}$.\\ Since $a\geq 0$ and $d\geq0$ from the earlier assumptions and using equations (\ref{cond1}) and (\ref{cond2}), we can infer that $k_{1}\geq 0$ and $k_{2}\geq 0$. Thus, it is clear from Descarte's rule of sign that the two roots of the characteristic equation given by (\ref{char}) will be positive. If $\lambda_{1}$ and $\lambda_{2}$ denote two positive eigenvalues of $V^{T}V$ then they are given by \begin{eqnarray} &&\lambda_1 = \frac{1}{2}(k_{1} + \sqrt{k_{1}^{2} - k_{2}})\nonumber\\&& \lambda_2 = \frac{1}{2}(k_{1} - \sqrt{k_{1}^{2} - k_{2}}) \label{eigen1} \end{eqnarray} Since both the eigenvalues are positive so $\|V\|=max\{\lambda_{1},\lambda_{2}\}=\lambda_{1}$. The condition (iii) says that $\|V\|\leq 1$ which implies \begin{eqnarray} 4(1 + \sqrt{k_{1}^{2} - k_{2}} -k_{2}) \geq 1 \label{cond300} \end{eqnarray} The map $\Phi_{\alpha,\beta}$ is positive if equations (\ref{cond100}) and (\ref{cond300}) holds simultaneously. In particular, the map $\Phi_{\alpha,\beta}$ will be positive for $\alpha\geq 0$ and $\beta=0$. \subsection{Is the map $\phi$ completely positive?} In this section, we will investigate the fact that whether the map $\phi$ is completely positive. To do this, we begin with the construction of Choi matrix corresponding to the positive operator $\Phi_{\alpha,\beta}$. The Choi matrix $C_{\Phi_{\alpha,\beta}}$ is defined as \cite{choi} \begin{eqnarray} C_{\Phi_{\alpha,\beta}}=\sum_{i,j=0}^{1}|i\rangle\langle j|\otimes \Phi_{\alpha,\beta}(|i\rangle\langle j|) \label{choimatrix} \end{eqnarray} where $|i\rangle$ represent the basis state in two-dimensional Hilbert space.\\ The Choi matrix $C_{\Phi_{\alpha,\beta}}$ can be re-expressed in terms of matrix as \begin{eqnarray} C_{\Phi_{\alpha,\beta}}= \begin{pmatrix} 2\alpha +\frac{\beta}{2} & 0 & 0 & 0 & \frac{\beta}{2} & 0 & \alpha & 0 \\ 0 & 2\alpha & \frac{\beta}{2} & 0 & 0 & 0 & \frac{\beta}{2} & \alpha \\ 0 & \frac{\beta}{2} & 0 & 0 & \alpha & \frac{\beta}{2} & 0 & 0\\ 0 & 0 & 0 & \frac{\beta}{2} & 0 & \alpha & 0 & \frac{\beta}{2} \\ \frac{\beta}{2} & 0 & \alpha & 0 & \frac{\beta}{2} & 0 & 0 & 0\\ 0 & 0 & \frac{\beta}{2} & \alpha & 0 & 0 & \frac{\beta}{2} & 0\\ \alpha & \frac{\beta}{2} & 0 & 0 & 0 & \frac{\beta}{2} & 2\alpha & 0\\ 0 & \alpha & 0 & \frac{\beta}{2} & 0 & 0 & 0 & 2\alpha +\frac{\beta}{2} \end{pmatrix}, \label{choimatrix1} \end{eqnarray} To show the completely positivity of a positive map $\Phi_{\alpha,\beta}$, we need to the show that the choi matrix $C_{\Phi_{\alpha,\beta}}$ corresponding to the positive map $\Phi_{\alpha,\beta}$ is positive semidefinite. We first express the choi matrix in block form as \begin{eqnarray} C_{\Phi_{\alpha,\beta}}= \begin{pmatrix} P & Q \\ Q^{*} & R \end{pmatrix} \label{block1} \end{eqnarray} where \begin{eqnarray} &&P= \begin{pmatrix} 2\alpha +\frac{\beta}{2} & 0 & 0 & 0 \\ 0 & 2\alpha & \frac{\beta}{2} & 0 \\ 0 & \frac{\beta}{2} & 0 & 0 \\ 0 & 0 & 0 & \frac{\beta}{2} \end{pmatrix}, Q=\begin{pmatrix} \frac{\beta}{2} & 0 & \alpha & 0 \\ 0 & 0 & \frac{\beta}{2} & \alpha \\ \alpha & \frac{\beta}{2} & 0 & 0 \\ 0 & \alpha & 0 & \frac{\beta}{2} \end{pmatrix},\nonumber\\&& R=\begin{pmatrix} \frac{\beta}{2} & 0 & 0 & 0 \\ 0 & 0 & \frac{\beta}{2} & 0 \\ 0 & \frac{\beta}{2} & 2\alpha & 0 \\ 0 & 0 & 0 & 2\alpha + \frac{\beta}{2} \end{pmatrix} \label{block2} \end{eqnarray} Following Result-3, we can show that the choi matrix $C_{\Phi_{\alpha,\beta}}$ is positive semidefinite if and only if the following conditions are satisfied: \begin{eqnarray} (i)~~ P\geq 0 ~~\textrm{holds when}~~\beta = 0~~\textrm{and}~~ \alpha \ge 0 \label{cond12} \end{eqnarray} \begin{eqnarray} (ii)~~ R \geq 0~~\textrm{holds when}~~\beta = 0~~ \textrm{and}~~ \alpha \ge 0 \label{cond22} \end{eqnarray} \begin{eqnarray} &&(iii)~~P - Q{R^{ - 1}}Q^{*} \ge 0~~ \textrm{holds for}~~ \textrm{either}~~\nonumber\\&& (\alpha=0~~ \textrm{and}~~ \beta \neq 0)~~ \textrm{or}~~ (\alpha > 0~~ \textrm{and}~~ 4\alpha +\beta < 0)\nonumber\\&&~~ \textrm{or} ~~ (\alpha > 0~~,~~ 3\alpha +2\beta \ge 0~~\textrm{and}~~\beta \neq 0) \label{cond23} \end{eqnarray} It can be easily observe that the conditions $(i)$, $(ii)$ and $(iii)$ does not hold simultaneously. Thus the map $\Phi_{\alpha,\beta}$ is not completely positive. \subsection{Conditions for which a map $\phi$ will be positive but not completely positive} In the previous sections, we have derived the condition for which the map $\Phi$ will be positive and later we proved that the positive map $\Phi$ cannot be completely positive. In this section, we will derive the common interval of $\alpha$ for which the map $\Phi$ will be positive but not completely positive simultaneously.\\ Without any loss of generality, let us consider the $2\times 2$ positive matrix $A_{1}\in M_{2}(\textrm{R})$ as \begin{eqnarray} A_{1}= \begin{pmatrix} \frac{1}{4} & \frac{1}{3} \\ \frac{1}{9} & 2 \end{pmatrix} \end{eqnarray} Further, taking $\beta=-\gamma (\gamma>0)$, the output of the mapping can be represented by the matrix as \begin{eqnarray} \Phi_{\alpha,-\gamma}(A_{1})= \begin{pmatrix} \frac{\alpha}{2} - \frac{\gamma}{2}& 0 & \frac{4\alpha}{9} & 0 \\ 0 & \frac{\alpha}{2} & -\frac{\gamma}{2} & \frac{4\alpha}{9} \\ \frac{4\alpha}{9} & -\frac{\gamma}{2} & 4\alpha & 0 \\ 0 & \frac{4\alpha}{9} & 0 & 4\alpha - \frac{\gamma}{2} \end{pmatrix} \label{mapmat300} \end{eqnarray} It can be easily shown that the map $\Phi_{\alpha,-\gamma}$ always produces a positive matrix at the output if $\gamma>0$ and $\alpha\geq \frac{9\gamma}{2\sqrt{146}}$. Thus $\Phi_{\alpha,-\gamma}$ represent a positive map if $\gamma>0$ and $\alpha\geq \frac{9\gamma}{2\sqrt{146}}$. Furthermore, the Choi matrix corresponding to the positive map $\Phi_{\alpha,-\gamma}$ is given by \begin{eqnarray} C_{\Phi_{\alpha,-\gamma}}= \begin{pmatrix} m_{1} & 0 & 0 & 0 & -\frac{\gamma}{2} & 0 & \alpha & 0 \\ 0 & 2\alpha & -\frac{\gamma}{2} & 0 & 0 & 0 & -\frac{\gamma}{2} & \alpha \\ 0 & -\frac{\gamma}{2} & 0 & 0 & \alpha & -\frac{\gamma}{2} & 0 & 0\\ 0 & 0 & 0 & -\frac{\gamma}{2} & 0 & \alpha & 0 & -\frac{\gamma}{2} \\ -\frac{\gamma}{2} & 0 & \alpha & 0 & -\frac{\gamma}{2} & 0 & 0 & 0\\ 0 & 0 & -\frac{\gamma}{2} & \alpha & 0 & 0 & -\frac{\gamma}{2} & 0\\ \alpha & -\frac{\gamma}{2} & 0 & 0 & 0 & -\frac{\gamma}{2} & 2\alpha & 0\\ 0 & \alpha & 0 & -\frac{\gamma}{2} & 0 & 0 & 0 & m_{1} \end{pmatrix} \label{choimatrix100} \end{eqnarray} where $m_{1}=2\alpha-\frac{\gamma}{2}$.\\ The eigenvalues of $C_{\Phi_{\alpha,-\gamma}}$ are given by \begin{eqnarray} &&\mu_{1}=\frac{-\gamma+\sqrt{4\alpha^{2}+\gamma^{2}}}{2},~~\mu_{2}=\frac{-\gamma-\sqrt{4\alpha^{2}+\gamma^{2}}}{2}\nonumber\\&& \mu_{3}=\frac{4\alpha-\gamma+\sqrt{4\alpha^{2}+\gamma^{2}}}{2},\nonumber\\&& \mu_{4}=\frac{4\alpha-\gamma-\sqrt{4\alpha^{2}+\gamma^{2}}}{2},\nonumber\\&& \mu_{5}=\alpha+\sqrt{\frac{4\alpha^{2}+\gamma^{2}+\sqrt{16\alpha^{2}+4\alpha^{2}\gamma^{2}+\gamma^{4}}}{2}}\nonumber\\&& \mu_{6}=\alpha+\sqrt{\frac{4\alpha^{2}+\gamma^{2}-\sqrt{16\alpha^{2}+4\alpha^{2}\gamma^{2}+\gamma^{4}}}{2}}\nonumber\\&& \mu_{7}=\alpha-\sqrt{\frac{4\alpha^{2}+\gamma^{2}+\sqrt{16\alpha^{2}+4\alpha^{2}\gamma^{2}+\gamma^{4}}}{2}}\nonumber\\&& \mu_{8}=\alpha-\sqrt{\frac{4\alpha^{2}+\gamma^{2}-\sqrt{16\alpha^{2}+4\alpha^{2}\gamma^{2}+\gamma^{4}}}{2}}\nonumber\\&& \label{cond320} \end{eqnarray} It can be observed that the Choi matrix $C_{\Phi_{\alpha,-\gamma}}$ has at least one negative eigenvalues for any $\alpha$ and $\gamma$. Therefore, $\Phi_{\alpha,-\gamma}$ is not completely positive map for any $\alpha$ and $\gamma$. Thus for $\gamma>0$ and $\alpha\geq \frac{9\gamma}{2\sqrt{146}}$, the map $\Phi_{\alpha,-\gamma}$ is positive but not completely positive. \section{Positive but not completely positive map act as witness operator for the detection of entangled states} In this section, we will construct a specific map which is positive but not completely positive and then use it to detect negative partial transpose entangled states and bound entangled states. We will construct the Choi matrix from the positive map that can be considered as a witness operator. A witness operator $W$ is a hermitian operator, which satisfies the following properties: \begin{eqnarray} &&(i)~~Tr(W \rho_{s})\geq 0,~~ \textrm{for all separable state}~~ \rho_{s}\nonumber\\&& (ii)~Tr(W \rho_{e})<0,~~ \textrm{for at least one entangled state}~~ \rho_{e}\nonumber\\&& \label{witnesscond} \end{eqnarray} \subsection{Detection of Bound Entangled State} To achieve our task, let us first fix $\gamma=2$ and then choose a value of $\alpha$ from the interval $\alpha\geq \frac{9}{\sqrt{146}}$. Taking $\alpha=\frac{3}{4}$, the matrix given in (\ref{mapmat300}) reduces to \begin{eqnarray} \Phi_{\frac{3}{4},-2}(A_{1})= \begin{pmatrix} \frac{11}{8}& 0 & \frac{1}{3} & 0 \\ 0 & \frac{3}{8} & -1 & \frac{1}{3} \\ \frac{1}{3} & -1 & 3 & 0 \\ 0 & \frac{1}{3} & 0 & 4 \end{pmatrix} \label{mapmat301} \end{eqnarray} In particular, the map $\Phi_{\frac{3}{4},-2}$ represent a positive map. Using this positive map, we can construct the Choi matrix which is given below: \begin{eqnarray} C_{\Phi_{\frac{3}{4},-2}}= \begin{pmatrix} \frac{1}{2} & 0 & 0 & 0 & -1 & 0 & \frac{3}{4} & 0 \\ 0 & \frac{3}{2} & -1 & 0 & 0 & 0 & -1 & \frac{3}{4} \\ 0 & -1 & 0 & 0 & \frac{3}{4} & -1 & 0 & 0\\ 0 & 0 & 0 & -1 & 0 & \frac{3}{4} & 0 & -1 \\ -1 & 0 & \frac{3}{4} & 0 & -1 & 0 & 0 & 0\\ 0 & 0 & -1 & \frac{3}{4} & 0 & 0 & -1 & 0\\ \frac{3}{4} & -1 & 0 & 0 & 0 & -1 & \frac{3}{2} & 0\\ 0 & \frac{3}{4} & 0 & -1 & 0 & 0 & 0 & \frac{1}{2} \end{pmatrix}, \label{choimatrix300} \end{eqnarray} The Choi matrix $C_{\Phi_{\frac{3}{4},-2}}$ has at least one negative eigenvalues and thus it does not represent a positive semidefinite matrix. Hence $\Phi_{\frac{3}{4},-2}$ is a positive but not completely positive map.\\ Next our task is to show that $C_{\Phi_{\frac{3}{4},-2}}$ act as witness operator and for this it is sufficient to show that there exist at least one entangled states described by the density operator $\rho_{e}$ for which $Tr(C_{\Phi_{\frac{3}{4},-2}} \rho_{e})<0$. Then we can say that the entangled state will be detected by $C_{\Phi_{\frac{3}{4},-2}}$.\\ Let us consider a quantum state described by the density operator $\rho_{b}$ which is given by \begin{eqnarray} \rho_{b}= \frac{1}{1+7b} \begin{pmatrix} b & 0 & 0 & 0 & 0 & b & 0 & 0 \\ 0 & b & 0 & 0 & 0 & 0 & b & 0 \\ 0 & 0 & b & 0 & 0 & 0 & 0 & b\\ 0 & 0 & 0 & b & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & \frac{1+b}{2} & 0 & 0 & \frac{\sqrt{1-b^{2}}}{2}\\ b & 0 & 0 & 0 & 0 & b & 0 & 0\\ 0 & b & 0 & 0 & 0 & 0 & b & 0\\ 0 & 0 & b & 0 & \frac{\sqrt{1-b^{2}}}{2} & 0 & 0 & \frac{1+b}{2} \end{pmatrix} \label{pptes} \end{eqnarray} where the state parameter satisfies $0\leq b\leq 1$. The state $\rho_{b}$ is shown to be a bound entangled state by range criterion \cite{phorodecki}.\\ We are now in a position to show the utility of the operator $C_{\Phi_{\frac{3}{4},-2}}$ in the detection of entanglement. To accomplish this task, we calculate $Tr(C_{\Phi_{\frac{3}{4},-2}} \rho_{b})$, which is given by \begin{eqnarray} Tr(C_{\Phi_{\frac{3}{4},-2}} \rho_{b})=\frac{b-1}{4(1+7b)}<0 \label{witnesscond1} \end{eqnarray} Thus the bound entangled state $\rho_{b}$ detected by the witness operator $C_{\Phi_{\frac{3}{4},-2}}$. \subsection{Detection of Negative Partial Transpose Entangled State} Let us consider a quantum state described by the density operator $\rho_{NPT}$ which is given by \begin{eqnarray} \rho_{NPT}= \frac{1}{3} \begin{pmatrix} 1 & 0 & 0 & 0 & 0 & 1 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\ 1 & 0 & 0 & 0 & 0 & 1 & 0 & 0\\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 \end{pmatrix} \label{nptes} \end{eqnarray} It can be easily shown that the state (\ref{nptes}) represent a negative partial transpose entangled state. Now, our task is to construct a witness operator which can detect it. To accomplish this task, let us start with the positive input matrix which is given by \begin{eqnarray} A_{2}= \begin{pmatrix} 3 & \frac{1}{3} \\ \frac{1}{9} & 2 \end{pmatrix} \end{eqnarray} Considering $\beta=-\gamma (\gamma>0)$ and applying the map on $A_{2}$, we get the output matrix in the form \begin{eqnarray} \Phi_{\alpha,-\gamma}(A_{2})= \begin{pmatrix} 6\alpha - \frac{\gamma}{2}& 0 & \frac{4\alpha}{9} & 0 \\ 0 & 6\alpha & -\frac{\gamma}{2} & \frac{4\alpha}{9} \\ \frac{4\alpha}{9} & -\frac{\gamma}{2} & 4\alpha & 0 \\ 0 & \frac{4\alpha}{9} & 0 & 4\alpha - \frac{\gamma}{2} \end{pmatrix} \label{mapmat400} \end{eqnarray} The map $\Phi_{\alpha,-\gamma}$ will be positive map if $\gamma>0$ and $\alpha\geq \frac{9\gamma}{90-2\sqrt{27}}$. In the next step, we fix $\gamma=1$ and then choose a value of $\alpha$ from the interval $\alpha\geq \frac{9}{90-2\sqrt{27}}$. Taking $\alpha=\frac{1}{8}$, the matrix given in (\ref{mapmat400}) reduces to \begin{eqnarray} \Phi_{\frac{1}{8},-1}(A_{2})= \begin{pmatrix} \frac{1}{4}& 0 & \frac{1}{18} & 0 \\ 0 & \frac{3}{4} & \frac{-1}{2} & \frac{1}{18} \\ \frac{1}{18} & \frac{-1}{2} & \frac{1}{2} & 0 \\ 0 & \frac{1}{18} & 0 & 0 \end{pmatrix} \label{mapmat301} \end{eqnarray} Therefore, the particular form of the map $\Phi_{\frac{1}{8},-1}$ represent a positive map. Using this positive map, we can construct the Choi matrix as \begin{eqnarray} C_{\Phi_{\frac{1}{8},-1}}= \begin{pmatrix} \frac{-1}{4} & 0 & 0 & 0 & \frac{-1}{2} & 0 & \frac{1}{8} & 0 \\ 0 & \frac{1}{4} & \frac{-1}{2} & 0 & 0 & 0 & \frac{-1}{2} & \frac{1}{8} \\ 0 & \frac{-1}{2} & 0 & 0 & \frac{1}{8} & \frac{-1}{2} & 0 & 0\\ 0 & 0 & 0 & \frac{-1}{2} & 0 & \frac{1}{8} & 0 & \frac{-1}{2} \\ \frac{-1}{2} & 0 & \frac{1}{8} & 0 & \frac{-1}{2} & 0 & 0 & 0\\ 0 & 0 & \frac{-1}{2} & \frac{1}{8} & 0 & 0 & \frac{-1}{2} & 0\\ \frac{1}{8} & \frac{-1}{2} & 0 & 0 & 0 & \frac{-1}{2} & \frac{1}{4} & 0\\ 0 & \frac{1}{8} & 0 & \frac{-1}{2} & 0 & 0 & 0 & \frac{-1}{4} \end{pmatrix}, \label{choimatrix500} \end{eqnarray} The Choi matrix $C_{\Phi_{\frac{1}{8},-1}}$ has at least one negative eigenvalues and thus it does not represent a positive semidefinite matrix. Hence $\Phi_{\frac{1}{8},-1}$ is a positive but not completely positive map.\\ We will now show that $C_{\Phi_{\frac{3}{4},-2}}$ act as witness operator and it detect the state (\ref{nptes}). To detect the state described by the density operator $\rho_{NPT}$, we calculate $Tr(C_{\Phi_{\frac{3}{4},-2}} \rho_{NPT})$, which is given by \begin{eqnarray} Tr(C_{\Phi_{\frac{1}{8},-1}} \rho_{NPT})=\frac{-1}{6}<0 \label{witnesscond1} \end{eqnarray} Thus the negative partial transpose entangled state $\rho_{NPT}$ detected by the witness operator $C_{\Phi_{\frac{1}{8},-1}}$. \section{Conclusion} To summarize, we have constructed a map which is applied on $n\times n$ matrix and as a result, we obtain $n^{2}\times n^{2}$ matrix at the output. The mapping constructed here is general and work for higher order matrices also. But to simplify the discussion, we have taken $n=2$ and then showed that the map is positive under certain conditions. Further, we have shown that the constructed map can never be completely positive and also obtained the conditions for which the map is positive but not completely positive. Lastly, we have discussed that the Choi matrix constructed from the positive map can act as a witness operator and take part in the detection of bound entangled state and negative partial transpose entangled state.

ce7faba41a8e8b793448343d1bb205db2dcf0f6f

\section{Introduction} Standard predator-prey models treat predators and prey as non-evolving, non-learning entities. In reality, over time both predators and prey start a co-evolutionary chase which is fueled by continual evolution and learning. A simple example is that of a literal chase where the predators consume the slower prey which are the easiest to catch. The remaining prey population is faster on average and succeeding generations of prey are naturally selected to be faster in a Darwinian evolution. As a result of quicker prey, under negative selection pressure, the slower predators are pruned out over time, which results in faster predators. It's the beginning of an adversarial chase! Similarly toxicity of prey and resistance of predator evolve through generations, if we consider the case where the lifespan of a predator covers a significantly large number of generations of prey. Such a co-evolution may be further limited by the premise that it will not let the predator gain ability to evolve fast enough to develop immunity to the toxic prey. But instead, as seen in nature, we may endow the predators with a learning mechanism and thus allow them to learn and continually adapt over time.[1] When a predator consumes a prey they update their inner beliefs about the prey based on the effect the prey has on the predator. To this extent, this paper treats the predator as a learning agent who uses reinforcement learning to update their policy of choosing to eat or not eat a prey of the same type when encountering it again in the future. But to effectively execute such strategies, a necessary requirement is the ability of the predator to distinguish between multiple prey types so that they can update the respective belief regarding that type. This ability to distinguish arises from the signal sent by the corresponding prey. A signal can be visual, acoustic or even a smell. For example a red bird sends a visual signal, different from a blue bird.[2] \\ Breaking it down, the predators have prior beliefs which may be learnt during the predator's lifetime, or genetically programmed. Based on the signal the predator receives from the prey they choose to either consume or avoid the prey that sent the signal. Due to prey evolution and natural selection the signals of a prey species changes through the generations. It gets even more interesting when two different species evolve to send the same signal. Effectively the predator cannot distinguish between the two and will thus share a common belief with respect to the common signal. This is commonly known as prey mimicry.[3] Some common types of mimicry are Batesian and Mullerian mimicry which arise due to prey toxicity. For a non evolving predator-prey system one can find equilibrium conditions and dynamics but it gets interesting in our current situation of learning predators and evolving prey.[4] \\ The model has many direct applications to vaccine design that mimics a yet-to-be-encountered (but known) virus. Guided by the model, the designed vaccine embeds itself in a Mullerian mimicry-ring, thus developing tolerance from and memory in the host-immune system. A quasi-Batesian virus then has difficulties to invade the host despite its previously-acquired abilities to deceive the host system via mimicry. As the adversarial chase ensues, virus evolves additional variants, which need to be recognized, analyzed and learned (AI/ML), leading to booster vaccines, resulting in a complex game of Whack-a-Mole. What would be the best learning strategies, if one's goal is to achieve a herd immunity rapidly in the population with respect to the collection of variants? \\ Thus, we see a fascinating variety of phenomena arising due to differences in efficacy-values the predator (resp. vaccine) has with respect to the different prey-species (resp. virus). We effectively model this using utility. For example, the utility of a toxic prey to a predator will be negative while a nutritious prey will be positive. Furthermore the utility function encodes preference and is exactly what we need to model the behaviour of evolving predators and prey. In regards to two mimicking species that share a common signal, the utility the predator associates to the signal will be the expected utility from the two species in question.\\ To keep matters simple, we analyse the following somewhat idealized system. We consider a signalling game with multiple species of predator and prey. The prey (Lifespan of Mayfly $\approx$24hours) has a relatively short lifespan compared to the the predator (Lifespan of Common Toad $\approx$12years). This way the prey can undergo mutations in its signal and adapt while the predator can only learn about the prey but cannot change genetically by germ-line evolution in the short period. As the signals of the prey mutate and evolve there will be convergence and divergence of signals of different species. We try to understand the conditions and extent of mimicry with respect to the utility value of different prey species to the predator. Here, the predators will all be modeled as multi-armed bandit agents. This dynamics results in a few interesting phenomena that arise as a mathematical consequence of our assumptions.\\ \vspace*{0.08in}\noindent{\bf New work}\\ Bio-mimicry has scarcely been interpreted using a formal mathematical framework of utility. Another novel idea is using multi-armed bandit predators that can continuously learn and update their belief regarding the non-stationary distribution of multi-dimensional signals. \\ Most of the following phenomena explained below are just mathematically feasible scenarios. Certain phenomenon such as: `Mimicry can be bad to both species' is discussed here for the first time, albeit, such hypothetical scenarios may or may not have natural/biological feasibility. Thus, `Mimicry can be bad to both species' may only refer to a novel but refutable hypothesis suggesting that mimicry is bad resulting in the extinction of both species involved, and will require falsifying experiments.\\ \section{Model} \subsection{Ecosystem} The ecosystem consists of multiple prey and multiple predators. On a given day a predator encounters some number of prey. Each prey gives out a respective signal. Based on the predators prior knowledge it chooses to consume a prey emitting a specific signal. After eating the prey, based on the utility it got from the prey it updates its information about the signal. This continues over time leading to a variety of phenomena. \\ It is to be noted that the prey mutate and evolve over each generation but the predator whose lifespan is significantly longer than the prey continuously learns and updates its knowledge. As naturally seen, the predator comes across a randomly sampled set of prey and chooses to consume one prey out of them. \\ There are two types of ecosystems we can focus on. In the first the prey also compete with each other for resources. In the second one the prey have a reproduction rate and can indefinitely grow if there is no predation. Both ecosystems have very distinct properties and showcase very different phenomena under similar conditions. We mainly focus on the indefinite growth in the examples seen in this paper. \subsection{Prey} Every prey has the following properties which will be defined below. \begin{itemize} \item Gene Vector, which encodes toxicity, evasion rate, etc. \item Signal, which is function of the gene vector \item Mutation rate \item Reproduction rate \end{itemize} \paragraph{Species} Prey divide into species. Every prey of a specific species share common traits such as reproduction rate, mutation rate and genetic constraints. In the paper each species will be represented by a different color (signaling scheme). \paragraph{Gene vector, Signal and Utility to Predator} Every prey has some genetic values, which we can encode as a gene vector. The gene vector is bound by the genetic constraints of the prey species. A natural analogy is the set of chromosomes and the genomic information therein. The allelic values assumed by the set of genes in the gene vector determine a variety of traits -- from how the prey looks to its speed of evasion from the predator. It essentially is analogous to a change in chemical composition which changes the taste, toxicity and nutritional value to the predator. To sum it up, the gene vector determines the signal the prey sends and also the utility a predator receives from consuming the respective prey.\\ We consider a signaling range bounded in the range $[L,R]$. Let $g$ be the signal function that takes gene vector as input and returns signal. Example 1: Consider two prey species $A$ and $B$. Let the gene vector of $A$ be [2,1] and the gene vector of $B$ be [1,2]. Suppose $g([x,y])=x+y$ is the signaling mechanism of both species, then prey both $A$ and $B$ give out the same signal 3. Given a predator $P$ with utility function $f([x,y])=x^2-y$, it can be seen that the utility of $A$ to the predator is 3 while the utility of $B$ to the predator is -2 which implies that $B$ is toxic to the predator $P$. $P$ cannot distinguish between the perfect mimics $A$ and $B$ as it receives the same signal 3. \\ Example 2: Consider two prey species $A$ and $B$. Let the gene vector of $A$ be [2,1] and the gene vector of $B$ be [1,2]. Suppose $g([x,y])=2x+y$ is the signaling mechanism of the species $A$ and $g([x,y])=x+y$ is the signaling mechanism of the species $B$. Then prey of $A$ signal 5 while $B$ signals 3. Given a predator $P$ with utility function $f([x,y])=x^2-y$, it can be seen that the utility of $A$ to the predator is 3 while the utility of $B$ to the predator is $-2$ which implies that $B$ is toxic to the predator $P$. $P$ can clearly distinguish between the two prey and will learn that signal 5 corresponds to a good prey while signal 3 corresponds to bad prey. \\ Since the signal is the only trait visible to the predator, we look specifically at parts of the gene vector affected by the signal. For the sake of simplicity we assume that the signaling function ($g$) to be a injective function. This assumption implies that the gene vector can be inferred from the signal. As the signaling function is injective, we can assume that the utility to the predator which depends on gene vector depends on the signal. We thus use the notation $u_{AP}(s)$ which refers to the utility of prey $A$ with signal $s$ to the predator $P$.\\ When we say constant utility we imply that the mutations in the gene vector result in changes in signal but not in utility for the predator. An example would be that the change in the first gene could make a prey more red in color but the first gene does not affect the toxicity or taste of the prey to predator in any way and thus does not influence utility.\\ If utility to predator is proportional to gene vector which in turn is proportional to signal, then utility will increase with increasing signal. Signal can be appropriately scaled along with mutation rate. \paragraph{Mutation} Every prey species has a natural system that has evolved with some constraints. There is a cost attributed to most changes. Some changes affect reproduction rate, some affect color, some affect speed or toxicity and other evasion mechanisms. This is a result of mutations in the gene vector. Effectively mutations in the gene vector result in new signals and utilities. \\ Mutations can be Gaussian, which implies that the most of the mutations will be concentrated around the point of origin with a very small probability of large mutations. We consider a simplified version where mutations are restricted to the current signal and the neighbouring two signals. Given mutation rate $p$, then the probability of signal $s$ mutating to $s-1$ or $s+1$ is $p$, while the probability of staying at $s$ is $1 - 2p$.\\ For example, if we consider $u(s) = s$ then it means that a mutation in the gene vector will result in a change in $s$ and a proportional change in utility. \paragraph{Signaling distribution} Every prey has a signal. So a prey species will have a signaling distribution which is a collection of signals from all the prey of the species. As the species are naturally selected the signal distribution changes shape. this will be described in detail below. \subsection{Predator} A predator has the following properties. \begin{itemize} \item Prior Knowledge \item Learning system and related parameters \item Signal perception function \item Utility function \item Consumption rate \end{itemize} \paragraph{Species} We also have multiple species of predators each with their own learning systems, utility and perception. \paragraph{Utility} The utility function takes in the gene vector of prey as input and return utility to predator for consuming prey as output. As defined in the Prey section, we use the notation $u_{AP}(s)$ which refers to the utility of prey $A$ with signal $s$ to the predator $P$. \paragraph{Signal from predator perspective} There are multiple prey species, each with their own signal distribution. The predator only sees the combined distribution of signals. An example has been illustrated in Figure ~\ref{fig:predator_view}. \begin{figure}[h] \begin{center} \subfigure[Actual signal distribution where each color corresponds to a separate prey species.]{\includegraphics[scale=0.45]{separate.png}} \subfigure[The signal distribution the predator sees as it cannot distinguish between the two prey species.]{\includegraphics[scale=0.45]{joint.png}} \\ \end{center} \caption{The two plots contrast difference between actual signal distribution of prey(a) vs the the perceived signal distribution of the predator(b). } \label{fig:predator_view} \end{figure} \paragraph{Signal perception function} In higher dimensions, predators see a projection of the signal onto a smaller space. For example given a 2 dimensional signal $s=(s_1,s_2)$ of a prey species $A$, a predator $P$ that sees only the first dimension will not be able to distinguish $(10, 3)$ and $(10, 20)$ but it will be able to distinguish between $(10,3)$ and $(5,3)$. To encode this property we use a signal perception function. A real world example is that off color blind predators. Another example is a species that can has a weak smell and weak eyesight and thus distinguishes a prey as a function of both.\\ So a signal perception function $f(s_1,s_2)=s_1+s_2$ will not be able to distinguish between $(10,20)$ and $(25,5)$. This essentially reduces a 2 dimensional signal to a one dimensional signal. This arrangement is sometimes more efficient due to the cost of evolving more complex senses or a more robust learning system. \paragraph{Consumption Rate} This refers to the number of prey the predator needs to consume at a given day. This can also be easily modified to required amount of utility instead of a required number of prey. \paragraph{Learning and Prior Knowledge} We assume that a new predator with no prior knowledge is introduced into a new environment where it has to learn the utilities from the signal.\\ Due to evolution, the signal distributions must be modeled as non-stationary and utility corresponding to a signal keeps changing. Thus the predator needs to forget over time and thus needs a discounting parameter. Clearly neighboring signals will have similar utilities for simple functions. Thus the predator needs to update its learning from a specific signal to surrounding signals as well. This further helps stabilise the learning for sparse signals. Finally we need a predator that explores for new opportunities but also exploits as is naturally seen. \paragraph{UCB (Upper Confidence Bounded) Bandit} We consider a standard discounted UCB agent with a bandit for every signal in the range $[L,R]$.[5] During update, the bandits of surrounding signals are also updated. The predator chooses the signal $s$ that maximizes the following value $v_s$. \begin{center}$ v_s = \mu_s + \sqrt{\left(\frac{\alpha log(n)}{2n_s}\right)}, \quad \quad s \in [L,R]$\end{center} Where, \begin{itemize} \setlength\itemsep{0em} \item $\mu_s$ is the discounted average utility gotten so far from signal s and its neighbours. \item $\alpha$ is a scaling constant \item $n$ is the total number of prey consumed so far \item $n_s$ corresponds to the number of prey of signal s consumed so far \end{itemize} \section{Signal Drift} As predators start exploring different signals they learn the corresponding utility over time and start exploiting signals of high utility. This evolutionary game uses essentially natural selection (with replicator dynamics) and results in the drift of the signal distribution for the particular prey species. But there are various conditions that result in the magnitude and direction of the drift. Remainder of the paper classifies the major ones. \\ There are two main types of signal drift. Signal drift of a prey species due to a signaling range corresponding to a lower utility. The other one being signal drift due to a signaling range with lower expected utility. Depending on the ecosystem one may dominate over the other producing interesting phenomena. \subsection{Signal Drift to lower prey utility} Consider a prey species $A$ with constant utility to predator $P$, meaning that the utility does not change with signal. For example a prey could change its voice frequency but its utility to the predator might not change at all. This situation will result in the predator sampling the prey uniformly, since it prefers every signal exactly the same. An example of which has been shown in Figure ~\ref{fig:flatten}\\ \begin{figure}[h] \centering \includegraphics[trim={0 0 0 0},clip,width=0.8\textwidth]{flat} \caption{Signal distribution over a 600 generations plotted at different time intervals. The distribution gets flatter and fatter over time. Int this specific example, the rate of consumption by predators is exactly equal to the reproduction rate so the total number of prey at any given generation is approximately 30000.} \label{fig:flatten} \end{figure}\\ We model utility as a function of signal. Given a signal distribution for the prey species, the prey species mutate and slowly change the signaling distribution. The predator will choose signals that induce a higher utility in the prey. This process results in the movement of the signaling distribution in the other direction. An example of such a adrift is shown in Figure ~\ref{fig:right_shift}. We consider a prey with utility proportional to the signal resulting in a left shift (lower end of the signalling range) because the utility on the left will be lower. \\ \begin{figure}[h] \centering \includegraphics[trim={0 0 0 0},clip,width=1\textwidth]{left_drift} \caption{Signal distribution over a 600 generations plotted at different time intervals. We see that the signal distribution shifts left as right has higher utility.} \label{fig:right_shift} \end{figure} An uniform movement in on direction specifically happens in increasing or decreasing functions. If the utility (which is a function of signal) is non monotonic (neither increasing nor decreasing) then the prey population will greedily descend in the direction of lower utility. Figure ~\ref{fig:non_monotonic} illustrates an example with a non monotonic utility function. \\ \begin{figure}[h] \centering \includegraphics[trim={0 0 0 0},clip,width=0.8\textwidth]{split} \caption{Signal distribution over a 600 generations plotted at different time intervals. The utility function is $u(s) = k - (s-30)^2$ see that the signal distribution splits at the point s=30.} \label{fig:non_monotonic} \end{figure} \subsection{Signal drift due to lower expected utility} Consider two prey species $A$, $B$ and a predator $P$.\\ $A$(Blue) :Constant high utility to $P$\\ $B$(Green) :Constant low utility to $P$ Over generations, due to exploiting of prey with highest utility by predator we see a natural shift in the signaling range of the population.\\ \begin{figure}[h] \begin{center} \subfigure[Signal distribution over a 1000 generations plotted at different time intervals.]{\includegraphics[scale=0.35]{drift_exp_util.png}} \subfigure[Population time series.]{\includegraphics[scale=0.3]{drift_exp_util_ts.png}} \\ \end{center} \caption{The plot shows species A(Blue) which is favoured by the predator is predated upon resulting in a right drift of the signal distribution as shown in subplot (a). In this specific example, the total consumption of predator is higher then the growth rate of prey species combined. This results in a decline of total population over time captured by subplot (b).} \label{fig:drift} \end{figure} The predator preys on the signaling range with maximum utility. Thus signals that are produced by only the species with high utility will be the prime targets for the predator. On exploitation the prey species with high utility will naturally evolve to signaling range that is shared with a prey species of lower utility. The predator will learn that the common signaling range and associate it with the expected utility. Over generations, a drift in the signal is achieved (Figure ~\ref{fig:drift}). Under the same notion, a prey species with low utility will drift away from prey species with higher utility. But this drift will be slower than the drift in case of a prey species with high utility. \section{Interesting Phenomena in 1-Dimensional signalling} We consider a one dimensional gene vector for the prey which corresponds to a one dimensional signal. All species can signal on a band from $L$ to $R$. \subsection{Pooling Equilibrium} This phenomena occurs when multiple prey species pool together which results in the predator not being able to distinguish between any of the species. This strategy results in an equilibrium as no species can be naturally selected and thus there is no induced drift.\\ Consider 3 prey species $A$,$B$, $C$ and predator $P$.\\ $A$ (Blue): constant high utility to $P$\\ $B$ (Green): constant low utility to $P$\\ $C$ (Red): constant medium utility to $P$ \begin{center}$u_{BP}(s)<u_{CP}(s)<u_{AP}(s)$\end{center} \begin{center}$\alpha_B=\alpha_C=\alpha_A$\end{center} \begin{figure}[h] \centering \includegraphics[trim={0 0 0 0},clip,width=1\textwidth]{pool} \caption{This results in a stable Pooling equilibrium of the three prey species.} \label{fig:pooling_eq} \end{figure} As seen in Figure ~\ref{fig:pooling_eq} the prey species enter a stable pooling equilibrium. This equilibrium is stable and persists because any change in the signal distribution of a single species will change the distribution of expected utility. This effectively will trigger a drift of all prey species such that the equilibrium is maintained. \subsection{Multi-peak Mimicry} This phenomena is the splitting of a single peaked signal distribution of a prey species to a multi-peak distribution. This dynamics can easily be seen in the case of non monotonic functions. Interestingly it is possible to generate the same phenomena using only simple linear monotonic functions.\\ We consider a situation where prey choose a trade-off between reproduction rate and toxicity. Higher toxicity implies a lower reproduction rate but also a lower utility to the predator. Consider 3 prey species $A$, $B$, $C$ and predator $P$.\\ $A$ (Blue): Toxicity is high, utility to predator $P$ is low and reproduction rate is low\\ $B$ (Green):Toxicity is low, utility to predator $P$ is high and reproduction rate is high \\ $C$ (Red): Toxicity is moderate, utility to predator $P$ is moderate and reproduction rate is moderate \begin{center}$u_{BP}(s)>u_{CP}(s)>u_{AP}(s)$\end{center} \begin{center}$\alpha_B<\alpha_C<\alpha_A$\end{center} In the scenario (Figure ~\ref{fig:multi_peak}), prey species $B$ would have gone extinct. But instead we observe that in the presence of mimicry not only does prey species $B$ thrive at the end but also its signal splits into two peaks. \begin{figure}[h] \begin{center} \subfigure[Signal distribution over the generations.]{\includegraphics[scale=0.45]{poly1.png}} \subfigure[Population time series.]{\includegraphics[scale=0.45]{poly2.png}} \\ \end{center} \caption{Initially prey species B is predated upon due to it's utility. This results in the prey species B splitting it's signal and drifting into two separate peaks one mimicking species A and the other peak mimicking species B. Subplot (a) illustrates the changes in the signal distribution while subplot (b) shows the corresponding changes in population of the prey species.} \label{fig:multi_peak} \end{figure} \subsection{Mimicry can be bad to both species} We show a situation where the mimicry of $A$ and $B$ results in mutual extinction. In the control ecosystem where the other does not exist, they actually thrive.\\ Consider 3 prey species $A$, $B$, $C$ and predator $P$.\\ $A$ (Blue): Utility to predator $P$ is negatively correlated to signal : $(u_{AP}(s) = a_0 - a_1s) $ \\ $B$ (Green): Utility to predator is constant and rate of mutation is insignificant : $(u_{BP}(s) = b_0) $\\ $C$ (Red): Utility to predator is correlated to signal : $(u_{CP}(s) = c_0 + c_1s) $\\ Since utility of $A$ is negatively correlated with signal it drifts towards the right ($R$) because the predator exploits $A$ of lower signal. Since utility of $C$ is correlated with signal it drifts towards the left ($L$) because the predator selects species $C$ of higher signal. Since, utility of $B$ is constant there is no drift but instead just a flattening of the signal distribution. \paragraph{Case 1: An ecosystem with only prey species $A$ and $B$:} $A$ is naturally selected towards a signal of lower utility and thus drifts right ($R$). Initially the utility of $A$ is higher and thus the corresponding signals are exploited but as $A$ drifts right its utility decreases till a tipping point where its utility equals the utility of the agents of species $B$. As a result we see an initial drop in population for $A$ followed by recovery which is shown in Figure ~\ref{fig:bad1}. The turning point is when a critical chunk of species $A$ drifts past to signal higher than $s = \frac{a_0-b_0}{a_1}$. Thus effectively resulting in a lower utility for the predator. A rational predator will make the choice to choose signals corresponding to $B$ compared to $A$. \\ \begin{figure}[h] \centering \includegraphics[trim={0 0 0 0},clip,width=0.7\textwidth]{bad1} \caption{Right drift in signal distribution for species A(Blue) according to the conditions presented in Case 1.} \label{fig:bad1} \end{figure} \paragraph{Case 2: An ecosystem with only prey species $B$ and $C$:} This is very similar to the previous case. $C$ is naturally selected towards a signal of lower utility and thus drifts left ($L$). Initially the utility of $C$ is higher and thus the corresponding signals are exploited but as $C$ drifts left its utility decreases till a tipping point where its utility equals the utility of the agents of species $B$. As a result we see an initial drop in population for $C$ followed by recovery as shown in ~\ref{fig:bad2}.The turning point is when a critical chunk of species $A$ drifts past to signal lower than $s = \frac{c_0-b_0}{c_1}$. \\ \begin{figure}[h] \centering \includegraphics[trim={0 0 0 0},clip,width=0.7\textwidth]{bad2} \caption{Left drift in signal distribution for species C(Red) according to the conditions presented in Case 2.} \label{fig:bad2} \end{figure}\\ \paragraph{Case 3: An ecosystem with all prey species $A$,$B$,$C$:} As a result of the pooling equilibrium formed by the signals of $A$ and $C$. The predator $P$ cannot distinguish between the signals of $A$ and $C$ and thus cannot naturally select them.\\ At the point of pooling, if the expected utility of the signals in the pooling equilibrium is lesser than the utility of the signals of prey species $B$. This leads to exploitation and thus eventually extinction.\\ \begin{figure}[h] \centering \includegraphics[trim={0 0 0 0},clip,width=0.5\textwidth]{bad3.png} \caption{In a combined setting of all 3 prey species A,B and C. We see that the drift of species A and C result in a pooling equilibrium resulting in the extinction of both species. We considered $u_{AP}(s) = 3 - 0.05s, u_{BP}(s) =1.3, u_C(S)=0.05s$ and initial distribution means at points 25 for A and 35 for C. $s=30$ will be the mean of the pooling equilibrium between $A$ and $C$ after drifting. This dynamics is because initially both $A$ and $C$ have the same utility and are thus selected together The expected utility of any signal in this pooling equilibrium will be 3/2. This is greater than the utility of prey $B$ and thus the pooling equilibrium will be exploited till extinction.} \label{fig:bad3} \end{figure} Figure ~\ref{fig:bad3} represents Case 3 which is a combination of both Case 1 and 2. Although a specific example, it can be easily generalized to show that a pooling equilibrium can form. Thus resulting in the stopping of natural selection towards a lower utility, effectively leading to extinction. This extinction is mainly because of the lack of freedom in the 1-Dimensional signaling space. \subsection{Oscillating Mimicry} Oscillating mimicry where signals of two species is constantly oscillating, is not possible when we consider a one dimensional gene vector, identity signal function ($g$) and a fixed predator population. This situation is due to the formation of a pooling equilibrium.\\ \section{Interesting Phenomena in 2-Dimensional signalling} 2-Dimensional signaling opens up a whole new world phenomena. In 1-Dimensional signaling due to a lack of freedom we are restricted to certain type of phenomena. here we present two basic ideas of how far we can go when we increase the signalling dimension. \subsection{Drift in 2-Dimensions} Drift in higher dimensions is similar to 1-dimensional case except with more freedom.The drift is determined by the signal perception function, utility function of predator and genetic vector of prey.\\ Consider a prey species $A$ and a predator $P$. $A$ has a genetic constraint that it cannot reproduce if its genes are too different from the rest. This constraint stops the the signaling distribution from spreading out in all directions. The predator $P$ can see signals only along the line $s_1 - s_2 = k$ because of the signal perception function $f(s_1,s_2) = s_1+s_2$. Furthermore a utility function $u_{AP}(s_1,s_2) =s_1+s_2$ results in a diagonally left-down drift for the prey species $A$. The resulting drift of a 2-dimensional signal distribution has been illustrated in Figure ~\ref{fig:2D} \begin{figure}[h] \centering \includegraphics[trim={0 0 0 0},clip,width=0.7\textwidth]{2d_drift.png} \caption{Drift in the 2-dimensional signal distribution of a prey species perpendicular to the predator's signal perception function.} \label{fig:2D} \end{figure} \subsection{Switching directions} In the presence of two prey species and two predator species, we can create very interesting phenomena which involve changes in drift direction. Such directional change is not possible in 1-Dimensional signaling due to the creation of a Pooling Equilibrium.\\ Consider two prey species $A$, $B$ and two predator species $P$ and $Q$. Let the predators and prey have the following functions. \begin{itemize} \item $P$ signal perception function : $f_P(s_1,s_2)=s_1$. This subcase means that $P$ can distinguish only the $x$ coordinate of the signal. \item $Q$ signal perception function : $f_Q(s_1,s_2)=s_2$. This subcase means that $Q$ can only distinguish the $y$ coordinate of the signal. \item The signaling function of $A$ as a function of the gene vector : $g_A(v_1,v_2)=(v_2,-v_1)$ \item The signaling function of $B$ as a function of the gene vector : $g_B(v_1,v_2)=(-v_1,-v_2)$ \item The utility function of $P$ as a function of gene vector : $u(v_1,v_2) \propto v_1+v_2$ \item The utility function of $Q$ as a function of gene vector : $u(v_1,v_2) \propto v_1-v_2$ \end{itemize} The above functions result in the following utility functions. \begin{itemize} \item $u_{AP}(s_1,s_2)\propto s_1-s_2$ \item $u_{AQ}(s_1,s_2)\propto -s_1-s_2$ \item $u_{BP}(s_1,s_2)\propto -s_1-s_2$ \item $u_{BQ}(s_1,s_2)\propto -s_1+s_2$ \end{itemize} Starting at the appropriate positions with correct scaling parameters such that species $A$ has all the $x$ coordinates of its signal less then $B$ but $y$ coordinates of its signal greater than $B$. If initially $u_{AP} > u_{BP}$ and $u_{BQ} > u_{AQ}$, then $A$ will drift left but to predation by $P$ and $B$ will drift down due to predation by $Q$. The drift results in a mean x coordinate decrease for species A and a mean y coordinate decrease in signal for species B. As shown by the utility function the drift in species A results in decrease of $u_{AP}$ but an increase in $u_{AQ}$. Similarly, the drift in species B results in decrease of $u_{BQ}$ but an increase in $u_{BP}$. Thus after a point when $u_{AP} = u_{BP}$ and $u_{BQ} = u_{AQ}$, the predators start gaining interest in the other prey which they were not predating originally. Slowly the predators reduce the predating of their original prey and begin predating the other prey. This will result in $A$ drifting upward due to predation by $Q$ and $B$ drifting right due to predation by $P$.\\ So $A$ initially starts drifting left, followed by a gradual change in direction resulting in an upward drift. Similarly $B$ initially drifts downward and gradually shifts to drifting right. \section{Concluding Remarks} \label{sec:conclusion} In most of the phenomena we have focused on using only linear functions. Although this may not be the case, it only shows that a simple set of linear rules can result in complex mechanisms like John Conway's ``The Game of Life.'' If we introduce higher degree or more complex functions we will be able to construct even more fascinating phenomena for bio-mimicry. Furthermore if we increase the dimension of signaling and the nature of genetic constraints, we can explore an even wider variety of counter-intuitive results.\\ This paper gives us a broad framework to understanding the existence of multiple mathematically feasible mimicry based phenomena. This can easily be extended to the dynamics between multiple agents that can systematically be described with learning predators and evolving prey.\\ This framework can be used is in understanding the introduction of Artificial intelligence or learning systems into new ecosystems. An example is the case of malicious bots getting better over time while security systems learn to deal with them. Apart from this, multiple other systems can effectively be modeled using this framework. By understanding these dynamics better we can potentially stop certain types of behaviour if required, by introducing minimal changes. \section{References} \begin{enumerate} \item Rowland HM, Fulford AJT, Ruxton GD. 2017. Predator learning differences affect the survival of chemically defended prey. Animal Behaviour. 124:65–74.\\ doi:10.1016/j.anbehav.2016.11.029. \\ \item Skelhorn J, Halpin CG, Rowe C. 2016. Learning about aposematic prey. Behavioral Ecology. 27(4):955–964. doi:10.1093/beheco/arw009. \\ \item Müller, Fritz (1879). "Ituna and Thyridia; a remarkable case of mimicry in butterflies. (R. Meldola translation)". Proclamations of the Entomological Society of London. 1879: 20–29.\\ \item Enaganti I, Mishra B. 2021 Mar 25. Lotka-Volterra Equations in the Presence of Mimicry. doi:10.1101/2021.03.25.436931. \\ \item Garivier A, Moulines E. 2008 May 22. On Upper-Confidence Bound Policies for Non-Stationary Bandit Problems. arXiv:08053415 [math, stat]. [accessed 2021 Apr 17].\\ https://arxiv.org/abs/0805.341\\ \end{enumerate} \end{document}

bb8770a3dd26bfcca99bf1c3550499db84ac278f

\section{Introduction} The bounded derived category of coherent sheaves $\D^b(X)$ on a projective variety has received a large amount of attention in the last 30 years or so. As in many areas of mathematics, it is often convenient and enlightening to ``decompose" the derived category into pieces. To be precise, for a projective variety $X$, we say that the derived category $\D^b(X)$ admits a semiorthogonal decomposition into two full triangulated subcategories $\A$ and $\B$ if there are no nonzero morphisms from objects of $\B$ to objects of $\A$, and the smallest full triangulated subcategory containing $\A$ and $\B$ is all of $\D^b(X)$. Of course, this definition also makes sense for any triangulated category $\T$ in place of $\D^b(X)$. If any semiorthogonal decomposition of a triangulated category $\T$ is trivial, that is, one of the components is zero, then we say that $\T$ is indecomposable. In the geometric setting, semiorthogonal decompositions are more then just convenient tools for the study of $\D^b(X)$ and $\perf X$. For example there are a number of interesting conjectures relating the existence of certain semiorthogonal decompositions to the rationality of $X$ (see \cite{Kuznetsov-rationality} for a somewhat recent survey). Despite the interest however, semiorthogonal decompositions can be hard to produce without inspiration, however an interesting special case is when the derived category is indecomposable. This occurs, for example, for smooth varieties $X$ with $\omega_X \cong \O_X$, or for smooth curves of positive genus \cite{Okawa-semiorthogonal_decomposability_of_curves}. The most general indecomposability criteria to date was proven in \cite[Theorem 3.1]{KawataniOkawa-Nonexistence}. There, the authors show that any smooth projective variety (more generally, any smooth and proper DM stack with a non-stacky point) with small enough canonical base locus has an indecomposable derived category. In this note we observe that these techniques generalize, with suitable assumptions, to the case of Cohen-Macaulay projective varieties. To be more specific, let $X$ be a Cohen-Macaulay projective variety and $x\in X$ any closed point. The key notion in the proof is that of a \emph{Koszul zero-cycle}, which is a closed subscheme $Z_x$ of $X$ whose underlying topological space is just the closed point $\{x\}$, but whose structure sheaf $\O_{Z_x}$ is perfect (that is, a bounded complex of locally free sheaves of finite rank) as an object of $\D^b(X)$, and whose support is exactly $\{x\}$. These skycraper sheaves form a spanning class for both $\perf X$ and $\D^b(X)$, and are a replacement for the sheaves $k(x)$, which are not perfect if $x\in X$ is a singular point. Our main result is an application of these objects, and generalizes \cite[Theorem 3.1]{KawataniOkawa-Nonexistence}. \begin{theorem}\label{Main Theorem} Let $X$ be a Cohen-Macaulay projective variety with dualizing sheaf $\omega_X$. Suppose that $\perf X = \langle \A, \B \rangle$ is a semiorthogonal decomposition. Then: \begin{enumerate} \item for any closed point $x \in X \setminus \operatorname{Bs}|\omega_X|$, and all Kozsul zero-cycles $Z_x$ at $x$, exactly one of the following holds: \begin{enumerate} \item $\O_{Z_x} \in \A$, or \item $\O_{Z_x} \in \B$. \end{enumerate} \item When (1a) (resp. (1b)) is satisfied, then the support of any object in $\B$ (resp. $\A$) is contained in $\operatorname{Bs}|\omega_X|$. \end{enumerate} \end{theorem} Along the way, we make sense of what the base locus of $\omega_X$ is, given that in the Cohen-Macaulay setting, it is not a line bundle. Combining this result with some auxiliary statements regarding the support of semiorthogonal components, we are then able to apply this to curves, yielding the following. \begin{corollary} Let $C$ be a projective curve of arithmetic genus at least one. Then $\perf C$ is indecomposable and $\D^b(C)$ is weakly indecomposable. \end{corollary} \noindent Here weakly indecomposable refers to the nonexistence of semiorthogonal decompositions where the components are admissible (see Definition \ref{def_admissible}). \subsection{Acknowledgements} The author wishes to thank Valery Lunts for his support, encouragement, and advice. The author also wishes to thank A. Thomas Yerger for helpful discussions. \subsection{Conventions} Throughout we work over a fixed algebraically closed field $k$. In addition, unless explicitly stated otherwise, any scheme or variety $X$ is assumed to be a projective variety; in particular for us this means an integral separated scheme of finite type over $\spec k$, such that $X$ admits a closed immersion into some projective space. A curve refers to a projective variety of dimension one. By $\D^*(X)$ we always mean the derived category of complexes of coherent sheaves respectively, with appropriate decoration $* \in \{\phantom{0},-,+,b\}$ meaning unbounded, bounded above, below, and bounded respectively. In particular, we identify $\D^b(X)$ with the full subcategory of $\D(\qcoh X)$ consisting of complexes of quasi-coherent sheaves with bounded coherent cohomology sheaves. By $\perf X$ we mean the full triangulated subcategory of $\D(\qcoh X)$ consisting of complexes which are isomorphic (in $\D(\qcoh X)$) to a bounded complex of vector bundles. \section{Preliminaries} \subsection{Base loci for coherent sheaves} To formulate our results, we need to define what we mean by the base locus of a sheaf. Let $X = \cup_j Y_j$ be a reduced quasi-projective scheme of finite type over $k$ (that is, $X$ is the finite union of the quasi-projective varieties $Y_j$), $\mathcal{F}$ a coherent sheaf on $X$, and $s \in H^0(X,\mathcal{F})$ a global section. We define the set $$D(s) = \{x\in X \text{ closed}\, | \, 0 \neq s(x) \in \mathcal{F}(x) \},$$ where $\mathcal{F}(x)$ is the fiber of $\mathcal{F}$. We denote the complement as $V(s) = D(s)^c$, which we refer to as the zero set of $s$. Suppose for the sake of illustration that $X$ was integral (so that by our conventions it is a quasi-projective variety), and that $\mathcal{F}$ is locally free. Then the definition of $V(s)$ and $D(s)$ obviously agree with the corresponding definitions for the zero set (respectively, its complement) of a section of a vector bundle. In this setting, so long as the section is nonzero, the sets are closed (respectively, open), but if $\F$ is not a vector bundle, this is no longer true. Instead, these sets are constructible. Recall that a locally closed subset of $X$ is the intersection of an open and closed subset. A set is said to be \emph{constructible in $X$} if it is a finite union of locally closed subsets of $X$. Constructible sets are closed under finite unions, finite intersections, and complements. \begin{lemma}\label{lem_constructible} Let $X = \cup_j Y_j$ be a reduced quasi-projective scheme of finite type over $k$, $\mathcal{F}$ a coherent sheaf on $X$, and $s \in H^0(X,\mathcal{F})$ a global section. Then $D(s)$ is a constructible subset of $X$. \end{lemma} \begin{proof} We proceed by induction on the dimension of $X$. Clearly, if $X$ is a finite number of points, then the set $D(s)$ is a finite disjoint union of points (possibly empty), and so is constructible. Now if $\dim X \geq 0$, let $Y \subset X$ be an irreducible component of positive dimension. Then via pullback along $Y \hookrightarrow X$, we assume for now that $X$ is integral and $\mathcal{F}$ is a coherent sheaf on $X$. By well-known facts, there is an open dense subset $j:U\hookrightarrow X$ such that $\mathcal{F}$ is locally free of non-negative rank. Denoting $s|_U := j^* s$, we see $D(s|_U) = D(s) \cap U$ is constructible in $X$. Indeed, $D(s|_U)$ is open in $U$, since $j^*\mathcal{F}$ is locally free, and thus $U \setminus D(s|_U)$ is closed in $U$, and so is of the form $V \cap U$ for some closed subset $V$ of $X$, and so $(X \setminus V) \cap U = D(s|_U)$ is an open subset of $X$, hence constructible. Now, still assuming $X$ is integral, the closed subset $X \setminus U = Z$ is a finite union of subvarieties (after possibly giving them their natural reduced subscheme structure) of lower dimension. Let $i:Z \hookrightarrow X$ be the closed immersion and set $s|_{Z} : = i^* s$ as before. Then by the induction hypothesis, the set $D(s|_{Z}) = D(s) \cap Z$ is constructible and we see that $D(s) = D(s) \cap U \coprod D(s) \cap Z$ is constructible. Now if $X = \cup_j Y_j$ is not integral with finitely many irreducible components $Y_j$, $D(s)$ is the union of its intersections with each $Y_j$, which we have shown above are each constructible. Using that constructible sets are closed under finite unions we have that $D(s)$ is constructible. \end{proof} \begin{remark}\label{rem_denseopen} Let $X$ be as above, $\mathcal{F}$ a torsion-free coherent sheaf on $X$, $s$ a global section of $\mathcal{F}$, and $Y \subset X$ is an irreducible component where $s|_Y$ is nonzero (so that $D(s|_Y)$ is not empty), then $D(s)$ is dense in $Y$. This follows as on $Y$, there is a dense open subset $U\subset Y$ where $\mathcal{F}$ is locally free, and by elementary topology, any dense subset of $U$ will also be dense in $Y$. Since $D(s|_U)$ is dense in $U$, this proves the claim. Further, since $D(s|_Y)$ is constructible and dense for $s$ nonzero, it will contain a dense open subset of $Y$, as given any topological space, a constructible subset always contains a dense open subset of its closure (see for example \cite[Lemma 2.1]{An-Rigid_geometric}). \end{remark} With the above in mind, we make the following definition. \begin{definition} Let $X$ be as in Lemma \ref{lem_constructible}. Given a coherent sheaf $\mathcal{F}$ on $X$, we define the base locus of $\mathcal{F}$ to be $$\operatorname{Bs}|\mathcal{F}| = \bigcap_{s\in H^0(X,\mathcal{F})} \overline{V(s)}.$$ The sheaf $\mathcal{F}$ is then base point free if this intersection is empty, or equivalently, if the interiors of the sets $D(s)$, ranging over $H^0(X,\mathcal{F})$, are an open cover of $X$. \end{definition} \noindent Note that by definition, the base locus of a sheaf is always a closed subset of $X$. In the case of curves however, things simplify nicely. \begin{lemma} Let $\F$ be a torsion-free sheaf on an curve $C$. Then $D(s)$ is open for any global section $s\in H^0(X,\mathcal{F})$. \end{lemma} \begin{proof} This follows from the fact that any constructible subset of a one dimensional variety is either open or closed (and hence a finite number of points), and if $D(s)$ is nonempty, it contains a dense open subset by Remark \ref{rem_denseopen}. If $D(s)$ is empty then we are done. \end{proof} \subsection{Cohen-Macaulay schemes} In the following we will only need the most basic facts about Cohen-Macaulay rings, a more thorough treatment can be found in \cite{BrunsHerzog-CM_rings}. Let $R$ be a commutative noetherian local ring with maximal ideal $\mathfrak{m}$, and let $M$ be a finitely generated $R$-module. A sequence of elements $(x_1,...,x_n) \subset \mathfrak{m}$ in a noetherian local ring $R$ is called a $M$-regular sequence if $x_i$ is a non-zero divisor on $M/(x_1,...,x_{i-1})M$ for all $i \leq n$ and $M/(x_1,...,x_n)M \neq 0$. The length of any maximal regular sequence is constant and is called the depth of $M$, denoted by $\operatorname{depth}M$. The depth should be thought of as an algebraic notion of dimension; in general the depth and dimension do not agree, but they are be related via $\operatorname{depth} M \leq \dim M$, where the dimension of a module is the dimension of its support as a sheaf on $\spec R$. Further, if $M$ has finite projective dimension over $R$, there is a close relationship between the depth of $R$ (as a module over itself) and the projective dimension of of $M$; $$\operatorname{proj.dim.} M + \operatorname{depth} M = \operatorname{depth} R,$$ known as the Auslander-Buchsbaum formula \cite[Theorem 1.3.3]{BrunsHerzog-CM_rings}. Finally, a module is said to be a Cohen-Macaulay module if $\operatorname{depth} M = \dim M$, and a ring is Cohen-Macaulay if it is so as a module over itself. We can globalize this, namely a locally noetherian scheme $X$ is said to be a Cohen-Macaulay scheme if $\O_{X,x}$ is a Cohen-Macaulay ring for all $x\in X$. Our primary example is when $R$ is a reduced local ring of dimension one, where such rings are automatically Cohen-Macaulay (this is \cite[Exercise 2.1.20(a)]{BrunsHerzog-CM_rings}). Since the definition of a Cohen-Macaulay scheme is local, it follows that any reduced curve is Cohen-Macaulay. Cohen-Macaulay varieties can be singular, and any regular or Gorenstein variety is automatically Cohen-Macaulay. Suppose we are given a noetherian local ring $R$ which is Cohen-Macaulay. The ideal generated by a maximal regular sequence $\mathbf{x} = (x_1,...,x_n)$, $n =\dim R$, satisfies $\dim R/ (x_1,...,x_n) =0$ and the associated Koszul resolution $K(\mathbf{x})$ \cite[Section 1.6]{BrunsHerzog-CM_rings} gives a finite free resolution of the quotient. Further, in $\D^b(R)$, the bounded derived category of finitely-generated $R$-modules, the Koszul resolution satisfies $K(\mathbf{x})^\vee \cong K(\mathbf{x})[-\dim R]$ \cite[Proposition 1.6.10]{BrunsHerzog-CM_rings}. Now we give a definition orginally given in \cite[Section 1.3]{Ruiperez-Fourier-Mukai_transforms_for_Gorenstein_schemes}. \begin{definition} Let $X$ be a Cohen-Macaulay scheme. Given any closed point $x\in X$, the local ring $\O_{X,x}$ is a Cohen-Macaulay local ring, and hence has a regular sequence of length $\dim X$. A Koszul zero-cycle at $x$ is the closed subscheme $Z_x$ defined by the quotient of $\O_{X,x}$ by the ideal $\mathcal{I}_x$ generated by this regular sequence. \end{definition} Clearly the structure sheaf $\O_{Z_x}$ of a Koszul zero-cycle $Z_x$ is a coherent sheaf supported only at the closed point $x\in X$, and is an object of $\perf X$ via the Koszul resolution. By the discussion above in the local case, we have that $\R\shom_{\O_X} (\O_{Z_x},\O_X) := \O_{Z_x}^\vee \cong \O_{Z_x}[-\dim X]$ in $\D^b(X)$. Since we mainly have singular varieties in mind for applications, it is worth noting that the Koszul zero-cycles are a suitable replacement for the structure sheaves of closed points, for the main reason that they form a spanning set for both $\perf X$ and $\D^b(X)$. Recall that a collection of objects $\Omega$ in a triangulated category $\mathcal{T}$ is said to be spanning if for all $T\in \mathcal{T}$, $\hom(T,A[i])=0$ for all $i \in \Z$ and $A \in \Omega$ implies $T=0$, and $\hom(A,T[i])=0$ for all $i \in \Z$ and $A \in \Omega$ implies $T=0$. Before we prove that these objects are a spanning set, we include the following elementary lemma for the reader's convenience. \begin{lemma}\label{lemma_local maps} Let $(R,\mathfrak{m},k)$ be a local noetherian ring and $M$ a finitely generated module. If $\mathfrak{m} \in \operatorname{supp} M$, $M$ admits a surjection $M \to k$. Further if $\{\mathfrak{m}\} = \operatorname{supp}M$, then $M$ also admits an injection $k \to M$. \end{lemma} \begin{proof} If $M$ is finitely generated and supported at the maximal ideal, then Nakayama's lemma implies that $M/\mathfrak{m}M \neq 0$, and further is a finite dimensional $k$-vector space. Choosing any further projection to a one-dimensional subspace, we have a surjection $M \to k$. Consider a prime ideal $\mathfrak{p} \subset R$. Recall that $\mathfrak{p}$ is said to be associated to $M$ if there exists an element $m \in M$ such that $\mathfrak{p} = \operatorname{Ann}_R(m)$. Equivalently, if and only if $R/\mathfrak{p}$ is isomorphic to a submodule of $M$ . Further, since $R$ is noetherian, standard results in commutative algebra (\cite[\href{https://stacks.math.columbia.edu/tag/0586}{Lemma 0586},\href{https://stacks.math.columbia.edu/tag/0587}{Lemma 0587}]{stacks-project}) imply that the set of associated primes is always nonempty and is a subset of the support of $M$. Now if $\{\mathfrak{m}\} = \operatorname{supp}M$, it follows that $\mathfrak{m}$ is an associated prime to $M$, and hence there is a submodule of $M$ which is isomorphic to $R/m \cong k$. The inclusion of submodules $k \to M$ yields the claim. \end{proof} For the following lemma, recall that the support of a complex in $\D^b(X)$ is by definition the union of the supports of its cohomology sheaves. It is a closed subset of $X$, possibly non-reduced. We will always work with the induced reduced subscheme structure when necessary. \begin{lemma}\label{lem_detect_support} Let $X$ be a Cohen-Macaulay quasi-projective variety, $x\in X$ be a closed point, $Z_x$ a Koszul zero cycle supported at $x$, and $Q \in \D^b(X)$. Then $x\in \operatorname{supp} Q$ if and only if there is some $i\in \Z$ such that $\hom(Q,\O_{Z_x}[i]) \neq 0$ if and only if there is some $j \in \Z$ such that $\hom(\O_{Z_x},Q[j]) \neq 0$. \end{lemma} \begin{proof} Since $X$ is assumed Cohen-Macaulay, for any closed point there is a Koszul zero-cycle supported on it. Suppose a closed point $x\in X$ belongs to the support of $Q$ and denote the cohomology sheaves of $Q$ by $\mathcal{H}^i$. We consider the spectral sequence \begin{equation}\label{eqn_hyperext1} E^{p,q}_2 = \hom_{\D^b(X)}(\mathcal{H}^{-q},\O_{Z_{x}}[p]) \implies \hom_{\D^b(X)} (Q , \O_{Z_x}[p+q]). \end{equation} Since $x$ belongs to the support of $Q$, it must also belong to the support of at least one cohomology sheaf, say $\mathcal{H}^{q}$. Now using that structure sheaf of a Koszul zero-cycle is supported at a single closed point: $$\hom_{\O_{X,x}}(\mathcal{H}^{q}_{x},\O_{Z_{x}}) \cong \hom_{\O_X}(\mathcal{H}^{q},\O_{Z_{x}}).$$ To show that these spaces are nonzero it is clearly enough to see that the left-hand side is nonzero. In particular the stalk $\mathcal{H}^q_x$ is a finitely generated $\O_{X,x}$-module which is supported on the maximal ideal $\mathfrak{m}_x$ by assumption, and $\O_{Z_x}$ is a finitely generated $\O_{X,x}$ module whose support is precisely the maximal ideal. Hence by Lemma \ref{lemma_local maps} there is a nonzero morphism $$\mathcal{H}^q_x \to k(x) \to \O_{Z_x},$$ which shows that the morphism space is nonzero. So now taking the maximal $q_0$ such that $\mathcal{H}^{q_0}_x \neq 0$ at $x$, we see that the term $E^{0,-q_0}_2$ survives to the $E_\infty$ page as there are no negative Ext groups between two sheaves, so the differential ending at $E^{0,-q_0}_2$ starts from zero, and the differential leaving $E^{0,-q_0}_2$ hits terms which are zero simply because we had chosen $q_0$ maximal, and all Ext groups between sheaves with disjoint support are trivial (using the local-to-global spectral sequence). Hence this term survives, and since it is nonzero, we see $\hom_{\D^b(X)}(Q,\O_{Z_x}[-q_0]) \neq 0$. Conversely, suppose that $\hom_{\D^b(X)}(Q,\O_{Z_x}[m]) \neq 0$ for some $m \in \Z$. Suppose by way of contradiction that $x \notin \operatorname{supp} Q$, equivalently, $x \notin \operatorname{supp}\mathcal{H}^i$ for all $i \in \Z$. But since sheaves with disjoint support have no nontrivial Ext groups by the local-to-global spectral sequence, the spectral sequence \ref{eqn_hyperext1} yields a contradiction. This shows the first equivalence. To check the latter equivalence, we first observe as above that if the support of $Q$ and $\O_{Z_x}$ are disjoint, then the local-to-global spectral sequence shows that $\hom(\O_{Z_x},Q[j]) = 0$ for all $j\in \Z$. Conversely, assume that $x\in \operatorname{supp}Q$, and now we wish to use the spectral sequence $$E^{p,q}_2 = \hom(\O_{Z_x},\mathcal{H}^q[p]) \implies \hom(\O_{Z_x},E[p+q])$$ where we as above denote $\mathcal{H}^i(Q)$ by just $\mathcal{H}^i$. Since $Q \in \D^b(X)$, we have that $E^{p,q}_2 = 0$ for $|q|>>0$. Noting that $\O_{Z_x}$ is defined via a Koszul resolution, we see that $E^{p,q}_2 =0$ for $p \notin [0,\dim X]$. So now let $q_0$ be the maximal integer such that $\mathcal{H}^{q_0}$ is supported at $x\in X$. Note that for all $q>q_0$, since the supports are disjoint, $E^{p,q}_2 =0$. We claim that $E^{\dim X,q_0} \neq 0$. If so, it is clear that this term survives the spectral sequence and hence proves the proposition. To prove the claim, consider the following chain of isomorphisms: \begin{align*} E^{\dim X,q_0} & \cong \ext^{\dim X}_{\O_X} (\O_{Z_x},\mathcal{H}^{q_0}) \\ & \cong \sext^{\dim X}_{\O_X} (\O_{Z_x},\mathcal{H}^{q_0}) \\ & = \mathcal{H}^{\dim X}(\R \shom_{\O_X}(\O_{Z_x},\mathcal{H}^{q_0})) \\ & \cong \mathcal{H}^{\dim X} (\O_{Z_x} \tens{\L} \mathcal{H}^{q_0}[-\dim X]) \\ & \cong \mathcal{H}^{0} (\O_{Z_x} \tens{\L} \mathcal{H}^{q_0}) \\ & \cong \O_{Z_x} \otimes \mathcal{H}^{q_0} \end{align*} where the second isomorphism follows from the local-to-global spectral sequence and the fourth follows as since $\O_{Z_x}$ is a Koszul zero cycle, $\O_{Z_x}^\vee \cong \O_{Z_x} [-\dim X]$. In the last line, since both are nonzero finitely generated modules over the local ring $\O_{X,x}$, this tensor product is nonzero, hence the claim. \end{proof} We also have the following result using very similar techniques. See \cite[Proposition A.91]{Bartocci-FM_in_geometry_and_physics}. \begin{lemma}\label{lem_detect support2} Let $X$ be a quasi-projective variety, $x\in X$ be a closed point, and $E \in \D^b(X)$. Then $x\in \operatorname{supp} E$ if and only if there is some $i\in \Z$ such that $\hom(E,k(x)[i]) \neq 0$. \end{lemma} \begin{remark}\label{remark_indecomposable} As a final remark on the Koszul zero cycles, note that by their construction, they are local $k$-algebras, and hence indecomposable. That is, whenever $\O_{Z_x} \cong A \oplus B$ with $A,B \in \D^b(X)$, we must have that one of $A$ or $B$ is zero. \end{remark} \subsection{Rouquier functors} All results in this section are contained in \cite[Section 4]{Rouquier-Dimensions_of_triangulated_categories} and \cite[Section 5]{Ballard-Derived_categories_of_singular_schemes_and_reconstruction}. Given a smooth projective variety $X$, the bounded derived category $\D^b(X)$ possesses a distinguished autoequivalence given by $$S_X(-) = (-) \tens{} \omega_X[\dim X] : \D^b(X) \to \D^b(X).$$ This functor captures the notion of Serre duality, and as such is called a Serre functor. More generally given a $k$-linear triangulated category $\T$, a Serre functor is a triangulated autoequivalence $S:\T \to \T$ for which there are natural isomorphisms $$\eta_{A,B}:\hom_{\T} (B,S(A)) \to \hom_{\T}(A,B)^*,$$ where ``$*$" indicates the vector space dual. When $X$ is singular, $\D^b(X)$ no longer possesses a Serre functor. When $X$ is Gorenstein however, we can instead work with $\perf X$. \begin{proposition}[\cite{Ballard-Derived_categories_of_singular_schemes_and_reconstruction, SalasSalas-Reconstruction_from_the_derived_category}] If $X$ is a proper variety over a field $k$, the category $\perf X$ has a Serre functor if and only if $X$ is Gorenstein. In particular, the functor has the formula: $S_X(-) = (-) \tens{} \omega_X [\dim X]$. \end{proposition} The failure of $\perf X$ to have a Serre functor in the non-Gorenstein case is essentially a failure of existence, as the functors $\hom(A,-)$ are no longer represented by objects in $\perf X$. However $\hom(A,-)$ \emph{can} be represented by objects in $\D^b(X)$. This observation is due to Rouquier \cite[Section 4]{Rouquier-Dimensions_of_triangulated_categories}, where the following more general definition was given. \begin{definition} Let $\mathcal{S}$ and $\T$ be arbitrary $k$-linear triangulated categories, and $F:\mathcal{S} \to \T$ be a $k$-linear triangulated functor. A Rouquier functor associated to $F$ is a $k$-linear triangulated functor $R_F : \mathcal{S} \to \T$ for which there are natural isomorphisms $$\eta_{A,B}: \hom_{\T} (B,R_F(A)) \to \hom_{\T}(F(A),B)^*.$$ \end{definition} Our main example is the case of projective schemes. For the following, see Proposition 7.47 and Remark 7.48 in \cite{Rouquier-Dimensions_of_triangulated_categories}, see also \cite[Example 5.12]{Ballard-Derived_categories_of_singular_schemes_and_reconstruction} for more details. \begin{lemma}\label{lemma_Rouquier functor is dualizing complex} Let $f:X \to \spec k$ be a projective scheme over the field and $\iota:\perf X \hookrightarrow D(\qcoh X)$ the inclusion functor. Then the Rouquier functor $R_X := R_\iota$ exists and is isomorphic to $$(-) \tens{\L} f^! \O_{\spec k}.$$ Furthermore, it is a fully faithful functor mapping $\perf X$ into $\D^b(X)$. \end{lemma} In our situation, where the projective scheme $X$ is a Cohen-Macaulay variety, the Rouquier functor becomes $$R_X(-) = (-) \tens{\L} \omega_X [\dim X],$$ where $\omega_X$ is the dualizing sheaf in the sense of \cite{Hartshorne-AG}. It is well-known that for a Cohen-Macaulay variety, the dualizing sheaf is a divisorial sheaf, that is, reflexive (in particular, torsion-free) and generically rank 1. When the variety is Gorenstein, the dualizing sheaf is actually a line bundle (\cite[Proposition V.9.3]{Hartshorne-Residues_and_Duality}). \subsection{Semiorthogonal decompositions} Recall the following definition. \begin{definition}\label{def_admissible} Given a triangulated category $\T$, a triangulated subcategory $\A \subset \T$ is said to be (left) right admissible if the inclusion functor $i:\A \to \T$ admits a (left) right adjoint ($i^*$) $i^!$. We say simply that $\A$ is admissible if it is both left and right admissible. \end{definition} \noindent Now let us remind the reader of the notion of a semiorthogonal decomposition. \begin{definition}\label{def_semi_dec} Let $\T$ be a triangulated category. A pair of strictly full (admissible) triangulated subcategories $\A, \B$ of $\T$ is a (admissible) semiorthogonal decomposition of $\T$ if the following conditions are satisfied: \begin{enumerate} \item $\A \subset \B^\perp$, and \item for any $T \in \T$, there is a triangle $$B \to T \to A \to B[1]$$ with $A \in \A$ and $B \in \B$. \end{enumerate} We then write $\T = \langle \A,\B \rangle$. \end{definition} \begin{remark} Given a left (right) admissible subcategory $\A \subset \T$, we have a semiorthogonal decomposition $\langle \A, \prescript{\perp}{}{\A}\rangle$ (respectively $\langle \A^\perp,\A \rangle$). Conversely if $\T = \langle \A,\B\rangle$, then $\A$ ($\B$) is automatically left (right) admissible \cite{Bondal-Quivers}. \end{remark} \begin{remark} When $X$ is smooth, any semiorthogonal decomposition $\D^b(X) = \langle \A, \B \rangle$ is automatically admissible. However in the singular case, this is no longer true. Admissible semiorthogonal decompositions are significantly better behaved, as the following very useful lemma suggests. \end{remark} \begin{lemma}[\cite{Orlov-Triangulated_cats_of_singularities}, Lemma 1.10]\label{lemma-restrict-sod} Let $X$ be a projective variety and suppose that $\D^b(X) = \langle \A_1,...,\A_n\rangle$ is an admissible semiorthogonal decomposition. Then $$\perf X = \langle \A_1\cap \perf X,...,\A_n \cap \perf X\rangle$$ is an admissible semiorthogonal decomposition for $\perf X$. \end{lemma} \begin{remark} If, in Definition \ref{def_semi_dec}, we also required the dual condition that $\B \subset \prescript{\perp}{}{\A}$, so that there are no morphisms from objects of $\A$ to objects of $\B$, then the corresponding notion is that of an orthogonal decomposition. It is well known that a connected scheme admits no nontrivial orthogonal decompositions of either $\perf X$ or $\D^b(X)$. \end{remark} \section{Main results} \begin{definition} Let $\T$ be a triangulated category. We say that $\T$ is (weakly) indecomposable if any (admissible) semiorthogonal decomposition is trivial. That is, whenever $\T = \langle \A,\B \rangle$, we have that either $\A=0$ or $\B=0$. \end{definition} \noindent The original version of the following theorem can be found \cite[Theorem 3.1]{KawataniOkawa-Nonexistence}, where it is proven for smooth DM stacks with at least one non-stacky point. Here we prove an extension to Cohen-Macaulay projective varieties. \begin{theorem}\label{theorem-main} Let $X$ be a Cohen-Macaulay projective variety with dualizing sheaf $\omega_X$. Suppose that $\perf X = \langle \A, \B \rangle$. Then: \begin{enumerate} \item for any closed point $x \in X \setminus \operatorname{Bs}|\omega_X|$, and all Kozsul zero-cycles $Z_x$ at $x$, exactly one of the following holds: \begin{enumerate} \item $\O_{Z_x} \in \A$, or \item $\O_{Z_x} \in \B$. \end{enumerate} \item When (1a) (resp. (1b)) is satisfied, then the support of any object in $\B$ (resp. $\A$) is contained in $\operatorname{Bs}|\omega_X|$. \end{enumerate} \end{theorem} The following is nearly verbatim from \cite[Theorem 3.1]{KawataniOkawa-Nonexistence}, but we include the proof with our modifications for the convenience of the reader. \begin{proof} Let $x\in X \setminus \operatorname{Bs}|\omega_X|$ be a closed point and, using that $X$ is Cohen-Macaulay, a Koszul zero cycle $Z_x$ supported at $x$. Suppose that $\perf X = \langle \A, \B \rangle$, and consider the triangle $$B \to \O_{Z_x} \to A \overset{f}{\to} B[1]$$ which is obtained from the semiorthogonal decomposition. If $f=0$, then $\O_{Z_x} \cong A \oplus B$, but this implies that one of $A$ or $B$ is zero (see Remark \ref{remark_indecomposable}), and hence $\O_{Z_x} \in \B$ or $\A$ respectively, so we may assume that $f \neq 0$. Now choose a global section $s\in H^0(X,\omega_X) \cong \hom(\O_X,\omega_X)$ such that $s(x) \neq 0$ and $x\in D(s)^\circ$, the interior of $D(s)$. We may choose such a global section as $x \notin \operatorname{Bs}|\omega_X|$, and by definition, the base locus of $\omega_X$ is the complement of the union of the interiors of the sets $D(s)$ (and in particular, is a closed subset of $X$). We now consider the composition $$A \otimes \O_X \overset{f \otimes \mathbbm{1}}{\to} B[1] \otimes \O_X \overset{\mathbbm{1} \otimes s}{\to} B[1] \otimes\omega_X.$$ Using the Rouqiuer functor, the composition lies in $$\hom_{\D^b(X)}(A,B\otimes \omega_X[1]) \cong \hom_{\perf X}(B,A[\dim X -1])^* =0,$$ where the latter space is zero by the assumption that $\perf X = \langle \A ,\B \rangle$. Now let $U \subset D(s)^\circ$ be an open neighborhood of $x$, and set $j :U \hookrightarrow X$ to be the open immersion. We now examine $j^* (f \otimes s) : = (f \otimes s)|_U = f|_U \otimes s|_U$ in exactly the same manner as \cite[Theorem 3.1]{KawataniOkawa-Nonexistence}. We know that this composition must be identically zero, and further we know that $j^* s := s|_U \neq 0$ by our choice of $U$. So we see that $f|_U = 0$; then since $j^*$ is triangulated, the restriction of the triangle $$B|_U \to \O_{Z_x}\to A|_U \overset{0}{\to} B[1]|_U$$ implies that $\O_{Z_x} \cong A|_U \oplus B|_U$. As before, one of the two summands must be trivial. Suppose that $A|_U \cong 0$. Then $\O_{Z_x} \cong B|_U$, and since the support of $A$ and $\O_{Z_x}$ is disjoint, it follows that the morphism $\O_{Z_x} \to A$ is zero, and hence $B \cong \O_{Z_x} \oplus A[-1]$. If $A[-1] \neq 0$, then there must exist a nonzero projection $B \to A[-1]$, contradicting that the decomposition was semiorthogonal. Hence $A[-1] \cong 0$, and $\O_{Z_x} \in \B$. If instead $B|_U \cong 0$, we similarly obtain $\O_{Z_x} \in \A$. For the second claim, we also proceed in the same manner as in \cite[Theorem 3.1]{KawataniOkawa-Nonexistence}, but we use Lemma \ref{lem_detect_support} in place of the corresponding \cite[Lemma 2.8]{KawataniOkawa-Nonexistence}. Notice that if $\O_{Z_x} \in \A$ (or $\B$) for some closed point $x \in X$, then any object in $\B$ (or $\A$) must have proper support, by Lemma \ref{lem_detect_support}. So now suppose that both $\A$ and $\B$ contain Koszul zero cycles, and consider the canonical filtration of the structure sheaf $$B \to \O_X \to A \to B[1].$$ It then follows that the support of $\O_X$ is the union of the supports of $A$ and $B$. But since both $\A$ and $\B$ contain Koszul zero cycles the supports of $A$ and $B$ are proper, which is a contradiction since we have assumed $X$ is irreducible. Hence only one of the two subcategories can contain any of the Koszul zero cycles. \end{proof} \begin{corollary}\label{corollary-main} Suppose that $X$ is a Cohen-Macaulay projective variety such that every connected component of $\operatorname{Bs}|\omega_X|$ is contained in an open subset on which $\omega_X$ is trivial. Then $\perf X$ is indecomposable. If $\omega_X$ is globally generated, then $\D^b(X)$ is weakly indecomposable. \end{corollary} \begin{proof} Suppose that every connected component of $\operatorname{Bs}|\omega_X|$ is contained in an open subset on which $\omega_X$ is trivial. Then by Theorem \ref{theorem-main}, given any nontrivial semiorthogonal decomposition $\perf X = \langle \A, \B \rangle$, one of the components must satisfy the condition that all of its objects must be supported in $\operatorname{Bs}|\omega_X|$. Without loss of generality we can assume that this component is $\B$. Then for any $B \in \B$, since its support is contained in an open subset where $\omega_X$ is trivial, $R_X(B) \cong B$. Hence for any object $A \in \A$, and any $n \in \Z$, $$\hom (B,A[n]) \cong \hom(A,B[n+\dim X])^* =0$$ by semiorthogonality, and so the decomposition $\perf X = \langle \A,\B \rangle$ is actually an orthogonal decomposition. But this contradicts that $X$ is connected (as we have assumed it is integral). Now suppose that $\operatorname{Bs}|\omega_X|$ is empty. Consider an admissible semiorthogonal decomposition $\D^b(X) = \langle \A, \B \rangle$. Then this induces a semiorthogonal decomposition of $\perf X$ by Lemma \ref{lemma-restrict-sod}, and by Theorem \ref{theorem-main}, one of the components, say $\tilde \A$ for concreteness, must contain the spanning set of Koszul zero-cycles $\{\O_{Z_x}\,|\, x\in X \text{ closed}\}$. It then follows that $\A$ also contains the spanning set of Koszul zero cycles since $\tilde \A = \perf X \cap \A$, and hence both $\B$ and $\tilde \B$ are zero. In particular, the admissible decomposition $\D^b(X) = \langle \A, \B \rangle$ must be trivial. \end{proof} These results alone are interesting, but due to the nature of the dualizing sheaf, the base locus $\operatorname{Bs}|\omega_X|$ can be difficult to analyze in many situations if $\omega_X$ is not globally generated. The next simplest situation is when $\operatorname{Bs}|\omega_X|$ is a finite number of points. To analyze this, we formulate two more results of possibly independent interest. \begin{lemma} \label{prop_main} Let $X$ be a projective variety, and let $Y \subset X$ be a finite collection of isolated closed points. Suppose that $\A \subset \D^b(X)$ is a left admissible subcategory such that the support of every object in $\A$ is contained in $Y$. Then $\A=0$. \end{lemma} \begin{proof} Suppose that the support of every object in $\A$ is contained in $Y$. Since $\A$ is left admissible, there is a semiorthogonal decomposition $\langle \A ,\prescript{\perp}{}{\A} \rangle = \D^b(X)$ (where $\prescript{\perp}{}{\A}$ is only right admissible in general). Assume by contradiction that $\A \neq 0$, and choose some nonzero object $A\in \A$. Using that the support is isolated, $A \cong \bigoplus_{y_i \in Y} A_i$, where the support of $A_i$ is exactly the closed point $y_i \in Y$. Now choose $B\in \prescript{\perp}{}{\A}$ so that its support is all of $X$. We can guarantee that such an object exists by looking at the distinguished filtration of $\O_X$ induced by $\langle \A ,\prescript{\perp}{}{\A} \rangle$ and the long exact sequence of cohomology sheaves. By Lemma \ref{lem_detect support2}, we then see that there are integers $j_i \in \Z$ such that $$\hom_{\D^b(X)}(B[j_i],k(y_i)) \neq 0.$$ Similarly, if $l_i$ is the minimal $l_i \in \Z$ such that $\mathcal{H}^{l_i}(A_i) \neq 0$ (recall the support of $A_i$ is exactly $y_i$), we have, by Lemma \ref{lemma_local maps}, a nonzero morphism $k(y_i) \to \mathcal{H}^{l_i}(A_i)$ realizing an isomorphism between the residue field and a submodule. The composite morphism $B[j_i] \to \mathcal{H}^{l_i}(A_i)$ is nonzero. Now composing with the morphisms $\mathcal{H}^{l_i}(A_i) \to A_i[l_i] \to A[l_i]$, where the first morphism uses that $l_i$ was minimal, we can see that the composition $B[j_i] \to A[l_i]$ is a nonzero map between two objects of $\prescript{\perp}{}{\A}$ and $\A$, which contradicts semiorthogonality. Hence we must have $\A=0$. \end{proof} \begin{lemma}\label{lem_perf support} Let $X$ be a projective variety and $F,G \in \perf X$ such that $\operatorname{supp}F \cap \operatorname{supp}G$ contains an isolated point. Then $\R\hom(F,G) \neq 0$. \end{lemma} \begin{proof} This follows as $\R \hom_{\perf X}(F,G) \cong \R \Gamma(X,F^* \otimes G)$. Since the support of $F^* \otimes G$ contains an isolated point, this complex decomposes into a nontrivial direct sum, with the support of one summand exactly the isolated point. Any object with zero-dimensional support has a nonvanishing (hyper)cohomology class given by a nonzero global section of the lowest degree cohomology sheaf, so we are done. \end{proof} \begin{corollary}\label{cor_point support sod} Let $X$ be a projective variety and assume that $\perf X =\langle \A,\B \rangle$ or $\D^b(X)=\langle \A,\B \rangle$, is a semiorthogonal decomposition, respectively, admissible semiorthogonal decomposition. Suppose that $Y \subset X$ be a finite collection of isolated closed points and that the support of every object in $\A$ or $\B$ is contained in $Y$. Then the semiorthogonal decomposition is trivial. \end{corollary} \begin{proof} For $\D^b(X)$, since we have assumed that the decomposition is admissible, this follows from Lemma \ref{prop_main}. Now for $\perf X$, this is handled by Lemma \ref{lem_perf support}. \end{proof} As we hinted at earlier, these results allow us to also resolve the question of indecomposability when $\operatorname{Bs}|\omega_X|$ is finite. \begin{corollary}\label{corollary_isolated} Let $X$ be a Cohen-Macaulay projective variety such that $\operatorname{Bs}|\omega_X|$ is a finite number of isolated closed points. Then $\perf X$ is indecomposable and $\D^b(X)$ is weakly indecomposable. \end{corollary} \begin{proof} Suppose we have a decomposition $\perf X = \langle \tilde\A,\tilde\B \rangle$. By Theorem \ref{theorem-main}, either $\tilde \A$ or $\tilde \B$ contains all Koszul zero-cycles $\O_{Z_x}$ for $x \notin \operatorname{Bs}|\omega_X|$. Without loss of generality we can assume its $\tilde\A$. Hence $\tilde \B$ satisfies the hypotheses in Corollary \ref{cor_point support sod}, so the decomposition of $\perf X$ is trivial. We see that this category is indecomposable. Now suppose we have an admissible decomposition $\D^b(X) = \langle \A,\B\rangle$. This, by Lemma \ref{lemma-restrict-sod}, induces $\perf X = \langle \tilde\A,\tilde\B \rangle$, where $\tilde \A = \perf X \cap \A$ and $\tilde \B = \perf X \cap \B$. The above paragraph implies (again, without loss of generality) that $\A \subset \D^b(X)$ also contains all Koszul zero-cycles $\O_{Z_x}$ for $x \notin \operatorname{Bs}|\omega_X|$. So $\B$ satisfies the hypotheses in Corollary \ref{cor_point support sod}, hence $\B =0$. Since this decomposition was arbitrary, we are done. \end{proof} The following corollary relates these observations to curves. Note that it is known that smooth and Gorenstein curves of arithmetic genus larger then one have empty canonical base locus (\cite[Theorem 1.6]{Hartshorne-Generalized_divisors} and \cite[Theorem D]{Catanese-Pluricanonical_gorenstein_curves}). So Corollary \ref{corollary-main} resolves this question for Gorenstein curves. In the Cohen-Macaulay case however, we do not know if the canonical base locus is empty, but we can still conclude the same result. \begin{corollary}\label{corollary-curves} Let $C$ be a projective curve of arithmetic genus at least one. Then $\perf C$ is indecomposable and $\D^b(C)$ is weakly indecomposable. \end{corollary} \begin{proof} First note that any projective curve is Cohen-Macaulay. By Corollary \ref{corollary_isolated}, it is enough to see that the base locus of the dualizing sheaf $\omega_C$ consists of isolated closed points. Let $p \in C$ be a regular point; given the exact sequence $$0 \to \omega_C(-p) \to \omega_C \to k(p) \to 0,$$ we see that $p$ is a base point of $\omega_C$ if and only if the map $$H^1(\omega_C(-p)) \to H^1(\omega_C) \to 0$$ is not an isomorphism, equivalently $h^1(\omega_C(-p)) \geq 2$. We claim that in fact $h^1(\omega_C(-p)) = 2$. Indeed, since $p$ is regular, Riemann-Roch informs us that $h^0(p) - h^1(p) = 2-p_a(C)$. Further, from the long exact sequence of cohomology arising from the short exact sequence above, we have $h^1(p) = h^0(\omega_C(-p)) \leq p_a(C)$. Rearranging, we have $$h^0(p) - 2 = h^0(\omega_C(-p)) - p_a(C),$$ where the left hand side is nonnegative, and the right hand side is nonpositive, hence they must both be equal, and $h^0(p) = 2$. Thus $p$ is a base point of $\omega_C$ if and only if $h^0(C,\O_C(p)) = 2$. However this latter equality holds only if $p_a(C) =0$. This shows that if $C_{reg}$ is the locus of regular points, $C_{reg} \cap \operatorname{Bs}|\omega_C|= \emptyset$, in particular, $\operatorname{Bs}|\omega_C|$ is isolated. \end{proof}

28aeb31452ea9cf5e703ad764a8cfde2ddbbf262

\section*{Appendix} \addcontentsline{toc}{section}{Appendix} When placed at the end of a chapter or contribution (as opposed to at the end of the book), the numbering of tables, figures, and equations in the appendix section continues on from that in the main text. Hence please \textit{do not} use the \verb|appendix| command when writing an appendix at the end of your chapter or contribution. If there is only one the appendix is designated ``Appendix'', or ``Appendix 1'', or ``Appendix 2'', etc. if there is more than one. \begin{equation} a \times b = c \end{equation} \section{Data} \label{sec:data} Taxpayers in Mexico are identified by their tax number (`{\it Registro Federal de Contribuyentes}', or RFC), a unique string of alphanumerical characters used to emit and receive invoices, submit tax statements and engage in other procedures. Since 2014, a large number of income and outcome transactions between taxpayers in Mexico have been recorded in CFDIs and stored by SAT. The data used in this work includes: \begin{itemize} \item A set of 81,511,015 taxpayer identifiers, anonymized to protect individual identities, which we denote RFCAs, and categorical information for each one of them that includes: taxpayer type (individual or legal entity), location, date of registration, and economical sector and activity. \item A total of 6,823,415,757 monthly CFDI aggregated emissions between taxpayers distributed between January 2015 and December 2018, which include the RFCA of both the emitter and receiver of the transaction, the month and year, type (either income or outcome), the number of transactions for that month, and the total amounts associated with them. \item A list of 8,570 RFCAs previously identified by Mexican government authorities as definitive or alleged EFOS. We use this data to train machine learning models and as focal points in the network science approach. \end{itemize} In the 48 months of CFDI emissions we analyze, 7,571,093 RFCAs emitted at least one CFDI, so that the already identified evaders (EFOS) account only for 0.0072\% of active taxpayers (those who at least emitted one receipt during the period of study). This means that the data is highly unbalanced: the ratio between the identified class (EFOS) and the undetermined ones is quite different, which has an impact in the design of the models and approaches we use. In what follows, when we refer to a RFCA as an EFOS, we mean those taxpayers already identified by the authorities as either definitive or alleged evaders. Unclassified RFCAs correspond to those that have not been classified as EFOS by the authorities, and we refer as suspects to those RFCA which are classified as possible evaders by our methods. \section{Discussion} \label{sec:discussion} We have shown that it is possible to use tools from network science and machine learning to automatically identify patterns of tax evaders similar to those that humans already identified. This is promising, but should also be taken with caution. It is promising, because similar techniques could be applied in a broad variety of areas: money laundering, bribery practices, and other illegal activities. In principle, this would benefit society. However, one should also be aware of the limits of these methods. First, the patterns identified are based on those already known to humans. This means that different patterns of tax evasion will not be identified. Second, using these techniques to curb illegal activities would promote criminals to use alternative patterns that are not identified, so an arms race would ensue. The tools have a relevant potential, but by themselves, are not a final solution. Finally, there is always the probability of misidentifying honest citizens or companies for wrongdoers, simply because they have similar statistical patterns. Thus, these methods can be used to identify and prioritize \emph{potential} suspects from a huge pool, but the final decision has to be made by humans. In general, our work illustrates the potential of recent technology to solve different problems exploiting large data sets, computing power, and sophisticated statistical techniques. Still, the limits of this technology are yet to be defined properly, which has led to much hype and overconfidence. Also, ethical issues must be considered for the use of this technology. As more applications are developed, we can learn from them to better situate the benefits and risks of using network science, machine learning, and related techniques to address social problems. \subsection{Merging the classification methods} \label{sec:integracion} We obtain a list of suspect RFCAs from each Machine Learning classification method (ANN and Random Forest). It is to be noted that the training set for both of these methods, consisted of examples from the evaders already identified by the authorities, therefore, there is an inherent and implicit bias in our methods towards the mechanisms and assumptions used by the authorities to identify evaders. This bias in unavoidable due to the data we had available and it is important to take into account further methods to make a wider characterization of these and other evasion mechanisms. We report the number of suspect RFCAs identified by each method in Table~\ref{tab:sosps_metodos}. The list obtained from the ANN corresponds to those RFCAs with an probability of belonging to the same class as EFOS \(> 0.8\). We use the same threshold for the probability assigned by the Random Forest model. The intersection of both suspect lists is contains 43,650 RFCA, which we consider to have a higher probability of being suspect evaders, as they were identified independently by two different methods. \begin{table}[th!] \centering \begin{tabular}{ c|c|c } Method & Classified as suspect & Unclassified \\ \hline Artificial Neural Network & 149,921 & 7,416,754 \\ Random Forest & 128,227 & 7,438,448 \\ \hline \end{tabular} \caption{ Number of suspect RFCA identified by each classification method. }\label{tab:sosps_metodos} \end{table} \subsubsection{Suspect RFCA comparison with EFOS} To compare the features of the suspect RFCAs identified by our classification methods with those of EFOS, we compute the distributions of the active and canceled amounts associated to their CFDI emissions. Active amounts correspond to the emitted and processed invoices, and canceled amounts correspond to those invoices that were canceled and not processed. The cancellation of CFDIs can be done freely and independently by taxpayers without an explicit authorization. As can be seen in Fig.~\ref{fig:boxPlots}, the distributions of these two variables are very similar for EFOS and the suspect RFCAs identified by our methods, while for unclassified RFCA, the distributions of these variables are different. Furthermore, in Fig.~\ref{fig:sops_trend}, we show the time behavior of the active and canceled amounts of the CFDIs emitted by the different RFCA groups (definitive EFOS, alleged EFOS and unclassified RFCAs). It can be seen that the difference between suspect and unclassified RFCAs is consistent for the whole analysis period, which sets these two groups (EFOS and suspect RFCA) apart from unclassified RFCA. \begin{figure}[th!] \centering \subfloat[\label{fig:box_mt_act}]{\includegraphics[width=0.45\linewidth]{images/resultados/boxplots/box_mt_can_15_12_sosps.eps}} \hspace{1cm} \subfloat[\label{fig:box_sub_act}]{\includegraphics[width=0.45\linewidth]{images/resultados/boxplots/box_sub_act_15_12_sops.eps}} \caption{ Boxplots for the distributions of the (a) logarithm of the total cancelled amount and (b) the active subtotal amount of the CFDI emitted by EFOS, suspect and unclassified RFCA. It can be seen that the distributions for EFOS and suspect RFCA are more similar between each other than when being compared with the distributions of unclassified RFCA. }\label{fig:boxPlots} \end{figure} \begin{figure}[th!] \centering \subfloat[\label{fig:no_quant_mt_can}]{\includegraphics[width=0.45\linewidth]{images/resultados/temp_sosps_sub_act.pdf}} \hspace{0.04\linewidth} \subfloat[\label{fig:no_quant_sub_act}]{\includegraphics[width=0.45\linewidth]{images/resultados/temp_sosps_mt_can.pdf}} \caption{ Temporal behavior of the (a) Active subtotal amount and (b) Cancelled total amount for each one of the groups of RFCAs: definitive EFOS (blue), alleged EFOS (red), unclassified RFCAs (green) and suspect RFCAs (cyan). The vertical dashed lines correspond to December for each of the years studied. It can be seen that the EFOS and suspect RFCAs behavior differentiates from unclassified RFCAs. }\label{fig:sops_trend} \end{figure} \subsubsection{Number of close EFOS to suspect RFCA} As it was discussed in Section~\ref{sec:redes}, the EFOS reach in the interaction networks allows us to identify the number of close EFOS to unclassified RFCA in the network (where close means at a distance \(d \leq 3\)) and use it to select those RFCAs more immersed in the EFOS operation networks. Now, if we look at the suspect RFCAs identified by the classification methods, and compute the number of close EFOS in a whole year, we observe that they are close to a large number of EFOS (see Fig.~\ref{fig:efos_cercanos_sosps}), which suggests that the suspect RFCAs in the intersection of the lists obtained from both classification method are colluded with the EFOS identified by the authorities and gives us confidence about our methods. It is to be noted that the closeness to EFOS wasn't part of the set of features used by the classification methods (ANN and Random Forest) to identify suspect RFCAs, as they were based only on CFDI data. \begin{figure}[th!] \centering \subfloat[\label{fig:sosps_num_efos_y_15}]{\includegraphics[width=0.49\linewidth]{images/redes/figs_reach/sosps_num_efos_y_15.pdf}} \subfloat[\label{fig:sosps_num_efos_y_16}]{\includegraphics[width=0.49\linewidth]{images/redes/figs_reach/sosps_num_efos_y_16.pdf}} \caption{ Total of close EFOS the suspect RFCAs identified by our classification methods for (a) 2015 and (b) 2016. It can be seen that a suspect RFCA are close to a large number of EFOS, which indicates that these RFCAs are immersed into the EFOS operation networks. }\label{fig:efos_cercanos_sosps} \end{figure} As we show in Table~\ref{tab:frac_pers}, 81.52\% of the suspect RFCA are legal persons, which suggests that most part of the emitted CFDI linked with allegedly simulated operations is made between companies or businesses. This can be due to the fact that, in this case the legal responsibility of possible illicit operations, falls to a legal entity and not a natural person. In table~\ref{tab:frac_pers}, we also show that 91.15\% of the suspect RFCA were active at the moment of the elaboration of the study, and less than 1\% of them had a cancelled or suspended status, which shows that most of the suspect RFCA are economically active, and thus susceptible of being investigated. \begin{table} [th! \centering \begin{tabular}{ c|c } Taxpayer type & Percentage of suspect RFCAs \\ \hline Legal & 81.52\% \\ Natural & 10.22\% \\ Without info & 8.3\% \\ \hline \end{tabular} \hspace{0.25cm} \begin{tabular}{ c|c } Status & Percentage of suspect RFCAs \\ \hline Active & 91.15\% \\ Cancelled & 0.13\% \\ Suspended & 0.46\% \\ Without info & 8.3\% \\ \hline \end{tabular} \caption{ Type of taxpayer and status of the suspect RFCA identified by the classification methods. Most of them were active legal taxpayers at the moment, which made them susceptible of being investigated by the authorities. }\label{tab:frac_pers} \end{table Using the data contained in the CFDI, we define the potential tax collection associated to an arbitrary RFCA, \(\mathrm{rec_{VAT}}\phi_i\), as the difference between the total nominal tax obtained by the income CFDI emitted during the year, \(\mathrm{VAT_{Nom}}_i\), and the payed tax reported in their tax statements, \(\mathrm{VAT_{Payed}}_i\), i.e. \begin{equation} \mathrm{rec_{VAT}}\phi_i = \sum \mathrm{VAT_{Nom}}_i - \mathrm{VAT_{Payed}}_i, \end{equation} so that, we define the total nominal tax collection that correspond to the set of suspect RFCA, \(\mathrm{REC_{IVA}}\phi \) as: \begin{equation} \mathrm{REC_{VAT}}\phi = \sum_i \mathrm{rec_{VAT}}\phi_i. \label{eq:def_rec} \end{equation} We use the income CFDI emitted by the identified suspect RFCA between the years 2015 and 2018, and their yearly tax statements, which include the total VAT payed by them. We believe that the information of the emitted CFDI allows us to have a better description of the economical activity and evasion mechanisms, because the income amounts declared in the tax statements is vulnerable of manipulation, and may not reflect reality, mostly because we are dealing with taxpayers suspects of simulating operations. A significative difference between income amounts from CFDI emissions and tax statements could be a indicator of possible illicit practices. \subsection{Yearly evasion estimates} As we discussed in section~\ref{sec:redes}, the number of close EFOS to a RFCA in the interaction networks, is an indicator of their collusion level in the sub-networks associated with EFOS, so that we can assume that a suspect RFCA close to a larger number of definitive and alleged EFOS, is much more susceptible to conduct the same practices as EFOS. With this in mind, we calculate evasion estimates for each year between 2015 and 2018, based on the suspect RFCA closest to EFOS in the monthly interaction networks for each year, which correspond between 28\% and 38\% of all suspect RFCA. \begin{figure}[th!] \centering \includegraphics[width=0.75\linewidth]{images/evasion/montos_cotas_mdp.pdf} \begin{tabular}{ c | c | c | c } \hline \hline\noalign{\smallskip} \multicolumn{4}{c}{VAT evation estimates (MP)} \\ \hline\noalign{\smallskip} \hline Year & Minimum & Average & Maximum \\ \hline 2015 & 40,097.27 & 111,048.36 & 185,087.23 \\ 2016 & 60,626.86 & 140,041.13 & 220,922.03 \\ 2017 & 64,377.11 & 173,717.06 & 286,273.35 \\ 2018 & 77,318.59 & 215,518.71 & 346,106.32 \\ \hline \hline\noalign{\smallskip} Year average & 60,604.96 & 135,081.31 & 259,597.23 \\ \hline \hline \end{tabular} \hspace{1.5cm} \begin{tabular}{ c | c | c } \hline \hline\noalign{\smallskip} \multicolumn{3}{c}{Unique suspect RFCA} \\ \hline\noalign{\smallskip} \hline Year & Minimum estimate & Maximum estimate \\ \hline 2015 & 2,686 & 10,767 \\ 2016 & 3,132 & 12,510 \\ 2017 & 3,461 & 13,743 \\ 2018 & 3,541 & 14,080 \\ \hline \hline\noalign{\smallskip} Total & 7,677 & 17,769 \\ \hline \hline \end{tabular} \caption{ Maximum and minimum VAT evasion estimates in millions of pesos (MP) associated to the emission of CFDI of potentially simulated operations from suspect RFCA in the period 2015 -2018. We present as well the number of unique RFCA used for the calculations for each year. We estimate the number of unique suspect RFCA to be between 7,677 and 17,769. }\label{fig:montos_cotas} \end{figure} The EFOS proximity index threshold, \(\theta_{\sigma}\), allows us to define minimum and maximum values for VAT evasion estimates. The maximum estimated value corresponds to all identified suspect RFCA, and minimum values to estimates based on suspect RFCA of the last quartile (top 25\%) of the EFOS proximity index distribution, which correspond to 7,677 unique suspect RFCA with CFDI emissions between 2015 and 2018, which corresponds to an average estimate of 60,604.96 millions of pesos per year. The VAT evasion estimates in Fig.~\ref{fig:montos_cotas} should not be considered as final values, as there might exist VAT evasion mechanisms different from the simulation of operations, which we don't take into account in this work. Furthermore, it is important to mention that there is an additional uncertainty because, it is not possible to determine precisely the fraction of simulated and real transactions made by the suspect EFOS with the available information. Because of this situation, we assume the 100\% of the emitted CFDI by suspect RFCA as simulated operations. A follow up study focused in the traceability of the emitted CFDI could help determine the fraction of simulated operations, which would allow us to make a more precise VAT evasion estimate. \section{Introduction} \label{sec:introduction} Tax has a crucial role in the economic growth and welfare of the general population. Paid taxes allow for government spending and public expenditures (in the short, medium, and long terms) such as education, healthcare, housing, pensions, security, and infrastructure~\cite{CPEUM}. Tax collection in Mexico is determined by a series of laws ({\it Ley de Ingresos de la Federación}, in Spanish) specifying the eligibility of taxpayers, tax rates and types, as well as the periodicity of fees and means of payment. Despite the general view that taxes are justified since they allow the Mexican state to fund beneficial activities for society, some individuals, corporations, or trusts may decide to illegally evade taxes by misrepresenting their state of affairs to the Tax Administration Service in Mexico ({\it Servicio de Administración Tributaria}, or SAT). In this sense, tax evasion is the reduction of constitutional tax liability by dishonest reporting, such as understating financial gains or overstating deductions \cite{tovar2000evasion}. Since 2014, Mexico has kept electronic records of all taxable transactions by means of a digital receipt or invoice known as {\it Comprobante Fiscal Digital por Internet} (CFDI). Each of these mandatory receipts includes data on the product or service transferred between taxpayers, date of transaction, cost, and corresponding tax amount. CFDIs are XML documents with technical specifications updated annually by SAT, including a certification seal that can only be produced by authorized parties \cite{CFDISAP}. Since CFDIs potentially uncover networks of individuals and legal entities involved in commercial transactions leading to tax revenue, they are an integral part of formal investigations by SAT in tax evasion, money laundering, and other tax-related illegal activities. Among all forms of tax evasion, here we focus in situations where taxpayers issue CFDIs in the absence of actual economic activity to increase tax deductions. Such legal entities, known as `enterprises billing simulated operations' (`{\it empresas que facturan operaciones simuladas}', or EFOS), are typically characterized by a lack of employees or infrastructure, as well as a fake or constantly changing address used to avoid detection. As a result of investigations already led by SAT, EFOS can be classified as {\it definitive} or {\it alleged} tax evaders \cite{DefEFO}. In order to operate, EFOS require the participation of `enterprises deducting simulated operations' (`{\it empresas que deducen operaciones simuladas}', or EDOS). Even if EDOS engage in illegal activity alongside EFOS, they tend to have demonstrable stability in their workforce, assets, and tax contributions. By receiving CFDIs associated with simulated operations, EDOS aim to reduce their tax rate to avoid payments to SAT and eventually obtain further fiscal benefits. In order to decrease the risk of systematic tax evasion, the Mexican federal government has established fiscal law (known as `{\it Código Fiscal de la Federación}') describing the official procedure to identify taxpayers as EFOS: a) First, SAT determines the lack of actual economic activity behind a set of CFDIs. b) Then, the government notifies the relevant legal entities via its official publication (`{\it Diario Oficial de la Federación}'). c) Alleged EFOS have 15 days to contest the claim. d) Associated EDOS (that have received the suspicious CFDIs) can correct their standing with SAT by resubmitting appropriate tax forms. e) Finally, if any tax revenue has been lost to illegal activity, SAT classifies relevant taxpayers as definitive EFOS and specifies the type of crime following fiscal law. Definitive EFOS are unable to emit further CFDIs. Despite its effectiveness, the official procedure of detecting tax evasion is time and resource consuming, particularly in initial stages of the process where suspicious CFDIs and alleged EFOS need to be manually selected from millions of contributors and billions of transactions each year. In order to complement these efforts, automated computational and statistical techniques (with tools from network science and machine learning) can be used to characterize the network of issued/received CFDIs between definitive EFOS and other taxpayers. By analyzing the temporal properties of the network of all taxable transactions in Mexico from 2015 to 2018, here we show evidence of a group of highly suspicious contributors with behavior statistically similar to that of definitive EFOS, as well as estimates of previously undetected tax evasion. When combined with current practices at SAT, this information has the potential to increase the efficacy of the governmental response to illegal tax activity in Mexico. \section{Results} \label{sec:results} \input{redes.tex} \input{network_science.tex} \input{integration.tex} \input{discussion.tex} \begin{acknowledgement} This manuscript describes research associated with a project advising the Tax Administration Service (SAT) of the Mexican federal government. The official report of the project (in Spanish) is available in the SAT website at \url{http://omawww.sat.gob.mx}. We thank Juan Pablo de Bottom, Alejandra Cañizares Tello, Leonardo Ignacio Arroyo Trejo, and Aline Jacobo Serrano at SAT, as well as Alejandro Frank Hoeflich, José Luis Mateos Trigos, Juan Claudio Toledo Roy, Ollin Langle, Juan Antonio López Rivera, Eric Solís Montufar, Octavio Zapata Fonseca, Romel Calero, José Luis Gordillo, and Ana Camila Baltar Rodríguez at UNAM. G.I. acknowledges partial support from the Air Force Office of Scientific Research under award number FA8655-20-1-7020, and by the EU H2020 ICT48 project Humane AI Net under contract 952026. C.P. and C.G. acknowledge support by projects CONACyT 285754 and UNAM-PAPIIT IG100518, IG101421, IN107919, and IV100120. \end{acknowledgement} \input{references.tex} \end{document} \subsection{Complex network approach}\label{sec:redes} In this section we describe the way in which we define interaction networks between EFOS and unclassified RFCA based on the emission and reception of CFDIs. We also describe the analysis we perform on the topology of the network and the roles EFOS and the rest of the RFCA play in them. This analysis allowed us to build the metrics used to define suspect evaders in the interaction networks. \subsubsection{Interaction network definition} The taxpayers activity records allows us to define interaction networks between them in which nodes correspond to taxpayers (identified by their RFCA) and are classified in one of three categories: definitive EFOS (those evaders already identified by the authorities), alleged EFOS (suspect evaders identified by the authorities), and unclassified RFCA. Links in the network correspond to directed transactions between taxpayers, which as CFDI themselves, can represent either income or outcome transactions (see Fig.~\ref{fig:def_enlace}). \begin{figure} [th!] \centering \includegraphics[width=0.5\linewidth]{images/redes/def_enlace.eps} \caption{ Directed links in the network correspond to a CFDI emitted between RFCA. This digital receipts (CFDI) can represent incomes and outcomes. } \label{fig:def_enlace} \end{figure} The topology of these networks represent the relationships between groups of taxpayers, which we assume reflect some of the association patterns and mechanisms EFOS and other RFCA have used for their practices and which we use to identify suspect evaders. With the available data we construct yearly and monthly interaction networks. On one hand, the year timescale allow us to identify a set of RFCA with whom EFOS interact more regularly. On the other hand, we have identified on the month timescale, that the amounts associated with transactions between EFOS occur more frequently inside an interval we have termed \textit{EFOS activity regime}, which correspond to higher amounts than those observed in transactions between unclassified RFCA. We build the interactions networks between RFCA, only taking into account the transactions with amounts inside this interval and characterize their topology and structure. \subsubsection{Yearly interaction networks}\label{ssect:redes_anuales} We first consider the interaction network built only from the income CFDI emissions and receptions of the nodes associated to EFOS with at least 10 transactions in a year. This restriction selects the nodes that interact more frequently during a year (at least once a month which), according to the homophily principle in social networks~\cite{aiello2012friendship, mcpherson2001birds, currarini2016simple,asikainen2020cumulative}, correspond to nodes which are more similar between them. \begin{figure} [th!] {\includegraphics[width=0.48\linewidth]{images/redes/com_3_2015.png}} {\includegraphics[width=0.48\linewidth]{images/redes/com_3_2016.png}} \caption{ Examples of the strongly connected components observed in interaction networks at year timescale, the left panel corresponds to 2015 and the right panel to 2016. Red nodes correspond to confirmed EFOS, yellow to suspect EFOS and blue nodes to unclassified RFCA. } \label{fig:ejemplos_coms} \end{figure} We identify the strongly connected components (SCC) in these networks with organized sets of taxpayers which can be related to anomalous financial activity. We show the largest SCC observed in 2015 and 2016 in Fig.~\ref{fig:ejemplos_coms}. Remembering that links in the network correspond to transactions, the presence of these structures imply a circular flow of goods and services, which being related to nodes associated to EFOS, is possible that these sets of nodes carry out tax evasion practices such as the exchange of receipts of simulated transactions, which suggests that the unclassified RFCA in the SCC might be suspect of carrying out the same practices. \subsubsection{Monthly interaction networks}\label{ssec:nivel_oper} In this section we consider the interaction networks between taxpayers at the month timescale. Unlike the yearly interaction networks where we only took into account the emissions and receptions of the nodes associated with EFOS, in this case we consider the transactions between all three kinds of nodes (EFOS, alleged EFOS and unclassified RFCA). Nonetheless, as the whole set of transactions is huge, we need to define criteria to filter transactions to reduce the network into a manageable size. \begin{figure}[th!] \centering \includegraphics[width=0.495\linewidth]{images/redes/temp_defs.pdf} \includegraphics[width=0.495\linewidth]{images/redes/temp_pres.pdf} \caption{ Time behaviour of the logarithm of the subtotal of the transactions associated to emissions from EFOS (left panel) and alleged EFOS (right panel), the vertical dashed lines indicate the month of December of each year. Solid lines show the median and the dashed lines the interquartile range of the distribution. It can be seen that the transactions between EFOS, either definitive or alleged, correspond to amounts around tens of thousands and millions of pesos. We define this range of amounts as the EFOS activity regime, which we use to filter the links in the interaction networks. }\label{fig:emision_dif} \end{figure} To this end, we obtain the distribution of the subtotal amounts before taxes of the transactions emitted by EFOS (for both definitive and suspect) to the remaining types of nodes. As can be seen in Fig.~\ref{fig:emision_dif}, the median of the distribution changes over time, showing an increase towards the end of the year. It is to be noted that the amounts of the transactions between EFOS are higher than those of the transactions between EFOS and unclassified RFCA, which suggests that EFOS make selective emissions whether the receiver of the transaction is an EFOS or arbitrary RFCA. We define the EFOS activity regime, as the amounts interval defined by the interquartile ranges of the amounts distribution obtained from the EFOS transactions. We only take into account links in the monthly interaction networks whose amount lies inside the EFOS activity regime. Once we have selected the relevant links in the monthly interaction networks between taxpayers, we obtain the largest SCC of the network, which e.g. for 2015 consists of 635,588 nodes. We define the reach of the \(i\)-th node in the network at distance \(d\), \(r_i(d)\), as the fraction of nodes in the SCC up to a distance \(d\) from the node, i.e. \begin{subequations} \begin{equation} \omega_i = \{ j \;|\; d_{ij} \leq d\}, \end{equation} \begin{equation} r_i(d) = \frac{|\omega_i|}{N}, \end{equation} \end{subequations} where, \(d_{ij}\) is the length of the shortest path length between nodes \(i\) and \(j\), \(|\omega_i|\) is the number of elements in \(\omega_i\), and \(N\) is the number of nodes in the SCC. If we now consider a set of nodes \(\Omega\), we can calculate the mean reach of the set of nodes at distance \(d\), \(\langle R(d) \rangle_{\Omega}\), by \begin{equation} \langle R(d) \rangle_{\Omega} = \frac{1}{|\Omega|} \sum_{i \in \Omega} r_i(d), \end{equation} where \(\Omega\) represents either the set of nodes that correspond to EFOS or unclassified RFCA. \begin{figure} [th!] \centering \includegraphics[width=0.49\linewidth]{images/redes/figs_reach/fig_reach_y_17_m_1.pdf} \includegraphics[width=0.49\linewidth]{images/redes/figs_reach/fig_reach_y_17_m_5.pdf} \caption{ Mean reach, \(R(d)\), as a function of distance \(d\), for nodes that correspond to EFOS and to unclassified RFCA. \(R(d)\) is higher for EFOS than for unclassified RFCA, which suggests that EFOS are more efficient to distribute their transactions in the network. Data correspond to January 2017 (left panel) and May 2017 (right panel). } \label{fig:reach_month} \end{figure} The topology of the network is such that, as can be seen in Fig.~\ref{fig:reach_month}, the reach of EFOS is higher for distances \(3 < d < 7\) than for the rest of the nodes. This suggests that EFOS are more efficient to distribute their transactions in the network, which can be related to mechanisms aimed to limit the traceability of their transactions. Following the reach behavior, we define the set of nearest neighbors of a node as the set of nodes at \(d_{ij} < 3\). We plot the distribution of close EFOS, for both a month and one year's aggregates. From unclassified RFCA in Fig.~\ref{fig:cercania}, we see there are cases in which unclassified RFCA are close to more than 100 EFOS in one month. \begin{figure} [th! \centering \includegraphics[width=0.49\linewidth]{images/redes/figs_reach/hist_efos_cerc_y_18_m_6.pdf} \includegraphics[width=0.49\linewidth]{images/redes/figs_reach/hist_efos_cerc_agg_y_18.pdf} \caption{ Nearest EFOS to unclassified RFCA distribution for the monthly interaction networks (left panel) and one year's aggregate (rigt panel). The number of close EFOS is a collusion indicator of unclassified RFCA in the EFOS operation networks. As can be seen in both panels, there are unclassified RFCA close to a large number of EFOS in both timescales. }\label{fig:cercania} \end{figure} We use the number of close EFOS to unclassified RFCA in the interaction network as an indicator of their collusion in the EFOS operation networks, so that we can assume that an arbitrary RFCA close to a large number of EFOS, is more susceptible to take part in the same corrupt practices as EFOS, when compared to RFCA which are not as close to EFOS in the network. We define for each one of the identified suspect RFCA, the \textit{EFOS proximity index}, \(\sigma_i(y)\), for the \(i\)-th node in the network for the year \(y\), as que quotient of the total number of EFOS at \(d_{ij} < 3\) during the year, and the number of months these EFOS were close to the RFCA, i.e. \begin{equation} \sigma_i(y) = \frac{\mathrm{Close\;EFOS\;in\;} y}{ \mathrm{Number\;of\;months\;they\;were\;close} }. \end{equation} Note that the denominator can be less than 12, as there can be months in which the RFCA wasn't close to any of the EFOS in the network. As the number of active EFOS in the network changes over time, we normalize the EFOS proximity index, which we express by \(\hat{\sigma}_i\), with respect to the maximum observed during the year, i.e. \begin{equation} \hat{\sigma}_i(y) = \frac{\sigma_i(y)}{\max(\sigma_i(y))}, \end{equation} where \(\hat{\sigma}_i(y) \in [0,1]\), this allows us to define a threshold to be used in all periods, \(\theta_\sigma(y)\), which through the condition \(\hat{\sigma}_i(y) \geq \theta_\sigma(y) \) with \(\theta_\sigma(y) \approx 1\), allows us to select the more colluded suspect RFCAs for each year. \begin{figure}[th!] \centering \subfloat[\label{fig:indice_y_15}]{\includegraphics[width=0.49\linewidth]{images/evasion/indice_y_15.pdf}} \subfloat[\label{fig:indice_y_16}]{\includegraphics[width=0.49\linewidth]{images/evasion/indice_y_16.pdf}} \caption{ Normalized proximity index to EFOS distribution for suspect RFCA identified by the ANN and Random Forest models. This index is used as an extra validation criteria of the suspect list obtained from the Machine Learning models using characteristics observed in the interaction networks. We show the results for (a) 2015 and (b) 2016. Dashed lines represent the 25, 50 and 75\% quartiles of the proximity index distribution. }\label{fig:cercania_sosps} \end{figure} The description we've made of both yearly and monthly interaction networks between taxpayers, allowed us to identify features of the EFOS organization mechanisms and the local structure of the network around them, such as: their organization in small operation networks related to closed flows of CFDIs between them, and selective CFDI emissions between EFOS of much larger amounts than the transactions they make with unclassified RFCAs, which suggests that EFOS operate inside an activity regime defined by the amounts of their transactions (see Fig.~\ref{fig:emision_dif}). We have been able as well to quantify, by means of the reach of nodes in the network and the proximity index, the collusion of unclassified RFCAs inside the EFOS operation networks. These results suggest that complex networks analysis provides useful techniques with ample potential to describe and characterize corruption networks and operational mechanisms, which are yet to be fully explored. \subsection{Deep Neural Networks} \label{subsec:deppNN} As a first classification method of unknown RFCs on whether they behave similarly (or not) to definitive EFOS, we implement an {\it artificial neural network} (ANN). ANNs are models of automatic learning inspired by the human brain. They consist of a collection of interconnected mathematical functions with characteristics analogous to those of biological neurons and are thus called neurons. Just like biological neurons, an artificial neuron collects and classifies information based on its input connections with other neurons, and thereby alternates between an active and inactive state. The connections between neurons have a weight associated with them representative of the intensity of the interaction, such that highly weighted connections are more relevant than those associated with a low weight to modify their activation state. It is via modification of the connections weight between neurons that a neural network learns to identify patterns, a process referred to as {\it training}. Neurons of an artificial neural network are often divided into different layers: an input layer that receives the data for classification; hidden layers that undertake the classification process through modification of the weights among neurons and adjustment of the input of data weights until the classification undertaken by the network is optimal; and an output layer, from which the final result of the networks' classification of input data is derived. The output of the network is compared with the desired outcome via a loss function that yields an error quantifier. During training, these errors are propagated through the network to update the weights and minimize the loss function. ANNs have been used in a variety of tasks, including computer vision~\cite{hongtao2016applications}, voice recognition~\cite{venayagamoorthy1998voice}, automatic translation \cite{zhang2015deep}, board and video games \cite{tesauro1988neural, clark2015training, NeuralState} and medical diagnostics \cite{amato2013artificial}. They have also been used in a variety of applications in financial services, from forecasting to market studies \cite{kimoto1990stock, mizuno1998application, wilson1994bankruptcy}, to fraud detection \cite{CreditCardFraud} and risk assessment \cite{trippi1992neural, yu2008credit}. A neural network can evaluate price data and discover opportunities to make commercial decisions based on data analysis. Networks can distinguish subtle, non-linear inter-dependencies and patterns that other methods of technical analysis cannot. \subsubsection{Data preparation} In our implementation, we design an ANN that receives input data from all CFDIs associated with an issuing RFCA. By means of a technique referred to as {\it re-sampling}~\cite{japkowicz2000class}, we form a balanced sample of unknown RFCAs and definitive EFOS. The re-sampling method considered in this implementation is comprised of random re-sampling of the small class (CFDIs issued by definitive EFOS) until it contains as many examples as the other class, in order to finally obtain a large dataset with the same quantity of CFDIs issued by unknown RFCAs and definitive EFOS. The model associates each RFCA with a value between 0 and 1 related to the probability that it will be an EFOS. In what follows, we describe the procedure used to design, train and evaluate the ANN. We will follow by presenting some results and conclusions. \subsubsection{Modeling} \paragraph{ANN design} A {\it dynamic recurrent neuronal network} (DRNN) is a special type of neural network that allows introduction of an arbitrary number of rows of data (input variables) at the same time, which is useful in this context since RFCAs have varying amounts of issued CFDIs. Recurrent neural networks are structures in which the output of each execution step provide the input to the following step; this enables them to retain learned information over time. {\it Long short term memory} (LSTM)\footnote{LSTM cells are a network topology developed for the first time by Hochreiter and Schmidhuber~\cite{Hochreiter:1997:LSM} to eliminate the problem of vanishing gradient~\cite{Hochreiter:1991} through the introduction of a memory mechanism. A gradient measures how much the output of a function changes if the input changes a little. The problem is that, for very deep networks, the gradient of errors dissipates rapidly over time, ending up being very small and this prevents a change in the weighted values. Networks with this problem are capable of learning short-term dependencies, but often have difficulty learning long-term dependencies.} describes the design of artificial neurons, i.e., those that give memory to the ANN. These neurons have the best known performance to date and are particularly effective for datasets derived from time series~\cite{greff2016lstm, yin2017comparative, chung2014empirical}. In particular, out of several structures tested, the best performance was obtained with an DRNN with three LSTM cell layers, each with 256 neurons, using a hyperbolic tangential function to calculate their internal state\footnote{These three layers correspond to the hidden layers that classify the input data. In addition to the hidden layers, the network has an input and an output layer.}. It is worth noting that the connections of an DRNN are not only among different layers, but are also connections from a neuron to itself over time. This means that the error propagation for weight adjustment occurs, not only among different nodes, but also between the same node and different time steps, as shown in Fig.~\ref{fig:RNN}. \begin{figure} \centering \includegraphics[width=0.75\textwidth]{images/RNA/RNNunrolled} \caption{ One part of neural network $A$ observes an input $x_t$ and calculates a value $h_t$. The cycle allows information to flow from one step of the network to the next. If we unroll the cycle, one recurrent neural network can be considered as multiple copies of the same network, such that each one passes a message to a successor. } \label{fig:RNN} \end{figure} An LSTM cell is controlled by three gates: the Forget gate, the Input gate and the Output gate. Each gate within the cell is a different neural network that decides what information is allowed in the cell's state, and which functions as the network's memory. The gates can learn what information is relevant to save or forget during the exercise. The Forget gate controls the amount of information that will be stored in the memory and discards irrelevant information. The Input gate controls the amount of new input that will be stored in the memory, i.e., determines the importance of new information. Finally, the Output gateway determines the characteristics of the analyzed information to obtain an output that will allow correct classification. \par The architecture of the neuronal network used to classify RFCA as possible EFOS is made up of three hidden LSTM cell layers each with 256 neurons connecting each neuron in a layer to a neuron in the following layer. The network unrolls over time to analyze all the invoices issued by an RFCA and, from what has been analyzed, classifies it into EFOS or non-EFOS. \par \paragraph{ANN training} The training of the ANN is carried out from the following steps. All the RFCAs previously identified as definitive EFOS are divided into two sets, one with $2,981$ of the RFCAs, referred to as training set, and another one with $745$ EFOS, the so-called test set. The same quantity of unknown RFCAs is added to the test set. One million unknown RFCAs are added to the training set and then the $2,981$ definitive EFOS are added until we obtain the same quantity of unknown RFCAs, ending up with a set of $2,000,000$ RFCAs. Thus, both sets will be made up of 50\% of the data from definitive EFOS, corresponding to the randomly selected EFOS income-type CFDI records, and 50\% of the unknown RFCAs randomly selected from the total population. Data of the associated CFDI is obtained for each RFCA. These are the data that are supplied to the ANN and from which the ANN is trained by adjusting internal parameters. Following the training process, the ANN is presented the set of test data that it has never seen previously, to evaluate its performance. \par \paragraph{Additional variables considered} In addition to incorporating the quantitative variables mentioned in section~\ref{sec:data}, we attempted to incorporate categorical data such as the type and situation of the contributor, the situation-description, the status of the contributor, the start date of operations, the sector and federal entity. We also considered incorporating data related to interactive networks (see section~\ref{sec:redes}), such as the degree of output and input, betweenness, closeness, stress, radiality and page rank. However, all the ANN trained with these variables performed equally or worse than the ANN that only used CFDI data. \subsubsection{Performance evaluation} \newcommand{\textrm{TP}}{\textrm{TP}} \newcommand{\textrm{TP}}{\textrm{TP}} \newcommand{\textrm{FP}}{\textrm{FP}} \newcommand{\textrm{FN}}{\textrm{FN}} \newcommand{\textrm{TN}}{\textrm{TN}} \newcommand{\textrm{TN}}{\textrm{TN}} We used the F1-score~\cite{Rijsbergen:1979:IR} as a measure to evaluate the competence of the trained model. The F1-score is obtained by calculating the precision harmonic mean and the recall. Precision is the proportion of relevant instances correctly classified out of all the instances that the model believes are relevant. If TP are the true positives and FP the false positives, precision would be given by $\textrm{TP}/(\textrm{TP}+\textrm{FP})$ (see Table \ref{tab:MatrizConfusion}). Precision answers the question: {\it How many of the selected RFCAs are actually EFOS?} Recall is the proportion of the incorrectly classified relevant instances out of all the actually relevant instances, $\textrm{TP}/(\textrm{TP}+\textrm{FN})$, where FN are the false negatives, that answers the question of all RFCAs that are actually EFOS: {\it How many were correctly classified?} The harmonic mean is defined as the value obtained when the number of values in the dataset is divided by the sum of its reciprocals. It is a type of mean generally used for numbers that represent a ratio or proportion (like the precision and recall) as it equalizes the weight of each datapoint. An F1-score attains its best value at 1 (perfect precision and recall) and the worst at 0. Table \ref{tab:MatrizConfusion} shows a way of separating the classifications made by the neuronal network to allow their evaluation. \par \begin{table} [th! \centering \begin{tabular}{ cc|c|c } & \multicolumn{3}{c}{{Predicted class }}\\ & & P & N \\ \cline{2-4} \multirow{2}{*}{{Real class}} & P & True positives (TP) & False negatives (FN)\\ \cline{2-4} & N & False positives (FP) & True negatives (TN) \\ \cline{2-4} \end{tabular} \caption{\label{tab:MatrizConfusion} Confusion matrix for binary classification. The true positives (TP) are the examples that the model correctly classified as EFOS. The false negatives (FN) are the examples that the model classified as non-EFOS, but that are actually EFOS. The true negatives (TN) are examples that the model classified as non-EFOS and have not been previously classified as EFOS. The false positives (FP) are examples that the model classified as EFOS, but that were not previously detected as such } \end{table} For example, if we take 500 definitive EFOS and 500 unknown ones, and we feed them to our network, we find that $\textrm{TP} = 448$, $\textrm{FN} = 52$, $\textrm{TN} = 416$ and $\textrm{FP} = 84$. Therefore the precision was $0.845$, the recall was $0.896$, and we obtained an F1-score of $0.87$. If we make the calculation with 1000 alleged EFOS, we obtain $TP = 881$, $FN = 119$ ($TN = FP = 0$ by definition), so the precision is 1, while the recall is $0.881$. We thus obtain an F1-score of 0.94. \par The RFCAs in the set of ``alleged'' show the same behavior that the model indentified by training with the set of ``definitive'', and it ends up identifying 88\% as EFOS. \par In Figure \ref{fig:hists_prob} we observe that the model is sure of its decision most of the time (i.e. ends with a very high or very low probability, with a bimodal distribution). Additionally, in the probability distribution of the non-identified RFCAs, there is a percentage that the model is classifying with high probability (i.e., the model is sure that they are EFOS) but has not previously classified them as EFOS. \par \begin{figure}[th!] \centering \includegraphics[width=\textwidth]{images/Fighistogramas_probablidadesFinal.pdf} \caption{ Histograms of the probabilities assigned by the neural network to different sets of RFCAs. (left) We observe that the network correctly assigns most of them a high probability of being EFOS. (right) We observe a bimodal distribution in which there is a considerable percentage of RFCAs that are assigned a high probability of being EFOS. } \label{fig:hists_prob} \end{figure} One of the greatest challenges in neural networks is to interpret what the network is learning from the data. It is not only important to develop a solid solution with great predictive power; it is also of interest to understand the functioning of the developed model, i.e., which variables are the most relevant, the presence of correlations, possible causal relationships, etc. To deepen our understanding of the results, we applied two techniques to determine the more relevant variables that we detail below. The first technique is based on hypothetical analysis or simulation, and is used to measure the relative influence of the input variables on the model results. In particular, to measure the importance of the variables, we took a sample of our data $X$ and calculated the model predictions once $Y$ is trained. Then, for each variable $x_i$ we cause a perturbation of this variable (and only this variable) via a normal random distribution centered on 0 with a scale 20\% of the variable mean, and calculate the prediction $Y_i$. We measure the effect exerted by this perturbation by calculating the difference of the quadratic root mean between the original output $Y$ and the perturbed $Y_i$. A larger mean square root difference means that the variable is ``more important''. Table \ref{tab:pert_var} (left) shows the five variables of greatest importance for the neuronal network. The second technique consists of the analysis of principal components, a statistical technique to convert high-dimensional to low dimensional data by selecting the most important characteristics that capture most of the information about the dataset. Characteristics are selected as a function of the variation they cause on the output. The characteristic that causes the highest variance is the first component, the characteristic responsible for the second highest variance is considered the second principal component, and so on. It is important to mention that the principal components are not correlated with each other. The importance of each variable is reflected in the magnitude of the corresponding values in vectors that characterize a linear transformation (greater magnitude indicates greater importance). Table~\ref{tab:pert_var} (right) lists the five variables that best characterize the dataset based on the first principal component, which contributes 99\% of the variance. The variable magnitudes are normalized so that the sum of squares is equal to 1. \begin{table}[th! \centering \begin{tabular}{ c | c } Variable & Perturbation effect\\ \hline Sub-active& 0.2099 \\ Active total& 0.1813 \\ Total amount after active & 0.1419 \\ IVA after active & 0.1083 \\ Cancelled amount & 0.0748 \\ \hline \end{tabular} \quad \begin{tabular}{ c | c } Variable & Magnitude \\ \hline Active total & 0.74925125 \\ Sub-active & 0.64598303 \\ Total amount after active & 0.10326791 \\ IVA after active & 0.1032678 \\ Cancelled amount & 0.0000125 \\ \hline \end{tabular} \caption{\label{tab:pert_var} (left) Effect of the perturbation on the probability assigned by the neural network; (right) importance of the variables based on the absolute value of the magnitude of the first principal component used to characterize the dataset. } \end{table} \subsubsection{Model results } The ANN efficiently classifies the identified EFOS that it has been presented with, and using the trained model, we classify as ``suspicious'' the unknown RFCAs to which the ANN assigns a greater probability of behaving similarly to published EFOS. The ANN classified $149,921$ unknown RFCAs, corresponding to 1.98\% of the total, as suspicious, using the threshold probability ($> 0.8$). \subsection{Random Forest} \label{subsec:randomForest} As a second method of classification, we use the automatic technique named {\it Random Forest} (RF). Techniques of automatic classification, including RF, detect groups of elements with similar statistical patterns in an available dataset, and from the knowledge acquired, make decisions about the membership of new elements in these groups. In our case, we consider the characteristics of EFOS published by SAT and we compare them to unknown RFCAs. A RF is constructed by randomly combining different {\it decision trees} in order to obtain results robust to noise sources inherent to the algorithm. A decision tree is a mathematical algorithm made up of a set of questions ordered and connected to each other through their responses (i.e., the formulation of a question depends on the answer to a preceding question). These questions involve the variables or characteristics of the data utilized. In constructing a decision tree each node represents one of the questions and each fork depends on its answer. Thus, in finishing the construction of a decision tree we can follow a path determined by questions and answers, finally answering the main question: how likely is this RFCA to be an EFOS? In statistical models like RF it is necessary to maintain a balance between measures such as the {\it variance} (variability in the prediction of models for different elements) and {\it bias} (the difference between actual and predicted value). To achieve this balance, an effective technique is the combination of various models, e.g., the combination of decision trees to form a RF. In this way, each decision tree issues a classification (i.e., a probability of suspecting it is an EFOS associated to a RFCA) and the final result of the RF is the most probable classification among all the trees constructed. One of the tasks to resolve in constructing a RF is to find the optimum number of decision trees used to determine the combination that generates the final solution. In our case, the RF technique is considered adequate since it offers the following advantages: \begin{itemize} \item Data preparation is minimal. It is only necessary to rely on a dataset where each element to classify, in this case each RFCA, is unique and has a fixed number of characteristics associated with each one of the classes involved, in this case definitive or unknown. \item It can handle a large number of variables without discriminating any one of them. \item It has been demonstrated that it is one of the methods with highest precision among the classification algorithms~\cite{breiman2001random}. \item It performs well with large-volume databases (which applies to the present case study) The result of the RF is a number between 0 and 1 for each evaluated RFCA, which will be interpreted as the probability that each unknown RFCA is a potential EFOS. \end{itemize} \subsubsection{Data preparation} For the implementation of the RF algorithm, the information by issuer is initially grouped, since this analysis is focused on classification of issuing RFCAs. As a result, a unique record for each issuing RFCA is obtained for each of the 48 months considered. Subsequently, by means of a technique called {\it undersampling}~\cite{liu2008exploratory} a balanced sample of unknown RFCA and definitive EFOS is generated. This technique seeks to determine the optimal number of RFCAs that will allow obtaining a balanced sample of the data (i.e., has the same quantity of unknowns and definitives) as well as a representative one (i.e., that will capture the characteristics of the whole population with the number of RFCAs selected. This process yields a sample with 1561 definitive EFOS and 1561 unknown RFCAs. The sample obtained so far is the baseline dataset used for implementation of the RF algorithm. As part of the data preparation phase, two independent treatments are applied to the previously generated sample: \begin{enumerate} \item The data were analyzed to determine what type of data transformation is viable for each of the sample variables. The family of {\it box cox} transformations was used to improve the normality of the data and equalize the variance in order to improve the algorithm’s performance~\cite{osborne2010improving}. \item Principal components analysis (PCA) was used. This consists in reduction of the dimensionality of the dataset by unifying existing variables to create new ones. This method is recommended to improve the performance of the algorithms in question~\cite{wold1987principal}. \end{enumerate} \subsubsection{Model construction} Using the RF algorithm three models were constructed that correspond to the following scenarios and use the sample generated in the previous section: \begin{itemize} \item First scenario: implementation of the RF algorithm without transformation \item Second scenario: implementation of the RF algorithm using the data sample to which PCA was applied. \item Third scenario: implementation of the RF algorithm using the data sample to which the {\it box cox} transformation was applied. \end{itemize} For each of the above scenarios, training of the RF algorithm aims to find the optimum number of constituent decision trees. This is achieved by performing iterations of the algorithm, modifying the number of trees used, and determining when the error generated stabilizes at a minimum value. It was concluded that the optimal number of decision tress was 100. \subsubsection{Performance evaluation} The following measures were used to evaluate the above scenarios: \begin{itemize} \item Receiver operating characteristic (ROC) curve: is a performance measure with values between 0 and 1; the higher the value, the better the performance is considered. An ROC curve is constructed using the information from two characteristics: the sensitivity (possibility of appropriately classifying a positive individual, in this case a definitive EFOS) and the specificity (possibility of appropriately classifying a negative individual, in this case an unknown RFCA that is not actually a definitive RFCA)~\cite{martinez2003curva}. \item Error: is a penalty measure. The closer it is to 0, the better it is considered. The error quantifies the part of the model that is making a mistake in classifying the RFCAs, and in the case of % a RF is obtained through a combination of the error generated by each one of the individual trees, as well the correlation between them~\cite{breiman2001random}. \end{itemize} \begin{table} [th! \centering \begin{tabular}{ l|c|c } {Escenario} & {ROC} & {Error} \\ \hline Random forest & 0.912 & 0.164 \\ Random forest plus principal components& 0.886 & 0.161 \\ Random forest plus variable transformation & 0.893 & 0.157 \\ \hline \end{tabular} \caption{\label{tab:machine_learning} Comparison of performance measures for the different ways in with the input data were transformed.} \label{tab:BAmodelos} \end{table} As shown in Table~\ref{tab:BAmodelos}, even though there is an improvement in the performance for the first scenario, error reduction is favored; therefore the model selected was the one that included the Box Cox transformation of data. This was the model used in the following steps. \begin{table} [th! \centering \begin{tabular}{ c|c|c|c|c } \multicolumn{1}{c}{}& \multicolumn{4}{c}{Years with activity} \\\hline Years classified as EFOS & 1 year & 2 years & 3 years & 4 years \\ \hline 0 & 17\% (133)& 5\% (56) & 3\% (11)& 6\% (8) \\ 1 & 83\% (631)& 13\% (143)& 6\% (24)& 4\% (6) \\ 2 & & 82\% (893)& 17\% (71)& 11\% (16) \\ 3 & & & 74\%(307)& 26\% (37) \\ 4 & & & & 53\% (77) \\ \hline \end{tabular} \caption{\label{tab:efos_anual} We study the performance of the RF algorithm over the years. We considered the definitive EFOS, separated by the number of years with activity (columns). In the different rows, we consider the number of years in which the algorithm classifies RFCAs as EFOS; thus, a definitive EFOS should be detected by the algorithm in at least one of the years of activity. For example, RF erroneously classified 3\% out of the total definitive EFOS with activity reported over 3 years, which corresponds to 11 definitive EFOS. } \end{table} An additional validation was conducted considering the selected model, which consisted of classifying the definitive EFOS using the model (which we know {\it a priori} should have have a high probability), and observing the outcome. A cut-off point of 0.8 was established, i.e., if the risk index is greater or equal than 0.8, the classified RFCA is considered an EFOS, otherwise it is not. Additionally, the years of activity of each definitive EFOS were considered for the final diagnosis, e.g., if it was active for two years, the two qualifications are considered, and so on. Results shown in Table \ref{tab:efos_anual} were obtained by developing the above, where it can be observed that close to 92\% of the definitive EFOS are being correctly classified by the algorithm, and the error is only 8\%. \begin{table} [th! \centering \begin{tabular}{ l|c|c } {Classification} & {Frequency} & {Percentage} \\ \hline EFOS & 1,908 & 79\% \\ No EFOS & 505 & 21\% \\ \hline \end{tabular} \caption{\label{tab:calificacion} Classifying the different types of contributors. } \end{table} Combining the above results, those RFCAs that in all years of activity were detected by the model were classified as possible EFOS, and as non-EFOS if not. Table \ref{tab:calificacion} shows that out of all definitive EFOS, only 505 were classified as non-EFOS, which means that they are the only ones about which the algorithm is completely wrong. This behavior is considered normal due to the possibility that the EFOS may have engaged in illegal activities only in some years. \subsubsection{Results} Using the model developed and validated in the previous sections (third scenario), we take four groups of unknown RFCAs (one per study year) and determine the risk index. Note that if the RFCA has been active for more than one year, it will have a different index each year. Based on the previous results the following groups were defined for all unknown RFCAs: \begin{itemize} \item Suspicious: all those unknown RFCAs that in each year of activity have a risk index greater or equal than $0.8$. \item Not suspicious: all unknown RFCAs that in at least one of the years of activity have a risk index $< 0.8$. \end{itemize} Based on these definitions the algorithm classified 7,438,448 RFCAs (98.3\%) as not suspicious and 128,227 RFCAs (1.7\%) as suspicious of being EFOS.

e24b5b0dbee6110a53952c70bb52b1f102723f00

\section{Introduction} \label{Sec:Introduction} Modern neural networks are typically trained on big data in a supervised fashion. This is commonly achieved by minimizing a loss function that measures the distance between network output and teaching input in several iterations so that the network output approaches the training input. A widely used method to do this is backpropagation, which descends along the gradient of the loss function by propagating weight changes through the network~\cite{rumelhart1986learning,bengio2012practical,lecun2012efficient}. While backpropagation has been applied very successfully, there are still several open questions requiring a definite answer. In particular, the type of loss function to be used and the optimal way of following the gradient are still open problems. A variety of loss functions has been used in the literature to judge the quality of a network output. Similarly, many optimization strategies have been investigated to let gradient descent converge faster or avoid local minima, including stochastic gradient descent, second order methods, and others~\cite{bengio2012practical,sutskever2013importance}. However, most design decisions are still largely based on empirical experiments and experience. This paper presents a theoretical study of one of the most popular loss functions, cross-entropy. Technically, the paper is an advancement of an approach developed in~\cite{jaeger2013neurological, jaeger2020arXiv}. Specifically, it discusses the theoretical ramifications when cross-entropy is minimized by two dual processes, which minimize the Kullback–Leibler divergence and the Shannon entropy, respectively. The advantage of this approach is that it leads to a loss function and weight space for which learning rate and momentum weight can be derived theoretically. These regularization parameters are important parameters for weight updates during backpropagation and thus have a strong influence on gradient decent. Although rules of thumb and effective heuristics exist for choosing these parameters, including dynamic parameter updates during backpropagation, they have eluded a conclusive theoretical analysis so far. The theoretical values derived in the following are similar to empirical values in the literature. Another advantage of the dual process model proposed in this paper is that it includes the golden ratio and complex numbers, which are both novel concepts in machine learning. Complex numbers in particular seem to be largely ignored by the vast majority of publications in computational machine learning. However, incorporating complex numbers may be necessary to achieve a full understanding of neural networks, in view of brain waves or similarities with physical systems, for example. The model proposed in this paper is an attempt at doing so. The paper is structured as follows: Section~\ref{Sec:MotivationApproach} motivates the paper, discussing intrinsic uncertainty in human perception. Section~\ref{Sec:CrossEntropy} recalls the well-known definitions of cross-entropy, Shannon entropy, and Kullback–Leibler divergence. Then, Section~\ref{Sec:GoldenRatio} lists the mathematical features of the golden ratio, borrowing text from an earlier work~\cite{jaeger2020arXiv}, before Section~\ref{Sec:IntrinsicUncertainty} formalizes the intrinsic uncertainty between observed and true probabilities. Based on these results, Section~\ref{sec:LossFunction} will derive a new loss function. Next, Section~\ref{Sec:ProcessModel} will present the dual process model from which Section~\ref{Sec:Regularization} will then derive the two regularization parameters, learning rate and momentum weight. Finally, a discussion and conclusion will summarize the results of the paper. \section{Motivation and Approach} \label{Sec:MotivationApproach} The main motivation of the approach developed in this paper lies in the observation that human perception is fraught with intrinsic uncertainty. A good example is Rubin's Vase, an ambiguous visual perception made popular by the Danish psychologist Edgar Rubin around 1915, which is shown in Figure~\ref{RubinVaseFigure}~\cite{rubin2015}. \begin{figure}[ht] \vskip 0.2in \begin{center} \centerline{\includegraphics[width=\columnwidth]{Facevase.png}} \caption{Rubin's Vase} \label{RubinVaseFigure} \end{center} \vskip -0.2in \end{figure} This figure could be interpreted either as a vase, or as two faces looking at each other, depending on what is considered background and foreground. One interpretation is the complement of the respective other. To develop this idea further in a more formal approach, let $\mathcal{X}$ be a discrete random variable with two possible values, $\mathcal{X} = \left\{x, \neg x \right\}$. Furthermore, let $\mathrm{P} (\mathcal{X})$ be a probability mass function on $\mathcal{X}$ that assigns probability values as follows: $\mathrm{P}(x) = p$, and $\mathrm{P}(\neg x) = 1-p$. Assuming that $\mathcal{X}$ underlies the intrinsic uncertainty of our perception, viewers do not know whether they observe $x$ or $\neg x$, or in information-theoretical terms, they do not know whether the information content of their observation is $-\ln(p)$ or $-\ln(1-p)$. Assuming, without loss of generality, that $p$ is the true probability, viewers do not know whether the information they can expect is $-p \cdot \ln(p)$ or $-p \cdot \ln(1-p)$. Only if both terms are equal, which is the case for $p=0.5$ when $p \in \; ]0\,;1[\,$, is there no uncertainty between them: \begin{equation} -p \cdot \ln(p) = -p \cdot \ln(1-p), \label{motivationEquation1} \end{equation} or equivalently: \begin{equation} 0 = -p \cdot \ln\bigg(\frac{1-p}{p}\bigg) \label{motivationEquation2} \end{equation} In Eq.~\ref{motivationEquation2}, the intrinsic uncertainty shows as follows: If an observer knows the value of~$p$, the observer does not know whether the fraction $(1-p)/p$ or its inverse needs to be used as argument to the logarithmic term, and for that matter, does not know the sign of the difference. Conversely, if the observer knows the sign of the difference, the observer cannot know the value of~$p$, which could be either~$p$ or~$1-p$. Eq.~\ref{motivationEquation2} also shows the basic equation structure of the Kullback–Leibler divergence, which will be discussed in the next section and Section~\ref{Sec:IntrinsicUncertainty}. \section{Cross Entropy} \label{Sec:CrossEntropy} Cross-entropy is a common loss function used for training of artificial neural networks. It quantifies the difference between two probability distributions, say~$p$ and~$q$. In communication theory, it measures the average number of bits needed to encode data coming from a source with distribution~$p$, when the model (encoding) is optimized for an estimated probability distribution~$q$, rather than the true distribution~$p$. Mathematically, for discrete probability distributions~$p$ and~$q$, defined on the same probability space $\mathcal{X}$, the cross-entropy~$H(p,q)$ is computed as follows: \begin{equation} H(p,q) = -\sum_{x \in \mathcal{X}} p(x) \cdot \ln\Big( q(x)\Big) \label{crossEntropyEquation} \end{equation} The cross-entropy~$H(p,q)$ can also be expressed as the sum of the Kullback–Leibler divergence~$D(p,q)$ from~$p$ to~$q$ and the Shannon entropy~$H(p)$: \begin{equation} H(p,q) = D(p,q) + H(p) \label{crossEntropySumEquation} \end{equation} The next subsections will briefly write out the definitions for these two information measures, including the special case of a two-valued random variable~$\mathcal{X}$ with outcome probabilities~$p$ and~$1-p$. \subsection{Kullback–Leibler divergence} The Kullback–Leibler divergence~$D_{KL}(p,q)$ describes the difference between a probability distribution~$p$, say a measured observation, and a second probability distribution, $q$, serving as a reference or model distribution. The Kullback–Leibler divergence can then be interpreted as the average difference of the number of bits required for encoding samples of~$p$ using the optimal encoding given by~$q$ (rather than the optimal coding for~$p$). For the two distributions, $p$ and~$q$, the Kullback–Leibler divergence~$D_{KL}(p,q)$ is then defined as follows \begin{equation} D_{KL}(p,q) = -\sum_{x \in \mathcal{X}} p(x) \cdot \ln \Bigg(\frac{q(x)}{p(x)}\Bigg) \label{KullbackLeiblerEquation} \end{equation} for a probability space $\mathcal{X}$. In the specific case of a two-valued random variable~$p$, the Kullback–Leibler divergence $D_{KL}(p,\neg p)$ between~$p$ and its complement~$1-p$ therefore computes as follows \begin{equation} D_{KL}(p,\neg p) = -p \cdot \ln \Bigg(\frac{1-p}{p}\Bigg) - (1-p) \cdot \ln \Bigg(\frac{p}{1-p}\Bigg) \label{KullbackLeiblerEquation2} \end{equation} For this specific case, the divergence $D(p) = D_{KL}(p,\neg p)$, assumes its minimum of zero when both distributions $p$ and $\neg p$ are identical, and thus when both outcomes of the random variable have a probability of~$0.5$. \subsection{Shannon Entropy} The Shannon entropy, or simply entropy, of a random variable $\mathcal{X}$ is the average information conveyed by its possible outcomes~\cite{shannon1948mathematical}. In mathematical terms, the entropy $H(\mathcal{X})$ of~$\mathcal{X}$ can be computed as follows: \begin{equation} H(\mathcal{X}) = -\sum_{x \in \mathcal{X}} p(x) \cdot \ln\Big(p(x)\Big), \label{entropyEquation} \end{equation} where~$p(x)$ is a probability distribution defined on all outcomes~$x$ of~$\mathcal{X}$. For a random variable~$\mathcal{X}$ with two possible outcomes, with probabilities~$p$ and $1-p$, the entropy thus computes as follows: \begin{equation} H(\mathcal{X}) = -p \cdot \ln(p) - (1-p) \cdot \ln(1-p) \label{entropyEquation2} \end{equation} In this case, contrary to the Kullback–Leibler divergence, the entropy assumes its maximum (not its minimum) when the outcome and its complement have the same probability, namely $p=0.5$. \section{Golden Ratio} \label{Sec:GoldenRatio} This section highlights a connection between Eq.~\ref{motivationEquation2} and the golden ratio, as shown in~\cite{jaeger2020arXiv}. In Eq.~\ref{motivationEquation2}, the argument to the logarithm, $(1-p)/p$, can be regarded as the perceived (or measured) probability, whereas the multiplier~$p$ is the true probability. If an observer wants to measure the true probability, both terms need to be equal, meaning \begin{equation} p = \frac{1-p}{p} \label{goldenRatioEquation} \end{equation} This holds true if~$p$ is the golden ratio~\cite{livioGoldenRatioBook}, as discussed next. Eq.~\ref{goldenRatioEquation} is equivalent to the following quadratic equation: \begin{equation} p^2 + p - 1 = 0, \label{plusGoldenRatioEquation} \end{equation} which has two irrational solutions~$p_1$ and~$p_2$: \begin{equation} p_1 = \frac{\sqrt{5}-1}{2} \approx 0.618, \label{p1Equation} \end{equation} and \begin{equation} p_2 = \frac{-\sqrt{5}-1}{2} \approx -1.618 \label{p2Equation} \end{equation} An important feature of both solutions is that their complement, $1-p$, equals their square, \begin{equation} 1-p = p^2 \label{complementPlusEquation} \end{equation} Alternatively, the golden ratio can be derived from another quadratic equation, which may be more commonly found in textbooks, and which results from replacing $p$ by $-p$ in Eq.~\ref{plusGoldenRatioEquation}: \begin{equation} p^2 - p - 1 = 0 \label{minusGoldenRatioEquation} \end{equation} This equation also has two irrational solutions, which are the negatives of~$p_1$ and~$p_2$: \begin{equation} -p_1 \approx -0.618 \mbox{\quad and} -p_2 \approx 1.618 \label{minusp1p2Equation} \end{equation} However, for both solutions of the second quadratic equation, Eq.~\ref{minusGoldenRatioEquation}, the complement $1-p$ is the negative reciprocal: \begin{equation} 1-p = -\frac{1}{p} \label{complementMinusEquation} \end{equation} As mentioned in~\cite{jaeger2020arXiv}, the sum of both solutions for Eq.~\ref{plusGoldenRatioEquation} and Eq.~\ref{minusGoldenRatioEquation} is either minus one or one, respectively: \begin{equation} p_1 + p_2 = -1 \mbox{\quad and \quad} -p_1 - p_2 = 1 \label{sumOneEquation} \end{equation} The literature sometimes uses the letter $\varphi$ for the golden ratio, and typically defines the golden ratio as a single value, often with $\varphi \approx 1.618$ and discarding negative values. In this paper, all four solutions to the quadratic equations above will be referred to as the golden ratio. \section{Intrinsic Uncertainty} \label{Sec:IntrinsicUncertainty} For Eq.~\ref{motivationEquation2}, the previous section showed that the measured probability equals the true probability in the golden ratio. This section develops Eq.~\ref{motivationEquation2} in a way that allows computing all possible measurements in a systematic way~\cite{jaeger2013neurological, jaeger2020arXiv}. By coupling the measured probability with the true probability, according to the relationship in~Eq.\ref{goldenRatioEquation}, and using the letter~$E$ (Energy) for the information difference, Eq.~\ref{motivationEquation2} can be developed as follows: \begin{eqnarray} E & = & -p \cdot \ln\bigg(\frac{1-p}{p}\bigg) \label{computationalModelEquation}\\ & \Leftrightarrow & -p^2 \cdot \ln\Big(1-p\Big) \\ & \Leftrightarrow & -p \cdot \ln\Big(1-p^2\Big) \\ & \Leftrightarrow & -p \cdot \ln\Big(\sqrt{1-p^2}\Big) \cdot 2 \label{squareRootPythagoreanEquation}\\ & \Leftrightarrow & -\sin(\phi) \cdot \ln\big(\cos(\phi)\big) \cdot 2, \label{pythagoreanEquation} \end{eqnarray} where the last expression holds for an angle~$\phi \in \big[0\,;\frac{\pi}{2}\big]$. Varying~$\phi$ in this range will produce all possible measurements, which are points on the unit circle. Using this scheme, the measured probability equals the true probability for $\phi = \pi/4$, and a measured probability of $\sin(\pi/4) = \cos(\pi/4) = 1/\sqrt{2}$. The letter~$E$ for energy is used in Eq.~\ref{computationalModelEquation} to emphasize that information is related to energy in the physical sense and also to emphasize that this energy, or information, can be absorbed or released, depending on the sign. In Eq.~\ref{pythagoreanEquation}, the energy~$E$ attains its minimum of zero for $\phi=0$, when $\sin(\phi)=0$ and $\cos(\phi)=1$. On the other hand, $E$ reaches its maximum, infinity, for $\phi=\pi/2$, when $\sin(\phi)=1$ and $\cos(\phi)=0$. Now, a dual energy for a second observer can be computed by swapping $\sin$ for $\cos$ in Eq.~\ref{pythagoreanEquation}, which amounts to using the main diagonal as a mirror axis. This dual energy will reach its maximum, when the original energy uses its minimum; and vice versa, it will be minimum when the original energy is maximum. Most importantly, it is possible to establish a formal analogy to the intrinsic uncertainty in observations, as motivated in Section~\ref{Sec:MotivationApproach}. For this, let there be two dual and intertwined processes based on the dual energies above. While one process considers all energies for $\phi > \pi/4$ as released, and all energies for $\phi < \pi/4$ as absorbed; its dual counterpart considers energies for $\phi > \pi/4$ as absorbed, and all energies for $\phi < \pi/4$ as released. The intrinsic uncertainty then shows as follows: If one process knows whether energy is released or absorbed, it does not know the magnitude of the energy. Conversely, if a process knows the magnitude of the energy, it does not know whether the energy is released or absorbed. Only when both processes work hand in hand, synchronously, can there be knowledge about the direction and magnitude of energy. However, an observer can truly measure only one or the other, which shows a connection to Heisenberg's uncertainty principle~\cite{jaeger2013neurological}. Nevertheless, an observer can determine the degree of uncertainty in a measurement by varying the angle $\phi$. For example, the observer could choose an angle of $\phi = \pi/4$ for which the observer may know the magnitude of the energy but not its direction, and then vary the angle to learn more about the direction at the expense of knowing about magnitude. Mathematically, this regularization can be achieved by a multiplier that regulates the energy~$E$, as shown in the following modification of Eq.~\ref{computationalModelEquation}: \begin{equation} Ei = -p \cdot \ln\bigg(\frac{1-p}{p}\bigg) \label{computationalModelEquation-i} \end{equation} Here, the term~$i$ is a placeholder for a parameter regulating~$E$. The remainder of this paper will discuss how a neural network can use this parameter for learning. Moreover, choosing~$i$ as the character for the parameter is not by chance. Section~\ref{Sec:ProcessModel} will consider this to be a complex number and will take advantage of the fact that~$i^2=-1$. \section{Loss function} \label{sec:LossFunction} This section will develop a loss function based on the theoretical results derived above. Section~\ref{Sec:IntrinsicUncertainty} showed that for an angle of $\phi = \pi/4$, one of the two dual intertwined processes will know the quantity of the input, but not the sign. Typically, for a machine learning task, the goal is to match the magnitude of the teaching input with the output of a network. The sign of the teaching input is ignored, or tacitly assumed to be positive. As discussed above, there is no way of knowing both the magnitude and the sign of the teaching input. Therefore, assuming a binary teaching input~$t$, let the difference function $d(y,t)$ between~$t$ and network output~$y$ be defined as follows: \begin{equation} d(y,t) = (y-t+1) \cdot \frac{\pi}{4} \label{differenceEquation} \end{equation} This function will output values in the interval $\big[0;\frac{\pi}{2}\big]$, and will be equal to one of the outer boundaries of the interval when the difference between network output and teaching input is maximum, with $|y-t|=1$. In case the network output is identical to the teaching input, $d(y,t)$ will be in the middle of the interval, $d(y,t)=\pi/4$. The angle defined by $d(y,t)$ can now be used as $\phi$, and its cosine as the observed input. Next, resolving Eq.~\ref{computationalModelEquation} for the true probability~$p$ leads to the following sigmoidal expression for~$p$: \begin{equation} p = \frac{1}{1+\exp\left(-E/p\right)} \label{sigmoidEquation} \end{equation} In this expression, the denominator~$p$ in the argument to the exponential function can be written as~$\sin(d)$, according to Eq.~\ref{pythagoreanEquation}. Furthermore, the energy~$E$ can be computed as the Kullback-Leibler divergence based on an input probability of~$\cos^2d$, using Eq.~\ref{KullbackLeiblerEquation2}. Here, the cosine value is squared to include the multiplication by two in Eq.~\ref{pythagoreanEquation}. Performing these substitutions in Eq.~\ref{sigmoidEquation} produces the following sigmoid function: \begin{equation} p = \frac{1}{1+\exp\left(-D(\cos^2(d))/\sin(d)\right)} \label{sigmoidEquationSinCos} \end{equation} With the difference between network output and teaching input defined according to Eq.~\ref{differenceEquation}, this function will output values in the interval $\left[\frac{1}{2}\,;1\right]$. When the network output is equal to the teaching input, the Kullback-Leibler divergence~$D$ in Eq.~\ref{sigmoidEquationSinCos} will be zero ($\cos^2(d)=0.5$), and the function will thus attain its minimum value of $0.5$. Therefore, the function defined by Eq.~\ref{sigmoidEquationSinCos} can be used as a loss function, which reaches its minimum of $0.5$ if, and only if, the Kullback-Leibler divergence is zero, assuming a radial distance measure between network output and teaching input, as defined in Eq.~\ref{differenceEquation}. \section{Dual Process Model} \label{Sec:ProcessModel} Section~\ref{sec:LossFunction} showed that minimizing the Kullback-Leibler divergence towards~$p=0.5$, and thus maximizing the entropy, allows learning the magnitude of the teaching input. However, according to the theory set forth here, there is a dual process that maximizes the Kullback-Leibler divergence, and thus minimizes the entropy, to learn the sign of the teaching input. Both processes run in parallel, and are intertwined, with the output of one process being the input to the respective other process. They are complementary in the sense that their underlying true probabilities are complements, either~$p$ or $1-p$, and so are their measurements, as motivated in Section~\ref{Sec:MotivationApproach}. To illustrate this principle, Figure~\ref{dualProcessFigure} shows the transition scheme between different equilibrium states from the viewpoint of a process. \begin{figure*}[ht] \vskip 0.2in \begin{center} \centerline{\includegraphics[width=\textwidth]{DualProcessModel}} \caption{Equilibrium states and transition scheme of the dual process model, with the golden ratio $\varphi \approx 1.618$} \label{dualProcessFigure} \end{center} \vskip -0.2in \end{figure*} Each transition from one corner to another in the diagram in Figure~\ref{dualProcessFigure}, represents either a change in the underlying complementary, true probability, or a change in measurement, meaning measuring the complement instead. For example, the equation in the lower-left corner in Figure~\ref{dualProcessFigure} is related to Eq.~\ref{pythagoreanEquation}. This becomes clear when looking at the following equation: \begin{equation} \frac{Ei}{2} = -\frac{1}{\sqrt{2}} \cdot \ln\left(\frac{1}{\sqrt{2}}\right) \label{GoldenRatioPythagoreanEquation} \end{equation} Eq.~\ref{GoldenRatioPythagoreanEquation} is almost identical to Eq.~\ref{pythagoreanEquation}, with $\phi = \pi/4$, except that the complex term~$i$ has been added, similar to Eq.~\ref{computationalModelEquation-i}. Thus, Eq.~\ref{GoldenRatioPythagoreanEquation} describes an equilibrium state in which the amount of released energy equals the amount of absorbed energy. According to Section~\ref{Sec:IntrinsicUncertainty}, this is the state where the measured probability equals the true probability, which is the case for the golden ratio. Therefore, the term $1/\sqrt{2}$ needs to be multiplied by the constant $\sqrt{2} \cdot p_1$, where $p_1$ is the value of the golden ratio as defined in Eq.~\ref{p1Equation}, in order to obtain the golden ratio in a state of equilibrium. Expressing the minus sign with the inverse argument of the logarithm, in Eq.~\ref{GoldenRatioPythagoreanEquation}, and writing $p_1$ as $1/\varphi$, with $\varphi \approx 1.618$, produces the equation in the lower-left corner of Figure~\ref{dualProcessFigure}: \begin{equation} \frac{Ei}{2} = \frac{\ln(\varphi)}{\varphi} \label{GoldenRatioEquilibriumEquation} \end{equation} By going up vertically in the diagram in Figure~\ref{dualProcessFigure}, reaching the equation of the dual process in the upper-left corner, the original input, $\varphi$, becomes what can be considered the output of the dual process, and the original output, $1/\varphi$, is now the input, or measurement, of the dual process. The dual process is based on the complement of the true probability, $1-p$, thus the double arrow in Figure~\ref{dualProcessFigure}. Furthermore, the energy in the dual process flows in the opposite direction. Therefore, a multiplication by the complex number~$i$ introduces a minus sign for the energy and brings the complex number to the other side of the equation ($i^2=-1$). Moving further along to the upper-right corner of Figure~\ref{dualProcessFigure} by measuring the complement of the input, which involves squaring the input according to Eq.~\ref{complementPlusEquation}, returns to the energy computation of the original process. Therefore, a multiplication by~$i$ removes the minus sign and brings~$i$ back to the other side of the equation. Taking the complement twice in a row, once for the true probability ($1-p$) and once for the measured input ($p^2$), is equivalent to looking at the same process from a different point of view. The next section will show how to exploit this fact for machine learning. The dashed lines in Figure~\ref{dualProcessFigure} show the corresponding state transitions for the opposing dual process, which uses the relation in Eq.~\ref{complementMinusEquation} to compute the complement of its input ($p=-1/p$). \section{Regularization} \label{Sec:Regularization} A training method based on backpropagation adapts the network weights in a way that minimizes the loss, meaning the difference between network output and teaching input~\cite{lecun2012efficient}. Using gradient descent, training implies computing the gradient of a loss function~$L$, such as the loss given by Eq.~\ref{sigmoidEquationSinCos}, with respect to each network weight. A backpropagation method accomplishes this for one network layer at a time, iteratively, propagating the gradient back from the output layer to the input layer. To move along the gradient towards the minimum of the loss function, a delta is added to each weight, which has the following form, when adding also a momentum term: \begin{equation} \Delta w_{ij}(t) = -\eta \frac{\partial L}{\partial w_{ij}(t)} + \alpha \cdot \Delta w_{ij}(t-1) \label{deltaWithMomentumEquation} \end{equation} In Eq.~\ref{deltaWithMomentumEquation}, $\Delta w_{ij}(t)$ denotes the delta added to each weight $w_{ij}$ between a node $i$ and a node $j$ in the network, at training iteration (or time)~$t$. The term $\partial L/\partial w_{ij}(t)$ is the partial derivative of the loss function with respect to $w_{ij}$, at time~$t$, which is multiplied with the learning rate~$\eta$. The sign of~$\Delta w_{ij}(t)$ is negative so that the loss function approaches its minimum. In practice, a momentum term describing the weight change at time~$t-1$, $\Delta w_{ij}(t-1)$, is commonly added. This term is typically multiplied by a weighting factor~$\alpha$, as seen in Eq.~\ref{deltaWithMomentumEquation}. The general conception is that the momentum term improves stochastic gradient descent by dampening oscillations. However, according to the dual process model developed here, the actual reason for the performance improvement brought about by the momentum term lies in the gradient of the dual process, as discussed below. As of yet, a conclusive theory for the optimal values of the learning rate~$\eta$ and the momentum weight~$\alpha$ has been lacking, although second order methods have been tried for example~\cite{bengio2012practical,sutskever2013importance}. Both parameters are often determined heuristically, either through empirical experiments or through search. Training results can be very sensitive to the value of the learning rate. For example, a small learning rate may produce a slow convergence, whereas a larger learning rate may result in the search passing over the minimum loss. Negotiating this delicate trade-off in the regularization of the training process can be time-consuming in practical applications. Literature seems to prefer learning rates around~$0.01$, although reported values differ by several orders of magnitude. For the momentum weight, higher values around $0.9$ are more common. The proposed dual process model allows deriving theoretical values for both regularization parameters, learning rate~$\eta$ and momentum weight~$\alpha$. Eq.~\ref{GoldenRatioPythagoreanEquation} and Figure~\ref{dualProcessFigure} help to understand this. As explained in Section~\ref{Sec:ProcessModel}, the equilibrium equations in the lower-left and upper-right corner of Figure~\ref{dualProcessFigure}, represent the same process from different point of views. Bringing the factor $1/2$ from the left-hand side of Eq.~\ref{GoldenRatioPythagoreanEquation} to its right-hand side, by squaring the argument of the logarithm, it becomes evident that the measured probability is $1/2$ in the state of equilibrium. Therefore, this process minimizes the Kullback–Leibler divergence and maximizes the entropy. On the other hand, the dual process represented by the equilibrium equations in the upper-left and lower-right corner of Figure~\ref{dualProcessFigure}, does the opposite because it features a minus sign in front of both equations. It maximizes the Kullback–Leibler divergence and minimizes the entropy. A gradient in the dual process model is therefore a composite of the gradients of both processes, involving the gradient of one process and the negative gradient of its dual process. Each summand in the weight adjustment defined by Eq.~\ref{deltaWithMomentumEquation}, namely the partial derivative $\partial L/\partial w_{ij}(t)$ and the momentum term $\Delta w_{ij}(t-1)$, corresponds to a gradient of one of the dual processes. The momentum weight~$\alpha$ follows from the results in Section~\ref{Sec:ProcessModel}. The multiplier for the equilibrium in Eq.~\ref{GoldenRatioPythagoreanEquation}, $1/\sqrt{2}$, is a gradient, when Eq.~\ref{GoldenRatioPythagoreanEquation} is considered a linear function with information input, $-\ln(x)$, and information output, $E/2$. The dual process has the same gradient, but with input and output reversed. Because both dual processes are intertwined, it is fair to assume that the dual process happens at time $t-1$, and that the current process at time~$t$ observes the output of its dual counterpart. Therefore, the multiplier in Eq.~\ref{GoldenRatioPythagoreanEquation} represents the gradient from the previous iteration. This gradient, and thus the delta at $t-1$, $\Delta w_{ij}(t-1)$, needs to be multiplied by a constant to obtain the golden ratio for the state of equilibrium, in Eq.~\ref{computationalModelEquation}, for which the measured probability equals the true probability. As in Section~\ref{Sec:ProcessModel}, this regularization can be computed as follows: \begin{equation} \alpha = \sqrt{2} \cdot p_1 \approx 0.874, \label{momentumWeightEquation} \end{equation} where $p_1$ is again the value of the golden ratio in Eq.~\ref{p1Equation}, which provides the value for the momentum weight $\alpha \approx 0.874$. The learning rate~$\eta$ can be derived from the momentum weight~$\alpha$ by converting the latter to the corresponding value of the dual process, following a processing chain in Figure~\ref{dualProcessFigure} of Section~\ref{Sec:ProcessModel}. For $\phi=\pi/2$, in Eq.~\ref{pythagoreanEquation}, $\sin(\phi)$ becomes one. Therefore, after multiplying with the momentum weight, the multiplier, or true probability, will be~$\alpha$. The true probability of the dual process is then given by the complement of~$\alpha$: $1-\alpha$. To make probabilities consistent with each other, the measured probability of the dual process needs to be squared according to Eq.~\ref{complementPlusEquation} in order to compute its complement. Taking the complement of the true probability and of the observed probability can be understood as looking at the same process from a dual point of view; see also Section~\ref{Sec:ProcessModel}. Consequently, applying these steps to the momentum weight~$\alpha$ results in the following expression for the learning rate~$\eta$: \begin{equation} \eta = (1-\alpha)^2 \approx 0.016 \label{learningRateEquation} \end{equation} This provides the value for the second regularization term, learning rate~$\eta$, with~$\eta \approx 0.016$. \section{Discussion} The previous section has shown how the regularization parameters of the delta learning rule given by Eq.~\ref{deltaWithMomentumEquation}, which are learning rate~$\eta$ and momentum weight~$\alpha$, can be derived from the dual process model proposed in Section~\ref{Sec:ProcessModel}. According to the theoretical framework set forth here, the delta rule combines the gradients of two dual processes. As already mentioned in~\cite{jaeger2020arXiv}, this goes beyond the traditional understanding according to which the momentum term produces a more stable gradient descent by smoothing weight changes over several iterations. While one process minimizes the Kullback–Leibler divergence ($p=0.5$) and maximizes the Shannon entropy ($p=0.5$), its dual counterpart does the opposite, by maximizing the Kullback–Leibler divergence and minimizing the entropy. This becomes evident when looking at Eq.~\ref{pythagoreanEquation} for which the equilibrium is reached when the observed probability inside the logarithm equals the true probability outside the logarithm, with $p=1/\sqrt{2}$. Squaring this value, meaning multiplying with the same value from the dual process for which observed and true probability switch their roles, leads to $p=0.5$. The process that minimizes the Kullback–Leibler divergence ensures that the output equals the training input, while its dual counterpart that minimizes the Shannon entropy ensures that there is no uncertainty in the output. However, only both processes taken together can minimize the cross entropy. Each process alone has limitations. The process minimizing the Kullback–Leibler divergence may know that the output equals the training input, but it cannot know with absolute certainty whether the output should be zero or one. This is why the loss function given by Eq.~\ref{sigmoidEquationSinCos} is defined in such a way that it assumes its minimum for an angle of of~$45^\circ$. On the other hand, the process minimizing the Shannon entropy may know that the output is zero, but it cannot know if this is equal to the intended teaching input. This reveals an inherent problem of the teaching input that has largely gone unnoticed so far in the literature. It is only possible to know either the teaching input signal, zero or one, or the actual teaching input, which could be identical to the teaching input signal or could equally well be its complement. It is only possible to know one or the other, but not both, which is reminiscent of Heisenberg's Uncertainty principle in physics~\cite{jaeger2013neurological}. Under these theoretical considerations, the gradient adjustment by means of the delta learning rule, as defined by Eq.~\ref{deltaWithMomentumEquation}, becomes a composite of two gradient adjustments. On the one hand, the gradient is followed to minimize the Kullback–Leibler divergence. On the other hand, the reversed gradient of the dual process maximizing the Kullback–Leibler divergence is followed to minimize the entropy. This again explains the two different point of views, or directions for a process in Section~\ref{Sec:ProcessModel} and Section~\ref{Sec:Regularization}. After a successful training, both processes together have minimized the cross-entropy. However, their knowledge is distributed among them. While one process has learned to mimic the teaching input, the dual process has learned whether the teaching signal needs to be taken at face value or if it needs to be complemented. The theoretical results in this paper confirm that cross-entropy is a profound loss function. However, rather than using cross-entropy directly as a loss function, it may be more appropriate to use it indirectly, via the sum of Kullback–Leibler divergence and Shannon entropy, following the dual process model laid out in Section~\ref{Sec:ProcessModel}. This can be achieved by using the loss function defined by Eq.~\ref{sigmoidEquationSinCos}, and applying the delta learning rule with momentum, as given by Eq.~\ref{deltaWithMomentumEquation}, with the specific values for learning rate~$\eta$ and momentum weight~$\alpha$ derived in Section~\ref{Sec:Regularization}. \section{Conclusion} This paper presents a theoretical analysis for the minimization of cross-entropy, which is one of the most popular loss functions used in combination with backpropagation for training artificial neural networks. The main result is a model comprising two dual processes, with one process minimizing the Kullback–Leibler divergence and the other process minimizing the Shannon entropy. The golden ratio as well as complex numbers play a major role in this model. Both are novel concepts in machine learning according to current knowledge. Moreover, specific values for learning rate and momentum weight follow from the model. The order of magnitude of their values is very similar to empirical values often used in practical experiments. Therefore, choosing these values for both regularization parameters improves the performance of gradient descent, and may help render an expensive hyper-parameter grid search redundant. \section*{Acknowledgement} This research was supported by the Intramural Research Program of the National Library of Medicine, National Institutes of Health.

b20c69228dcb4ed74a127aba4531687c60a677d7

\section{Introduction}\label{sec:intro} The low-rank matrix recovery problem is to recover an unknown low-rank ground truth matrix from certain measurements. This problem has a variety of applications in machine learning, such as recommendation systems \citep{KBV2009} and motion detection \citep{ZYY2013,FS2020}, and in engineering problems, such as power system state estimation \citep{ZML2018}. In this paper, we consider two variants of the low-rank matrix recovery problem with a general measurement model represented by an arbitrary smooth function. The first variant is the symmetric problem, in which the ground truth $M^* \in \mathbb R^{n \times n}$ is a symmetric and positive semidefinite matrix with $\rank(M^*)=r$, and $M^*$ is a global minimizer of some loss function $f_s$. Then, $M^*$ can be recovered by solving the optimization problem: \begin{equation}\label{eq:symrecorig} \begin{aligned} \min \quad & f_s(M) \\ \st \quad & \rank(M) \leq r, \\ & M \succeq 0, \: M \in \mathbb R^{n \times n}. \end{aligned} \end{equation} Note that minimizing $f_s(M)$ over positive semidefinite matrices without the rank constraint would often lead to finding a solution with the highest-rank possible rather than the rank-constrained solution $M^*$. The second variant of the low-rank matrix recovery problem to be studied is the asymmetric problem, in which $M^* \in \mathbb R^{n \times m}$ is a possibly non-square matrix with $\rank(M^*)=r$, and it is a global minimizer of some loss function $f_a$. Similarly, $M^*$ can be recovered by solving \begin{equation}\label{eq:asymrecorig} \begin{aligned} \min \quad & f_a(M) \\ \st \quad & \rank(M) \leq r, \\ & M \in \mathbb R^{n \times m}. \end{aligned} \end{equation} As a special case, the loss function $f_s$ or $f_a$ can be induced by linear measurements. In this situation, we are given a linear operator $\mathcal A:\mathbb R^{n \times n} \to \mathbb R^p$ or $\mathcal A:\mathbb R^{n \times m} \to \mathbb R^p$, where $p$ denotes the number of measurements. To recover $M^*$ from the vector $d=\mathcal A(M^*)$, the function $f_s(M)$ or $f_a(M)$ is often chosen to be \begin{equation}\label{eq:linearobj} \frac{1}{2}\norm{\mathcal A(M)-d}^2. \end{equation} Besides, there are many natural choices for the loss function, such as a nonlinear model associated with the 1-bit matrix recovery \citep{DPBW2014}. The symmetric problem \eqref{eq:symrecorig} can be transformed into an unconstrained optimization problem by factoring $M$ as $XX^T$ with $X \in \mathbb R^{n \times r}$, which leads to the following equivalent formulation: \begin{equation}\label{eq:symrec} \min_{X \in \mathbb R^{n \times r}}f_s(XX^T). \end{equation} In the asymmetric case, one can similarly factor $M$ as $UV^T$ with $U \in \mathbb R^{n \times r}$ and $V \in \mathbb R^{m \times r}$. Note that $(UP,V(P^{-1})^T)$ gives another possible factorization of $M$ for any invertible matrix $P \in \mathbb R^{r \times r}$. To reduce the redundancy, a regularization term is usually added to the objective function to enforce that the factorization is balanced, i.e., $U^TU=V^TV$ is satisfied \citep{TBSS2016}. Since every factorization can be converted into a balanced one by selecting an appropriate $P$, the original asymmetric problem \eqref{eq:asymrecorig} is equivalent to \begin{equation}\label{eq:asymrec} \min_{U \in \mathbb R^{n \times r},V \in \mathbb R^{m \times r}}f_a(UV^T)+\frac{\phi}{4}\norm{U^TU-V^TV}_F^2, \end{equation} where $\phi>0$ is an arbitrary constant. To handle the symmetric and asymmetric problems in a unified way, we will use the same notation $X$ to denote the matrix of decision variables in both cases. In the symmetric case, $X$ is obtained from the equation $M=XX^T$. In the asymmetric case, $X$ is defined as \[ X=\begin{bmatrix} U \\ V \end{bmatrix} \in \mathbb R^{(n+m) \times r}. \] To rewrite the asymmetric problem \eqref{eq:asymrec} in terms of $X$, we apply the technique in \citet{TBSS2016} by defining an auxiliary function $F:\mathbb R^{(n+m) \times (n+m)} \to \mathbb R$ as \begin{multline}\label{eq:asymtransfunc} F\left(\begin{bmatrix} N_{11} & N_{12} \\ N_{21} & N_{22} \end{bmatrix}\right)=\frac{1}{2}(f_a(N_{12})+f_a(N_{21}^T)) \\ +\frac{\phi}{4}(\norm{N_{11}}_F^2+\norm{N_{22}}_F^2-\norm{N_{12}}_F^2-\norm{N_{21}}_F^2), \end{multline} in which the argument of the function $F$ is partitioned into four blocks, denoted as $N_{11} \in \mathbb R^{n \times n}$, $N_{12} \in \mathbb R^{n \times m}$, $N_{21} \in \mathbb R^{m \times n}$, $N_{22} \in \mathbb R^{m \times m}$. The problem \eqref{eq:asymrec} then reduces to \begin{equation}\label{eq:asymtrans} \min_{X \in \mathbb R^{(n+m) \times r}}F(XX^T), \end{equation} which is a special case of the symmetric problem \eqref{eq:symrec}. Henceforth, the objective functions of the two problems will be referred as to $g_s(X)=f_s(XX^T)$ and $g_a(X)=F(XX^T)$, respectively. The unconstrained problems \eqref{eq:symrec} and \eqref{eq:asymrec} are often solved by local search algorithms, such as the gradient descent method, due to their efficiency in handling large-scale problems. Since the objective functions $g_s(X)$ and $g_a(X)$ are nonconvex, local search methods may converge to a spurious (non-global) local minimum. To guarantee the absence of such spurious solutions, the restricted isometry property (RIP) defined below is the most common condition imposed on the functions $f_s$ and $f_a$ \citep{BNS2016,GJZ2017,ZLTW2018,ZWYG2018,ZJSL2018,ZSL2019,HLB2020,ZZ2020,BL2020-pb,ZBL2021-p,Zhang2021-p}. \begin{definition}[\citet{RFP2010,ZLTW2018}]\label{def:rip} A twice continuously differentiable function $f_s:\mathbb R^{n \times n} \to \mathbb R$ satisfies the \emph{restricted isometry property} of rank $(2r_1,2r_2)$ for a constant $\delta \in [0,1)$, denoted as $\delta$-$\RIP_{2r_1,2r_2}$, if \[ (1-\delta)\norm{K}_F^2 \leq [\nabla^2f_s(M)](K,K) \leq (1+\delta)\norm{K}_F^2 \] holds for all matrices $M,K \in \mathbb R^{n \times n}$ with $\rank(M) \leq 2r_1$ and $\rank(K) \leq 2r_2$. In the case when $r_1=r_2=r$, the notation $\RIP_{2r,2r}$ will be simplified as $\RIP_{2r}$. A similar definition can be also made for the asymmetric loss function $f_a$. \end{definition} The state-of-the-art results on the non-existence of spurious local minima are presented in \citet{ZBL2021-p,Zhang2021-p}. \citet{ZBL2021-p} shows that the problem \eqref{eq:symrec} or \eqref{eq:asymrec} is devoid of spurious local minima if i) the associated function $f_s$ or $f_a$ satisfies the $\delta$-$\RIP_2$ property with $\delta<1/2$ in case $r=1$, ii) the function $f_s$ or $f_a$ satisfies the $\delta$-$\RIP_{2r}$ property with $\delta \leq 1/3$ in case $r>1$. \citet{Zhang2021-p} further shows that a special case of the symmetric problem \eqref{eq:symrec} does not have spurious local minima if iii) $f_s$ is in the form \eqref{eq:linearobj} given by linear measurements and satisfies the $\delta$-$\RIP_{2r}$ property with $\delta<1/2$. The absence of spurious local minima under the above conditions does not automatically imply the existence of numerical algorithms with a fast convergence to the ground truth. In this paper, we will significantly strengthen the above two results by establishing the global linear convergence for problems with an arbitrary rank $r$ and a general loss function $f_s$ or $f_a$ under the almost same RIP assumption as above, i.e., $\delta<1/2$ for the symmetric problem and $\delta<1/3$ for the asymmetric problem. One common approach to establish fast convergence is to first show that the objective function has favorable regularity properties, such as strong convexity, in a neighborhood of the global minimizers, which guarantees that common iterative algorithms will converge to a global minimizer at least linearly if they are initialized in this neighborhood. Second, given the local convergence result, certain algorithms can be utilized to reach the above neighborhood from an arbitrary initial point. Note that randomization and stochasticity are often needed in those algorithms to avoid saddle points that are far from the ground truth, such as random initialization \citep{LSJR2016} or random perturbation during the iterations \citep{GHJY2015,JGNK2017}. In this paper, we deal with the two above-mentioned aspects for the low-rank matrix recovery problem separately. \subsection{Notations and Conventions} In this paper, $I_n$ denotes the identity matrix of size $n \times n$, $A \otimes B$ denotes the Kronecker product of matrices $A$ and $B$, and $A \succeq 0$ means that $A$ is a symmetric and positive semidefinite matrix. $\sigma_i(A)$ denotes the $i$-th largest singular value of the matrix $A$. $\mathbf A=\vect(A)$ is the vector obtained from stacking the columns of a matrix $A$. For a vector $\mathbf A$ of dimension $n^2$, its symmetric matricization $\mat_S(\mathbf A)$ is defined as $(A+A^T)/2$ with $A$ being the unique matrix satisfying $\mathbf A=\vect(A)$. For two matrices $A$ and $B$ of the same size, their inner product is denoted as $\langle A,B\rangle=\tr(A^TB)$ and $\norm{A}_F=\sqrt{\langle A,A\rangle}$ denotes the Frobenius norm of $A$. Given a matrix $M$ and a set $\mathcal Z$ of matrices, define \[ \dist(X,\mathcal Z)=\min_{Z \in \mathcal Z}\norm{X-Z}_F. \] Moreover, $\norm{v}$ denotes the Euclidean norm of a vector $v$. The action of the Hessian $\nabla^2f(M)$ of a matrix function $f$ on any two matrices $K$ and $L$ is given by \[ [\nabla^2f(M)](K,L)=\sum_{i,j,k,l}\frac{\partial^2f}{\partial M_{ij}\partial M_{kl}}(M)K_{ij}L_{kl}. \] \subsection{Summary of Main Contributions} For the local convergence, we prove in Section~\ref{sec:local} that a regularity property named the Polyak--{\L}ojasiewicz (PL) inequality always holds in a neighborhood of the global minimizers. The PL inequality is significantly weaker than the regularity condition used in previous works to study the local convergence of the low-rank matrix recovery problem, while it still guarantees a linear convergence to the ground truth. Hence, as will be compared with the prior literature in Section~\ref{sec:related}, not only are the obtained local regularity regions remarkably larger than the existing ones, but also they require significantly weaker RIP assumptions. Specifically, if $f_s$ satisfies the $\delta$-$\RIP_{2r}$ property for an arbitrary $\delta$, we will show that there exists some constant $\mu>0$ such that the objective function $g_s$ of the symmetric problem \eqref{eq:symrec} satisfies the PL inequality \[ \frac{1}{2}\norm{\nabla g_s(X)}_F^2 \geq \mu(g_s(X)-f_s(M^*)) \] for all $X$ in the region \[ \{X \in \mathbb R^{n \times r} | \dist(X,\mathcal Z) \leq \tilde C\} \] with \[ \tilde C<\sqrt{2(\sqrt 2-1)}\sqrt{1-\delta^2}\sigma_r(M^*)^{1/2}. \] Here, $\dist(X,\mathcal Z)$ is the Frobenius distance between the matrix $X$ and the set $\mathcal Z$ of global minimizers of the problem \eqref{eq:symrec}. A similar result will also be derived for the asymmetric problem \eqref{eq:asymrec}. Based on these results, local linear convergence can then be established. Compared with the previous results, our new results are advantageous for two reasons. First, the weaker RIP assumptions imposed by our results allow them to be applicable to a much broader class of problems, especially those problems with nonlinear measurements where the RIP constant of the loss function $f_s$ or $f_a$ varies at different points. In this case, the region in which the RIP constant is below the previous bounds may be significantly small or even empty, while the region satisfying our bounds is much larger since the radius of the region is increased by more than a constant factor. Second, when the RIP constant is large and global convergence cannot be established due to the existence of spurious solutions, the enlarged local regularity regions identified in this work can reduce the sample complexity to find the correct initial point converging to the ground truth. This has a major practical value in problems like data analytics in power systems \citep{JLSB2021} in which there is a fundamental limit to the number of measurements due to the physics of the network. For the global convergence analysis, in Section~\ref{sec:global}, we first study the symmetric problem \eqref{eq:symrec} with an arbitrary objective and an arbitrary rank $r$ and prove that the objective $g_s$ satisfies the strict saddle property if the function $f_s$ has the $\delta$-$\RIP_{2r}$ property with $\delta<1/2$. Note that this result is sharp, because in \citet{ZJSL2018} a counterexample has been found that contains spurious local minima under $\delta=1/2$. Using the above strict saddle property and the local convergence result proven in Section~\ref{sec:local}, we show that the perturbed gradient descent method with local improvement will find an approximate solution $X$ satisfying $\norm{XX^T-M^*}_F \leq \epsilon$ in $O(\log 1/\epsilon)$ number of iterations for an arbitrary tolerance $\epsilon$. Moreover, the convergence result for symmetric problems also implies the global linear convergence for asymmetric problems under the $\delta$-$\RIP_{2r}$ condition with $\delta<1/3$. \subsection{Assumptions} The assumptions required in this work will be introduced below. To avoid using different notations for the symmetric and asymmetric problems, we use the universal notation $f(M)$ henceforth to denote either $f_s(M)$ or $f_a(M)$. Similarly, $M^*$ denotes the ground truth in either of the cases. \begin{assumption}\label{ass:smooth} The function $f$ is twice continuously differentiable. In addition, its gradient $\nabla f$ is $\rho_1$-restricted Lipschitz continuous for some constant $\rho_1$, i.e., the inequality \[ \norm{\nabla f(M)-\nabla f(M')}_F \leq \rho_1\norm{M-M'}_F \] holds for all matrices $M$ and $M'$ with $\rank(M) \leq r$ and $\rank(M') \leq r$. The Hessian of the function $f$ is also $\rho_2$-restricted Lipschitz continuous for some constant $\rho_2$, i.e., the inequality \[ \abs{[\nabla^2f(M)-\nabla^2f(M')](K,K)} \leq \rho_2\norm{M-M'}_F\norm{K}_F^2 \] holds for all matrices $M,M',K$ with $\rank(M) \leq r$, $\rank(M') \leq r$ and $\rank(K) \leq 2r$. \end{assumption} \begin{assumption}\label{ass:rip} The function $f$ satisfies the $\delta$-$\RIP_{2r}$ property. Furthermore, $\rho_1$ in Assumption~\ref{ass:smooth} is chosen to be large enough such that $\rho_1 \geq 1+2\delta$. \end{assumption} \begin{assumption} The ground truth $M^*$ satisfies $\norm{M^*}_F \leq D$, and the initial point $X_0$ of the local search algorithm also satisfies $\norm{X_0X_0^T}_F \leq D$, where $D$ is a constant given by the prior knowledge (every large enough $D$ satisfies this assumption). \end{assumption} \begin{assumption}\label{ass:regcoeff} In the asymmetric problem \eqref{eq:asymrec}, the coefficient $\phi$ of the regularization term is chosen to be $\phi=(1-\delta)/2$. \end{assumption} Note that the results of this paper still hold if the gradient and Hessian of the function $f$ are restricted Lipschitz continuous only over a bounded region. Here, for simplicity we assume that these properties hold for all low-rank matrices. As mentioned before Definition~\ref{def:rip}, the RIP-related Assumption~\ref{ass:rip} is a widely used assumption in studying the landscape of low-rank matrix recovery problems, which is satisfied in a variety of problems, such as those for which $f$ is given by a sufficiently large number of random Gaussian linear measurements \citep{CP2011}. Moreover, in the case when the function $f$ does not satisfy the RIP assumption globally, it often satisfies RIP in a neighborhood of the global minimizers, and the theorems in this paper can still be applied to obtain local convergence results. For the asymmetric problem, it can be verified that, by choosing the coefficient $\phi$ of the regularization term as in Assumption~\ref{ass:regcoeff}, the function $F$ in \eqref{eq:asymtrans} satisfies the $2\delta/(1+\delta)$-$\RIP_{2r}$ property after scaling (see \citet{ZBL2021-p}). Other values of $\phi$ can also lead to the RIP property on $F$, but the specific value in Assumption~\ref{ass:regcoeff} is the one minimizing the RIP constant. Furthermore, if $M^*=U^*V^{*T}$ is a balanced factorization of the ground truth $M^*$, then \begin{equation}\label{eq:auggroundtruth} \tilde M^*=\begin{bmatrix} U^* \\ V^* \end{bmatrix}\begin{bmatrix} U^{*T} & V^{*T} \end{bmatrix} \in \mathbb R^{(n+m) \times (n+m)} \end{equation} is called the augmented ground truth, which is obviously a global minimizer of the transformed asymmetric problem \eqref{eq:asymtrans}. $\tilde M^*$ is independent of the factorization $(U^*,V^*)$, and \[ \norm{\tilde M^*}_F=2\norm{M^*}_F \leq 2D, \quad \sigma_r(\tilde M^*)=2\sigma_r(M^*). \] We include the proofs of the above statements in Appendix~\ref{app:factor} for completeness. In addition, we prove in Appendix~\ref{app:factor} that the gradient and Hessian of the function $g_s$ in the symmetric problem \eqref{eq:symrec} and those of the function $g_a$ in the asymmetric problem \eqref{eq:asymrec} share the same Lipschitz property over a bounded region. Using the above observations, one can translate any results developed for symmetric problems to similar results for asymmetric problems by simply replacing $\delta$ with $2\delta/(1+\delta)$, $D$ with $2D$, and $\sigma_r(M^*)$ with $2\sigma_r(M^*)$. \section{Related Works}\label{sec:related} The low-rank matrix recovery problem has been investigated in numerous papers. In this section, we focus on the existing results related to the linear convergence for the factored problems \eqref{eq:symrec} and \eqref{eq:asymrec} solved by local search methods. The major previous results on the local regularity property are summarized in Table~\ref{tab:localreg}. In this table, each number in the last column reported for the existing works denotes the radius $R$ such that their respective objective functions $g$ satisfy the $(\alpha,\beta)$-regularity condition \[ \langle\nabla g(X),X-\mathcal P_{\mathcal Z}(X)\rangle \geq \frac{\alpha}{2}\dist(X,\mathcal Z)^2+\frac{1}{2\beta}\norm{\nabla g(X)}_F^2 \] for all matrices $X$ with $\dist(X,\mathcal Z) \leq R$. Here, $\mathcal Z$ is the set of global minimizers, and $\mathcal P_{\mathcal Z}(X)$ is a global minimizer $Z \in \mathcal Z$ that is the closest to $X$. The $(\alpha,\beta)$-regularity condition is slightly weaker than the strong convexity condition, and it can lead to linear convergence on the same region. In Table~\ref{tab:localreg}, we do not include specialized results that are only applicable to a specific objective \citep{JGNK2017,HLZ2020-p}, or probabilistic results for randomly generated measurements \citep{ZL2015}. Moreover, \citet{LL2020,ZCG2020} used the accelerated gradient descent to obtain a faster convergence rate, but their convergence regions are even smaller than the ones based on the $(\alpha,\beta)$-regularity condition as listed in Table~\ref{tab:localreg}. Each number in the last column reported for our results refers to the radius of the region satisfying the PL inequality, which is a weaker condition than the $(\alpha,\beta)$-regularity condition while offering the same convergence rate guarantee. It can be observed that we have identified far larger regions than the existing ones under weaker RIP assumptions by replacing the $(\alpha,\beta)$-regularity condition with the PL inequality. \begin{table*} \centering \renewcommand*{\arraystretch}{1.85} \begin{tabular}{cc>{\centering\arraybackslash}m{2cm}c} \toprule Paper & Objective & Assumption & Radius of Local Regularity Region \\\midrule \citet{BKS2016} & S/G & $f_s$ Convex, \newline $\delta_{2r} \leq \delta$ & $\dfrac{1}{100}\dfrac{1-\delta}{1+\delta}\dfrac{\sigma_r(M^*)}{\sigma_1(M^*)}\sigma_r(M^*)^{1/2}$ \\ \citet{TBSS2016} & S/L & $\delta_{6r} \leq 1/10$ & $\dfrac{1}{4}\sigma_r(M^*)^{1/2}$ \\ \citet{TBSS2016} & A/L & $\delta_{6r} \leq 1/25$ & $\dfrac{1}{4}\sigma_r(M^*)^{1/2}$ \\ \citet{PKCS2018} & A/G & $f_a$ Convex, \newline $\delta_{2r} \leq \delta$ & $\dfrac{\sqrt 2}{10}\sqrt{\dfrac{1-\delta}{1+\delta}}\sigma_r(M^*)^{1/2}$ \\ \citet{ZLTW2021} & A/G & $\delta_{2r,4r} \leq 1/50$ & $\sigma_r(M^*)^{1/2}$ \\\midrule Ours & S/G & $\delta_{2r} \leq \delta$ & $0.91\sqrt{1-\delta^2}\sigma_r(M^*)^{1/2}$ \\ Ours & A/G & $\delta_{2r} \leq \delta$ & $1.29\dfrac{\sqrt{1+2\delta-3\delta^2}}{1+\delta}\sigma_r(M^*)^{1/2}$ \\\bottomrule \end{tabular} \caption{Previous local regularity results for the low-rank matrix recovery problems and the comparison with our results (``S'', ``A'', ``L'' and ``G'' stand for the symmetric case, asymmetric case, linear measurement and general nonlinear function).}\label{tab:localreg} \end{table*} Regarding the existing global convergence results for the low-rank matrix recovery problem, \citet{JGNK2017} established the global linear convergence for the symmetric problem in a very specialized case with $f_s$ being a quadratic loss function, and \citet{TBSS2016} proposed the Procrustes flow method with the global linear convergence for the linear measurement case under the assumption that the function $f_s$ satisfies the $1/10$-$\RIP_{6r}$ property for symmetric problems or the function $f_a$ satisfies the $1/25$-$\RIP_{6r}$ property for asymmetric problems under a careful initialization. \citet{ZWL2015} established the global linear convergence for asymmetric problems with linear measurements under the assumption that $f_a$ satisfies $\delta$-$\RIP_{2r}$ with $\delta \leq O(1/r)$ using alternating exact minimization over variables $U$ and $V$ in \eqref{eq:asymrec}. In addition, the strict saddle property proven in \citet{GJZ2017} leads to the global linear convergence of perturbed gradient methods for the linear measurement case under the $1/10$-$\RIP_{2r}$ assumption for symmetric problems and the $1/20$-$\RIP_{2r}$ assumption for asymmetric problems. Later, \citet{ZLTW2018} proved a weaker strict saddle property under the $1/5$-$\RIP_{2r,4r}$ assumption for symmetric problems with general objectives, while \citet{LZT2017-p} proved the same weaker property under the $1/5$-$\RIP_{2r,4r}$ assumption for asymmetric problems with general objectives and nuclear norm regularization. Our results requiring the $\delta$-$\RIP_{2r}$ property with $\delta<1/2$ for symmetric problems with general objectives and the $\delta$-$\RIP_{2r}$ property with $\delta<1/3$ for asymmetric problems with general objectives depend on significantly weaker RIP assumptions and thus can be applied to a broader class of problems, which is a major improvement over all previous results on the global linear convergence. Besides local search methods for the factored problems, there are other approaches for tackling the low-rank matrix recovery. Earlier works such as \citet{CR2009,RFP2010} solved the original nonconvex problems based on convex relaxations. Although they can achieve good performance guarantees under the RIP assumptions, they are not suitable for large-scale problems. Other approaches for solving the low-rank matrix recovery include applying the inertial proximal gradient descent method directly to the original objective functions without factoring the decision variable $M$ \citep{DHLR2020}. However, it may converge to an arbitrary critical point, while in this paper we show that RIP-based local search methods can guarantee the global convergence to a global minimum. \section{Local Convergence}\label{sec:local} In this section, we present the local regularity results for problems \eqref{eq:symrec} and \eqref{eq:asymrec}, which state that the functions $g_s$ and $g_a$ satisfy the PL inequality locally, leading to local linear convergence results for the gradient descent method. The proofs are delegated to Appendix~\ref{app:localproof}. First, we consider the symmetric problem \eqref{eq:symrec}. The development of the local PL inequality for this problem is enlightened by the high-level idea behind the proof of the absence of spurious local minima in \citet{ZSL2019,ZZ2020,BL2020-pb}. The objective is to find a function $f_s^*$ corresponding to the worst-case scenario, meaning that it satisfies the $\delta$-$\RIP_{2r}$ property with the smallest possible $\delta$ while the PL inequality is violated at a particular matrix $X$. This is achieved by designing a semidefinite program parameterized by $X$ with constraints implied by the $\delta$-$\RIP_{2r}$ property and the negation of the PL inequality. Denote the optimal value of the semidefinite program by $\delta_f^*(X)$. If a given function $f_s$ satisfies $\delta$-$\RIP_{2r}$ with $\delta<\delta_f^*(X)$ for all $X \in \mathbb R^{n \times r}$ in a neighborhood of the global minimizers, it can be concluded that the PL inequality holds for all matrices in this neighborhood. \begin{lemma}\label{lem:localpl} Consider the symmetric problem \eqref{eq:symrec} and an arbitrary positive number $\tilde C$ satisfying \begin{equation}\label{eq:plradius} \tilde C<\sqrt{2(\sqrt 2-1)}\sqrt{1-\delta^2}\sigma_r(M^*)^{1/2}. \end{equation} There exists a constant $\mu>0$ such that the PL inequality \[ \frac{1}{2}\norm{\nabla g_s(X)}_F^2 \geq \mu(g_s(X)-f_s(M^*)) \] holds for all matrices in the region \begin{equation}\label{eq:plneighbor} \{X \in \mathbb R^{n \times r} | \dist(X,\mathcal Z) \leq \tilde C\}, \end{equation} where $\mathcal Z$ is the set of global minimizers of the problem \eqref{eq:symrec}. \end{lemma} In the above, note that $\sigma_r(M^*)$ and thus $\tilde C$ are always positive because $M^*$ is assumed to be rank $r$. Both the $(\alpha,\beta)$-regularity condition used in the prior literature and the PL inequality deployed here guarantee a linear convergence if it is already known that the trajectory at all iterations remains within the region in which the associated condition holds. However, there is a key difference between these two conditions. The $(\alpha,\beta)$-regularity condition ensures that $\dist(X,\mathcal Z)$ is nonincreasing during the iterations under a sufficiently small step size, and thus the trajectory never leaves the local neighborhood. In contrast, the weaker PL inequality may not be able to guarantee this property. To resolve this issue, in our convergence proof we will adopt a different distance function given by $\norm{XX^T-M^*}_F$. By Taylor's formula and the definition of the $\delta$-$\RIP_{2r}$ property, we have \begin{equation}\label{eq:globalrip} \begin{aligned} \frac{1-\delta}{2}\norm{M-M^*}_F^2 &\leq f_s(M)-f_s(M^*) \\ &\leq \frac{1+\delta}{2}\norm{M-M^*}_F^2, \end{aligned} \end{equation} for all matrices $M \in \mathbb R^{n \times n}$ with $\rank(M) \leq r$. Therefore, if $M,M' \in \mathbb R^{n \times n}$ are two matrices such that $f_s(M) \leq f_s(M')$, then the inequality \eqref{eq:globalrip} implies that \begin{equation}\label{eq:levelset} \norm{M-M^*}_F \leq \sqrt\frac{1+\delta}{1-\delta}\norm{M'-M^*}_F. \end{equation} Therefore, the distance function $\norm{XX^T-M^*}_F$ is almost nonincreasing if the function value $g_s(X)$ does not increase. Combining this idea with the local PL inequality proved in Lemma~\ref{lem:localpl}, we obtain the next local convergence result. \begin{theorem}\label{thm:localconvsym} For the symmetric problem \eqref{eq:symrec}, the gradient descent method converges to the optimal solution linearly if the initial point $X_0$ satisfies \[ \norm{X_0X_0^T-M^*}_F<2(\sqrt 2-1)(1-\delta)\sigma_r(M^*) \] and the step size $\eta$ satisfies \[ 1/\eta \geq 12\rho_1r^{1/2}\left(\sqrt\frac{1+\delta}{1-\delta}\norm{X_0X_0^T-M^*}_F+D\right). \] Specifically, there exists some constant $\mu>0$ (which depends on $X_0$ but not on $\eta$) such that \begin{multline}\label{eq:convrate} \norm{X_tX_t^T-M^*}_F \leq (1-\mu\eta)^{t/2}\sqrt\frac{1+\delta}{1-\delta}\norm{X_0X_0^T-M^*}_F, \\ \forall t \in \{0,1,\dots\}, \end{multline} where $X_t$ denotes the output of the algorithm at iteration $t$. \end{theorem} Note that since the left-hand side of \eqref{eq:convrate} is nonnegative, we have $0 \leq 1-\mu\eta \leq 1$. As a remark, although our bound on the step size $\eta$ in Theorem~\ref{thm:localconvsym} seems complex, it essentially says that $\eta$ needs to be small, and the upper bound on the acceptable values of the step size can be explicitly calculated out routinely after all the parameters of the problem are given. Furthermore, using the transformation from asymmetric problems to symmetric problems, one can obtain parallel results for the asymmetric problem \eqref{eq:asymrec} as below. \begin{theorem}\label{thm:localconvasym} Consider the asymmetric problem \eqref{eq:asymrec}. The PL inequality is satisfied in the region \[ \{X \in \mathbb R^{(n+m) \times r} | \dist(X,\mathcal Z) \leq \tilde C\}, \] where $\mathcal Z$ denotes the set of global minimizers and \[ \tilde C<2\sqrt{\sqrt 2-1}\frac{\sqrt{1+2\delta-3\delta^2}}{1+\delta}\sigma_r(M^*)^{1/2}. \] Moreover, local linear convergence is guaranteed for the gradient descent method if the initial point $X_0$ satisfies \[ \norm{X_0X_0^T-\tilde M^*}_F<4(\sqrt 2-1)\frac{1-\delta}{1+\delta}\sigma_r(M^*) \] and the step size $\eta$ satisfies \[ 1/\eta \geq 12\rho_1r^{1/2}\left(\sqrt\frac{1+3\delta}{1-\delta}\norm{X_0X_0^T-\tilde M^*}_F+2D\right). \] \end{theorem} \section{Global Convergence}\label{sec:global} Having developed local convergence results, the next step is to design an algorithm whose trajectory will eventually enter the local convergence region from any initial point. The major challenge is to deal with the saddle points outside the local regularity region. One common approach is the perturbed gradient descent method, which adds random noise to jump out of a neighborhood of a strict saddle point. Using the symmetric problem as an example, the basic idea is to first use the analysis in \citet{JGNK2017} to show that the perturbed gradient descent method will successfully find a matrix $X$ that approximately satisfies the first-order and second-order necessary optimality conditions, i.e., \begin{equation}\label{eq:approxoptimalitycond} \norm{\nabla g_s(X)}_F \leq \kappa, \quad \lambda_{\min}(\nabla^2g_s(X)) \geq -\kappa, \end{equation} after a certain number of iterations where the number depends on $\kappa$. Here, $\lambda_{\min}(\nabla^2g_s(X))$ denotes the minimum eigenvalue of the matrix $\mathbf G$ that satisfies the equation \[ (\vect(U))^T\mathbf G\vect(V)=[\nabla^2g_s(X)](U,V), \] for all $U,V \in \mathbb R^{n \times r}$. The second step is to prove the strict saddle property for the problem, which means that for appropriate values of $\kappa$ the two conditions in \eqref{eq:approxoptimalitycond} imply that $\norm{XX^T-M^*}_F$ is so small that $X$ is in the local convergence region given by Theorem~\ref{thm:localconvsym}. After this iteration, the algorithm switches to the simple gradient descent method. This two-phase algorithm is commonly called the perturbed gradient descent method with local improvement \citep{JGNK2017}, whose details are given by Algorithm~\ref{alg:perturbed} in Appendix~\ref{app:globalproof}. The proofs in this section are also given in Appendix~\ref{app:globalproof}. In this section, we will present two conditions that guarantee the global linear convergence of the above algorithm. For symmetric problems, the next lemma provides the strict saddle property and fulfills the purpose for the second step mentioned above. Its proof is a generalization of the one for the absence of spurious local minima under the same assumption in \citet{Zhang2021-p}. \begin{lemma}\label{lem:strictsaddlelin} Consider the symmetric problem \eqref{eq:symrec} with $\delta<1/2$. For every $C>0$, there exists some $\kappa>0$ such that for every $X \in \mathbb R^{n \times r}$ the two conditions given in \eqref{eq:approxoptimalitycond} will imply $\norm{XX^T-ZZ^T}_F<C$. \end{lemma} The remaining step is to show that the trajectory of the perturbed gradient descent method will always belong to a bounded region in which the gradient and Hessian of the objective $g_s$ are Lipschitz continuous (see Appendix~\ref{app:globalproof}). Combining the above results with Theorem~3 in \citet{JGNK2017}, we can obtain the following global linear convergence result. \begin{theorem}\label{thm:globalconvlin} Consider the symmetric problem \eqref{eq:symrec} with $\delta<1/2$. For every $\epsilon>0$, the perturbed gradient descent method with local improvement under a suitable step size $\eta$ and perturbation size $w$ finds a solution $\hat X$ satisfying $\norm{\hat X\hat X^T-M^*}_F \leq \epsilon$ with high probability in $O(\log(1/\epsilon))$ number of iterations. Here, $\eta$ and $w$ are defined in Algorithm~\ref{alg:perturbed} in Appendix~\ref{app:globalproof}. \end{theorem} In the above theorem, the order $O(\log(1/\epsilon))$ of the convergence rate is determined by the number of iterations spent in the second phase of the algorithm, because the number of iterations in the first phase is independent of $\epsilon$. Note that we only show the relationship between the number of iterations and $\epsilon$, but the convergence rate also depends on the initial point $X_0$ and the loss function $f_s$. Moreover, although not being related to the final convergence rate, Theorem~3 in \citet{JGNK2017} also shows that the number of iterations in the first phase is polynomial with respect to the problem size. For asymmetric problems with arbitrary objectives and rank $r$, if we apply the transformation from asymmetric problems to symmetric problems and replace $\delta$ in Theorem~\ref{thm:globalconvlin} with $2\delta/(1+\delta)$, Theorem~\ref{thm:globalconvlin} immediately implies the following global linear convergence result. \begin{theorem}\label{thm:globalconv} Consider the asymmetric problem \eqref{eq:asymrec} with $\delta<1/3$. For every $\epsilon>0$, the perturbed gradient descent method with local improvement under a suitable step size $\eta$ and perturbation size $w$ finds a solution $\hat X$ satisfying $\norm{\hat X\hat X^T-\tilde M^*}_F \leq \epsilon$ with high probability in $O(\log(1/\epsilon))$ number of iterations. \end{theorem} \section{Numerical Illustration}\label{sec:numerical} \begin{figure*}[t] \centering \subfloat[]{\includegraphics[width=6cm]{numericalsym}}\qquad \subfloat[]{\includegraphics[width=6cm]{numericalasym}} \\ \subfloat[]{\includegraphics[width=6cm]{numericalonebit}}\qquad \subfloat[]{\includegraphics[width=6cm]{numericalonebitlarge}} \caption{The trajectory of the perturbed gradient descent method for solving the low-rank matrix recovery problem. The marker in each figure shows the boundary of the local convergence region provided by Theorem~\ref{thm:localconvsym}. (a) A symmetric linear problem with $r=1$, $n=40$, $p=120$ and $\delta$ estimated to be $0.49$. (b) An asymmetric linear problem with $r=5$, $n=10$, $m=8$, $p=220$ and $\delta$ estimated to be $0.32$. (c) The 1-bit matrix recovery problem with $r=5$, $n=10$. (d) The 1-bit matrix recovery problem with $r=2$, $n=600$.}\label{fig:numerical} \end{figure*} In this section, we conduct numerical experiments to demonstrate the behavior of the perturbed gradient descent algorithm for solving low-rank matrix recovery problems. The linear convergence rate observed for the examples below supports our theoretical analyses in Section~\ref{sec:local} and Section~\ref{sec:global}. In the first experiment, we consider the loss function \eqref{eq:linearobj} induced by a linear operator $\mathcal A$ with \[ \mathcal A(M)=(\langle A_1,M\rangle,\dots,\langle A_p,M\rangle). \] Here, each entry of $A_i$ is independently generated from the standard Gaussian distribution. As shown in \citet{CP2011}, such linear operator $\mathcal A$ satisfies RIP with high probability if the number of measurements is large enough. Since it is NP-hard to check whether the resulting loss function $f_s$ or $f_a$ satisfies the $\delta$-$\RIP_{2r}$ for certain $\delta$, the $\delta$ parameter is estimated as follows: For the symmetric problem \eqref{eq:symrec}, we first generate $10^4$ random matrices $X \in \mathbb R^{n \times 2r}$ with each entry independently selected from the standard Gaussian distribution, and then find the proper scaling factor $a \in \mathbb R$ and the smallest $\delta$ such that \[ (1-\delta)\norm{XX^T}_F^2 \leq \norm{a\mathcal A(XX^T)}^2 \leq (1+\delta)\norm{XX^T}_F^2 \] holds for all generated matrices $X$. The $\delta$ parameter for the asymmetric problem \eqref{eq:asymrec} can be estimated similarly. After that, the ground truth $M^*=XX^T$ or $M^*=UV^T$ is generated randomly with each entry of $X$ or $(U,V)$ independently selected from the standard Gaussian distribution. The initial point is generated in the same way. Figure~\ref{fig:numerical}(a) and (b) show the difference between the obtained solution and the ground truth together with the norm of the gradient of the objective function at different iterations. The convergence behavior clearly divides into two stages. The convergence rate is sublinear initially and then switches to linear when the current point moves into the local region associated with the PL inequality. In Figure~\ref{fig:numerical}(a) and (b), the marker shows the first time when the current point falls into the local convergence region provide in Theorem~\ref{thm:localconvsym} or Theorem~\ref{thm:localconvasym}. It can be seen that these theorems predict the boundary of the transition from a sublinear convergence rate to the linear convergence rate fairly tightly. After this point, $O(\log(1/\epsilon))$ additional iterations are needed to find an approximate solution with accuracy $\epsilon$. On the other hand, the occasion when perturbation needs to be added is rare in practice since it is unlikely for the trajectory to be very close to a saddle point. However, such perturbation is necessary theoretically to deal with pathological cases. Second, we consider the 1-bit matrix recovery \citep{DPBW2014} with full measurements, which is a nonlinear low-rank matrix recovery problem. In this problem, there is an unknown symmetric ground truth matrix $\hat M \in \mathbb R^{n \times n}$ with $\hat M \succeq 0$ and $\rank(\hat M)=r$. One is allowed to take independent measurements on every entry $\hat M_{ij}$, where each measurement value is a binary random variable whose distribution is given by $Y_{ij}=1$ with probability $\sigma(\hat M_{ij})$ and $Y_{ij}=0$ otherwise. Here, $\sigma(x)$ is commonly chosen to be the sigmoid function $\mathrm e^x/(\mathrm e^x+1)$. After a number of measurements are taken, let $y_{ij}$ be the percentage of the measurements on the $(i,j)$-th entry that are equal to $1$. The goal is to find the maximum likelihood estimator for the ground truth $\hat M$, which can be formulated as finding the global minimizer $M^*$ of the problem \eqref{eq:symrec} with \[ f_s(M)=-\sum_{i=1}^n\sum_{j=1}^n(y_{ij}M_{ij}-\log(1+\mathrm e^{M_{ij}})). \] To establish the RIP condition for the function $f_s$ above, consider its Hessian $\nabla^2f_s(M)$ that is given by \[ [\nabla^2f_s(M)](K,L)=\sum_{i=1}^n\sum_{j=1}^n\sigma'(M_{ij})K_{ij}L_{ij}, \] for every $M,K,L \in \mathbb R^{n \times n}$. On the region \begin{equation}\label{eq:onebitregion} \{M \in \mathbb R^{n \times n} | \, \abs{M_{ij}} \leq 2.29, \: \forall i,j=1,\dots,n\}, \end{equation} we have $1/12<\sigma'(M_{ij}) \leq 1/4$, and thus the function $f_s$ satisfies the $\delta$-$\RIP_{2r}$ property with $\delta<1/2$. Note that due to the noisy measurements the global minimizer $M^*$ is not equal to $\hat M$ in general. However, for demonstration purposes we should know $M^*$ a priori, and hence we consider the case when the number of measurements is large enough such that $y_{ij}=\sigma(\hat M_{ij})$ and $M^*=\hat M$. In Figure~\ref{fig:numerical}(c), the ground truth and the initial point are generated randomly in the region \eqref{eq:onebitregion}. Here, we can observe a similar two-stage convergence behavior as in the example with linear measurements. We experiment on the same problem with a larger matrix size $n$ as shown in Figure~\ref{fig:numerical}(d), which also gives similar results. \section{Conclusion} In this paper, we study the local and global convergence behaviors of gradient-based local search methods for solving low-rank matrix recovery problems in both symmetric and asymmetric cases. First, we present a novel method to identify a local region in which the PL inequality is satisfied, which is significantly larger than the region associated with the regularity conditions proven in the prior literature. This leads to a linear convergence result for the gradient descent method over a large local region. Second, we develop the strict saddle property for symmetric problems under the $\delta$-$\RIP_{2r}$ property with $\delta<1/2$. Then, we prove the global linear convergence of the perturbed gradient descent method for symmetric problems under the $\delta$-$\RIP_{2r}$ property with $\delta<1/2$, and the same convergence property can also be guaranteed for asymmetric problems with $\delta<1/3$. Compared with the existing results, these conditions are remarkably weaker and can be applied to a larger class of problems.

89d8bb5f0ebb76370af904cad3d9a263f0f9e6ab

\section{Introduction} Discrete Logarithm problem in field $F_p^*$ , where p is prime ( $ g,\beta \in F_p^*$ ), where $g$ a generator, can be defined as finding x for which $g^x \equiv \beta \pmod {p} $. \\ Traditional methods avaialble for finding discrete logarithm are `baby steps giant steps' \cite{babystepgiantstep} , Silver Pohlig-Hellman \cite {silver-pohilg} and Pollard's rho/Kangaroo Method \cite {pollard} \cite{rho}. Silver-pohlig-hellman being the fastest if prime is smooth, i.e prime factors of p-1 can be easily found and such factors are far smaller then p. Run time of pohlig-hellman method in worst case is $\mathcal{O}(\sqrt{p})$, however what is interesting is that it is a lot more efficient when the order of p is smooth. A smooth order means that the order has prime factors, which are far smaller. If the prime factorisation of $p-1 = p_1^{e_1}p_2^{e_2}\dot{...} p_k^{e_k}$ the complexity is $ \mathcal{O} {(\sum_i^k e^i(\log{p}+ \sqrt{p_i}))} $ Our proposed method uses the idea of calculating the $p_i$th root for all prime factors which eventually leads us to the desired result. The result is obtained directly without the need of combining results for all $p_i^{e_i} $ through Chinese remainder theorem. The run time complexity of proposed method , for prime p whose order is B-smooth is $ \mathcal{O} {(\sum_i^k e^i( \sqrt{ p_i}))}$ , almost same as that of Pohlig-Hellman method. However in practical terms absence of extra work done in Pohlig-Hellman method makes this a better candidate for solving discrete logarithm problem , although there is no theoretical change in the upper bound of the run-time complexity. \section{Basic Background} Let $g,\beta \in F^*_p $ where $g$ is a generator and $ p-1= p_1^{e_1}p_2^{e_2}\dot{...} p_k^{e_k}$. Now by definition , of $g$ being generator , we can always find $x \in \{0,..,p-1\}$ such that $ g^x \equiv \beta$.\\ \\Also , by Fermat's Theorem $ g^{(p-1)} \equiv 1$. In $F^*_p$ , if $i^{th}$ root of any element $\beta$ exists, there are exactly $i$ such roots. Also all such i roots $ \{ \beta_1,\beta_2,..\beta_i\}$ form a cyclic group. Other elements of this group can be generated from $\beta_1$ by successively multiplying it with $g ^{(p-1)/i}$ for any $g$, a generator $ \in F^*_p $. However \section{Algorithm } Given in field $F_p^*$ , where p is prime ( $ g,\beta \in F_p^*$ ), where $g$ a generator and $ p-1= p_1^{e_1}p_2^{e_2}\dot{...} p_k^{e_k}$. Let $\sum_i^k e_i = s(say)$. First we note that in s steps of exponentiation we get $\{\beta,\beta^{p_1}, \dot{...},\beta^{ p_1^{e_1}p_2^{e_2}\dot{...} p_k^{e_{k-1}}}, 1\}$ or $\{\beta, \beta_1,\beta_2,..\beta_{s-1},1\}$. We note that $\beta^{p_1^{e_1}p_2^{e_2}\dot{...} p_k^{e_k}} \equiv 1$. We can write these $s+1$ elements as trace back list.$\{1, \beta_{s-1}, \beta_{s-2},\beta_{s-3},..,\beta_1,\beta \}$ . We know by Fermat's Theorem that $g^{(p-1)}\equiv 1 $. Now idea is to keep taking $p_i^{th}$ , root starting from $g^{(p-1)}$ , so that we trace back the list generated from $\beta$. Basic idea is to generate a list equivalent to trace back list where elements at each positions are equal , by taking the appropriate $p_i^{th} $ root .\\ i.e. $ \{g^{p-1}, g^{x_1}, g^{x_2},..,g^x \} \equiv \{1, \beta_{s-1}, \beta_{s-2},\beta_{s-3},..,\beta_1,\beta \} $, trace back list.\\ $x_1,x_2,..x $ are obtained by taking successive $ p_i^{th}$ root starting from $g^{p-1}$\\ We note that one of the $p_i^{th}$ root of $ g^{(p-1)}$ would be given $g^{(p-1)/p_i}$ and it can be clearly seen that such root exist as $p_i\text{ } |\text{ } (p-1)$.Now we start with by taking $p_k$ root of $ g^{p-1} \equiv 1$ .Let $g^{(p-1)/p_k} \equiv \delta $. Now $ \beta_{s-1} $ would be member of the group of order $p_k$ i.e. \\ $ \beta_{s-1} \in \{ \delta, \delta g^{(p-1)/p_k},\delta g^{2(p-1)/p_k},..\delta g^{(p_k-1)(p-1)/p_k} \} $ \\ or \\ $ \beta_{s-1} \in \{\delta, \delta \epsilon_0,\delta \epsilon_0^2,..\delta \epsilon_0^{p_k-1} \} $ where $ \epsilon_0 \equiv g^{(p-1)/p_k} $. \\ \\$\beta_{s-1} $ can be located in the cyclic group of order $p_k$ through any of the baby step - giant steps method or pollard's rho method in $\mathcal{O}{(\sqrt{p_k})}$ time.\\ Following this first step we will get $g^r \equiv \beta_{s-1}$ where $ r$ is multiple of $p_1^{e_1},p_2^{e_2},..,p_k^{e_k-1}$. We can continue taking $p_i^{th}$ root of $g^r$ and we retrace the path from $\beta_{s-1}$, $\beta_{s-2} ,..\beta_1$ till we reach $ g^x \equiv \beta$. \subsection{Pseudo Code} Please note pseudo code is for illustrative purposes only. \begin{algorithm}[H] \caption{Pseudo Code to calculate x such that $ g^x \equiv \gamma $ i.e. $\log{\gamma}$} \KwInput{$g ,\gamma \in F_p^* , p$; $g $ is a generator and $ p = \prod_i^kp_i^{e_i}+1 $ where $p_i$ are all primes.} \KwOutput{$x \text{ which is a solution of } g^x\equiv \gamma \pmod p $ } \begin{algorithmic} \STATE $Stack\_S \gets \gamma, elem \gets \gamma $ \STATE $plist \gets [p_1,p_1,..p_k,p_k] \text{ i.e. list of all prime factors ,as many times as they appear.} $ \WHILE{$ plist \text{ is not empty } $} \STATE $ elem \gets elem^{pop(plist)} , Stack \gets elem $ \ENDWHILE \STATE $ gpow \gets p-1, plist \gets [p_1,p_1,..p_k,p_k] $ \WHILE{ $ plist \text{ is not empty }$} \STATE $ elem \gets pop(plist) , k \gets 0 $ \STATE $ gpow \gets gpow/elem , \delta \gets g^{gpow}$ \IF {$\delta != \text{top of the Stack\_S } $ } \STATE Find $k$ such that $ \delta g^{k(p-1)/elem}\equiv \text{ Top of the stack}$ , through traditional methods \ENDIF \STATE $ \text{Remove top element of Stack\_S} $ \STATE $gpow \gets gpow + k*(gpow/elem)$ \ENDWHILE \RETURN{$gpow$} \end{algorithmic} \end{algorithm} \subsection{Example of algorithm's working} We note $41=2^3.5^1+1 $ i.e list of prime factors of $p-1$ counting each appearance is $ = [2,2,2,5] $ . Let us assume that $g=13$ is given as generator of the group and we have to find the discrete logarithm of 8, i.e we have to find $x$ such that $13^x \equiv 8 \pmod{41} $.\\ \\Now using the list of prime factors we obtain the trace back list $ [8, 8^2,8^{4},8^{8},8^{40}]$ or $[ 8,23,37,16,1]$ , all values calculated in field $ F^*_{41}$. So our trace back list is $[ 8,23,37,16,1]$. \\ \\ We start with generator now , we know $ 13^{40} \equiv 1$ \\ \\We take off the top element of trace back list , which is equal to $13^{40} $, now trace back list is $[ 8,23,37,16]$. \\Taking the fifth root , we get $ 13^{40/5} \equiv 13^{8} \equiv 10$ \\ All possible $5^{th}$ root of unity in field $ F^*_{41}$ roots can be written as $ \{13^8, 13^8*13^8,13^8*13^{16}, 13^8*13^{24},13^8*13^{32}\}$ or $[10, 18, 16, 37, 1] $. These are all possible $5^{th}$ root of unity in field $ F^*_{41}$ , hence top of the trace back list $16$ must be member of this group. we get $13^{24} \equiv 16 $. Locating this can be done by traditional methods like `baby steps giant steps' \cite{babystepgiantstep} and Pollard's rho/Kangaroo Method \cite {pollard} \cite{rho} in $ \mathcal{O}(\sqrt{p_i})$, where $p_i$ is the prime factor and order of the group. \\We take off the top element of trace back list , which is equal to $13^{24} $, now trace back list is $[ 8,23,37]$ \\ Backtracking, we now take square root , we $ 13^{24/2} \equiv 13^{12} \equiv 4 $\\ All possible square roots are $ [ 13^{12} , 13^{12}*13^{20}]$ or $[4,37]$ . As our top of the trace back is $37$ we get $ 13^32 \equiv 37$. \\ \\We take off the top element of trace back list ,which is equal to $13^{32} $, now trace back list is $[ 8,23]$ Now again backtracking , we now take square root , we get $ 13^{32/2} \equiv 13^{16} \equiv 18 $\\.As our top of the trace back is $18$. All possible square roots are $ [ 13^{16} , 13^{16}*13^{20}]$ or $[18,23]$ . As element we are looking from our trace back list is 23 , we get $13^{36} \equiv 23 $ \\We take off the top element of trace back list ,which is equal to $13^{36} $, now trace back list is $[ 8]$ \\ Now for last time backtracking , we again take square root , we get $ 13^{36/2} \equiv 13^{18} \equiv 8$. Our desired result would definitely have been member of $[ 13^{18} , 13^{18}*13^{20}] $ or $[8,33]$. Here we get $ 13^{18} \equiv 8$ , so discrete $\log_{13}{8} \equiv 18 $ in $F_{41}^*. $\\ \section{Run-time Analysis} We see trace back list, in field $F^*_p$ can be generated $ \mathcal{O}( \log {p})$. taking $p_i^{th}$ root and locating the exact root we are looking as per our trace back list can be done in $ \mathcal{O}(\sqrt {p_i})$ by baby-step-giant step algorithms for one particular $ p_i$. However this step has to be repeated for $e_i$ time for each $p_i$, hence total runtime for proposed method is $ \mathcal{O}(\log{p}+ \sum_i^k e_i\sqrt {p_i})$, which is bounded by same limits as provided by pohlig-hellman method. However , in practical terms, no combination of various group results through CRT is to be done and other lesser book keeping during the algorithm leads to much faster running time. \subsection{Run time comparison} Python code implementation of both the proposed method and Pohlig-Hellman method has been provided for comparison. Test run were run 10000 times for each example and averaged time in micro seconds is given. It is easy to see that considerable practical advantages are offered by proposed method. \begin{center} \begin{tabular}{||c c c c||} \hline & Trial 1 & Trial 2& Trial 3 \\ [0.5ex] \hline $prime\ p =$& 41 & 8101 & 200560490131 \\ \hline $generator\ g=$ & 13 & 6 & 79 \\ \hline $ h= $ & 8 & 7531 & 23 \\ \hline $\log{h} \in F_p^* =$ & 18 & 6689 & 127013812855\\ \hline\hline Dlp time & 16.3 & 47 & 274 \\ \hline Pohlig-hellman & 33.2 & 87.2 & 475 \\ [1ex] \hline \end{tabular} \end{center} \section{Advantages of Proposed Algorithm } Before we count the advantages , let us explore why proposed method gives better run-times compared to pohlig-hellman. \subsection{Why faster run-time ?} Pohlig-Hellman method calculates discrete log for each prime factor $p_i^{e_i}$ and combines the result so obtained through Chinese remainder theorem. However proposed method keeps taking $p_i^{th}$ root starting from $ g^{p-1}$ and every level it selects one out of $p_i$ values depending upon the trace back list already prepared by raising $\beta $, whose discrete logarithm is sought. This way $g^{p-1}$ reduces to $g^x$ mimicking $\beta^{p-1}$'s journey to $\beta$. And $x$ is the desired result. Taking advantage of the property of $\beta$ a direct approach to calculating discrete Logarithm lends to faster run-time. \subsection{Advantages} 1. Advantages of proposed algorithm is more on practical side as it provides faster calculation of discrete logarithm put does not improve the worst case performance bound of Pohlig-Hellman method. 2. Pohlig-Hellman's run time is agnostic to value whose log is being calculated. Proposed algorithm however may will give further lower run times for certain values in $F_p^*$ 3. Proposed method can be modified with same run time advantages for Elliptic curves too. \section{Conclusion and Acknowledgements} Discrete Logarithm problem is classical problem , cryptographic systems based on hardness of this problem still remain very popular. However special cases where p is smooth prime proposed method provides practically faster algorithm than Pohlig-Hellman . However reducing the upper bound of the worst case still remain a challenge. \\ I would like to thank Prof G.P.Biswas of IIT Dhanbad and Prof R.Munshi of ISI Kolkata, for their constant support and guidance. I owe special thanks to Gaurav sinha, IRS , and Syed Waquar Raza,IPS, for reading the proof of the paper and giving valuable suggestions.

13bb07be66c89ece053ce580d7741ef35fca767e

\section{Introduction} \label{introduction} In machine learning, models are categorized as discriminative models or generative models. From its inception, deep learning has focused on classification and discriminative models \citep{krizhevsky2012imagenet, hinton2012deep, collobert2011natural}. Another perspective came with the construction of generative models based on neural networks \citep{kingma2013auto, goodfellow2014generative, van2016conditional, kingma2018glow}. Both kinds of models give us information about the data and the similarity between examples. In particular, generative models introduce a geometric structure on generated data. Such models transform a random low-dimensional vector to an example sampled from a probability distribution approximating the one of the training dataset. As proved by \citet{arjovsky}, generated data lie on a countable union of manifolds. This fact supports the human intuition that data have a low-dimensional manifold structure, but in generative models the dimension of such a manifold is usually a hyper-parameter fixed by the experimenter. A recent algorithm by \citet{peebles2020hessian} provides a way to find an approximation of the number of dimensions of the data manifold, deactivating irrelevant dimensions in a GAN (see \citet{dollar}. Similarly, here we try to understand if a discriminative model can be used to detect a manifold structure on the space containing data and to provide tools to navigate this manifold. The implicit definition of such a manifold and the possibility to trace paths between points on the manifold can open many possible applications. In particular, we could use paths to define a system of coordinates on the manifold (more specifically on a chart of the manifold). Such coordinates would immediately give us a low-dimensional parametrization of our data, allowing us to do dimensionality reduction. In supervised learning, a model is trained on a labeled dataset to identify the correct label on unseen data. A trained neural network classifier builds a hierarchy of representations that encodes increasingly complex features of the input data \citep{olah2017feature}. Through the representation function, a distance (e.g. euclidean or cosine) on the representation space of a layer endows input data with a distance. This pyramid of distances on examples is increasingly class-aware: the deeper is the layer, the better the metric reflects the similarity of data according to the task at hand. This observation suggests that the model is implicitly organizing the data according to a suitable structure. Unfortunately, these intermediate representations and metrics are insufficient to understand the geometric structure of data. First of all, representation functions are not invertible, so we cannot recover the original example from its intermediate representation or interpolate between data points. Moreover, the domain of representation functions is the entire data domain $\mathbb{R}^n$. This domain is mostly composed of meaningless noise and data occupy only a thin region inside of it. So, even if representation functions provide us a distance, those metrics are incapable of distinguishing between meaningful data and noise. We find out that a ReLU neural network implicitly identifies a low-dimensional submanifold of the data domain that contains real data. We prove that if the activation function is piecewise-linear (e.g. ReLU), the neural network decomposes the data domain $\mathbb{R}^n$ as the disjoint union of submanifolds (the leaves of a foliation, using the terminology of differential geometry). The dimension of every submanifold (every leaf of the foliation) is bounded by the number of classes of our classification model, so it is much smaller than $n$, the dimension of the data domain $\mathbb{R}^n$. Our main theoretical result, Theorem \ref{mainthm}, stems from the study of the properties of a variant of the Fisher Information matrix, the {\sl local data matrix}. However, Theorem \ref{mainthm} cannot tell us which leaves of this foliation are meaningful or useful for practical applications: the interpretation of this geometric structure can only come from experiments. We report experiments performed on MNIST dataset. We choose to focus on MNIST because it is easily interpretable. \vskip 0.05in \begin{figure}[ht] \begin{center} \centerline{\includegraphics[width=\columnwidth]{foliation.eps}} \vskip 0.05in \caption{We provide algorithms to move along a leaf and orthogonally to it. In our experiments, we observe that moving orthogonally to the leaves changes the amount of noise in images. On the other hand, a path on a leaf is able to transform an image into a different one preserving the amount of noise of the starting image.} \label{fig:summary} \end{center} \vskip -0.2in \end{figure} Our experiments suggest that all valid data points lie on only one leaf of the foliation, the {\sl data leaf}. To observe this phenomenon we take an example from the dataset and we try to connect it with another random example following a path along the leaf containing the starting point. If such a path exists, it means that the destination example belongs to the same leaf of the foliation; otherwise, the two data points belongs to different leaves. \ref{fig:summary} pictures a summary of our experiments. Visualizing the intermediate points on these joining paths, we see that the low-dimensional data manifold defined by the model is not the anthropocentric data manifold composed of data meaningful for a human observer. The model-centric data manifold comprises images that do not belong to a precise class. The model needs those transition points to connect points with different labels. At the same time, it understands that such transition points represent an ambiguous digit: on such points, the model assigns a low probability to every class. The experiments also show that moving orthogonally to the data leaf we find noisy images. That means that the other leaves of the foliation contain images with a level of noise that increases with the distance from the data leaf. These noisy images become soon meaningless to the human eye, while the model still classifies them with high confidence. This fact is a consequence of the property of the local data matrix: equation (\ref{kl-g}) prescribes that the model output does not change if we move in a direction orthogonal to the tangent space of the leaf on which our data is located. This remark points us to other possible applications of the model-centric data manifold. We could project a noisy point on the data leaf to perform denoising, or we could use the distance from the data leaf to recognize out-of-distribution examples. Such applications require further research on the characterization of the model-centric data manifold. The main contributions of the paper are: \\ 1. the definition of the local data matrix $G(x,w)$ at a point $x$ of the data domain and for a given model $w$, and the study of its properties; \\ 2. the proof that the subspace spanned by the eigenvectors with non-zero eigenvalue of the local data matrix $G(x,w)$ can be interpreted as the tangent space of a Riemannian manifold, whose dimension is bounded by the number of classes on which our model is trained; \\ 3. the identification and visualization of the model-centric data manifold through paths, obtained via experiments on MNIST. \par{\bf Organization of the paper}. In Section \ref{ig-sec}, we review the fundamentals of information geometry using a novel perspective that aims at facilitating the comprehension of the key concepts of the paper. We introduce the {\sl local data matrix} $G(x,w)$ and we summarize its properties in Prop. \ref{prop1}. In Section \ref{model-sec}, we show that, through the local data matrix, under some mild hypotheses, the data domain foliates as a disjoint union of leaves, which are all Riemannian submanifolds of $\mathbb{R}^n$, with metric given via $G(x,w)$. In Section \ref{exp-sec}, we provide evidence that all our dataset lies on one leaf of the foliation and that moving along directions orthogonal to the data leaf amounts to adding noise to data. \section{Information Geometry} \label{ig-sec} Here we collect some results pertaining to information geometry \citep{amari, nielsen}, using a novel perspective adapted to our question, namely how to provide a manifold structure to the space containing data. Let $p(y|x,w)$ be a discrete probability distribution on $C$ classification labels, i.e. $p(y|x,w)=(p_i(y|x,w))_{i=1,\ldots, C}$, $x \in \Sigma \subset \mathbb{R}^n$, $w \in \mathbb{R}^d$. In the applications, $x$ represent input data belonging to a certain dataset $\Sigma$, while $w$ are the learning parameters, i.e. the parameters of the empirical model. As we are going to see in our discussion later on, it is fruitful to treat the two sets of variables $x$ and $w$ on equal grounds. This will naturally lead to a geometric structure on a low dimensional submanifold of $\mathbb{R}^n$, that we can navigate through paths joining points in the dataset $\Sigma$ (see Section \ref{exp-sec}). In order to give some context to our treatment, we define, following \citet{amari} Section 3, the \textit{information loss} $I(x,w)= - \log(p(y|x,w))$ and the \textit{loss function} $L(x,w)=\mathbb{E}_{y\sim q}[I(x,w)]$. Typically $L(x,w)$ is used for practical optimizations, where we need to compare the model output distribution $p(y|x,w)$ with a certain known true distribution $q(y|x)$. We may also view $L(x,w)$ as the Kullback-Leibler divergence between $p(y|x,w)$ and $q(y|x)$ up to the constant $-\sum_i q_i(y|x) \log q_i(y|x)$, irrelevant for any optimization problem: \begin{align} & L(x,w) =\mathbb{E}_{y \sim q}[- \log(p(y|x,w))]= \nonumber \\ &= \sum_{i=1}^C q_i(y|x) \log \frac{q_i(y|x)}{p_i(y|x,w)} - \sum_{i=1}^C q_i(y|x) \log q_i(y|x)= \nonumber \\ &=\mathrm{KL}(q(y|x)||p(y|x,w)) - \sum_{i=1}^C q_i(y|x) \log q_i(y|x). \end{align} A popular choice for $p(y|x,w)$ in deep learning classification algorithms is \begin{equation}\label{prob} p_i(y|x,w)=\text{softmax}(s(x,w))_i = \frac{e^{s_i(x,w)}}{\sum_{j=1}^{C} e^{s_j(x,w)}}, \end{equation} where $s(x,w) \in \mathbb{R}^C$ is a score function determined by parameters $w$. From such $p(y|x,w)$ we derive the cross-entropy with softmax loss function: \begin{align} L(x,w)&=\mathbb{E}_{y\sim q}[I(x,w)]=\mathbb{E}_{y\sim q}[-\log p(y|x,w)]= \nonumber \\ &= - s_{y_x}(x,w) + \log{\sum_{j=1}^C e^{s_j(x,w)}}, \label{softmax} \end{align} where $L(x,w)$ is computed with respect to the probability mass distribution $q(y|x)$ assigning $1$ to the correct label $y_x$ of our datum $x$ and zero otherwise. Going back to the general setting, notice that \begin{align} \label{eq:expected_value} & \mathbb{E}_{y\sim p}[\nabla_w I(x,w))]=\sum_{i=1}^C p_i(y|x,w) \nabla_w \log p_i(y|x,w)= \nonumber \\ &= \! \sum_{i=1}^C \nabla_w p_i(y|x,w)= \nabla_w \! \! \left(\sum_{i=1}^C p_i(y|x,w)\!\right) \! = \! \nabla_w 1 = 0. \end{align} Let us now define the following two matrices: \begin{align} & F(x,w) = \mathbb{E}_{y\sim p}[\nabla_w\log p(y|x,w) \cdot (\nabla_w \log p(y|x,w))^T] \label{fisher} \\ & G(x,w) = \mathbb{E}_{y\sim p}[\nabla_x\log p(y|x,w) \cdot (\nabla_x \log p(y|x,w))^T]. \label{metric-g} \end{align} We call $F(x,w)$ the \textit{local Fisher matrix} at the datum $x$ and $G(x,w)$ the \textit{local data matrix} given the model $w$. The Fisher matrix \citep{amari} is obtained as $F(w)=\mathbb{E}_{x \sim \Sigma}[F(x,w)]$ and it gives information on the metric structure of the space of parameters. Similarly, we can reverse our perspective and see how $G(x,w)$ allows us to recognize some structure in our dataset. The following observations apply to both $F(x,w)$ and $G(x,w)$ and provide the theoretical cornerstone of Section \ref{model-sec}. \begin{proposition} \label{prop1} Let the notation be as above. Then: \begin{enumerate} \item \label{thm:positive} $F(x, w)$ and $G(x, w)$ are positive semidefinite symmetric matrices. \item \label{thm:ker} $\ker F(x,w)= (\mathrm{span}_{i=1, \ldots, C}\{ \nabla_w \log p_i(y|x,w)\})^\perp$;\\ $\ker G(x,w)= (\mathrm{span}_{i=1, \ldots, C}\{ \nabla_x \log p_i(y|x,w)\})^\perp$. \item \label{thm:rk} $\mathrm{rank}\ F(x,w) < C$, \quad $\mathrm{rank}\ G(x,w) < C$. \end{enumerate} \end{proposition} \begin{proof} See Appendix \ref{app-sec2}. \end{proof} This result tells us that the rank of both $F(x,w)$ and $G(x,w)$ is bounded by $C$, the number of classes in our classification problem. The bound on $\mathrm{rank}\ G(x,w)$ will allow us to define a submanifold of $\mathbb{R}^n$ of dimension $\mathrm{rank}\ G(x,w)$ (Section \ref{model-sec}), that our experiments show contains our dataset (Section \ref{exp-sec}). In practical situations, this dimension is much lower than the size of $G(x, w)$, i.e. the input size $n$, as shown in Table \ref{table:g-bound}. \vskip -0.1in \begin{table}[ht] \caption{Bound on the rank of $G(x,w)$ for popular image classification tasks.} \label{table:g-bound} \begin{center} \begin{small} \begin{sc} \begin{tabular}{ccc} \toprule Dataset & $G(x, w)$ size & $\mathrm{rank}\ G(x, w)$ bound \\ \midrule MNIST & 784 & 10 \\ CIFAR-10 & 3072 & 10 \\ CIFAR-100 & 3072 & 100 \\ ImageNet & 150528 & 1000 \\ \bottomrule \end{tabular} \end{sc} \end{small} \end{center} \vskip -0.1in \end{table} \section{The Model View on the Data Manifold}\label{model-sec} We now turn to examine some properties of the matrices $F(x,w)$ and $G(x,w)$ that will enable us to discover a submanifold structure on the portion of $\mathbb{R}^n$ occupied by our dataset (see Section \ref{exp-sec}) and to prove the claims at the end of the above section. We recall that, given a perturbation of the weights $w$, the Kullback-Leibler divergence, gives, to second-order approximation, the following formula: \begin{align} &\mathrm{KL}(p(y|x,w+\delta w)||p(y|x,w)) = \nonumber \\ &= (\delta w)^T F(x,w) (\delta w) +\mathcal{O}(||\delta w||^3)\label{kl-f}. \end{align} Equation (\ref{kl-f}), together with Proposition \ref{prop1}, effectively expresses the fact that during SGD dynamics with a mini-batch of size 1, we have only a very limited number of directions, namely $C - 1$, in which the change $\delta w$ affects the loss. Taking the expectation with respect to $x \sim \Sigma$ on both sides of the equation, we obtain the analogous property for the Fisher matrix $F(w)=\mathbb{E}_{x \sim \Sigma}[F(x,w)]$. While $F(x,w)$ has a low rank, the rank of $F(w)$ is bounded by $|\Sigma|(C-1)$, a number that is often higher than the size of $F(w)$. Thus $F(w)$, when non-degenerate, is an effective metric on the parameter space. It allows us to measure, according to a certain step $\delta w$, when we reach a stable predicted probability and thus the end of model training (see \citet{martens} and refs. within). It must be however noted that $F(w)$ retains its information content only {\sl away} from the trained model, that is, well before the end of the training phase (see \citet{as} for an empirical validation of such statements). We are going to see, with our experiments in Section \ref{exp-sec}, that a similar phenomenon occurs for $G(x,w)$. We now turn to the local data matrix $G(x,w)$, thus interpreting equation (\ref{kl-f}) in the data domain $\mathbb{R}^n$. For a perturbation $\delta x$ of the data $x$, we have, up to second order approximation: \begin{align} &\mathrm{KL}(p(y|x+\delta x,w)||p(y|x,w)) = \nonumber \\ &= (\delta x)^T G(x,w) (\delta x) +\mathcal{O}(||\delta x||^3) \label{kl-g}. \end{align} This equation tells us that, if we move along the directions of $\ker G(x,w)$, the probability distribution $p(y|x,w)$ is constant (up to a second order approximation) while our data is changing. Those are the vast majority of the directions, since $\mathrm{rank}\ G(x,w)<C$ and typically $C \ll n$, hence, we interpret them as the \textit{noise directions}. On the other hand, if we move from a data point $x$ along the directions in $(\ker G(x,w))^\perp$, data will change along with the probability distribution $p(y|x,w)$ associated with it. These are the directions going toward data points that the model classifies with confidence. Equation (\ref{kl-g}) suggests to view $G(x,w)$ as a metric on the data domain. However, because of its low rank (see Proposition \ref{prop1}), we need to restrict our attention to the subspace $(\ker G(x,w))^\perp$, where $G(x,w)$ is non-degenerate. $G(x,w)$ allows us to define a \textit{distribution} $\mathcal{D}$ on $\mathbb{R}^n$. In general, in differential geometry, we call distribution on $\mathbb{R}^n$ an assignment: $$ x \mapsto \mathcal{D}_x \subset \mathbb{R}^n, \qquad \forall x \in \mathbb{R}^n $$ where $\mathcal{D}_x$ is a vector subspace of $\mathbb{R}^n$ of a fixed dimension $k$ (see Appendix \ref{app-sec1} for more details on this notion in the general context). Assume now $G(x,w)$ has constant rank; as we shall see in our experiments, this is the case for a non fully trained model. We thus obtain a \textit{distribution} $\mathcal{D}$: \begin{equation}\label{distr-g} x \mapsto \mathcal{D}_x =(\ker G(x,w))^\perp \subset \mathbb{R}^n. \end{equation} We now would like to see if our distribution (\ref{distr-g}) defines a \textit{foliation structure}. This means that we can decompose $\mathbb{R}^n$ as the disjoint union of submanifolds, called \textit{leaves} of the foliation, and there is a {\sl unique} submanifold (leaf) going through each point $x$ (see Figure \ref{app-fig1} in Appendix \ref{app-sec1}, where the distribution is generated by the vector field $X$ and $\mathbb{R}^2$ is the disjoint union of circles, the leaves of the foliation). The distribution comes into the play, because it gives the tangent space to the leaf through $x$: $\mathcal{D}_x= (\ker G(x,w))^\perp$. In this way, moving along the directions in $\mathcal{D}_x$ at each point $x$, will produce a path lying in one of the submanifolds (leaves) of the foliation. The existence of a foliation, whose leaf through a point $x$ has tangent space $\mathcal{D}_x$, comes through Frobenius theorem, which we state in the Appendix \ref{app-sec1} and we recall here in the version that we need. {\bf Frobenius Theorem}. \textit{Let $x \in \mathbb{R}^n$ and let $\mathcal{D}$ be a distribution in $\mathbb{R}^n$. Assume that in a neighbourhood $U$ of $x$: \begin{equation}\label{inv-prop} [X,Y] \in \mathcal{D}, \qquad \forall\ X,Y \in \mathcal{D}. \end{equation} Then, there exists a (local) submanifold $N \subset \mathbb{R}^n$, $x \in N$, such that $T_zN=\mathcal{D}_z$, for all $z \in N$.} It is not reasonable to expect that a general classifier satisfies the involutive property (\ref{inv-prop}), however it is remarkable that for a large class of classifiers, namely deep ReLU neural networks, this is the case, with $p$ given by softmax as in equation (\ref{prob}). \begin{theorem} \label{mainthm} Let $w$ be the weights of a deep ReLU neural network classifier, $p$ given by softmax, $G(x,w)$ the local data matrix. Assume $G(x,w)$ has constant rank. Then, there exists a local submanifold $N \subset \mathbb{R}^n$, $x \in N$, such that its tangent space at $z$, $T_z N=(\ker G(z,w))^\perp$ for all $z \in N$. \end{theorem} \begin{proof} See Appendix \ref{app-sec2}. \end{proof} Through the application of Frobenius theorem, Theorem \ref{mainthm} gives us a foliation of the data domain $\mathbb{R}^n$. $\mathbb{R}^n$ decomposes into the disjoint union of $C-1$ dimensional submanifolds, whose tangent space at a fixed $x \in \mathbb{R}^n$ is $\mathcal{D}_x=(\ker G(x,w))^\perp$ (see Appendix \ref{app-sec1}). Every point $x$ determines a unique submanifold corresponding to the leaf of the foliation through $x$. We may extend this local submanifold structure to obtain a global structure of manifold on a leaf, still retaining the above property regarding the distribution. As we shall see in Section \ref{exp-sec}, we can move from a point $x$ in our dataset $\Sigma$ to another point $x'$ also in $\Sigma$ with an \textit{horizontal} path, that is a path tangent to $\mathcal{D}_x$, hence lying on the leaf of $x$ and $x'$. Our experiments show that we can connect every pair of points $(x, x') \in \Sigma \times \Sigma$ with horizontal paths. It means that all the dataset belongs to a single leaf, which we call the \textit{data leaf} $\mathcal{L}$. Our model, through the local data matrix $G(x,w)$, enables us to move on the low-dimensional submanifold $\mathcal{L}$, to which all of our dataset belongs. Of course not all the points of $\mathcal{L}$ correspond to elements of the dataset; however as we show in the experiments, on most points of $\mathcal{L}$ the model gives prediction compatible with human observers. We also notice that each leaf of our foliation comes naturally equipped with a metric given at each point by the matrix $G(x,w)$, restricted to the subspace $(\ker G(x,w))^\perp$, which coincides with the tangent space to the leaf, where $G(x,w)$ is non-degenerate. Hence $G(x,w)$ will provide each leaf with a natural Riemannian manifold structure. We end this section with an observation, comparing our approach to the geometry of the data domain, with the parameter space. \par{\bf Remark}. Equation (\ref{kl-f}) provides a metric to the parameter space $\mathbb{R}^d$, motivating our approach to the data domain. For each $w \in \mathbb{R}^d$, we can define, as we did for the data domain, a distribution $w \mapsto \mathcal{D}'_w:=(\ker F(w))^\perp$, using the Fisher matrix. However, it is easy to see empirically that this distribution is {\sl not involutive}, i.e. there is no foliation and no submanifold corresponding to it. \section{Experiments}\label{exp-sec} We performed experiments on the MNIST dataset \citep{lbbh} and on the CIFAR-10 dataset \citep{krizhevsky2010cifar}. We report in this section the experiments on the MNIST dataset only, because they are easier to interpret in the geometrical framework (foliation and leaves) introduced in our previous section. CIFAR-10 experiments are reported in the dedicated Appendix \ref{app-sec3}. All the following experiments use the same simple CNN classifier trained on MNIST. The neural network is similar to LeNet, with 32 and 64 channels in the two convolutional layers and 128 hidden units in the fully-connected layer. The architecture is identical to the one proposed in the official MNIST example of the PyTorch framework \citep{pytorch}. The network has ReLU activation function, so it satisfies the hypothesis of Theorem \ref{mainthm}. We train this network with SGD, a batch size of 60 and a fixed learning rate of 0.01. \subsection{Rank and Trace of the Local Data Matrix} The definition of distribution in Section \ref{model-sec} and the consequent results require the rank of $G(x, w)$ to be constant on every point $x$ for a certain parameter configuration $w$. We remark that the rank of $G(x, w)$ can be at most $C-1$ (Proposition \ref{thm:rk}), i.e. much lower than the size of $G(x, w)$ (Table \ref{table:g-bound}). To check if our assumption is satisfied by a model, we should calculate the rank of several local data matrices varying the data point $x$. Unfortunately, it is very difficult to establish the rank of a matrix in numerical experiments, because the rank is not robust with respect to small perturbations. Knowing that the rank correspond to the $\ell_0$ norm of the eigenvalues of a matrix, we can consider the $\ell_1$ norm of the eigenvalues as a soft version of the rank. For a positive semidefinite symmetric matrix like $G(x, w)$ (see Proposition \ref{prop1}.\ref{thm:positive}), the $\ell_1$ norm of the eigenvalues corresponds to the trace of the matrix. We can suppose that at the beginning of the training $\mathrm{rank}\ G(x, w) = C - 1$ for almost all $x \in \mathbb{R}^n$, because the vectors $\nabla_x \log p_i(y|x, w)$ with $i=1,\ldots, C$ point toward random directions. So, if the mean trace measured for a certain parameter configuration $w$ is above the value measured at the beginning of the training, we can be confident that the number of non-null eigenvalues in $G(x, w)$ has not changed. When the mean trace goes below the reference starting value during training, we can expect a non-constant value of $\mathrm{rank}\ G(x, w)$ because for some points $x$ the model has converged to its final prediction, while other points require more training step to reach convergence. \begin{figure}[ht] \begin{center} \centerline{\includegraphics[width=\columnwidth]{g_trace.eps}} \vskip -0.1in \caption{The blue line is the mean trace of $G(x, w)$ for a batch of images during training. The red line represent the trend of the mean trace obtained through an exponential moving average.} \label{fig:trace} \end{center} \vskip -0.2in \end{figure} \begin{figure*}[ht] \begin{center} \includegraphics[width=\textwidth]{06311_noise.eps} \hfill \vspace*{-1em} \includegraphics[width=\textwidth]{03867_noise.eps} \end{center} \vskip -0.1in \caption{Paths across leaves of the foliation. Notice how the images become increasingly indistinguishable from noise, while the classifier continues to assign a high probably to the class of the initial image.} \label{fig:noise} \vskip -0.1in \end{figure*} Figure \ref{fig:trace} shows the trend of the mean trace of $G(x, w)$ for a batch during the training. The plot is similar to the ones reported in \citet{as} for the Fisher information matrix and it tells us that the local data matrix $G(x, w)$ loses its informative content at the end of training as well. Since \begin{align} &\Tr G(x, w)\! =\! \Tr \mathbb{E}_{y\sim p}[\nabla_x\! \log p(y|x) \! \cdot \! (\nabla_x\! \log p(y|x))^T] \! = \nonumber \\ &= \mathbb{E}_{y\sim p}[\Tr(\nabla_x\log p(y|x,w) \cdot (\nabla_x \log p(y|x,w))^T)] = \nonumber \\ &= \mathbb{E}_{y\sim p}[||\nabla_x\log p(y|x,w)||^2], \end{align} we could expect such a decreasing trend. In fact a small perturbation of the data point $x$ should not influence the prediction of a fully-trained model, thus $||\nabla_x\log p_i(y|x,w)||^2$ should be almost null for every class $i$ at the end of training. All these observations motivate us to perform subsequent experiments on a partially trained model. In particular, we use the checkpoint at step 10000, i.e. at the end of the 10\textsuperscript{th} epoch. At this step, the smoothed mean trace of $G(x, w)$ is very similar to the one measured at the beginning of the training, so the rank of the local data matrix should be $C-1$ on almost every point. \subsection{Characterization of Leaves} \label{characterization-subsec} Since our neural network classifier satisfies all the hypothesis of Theorem \ref{mainthm}, it views the data domain $\mathbb{R}^n$ as a {\sl foliation}. Nevertheless, this result does not give us any clue about the characterization of leaves. We know, however, that, moving away from a data point $x$ while remaining on the same leaf, we can obtain points with different labels, while moving in a direction orthogonal to the tangent space to the leaf of $x$, we obtain points with the same label as $x$, as long as the estimate (\ref{kl-g}) holds. To understand the distinguishing factors of a leaf, we move from a data point $x$ across leaves and we inspect the crossed data points. More specifically, we start from an image in MNIST test set and we let it evolve moving orthogonally to the tangent space $\mathcal{D}_{x_t} = (\ker\ G(x_t, w))^\perp$. Since the dimension of the orthogonal space is $n - C + 1$, there are many possible directions. Thus, to monotonically increase the distance from the starting point $x$, we use a fixed random direction and at every step we project it on $\ker G(x_t, w)$, i.e on the orthogonal space of the tangent space of current leaf. We observe in Figure \ref{fig:noise}, that the noise in the images increases steadily. Moreover, we can see that, as predicted by equation (\ref{kl-g}), model predictions remain very certain even when the digit is indistinguishable for the human eye. This experiment makes us speculate that a leaf is characterized by a constant amount of noise. If that is the case, all valid data should reside on the same leaf characterized by the absence of noise, the {\sl data leaf}. \subsection{Horizontal Paths on Leaves} We follow horizontal paths connecting two images from MNIST test set. If it is possible to join two inputs with a horizontal path, then we know that those points are on the same leaf. In our case a horizontal path is tangent to the distribution $\mathcal{D}$ described in the previous section, i.e. $\mathcal{D}_x=(\ker G(x,w))^\perp = \mathrm{span}_{i=1, \ldots, C}\{ \nabla_w \log p_i(y|x,w)\}$ from Proposition \ref{thm:ker}. We remark that $\mathcal{D}_x$ coincides with the row space of the Jacobian matrix $\mathrm{Jac}_x \log p(y|x, w)$. We use this equivalence to compute the projection of a vector on the distribution $\mathcal{D}_x$. \vskip -0.1in \begin{algorithm}[H] \caption{Find a horizontal path between points} \label{alg:path} \begin{algorithmic} \STATE {\bfseries Input:} source $s$, destination $d$, step size $\alpha$, number of iterations $T$, model parameters $w$ \STATE $x_0 = s$ \FOR{$t=1,\ldots, T$} \STATE \# Calculate Jacobian of $\log p(y|x_{t-1}, w)$ w.r.t. $x_{t-1}$ \STATE $j = \mathrm{Jac}_x \log p(y|x_{t-1}, w)$ \STATE \# Project the gradient of $||d -x_{t-1}||^2$ on $\mathcal{D}_{x_{t-1}}$) \STATE $v = \text{projection}(d - x_{t-1}, j)$ \STATE \# Update $x_{t-1}$ to obtain the next point $x_{t}$ on the path \STATE $x_{t} = x_{t-1} + \alpha \frac{v}{||v||}$ \ENDFOR \STATE {\bfseries Output:} $\{x_t\}_{t=0,\ldots, T}$ \end{algorithmic} \end{algorithm} \vskip -0.1in \begin{figure*}[tb] \begin{center} \includegraphics[width=\textwidth]{03898_09709.eps} \hfill \vspace*{-1em} \includegraphics[width=\textwidth]{04185_05874.eps} \end{center} \vskip -0.3in \caption{Horizontal paths between two images in MNIST test set. Here, the paths require at most 5000 steps, different source-destination pairs can need more steps and a longer path. The sequences show that a horizontal path is very different from an interpolation.} \label{fig:horizontal} \end{figure*} \begin{figure*}[tb] \begin{center} \includegraphics[width=\textwidth]{00125_flip_06349.eps} \hfill \vspace*{-1em} \includegraphics[width=\textwidth]{04200_flip_09792.eps} \end{center} \vskip -0.1in \caption{Horizontal paths between a valid image and a mirrored image from MNIST test set. The mirrored images are chosen to resemble letters (E and P). Notice how in the first sequence the path passes through the digit 5 before reaching the destination point.} \label{fig:letter} \vskip -0.1in \end{figure*} Algorithm \ref{alg:path} finds an approximate horizontal path from a source point to a destination point. The algorithm is obtained as a simplification of Riemannian gradient descent \citep{gabay1982minimizing} using the squared euclidean distance from the destination point as loss function. Since we do not have an explicit characterization of the leaves, we cannot perform the retraction step commonly used in optimization algorithms on manifolds. To circumvent this problem, we normalize the gradient vector $v$ and use a small step size $\alpha = 0.1$ in our experiments. The normalization assures that the norm of the displacement at every step is controlled by the step size. Figure \ref{fig:horizontal} shows the results of the application of Algorithm \ref{alg:path} on some pairs of images in MNIST test set. We observe that it is possible to link different images with an horizontal path, confirming our conjecture about the existence of the data leaf. This experiment shows that the model sees the data on the same manifold, namely one leaf of the foliation determined by our distribution $\mathcal{D}$. The observation of the paths on the data leaf gives us a novel point of view on the generalization property of the model. First of all, while we could expect to find all training data on the same leaf, it is remarkable that test data are placed on the same leaf too. Furthermore, we observe that the data leaf is not limited to digits and transition points between them. In fact, we can find paths connecting valid images with images of non-existent digits similar to letters, as shown in Figure \ref{fig:letter}. This fact suggests that the model-centric data manifold is somewhat more general than the classification task on which the model is trained. Our final experiment gives us another remarkable confirmation of our theoretical findings: it is impossible to link with a horizontal path a noisy image outside the data leaf and a valid data image. A noisy image is generated using the same strategy presented in Section \ref{characterization-subsec}, i.e. modifying a valid image along a direction in $\ker G(x, w)$. Figure \ref{fig:noise_horizontal} shows that even after 10000 iterations of Algorithm \ref{alg:path}, it is impossible to converge to the valid destination point. Indeed, the noise is preserved along the path and the noise pattern stabilizes after 5000 iterations. The final point reached is a noisy version of the actual destination point. \begin{figure*}[ht] \begin{center} \includegraphics[width=\textwidth]{00747_noise_03333.eps} \hfill \vspace*{-1em} \includegraphics[width=\textwidth]{02386_noise_07528.eps} \end{center} \vskip -0.1in \caption{Horizontal paths unable to reach a valid image in MNIST test set from a noisy image.} \label{fig:noise_horizontal} \vskip -0.1in \end{figure*} All those experiments confirm that there exists a model-centric data manifold, the data leaf of the foliation. In addition, they exhibit that the leaves in the foliation of the data domain are characterized by the noise content. \section{Related Works} \par{\bf Fisher Information Matrix}. In \citet{as} the authors discuss the information content of the Fisher matrix, and they show that such content changes during the training of a neural network, decreasing rapidly towards the end of it. This shows that the Fisher-Rao metric acquires importance during the training phase only (see also \citet{kirk}). More on the geometry of the parameter space, as a Riemannian manifold, is found in \citet{bronstein}. In this paper, we take an analog of the Fisher-Rao metric, but on the data domain. In our examples, we see a phenomenon similar to the one observed in \citet{as}: the trace of the local data matrix decreases rapidly, as the model completes its training. Other metrics on datasets were suggested, for example see \citet{montufar} and refs. within, but with different purposes. Here, our philosophy is the same as in \citet{frosini}: we believe that data itself is not equipped with a geometric structure, but such structure emerges only when the model views data, with a given classification task. \par{\bf Intrinsic Dimension}. \citet{ansuini2019intrinsic} measure the intrinsic dimension of layer representations for many common neural network architecture. At the same time, they measure the intrinsic dimension of MNIST, CIFAR-10 \citep{krizhevsky2010cifar} and ImageNet \citep{deng2009imagenet} datasets. Our objective is similar, but we do not specifically quantify the dimension of the data manifold. The data leaf reflects how a classifier sees a geometric structure on the discrete data points from a dataset. Its dimension is intimately linked to the classification task. For this reason the results are not directly comparable. \citet{ansuini2019intrinsic} show that MNIST and neural networks trained on it behave very differently from networks trained on CIFAR-10 or ImageNet. While our experiments focus on MNIST, additional experiments on CIFAR-10 are shown in the Appendix \ref{app-sec3}. \par{\bf Adversarial Attacks}. Our method to navigate the leaves of the foliation is very similar to common adversarial attack methods, like Fast Gradient Sign Method \citep{fgsm} or Projected Gradient Descent. Adversarial attacks and our navigation algorithm both rely on gradients $\nabla_x \log p_i$, but adversarial generation algorithms perturb the original image by $\sign(\nabla_x \log p_i)$. In general, $\sign(\nabla_x \log p_i) \notin \ker(G(x,w))^\perp$, so adversarial examples are created perturbing the image outside the data leaf. \section{Conclusions} In this paper, we introduce the local data matrix, a novel mathematical object that sheds light on the internal working of a neural network classifier. We prove that the model organizes the data domain according to the geometric structure of a foliation. Experiments show that valid data are placed on the same leaf of the foliation, thus the model sees the data on a low-dimensional submanifold of the data domain. Such submanifold appears more general than the model itself, because it includes meaningless, but visually similar, images together with training and test data. In the future, we aim to characterize the data leaf and to study the Riemannian metric given by the local data matrix. If we could analytically characterize the leaves of the foliation by the degree of noise, we could distinguish noisy data from valid examples. That can be used in the inference phase to exclude examples on which model predictions are not reliable. Furthermore, it could pave the way to the creation of novel generic denoising algorithms applicable to every kind of data.

2e255bb1ddbd860b796f4f68e3714c22391667bb

\section{Introduction}\label{sec:Intro} The $\Lambda$CDM paradigm has been demonstrably successful in describing the observed large-scale structure, accounting for the existence of the cosmic microwave background, and naturally explaining the current accelerated expansion of the Universe. Many predictions of the standard cold dark matter (CDM) model on sub-galactic scales, however, appear inconsistent with observations (e.g. the Core-cusp, Missing Satellites, and Too-Big-to-Fail problems) \cite{Bullock:2017xww, Weinberg:2013aya, Spergel:1999mh, Kamionkowski:1999vp, SommerLarsen:1999jx}. As a well-motivated alternative DM candidate, the fuzzy dark matter (FDM) \cite{Hu:2000ke, Marsh2016, Hui:2016ltb, Niemeyer:2019aqm} is composed of ultralight bosons with a particle mass $m_b \sim 10^{-22}$\textendash$10^{-20}$ eV. The model offers both observationally consistent descriptions of small-scale structure and large-scale predictions consistent with the $\Lambda$CDM framework \cite{Schive:2014dra, Marsh:2015wka, Calabrese:2016hmp, Chen:2016unw, Gonzales-Morales:2016mkl}. Exhibiting astrophysically large de Broglie wavelength, FDM features a standing wave soliton supported by quantum pressure forming at the center of every galaxy, accompanied by an extensive NFW-like halo \cite{Schive:2014dra, Schive:2014hza}. The rich phenomenology of FDM has been explored actively via numerical simulations \cite{Schwabe:2016rze, Mocz:2017wlg, Lin:2018whl, Veltmaat:2018dfz, Veltmaat:2019hou, Li:2020ryg}. However, despite the substantial progress by simulations in revealing many intriguing aspects of FDM, mechanisms such as soliton coherent density oscillations, stochastic density fluctuations (halo granules), and soliton random walk excursions still require further theoretical investigations and observational support. A recent investigation into the survival of the ancient central star cluster (SC) of the ultra-faint dwarf galaxy Eridanus II (Eri II) by Marsh \& Niemeyer (2019) \cite{Marsh:2018zyw}, henceforth abbreviated as \MN, estimated the gravitational heating effect caused by stochastic fluctuations and soliton density oscillations. By approximating the dynamical timescales of both mechanisms, a lower bound $m_b \gtrsim 10^{-19}$ eV (with partial exclusions in the range $10^{-20} < m_b < 10^{-19}$ eV) was derived together with the constraint from the subhalo mass function. This bound is rather stringent compared with other constraints from, for example, Lyman-$\alpha$ forest $m_b \gtrsim 2\times10^{-21}$ eV \cite{Kobayashi:2017jcf, Irsic:2017yje, Armengaud:2017nkf}, DM (sub)halo abundance $m_b \gtrsim 2\textendash3 \times 10^{-21}\eV$ \cite{Nadler:2019zrb, Schutz:2020jox, Nadler:2020prv}, galaxy luminosity function at high redshifts $m_b \gtrsim 1\textendash8\times10^{-22}\eV$ \cite{Schive:2015kza, Corasaniti:2016epp, Menci:2017nsr}, and heating of galactic disks $m_b \gtrsim 1\textendash2\times10^{-22}\eV$ \cite{Church2019, BarOr2019, Amr2020}. In this work, we present a theoretical framework that pins down the dynamical timescales of both the ground-state soliton wavefunction $\tau_{00}$ and the coherent density oscillations $\tau_\text{soliton}$ for a perturbed FDM soliton. Concerning the stability of Eri II SC, \cite{Schive:2019rrw} demonstrates that the SC undergoes a complete tidal disruption within $\sim 1\Gyr$ due to soliton random walk excursions in the presence of a background halo; the effect can be largely mitigated if the background halo is stripped away by the Milky Way tides. Observational constraints on Eri II \cite{Crnojevi:2016ApJ, Li:2016utv} indicate that the mass contribution within its half-light radius is predominantly sourced by the soliton core for $m_b \lesssim 10^{-21}$ eV. We confirm both analytically and with N-body and FDM simulations that the effect of gravitational heating on Eri II SC due to soliton density oscillations is negligible for $m_b \lesssim 4\times10^{-21}\eV$, contrary to the conclusion drawn by \MN. We further show that the FDM constraints derived by \MN\, from the subhalo mass function is similarly invalid. We begin with an overview on the analytical descriptions of FDM and the astrophysical constraints on the soliton profile from Eri II in Sec. \ref{OscillationFrequencyRevisit}. In Secs. \ref{Heating} and \ref{sec:sim}, we present both analytical and numerical calculations on the SC heating and demonstrate that orbital resonances are inefficient for $m_b \lesssim 4\times10^{-21}$ eV, which is followed by remarks concerning \MN\, in Sec. \ref{app:TragetPaper}. In Sec. \ref{sec:conclude} we conclude. Technical details are organized as follows: derivations of and discussions on relevant features of an unperturbed soliton (Appendix \ref{app:SolitonUnperturbed}), linear soliton density perturbations (Appendix \ref{app:PerturbationAppendix}), gravitational heating of the SC (Appendix \ref{app:HamiltonianPerturbation}), and resonance analysis (Appendix \ref{app:Resonance}). \section{Fuzzy Dark Matter Solitons: Properties and Constraints}\label{OscillationFrequencyRevisit} In this section, we examine aspects of the FDM soliton and the astrophysical constraints from Eri II on the soliton density profile. We identify two distinct timescales associated with the ground-state soliton wavefunction and solitary coherent density fluctuations respectively. \subsection{Soliton Oscillations}\label{sec:SolitonTimeScale} The evolution of FDM is described by a classical complex scalar $\Psi$ that obeys the coupled Schr\"{o}dinger-Poisson (SP) equations \cite{Hu:2000ke, Hui:2016ltb, Niemeyer:2019aqm} \begin{eqnarray} i \hbar \frac{\partial \Psi}{\partial t} &=& -\frac{\hbar^2}{2m_b} \nabla^2 \Psi + m_bV\Psi,\label{eqn:SPeqnS}\\ \nabla^2 V &=& 4\pi G m_b |\Psi|^2, \label{eqn:SPeqnP} \end{eqnarray} where $V$ denotes the gravitational potential, $G$ the gravitational constant, and $m_b|\Psi|^2 = \rho_\text{FDM}$ the mass density. The density profile of an unperturbed (ground-state) soliton $\rho_\text{soliton}$ and the characteristic radius of a soliton $r_c$ at which $\rho_\text{soliton}=\rho_c/2$ read \cite{Schive:2014dra} \begin{eqnarray} &&\rho_\text{soliton}(\gamma; \rho_c, m_b) =\rho_c\big(1+9.1\times 10^{-2}\gamma^2\big)^{-8},\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\:\:\:\:\label{eqn:gammaDensityDist}\\ &&r_c(\rho_c, m_b)=\Big(\frac{\rho_c}{1.9 \text{ M}_\odot \text{pc}^{-3}}\Big)^{-1/4} \Big(\frac{m_b}{10^{-23}\text{eV}}\Big)^{-1/2} \text{kpc}, \;\;\;\;\label{eqn:SolitonRadiusEriII} \end{eqnarray} where $\rho_c$ denotes the soliton central density and we have introduced a dimensionless spatial variable \begin{eqnarray} \gamma &\equiv& \frac{r}{r_c}. \label{eqn:ScaledRadius} \end{eqnarray} The gravitational potential of a soliton takes the form \begin{eqnarray}\label{eqn:VExcitedStatesAnalysis} &&V_0(\gamma; \rho_c, m_b) = -G r_c^2 \rho_c \Bigg\{\bigg[\frac{4.17\times 10^{-17}}{(1+9.1\times 10^{-2}\gamma^2)^6}\bigg]\big(1.83 \\ &&\times 10^{17}+5.14\times 10^{16}\gamma^2 +7.09\times 10^{15}\gamma^4+ 5.30\times 10^{14}\gamma^6 \nonumber \\ &&+2.07\times 10^{12}\gamma^{8}+3.33\times 10^{11} \gamma^{10}\big) + \frac{7.38\arctan{\big(0.302\gamma\big)}}{\gamma}\Bigg\},\nonumber \end{eqnarray} which suggests a cutoff radius $\gamma_\text{cut} = 3.3$ (Fig. \ref{fig:SolitonPotential}), beyond which the soliton acts gravitationally as a point mass. Relatedly, $\gamma_\text{cut}$ also offers a natural explanation to the observed universal soliton-halo profile transition radius $\gamma \simeq 3.5$ \cite{Mocz:2017wlg}. The eigenfunctions of an unperturbed soliton $\psi_{nlm}(\gamma, \theta, \phi) = (r_c \gamma)^{-1}u_{nl}(\gamma) \times Y_l^m(\theta, \phi)$ (\eref{eqn:WaveFDecomposition}) are derived in Appendix \ref{app:SolitonUnperturbed}. The associated energy levels $E_{nl}$ scales as $\rho_c^{1/2}$, as $m_b V_0 \propto m_b r_c^2 \rho_c\propto\rho_c^{1/2}$. The oscillation period of the ground-state soliton wavefunction is given by \begin{eqnarray} \tau_{00}(\rho_c)&=& \frac{2\pi \hbar}{|E_{00}|} \simeq 39.9 \bigg(\frac{\rho_c}{\text{M}_\odot \text{pc}^{-3}}\bigg)^{-1/2} \text{ Myr},\;\;\;\;\; \label{eqn:NumGroundPeriod} \end{eqnarray} which expectedly agrees within $5 \%$ to the value observed in the simulated soliton wavefunction oscillations \cite{Schive:2019rrw} ($\tau_{00} = 116$ Myr and $\rho_c = 0.107$ M$_\odot$pc$^{-3}$). Irrespective of the presence of a background halo, the ground-state soliton both follows \eref{eqn:gammaDensityDist} in density distribution and appears well virialized. In physically well-motivated setups where the gravitational potential vanishes at infinity, this scaling relation is expected to hold as long as the halo contribution to the total potential is subdominant or marginally comparable to that of soliton. Granted that the solitonic core arises as the ground-state eigenfunction to the coupled SP equations, the de Broglie relation has been used to estimate the fluctuation timescale of FDM, giving $2\pi\hbar/m_b \sigma_\text{FDM}^2 \sim \tau_{00}$ (e.g. \cite{Marsh:2018zyw, Hui:2020hbq}), where $\sigma_\text{FDM}$ denotes the velocity dispersion of FDM. Note that however $\tau_{00}$ is not a gauge-invariant physical observable. Even more crucially, the true soliton oscillation timescale has $\tau_\text{soliton} \simeq 2.3 \tau_{00}$, as detailed below. We simulate the evolution of a perturbed soliton using the code \texttt{GAMER} \cite{Schive:2017sdo}, where the initial density profiles of six isolated solitons with $m_b = 10^{-22.5}$\textendash$10^{-20.5}$ eV are artificially perturbed and then evolved for $\sim 8$\textendash10 Gyr (top panel of Fig \ref{fig:Gamer2}). For a given unperturbed soliton, we fix $\rho_c$ (adopted from Fig. \ref{fig:SolitonMass}) and slightly alter the outer slope of the soliton profile (e.g. changing the power in \eref{eqn:gammaDensityDist} from $-8$ to $-8.5$). The perturbed soliton then evolves till the oscillations stabilize, after which we adopt the time-averaged peak density as the genuine $\rho_c$ of this oscillating soliton. The soliton retains spherical symmetry during evolution. Our simulations show that the solitary density oscillations exhibit a characteristic timescale \begin{eqnarray} \tau_\text{soliton}(\rho_c) \simeq 92.1 \bigg(\frac{\rho_c}{\text{M}_\odot \text{pc}^{-3}}\bigg)^{-1/2} \text{ Myr}, \label{eqn:NumFirstExcitedPeriod} \end{eqnarray} with a chi-square $\chi^2 = 3.79\times 10^{-3}$ by fitting the frequency spectra (bottom panel of Fig. \ref{fig:Gamer2}). The result is independent of both $m_b$ and the oscillation amplitude $\mathcal{C}$; $\tau_\text{soliton}\rho_c^{1/2}$ varies less than $\mathcal{O}(10^{-2})$ across the range $\mathcal{C} = 0.08$\textendash$0.35$ for different $m_b$ in our simulations. This genuine dynamical timescale of soliton density oscillations is universal, agreeing within 1\% to the values observed in fully self-consistent simulations: soliton-halo systems ($\mathcal{C} \simeq 0.33$) \cite{Veltmaat:2018dfz} and a stabilized soliton system after experiencing significant halo stripping ($\mathcal{C} \simeq 0.25$) \cite{Schive:2019rrw} (see also Sec. \ref{subsec:sim_fdm}). Note that the oscillation amplitude can nevertheless exhibit small variations. \begin{figure} \includegraphics[width=\linewidth]{./FIG1_1_DensityOscillation.pdf} \includegraphics[width=\linewidth]{./FIG1_2_FFTSimulation.pdf} \caption{\textit{Top:} Peak core density $\rho_c$ in a perturbed soliton for $m_b=10^{-22.5}$ (blue) and $10^{-21.3}$ eV (red). The time-averaged densities and extrema (determined by $\mathcal{C}$) are denoted by dashed and dotted lines respectively. \textit{Bottom:} Power spectra of the density fluctuations for $m_b = 10^{-22.5}$ (blue), $10^{-22.1}$ (yellow), $10^{-21.7}$ (green), and $10^{-21.3}$ eV (red). The period of soliton density oscillations depends only on the peak core density, scaling as $\tau_\text{soliton} \propto \rho_c^{-1/2}$ (\eref{eqn:NumFirstExcitedPeriod}).} \label{fig:Gamer2} \end{figure} To obtain an analytical estimate of $\tau_\text{soliton}$, we work perturbatively by considering $\mathcal{C} \rightarrow 0$. To the first-order analysis presented in Appendix \ref{app:PerturbationAppendix}, the coupled perturbations in the wavefunction and sourcing potential are related via \pCAeref{eqn:appRadialPerturbationEq}\pCCeref{app:CoupledDensity}. The eigenfunctions for a unperturbed soliton derived in Appendix \ref{app:SolitonEnergyLevels} conveniently offer a complete set of orthonormal basis to decompose the perturbed wavefunction $\delta u(\gamma, t)$. The resulting order-of-magnitude estimate \eref{eqn:PerturbationPeriod} yields $\tau_p(\rho_c) \simeq 0.58 \tau_\text{soliton}(\rho_c)$; whether high-order contributions can resolve the $\sim 40\%$ discrepancy between $\tau_p(\rho_c)$ and $\tau_\text{soliton}(\rho_c)$ is left to future work. In parallel, the interpretation proposed by Li et al. \cite{Li:2020ryg} attributes $\tau_\text{soliton}$ to the prominent interference timescale between ground-state and first-excited-state eigenfunctions. The reconstructed timescale, $\tau_\text{rec} \simeq 0.83 \tau_\text{soliton}$, depends on the energy difference $|E_{00}-E_{10}|$, which can also be directly read off from Table \ref{tab:SolitonTotalEnergy}. Importantly, the timescales of the ground-state soliton wavefunction \eref{eqn:NumGroundPeriod} and soliton density oscillations \eref{eqn:NumFirstExcitedPeriod} are distinct, the ratio of which is a constant $\tau_\text{soliton}/\tau_{00} \simeq 2.3$ independent of $\rho_c, m_b,$ and $r_c$. This subtlety has not been correctly taken into account by the previous studies on the gravitational heating of a SC by an oscillating soliton \cite{Marsh:2018zyw}. We will show in the succeeding sections that SC heating due to soliton density oscillations is negligible irrespective of $m_b$, given that $\tau_\text{soliton}/\tau_\text{SC}$ deviates noticeably from unity. \subsection{Constraints from Eridanus II}\label{sec:AstroConstraints} The half-light radius of Eri II is estimated to be $r_\text{EII}=277\pm14$ pc \cite{Crnojevi:2016ApJ}, within which the mass enclosed is $M_\text{EII}^{\le r_\text{EII}}=1.2^{+0.4}_{-0.3}\times10^7$ M$_\odot$, corresponding to an average total mass density $\rho_\text{EII} \sim 0.135$ M$_\odot$pc$^{-3}$ \cite{Li:2016utv}. The inferred mass-to-light ratio indicates Eri II is DM-dominated \cite{Li:2016utv}. The central SC of Eri II appears to be an intermediate-age population ($\sim 3$\textendash$13$ Gyr) with a half-light radius of $r_\text{SC}=13\pm1$ pc \cite{Crnojevi:2016ApJ, Simon:2020qsf} and a mass of $M_\text{SC}^{\le r_\text{SC}} \sim 2000$ M$_\odot$ (assuming that the stellar mass-to-light ratio is unity) \cite{Li:2016utv}. The corresponding stellar mass density is $\rho_\text{SC} = 0.217$ M$_\odot$pc$^{-3}$. \begin{figure} \centering \includegraphics[width=\linewidth]{./FIG2_EriIIMassFitting.pdf} \caption{Values of $\rho_c$ (top panel) and $r_c$ (bottom panel) yielding $M_\text{soliton}^{\leq r_\text{EII}}= 1.2^{+0.4}_{-0.3}\times10^7$ M$_\odot$ at 68\% CL. (red shaded). The yellow shaded regions indicate where the gravitational potential within $r_\text{SC}$ is dominated by the soliton. The possible mass contribution from a background halo is excluded in this analysis, and thus the constraints on $\rho_c$ and $r_c$ as plotted here are most robust for $m_b \leq 1.5\times 10^{-21}$ eV, where $3.3 r_c \geq r_\text{EII}$ and consequently $M_\text{soliton}^{\leq r_\text{EII}}/M_\text{halo}^{\leq r_\text{EII}} \gg 1$. A background halo may be required to remedy the unphysically concentrated mass distribution within $r_\text{SC}$ in the gray shaded areas, where $r_c < r_\text{SC}$.} \label{fig:SolitonMass} \end{figure} The measurements of $M_\text{EII}^{\le r_\text{EII}}=1.2^{+0.4}_{-0.3}\times10^7$ M$_\odot$ place constraints on both $\rho_c$ and $r_c$, as plotted in Fig. \ref{fig:SolitonMass}; the soliton density profile thus reduces to an one-parameter scaling, uniquely determined by $m_b$ (or equivalently $\rho_c$). Note that the mass contribution for $\gamma \lesssim 3.3$ is completely dominated by the soliton \cite{Schive:2014dra, Mocz:2017wlg, Pozo:2020ukk}. We hence exclude the mass contribution of a possible background halo, the modeling of which is irrelevant in this analysis for $m_b\leq 1.5\times10^{-21}$ eV as $3.3r_c \geq r_\text{EII}$. The range of $m_b$ shown in Fig. \ref{fig:SolitonMass} can be partitioned into four regions: (1) For $m_b \lesssim 4.6\times10^{-22}$ eV, $r_c > r_\text{EII}$ and the SC is gravitationally self-bound. The peak core density $\rho_c$ asymptotically approaches $\rho_\text{EII}$; $\rho_c \sim $ const then implies $r_c \propto m_b^{-1/2}$. (2) For $4.6\times 10^{-22}\eV \lesssim m_b \lesssim 1.5\times 10^{-21}$ eV, $r_\text{EII}/3.3 < r_c \leq r_\text{EII}$ and $\rho_c$ increases with $m_b$ as the soliton starts entering $r_\text{EII}$. As a result, the SC no longer remains self-bound (yellow shaded regions). (3) For $1.5\times 10^{-21}\eV \lesssim m_b \lesssim 3.8\times 10^{-21}$ eV, $r_\text{SC} < r_c \leq r_\text{EII}/3.3$ and the soliton is well within Eri II. As $M_\text{soliton} \simeq$ const, $\rho_c \propto m_b^6$ and $r_c \propto m_b^{-2}$. (4) Lastly for $m_b \gtrsim 3.8\times 10^{-21}$ eV, $r_c < r_\text{SC}$, indicating a soliton of mass $M_\text{EII}^{\leq r_\text{EII}}$ is contained within $r_\text{SC}$. The FDM density typically drops by 1.5 to 3 orders of magnitude from $\rho_c$ at the soliton-halo profile transition for a wide range of possible distributions \cite{Schive:2014hza, Pozo:2020ukk, Mocz:2017wlg}. A conservative estimate suggests $M_\text{halo}^{\leq r_\text{EII}}/M_\text{soliton}^{\leq r_\text{EII}} \gtrsim \mathcal{O}(1)$ for $r_c \lesssim r_\text{EII}/4$, which implies the mass contribution of a background halo is still subordinate, albeit possibly not negligible for $1.5\times10^{-21} \lesssim m_b \lesssim 3.8\times10^{-21}$ eV. For $m_b \gtrsim 3.8\times 10^{-21}$ eV (gray shaded regions in Fig. \ref{fig:SolitonMass}), the soliton-only constraints become likely unphysical as the mass of entire Eri II is mostly contained within $r_\text{SC}$, which signals the presence of a background halo to prevent $r_c$ from shrinking below $r_\text{SC}$. Once the halo is introduced, however, the SC could be vulnerable to tidal disruption due to soliton random walk excursions \cite{Schive:2019rrw} or diffusion heating due to granular density fluctuations in the halo \cite{Marsh:2018zyw, Amr2020}. A more thorough numerical investigation is needed nevertheless, as \cite{Schive:2019rrw} only tested the impact of soliton random walk for $m_b = 8\times 10^{-23}\eV$ and the efficiency of diffusion heating has not been rigorously confirmed. \section{Star Cluster Heating}\label{Heating} \begin{figure} \includegraphics[width=\linewidth]{./FIG3_TimeDomain.pdf} \caption{Dynamical timescales of Eri II star cluster (SC) $\tau_\text{SC}$ (\eref{eqn:MassOrbitalPeriod}, red), soliton density fluctuations $\tau_\text{soliton}$ (\eref{eqn:NumFirstExcitedPeriod}, orange), and the ground-state soliton wavefunction $\tau_{00}$ (\eref{eqn:NumGroundPeriod}, purple) in our analysis (solid). The halo mass contribution is excluded (Fig. \ref{fig:SolitonMass}). The apparent discrepancy between the aforementioned timescales (solid) and those adopted by \MN\,\cite{Marsh:2018zyw} (dashed, $^\dagger$) are discussed in Sec. \ref{app:TragetPaper}.} \label{fig:Gamer2PLOT2} \end{figure} The survival of Eri II SC was previously investigated by \MN\, to constrain the viable FDM particle mass. They considered the effect of gravitational heating on the SC caused by soliton density oscillations and identified several resonance bands with efficient heating between $10^{-21} \text{ eV}\lesssim m_b \lesssim 5\times 10^{-21}$ eV. On the contrary, we demonstrate in this section that, with an improved estimate of $\tau_{\rm soliton}$, SC heating becomes negligible for $m_b \lesssim 3.8\times10^{-21}\eV$. We model the single-mode soliton density oscillations with an amplitude $\mathcal{C}$, frequency $\omega_\text{osc}=2\pi/\tau_\text{osc}$, and an arbitrary phase $\phi = \left[0,2\pi\right)$ by \begin{eqnarray} \rho_c \Big[1+ \mathcal{C}\sin(\omega_\text{osc}t+\phi)\Big]\equiv\rho_cs(t), \label{eqn:SolitonOsc} \end{eqnarray} which propagates the time-dependence to the soliton radius $r_c s(t)^{-1/4}$ (\eref{eqn:SolitonRadiusEriII}) and consequently the soliton enclosed mass $M_\text{soliton}^{\leq r}(\gamma = r r_c^{-1} s(t)^{1/4}; \rho_c s(t), m_b)$ (\eref{eqn:SolitonEncMass}). Here both $\rho_c$ and $r_c$ are the unperturbed (time-averaged) values; the core is assumed to retain the soliton profile during oscillations. We treat $\omega_\text{osc}$ as a free variable to locate the resonance bands. \begin{figure} \includegraphics[width=\linewidth]{./FIG4_SemiMajorEvoC1.pdf} \caption{Time evolution of test star's orbital radius normalized to $r_\text{SC} = 13$ pc, for $\mathcal{C} = $ 0.1 (green), 0.3 (red), 0.5 (yellow), and 0.7 (blue), fixing $m_b = 10^{-22}$ eV. The top, middle, bottom panels has $\tau_\text{osc} = \tau_\text{soliton}$, $\tau_\text{s.r.,2}$, and $\tau_\text{d.r.}$ (\pAeref{eqn:DirectResonanceaeff}\pBeref{eqn:Superharmonics}) respectively; \eref{eqn:MassOrbitalPeriod} yields $\tau_\text{SC} = 75.8$ Myr. Gravitational heating of Eri II SC is ineffective, suggested by the realistic modeling of soliton oscillations with $\tau_\text{osc} = \tau_\text{soliton}$ (see also Sec. \ref{subsec:sim_fdm}).} \label{fig:TidalHeatingResultsC} \end{figure} Consider a star of mass $m_\star$ at an initial distance $r_0=r_\text{SC} = 13$ pc with respect to the center of mass of SC, following a similar setup in \MN\, (see also Sec. \ref{app:TragetPaper}). Here we assume the centers of the SC and Eri II coincide to a good approximation. The unperturbed circular orbit has an orbital frequency \footnote{The cored density profiles of both a soliton and the SC, roughly following $M^{\leq r} \propto r^3$ within their respective cutoff radii, suggest that this estimate of $\omega_{\text{SC},0}$ is insensitive to the choice of $r_\text{SC}$.} \begin{eqnarray} \omega_{\text{SC},0} &&=\sqrt{\frac{G\big(M_\text{SC}^{\le r_\text{SC}}+M_\text{soliton}^{\leq r_\text{SC}}\big)}{r_\text{SC}^3}}. \label{eqn:MassOrbitalPeriod} \end{eqnarray} Figure \ref{fig:Gamer2PLOT2} compares the relevant timescales of soliton dynamics under the soliton-only constraints (i.e. ignoring the mass contribution from an outer halo), where $\tau_\text{SC} = 2\pi/\omega_{\text{SC},0}$ (red), $\tau_\text{soliton}$ (orange), and $\tau_{00}$ (purple) are plotted in solid. For $m_b \lesssim 4.6\times 10^{-22}$ eV, the SC orbital period $\tau_\text{SC}$ roughly remains constant (so are $\tau_{00}$ and $\tau_\text{soliton}$) since $\rho_c \sim$ const. Moreover, $\tau_\text{soliton}/\tau_\text{SC} \sim 2$\textendash3 for $m_b \lesssim3.8\times 10^{-21}$ eV. For $m_b \gtrsim 3.8\times 10^{-21}$ eV, $\tau_\text{SC}$ also approaches a constant value as $r_c \lesssim r_\text{SC}$ and $M_\text{soliton}^{\leq r_\text{SC}}\sim M_\text{EII}^{\leq r_\text{EII}}$, leading to $\tau_\text{soliton}/\tau_\text{SC} \lesssim 1$. However if a background halo is introduced to ensure $r_c \geq r_\text{SC}$, then $\tau_\text{soliton}/\tau_\text{SC} \gtrsim 2$ even for $m_b \gtrsim 3.8\times10^{-21}$ eV. See Sec. \ref{app:TragetPaper} regarding remarks about $^\dagger\tau_{00} \equiv {^\dagger\tau_\text{soliton}}$ (blue, dashed) and $^\dagger\tau_\text{SC}$ (green, dashed) presented in \MN, which are apparently inconsistent with our analysis. \begin{figure} \includegraphics[width=\linewidth]{./FIG5_SemiMajorEvoTauRatio.pdf} \caption{Time-averaged radial migration of a test star over 10 Gyr. \textit{Top:} We compare three fiducial values of $\mathcal{C} = 0.1$ (red), $0.3$ (yellow), and $0.5$ (green), fixing $m_b = 10^{-22}$ eV. Direct resonance $\tau_\text{d.r.}$ and superharmonic resonances $\tau_\text{s.r.}$ are given by \pAeref{eqn:DirectResonanceaeff}\pBeref{eqn:Superharmonics}. \textit{Bottom:} We take $m_b = 10^{-22}$ eV (yellow), $10^{-21.5}$ eV (light blue), and $10^{-21}$ eV (dark blue), fixing $\mathcal{C} = 0.3$. The efficiency of gravitational heating peaks at $\tau_\text{osc}/\tau_\text{SC} \simeq 0.5$\textendash$0.8$ and is negligibly low for the realistic case with $\tau_\text{osc}/\tau_\text{SC} \simeq 2\textendash3$.} \label{fig:NumericalRAverage} \end{figure} The time evolution of the orbital radius reads \begin{eqnarray} \ddot{r} = \omega_{\text{SC}}^2 r- \frac{G\big[M_\text{SC}^{\le r_\text{SC}}+M_\text{soliton}^{\leq r(t)}(\gamma;\rho_cs(t),m_b)\big]}{r^2}, \;\;\;\;\;\;\;\label{eqn:TidalHeatingMass} \end{eqnarray} where the conservation of angular momentum gives $\omega_\text{SC}^2r = \omega_{\text{SC},0}^2r_\text{SC}^4r^{-3}$.The same radial equation of motion (\eref{eqn:appTidalHeatingFullyGeneral}) can also be obtained via the Hamiltonian formalism by treating the density fluctuations as a time-dependent perturbation in the ground-state soliton potential (see Appendix \ref{app:HamiltonianPerturbation}). We then linearize \eref{eqn:TidalHeatingMass} to the first non-vanishing order in $\gamma$ and identify two types of instabilities: the direct resonance $\tau_\text{d.r.}$ (\eref{eqn:DirectResonanceaeff}) and a series of secondary superharmonic resonances $\tau_{\text{s.r.},n} \simeq n \tau_\text{d.r.}$ for $n \geq 2, n \in \mathbb{Z}^{+}$ (\eref{eqn:Superharmonics}), as detailed in Appendix \ref{app:Resonance}. Figure \ref{fig:TidalHeatingResultsC} compares the time evolution of the orbital radius obtained by numerically integrating \eref{eqn:TidalHeatingMass} for $\mathcal{C} = 0.1$ (green), $0.3$ (red), $0.5$ (yellow), and $0.7$ (blue), fixing $r_0 = r_\text{SC} = 13$ pc, $\phi =0$, and $m_b = 10^{-22}$ eV (corresponding to $\rho_c = 0.149$ M$_\odot$pc$^{-3}$). The top, middle, and bottom panels adopt $\tau_\text{osc} = \tau_\text{soliton}, \tau_\text{s.r.,2},$ and $\tau_\text{d.r.}$ respectively. The orbit is non-Keplerian for $\mathcal{C} \neq 0$. We observe that gravitational heating is inefficient for the realistic modeling of soliton oscillations with $\tau_\text{osc} = \tau_\text{soliton}$; this conclusion is further supported by the self-consistent numerical simulations elaborated in Sec. \ref{subsec:sim_fdm}. The effect of orbital resonances becomes more pronounced when $\tau_\text{osc}/\tau_\text{SC} \rightarrow 1$ falls on resonance bands and for large $\mathcal{C}$. Figure \ref{fig:NumericalRAverage} shows the shift in orbital radius $\smash{r_\text{SC,final} \equiv\int_{9 \text{ Gyr}}^{10 \text{ Gyr}}r(t) dt \big/1 \text{ Gyr}}$ \footnote{The resulting $r_\text{SC,final}$ converges for alternative choices of averaging over $8$\textendash$10$ Gyr or adopting all time maximum.}. The results converge for arbitrary choices of $\phi$ and $ -25\leq \dot{r}_0\leq 25 $. The top panel plots the time-averaged radial migration for $\mathcal{C} = 0.1$ (red), $0.3$ (yellow), and $0.5$ (green) as a function of $\tau_\text{osc}/\tau_\text{SC}$ with $m_b = 10^{-22}$ eV. In the bottom panel, we take $m_b = 10^{-22}$ eV (yellow), $10^{-21.5}$ eV (light blue), and $10^{-21}$ eV (dark blue), fixing $\mathcal{C} = 0.3$. Non-trivial orbital resonances are observed only for $\tau_\text{osc}/\tau_\text{SC} \simeq 0.5$\textendash$0.8$, which confirms a negligible heating effect of SC given $\tau_\text{soliton}/\tau_\text{SC} \sim 2$. The resonance bands widen as $\mathcal{C}$ increases. For large $m_b$, the constrained $\rho_c$ and consequently the (initial) angular momentum increase, which qualitatively shifts both $\tau_\text{d.r.}$ and $\tau_\text{s.r.,n}$ towards smaller $\tau_\text{osc}/\tau_\text{SC}$. \begin{figure} \includegraphics[width=\linewidth]{./FIG6_StabilityDiagramPeriods.pdf} \caption{Stability diagram of \eref{eqn:TidalHeatingMass}, fixing $\mathcal{C} = 0.3$ with soliton-only constrained $\rho_c$ and $r_c$ applicable for $m_b \lesssim 4\times 10^{-21}$ eV (Fig. \ref{fig:SolitonMass}). The soliton oscillation timescale $\tau_\text{soliton}$ (orange) shows no overlap with $\tau_\text{d.r.}$ or $\tau_{\text{s.r.},2}$ (red shaded regions where $r_\text{SC,final}/r_\text{SC} \geq 2$ over $10$ Gyr). The ineffective SC heating is in clear contradiction with the result by \MN\, \cite{Marsh:2018zyw} (purple resonance bands); see also Sec. \ref{app:Comment2}.} \label{fig:TidalHeatingResultsRhoc} \end{figure} Figure \ref{fig:TidalHeatingResultsRhoc} plots the SC stability diagram; the red shaded areas indicate the parameter space with efficient orbital heating, defined as $r_\text{SC,final}/r_\text{SC} \geq 2$ over 10 Gyr. The genuine oscillation timescale of an FDM soliton $\tau_\text{soliton}$ (orange) does not intersect with either $\tau_\text{d.r.}$ or $\tau_{\text{s.r.},2}$; the heating mechanism is therefore inefficient. The SC resonance bands identified in NM2019 are shown in purple; this stark inconsistency is further discussed in Sec. \ref{app:TragetPaper}. For $m_b \gtrsim 4\times 10^{-21}$ eV (gray shaded), the significant uncertainty in subhalo profile modeling propagates to the soliton-only constraints on $\rho_c$ and $r_c$, rendering $\tau_\text{soliton}/\tau_\text{SC}$ poorly constrained by observations. In summary, the single-test-star toy model firmly indicates that gravitational heating from an oscillating soliton is inefficient. For $m_b \lesssim 4.6\times 10^{-22}$ eV, the SC is self-bound as $\rho_c \sim$ const and $\rho_c\lesssim \rho_\text{SC} = 0.217$ M$_\odot$pc$^{-3}$. The $m_b$-independent relation in the dynamical timescales, $\tau_\text{soliton} > 2\tau_\text{SC}$, results in a negligible effect of SC heating. For $m_b \gtrsim 4.6\times 10^{-22}$ eV, we have $\rho_c \gtrsim \rho_\text{SC}$ as $\rho_c$ increases with $m_b$. The SC is not self-bound and consequently $\tau_\text{SC}$ is determined by $\rho_c$. The timescales $\tau_\text{SC}$ and $\tau_{00}$ differ $\lesssim 30\%$ under the soliton-only constraints for the applicable range of $m_b$, as detailed in Sec. \ref{sec:SolitonTimeScale}. Consequently, we have $\tau_\text{SC} \simeq \tau_{00} \sim \tau_\text{soliton}/2$ and hence $\tau_\text{soliton}/\tau_\text{SC} \simeq 2$, suggesting again ineffective heating. \begin{figure} \includegraphics[width=\linewidth]{./FIG7_ext_half-light-radius_evolution.pdf} \caption{Time evolution of the SC's projected half-light radius $\Rhl$ in an oscillating external soliton potential with period $\tosc$. $\Rhl$ is normalized to its initial value $\Rhli$ and $\tosc$ to the characteristic timescale of the SC $\tSC$. Among the four representative cases with different $\tosc/\tSC$, only the case $\tosc/\tSC=0.9$ exhibits significant heating, with $\Rhl$ increasing by an order of magnitude within $1\Gyr$. See also Figs. \ref{fig:simu_ext_proj_dens} and \ref{fig:simu_ext_profile}.} \label{fig:simu_ext_Rhl_evolution} \end{figure} \section{Simulations} \label{sec:sim} In this section, we perform three-dimensional simulations to compare against the single-particle toy model presented in Sec. \ref{Heating}. We first model the soliton oscillations as an external potential, followed by self-consistent FDM simulations. \subsection{N-body Simulations with an External Potential} \label{subsec:sim_ext} \subsubsection{Simulation setup} \label{subsubsec:sim_ext_setup} \begin{figure} \includegraphics[width=\linewidth]{./FIG8_ext_proj_dens.pdf} \caption{Projected mass of the SCs and the associated half-light radii (blue circles). The top left panel shows the initial condition. The other three show the results with different values of $\tosc/\tSC$ at the respective epoch with a maximum half-light radius $\Rhlmax$ (see \fref{fig:simu_ext_Rhl_evolution}). Heating by an oscillating soliton is only manifest for $\tosc/\tSC=0.9$. Visualization is done with \texttt{yt} \cite{yt}.} \label{fig:simu_ext_proj_dens} \end{figure} The SC is modeled by a Plummer sphere with a total mass $8.7\times10^3\Msun$ (within a cut-off radius of $100\pc$), a projected half-light radius $\Rhli = 22\pc$, and a corresponding peak stellar mass density $\rhosc \sim 0.22\Msun\pc^{-3}$. We follow the treatment in \cite{Aarseth1974AA37183A} to sample the stellar velocity. The soliton is modeled by an external potential following \eref{eqn:VExcitedStatesAnalysis} that fixes $m_b=10^{-22} \eV$ and $r_c=650\pc$, yielding a soliton mass density $\rhosol \sim 0.11\Msun\pc^{-3}$. The centers of the SC and the soliton coincide. The SC is self-bound but the gravitational influence of the soliton is non-negligible, as $\rhosc \sim 2 \rhosol$. We thus first evolve the SC in a \emph{fixed} soliton potential until it relaxes and then take this relaxed SC as the initial condition, which has a smaller projected half-light radius $\Rhli \sim 13\pc$ and a characteristic timescale $\tSC \sim 71\Myr$ (\eref{eqn:MassOrbitalPeriod}). The soliton then oscillates according to \eref{eqn:SolitonOsc} with varying $\Wosc=2\pi/\tosc$ for a fixed $\mathcal{C}=0.5$ over $2\Gyr$. We use a direct N-body method with a fourth-order Hermite scheme, individual time-step, and GPU acceleration \cite{Schive:2007mn} to evolve the SC. The gravitational softening length is set to $0.1 \textrm{--} 1.0\pc$. The total number of particles is $2\times10^4 \textrm{--} 8\times10^4$, leading to a particle mass resolution of $\sim 0.4 \textrm{--} 0.1\Msun$. We have verified that the simulation results are insensitive to the adopted softening length and particle mass resolution. The maximum relative error in angular momentum conservation is less than $0.2\%$ in all simulations. \subsubsection{Orbital evolution and heating efficiency} \label{subsubsec:sim_ext_results} First, we present the results of four representative cases with $\tosc/\tSC= 0.5, 0.9, 2.0,$ and $3.0$. Figure \ref{fig:simu_ext_Rhl_evolution} shows the time evolution of the projected half-light radius $\Rhl(t)$. Only the case $\tosc/\tSC=0.9$ exhibits prominent heating, with a maximum half-light radius $\Rhlmax \sim 11.5 \Rhli$ at $t \sim 0.7\Gyr$. For other three fiducial values of $\tosc/\tSC$ deviating notably from unity, the SC is stable against an oscillating external soliton potential, with $\Rhlmax \lesssim 1.3 \Rhli$. Figure \ref{fig:simu_ext_proj_dens} compares the initial projected stellar mass with the resulting configurations when $\Rhl(t)=\Rhlmax$ for $\tosc/\tSC = 0.5, 0.9, 2.0$. The corresponding density profiles are plotted in \fref{fig:simu_ext_profile}. The efficiency of gravitational heating is greatly attenuated for $\tosc/\tSC \neq \mathcal{O}(0.9)$. \begin{figure} \includegraphics[width=\linewidth]{./FIG9_ext_profile.pdf} \caption{Stellar density profiles. We plot the initial condition (black) and the configurations at the respective epoch with maximum half-light radii for $\tosc/\tSC = 0.5$ (purple), $0.9$ (red), $2.0$ (green), and $3.0$ (blue) (see \fref{fig:simu_ext_Rhl_evolution}). The vertical lines mark the initial half-light radius ($\Rhli$) and the maximum half-light radius for $\tosc/\tSC=0.9$ ($\Rhlmax^{\tau=0.9}$), only where the effect of heating is appreciable.} \label{fig:simu_ext_profile} \end{figure} We then perform an extensive set of simulations to probe the heating efficiency with $0.5 \le \tosc/\tSC \le 3.1$. Figure \ref{fig:simu_ext_Rhl_max} shows $\Rhlmax$ as a function of $\tosc/\tSC$. The heating efficiency peaks around $\tosc/\tSC \sim 0.9$ and drops dramatically when $\tosc/\tSC$ deviates from unity. This result is in good agreement with the single-particle toy model (\fref{fig:NumericalRAverage}), where the heating efficiency peaks at $\tosc/\tSC = 0.835$ for $m_b = 10^{-22}$ eV and $\mathcal{C} = 0.5$. The half-mass containment for a Plummer profile lies between $0.483\Rhli$\textendash$1.52\Rhli$, within which the variation in $\tau_\text{SC}(r)$ is less than $\sim 50\%$ for $\rho_c/\rho_\text{SC} \geq 0.622$ consistent with observations. The assumption of no-shell crossing in the single-particle description is therefore largely applicable to the more realistic SC modeling. \begin{figure} \includegraphics[width=\linewidth]{./FIG10_ext_half-light-radius_max.pdf} \caption{Maximum expansion of the SC's half-light radius caused by an external oscillating soliton potential with period $\tosc$. Gravitational heating is most prominent at $\tosc/\tSC \sim 0.9$ and becomes ineffective as $\tosc/\tSC$ deviates from unity.} \label{fig:simu_ext_Rhl_max} \end{figure} \subsection{Self-consistent FDM Simulations} \label{subsec:sim_fdm} To further test whether the SC can survive within a realistic FDM soliton, in this subsection we evolve both the SC and soliton in a \emph{self-consistent} approach. \subsubsection{Simulation setup} \label{subsubsec:sim_fdm_setup} The SC is embedded and evolved in the center of an FDM halo mimicking Eri II, following the simulation setup in \cite{Schive:2019rrw}. In this work, we focus on the case where the outer halo surrounding the soliton has been largely stripped away by an external tidal field; subsequently, soliton random walk excursions are greatly suppressed and the effect thereof can be safely ignored (see Figs. 3 and 4 therein). This assumption is crucial, for otherwise the off-center separation between the soliton and the SC can tidally disrupt the SC within $\sim 1\Gyr$, as demonstrated in \cite{Schive:2019rrw}. The adopted FDM particle mass $m_b=8.0\times10^{-23} \eV$ and the soliton half-density radius $r_c \sim 830\pc$ together give $\rhosol \sim 6.5\times10^{-2}\Msun\pc^{-3}$. The SC parameters are the same as those adopted in the previous subsection, except that we do not need to relax the SC further due to a lower value of $\rhosol$ here. The corresponding characteristic timescale of the SC is $\tSC \sim 120\Myr$ (\eref{eqn:MassOrbitalPeriod}). The initial density distribution is illustrated in \fref{fig:simu_dyn_proj_dens}. \begin{figure} \includegraphics[width=\linewidth]{./FIG11_dyn_proj_dens.pdf} \caption{Projected dark matter density of the initial condition in a self-consistent FDM simulation. The bright region represents the soliton and the blue solid circle marks the SC. The ambient medium of soliton has been largely stripped away by an external tidal field. Visualization is done with \texttt{yt} \cite{yt}.} \label{fig:simu_dyn_proj_dens} \end{figure} We use the code \texttt{GAMER} \cite{Schive:2017sdo} to evolve FDM and the SC self-consistently. It supports adaptive mesh refinement and hybrid CPU/GPU parallelization. We adopt a simulation box of size $250\kpc$ covered by a $128^3$ root grid and nine refinement levels, leading to a maximum spatial resolution of $\sim 3.8\pc$ capable of resolving the compact SC. The particle mass resolution is $3.6\times 10^{-2}\Msun$. See \cite{Schive:2019rrw} for more details. We run the simulations for $5\Gyr$. \subsubsection{Heating efficiency} \label{subsubsec:sim_fdm_results} Figure \ref{fig:simu_dyn_sol-osc} shows the time evolution of the peak soliton density. It demonstrates that the core oscillation is an intrinsic property that persists even in the absence of an outer halo. The gravitational influence of SC has no discernible impact on the density oscillations of the soliton, as $M_\text{soliton}/M_\text{SC} = \mathcal{O}(10^5)$. The observed oscillation amplitude is $C \sim 0.25$ and the oscillation period is $\tsol \sim 370\Myr$, in good agreement with \eref{eqn:NumFirstExcitedPeriod}. \begin{figure} \includegraphics[width=\linewidth]{FIG12_dyn_soliton-oscillation.pdf} \caption{Time evolution of the peak soliton density in a self-consistent FDM simulation. The soliton oscillates with a characteristic period of $\tsol \sim 370\Myr$, consistent with \eref{eqn:NumFirstExcitedPeriod}. The oscillation amplitude is $C \sim 0.25$.} \label{fig:simu_dyn_sol-osc} \end{figure} Most importantly, we note that $\tsol/\tSC \sim 3$, suggesting negligible heating of the SC from the soliton oscillations (see also Figs. \ref{fig:NumericalRAverage} and \ref{fig:simu_ext_Rhl_max}). This is confirmed by \fref{fig:simu_dyn_Rhl_evolution}, which plots the projected half-light radius as a function of time. The maximum expansion of the half-light radius within $5\Gyr$ is found to be only $\Rhlmax/\Rhli \sim 1.05$. Figure \ref{fig:simu_dyn_profile} compares the stellar density profiles at $t=0$ and $5\Gyr$, further demonstrating that the SC is stable against the soliton oscillations. \begin{figure} \includegraphics[width=\linewidth]{./FIG13_dyn_half-light-radius_evolution.pdf} \caption{Time evolution of the SC's half-light radius $\Rhl$, normalized to its initial value $\Rhli$, in a self-consistent oscillating soliton. The effect of heating appears to be negligible.} \label{fig:simu_dyn_Rhl_evolution} \end{figure} \section{Comparison with PREVIOUS FDM Constraints from Eridanus II}\label{app:TragetPaper} The constraints on FDM particle mass presented by \MN\, were set by the SC resonances, diffusion approximation, and subhalo mass function, together yielding $m_b \gtrsim 10^{-19}$ eV. In Sec. \ref{app:Comment1}, we discuss the dynamical timescales erroneously cited in \MN. We point out in Sec. \ref{app:Comment2} some mistakes and convergence issues with the perturbation calculations on gravitational heating presented therein. In Sec. \ref{app:Comment0}, we examine the validity of FDM constraints set by the subhalo mass function. \subsection{Dynamical Timescales}\label{app:Comment1} \begin{figure} \includegraphics[width=\linewidth]{./FIG14_dyn_profile.pdf} \caption{Density profiles in a self-consistent FDM simulation. The dashed and solid lines compare the stellar profiles at the beginning and end of the simulation, demonstrating that the SC is stable against an oscillating soliton. For reference, the dash-dotted and dotted lines show the soliton at its minimum and maximum densities.} \label{fig:simu_dyn_profile} \end{figure} As stated in \MN, the diffusion approximation was applicable in the regime that satisfied $^\dagger\tau_\text{SC} \gg$ $^\dagger\tau_\text{soliton}$. In the limit $^\dagger\tau_\text{SC} \ll$ $^\dagger\tau_\text{soliton}$, they investigated SC heating, treating the SC as a virial system. The dynamical timescale of stochastic fluctuations (halo granules) was taken to be the same as $\tau_\text{soliton}$ in \MN. The validity of calculations therein rest first and foremost upon correctly determining the scaling relations of both $\tau_\text{SC}$ and $\tau_\text{soliton}$ as functions of $m_b$. The accurate analytical modeling of the soliton-SC system is equivalently crucial in yielding robust constraints. The erroneous dynamical timescales of both the SC and an FDM soliton presented therein can be summarized as follows: (1) In computing $^\dagger\tau_\text{SC}$, the mass contribution of a soliton enclosed within $r_\text{SC}$ was completely neglected, which was only marginally correct for $m_b \lesssim 4.6\times 10^{-22}$ eV (albeit still with $\gtrsim 24\%$ error compared to $\tau_\text{SC}$, \eref{eqn:MassOrbitalPeriod}, in this limit). (2) The velocity dispersion of FDM $\sigma_\text{FDM}$ was assumed constant (i.e. independent of $m_b$) and equal to that of stars $\sigma_\text{EII}$, which was generally incorrect and led to the problematic scaling relation $^\dagger\tau_\text{soliton} \propto m_b^{-1}$. (3) The two distinct timescales of a soliton were conflated $^\dagger\tau_{00} =$ $^\dagger\tau_\text{soliton}$ in \MN. See Sec. \ref{sec:SolitonTimeScale} for discussions on this subtlety and Fig. \ref{fig:Gamer2PLOT2} for a comparison between the dynamical timescales adopted in \MN\, and this work; the discrepancy is further discussed below. The dynamics of Eri II SC is determined gravitationally by the averaged mass enclosed within the orbit of individual member star. In \MN, the characteristic timescale of the SC was estimated by taking a test star on a Keplerian orbit with a semi-major axis $r_\text{SC}$: $^\dagger\tau_\text{SC} = 2\pi r_\text{SC}^{3/2}/(GM^{\leq r_\text{SC}}_\text{total})^{1/2}$, where the total enclosed mass $M^{\leq r_\text{SC}}_\text{total} = M_\text{soliton}^{\leq r_\text{SC}}+ M_\text{SC}^{\le r_\text{SC}}$ should include contributions from both the SC and soliton. For the fiducial $\rho_c = 0.15\Msun\pc^{-3}$ adopted in \MN, the enclosed soliton mass shall read $M_\text{soliton}^{\leq r_\text{SC}} \simeq 1380\Msun$ instead of the value $330\Msun$ quoted therein. The underestimated $M_\text{soliton}^{\leq r_\text{SC}}$ led to the incorrect approximation $M_\text{total}^{\leq r_\text{SC}} \sim M_\text{SC}^{\leq r_\text{SC}}$ (i.e. Eq. (18) in \MN). We have however demonstrated that $M_\text{soliton}^{\leq r_\text{SC}}/M_\text{SC}^{\le r_\text{SC}} \gtrsim \mathcal{O}(1)$ always holds regardless of the value of $m_b$ (see Fig. \ref{fig:SolitonMass}). We have verified $\tau_\text{soliton} \simeq 2.3\tau_{00}$ in Sec. \ref{OscillationFrequencyRevisit}, as opposed to the assumption $^\dagger\tau_\text{soliton} = {^\dagger\tau_{00}}$ made by \MN. Operating under such an assumption, \MN\, then invoked the de Broglie relation to infer $^\dagger \tau_\text{soliton} = 2\pi \hbar/m_b \sigma_\text{FDM}^2$. The velocity dispersion of FDM was further assumed to be the same as that of Eri II, with $\sigma_{\text{1D,FDM}} = \sigma_\text{1D,EII} = 6.9 \text{ km s}^{-1}$ \cite{Li:2016utv}. The resulting scaling relation reads \begin{eqnarray} ^\dagger\tau_{00} &=& ^\dagger\tau_\text{soliton} \simeq 825\bigg(\frac{m_b}{10^{-22} \text{ eV}}\bigg)^{-1} \text{ Myr}, \end{eqnarray} with a parameter dependence that is generally incorrect (cf. \eref{eqn:NumGroundPeriod}) \footnote{The parametric dependence of $\sigma_{\text{FDM}}$ can be understood heuristically by noting that $\tau_{00} \propto m_b^{-1}\sigma_{\text{FDM}}^{-2}$ and $\tau_{00} \propto \rho_c^{-1/2}$, yielding $\sigma_{\text{FDM}} \propto \rho_c^{1/4}m_b^{-1/2}$.}. For instance, when the physical size of a soliton gets larger than $r_\text{EII}$ for $m_b \lesssim 4.6 \times 10^{-22}$ eV, we have $\tau_\text{soliton} \sim$ const (i.e. independent of $m_b$; see \fref{fig:Gamer2PLOT2}). For $m_b > 10^{-20}$ eV, the condition $^\dagger\tau_\text{SC} \gg$ $^\dagger\tau_\text{granules}$ ensured the validity of diffusion approximation calculated in \MN, where the timescales of halo granular density fluctuations and of soliton oscillations were assumed identical $^\dagger\tau_\text{granules} \equiv {^\dagger\tau}_\text{soliton}$. A comparable FDM constraint was derived in \cite{Amr2020} based on similar assumptions (e.g. anchoring the velocity dispersion of FDM to $\sigma_\text{EII}$). Our present analysis is not directly applicable to this large-mass regime, because the soliton oscillation timescale exhibits large uncertainty due to the unconstrained background halo profile modeling (gray shaded area in Fig.~\ref{fig:SolitonMass}). Additionally, the timescale of granular density fluctuations in relation to $\tau_\text{soliton}$ is also uncertain. Hence whether diffusion heating for $m_b > 10^{-20}\eV$ is as efficient as reported by \MN\, and \cite{Amr2020} remains to be investigated. \subsection{Star Cluster Heating Calculations}\label{app:Comment2} Gravitational heating caused by soliton density oscillations on the SC was investigated in \MN, where several (possibly) invalid assumptions were made in calculations therein. Our single-test-star model (\eref{eqn:TidalHeatingMass}) does not rely on these assumptions and yields more accurate predictions, consistent with both the N-body and FDM simulations. To formulate the SC-soliton system, \MN\, first explicitly assumed the SC to be virialized for all time and have the unperturbed Hamiltonian $H_0 = G M_\text{SC}^{\le r_\text{SC}} m_\star/(2 a_0)$ for a star with mass $m_\star$ (Eq. (17) therein) with the semi-major axis $a_0 = r_\text{SC}$. The condition $^\dagger \tau_\text{SC} \ll$ $^\dagger\tau_\text{soliton}$ must be satisfied to be physically self-consistent, or otherwise we would need to demand $\mathcal{C} \ll 1$, given $M_\text{soliton}^{\leq r_\text{SC}}/M_\text{SC}^{\le r_\text{SC}} \gtrsim \mathcal{O}(1)$ holds true for all $m_b$ range of interest. Neither assumption is valid however, as shown in Sec. \ref{sec:SolitonTimeScale}. Here the mass contribution of the soliton was also erroneously neglected under the assumption $M_\text{soliton}^{\leq r_\text{SC}}/M_\text{SC}^{\le r_\text{SC}} \ll 1$. Furthermore, the soliton mass enclosed within $r_\text{SC}$ was also taken fixed (an incorrect assumption; see Sec. \ref{Heating}) to yield the perturbation Hamiltonian $\Delta H = \mathcal{C} [-G M_{\text{soliton}}^{\leq r_\text{SC}} m_\star/r(t)]\cos ({^\dagger\omega_\text{soliton}t})$, where $r(t)$ was the unperturbed orbit. The semi-major axis then evolved as \begin{eqnarray} \dot{a} &=& 2\mathcal{C} \bigg(\frac{M_{\text{soliton}}^{\leq r_\text{SC}}}{M_\text{SC}^{\le r_\text{SC}}}\bigg) a_0^{2\dagger}\omega_\text{soliton}\frac{\sin ({^\dagger\omega_\text{soliton}t})}{r(t)}, \label{eqn:TidalResonanceIntegral} \end{eqnarray} which was integrated over $\text{Max}(^\dagger\tau_\text{orb}, {^\dagger\tau_\text{soliton}})$ with the initial condition $a = a_0$, fixing the eccentricity $e = 0.5$. The final semi-major axis $a_\text{final}$ was taken to be the average over $\tau_\text{ave} \eee10\times \text{Max}(^\dagger\tau_\text{orb}, {^\dagger\tau_\text{soliton}})$, and orbital resonances were considered efficient for $a_\text{final}/a_0 \geq 2$. In this formulation, $^\dagger \tau_\text{SC}$ appeared only in the denominator of $\Delta H$ for elliptical orbits. For $e=0$, the solution of \eref{eqn:TidalResonanceIntegral} now depends solely on $^\dagger \tau_\text{soliton}$ as $r(t) = r_\text{SC} = \rm{const}$. Namely, $^\dagger\tau_\text{SC}$ becomes irrelevant in determining the efficiency of orbital resonances, which is unphysical. Quantifying the potential perturbation induced by soliton density fluctuations by approximating $M^{\leq r(t)}_\text{soliton}$ as a point mass in $\Delta H$ effectively increases the oscillation amplitude $\mathcal{C}$ by a factor of two as detailed below. If one adopts the unperturbed soliton potential $V_0(\gamma; \rho_c, m_b)$, such an approach (perturbation Hamiltonian) to the first non-vanishing order in $\gamma$ yields (cf. \eref{eqn:appTidalHeating2ndODE}) \begin{eqnarray}\label{eqn:appMarsh} \ddot{r} =\omega_\text{SC}^2r- G\bigg\{\frac{M_\text{SC}}{r^2}+\frac{4\pi}{3}\rho_c r \Big[1+ 2\mathcal{C}\sin({^\dagger\omega_\text{osc}t+\phi)}\Big]\bigg\}.\nonumber\\ \end{eqnarray} However, the density oscillation amplitude should be $\mathcal{C}$ instead of $2\mathcal{C}$ by construction (see \pAeref{eqn:SolitonOsc}\pBeref{eqn:TidalHeatingMass}). \Seref{eqn:appMarsh} in turn places an unphysical upper bound on $\mathcal{C} \leq 0.5$, above which $\rho_c$ is no longer strictly positive over a complete cycle of density oscillation. Finally, in \MN, the resulting radial shifts, $a_\text{final}/a_0$, qualitatively diverge for other equivalently sensible choices of $\tau_\text{ave}$ (e.g. 3 Gyr). Similarly, varying the phase difference $\phi$ between the SC orbital motion and soliton density fluctuations, which was implicitly assumed to be zero in \MN, can also non-trivially alter the results for $e \neq 0$. In comparison, our approach does not suffer from these issues. \subsection{Subhalo Mass Function}\label{app:Comment0} In this subsection, we distinguish the unstripped halo mass at accretion $M_\text{halo}^{\leq r_\text{vir}}$ that obeys the core-halo mass scaling relation \eref{app:CoreHaloRelation} from the subhalo mass $M_\text{subhalo}^{\leq r_\text{vir}}$ after accounting for the effect of tidal stripping. Given the subhalo mass function $d n_\text{sub}(m_b)/d \ln M_\text{subhalo}$, the predicted number of Milky Way subhalos $n_\text{EII}(m_b)$ reads \begin{eqnarray}\label{eqn:SubhaloMassFunc} n_\text{EII}(m_b) = \int_{M_\text{subhalo}^{-2\sigma}}^{M_\text{subhalo}^{+2\sigma}} d \ln M_\text{subhalo} \bigg[\frac{dn_\text{sub}(m_b)}{d \ln M_\text{subhalo}}\bigg],\;\;\;\;\;\; \end{eqnarray} within 95\% confidence level (CL) of the Eri II halo mass $M_\text{subhalo}^{\leq r_\text{vir}}$, where $r_\text{vir}$ denotes the halo virial radius. The existence of Eri II asserts that $n_\text{EII}(m_b) \geq 1$; by directly equating $M_\text{subhalo}^{\leq r_\text{vir}}$ to $M_\text{EII}^{\leq r_\text{EII}} = 1.2^{+0.4}_{-0.3} \times10^7$ M$_\odot$ at 68\% CL, a lower bound $m_b \gtrsim 8\times 10^{-22}$ eV was derived by \MN. \begin{figure} \includegraphics[width=\linewidth]{./FIG15_MhaloFunction.pdf} \caption{Unstripped halo mass $M_\text{halo}^{\leq r_\text{vir}}$ (red) and soliton mass (orange) consistent with both observations (Fig. \ref{fig:SolitonMass}) and \eref{eqn:solitonhalomassratio}. The allowed values of subhalo mass $M_\text{subhalo}^{\leq r_\text{vir}}$ (pink shaded) are compared with that adopted by \MN\, \cite{Marsh:2018zyw}, which was set to be the enclosed mass of Eri II within its half-light radius $^\dagger M_{\rm subhalo}^{\le r_{\rm vir}} \equiv M_\text{EII}^{\leq r_\text{EII}}$ (light blue). Note that in general $M_{\rm subhalo}^{\le r_{\rm vir}} \gg M_\text{EII}^{\leq r_\text{EII}}$.} \label{fig:MhaloFunction} \end{figure} One major issue of the subhalo mass function calculation by \MN\, lies in the questionable assumption of $M_\text{subhalo}^{\leq r_\text{vir}} = M_\text{EII}^{\leq r_\text{EII}}$. From the core-halo relation and the observed mean density of Eri II within $r_\text{EII}$, we have $M_\text{halo}^{\leq r_\text{vir}} \gg M_\text{EII}^{\leq r_\text{EII}}$ as detailed below. First note the soliton mass contribution always dominates within $3.3 r_c$ compared with that of a background halo, as noted in Sec. \ref{sec:AstroConstraints}. The core-halo mass relation \cite{Schive:2014hza} and the total soliton mass (\eref{eqn:SolitonTotalMass}) together yield the soliton-halo mass ratio at redshift zero (see Appendix \ref{app:SolitonUnperturbed}) \begin{eqnarray} \frac{M_\text{halo}^{\leq r_\text{vir}}}{M_\text{soliton}} = 135\bigg(\frac{\rho_c}{\text{M}_\odot \text{pc}^{-3}}\bigg)^{1/2}, \label{eqn:solitonhalomassratio} \end{eqnarray} which carries no (explicit) dependence on $m_b$. $\rho_c \geq 0.135$ M$\odot$ pc$^{-3}$ for Eri II then implies $M_\text{halo}^{\leq r_\text{vir}} \geq 49.6M_\text{soliton}$. As demonstrated in Fig. \ref{fig:SolitonMass}, we have $3.3 r_c \geq r_{\rm EII}$ for $m_b \lesssim 10^{-21}$ eV, and thus $M_\text{halo}^{\leq r_\text{vir}} \gg M_\text{soliton}^{\leq 3.3 r_c} \geq M_\text{EII}^{\leq r_\text{EII}}$. Figure~\ref{fig:MhaloFunction} compares the values of $M_\text{halo}^{\leq r_\text{vir}}$ (red), $M_\text{soliton}$ (orange), and $M_\text{EII}^{\leq r_\text{EII}}$ that are consistent with both observations and \eref{eqn:solitonhalomassratio}. The maximum $m_b$ plotted, $\sim 10^{-21}\eV$, covers the lower bound $m_b \gtrsim 8\times10^{-22}\eV$ derived by \MN. Evidently, $M_\text{halo}^{\leq r_\text{vir}} \gg M_\text{soliton} \gtrsim M_\text{EII}^{\leq r_\text{EII}}$ for the entire range of $m_b$ of interest. Moreover, in our cosmological simulations, Eri-II-like halos (i.e. with $\rho_c \sim 0.135$ M$_\odot$pc$^{-3}$) do naturally form with $m_b \sim 10^{-22}$ eV \cite{Schive:2014dra, Schive:2014hza}, which further supports the consistency of the preceding analysis. An accurate estimate of subhalo mass $M_\text{subhalo}^{\leq r_\text{vir}}$ requires detailed information on the soliton-subhalo evolution history in response to the tidal field of the host galaxy. The pink shaded region in Fig. \ref{fig:MhaloFunction} indicates observationally compatible values of $M_\text{subhalo}^{\leq r_\text{vir}}$. In more extreme cases with substantial tidal stripping, the subhalo mass $M_\text{subhalo}^{\leq r_\text{vir}}$ can reduce to a mass scale comparable to $M_\text{soliton}$ \cite{Du:2018qor, Schive:2019rrw}. Here we have assumed that the soliton remains intact since a partially stripped soliton is unstable \cite{Du:2018qor}. Therefore, the inequality $M_\text{halo}^{\leq r_\text{vir}} \geq M_\text{subhalo}^{\leq r_\text{vir}} \geq M_\text{soliton} \gg M_\text{EII}^{\leq r_\text{EII}}$ always holds for $m_b < 8\times10^{-22}$ eV, even when assuming complete tidal disruption of the surrounding halo. The revised subhalo mass thus frees up the previously claimed lower bound of particle mass $m_b \gtrsim 8\times10^{-22}\eV$ by \MN. The other questionable aspect of \eref{eqn:SubhaloMassFunc} lies in the integration over the uncertainty range $\pm 2\sigma$, which gives $n_\text{EII} \rightarrow 0$ when $\sigma \rightarrow 0$ irrespective of the subhalo mass function. A more sensible approach is to either include more halo samples to estimate the halo number density around $M_\text{subhalo}^{\leq r_\text{vir}}$ or calculate the expectation value of $n_\text{EII}(<M_\text{subhalo}^{+2\sigma})$ from the cumulative subhalo mass function. The halo-to-halo variance should also be properly accounted for. \section{Summary and Conclusions}\label{sec:conclude} We have presented a detailed analytical and numerical study of soliton oscillations and identified two characteristic timescales associated, respectively, with the ground-state soliton wavefunction $\tau_{00}$ and the soliton density oscillations $\tau_\text{soliton}$; the astrophysical constraints from Eri II have also been reexamined. The main results are enumerated as follows. \begin{itemize} \item Both $\tau_{00}$ and $\tau_\text{soliton}$ depend only on the soliton peak density $\rho_c$, with a constant ratio $\tau_\text{soliton}/\tau_{00} \simeq 2.3$ irrespective of the boson mass $m_b$. \item The timescale of the central star cluster (SC) in Eri II, $\tau_\text{SC}$, has $\tau_\text{soliton} / \tau_\text{SC} \simeq 2\textendash3$ that deviates substantially from unity. \item Star cluster heating by an oscillating external gravitational field with a period $\tau_\text{osc}$ is noticeable only for $\tau_\text{osc}/\tau_\text{SC} \simeq 0.5$\textendash$1$. The heating by an oscillating soliton is hence ineffective, given $\tau_\text{soliton} / \tau_\text{SC} \simeq 2\textendash3$. \item Some of the FDM constraints derived by \MN\, appear invalid. The effect of orbital resonances is negligible over 10 Gyr as detailed above. The constraint from subhalo mass function cited therein is overly stringent, derived from a severely underestimated Eri II subhalo mass. \end{itemize} We have implicitly assumed in this work that the background halo of Eri II has been stripped away by the tidal field of the Milky Way. Otherwise, the presence of a background halo would lead to soliton random walk excursions, causing complete tidal disruption of the SC within $\sim$ 1 Gyr for $m_b \sim 10^{-22}\eV$ \cite{Schive:2019rrw}. Accordingly, the analysis here is mainly applicable for $m_b \lesssim 3.8\times10^{-21}\eV$, above which the halo contribution becomes non-negligible and requires further investigation. We have only used the average mass density within the half-light radius of Eri II ($\rho_\text{EII}$) to constrain the soliton profile, which however cannot break the degeneracy between $\rho_c$ and $m_b$. This limitation can be relaxed using the dynamics of Eri II member stars to determine the enclosed mass profile (e.g. see \cite{Zoutendijk:2021kee}). In addition, further observations capable of accurately determining the velocity dispersion of Eri II SC will conclude whether the SC is self-bound (or not), indicating $m_b \lesssim (\gtrsim) 4.6 \times 10^{-22}\eV$ (see Fig. \ref{fig:SolitonMass}). Lastly, we note that the tidal disruption of the SC due to soliton random walk excursions should be investigated more extensively with different $m_b$ and $M_{\rm halo}$. The particle mass constraint from the subhalo mass function discussed in \sref{app:Comment0} focuses on a specific halo mass range corresponding to Eri II. Nadler et al. \cite{Nadler:2020prv}, on the other hand, used the Milky-Way satellite galaxy luminosity functions, especially the faintest galaxies, to obtain a stringent constraint of $m_b \gtrsim 2.9 \times 10^{-21}\eV$. Cosmological FDM simulations capable of forming Milky Way-sized halos are indispensable for investigating this issue further, especially regarding whether the smaller Jeans wavelength and the extra small-scale powers developed at lower redshifts (e.g. see \cite{2021arXiv210101828M}) could increase the abundance of low-mass halos. \subsection*{Acknowledgments} We are happy to acknowledge useful conversations with Frank van den Bosch, Dhruba Dutta Chowdhury, Zhi Li, David J. E. Marsh, Jens C. Niemeyer, and Daniel W. Boutros. H. S. acknowledges funding support from the Jade Mountain Young Scholar Award No. NTU-109V0201, sponsored by the Ministry of Education, Taiwan. This work is partially supported by the Ministry of Science and Technology (MOST) of Taiwan under Grants No. MOST 107-2119-M-002-036-MY3 and No. MOST 108-2112-M-002-023-MY3, and the NTU Core Consortium project under Grants No. NTUCC-108L893401 and No. NTU-CC-108L893402.

2139935b72393285d7612dcdde1dcefaad5354b9

\section{Introduction} In the last decade deep and over-parametrized neural networks achieved breakthroughs in a variety of contexts~\cite{Alex12,He16,Vaswani17,GPT3,Silver17}. To keep pushing the boundaries of their capacity, these networks have grown in size to over billions of parameters~\cite{Tan19}. Network pruning approaches have recently received renewed attention as effective procedures to counterbalance this growth by reducing networks size without harming their performance. Several pruning methods have been introduced~\cite{Hoefler21}: they are all based on the idea of cutting weights from a trained, or partially trained, network in order to obtain a sparse sub-network that works at a comparable level of accuracy than the original one. For applications, pruning is very efficient in improving storage and computational cost. On the theoretical side, it led to the discovery that within a randomly-initialized network one can find a sub-network, {\it a winning lottery ticket}, that when trained in isolation can match the test accuracy of the original network after training for at most the same number of iterations~\cite{Frankle19}. The introduction of better network architectures, with geometries more suited to specific tasks, and which result in faster training and overall better generalizing properties can also be seen from a similar perspective. For instance, Convolutional Neural Networks~\cite{Alex12,He16,LeCun89,Simonyan15,Szegedy15} (CNN), which are very efficient for visual tasks, can be embedded in the Fully Connected Network (FCN) class. CNNs can therefore be interpreted as sparse and smaller FCNs trained from very specific initial conditions very well adapted to the task~\cite{Dascoli19}. Also in this case the size is drastically reduced (with respect to the FCN counterpart). The difference with pruning is that the main characteristics of CNNs---local connectivity and weight sharing---do not come from an automatic procedure but are ``hand-designed'' features introduced starting from analogies with the virtual cortex~\cite{Hubel62,Fukushima80}, studies of the invariant properties of natural images~\cite{Ruderman94}, and engineering improvements~\cite{LeCun89}. Pruning instead proceeds in an agnostic way in order to find smaller and efficient architectures; it takes advantage of properties which are imprinted in the network through training and which result from the interplay of data, task and optimization algorithm. Pruning induces an {\it inductive bias} customized to the learning task at hand, which leads to a network with better training and generalization properties compared to one of similar size trained from scratch. It has been actually shown that the same winning lottery ticket generalizes across training conditions and similar datasets~\cite{Morcos19}. This implies that the bias induced by pruning is sufficiently generic to transfer the pruned networks within the same data domain. Here, we characterize the nature of such inductive bias in the case of dense networks trained to classify natural images, and reveal that the winning lottery tickets of FCNs display the key features of CNNs. We apply \textit{iterative magnitude pruning} (IMP)~\cite{Frankle19} on FCNs trained on a low resolution version of ImageNet~\cite{ImageNet32}. We show that the sub-network obtained by IMP is characterized by {\it local connectivity}, especially in the first hidden layer, and masks leading to {\it local features} with patterns very reminiscent of the ones of trained CNNs~\cite{Zeiler14}. Deeper layers are made up of these local features with larger receptive fields hinting at the hierarchical structure found in CNNs. We study how the data and the task affect these properties. We show a crossover from a large dataset regime (large signal-to-noise ratio), where the inductive CNN-like bias is present, to a small dataset regime in which pruning loses performance and concomitantly the properties described above disappear. We show that a similar crossover takes place when going from a meaningful task to one with low semantic value. \section{Related works} Several recent works have studied pruning from the perspective of the lottery ticket hypothesis~\cite{Frankle19}: some investigated more general weight reinitializations~\cite{Zhou19}, whereas others actually questioned its validity on larger nets~\cite{Liu19,Gale19}. Later, Refs.~\cite{Frankle19-2,Frankle20} showed that in order to obtain good results for complex networks, datasets, and tasks it is important to modify the rewinding time compared to the initial formulation of the hypothesis. Other studies concerned hyperparameters modifications~\cite{Frankle20-2,Renda20}, and concentrated on the transferability~\cite{Morcos19,Sabatelli20} of the pruned networks. To the best of our knowledge, our work is the first to analyze the internal structure of winning tickets in order to highlight the inductive bias induced by pruning and relate it to architectural properties. Perhaps the most related work is Ref.~\cite{Neyshabur20}, which addresses the issue of learning CNN-like inductive bias from data and through training. It shows that training using a modified $\ell_1$ regularization is a way to induce local masks for visual tasks. This is similar to our results on pruning: enforcing sparsity during training allows to recover locality. Ref.~\cite{Dascoli19} studies the role of CNN-like inductive biases by embedding convolutional architectures within the general FCN class. It shows that enforcing CNN-like features in an FCN can improve performance even beyond that of its CNN counterpart. Finally, a manuscript~\cite{tolstikhin2021mlp} that appeared after the completion of our work shows that by considering a particular multilayer perceptron architecture, called MLP-mixer, some of the CNN features can be learned from scratch using a large training dataset. \section{Method, observables and notation}\label{ss:setup} \paragraph{Data, networks, and training procedure} Throughout this work our reference dataset, referred to as ImageNet32, consists of the almost 1.3M images classified in 1000 categories of the ILSVRC-12 image classification challenge~\cite{ILSVRC15}, cropped and scaled down to $32\times32$ resolution as detailed in~\cite{ImageNet32}. Considering the 3 color channels (RGB) the input size is $n_0=3072$ (we index the input as layer 0). The validation test contains 10k images. To keep the analysis simple, in the following we consider a FCN having 3 hidden layers of equal size $n_1=n_2=n_3=1024$. Generalizations to more general FCN architectures (more layers and different layer sizes) are presented in the SM \ref{SMss:arch}. Note that the aim of our work is to characterize the sub-networks induced by pruning in the simplest setup which allows a thorough analysis. For this reason, we use a small and fixed input size, which is useful to have a manageable size for the FCN, and we do not use any form of augmentation in order to avoid biases towards translational invariant features (the choice of a large dataset is important in this sense to obtain good statistics, as shown in Sec.~\ref{s:roledata}). We apply batch normalization~\cite{Ioffe15} and we use rectified linear unit (ReLU) nonlinearities for each layer. We denote $w^{l+1}_{ij}$ the weights between node $i$ of layer $l$ and node $j$ of layer $l+1$. Weights are initialized from a normal distribution $w^{l+1}\sim\mathcal{N}(0,2/(n_l+n_{l+1}))$~\cite{Glorot10}. Each node also has a bias $b^l_i$, initialized at 0. The size of the output layer equals the number of categories ($n_4=1000$ in this first experiment, 10 in Section~\ref{s:roletask}), which are converted to probabilities to compute the cross-entropy loss. The reported accuracy is the ratio of correct, most probable, labels. We train on mini-batches of 1000 images, and minimize through stochastic gradient descent with learning rate $\alpha=0.1$. Each training run corresponds to $10^5$ steps (around 78 epochs): while this is not sufficient to fit the training set, it is effective in identifying a winning lottery ticket~\cite{You20}. We checked that our results are robust while changing the optimization procedure, see the SM \ref{SMss:min}. \paragraph{Iterative Magnitude Pruning} We prune iteratively the 3 hidden layers. Each time, we remove $p=30\%$ of previously unpruned weights with the lowest absolute value at the end of training, per layer, as originally described in~\cite{Frankle19}. When analyzing the pruned networks, we denote $u$ the ratio of unpruned weights per layer: after $n$ iterations $u=(1-p)^n$. As suggested in~\cite{Frankle19-2}, we rewind the weights after each training not to the initial values, but to the ones obtained after 1000 steps of optimization in the first run. Instead of a winning lottery ticket the resulting network has been called a matching ticket~\cite{Frankle19-2} (we have verified that the main result holds when rewinding to 0, see SM \ref{SMss:IMPpar}). We iterate the training and pruning until less than 20\% of the nodes in any layer retain any connection to the previous layer, i.e.\ until more than 80\% of the nodes have been completely pruned away. More information on the pruning parameters, including total training time, is presented in SM \ref{SMss:IMPpar}. \paragraph{Observables} In order to analyze the structure of the subnetwork induced by pruning we focus on the following observables: \begin{itemize} \item \textit{Masks}: for each layer, we define masks by assigning 0 to pruned weights, and 1 to all the others. We indicate with $m^l_{ij}$ the mask formed by the unpruned weights $w^l_{ij}$. \item{\it Node connectivity}: for each node $j$ in layer $l$ we define the input connectivity $C^{{\rm in},l}_j=\sum_i m^l_{ij}$ as the number of input unpruned nodes, and $C^{{\rm out},l}_j=\sum_i m^{l+1}_{ji}$ the output connectivity. \item {\it Local distances}: in order to asses the locality of the masks, we count the number of pixels in a given relative position, ${\bf d}$, that are connected to the same hidden node, as shown schematically in Fig.~\ref{fig:2}\textbf{a}. More precisely: two inputs $i$ and $i'$ are connected through node $j$ if $m^1_{ij}m^1_{i'j}=1$. For such inputs we define the displacement ${\bf d}_{ii'}$: if input index $i$ identifies a pixel of spatial position $(x,y)$ and $i'$ corresponds to position $(x',y')$, we define their displacement ${\bf d}_{ii'}\equiv(x'-x,y'-y)$. The histogram of connected pixels at a given displacement is denoted $S({\bf d})=\sum_j\sum_{i,i'|{\bf d}_{ii'}={\bf d}}m^1_{ij}m^1_{i'j}$. We consider both the sum over pixels belonging to the same color channel ($S^{\rm sc}$) and over different color channels ($S^{\rm dc}$). \end{itemize} \section{Pruning sifts out local features and recovers CNN-like masks}\label{s:main} \begin{figure}[!ht] \centering \includegraphics[width=\columnwidth]{Fig1.pdf} \caption{ \textbf{a:} Highest validation accuracy at different percentage of unpruned weights. Averages and standard deviations over 4 independent IMP runs. \textbf{b:} Histograms of the number of $C^{{\rm in},1}$ for three configurations at different sparsities. The triangles represent the average connections $n_0u$ at each sparsity. \textbf{c:} Accuracy with respect to the full network for networks with a varying number of nodes removed, in increasing or decreasing order of $C^{{\rm in},1}_j$. }\label{fig:1} \end{figure} \begin{figure}[!ht] \centering \includegraphics[width=\columnwidth]{Fig2.pdf} \caption{ \textbf{a:} Sketch of the displacement and connectivity considered when computing $S({\bf d})$. \textbf{b:} $S^{\rm sc}$ (top) and $S^{\rm dc}$ (bottom) for each sparsity considered (columns). Relative positions are shown with respect to the center of each image (${\bf d}=0$) and the color scale of each plot goes from 0 to its maximum. For $S^{\rm sc}$, the single point at ${\bf d}=0$ is removed as it is trivially always connected. \textbf{c:} Same as \textbf{b}, but for the second hidden layer at the higher sparsity. }\label{fig:2} \end{figure} Fig.~\ref{fig:1}\textbf{a} shows the highest (early stop) validation accuracy as a function of $u$ throughout the IMP procedure. As expected~\cite{blalock2020state}, pruning allows to find sub-networks of considerably smaller sizes that work as well as the original ones. Actually, as shown in Fig.~\ref{fig:1}\textbf{a}, IMP leads to a strong improvement in the accuracy which goes from less than 9\% to more than 12\% when 99\% of the weights are pruned\footnote{The overall accuracy is low compared to the state of the art. This is due to the fact that we are focusing on a simple (hence fully analyzable) setup and by choice we avoid improvements that could lead to biases in the results. Yet, considering the low resolution, lack of augmentation, and simplicity of the network for a notoriously complex dataset, the improvement over the random $10^{-3}$ is a clear signal of learning. The top-5 accuracy is around $26\%$.}. \paragraph{The pruning-induced masks are local} To characterize these pruned network, we concentrate on 3 sparsity values close to the maximum accuracy reached by IMP: $u=4\%$, 1\%, and 0.3\%, as circled in Fig.~\ref{fig:1}\textbf{a}. We start by showing the histograms of $C^{{\rm in},1}$, i.e.\ the input connectivity per node of the first hidden layer. For a random pruning, this would be binomially distributed around the value $n_0u$, i.e.\ 123, 31, and 10 respectively for the sparsities considered. Fig.~\ref{fig:1}\textbf{b} shows that the distributions obtained after pruning instead have long positive tails, representing a population of nodes that preserve a large number of unpruned input weights. Since IMP should preserve ``important'' connections, it is interesting to check whether these nodes represent meaningful \textit{features} that have been sifted out by the procedure. To this end, we consider the relative accuracy of the final network when either the nodes with the largest or with the smallest $C^{{\rm in},1}$ have been completely removed (i.e.\ we set $m^1_{ik}=0$ for those nodes $k$, without any further training). As shown in Fig.~\ref{fig:1}\textbf{c}, the accuracy as a function of number of removed weights drops much more quickly when removing the highly connected nodes, thus confirming their relevance. In order to investigate the emerging local structure of the unpruned weights, we consider the number of connected pixels in a given relative position $S({\bf d})$. A 2D map for any possible displacement ${\bf d}$ is shown in Fig.~\ref{fig:2}\textbf{b}, for the 3 sparsities considered and for either same color or different color channels. Remarkably, we find that inputs connected to the same node are locally close, i.e. their distance is small, both within and between color channels. Note that the same 2D maps for the unpruned layer or pruned layers with random connections, are very different and non-local as shown in the SM \ref{SMs:masks}. We also find that straight horizontal and vertical relative displacements are slightly more likely. As the pruning progresses, the localization becomes stronger and the anisotropy more apparent. The same 2D maps are shown by only selecting inputs connected to nodes with a given connectivity in the SM \ref{SMss:locconn}. Our results show that the locality of the masks holds for all nodes except the ones with the smallest values of $C^{{\rm in},1}$. For the larger masks the effect is truly remarkable given their large number of unpruned weights. The set of results discussed above unveils one key feature of the matching (and winning) lottery ticket of FCNs: they display {\it local masks} that emerge by pruning. \begin{figure}[!hbt] \centering \includegraphics[width=\columnwidth]{Fig3.pdf} \caption{ \textbf{a:} Masks of the 30 nodes with largest $C^{{\rm in},1}_j$ for the best IMP iteration. Binary masks are directly converted to RGB intensities. \textbf{b:} Masks of 9 sample nodes around half in the $C^{{\rm in},1}_j$ ranking (having 20 connections each). Same iteration and representation as panel \textbf{a}. \textbf{c:} For a hand-picked selection of the most connected nodes (columns), the top row reports the mask and the second row represent the masked weights $w^1_{ij}m^1_{ij}$ normalized from minimum to maximum and converted to RGB. The last 3 rows represent the images with highest activation for that specific node. }\label{fig:3} \end{figure} \paragraph{The local features are CNN-like} We now concentrate on the nodes with the largest $C^{{\rm in},1}_j$, which have been found to be most relevant, and focus on the spatial structure of the associated masks. Fig.~\ref{fig:3}\textbf{a} shows the masks of the 30 most connected nodes for the best IMP iteration ($u\simeq 1\%$): they all present well localized distributions, either in a single location, symmetric with respect to the vertical axis, or sometimes around the edges of the images. Several masks also retain the same pixels through multiple color channels, often with a well defined pattern (e.g.\ 2 channels out of 3). Moreover, many of the masks show a specific preference for vertical and horizontal directions, in line with the previously observed anisotropy, with some consisting of parallel lines separated by one or two pixels (the effect is even more pronounced for stronger pruning, see SM \ref{SMss:bestm}, where we also show a larger sample of masks). A video of the evolution of first layer masks is linked in SM \ref{SMs:ext}. Masks with a lower number of connections, e.g.\ the ones positioned around the half of the $C^{{\rm in},1}_j$ ranking, do not show such well defined structures, as shown in Fig.~\ref{fig:3}\textbf{b}. In Fig.~\ref{fig:3}\textbf{c} we select a few nodes within the top 5\% most connected ones to highlight their specific structure: vertical lines, horizontal lines, oblique lines and complementary patches of color. To clarify the features selected by these nodes, we show not only the masks, but also the masks multiplied by the final weights and the 3 images in the validation with the highest activations over these nodes. The weights highlight further structure in the localized patches, with alternating positive and negative weights perpendicular to the preferential direction for long masks, or between different patches. These are very reminiscent of Gabor filters~\cite{Marcelja80}, and very similar to CNN masks, which respond preferentially to local high contrast details, e.g.\ edges in a specific direction or color patches~\cite{Zeiler14}. Moreover, very comparable filters can be seen at different locations, showing the emergence of partial translational invariance in the local features even without data augmentation\footnote{Note that images in the training set have a bias for structure in the center and at the boundary, a characteristic that emerges also from pruning: features tend to have more connections towards the center and edges of the image, as shown in SM \ref{SMs:masks}.}. Lastly, the example images unveil the response of these nodes also to local features not directly related to the final classification task. This directly relates to the good transfer properties of the subnetworks corresponding to the winning tickets \cite{Morcos19, Sabatelli20}. Further experiments highlighting the structure of the masks are presented in SM \ref{SMss:rot}, where the images are rotated, in SM \ref{SMss:IN64}, where higher resolution images are used as input, and in SM \ref{SMss:transl}, where we restore translational invariance by considering all possible image translations. To continue the parallel with CNNs, we now focus on the second hidden layer. In this case, for a given second layer node $j$, we define the associated effective mask as $\mu_{ij}=\theta(\sum_k m^1_{ik}m^2_{kj}-1/2)$ ($\theta$ being the step function), which equals 1 only if pixel $i$ is connected to \textit{any} first hidden layer node connected to $j$, and 0 otherwise. As done for the first hidden layer, we asses the locality of these masks via the local displacements, ${\bf d}_{i,i'}$, of the input nodes connected to the same mask. We show the 2D map of these displacements, $S({\bf d})$, in Fig.~\ref{fig:2}\textbf{c} for same and different channels for pruning $u=0.3\%$. Since the network is just 3 layers deep, and the receptive fields are already fairly large in the first layer, one does not expect the second layer nodes to be extremely localized. Yet, we find once again fairly local masks which favor horizontal and vertical directions, thus showing that pruning induces locality also beyond the first hidden layer. Masks from this layer, and some details on their composition, are further analyzed in SM \ref{SMss:bestm}. \section{The number of data controls pruning performance}\label{s:roledata} \begin{figure}[!htb] \centering \includegraphics[width=\columnwidth]{Fig4.pdf} \caption{ \textbf{a:} Gain in best validation accuracy (difference from the value corresponding to the unpruned network) as a function of $u$ for training on different percentages of the original dataset. The lines are averages of 4 independent runs, with the shaded area representing one standard deviation. The dashed line marks the iteration at $u\simeq0.5\%$ used in panels \textbf{c} and \textbf{d}. \textbf{b:} Largest gain in validation accuracy during the IMP procedure for different dataset sizes (percentage of the whole dataset). \textbf{c:} Histogram of $C^{{\rm in},1}$ for different dataset sizes (see label at the bottom) at the iteration $u\simeq0.5\%$. The dashed lines represent the theoretical binomial distribution for random pruning. \textbf{d:} $S^{\rm sc}({\bf d})$ for different dataset sizes at the iteration $u\simeq0.5\%$. Normalization and color map as show in Fig.~\ref{fig:2}. }\label{fig:4} \end{figure} We now study to what extent the effects found in the previous section depend on the properties of the dataset, in particular in the following we focus on the role of {\it the number of data} used to train the network. We repeat the training and the IMP with datasets of progressively smaller size, down to 2\% of the original. While this may seem like an extreme reduction of the dataset, even just 5\% of the data is actually larger than commonly used datasets such as CIFAR-10/100. What interests us here is the gain in accuracy due to the IMP, i.e. how much pruning leads to an increase of performance for a given number of data\footnote{Of course, the best validation accuracy decreases as we decrease the dataset size (to just around 2\% for 5\% of the data).}. In Fig.~\ref{fig:4}\textbf{a} we show the difference in accuracy with respect to the value obtained for the unpruned network during the IMP procedure. When decreasing the size of the dataset, the benefit of pruning diminishes, until no improvement at all is visible for the smallest datasets, see panel \textbf{b}, showing the maxima as a function of the dataset size. Decreasing the number of data used to train the network also drastically affects the connectivity and locality of the subnetworks found by pruning. In order to present this result we focus for each dataset size on an iteration close to the maximum of all curves of panel \textbf{a}, around $u=0.5\%$. Panels \textbf{c} show that decreasing the dataset size the connectivity distribution becomes essentially identical to the one corresponding to pruning weights at random independently of their magnitude (dashed line). Concomitantly, masks become more and more non-local, as shown in panels \textbf{d} (note that low connected nodes are affected earlier, see SM \ref{SMss:locconn}). Our results show that the dataset size clearly plays a major role for pruning. Our interpretation is that by decreasing the number of data one decreases the signal to noise ratio, hence effectively increasing random idiosyncratic fluctuations in the weights. When such fluctuations become of the same order of magnitude as the ones of the good subnetworks embedded in the original dense one, it becomes impossible to unveil the matching lottery ticket by pruning. In consequence, in this regime the subnetworks obtained by IMP are random, featureless and do not display any local feature. More results on the disruption of local features due to small dataset sizes are shown in SM \ref{SMss:bestm}. \section{The role of the task: the importance of being meaningful}\label{s:roletask} \begin{figure}[!htb] \centering \includegraphics[width=\columnwidth]{Fig5.pdf} \caption{ \textbf{a:} 10 sample images for each of the 10 macro classes. Each column shows a different \textit{semantic} cluster (rough label at the bottom) where images can be seen to have similar content, while each row is a different \textit{random} cluster, and no structure is visible. \textbf{b:} Gain in best validation accuracy (difference from the value corresponding to the unpruned network) for the random and semantic tasks. The lines are averages of 4 independent runs, with the shaded area representing one standard deviation. Each training was run for 500k steps for the semantic and 300k steps for the random clustering to ensure convergence. The dashed line marks the iteration at $u\simeq0.7\%$ used in panels \textbf{c} and \textbf{d}. \textbf{c:} Histogram of $C^{{\rm in},1}$ for the two cases at the iteration $u\simeq0.7\%$. The dashed line is the theoretical binomial distribution for random pruning. \textbf{d:} $S^{\rm sc}({\bf d})$ for the two cases at the iteration $u\simeq0.7\%$. Normalization and color map as shown in Fig.~\ref{fig:2}. }\label{fig:5} \end{figure} To further understand the role played by the properties of the dataset we now modify it in the following way. We cluster the original 1000 classes of the training set and test set in 10 balanced macro-classes in two different ways: the first one, called \textit{random}, groups the original categories in a non-meaningful way (on the basis of their numerical label in the dataset modulo 10). This produces 10 classes which are all very similar. The second one, called \textit{semantic}, clusters together similar original classes, into macro-categories such as ``dog'', ``device'', ``transport and furnishing'', etc. This produces 10 classes which characterize the data in a meaningful way. An example of the difference between these two ways of regrouping the 1000 classes is reported in Fig.~\ref{fig:5}\textbf{a}, where the rows represent examples from the ``random'' classes, and the columns from the ``semantic'' classes. For more details, see SM \ref{SMs:clust}. After training the accuracy obtained for the random task is at best 13\%, while it surpasses 36\% for the semantic one. This difference is expected since in the former case elements of different classes are very difficult to be distinguished and training mainly leads to memorization and not learning. We now compare the gain in accuracy due to pruning for the semantic and the random task. Fig.~\ref{fig:5}\textbf{b} shows that pruning is beneficial for the former but not for the latter. Moreover, the geometrical properties of the subnetwork induced by pruning are very different in the two cases. In panel \textbf{c} and panel \textbf{d} we show respectively the histogram of the connectivities $C^{{\rm in},1}$ and of the local displacements ${\bf d}_{ii'}$. These results reveal that the subnetwork obtained by pruning displays local features and non-trivial connectivity distribution for the semantic task only. In the other case, the subnetwork is essentially featureless, non-local and like one where weights have been pruned at random. Sample masks for the two tasks are shown in SM \ref{SMs:masks}. These results highlight the role of the task in shaping the properties of the network obtained by pruning: only for the task that the network can efficiently learn, and not just memorize, local features emerge. This result is a direct consequence of the \textit{qualitative} difference in the network solutions for the two tasks: features do not simply emerge from the data in an unsupervised fashion, but they represent structures useful for optimization with respect to a meaningful loss. If the task is solved in an unstructured way, e.g.\ through memorization, the features are simply not selected. In SM~\ref{SMss:10cl_more} we provide more evidence by focusing on other tasks in between the semantic and the random ones. \section{Discussion}\label{s:disc} CNNs are at the heart of modern machine learning for vision. They have been inspired by our visual system and they have proven their effectiveness in countless tasks. However, there is no \textit{a priori} justification for convolutional architectures, and no reason why other even more effective architectures should not exist. Indeed, several modification of the original convolutional idea have proven effective~\cite{He16,Szegedy16,Yu16}, and recently new architectures such as transformers have challenged the CNN supremacy in vision~\cite{Dosovitskiy20}. It is therefore interesting to see if the CNN architecture, or something with similar characteristics, could be naturally obtained from data. A recent work~\cite{Neyshabur20} takes a first important step in this direction, with a custom modification of the LASSO technique~\cite{Tibshirani96}, and working on a small but highly augmented dataset, it shows the emergence in training of local and sparse nonzero weights. In this work, we have followed a different perspective. With the aim of characterizing the inductive bias induced by pruning, we have limited ourselves to the original inputs, and we have applied iterative magnitude pruning. This approach has two benefits: it naturally leads to \textit{masks}, i.e.\ it irreversibly changes the architecture of the original network, and it leads to a final model that can be trained \textit{from scratch} to the final accuracy (i.e.\ from the original initialization values, although these are still correlated with the masks). In combination with the lack of augmentation, this makes features selected from this procedure the natural structure emerging from the dataset and task at hand. We have presented a thorough analysis of the architecture thus found: the pruning procedure does not affect all nodes equally, but a small number of them retains a disproportionate amount of inputs. These nodes are particularly important for the performance of the network and they correspond to input masks that are local in space and color channels, with a preference for specific patterns, found throughout the image (although for these images some features are clearly more likely to appear at some specific position than others). At visual inspection, these features bear a strong resemblance to those found in CNNs~\cite{Zeiler14}, the visual cortex~\cite{Marcelja80,Wolf05} and obtained with $\beta-$LASSO~\cite{Neyshabur20}. At high pruning, the features become more local and sparse, and they are combined in the next layer with a preference for similar patterns and close location. All these characteristics align well with the design choices of CNNs: selecting local features and combining them locally is the natural structure emerging from the image classification task---it represents the winning ticket that can be more easily trained. While other features like weight sharing simply cannot be reproduced in this model, the presence of similar masks at different locations is a sign of translational invariance (or the closest version of it available in these mostly centered images, an experiment with complete translational invariance is presented in SM \ref{SMss:transl}). In this sense, this procedure could suggest network structures even better aligned with the data from a real dataset, leading to more effective architectures. For instance the sparse sampling, even at this low image resolution, is reminiscent of dilated convolutions~\cite{Yu16}. As could be expected, only training on a large enough dataset leads to the emergence of these features. Interestingly, the success of the IMP itself seems to be related to their presence, hinting at the different structure of the solution, in line with recent findings about generalization and prunability~\cite{Kuhn21}. In the same direction, we have shown that locality and defined features disappear when we replace the task with a more vague one, where IMP is also ineffective. Features are thus not a simple consequence of the dataset, but depend on the structure imposed by the task. There are several directions which are worth exploring in future works. More complex starting architectures and larger or augmented training sets could be explored, to obtain even closer approximations to CNNs or other more effective architectures. The dependence of the architectural properties induced by pruning on the optimization protocol is another interesting aspect. For instance, we have found that just training until the beginning of overfitting is sufficient to obtain these features (see SM \ref{SMss:IMPpar}). We have also noticed that precursors are already visible at the early stages of pruning (see video link in SM \ref{SMs:ext}), suggesting the possibility of their early identification and more refined search strategies. In addition, it would be interesting to study more advanced pruning algorithms such as~\cite{Evci20,Lin20}, especially iterative or inherently sparse ones that could explore structures corresponding to prohibitively large networks. Finally, repeating our analysis for tasks different from visual ones, e.g. for audio~\cite{Oord16} and time series~\cite{Gamboa17}, and study the architectural bias induced by pruning is certainly an interesting direction for future research. This approach could also highlight useful architectures in fields that are still using more standard FCNs such as material modeling~\cite{Behler16}. In conclusion, our work underlines the effectiveness of pruning as a tool to uncover not only more efficient networks, but also specific architectural properties associated to the structure of the data and relevant for the training task. From this point of view, pruning schemes could offer a way to uncover effective new architectures for a wide category of problems. \begin{ack} We thank S. d'Ascoli, S. Goldt, M. Refinetti and L. Sagun for useful discussions. FP would especially like to thank E. K\"{u}\c{c}\"{u}kbenli for insightful ideas. We acknowledge funding from the French government under management of Agence Nationale de la Recherche as part of the ``Investissements d’avenir'' program, reference ANR-19-P3IA-0001 (PRAIRIE 3IA Institute) and from the Simons Foundation collaboration ``Cracking the Glass Problem'' (No. 454935 to G. Biroli). \end{ack} \bibliographystyle{unsrtnat}

8c4795516a90aa8eb49069676a09df9eb4d6aabd

\section{Introduction} In this article we study the ensemble Kalman approach for solving an inverse problem of the form \begin{align} y = A u^\dagger + \varepsilon ,\quad \varepsilon \sim \mathcal{N}(0,\Gamma), \end{align} where $u^\dagger \in \mathbb R^n$ is the unknown parameter of interest, $y\in \mathbb R^m$ are noisy measurements, $A\in \mathbb R^{m\times n}$ is a forward operator mapping the parameter space into the observation space, and $\Gamma \in \mathbb R^{m\times m}$ is a covariance matrix of the noise model in the measurement process which gives $y$. Following the Bayesian approach to inverse problems \cite{kaipio2006statistical}, we specify a prior measure $\mu_0$ on the set of feasible parameters $u$. Bayes' theorem then shows a way of incorporating the data $y$ into the prior, yielding a posterior measure $\mu$ such that \[\frac{\d\mu}{\d\mu_0}(u) \propto \exp\left(-\frac{1}{2}\|y-Au\|_\Gamma^2\right)\] Here (and in the following) we define the weighted inner product and norm as $\langle x,y\rangle_H:=\langle x,H^{-1}y\rangle$ and $\norm{x}_H^2 :=\langle x, H^{-1} x\rangle$ for a symmetric positive definite matrix~$H$. \paragraph{EKI} The ensemble Kalman methodology for inverse problems (EKI), see \cite{iglesias2013ensemble,schillings2017analysis,blomker2019well,herty2019kinetic}, generalized from a method for linear Gaussian state estimation \cite{kalman1960new}, was originally derived in the field of data assimilation \cite{evensen1994sequential}, see also \cite{evensen2009data,majda2012filtering,bergemann2012ensemble,law2015data,reich2015probabilistic,de2018long}, under the name of Ensemble Kalman-Bucy filter, was first applied in the context of Bayesian inverse problems in \cite{chen2012ensemble,emerick2013ensemble}, and was framed as a derivative-free optimization method with connections to sequential Monte Carlo in \cite{reich2011dynamical}. Its basic idea is to replace all measures involved with an empirical measure generated by an ensemble of particles: An initial ensemble of particles $\{u_0^j\}_{j=1}^J$ (usually sampled from the prior $\mu_0$) is considered a surrogate for the prior, and transitioning from prior to posterior amounts to moving the ensemble to new positions $u_1^j$, with the posterior ensemble $\{u_1^j\}_{j=1}^J$ standing in for the posterior. It can be shown that this particle update is given by the ensemble Kalman inversion (EKI) \begin{align}\label{eq:enk_it_1step} u^j_{1} = u^j_0 - C(u_0) A^T(A C(u_0) A^T + \Gamma)^{-1}(Au^j_0 - y), \end{align} where $C(u_0)$ is the sample covariance of the initial ensemble $\{u_0^j\}_{j=1}^J$. Under linear and Gaussian assumptions, and for an initial ensemble with empirical covariance matching exactly the prior covariance, it can be shown that the empirical measure given by the final particles $\{u_1^j\}_{j=1}^J$ is exactly the Gaussian measure identical to the posterior. Although only exact in this linear and Gaussian regime (see \cite{ernst2015analysis}), the computational benefit of replacing measures by empirical measures has prompted usage of this methodology in a more broader context, for example for nonlinear observation operators and non-Gaussian priors. If prior and posterior are very different from each other, then the particle updates \labelcref{eq:enk_it_1step} exhibit a large jump. Exchanging this one-step algorithm by a many-step iteration is supposed to yield a much smoother transition between prior and posterior, alongside a promise of computational improvements stemming from ``easing into'' the posterior. By introducing intermediate time steps (and thus intermediate measures $\mu_k$ ``interpolating'' between the prior $\mu_0$ and the posterior $\mu$), one obtains the iteration \begin{align}\label{eq:enk_it} u^j_{n+1} = u^j_n - C(u_n) A^T(A C(u_n) A^T + \tau^{-1}\Gamma)^{-1}(Au^j_n - \tilde y), \quad \tilde y \sim \mathcal N(y, \tau^{-1}\Sigma), \end{align} where $\tau=1/N$ with $N\in\mathbb N$ is a time step and $n=0,\ldots,N$. If $\Sigma = \Gamma$, this is the stochastic iterative EKI dynamics \cite{garbuno2020interacting}. If we set $\Sigma = 0$, we obtain the deterministic iterative EKI dynamics. The data $y$ is used in a perturbed form $\tilde y$ where the perturbation is done in such a way as to match the statistical properties of the observation noise. The idea is that perturbing the data additionally would help the dynamics explore the space state better. Also, we will see that without this perturbation, in contrast to~\labelcref{eq:enk_it_1step}, the continuous EKI dynamics does not recover the posterior. The case $\tau = 1$ corresponds to the ``vanilla'' ensemble Kalman inversion \labelcref{eq:enk_it_1step}, assimilating the data $y$ into the prior information given by the initial ensemble $\{u_0^j\}_{j=1}^J$, yielding the posterior given by the updated ensemble $\{u^j_1\}_{j=1}^J$. After some algebra it can be seen that \labelcref{eq:enk_it} can be equivalently reformulated as solution of a variational regularization problem: \begin{align}\label{eq:enk_it_optim} u^j_{n+1} = \argmin_{u} \frac{\tau}{2}\norm{Au - \tilde y}_\Gamma^2 + \frac{1}{2}\norm{u - u_n^j}_{ C(u_n)}^2. \end{align} Furthermore, by linearity of \labelcref{eq:enk_it} the sample means $m_n:=\frac{1}{J}\sum_{j=1}^Ju_n^j$ also satisfy \begin{align}\label{eq:enk_it_means_optim} m_{n+1} = \argmin_{u} \frac{\tau}{2}\norm{A u - \tilde y}_\Gamma^2 + \frac{1}{2}\norm{u - m_n}_{ C(u_n)}^2. \end{align} Hence the evolution of the whole ensemble and the sample means can be seen as minimizing movement discretization of the gradient flow of the functional $u\mapsto \frac12\norm{Au-\tilde y}^2_\Gamma$ with respect to the varying norm $\norm{\cdot}_{ C(u_n)}^2$, see for example the discussion in \cite{armbruster2020stabilization} and also \cite{weissmann2020particle}. \paragraph{Continuous EKI} Following \cite{iglesias2013ensemble,schillings2017analysis}, sending $\tau\to 0$ one arrives at the continuous ensemble Kalman inversion method \begin{equation} \label{eq:enkf_stoch} \dot{u}^j(t) = - C(t) A^T\Gamma^{-1}(Au^j(t)-y) + C(t)A^T\Gamma^{-1}\sqrt{\Sigma}\dot{\mathbf{W}}^j(t) \end{equation} for $j\in\{1,\ldots, J\}$, where $t\mapsto \mathbf{W}^j(t)$ are i.i.d. Wiener processes and $\Sigma$ is a symmetric positive definite matrix, with interesting special cases being $\Sigma=0$ (deterministic EKI) and $\Sigma=\Gamma$ (stochastic EKI). Note that by construction $t=0$ corresponds to the prior and $t=1$ is supposed to approximate the posterior (although in which sense has to be specified). For time $t > 1$, the influence of the prior decreases further. The quantities, \begin{align} \label{eq:sample_cov} C(t) &:=\frac{1}{J}\sum_{j=1}^J(u^j(t)-m(t)) \otimes (u^j(t)-m(t)),\\ \label{eq:sample_mean} m(t) &:= \frac{1}{J}\sum_{j=1}^J u^j(t), \end{align} denote the \emph{empirical covariance} and \emph{empirical mean} of the particles $u^j(t)$ at time $t>0$. Apart from time-continuous limits of the ensemble Kalman inversion \labelcref{eq:enk_it}, another interesting limit is the \emph{mean-field} limit as the number of particles $J\in\mathbb N$ in the ensemble goes to infinity. In this case, it is well known that the empirical measure $\rho_J(t) := \frac{1}{J}\sum_{j=1}^J \delta_{u^j(t)}$, where the particles $u^j(t)$ solve \labelcref{eq:enkf_stoch}, converge to a time-dependent measure $\rho(t,u)$ which solves the following Fokker-Planck equation, see, e.g., \cite{law2016deterministic,ding2019mean,herty2019kinetic}: \neww{% \begin{align}\label{eq:mean-field-limit} \partial_t \rho = \div\left(\rho~\mathfrak C(t)A^T\Gamma^{-1}(A u - y)\right) + \frac12 \operatorname{Tr}(D^2\rho~\mathfrak C(t)A^T\Gamma^{-1}A\mathfrak C(t)), \end{align} where $D^2\rho$ denotes the Hessian matrix of $\rho$ with respect to the variable $u$.} The mean and covariance matrix associated to such measures are defined as \begin{align} \mathfrak m(t) &= \int u \d\rho(t,u), \label{eq:meanfield_mean}\\ \mathfrak C(t) &= \int (u - \mathfrak m(t))\otimes(u - \mathfrak m(t)) \d\rho(t,u).\label{eq:meanfield_cov} \end{align} \paragraph{Various Notions of Mean and Covariance} In the context of different EKI methodologies it is very important to exactly specify what is meant by ``mean'' or ``covariance'': For given paths of Brownian motion, the ensemble given by \labelcref{eq:enkf_stoch} has an empirical (sample) covariance $ C$ \labelcref{eq:sample_cov} and an empirical (sample) mean $m$ \labelcref{eq:sample_mean}, i.e., the center of mass and a measure for the spread given by the particles. In addition, if we take the expectation $\mathbb E^\mathbf{W}$ with respect to the Wiener measure governing the Wiener noise terms $\mathbf{W}^j$, we get a different notion of mean and covariance, corresponding to the \emph{average empirical mean} $\mathbf{m}=\mathbb E^\mathbf{W} m$ and \emph{average empirical covariance} $\mathbf{C}=\mathbb E^\mathbf{W} C$. Finally, by considering the ``mean-field limit'' $J\to\infty$, one obtains a time-dependent measure $\rho_t$ with a mean and covariance (see \labelcref{eq:meanfield_mean,eq:meanfield_cov}). Depending on the type of mean and covariance one is referring two, the governing equations for these two quantities change. For instance, in the deterministic case, where $\Sigma=0$ in \labelcref{eq:enkf_stoch}, the empirical covariance and mean \labelcref{eq:sample_cov,eq:sample_mean} evolve according to (see~\cref{lem:derivation_ODEs}) \begin{align} \label{eq:ODE_cov_emp} \dot{ C}(t) &= -2 C(t) A^T \Gamma^{-1} A C(t),\\ \label{eq:ODE_mean_emp} \dot{m}(t) &= - C(t) A^T\Gamma^{-1}(A m(t)-y). \end{align} If $\Sigma=\Gamma$, these two empirical quantities evolve under a \emph{stochastic} differential equation, see \cref{lem:derivation_SDEs}. In contrast, as outlined in \cite{garbuno2020interacting} and shown in \cref{lem:derivation_meanfield} the mean-field covariance and mean \labelcref{eq:meanfield_cov,eq:meanfield_mean} evolve according to the differential equation \begin{align} \label{eq:ODE_cov_mean_field} \dot{\mathfrak C}(t) &= -\mathfrak C(t) A^T \Gamma^{-1} A \mathfrak C(t),\\ \label{eq:ODE_mean_mean_field} \dot{\mathfrak m}(t) &= -\mathfrak C(t) A^T\Gamma^{-1}(A\mathfrak m(t)-y). \end{align} Finally, the average empirical mean $\mathbf{m}(t):= \mathbb E^\mathbf{W} m(t)$ and average empirical covariance $\mathbf{C}(t) := \mathbb E^\mathbf{W} C(t)$ of the EKI for $\Sigma = \Gamma$ evolve as (see \cref{lem:derivation_SDEs}): \begin{align} \dot{\mathbf C}(t) &= -\frac{J+1}{J}\mathbf C(t) A^T \Gamma^{-1} A \mathbf C(t)\notag \\ \label{eq:ODE_cov_av_emp} &\qquad - \mathbb E^\mathbf{W}\left[(C(t)-\mathbf C(t))A^T\Gamma^{-1}A(C(t)-\mathbf C(t))\right],\\ \dot{\mathbf m}(t) &= -\mathbf C(t) A^T\Gamma^{-1}(A\mathbf m(t)-y) \notag \\ \label{eq:ODE_mean_av_emp} &\qquad -\mathbb E^\mathbf{W}\left[(C(t)-\mathbf C(t))A^T\Gamma^{-1}A(m(t)-\mathbf m(t))\right] \end{align} Note that the first and the second as well as the leading term of the third covariance equation only differ by constants. As shown in \cite{garbuno2020interacting}, the solution of the mean-field equations \labelcref{eq:ODE_cov_mean_field,eq:ODE_mean_mean_field} can be explicitly computed and is given by \begin{align} \label{eq:sol_mean_mean_field} \mathfrak m(t) &= \mathfrak C(t)\mathfrak C_0^{-1}\mathfrak m_0 + t \mathfrak C(t)A^T\Gamma^{-1}y,\\ \label{eq:sol_cov_mean_field} \mathfrak C(t) &= \mathfrak C_0(E+tA^T\Gamma^{-1}A\mathfrak C_0)^{-1}. \end{align} For $t=1$ these coincide with the posterior mean and covariance associated to \labelcref{eq:enk_it_1step}, which can be explicitly computed in this Gaussian and linear regime. In contrast, as we will show in this paper that the solutions to the other systems of equations \emph{do not} recover the posterior exactly. For instance, the solution to \labelcref{eq:ODE_cov_emp,eq:ODE_mean_emp} is given by \begin{align} \label{eq:sol_mean_emp} m(t) &= \sqrt{ C(t) C_0^{-1}} m_0 + \sqrt{ C(t) C_0^{-1}} \int_0^t \sqrt{ C(s) C_0^{-1}}\d s\, C_0A^T\Gamma^{-1}y,\\ C(t) &= C_0(E+2tA^T\Gamma^{-1}A C_0)^{-1}, \end{align} which does clearly not coincide with the posterior mean and covariance for any time $t\geq 0$. \paragraph{A Unified Framework for Ensemble Analysis} In this article we consider the following two differential equations \begin{alignat}{2} \label{eq:ODE_cov} \dot{\mathsf{C}}(t) &= -\alpha\mathsf{C}(t) A^T \Gamma^{-1} A \mathsf{C}(t), \quad &&\mathsf{C}(0) = \mathsf{C}_0,\\ \label{eq:ODE_mean} \dot{\mathsf{m}}(t) &= -\mathsf{C}(t) A^T\Gamma^{-1}(A\mathsf{m}(t)-y), \quad &&\mathsf{m}(0) = \mathsf{m}_0, \end{alignat} where $\alpha\geq 1$ is a free parameter. For $\alpha=2$ they reduce to the deterministic equations \labelcref{eq:ODE_cov_emp,eq:ODE_mean_emp} of the sample mean and covariance, for $\alpha=1$ to the mean-field equations \labelcref{eq:ODE_cov_mean_field,eq:ODE_mean_mean_field}, and for $\alpha=\frac{J+1}{J}$ to the leading term of the average sample equations \labelcref{eq:ODE_cov_av_emp,eq:ODE_mean_av_emp}. With this we continue a long line of articles which analyze fine deterministic properties of EKI. For instance, we refer to \cite{schillings2017analysis,schillings2018convergence,chada2020tikhonov,parzer2021convergence} for results, e.g., on existence, asymptotic behavior, and zero-noise consistency of EKI. In the whole article we assume that $\mathsf{C}_0$ is symmetric and positive definite. We record the most important bits of our notation in \cref{tab:notation}. As can be easily checked the ordinary differential equation \labelcref{eq:ODE_cov} for the covariance has the explicit solution \begin{align} \label{eq:sol_cov} \mathsf{C}(t) = \mathsf{C}_0(E+\alpha tA^T\Gamma^{-1}A\mathsf{C}_0)^{-1}, \end{align} which is a simple time-rescaling of the mean-field solution \labelcref{eq:sol_cov_mean_field}. However, as already seen above, the equation for the mean $\mathsf{m}(t)$ cannot be obtained as simple as that. In contrast, the formula for the sample mean which we obtain is given by the expression \begin{align}\label{eq:sol_mean} \mathsf{m}(t) = \left(\mathsf{C}(t)\mathsf{C}_0^{-1}\right)^\frac{1}{\alpha}\mathsf{m}_0 + \left(\mathsf{C}(t)\mathsf{C}_0^{-1}\right)^\frac{1}{\alpha}\int_0^t\left(\mathsf{C}(s)\mathsf{C}_0^{-1}\right)^{1-\frac{1}{\alpha}}\d s \,\mathsf{C}_0A^T\Gamma^{-1}y \end{align} and will be derived in \cref{sec:char_ode}. Obviously, for $\alpha=1$ this coincides with \labelcref{eq:sol_mean_mean_field} and for $\alpha=2$ with \labelcref{eq:sol_mean_emp}. We want to point out, though, that independent of $\alpha$ the asymptotic behaviour of particles and their covariance as $t\to\infty$ is the same since we can show that $\lim_{t\to\infty}\mathsf{C}(t)\mathsf{C}_0^{-1} =\lim_{t\to\infty}\left(\mathsf{C}(t)\mathsf{C}_0^{-1}\right)^p$ for every power $p>0$. \begin{table} \begin{tabular}[h]{l|l} Notation & Meaning \\ \hline \hline $C, m$ & \begin{minipage}{0.8\textwidth}Empirical covariance and mean of a given ensemble for EKI (deterministic or stochastic).\\[-0.8em]\end{minipage}\\ \hline $\mathbf C, \mathbf m$ & \begin{minipage}{0.8\textwidth}Average (w.r.t. the Wiener measure $\{\mathbf{W}^j\}_{j}$) empirical covariance and mean of a given ensemble for EKI.\\[-0.8em] \end{minipage}\\ \hline $\mathfrak C, \mathfrak m$ & Mean and covariance of mean-field limit of stochastic EKI.\\ \hline $\mathsf{C}, \mathsf{m}$ & \begin{minipage}{0.8\textwidth} Stand-in for $(C,m),(\mathfrak C, \mathfrak m)$ and $(\mathbf C, \mathbf m)$, respectively, for a unified calculation valid for all special cases.\\[-0.8em]\end{minipage} \\ \end{tabular} \caption{Notation in this manuscript} \label{tab:notation} \end{table} \paragraph{Time-invariant Diagonalization of $\mathsf{C}(t)\mathsf{C}_0^{-1}$} Although we will use the explicit expression \labelcref{eq:sol_cov} for $\mathsf{C}(t)$ in the following, we want to point out that many statements given can be rephrased in a form which only requires knowledge of the following fact: While $\mathsf{C}(t)$ changes its diagonalization over time (see \cref{sec:spectral}), the eigenvectors of $\mathsf{C}(t)\mathsf{C}_0^{-1}$ stay constant. This can be seen by defining $\mathsf{C}^c(t) = \mathsf{C}(t) \mathsf{C}_0^{-1}$, by which \labelcref{eq:ODE_cov} becomes $\dot\mathsf{C}^c(t) = -\alpha\mathsf{C}^c(t) \mathsf{C}_0A^T\Gamma^{-1}A \mathsf{C}^c(t)$. If we diagonalize $\mathsf{C}_0 A^T\Gamma^{-1}A = SDS^{-1}$ (which we justify in the following) and set $\mathsf{C}^d(t) := S^{-1}\mathsf{C}^c(t) S$, we obtain the matrix ordinary differential equation \[ \dot \mathsf{C}^d(t) = -\alpha \mathsf{C}^d(t) D \mathsf{C}^d(t).\] This decouples into a set of $n$ scalar ordinary differential equations since $\mathsf{C}^d(0) = E$ and $D$ is a diagonal matrix. \paragraph{Approximating the Posterior for $t=1$} Related to the ensemble Kalman inversion is the Bayesian approach of solving the inverse problem $Au=y$ by computing the Maximum A Posteriori (MAP) estimator \begin{align}\label{eq:MAP} u_\mathrm{MAP}(t) = \argmin \frac{t}{2} \norm{Au-y}_\Gamma^2 + \frac{1}{2} \norm{u-m_0}_{C_0}^2, \end{align} where the prior $u \sim \mathcal{N}(m_0,C_0)$ plays the role of a Tikhonov-type regularization. Using the optimality conditions of \labelcref{eq:MAP}, one can explicitly compute the MAP estimator as \begin{align}\label{eq:MAP_sol} u_\mathrm{MAP}(t) = m_0 + t C_0 (E +t A^T \Gamma^{-1} A C_0)^{-1}A^T \Gamma^{-1}(y-A m_0). \end{align} Here $t>0$ is a free parameter which is usually set to $1$. The MAP estimator has some heuristic relations to the discrete ensemble Kalman inversion \labelcref{eq:enk_it}. In the light of \labelcref{eq:enk_it_means_optim} it is obvious that the MAP estimator can be seen as one-step explicit Euler discretization of ordinary differential equation \labelcref{eq:ODE_mean} which drives the sample means $t\mapsto m(t)$ of the time-continuous ensemble Kalman inversion \labelcref{eq:enkf_stoch}. In particular, if one performs one iteration of \labelcref{eq:enk_it} with $\tau=1$ and $\Sigma=0$, then the sample mean coincides with $u_\mathrm{MAP}$ for $t=1$, \emph{even in the deterministic case}. One might hope that a similar property is true for the ensemble mean $\mathsf{m}(t)$, governed by \labelcref{eq:ODE_cov,eq:ODE_mean}. This is unfortunately not the case for general $\alpha>0$, as can be seen from the explicit solution \labelcref{eq:sol_cov,eq:sol_mean}, which we derive. It can be seen that, with the only exception of $\alpha=1$, there is no time $t\geq 0$ such that $\mathsf{C}(t)$ and $\mathsf{m}(t)$ constitute the mean and covariance of the posterior, not even in the linear Gaussian setting. The fundamental problem here is that one-step time discretizations are generally not exact. For readers interested in this issue we refer to \cite{burger2016spectral,bungert2019nonlinear}, where it was classified when one-step time discretizations and gradient flows of one-homogeneous functionals coincide. For illustration purposes, we compare an ensemble of particles which matches the true posterior with the result of the EKI for $t=1$ with and without noise perturbation in \cref{fig:covariances}. The set-up here is $A = \operatorname{diag}(4,1)$, the Gaussian prior has mean and covariance \begin{align*} \mathsf{m}_0 = \begin{pmatrix} 4\\4 \end{pmatrix},\quad \mathsf{C}_0 = \begin{pmatrix} 2 & -1 \\ -1 & 2 \end{pmatrix}, \end{align*} and data is given by $y=(0,0)^T$. \new{One can clearly see that the empirical mean and covariances of the unperturbed EKI are quite different from the posterior samples. The perturbed EKI approximates the posterior better but also not exactly. This is due to two reasons: First, there is perturbative noise in the stochastic EKI dynamics, so the ensemble cannot match the posterior exactly due to stochasticity (only the mean-field dynamics exactly recovers the posterior for $t=1$). Second, the factor of $\alpha = (J+1)/J$ is not equal to $1$ for finite ensemble sizes, hence even the average (w.r.t. Wiener measure) dynamics is not identical to the posterior-recovering dynamics \labelcref{eq:ODE_cov_mean_field,eq:ODE_mean_mean_field}. However, with increasing number of particles the approximation gets better since one then enters the mean-field regime, which recovers the posterior.} \begin{figure} \centering \includegraphics[width=0.49\textwidth,trim=1cm .5cm 1.5cm 1cm,clip]{figs/covariances_J45.pdf \hfill% \includegraphics[width=0.49\textwidth,trim=1cm .5cm 1.5cm 1cm,clip]{figs/covariances_J3.pdf}\\ \centering \hfill% \caption{Implementation of EKI dynamics for ensemble size 45 (left) and 3 (right) for time $t=1$. \textbf{Blue:} Ensemble matching the true posterior (triangles, dotted line). \textbf{Black:} Ensemble from deterministic EKI for $t=1$ (squares, solid line). \textbf{Green:} One realization of an ensemble from the stochastic EKI (circles, dash-dot line). This EKI approximately recovers the posterior measure, but only for large ensemble sizes (due to the noise and the fact that $J=3$ is not sufficiently close to the mean-field limit). Approximating the mean-field limit, i.e., taking a large ensemble $J=45$ improves the quality of approximation. Note that mean and covariance of deterministic EKI are identical between the two figures as they do not depend on $J$, unlike the stochastic EKI. Axes are scaled for visualization purposes. Ellipses visualize one standard deviation of the empirical covariances involved. } \label{fig:covariances} \end{figure} \paragraph{Ensemble Collapse and Non-monotonous Residual Convergence} Besides sample covariance and mean it is also interesting to consider the \emph{deviations} $e^j(t)$, the \emph{residuals} $r^j(t)$, and the \emph{residual mean} $r(t)$ of the deterministic EKI dynamics given by \labelcref{eq:enkf_stoch} for $\Sigma = 0$, i.e., \begin{equation*} \dot{u}^j(t) = - C(t) A^T\Gamma^{-1}(Au^j(t)-y). \end{equation*} Deviations, residuals, and residual mean are defined through \begin{align*} e^j(t) &:= u^j(t)-m(t),\\ r^j(t) &:=u^j(t) - u^\dagger,\\ r(t) &:= \frac{1}{J}\sum_{j=1}^J r^j(t) = m(t) - u^\dagger, \end{align*} respectively. Here, $u^\dagger$ is a suitable parameter value with respect to the data which will be specified in \cref{sec:particledynamics}. If $A$ has a nontrivial kernel, there is a whole subspace of possible candidates for $u^\dagger$. We will see that there is a canonical choice depending on the initial ensemble $\{u^j_0\}_{j=1}^J$. Naturally, one can also study similar quantities for stochastic or mean-field EKI (with very similar techniques) but this is beyond the scope of this paper. Following from \labelcref{eq:ODE_cov_emp,eq:ODE_mean_emp} deviations, residuals, and residual mean evolve according to the same differential equation \begin{align}\label{eq:ODE_residuals} \dot{x}(t) &= -C(t) A^T\Gamma^{-1}A x(t),\quad x\in\{e^j,r^j,r\},\\ \dot C(t) &= -2C(t)A^T\Gamma^{-1}AC(t), \end{align} and differ only in their initialization $x_0\in \{e^j_0, r^j_0,r_0\}$. As we will see, the explicit solution to \labelcref{eq:ODE_residuals} with initial condition $x(0)=x_0$ is given by \begin{align}\label{eq:sol_ODE_residuals} x(t) = \sqrt{C(t)C_0^{-1}} x_0. \end{align} Hence, for understanding the asymptotic behavior of the deviations, residuals, and residual mean, it suffices to characterize the asymptotic behavior of the empirical covariance matrix $C(t)$, or rather the preconditioned covariance matrix $C(t)C_0^{-1}$. Of particular interest in the context of ensemble Kalman inversion are the ensemble and residual spreads, defined as \begin{align} \label{eq:ensemble_spread} V_e(t) = \frac{1}{2J}\sum_{j=1}^J \norm{e^j(t)}^2,\\ \label{eq:residual_spread} V_r(t) = \frac{1}{2J}\sum_{j=1}^J \norm{r^j(t)}^2. \end{align} It is well-known that the ensemble spread $V_e$ is monotonically decreasing (this is the deterministic analogue of Theorem~3.4 in \cite{blomker2019well}), which we call \textit{ensemble collapse}. In contrast, the residual spread $V_r$ does not decrease monotonously in general (even in the zero-noise setting and for the optimal choice of $u^\dagger$ in the definition of $r^j$) and we devote \cref{sec:spreads} to a detailed study of this issue. Mapping the deviations and residuals into observation space, one can define the functions \begin{align} \label{eq:fwd_ensemble_spread} \mathfrak V_e(t) = \frac{1}{2J}\sum_{j=1}^J \norm{Ae^j(t)}_\Gamma^2,\\ \label{eq:fwd_residual_spread} \mathfrak V_r(t) = \frac{1}{2J}\sum_{j=1}^J \norm{Ar^j(t)}_\Gamma^2, \end{align} which indeed decrease monotonously, as shown in \cite{schillings2017analysis}. As far as the authors are aware, there has not been any exhaustive discussion of the convergence of deterministic particle dynamics and residuals of the ensemble Kalman inversion method for the infinite-time limit in parameter space: The introduction of the ensemble Kalman methodology was carried out in \cite{iglesias2013ensemble}, with no analysis of its dynamics for $t\to \infty$. In \cite{kelly2014well} it was proved (although in a more general data assimilation setting) that the dynamics do not blow up in finite time by bounding their growth exponentially, but numerical experiments soon suggested that a lot more would be feasible. The first long-time analysis of the EKI for Bayesian inverse problems (outside of the physically time-dependent data assimilation domain) was conducted in \cite{schillings2017analysis}, with a sequel in \cite{schillings2018convergence}, but is constrained to the observation space which -- for the case of lower-rank maps $A$ -- only allows control on a subspace of the parameter space (given by the orthogonal complement of $A$). Similarly, \cite{blomker2018strongly} and \cite{blomker2019well} proved convergence to zero of the ensemble spread but could prove convergence of the residuals only in observation space and under additional variance inflation. The missing piece, i.e., a full deterministic convergence analysis of the ensemble Kalman inversion methodology for linear Bayesian inverse problems, is provided with this manuscript. \paragraph{Contributions of this Article} \begin{itemize} \item \cref{thm:cov_dynamics,thm:asymptotic_profile,thm:deterministic_EKI,thm:mean_field}: An explicit characterization of the \textbf{dynamics of deterministic and mean-field EKI}, together with a detailed convergence analysis of these quantities for $t\to\infty$. The main idea here is that the diagonalization of $\mathsf{C}_0A^T\Gamma^{-1}A$ (where $\mathsf{C}_0$ is the initial ensemble's covariance, $A$ is the forward operator and $\Gamma$ is the observation noise covariance) also diagonalizes $\mathsf{C}(t)\mathsf{C}_0^{-1}$ and thus allows a complete analysis of the dynamics of the EKI. \item A negative answer to the question \textbf{whether the residuals decrease mono\-tonous\-ly}, even in the case of ``clean data'' (i.e., no observation noise) and for an optimal choice of the residual reference parameter $u^\dagger$ in $r^j(t) := u^j(t) - u^\dagger$. \item \cref{thm:eigenvectors}: A \textbf{spectral analysis of the covariance} $\mathsf{C}(t)$ itself, leading to a coupled differential algebraic system characterizing the evolution of eigenvalues and eigenvectors of the ensemble covariance. \item Our results in particular show that the deterministic EKI dynamics (and to a lesser extent the averaged stochastic EKI dynamics for \textit{small} finite ensemble size due to the fact that $(J+1)/J\neq 1$ and an additional stochastic correction term in \labelcref{eq:ODE_mean_av_emp,eq:ODE_cov_av_emp}) \textbf{do not recover the posterior measure for $t=1$} (or any other time). \end{itemize} \begin{remark}[Subspace Property]\label{rem:infdim} In \cite{iglesias2013ensemble}, it was shown that the EKI dynamics (both in deterministic and stochastic version) stays in the affine subspace spanned by the initial ensemble. This means that even for a infinite-dimensional inverse problem, the EKI method lives in finite-dimensional space which is why we assume this a-priori. The setting of ensemble Kalman inversion on an infinite-dimensional parameter space can be rephrased as a finite-dimensional problem by describing all quantities involved by the span of the initial ensemble. By the subspace property, we never leave this finite-dimensional span. Most realistic measurement processes generate finite-dimensional even for infinite-dimensional parameter models. This means that assuming both parameter and observation to be finite-dimensional quantities is not a strong restriction in most contexts. \neww{% Furthermore, the subspace property allows one to restrict all considerations to the affine subspace, spanned by the initial ensemble, where the covariance matrix $C_0$ is invertible. Hence, we assume the latter for the rest of the paper. Furthermore, all ground truth quantities must be interpreted as being projected onto this affine subspace of the parameter space.} \end{remark} We also remark that the ``time parameter'' $t$ rather corresponds to a regularization parameter and plays the role of supposedly (but in a strict sense only for the mean-field limit of the stochastic EKI) interpolating between prior ($t=0$) and posterior ($t=1$), with $t\to\infty$ corresponding to the limit of vanishing regularization. This means that $t$ is not physical time and does also not keep track of data accumulating over time, as it is in the setting of data assimilation. Relevant research on the ensemble Kalman-Bucy filter and the Ensemble Kalman inversion and their respective long-time behaviour include \cite{kelly2014well,amezcua2014ensemble,tong2016nonlinear,de2018long,ding2019mean,chada2020tikhonov}, with \cite{bergemann2009ensemble,bergemann2010localization,bergemann2010mollified,bergemann2012ensemble} being particularly relevant to our approach of analysing the ensemble's empirical mean and covariance. For a very well-written and extensive review of research on the stochastic ensemble Kalman methodology and its interpretation as metric gradient flow, we refer to \cite{garbuno2020interacting} and the references therein. \section{Spectral Decomposition of Preconditioned Covariance}\label{sec:spec_precond} We recall that we want to analyze \labelcref{eq:ODE_cov}, which is \begin{align*} \dot{\mathsf{C}}(t) = -\alpha\mathsf{C}(t) A^T \Gamma^{-1} A \mathsf{C}(t), \quad \mathsf{C}(0) = \mathsf{C}_0 \end{align*} with $\alpha > 0$ being a parameter unifying the treatment of deterministic, average stochastic, and mean-field EKI. We will use the closed-form solution~\labelcref{eq:sol_cov} \begin{align*} \mathsf{C}(t) = \mathsf{C}_0(E+\alpha tA^T\Gamma^{-1}A\mathsf{C}_0)^{-1} \end{align*} in order to derive a spectral decomposition of $\mathsf{C}(t)\mathsf{C}_0^{-1}$ and study asymptotic profiles. Since the explicit solution alone does not provide us with sufficient insight about the behavior of $\mathsf{C}(t)$, we pursue a different strategy for constructing a solution. Additionally, we hope that the methods used will extend to cases where explicit solutions can not be constructed (e.g., the nonlinear or stochastic case). The following two lemmas will be needed for the construction. We note that in this manuscript we will consider positive definite matrices which are not necessarily symmetric. \begin{lemma}\label{lem:product_diagonalizable} Given two symmetric matrices $V,W\in\mathbb R^{n\times n}$ with at least one of them being positive definite, the products $VW$ and $WV$ are diagonalizable. If the other matrix is also at least positive (semi)definite, then $VW$ and $WV$ are also positive (semi)definite. \end{lemma} \begin{proof} Without loss of generality assume that $W$ is positive definite. Thus there exists a non-singular square root matrix $W^{1/2}$. Then \[W^{-1/2}WVW^{1/2} = W^{1/2}VW^{1/2}.\] The matrix $W^{1/2}VW^{1/2}$ is symmetric and thus diagonalizable. The left hand side is a similarity transform of the product $WV$. This shows that $WV$ is diagonalizable. The statement is proven for $VW$ by arguing the same way for $W^{1/2}VWW^{-1/2}$. For the other statement note that if $V$ is positive (semi)definite in addition, then there also exists a square root $V^{1/2}$ and then $W^{1/2}VW^{1/2} = (W^{1/2}V^{1/2})(V^{1/2}W^{1/2})$ which is positive (semi)definite and hence this property holds for the matrices $VW$ and $WV$. \end{proof} \begin{lemma}\label{lem:algebraic_sol} Let $M$ be diagonalizable with non-negative eigenvalues such that $$M = S\operatorname{diag}(\mu_1,\ldots,\mu_k, 0, \ldots, 0)S^{-1}.$$ The columns of $S=(w_1,\ldots,w_n)$ are the eigenvectors such that $Mw_i = \mu_i w_i$. Then $R(t) = (E+tM)^{-1}$ exists for all $t\in \mathbb R\setminus\{-\mu_1^{-1},\dots,-\mu_n^{-1}\}$ and has the form \begin{align*} R(t) = S\operatorname{diag}\left(\frac{1}{1+t\mu_i}\right)_{i=1}^n S^{-1}. \end{align*} Also, $R_\infty := \lim_{t\to\infty}R(t)$ exists and has the same eigenvectors as $M$, but with \begin{align*} R_\infty w_i = \begin{cases} 0&\text{ for } i =1,\ldots, k,\\ w_i&\text{ for } i=k+1,\ldots,n. \end{cases} \end{align*} \end{lemma} \begin{proof} The invertibility of $(E+tM)$ for $t\neq -\mu_i^{-1}$, $i=1,\ldots,k$, follows directly from the spectral properties of $M$. Note that \[R(t) = (E+t S D S^{-1})^{-1} = S(E+tD)^{-1}S^{-1} = S\operatorname{diag}\left(\frac{1}{1+t\mu_i}\right)_{i=1}^n S^{-1}.\] This proves that indeed $R(t)$ has the same spectral decomposition as $M$ for all $t\in\mathbb R$ with \[R(t)w_i = \frac{1}{1+t\mu_i}w_i\] and the limit $t\to\infty$ as claimed. \end{proof} \subsection{Diagonalization of the Preconditioned Covariance}\label{sec:diagonalization} Now we are ready to formulate the first theorem of this section. The main idea here is that, while an eigenbasis of $\mathsf{C}(t)$ seems to behave erratically (stretching and rotating in a way to assimilate both the current covariance structure and the influence of the observation operator), the matrix $\mathsf{C}(t)\mathsf{C}_0^{-1}$ has fixed eigenvectors which equal those of $\mathsf{C}_0 A^T\Gamma^{-1}A$ (albeit with different eigenvalues). This already follows from \cite{garbuno2020interacting}, where it is shown that $\mathsf{C}(t)^{-1}$ is linear in time and that $\mathsf{C}(t)^{-1}\mathsf{C}_0$ has eigenvectors which do not depend on $t$ and allows to derive an explicit form of the solution of the EKI dynamics. \begin{theorem}[Covariance Dynamics]\label{thm:cov_dynamics} Let $\mathsf{C}(t) = (E+\alpha t\mathsf{C}_0A^T\Gamma^{-1}A)^{-1}\mathsf{C}_0$ denote the solution of \labelcref{eq:ODE_cov} with initial condition $\mathsf{C}_0$. Then it holds: \begin{itemize} \item The matrix $\mathsf{C}_0 A^T\Gamma^{-1}A$ is diagonalizable, meaning that $\mathsf{C}_0 A^T\Gamma^{-1}A = SDS^{-1}$ with $D = \operatorname{diag}(\mu_1,\dots,\mu_k,0,\dots,0)$ where $\mu_1\geq \cdots \geq\mu_k>0$ for some $k\leq n$. \item Letting $E(t):=\operatorname{diag}\left(\frac{1}{1+\alpha t\mu_i}\right)_{i=1}^n$, it holds that \begin{align*} \mathsf{C}(t) &= S E(t) S^{-1} \mathsf{C}_0,\\ \mathsf{C}(t)A^T\Gamma^{-1}A &= SD(t)S^{-1},\\ \mathsf{C}_0A^T\Gamma^{-1}A\mathsf{C}(t) &= SD(t)S^{-1}\mathsf{C}_0, \end{align*} where $D(t) := DE(t) = \operatorname{diag}\left( \frac{\mu_i}{1+\alpha t\mu_i}\right)_{i=1}^n$. Note that $D(0) = D$ and $E(0) = E$. \item $\mathsf{C}_\infty :=\lim_{t\to\infty}\mathsf{C}(t) = SE_\infty S^{-1} \mathsf{C}_0$, where $E_\infty = \operatorname{diag}(0,\ldots,0,1,\ldots, 1)$ has $k$ zero entries and $n-k$ entries of one, has the property that \[A\mathsf{C}_\infty=0.\] \end{itemize} \end{theorem} \begin{proof} Using \cref{lem:product_diagonalizable} the matrix $\mathsf{C}_0A^T\Gamma^{-1}A$ is positive semidefinite and diagonalizable as $\mathsf{C}_0A^T\Gamma^{-1}A = S \operatorname{diag}(\mu_1,\ldots,\mu_k, 0, \ldots, 0)S^{-1}$. From \labelcref{eq:sol_cov} we have an explicit solution of the covariance ordinary differential equation~\labelcref{eq:ODE_cov} given by $\mathsf{C}(t) = (E+\alpha t\mathsf{C}_0A^T\Gamma^{-1}A)^{-1}\mathsf{C}_0$. Hence, by \cref{lem:algebraic_sol} we can express this as $\mathsf{C}(t)= S E(t) S^{-1}\mathsf{C}_0$ for all $t\geq 0$, with \begin{align*} E(t)=\operatorname{diag}\left(\frac{1}{1+\alpha t\mu_i}\right)_{i=1}^n \end{align*} as claimed. The characterization of $\mathsf{C}(t)A^T\Gamma^{-1}A$ follows directly by seeing that \begin{align*} \mathsf{C}(t)A^T\Gamma^{-1}A &= SE(t)S^{-1}\mathsf{C}_0A^T\Gamma^{-1}A = SE(t)DS^{-1} = SD(t)S^{-1}. \end{align*} Using $D E_\infty = 0$, one calculates \begin{align*} \mathsf{C}_\infty A^T\Gamma^{-1}A &= SE_\infty S^{-1} \mathsf{C}_0 A^T\Gamma^{-1}A = SE_\infty S^{-1}SDS^{-1} = SE_\infty DS^{-1} = 0. \end{align*} This means that, by transposing this equality, we also have $A^T\Gamma^{-1}A\mathsf{C}_\infty = 0$, which is equivalent to $A \mathsf{C}_\infty=0$. \end{proof} \subsection{Asymptotic Profiles} In the previous section we have used a diagonalization to construct an explicit solution of the covariance matrix $\mathsf{C}(t)$ and understand its asymptotic behavior as $t\to\infty$. Now we study ``second-order'' asymptotics by investigating the asymptotic behavior of the time derivative $\dot{\mathsf{C}}(t)$ of the covariance. More precisely, we study asymptotic profiles of $\mathsf{C}(t)$ in the spirit of \cite{bungert2019asymptotic}, which are defined as limit of the approximate time derivative at infinity \begin{align*} \lim_{t\to\infty}t(\mathsf{C}(t) - \mathsf{C}_\infty), \end{align*} typically solve a nonlinear eigenvalue problem, and constitute self-similar solutions of the dynamics. To set the scene, we rewrite the covariance dynamics \labelcref{eq:ODE_cov} in the abstract form \begin{align*} \dot{\mathsf{C}}(t) = -\mathcal{A}(\mathsf{C}(t)), \qquad \mathsf{C}(0)=\mathsf{C}_0, \end{align*} where the nonlinear operator $\mathcal{A}$ is defined as \begin{align}\label{eq:operator_A} \mathcal{A}:\mathbb R^{n\times n}\to\mathbb R^{n\times n}, \quad \mathcal{A}(\mathsf{C}) =\alpha \mathsf{C} A^T \Gamma^{-1} A \mathsf{C}. \end{align} We will show that the rescaled solutions $t(\mathsf{C}(t)-\mathsf{C}_\infty)$ converge to an eigenvector $\hat{\mathsf{C}}$ of $\mathcal{A}$, meaning $\hat{\mathsf{C}}=\mathcal{A}(\hat{\mathsf{C}})$ Note that eigenvectors of $\mathcal{A}$, i.e., matrices with $\lambda\hat{\mathsf{C}}=\mathcal{A}(\hat{\mathsf{C}})$ give rise to self-similar solutions of \labelcref{eq:ODE_cov} (cf.~\cite{bungert2019asymptotic} for a general study). Indeed if $\mathsf{C}(0)=\hat{\mathsf{C}}$ then $\mathsf{C}(t)=a(t) \hat{\mathsf{C}}$, where $a:[0,\infty)\to\mathbb R$ solves the initial value problem $a'(t)=-\alpha\lambda a(t)^2, \;a(0)=1$. Hence, in this case \begin{align} \mathsf{C}(t) = \frac{1}{1+\alpha \lambda t} \hat{\mathsf{C}}. \end{align} In the context of the ensemble Kalman inversion this means that if the covariance matrix of the initial ensemble is an eigenvector of $\mathcal{A}$, then the shape of the ensemble remains unchanged during the flow. Hence, \cref{thm:asymptotic_profile} means that the rescaled covariance matrix of the ensemble Kalman inversion approaches a matrix which is a self-similar solution to~\labelcref{eq:ODE_cov}. \begin{theorem}[Asymptotic Profiles]\label{thm:asymptotic_profile} Let $\mathsf{C}(t)$ denote the solution of \labelcref{eq:ODE_cov} with initial condition $\mathsf{C}_0$. Then the limit \begin{align} \hat{\mathsf{C}} := \lim_{t\to\infty}t(\mathsf{C}(t)-\mathsf{C}_\infty) \end{align} exists, and satisfies $\mathcal{A}(\hat{\mathsf{C}})=\hat{\mathsf{C}}$. \end{theorem} \begin{proof} Plugging in the explicit representations $\mathsf{C}(t)= S E(t) S^{-1} \mathsf{C}_0$ and $\mathsf{C}_\infty=S E_\infty S^{-1} \mathsf{C}_0$, where $E(t)$ and $E_\infty$ are as in \cref{thm:cov_dynamics}, we obtain \begin{align*} t(\mathsf{C}(t)-\mathsf{C}_\infty) &= t S \left( E(t)-E_\infty \right) S^{-1} \mathsf{C}_0 \\ &= S \operatorname{diag}\left(\frac{t}{1+\alpha t\mu_1},\dots,\frac{t}{1+\alpha t\mu_k},0,\dots,0\right) S^{-1}\mathsf{C}_0 \\ &\longrightarrow S \underbrace{ \operatorname{diag}\left(\frac{1}{\alpha\mu_1},\dots,\frac{1}{\alpha\mu_k},0,\dots,0\right) }_{\hat{D}} S^{-1}\mathsf{C}_0, \quad t\to\infty. \end{align*} Hence, we can define the matrix $\hat{\mathsf{C}}:=S \hat{D}S^{-1}\mathsf{C}_0$ and observe that \begin{align*} \mathcal{A}(\hat{\mathsf{C}}) &= \alpha \hat{\mathsf{C}}A^T\Gamma^{-1}A\hat{\mathsf{C}} = \alpha S\hat{D}S^{-1}\mathsf{C}_0 A^T\Gamma^{-1} A S\hat{D}S^{-1} \mathsf{C}_0 \\ &= \alpha S\hat{D}S^{-1} S D S^{-1} S \hat{D}S^{-1}\mathsf{C}_0 = \alpha S\hat{D} D \hat{D}S^{-1}\mathsf{C}_0 = S\hat{D}S^{-1}\mathsf{C}_0 = \hat{\mathsf{C}}. \end{align*} Here we used that $\alpha \hat{D}$ is the pseudo-inverse matrix of $D$. \end{proof} \begin{example} Let $A=(0,1)$, $\Gamma=E$, and $\mathsf{C}_0=\begin{pmatrix} a&b \\ b&d\end{pmatrix}$. Then if $a\neq 0$ one can compute that \begin{align*} \mathsf{C}(t) &= \frac{1}{1+\alpha t d} \begin{pmatrix} a+\alpha t\det\mathsf{C}_0 & b \\ b & d \end{pmatrix}, \\ \mathsf{C}_\infty &= \begin{pmatrix} \tfrac{\det\mathsf{C}_0}{d} & 0 \\ 0 & 0 \end{pmatrix}, \\ \hat{\mathsf{C}} &= \frac1\alpha \begin{pmatrix} \tfrac{b^2}{d^2} & \tfrac{b}{d} \\ \tfrac{b}{d} & 1 \end{pmatrix}. \end{align*} Without loss of generality we can assume $d=1$ which yields $\hat{\mathsf{C}}=\frac1\alpha\begin{pmatrix}b^2&b\\b& 1\end{pmatrix}.$ This matrix has the eigenvectors $(1,-b)^T$ and $(b,1)^T$ with eigenvalues $0$ and $(b^2+1)/\alpha$, respectively. This means that the ensemble approaches its limit in a tilted way if $b\neq 0$ and parallely if $b=0$. We illustrate this in \cref{fig:profile}, where the first row shows the evolution of EKI with an initial ensemble aligned to the subspace of solutions (depicted in red), and the second row shows the evolution of a rotated ensemble. The eigenvectors of the respective asymptotic profiles are depicted in the rightmost plots. Here the pink eigenvectors, corresponding to the non-zero eigenvalue of $\hat{\mathsf{C}}$, show the tilting with which the ensembles hits their limiting configurations. \end{example} \begin{figure}[tb] \def0.31\textwidth{0.31\textwidth} \centering \fbox{\includegraphics[width=0.31\textwidth,trim=8cm 6.5cm 8cm 4cm,clip]{figs/profile_straight_it_1}}\hfill% \fbox{\includegraphics[width=0.31\textwidth,trim=8cm 6.5cm 8cm 4cm,clip]{figs/profile_straight_it_5}}\hfill% \fbox{\includegraphics[width=0.31\textwidth,trim=8cm 6.5cm 8cm 4cm,clip]{figs/profile_straight_it_final}}\\% \vspace*{0.01\textwidth} \fbox{\includegraphics[width=0.31\textwidth,trim=8cm 6.5cm 8cm 4cm,clip]{figs/profile_skew_it_1}}\hfill% \fbox{\includegraphics[width=0.31\textwidth,trim=8cm 6.5cm 8cm 4cm,clip]{figs/profile_skew_it_5}}\hfill% \fbox{\includegraphics[width=0.31\textwidth,trim=8cm 6.5cm 8cm 4cm,clip]{figs/profile_skew_it_final}}% \caption{Asymptotic Profiles of ensemble Kalman inversion. From left to right: inital condition, intermediate time step, converged state. \textbf{Top row:} symmetric prior. \textbf{Bottom row:} asymmetric prior. Red and green arrows indicate eigenvectors of the empirical covariance matrix. Magenta and black arrows indicate eigenvectors of the asymptotic profile.} \label{fig:profile} \end{figure} \section{Particle Dynamics}\label{sec:particledynamics} We start again by recalling that the system we are analysing the particle dynamics of the EKI: \begin{align*} \dot{u}^j(t) = - C(t) A^T\Gamma^{-1}(Au^j(t)-y) + C(t)A^T\Gamma^{-1}\sqrt{\Sigma}\dot{\mathbf{W}}^j(t). \end{align*} The evolution of particle mean and covariance (in the various senses described above) can be uniformly described by \labelcref{eq:ODE_cov,eq:ODE_mean} which we repeat for convenience: \begin{alignat*}{2} \dot{\mathsf{C}}(t) &= -\alpha\mathsf{C}(t) A^T \Gamma^{-1} A \mathsf{C}(t), \quad &&\mathsf{C}(0) = \mathsf{C}_0,\\ \dot{\mathsf{m}}(t) &= -\mathsf{C}(t) A^T\Gamma^{-1}(A\mathsf{m}(t)-y), \quad &&\mathsf{m}(0) = \mathsf{m}_0. \end{alignat*} With the complete analysis of the covariance dynamics from \cref{sec:spec_precond} at hand we can now analyse, inter alia, the deterministic EKI dynamics of particles and their empirical mean $m(t)$, the averaged empirical means $\mathbf{m}(t)$ of stochastic EKI, and the mean-field means $\mathfrak m(t)$. Furthermore, we will study the long-time behaviour of these quantities and apply our findinings to ensemble and residual spreads. We structure the main results of this section, the proofs of which follow directly from the statements in \cref{sec:char_ode,sec:asymp_rates}, into three theorems. The first one deals with deterministic EKI, the second one with the averaged quantities of stochastic EKI, and the third one with mean-field EKI. Before stating the theorems, we recall the definition of the pseudo-inverse of a diagonalizable matrix $M = U\Lambda U^{-1}$ as $M^+ = U\Lambda^+U^{-1}$ where $\operatorname{diag}(a_1\ldots,a_k,0,\ldots,0)^+ = \operatorname{diag}(a_1^{-1},\ldots,a_k^{-1},0,\ldots,0)$. This allows us to define a pseudo-inverse of the diagonalizable matrix $\mathsf{C}_0A^T\Gamma^{-1}A=SDS^{-1}$ as $(\mathsf{C}_0A^T\Gamma^{-1}A)^+:=SD^+S^{-1}$. For the matrix $A^T\Gamma^{-1}A$ we will therefore define $(A^T\Gamma^{-1}A)^-:=(\mathsf{C}_0A^T\Gamma^{-1}A)^+\mathsf{C}_0$ and remark that this is \emph{not} the Moore-Penrose pseudo-inverse since it does not fulfill a hermiticity condition. The Moore-Penrose pseudo-inverse $M^+$ of a matrix $M$ has to fulfill the following: \begin{align*} M^+MM^+ = M^+,\quad MM^+M = M,\quad (MM^+)^T = MM^+,\quad (M^+M)^T = M^+M. \end{align*} The matrix $(A^T\Gamma^{-1}A)^-$ satisfies the first two, as can be checked by elementary computation, but not the latter two conditions, since for example \begin{align*} (A^T\Gamma^{-1}A)^-(A^T\Gamma^{-1}A) &= (SD^+S^{-1})\mathsf{C}_0(\mathsf{C}_0^{-1}SDS^{-1}) =SD^+DS^{-1} \end{align*} and this matrix is not symmetric in general (unless $S$ is orthogonal or $A$ is one-to-one). \begin{theorem}[Deterministic EKI]\label{thm:deterministic_EKI} Let $y = Au^\dagger + \varepsilon$ and $\{u^j\}_{j=1}^J$ solve the deterministic EKI dynamics \labelcref{eq:enkf_stoch} with $\Sigma=0$. Then the particles $u^j$ and their empirical mean $m(t)$ satisfy: \begin{align*} u^j(t) &= u^j_0 + \left(E-\sqrt{C(t)C_0^{-1}}\right)(u^\dagger - u^j_0) + \\ &\qquad\sqrt{C(t)C_0^{-1}} \int_0^t \sqrt{C(s)C_0^{-1}}\d s\,C_0A^T\Gamma^{-1}\varepsilon,\\ m(t) &= m_0 + \left(E-\sqrt{C(t)C_0^{-1}}\right)(u^\dagger - m_0) + \\ &\qquad\sqrt{C(t)C_0^{-1}} \int_0^t \sqrt{C(s)C_0^{-1}}\d s\,C_0A^T\Gamma^{-1}\varepsilon,\\ \lim_{t\to\infty} u^j(t) &= u^{j,\dagger} + (A^T\Gamma^{-1}A)^-(A^T\Gamma^{-1}\varepsilon),\\ \lim_{t\to\infty} m(t) &= m^{\dagger} + (A^T\Gamma^{-1}A)^-(A^T\Gamma^{-1}\varepsilon), \end{align*} where $u^{j,\dagger} = u^j_0 + \left(E-{C_\infty C_0^{-1}}\right)(u^\dagger - u^j_0)$ and $m^{\dagger} = m_0 + \left(E-{C_\infty C_0^{-1}}\right)(u^\dagger - m_0)$ are equivalently characterized by \begin{align*} u^{j,\dagger} &= \argmin\left\lbrace\|{u-u^j_0}\|_{C_0} \,:\, u\in \mathbb R^n,\,Au = Au^\dagger\right\rbrace, \\ m^{\dagger} &= \argmin\left\lbrace\norm{m-m_0}_{C_0}\,:\, m\in\mathbb R^n,\,Am = Au^\dagger\right\rbrace. \end{align*} Alternatively, we can write \begin{align*} \lim_{t\to\infty}u^j(t) &= \argmin\left\lbrace\|{u-u^j_0}\|_{C_0} \,:\, u\in \mathbb R^n,\,Au = \Pi_{\mathrm{ran}(A)}^\Gamma(y)\right\rbrace, \\ \lim_{t\to\infty}m(t) &= \argmin\left\lbrace\|{m-m_0}\|_{C_0} \,:\, u\in \mathbb R^n,\,Am = \Pi_{\mathrm{ran}(A)}^\Gamma(y)\right\rbrace, \end{align*} where $\Pi_V^\Gamma$ is the $\Gamma$-orthogonal projection operator onto a closed subspace $V$. The rate of convergence of all quantities involved is given by ${1}/{\sqrt{t}}$. \end{theorem} \new{ \begin{corollary}[Consistency for Vanishing Noise] \cref{thm:deterministic_EKI} in particular shows that as $\varepsilon\to 0$, i.e., the noise in the data vanishes, the particle positions and the sample mean converge to the minimal-prior solutions of the inverse problem. For convergence rates of deterministic EKI in the vanishing noise limit we refer to the recent paper \cite{parzer2021convergence}. There also a discrepancy principle for early-stopping the EKI is derived, see also \cite{iglesias2013ensemble}. \end{corollary}} \begin{theorem}[Mean-Field EKI]\label{thm:mean_field} Let $y = Au^\dagger + \varepsilon$ and $\rho(t,u)$ solve the mean-field dynamics of stochastic EKI. Then the mean-field mean $\mathfrak{m}(t)$ satisfies: \begin{align*} \mathfrak m(t) &= \mathfrak m_0 + \left(E-{\mathfrak C(t)\mathbf C_0^{-1}}\right)(u^\dagger - \mathfrak m_0) + t\mathfrak C(t)A^T\Gamma^{-1}\varepsilon,\\ \lim_{t\to\infty} \mathfrak m(t) &= \mathfrak m^{\dagger} + (A^T\Gamma^{-1}A)^-(A^T\Gamma^{-1}\varepsilon), \end{align*} where $\mathfrak m^{\dagger} = \mathfrak m_0 + \left(E-{\mathfrak C_\infty \mathfrak C_0^{-1}}\right)(u^\dagger - \mathfrak m_0)$ is equivalently characterized by \begin{align*} \mathfrak m^{\dagger} &= \argmin\left\lbrace\norm{\mathfrak m-\mathfrak m_0}_{\mathfrak C_0}\,:\, \mathfrak m\in\mathbb R^n,\,A\mathfrak m = Au^\dagger\right\rbrace. \end{align*} Alternatively, we can write \begin{align*} \lim_{t\to\infty}\mathfrak{m}(t) &= \argmin\left\lbrace\norm{\mathfrak{m}-\mathfrak{m}_0}_{\mathfrak C_0} \,:\, \mathfrak{m}\in \mathbb R^n,\,A\mathfrak{m} = \Pi_{\mathrm{ran}(A)}^\Gamma(y)\right\rbrace, \end{align*} where $\Pi_V^\Gamma$ is the $\Gamma$-orthogonal projection operator onto a closed subspace $V$. The rate of convergence of $\mathfrak m(t)$ to $\mathfrak m_\infty$ is given by ${1}/{t}$. \end{theorem} \begin{remark} It is not obvious but can be verified with a short calculation that formula for $\mathfrak m(t)$ constructed in this theorem indeed coincides with \labelcref{eq:sol_mean_mean_field} from \cite{garbuno2020interacting} for $\varepsilon=0$. In particular, it coincides with the posterior for $t=1$. \end{remark} \begin{remark}[Asymptotic Behavior] \cref{thm:deterministic_EKI,thm:mean_field} show that the asymptotic behavior of all the different notions of mean as $t\to\infty$ coincide. Hence, in the case of vanishing noise ($\varepsilon\approx 0$), where on would like to let the evolution proceed to large times $t$, both deterministic, average stochastic, and mean-field EKI can be used comparably. \end{remark} \subsection{Fundamental Dynamical Properties}\label{sec:char_ode} \new{We can treat all equations arising in \cref{thm:deterministic_EKI,thm:mean_field}} in a uniform way by considering the ordinary differential equations \begin{alignat}{2} \dot{\mathsf{C}}(t) &= -\alpha\mathsf{C}(t) A^T \Gamma^{-1} A \mathsf{C}(t), \quad &&\mathsf{C}(0) = \mathsf{C}_0,\\ \label{eq:ODE_general} \dot{x}(t) &= -\mathsf{C}(t)A^T\Gamma^{-1}(Ax(t)-y),\quad &&x(0)=x_0. \end{alignat} In the following proposition we derive the solution of this ordinary differential equation and its asymptotic behavior. \new{For this we utilize the diagonalization of $\mathsf{C}(t)$, which is provided by \cref{thm:cov_dynamics}.} \begin{proposition}\label{prop:char_ODE} The solution of \labelcref{eq:ODE_general} is given by \begin{align}\label{eq:ODE_general_sol_no_source} x(t) = \left(\mathsf{C}(t)\mathsf{C}_0^{-1}\right)^\frac{1}{\alpha} x_0 + \left(\mathsf{C}(t)\mathsf{C}_0^{-1}\right)^\frac{1}{\alpha} \int_0^t \left(\mathsf{C}(s)\mathsf{C}_0^{-1}\right)^{1-\frac{1}{\alpha}}\d s\, \mathsf{C}_0 A^T\Gamma^{-1} y. \end{align} \end{proposition} \begin{proof} We define $L(t) := \left(\mathsf{C}(t)\mathsf{C}_0^{-1}\right)^{-\frac{1}{\alpha}}=S E(t)^{-\frac{1}{\alpha}}S^{-1}$. Using the definition of the matrices $E(t)$, $D$, and $D(t)$ as in \cref{thm:cov_dynamics}, we can compute \begin{align*} \dot{L}(t) &= S\operatorname{diag}\left((1+\alpha t\mu_i)^{\frac{1}{\alpha}-1}\mu_i\right)_{i=1}^nS^{-1}=SE(t)^{-\frac{1}{\alpha}}E(t)DS^{-1}\\ &=S E(t)^{-\frac{1}{\alpha}}D(t)S^{-1}\\ &=: S M_\alpha(t)S^{-1}. \end{align*} The product $L(t)x(t)$ then satisfies \begin{align*} &\phantom{=}\frac{\mathrm d}{\mathrm d t}\left[L(t)x(t)\right]\\ &=\dot{L}(t)x(t)+L(t)\dot{x}(t) \\ &= S M_\alpha(t) S^{-1} x(t) - S E(t)^{-\frac{1}{\alpha}} S^{-1} \mathsf{C}(t) A^T\Gamma^{-1}(Ax(t)-y)\\ &= S M_\alpha(t) S^{-1} x(t) - S E(t)^{-\frac{1}{\alpha}} S^{-1} S D(t) S^{-1}x(t) + SE(t)^{-\frac{1}{\alpha}}S^{-1}\mathsf{C}(t)A^T\Gamma^{-1}y\\ &= S M_\alpha(t) S^{-1} x(t) - S \underbrace{E(t)^{-\frac{1}{\alpha}}D(t)}_{=M_\alpha(t)} S^{-1}x(t) + SE(t)^{-\frac{1}{\alpha}}S^{-1}\mathsf{C}(t)A^T\Gamma^{-1}y \\ &= SE(t)^{-\frac{1}{\alpha}}S^{-1}\mathsf{C}(t) A^T\Gamma^{-1}y\\ &= SE(t)^{-\frac{1}{\alpha}}S^{-1}S E(t) S^{-1} \mathsf{C}_0 A^T\Gamma^{-1}y\\ &= SE(t)^{1-\frac{1}{\alpha}}S^{-1} \mathsf{C}_0 A^T\Gamma^{-1}y\\ &= \left(\mathsf{C}(t)\mathsf{C}_0^{-1}\right)^{1-\frac{1}{\alpha}} \mathsf{C}_0 A^T\Gamma^{-1}y. \end{align*} We can integrate this equation to the following one, which is equivalent to \labelcref{eq:ODE_general_sol_no_source}: \begin{align*} L(t)x(t) = x_0 + \int_0^t \left(\mathsf{C}(s)\mathsf{C}_0^{-1}\right)^{1-\frac{1}{\alpha}}\d s\, \mathsf{C}_0 A^T\Gamma^{-1} y. \end{align*} \end{proof} \new{ Let us now assume that the measured data is given by $y=Au^\dagger + \varepsilon$. Then we can split the integral in \labelcref{eq:ODE_general_sol_no_source} into two parts, where the first one can be evaluated using that $A u^\dagger$ is trivially in the range of the forward operator. \begin{proposition} Let $y=Au^\dagger + \varepsilon$. Then the solution of \labelcref{eq:ODE_general} is given by \begin{align}\label{eq:ODE_general_sol_source_and_noise} \begin{split} x(t) &= x_0 + \left(E-\left({\mathsf{C}(t)\mathsf{C}_0^{-1}}\right)^\frac{1}{\alpha}\right)(u^\dagger - x_0) \\ &\qquad + \left({\mathsf{C}(t)\mathsf{C}_0^{-1}}\right)^\frac{1}{\alpha} \int_0^t \left({\mathsf{C}(s)\mathsf{C}_0^{-1}}\right)^{1-\frac{1}{\alpha}}\d s\,\mathsf{C}_0A^T\Gamma^{-1}\varepsilon. \end{split} \end{align} \end{proposition} \begin{proof} From \cref{prop:char_ODE} we know that $x(t)$ has the expression \labelcref{eq:ODE_general_sol_no_source}. Plugging in $y=Au^\dagger+\varepsilon$ we can split the integral into two. The second one, featuring $\varepsilon$, just remains as it is. The first one can be computed as follows: \begin{align*} &\phantom{=}\left({\mathsf{C}(t)\mathsf{C}_0^{-1}}\right)^\frac{1}{\alpha}\int_0^t \left({\mathsf{C}(s)\mathsf{C}_0^{-1}}\right)^{1-\frac{1}{\alpha}} \d s\,\mathsf{C}_0A^T\Gamma^{-1}Au^\dagger\\ &= SE(t)^\frac{1}{\alpha}\int_0^t\operatorname{diag}\left({\mu_i}({1+\alpha s\mu_i})^{\frac{1}{\alpha}-1}\right)_{i=1}^n\d s\, S^{-1}u^\dagger\\ &= SE(t)^\frac{1}{\alpha}\operatorname{diag}\left(({1+\alpha t\mu_i})^\frac{1}{\alpha}-1\right)_{i=1}^n S^{-1}u^\dagger\\ &= S\left(E-E(t)^\frac{1}{\alpha}\right)S^{-1}u^\dagger\\ &= \left(E-\left({\mathsf{C}(t)\mathsf{C}_0^{-1}}\right)^\frac{1}{\alpha}\right)u^\dagger. \end{align*} Hence, using this together with \labelcref{eq:ODE_general_sol_no_source} we obtain the desired expression: \begin{align*} x(t) &= \left(\mathsf{C}(t)\mathsf{C}_0^{-1}\right)^\frac{1}{\alpha}x_0 + \left(E-\left({\mathsf{C}(t)\mathsf{C}_0^{-1}}\right)^\frac{1}{\alpha}\right)u^\dagger\\ &\qquad+ \left(\mathsf{C}(t)\mathsf{C}_0^{-1}\right)^\frac{1}{\alpha} \int_0^t \left(\mathsf{C}(s)\mathsf{C}_0^{-1}\right)^{1-\frac{1}{\alpha}}\d s\, \mathsf{C}_0 A^T\Gamma^{-1} \varepsilon\\ &=x_0 + \left(E-\left({\mathsf{C}(t)\mathsf{C}_0^{-1}}\right)^\frac{1}{\alpha}\right)(u^\dagger-x_0)\\ &\qquad+ \left(\mathsf{C}(t)\mathsf{C}_0^{-1}\right)^\frac{1}{\alpha} \int_0^t \left(\mathsf{C}(s)\mathsf{C}_0^{-1}\right)^{1-\frac{1}{\alpha}}\d s\, \mathsf{C}_0 A^T\Gamma^{-1} \varepsilon. \end{align*} \end{proof}} \subsection{Asymptotic Behavior and Convergence Rates}\label{sec:asymp_rates} If one tries to solve the integral in \labelcref{eq:ODE_general_sol_no_source} or \labelcref{eq:ODE_general_sol_source_and_noise} one obtains \begin{align*} &\int_0^t\left({\mathsf{C}(s)\mathsf{C}_0^{-1}}\right)^{1-\frac{1}{\alpha}}\d s \\ &= S\operatorname{diag}\left(\frac{({1+\alpha t\mu_1})^\frac{1}{\alpha}-1}{\mu_1},\dots,\frac{({1+\alpha t\mu_k})^\frac{1}{\alpha}-1}{\mu_k},t,\dots,t\right)S^{-1} \end{align*} and therefore \begin{align*} \left({\mathsf{C}(t)\mathsf{C}_0^{-1}}\right)^\frac{1}{\alpha}&\int_0^t\left({\mathsf{C}(s)\mathsf{C}_0^{-1}}\right)^{1-\frac{1}{\alpha}}\d s \\ &= S\operatorname{diag}\left(\frac{1-\frac{1}{(1+\alpha t\mu_1)^{\frac{1}{\alpha}}}}{\mu_1},\dots,\frac{1-\frac{1}{({{1+\alpha t\mu_k})^{\frac{1}{\alpha}}}}}{\mu_k},t,\dots,t\right)S^{-1}. \end{align*} The diagonal matrix in this expression blows up as $t\to\infty$ unless it is multiplied with $D=\operatorname{diag}(\mu_1,\dots,\mu_k,0,\dots,0)$. Remember that $D$ occured as diagonalization of $\mathsf{C}_0A^T\Gamma^{-1}A=SDS^{-1}$. This shows that the integral term has to be multiplied with $\mathsf{C}_0A^T\Gamma^{-1}A$ in order to exhibit a well-defined asymptotic behavior as $t\to\infty$. \new{ Since in \labelcref{eq:ODE_general_sol_no_source,eq:ODE_general_sol_source_and_noise} the integral is multiplied by $\mathsf{C}_0A^T\Gamma^{-1}y$ (or $\mathsf{C}_0A^T\Gamma^{-1}\varepsilon$ respectively), this shows that the data $y$ or the noise $\varepsilon$ has to lie in $\mathrm{ran}(A)$ for a well-defined asymptotic behavior. The previous observations can be collected in the following lemma. \begin{lemma}\label{lem:intergral} It holds that \begin{align*} \lim_{t\to\infty}\left({\mathsf{C}(t)\mathsf{C}_0^{-1}}\right)^\frac{1}{\alpha}&\int_0^t\left({\mathsf{C}(s)\mathsf{C}_0^{-1}}\right)^{1-\frac{1}{\alpha}}\d s\,\mathsf{C}_0A^T\Gamma^{-1}A = S(E-E_\infty)S^{-1}=E-\mathsf{C}_\infty\mathsf{C}_0^{-1}. \end{align*} \end{lemma} Using this lemma we can study the asymptotic behavior of \labelcref{eq:ODE_general_sol_source_and_noise} under the assumption that $\varepsilon\in\mathrm{ran}(A)$. This can be assumed without loss of generality thanks to the following lemma, which assumes that the data is split into a range component and a component in the orthogonal complement \begin{align} \mathrm{ran}(A)^{\bot,\Gamma}:=\left\lbrace y \in \mathbb R^m\,:\, \langle y, A u \rangle_\Gamma=0, \;\forall u\in\mathbb R^n \right\rbrace \new{=\ker(A^T\Gamma^{-1})}. \end{align} In our finite-dimensional observation setting this split is always possible since $\mathrm{ran}(A)$ is closed. \begin{lemma}\label{lem:orthogonal_range} Let $y= \bar y + y^\bot$ where $\bar y \in \mathrm{ran}(A)$ and $y^\bot \in \mathrm{ran}(A)^{\bot,\Gamma}$. Then the ensemble Kalman inversion \labelcref{eq:enkf_stoch} with datum $\bar y$ coincides with the one for $y$. \end{lemma} \begin{proof} The ensemble Kalman inversion \labelcref{eq:enkf_stoch} with datum $y$ reads \begin{align*} \dot{u}^j(t) &= -\mathsf{C}(t) A^T\Gamma^{-1}(Au^j(t)-\bar y - y^\bot) + C(t)A^T\Gamma^{-1}\sqrt{\Sigma}\dot{\mathbf{W}}^j(t) \\ &= -\mathsf{C}(t) A^T\Gamma^{-1}(Au^j(t)-\bar y) + \mathsf{C}(t) A^T\Gamma^{-1}y^\bot + C(t)A^T\Gamma^{-1}\sqrt{\Sigma}\dot{\mathbf{W}}^j(t). \end{align*} Now the claim follows from \begin{align*} \mathsf{C}(t)A^T\Gamma^{-1}y^\bot &= \frac{1}{J}\sum_{j=1}^J(u^j(t)-\mathsf{m}(t))\langle u^j(t) - \mathsf{m}(t), A^T\Gamma^{-1}y^\bot\rangle= 0. \end{align*} \end{proof} Hence we can formulate the following result on the asymptotic behavior. As it can be expected from a regularization method for inverse problems, up to a noise term the time asymptotic limit \labelcref{eq:ODE_general_sol_source_and_noise} can be interpreted as projection of the initial datum $x_0$ onto the solution set of $\{Au=y^\dagger\}$ where $y^\dagger=Au^\dagger$. In other words, $x(t)$ converges to a solution of $Au=y^\dagger$ with minimal value of the prior $x\mapsto\frac{1}{2}\norm{x-x_0}_{\mathsf{C}_0}^2$, which can be interpreted as the formal limit of $t\to\infty$ of the variational regularization method \labelcref{eq:MAP}, see \cite{bungert2019solution} for a rigorous study of this phenomenon. \begin{proposition}\label{prop:general_asymptotics} Let $y=Au^\dagger + \varepsilon$ where $\varepsilon=\bar \varepsilon + \varepsilon^\bot$ with $\bar \varepsilon\in\mathrm{ran}(A)$ and $\varepsilon^\bot\in\mathrm{ran}(A)^{\bot,\Gamma}$. Then the solution of \labelcref{eq:ODE_general} admits the asymptotic behavior \begin{align*} \lim_{t\to\infty}x(t) =x^\dagger + (A^T\Gamma^{-1}A)^-A^T\Gamma^{-1}\varepsilon, \end{align*} where $x^{\dagger} := x_0 + \left(E-{\mathsf{C}_\infty \mathsf{C}_0^{-1}}\right)(u^\dagger - x_0)$ is equivalently characterized by \begin{align*} x^{\dagger} = \argmin\left\lbrace\norm{x-x_0}_{\mathsf{C}_0}\,:\, x\in\mathbb R^n,\,Ax = Au^\dagger\right\rbrace. \end{align*} Finally, we can also write \begin{align*} \lim_{t\to\infty}x(t) = \argmin\left\lbrace\norm{x-x_0}_{\mathsf{C}_0}\,:\, x\in\mathbb R^n,\,Ax = \Pi_{\mathrm{ran}(A)}^\Gamma(y)\right\rbrace. \end{align*} where $\Pi_V^\Gamma$ is the $\Gamma$-orthogonal projection operator onto a closed subspace $V$. \end{proposition} \begin{proof} The proof is subdivided into four steps: First, we prove the asymptotic behavior, second, we show that $Ax^\dagger=Au^\dagger$, and third, we prove that $x^\dagger$ has minimal prior among all such parameters, after which we prove the alternative characterization of $\lim_{t\to\infty} x(t)$.\\ \textbf{Step 1:} Using $\mathsf{C}(t)\to\mathsf{C}_\infty$ in \labelcref{eq:ODE_general_sol_source_and_noise}, applying \cref{lem:intergral}, and utilizing the fact that \begin{align*} \left({\mathsf{C}_\infty\mathsf{C}_0^{-1}}\right)^\frac{1}{\alpha}=S {E_\infty}^\frac{1}{\alpha} S^{-1}=S E_\infty S^{-1} = \mathsf{C}_\infty\mathsf{C}_0^{-1} \end{align*} shows the identity \begin{align*} \lim_{t\to\infty}x(t)=x_0+\left(E-\mathsf{C}_\infty\mathsf{C}_0^{-1}\right)(u^\dagger-x_0)+\left(E-\mathsf{C}_\infty\mathsf{C}_0^{-1}\right)\xi, \end{align*} where $\xi\in\mathbb R^n$ is such that $A\xi=\bar\varepsilon$. Connecting this with the pseudo-inverse of $A^T\Gamma^{-1}A$ needs some thought: Using $(A^T\Gamma^{-1}A)^- = SD^+S^{-1}\mathsf{C}_0$ we obtain \begin{align*} (A^T\Gamma^{-1}A)^-A^T\Gamma^{-1}A &= S D^+ S^{-1}\mathsf{C}_0 \mathsf{C}_0^{-1} SDS^{-1} = SD^+DS^{-1} \\ &= S(E-E_\infty)S^{-1} = E - {\mathsf{C}_\infty\mathsf{C}_0^{-1}}. \end{align*} Finally, this implies \begin{align*} \left(E-\mathsf{C}_\infty\mathsf{C}_0^{-1}\right)\xi &= (A^T\Gamma^{-1}A)^-A^T\Gamma^{-1}A\xi\\ &=(A^T\Gamma^{-1}A)^-A^T\Gamma^{-1}\bar\varepsilon=(A^T\Gamma^{-1}A)^-A^T\Gamma^{-1}\varepsilon, \end{align*} since $\varepsilon=\bar\varepsilon+\varepsilon^\bot$ where $\varepsilon^\bot\in\mathrm{ran}(A)^{\bot,\Gamma}=\ker(A^T\Gamma^{-1})$.\\ \textbf{Step 2:} Now we show that $Ax^\dagger = Au^\dagger$. To this end we rewrite $x^\dagger$ as \begin{align*} x^\dagger = \mathsf{C}_\infty\mathsf{C}_0^{-1}(x_0-u^\dagger) + u^\dagger = SE_\infty S^{-1}(x_0-u^\dagger) + u^\dagger. \end{align*} Applying the matrix $A^T\Gamma^{-1}A$ to this equation yields \begin{align*} A^T\Gamma^{-1}A x^\dagger &= A^T\Gamma^{-1}ASE_\infty S^{-1}(x_0-u^\dagger) + A^T\Gamma^{-1}A u^\dagger\\ &= \mathsf{C}_0^{-1}SDE_\infty S^{-1}(x_0 - u^\dagger) + A^T\Gamma^{-1}Au^\dagger\\ &= A^T\Gamma^{-1}Au^\dagger. \end{align*} where we used that $D E_\infty=0$. Since $\Gamma^{-1/2}A$ is trivially surjective on its range, the transpose $A^T\Gamma^{-1/2}$ is injective there. This implies $\Gamma^{-1/2}A x^\dagger=\Gamma^{-1/2}Au^\dagger$ and multiplication with $\Gamma^{1/2}$ shows that $A x^\dagger=Au^\dagger$.\\ \textbf{Step 3:} We now show that $\norm{x^\dagger-x_0}_{\mathsf{C}_0}$ is minimal among all parameters with $Ax=Au^\dagger$. First, we note that $\mathsf{C}_\infty\mathsf{C}_0^{-1}=\mathsf{C}_\infty\mathsf{C}_0^{-1}\mathsf{C}_\infty\mathsf{C}_0^{-1}$ and hence $\mathsf{C}_\infty = \mathsf{C}_\infty\mathsf{C}_0^{-1}\mathsf{C}_\infty$. Second, it holds that $x^\dagger-x_0=(E-\mathsf{C}_\infty\mathsf{C}_0^{-1})(u^\dagger-x_0)$. We now claim that $u^\dagger$ can be replaced by any $x\in\mathbb R^n$ such that $Ax=Au^\dagger$. Indeed for $\xi=u^\dagger-x$ it holds $A\xi=0$ and hence $0=\mathsf{C}_0A^T\Gamma^{-1}A\xi=SDS^{-1}\xi$. Since $S$ is invertible and $D=\operatorname{diag}(\mu_1,\dots,\mu_k,0,\dots,0)$, we infer that $S^{-1}\xi=(0,\dots,0,\ast,\dots,\ast)$ and therefore \begin{align*} \left(E-{\mathsf{C}_\infty\mathsf{C}_0^{-1}}\right)\xi = S(E-E_\infty)S^{-1}\xi = 0. \end{align*} This allows us to compute for any $x\in\mathbb R^n$ with $Ax=Au^\dagger$: \begin{align*} \norm{x^\dagger-x_0}_{\mathsf{C}_0}^2 &= \norm{(E-{\mathsf{C}_\infty\mathsf{C}_0^{-1}})(x-x_0)}_{\mathsf{C}_0}^2 \\ &= \norm{x-x_0}_{\mathsf{C}_0}^2 - 2\left\langle x-x_0, {\mathsf{C}_\infty\mathsf{C}_0^{-1}}(x-x_0)\right\rangle_{\mathsf{C}_0} + \norm{{\mathsf{C}_\infty\mathsf{C}_0^{-1}}(x-x_0)}_{\mathsf{C}_0}^2\\ &=\norm{x-x_0}_{\mathsf{C}_0}^2 - 2\left\langle\mathsf{C}_0^{-1}(x-x_0), {\mathsf{C}_\infty\mathsf{C}_0^{-1}}(x-x_0)\right\rangle + \norm{{\mathsf{C}_\infty\mathsf{C}_0^{-1}}(x-x_0)}_{\mathsf{C}_0}^2. \end{align*} Using $\mathsf{C}_\infty = \mathsf{C}_\infty\mathsf{C}_0^{-1}\mathsf{C}_\infty$ and the symmetry of all covariance matrices, the inner product can be simplified as follows \begin{align*} \left\langle\mathsf{C}_0^{-1}(x-x_0), {\mathsf{C}_\infty\mathsf{C}_0^{-1}}(x-x_0)\right\rangle &= \left\langle\mathsf{C}_0^{-1}(x-x_0), \mathsf{C}_\infty\mathsf{C}_0^{-1}\mathsf{C}_\infty\mathsf{C}_0^{-1}(x-x_0)\right\rangle \\ &= \left\langle\mathsf{C}_0^{-1}\mathsf{C}_\infty\mathsf{C}_0^{-1}(x-x_0), \mathsf{C}_\infty\mathsf{C}_0^{-1}(x-x_0)\right\rangle \\ &= \norm{\mathsf{C}_\infty\mathsf{C}_0^{-1}(x-x_0)}_{\mathsf{C}_0}^2. \end{align*} Plugging this into the previous equation yields \begin{align*} \norm{x^\dagger-x_0}_{\mathsf{C}_0}^2 = \norm{x-x_0}_{\mathsf{C}_0}^2-\norm{{\mathsf{C}_\infty\mathsf{C}_0^{-1}}(x-x_0)}_{\mathsf{C}_0}^2\leq \norm{x-x_0}_{\mathsf{C}_0}^2, \end{align*} which proves the statement since $Ax^\dagger = y$ by Step 2 and $\norm{\cdot}_{\mathsf{C}_0}$ is strictly convex.\\ \textbf{Step 4:} The last statement follows from everything we have proven so far by writing $y = Au^\dagger + \varepsilon = Au^\dagger + \varepsilon^\dagger + \varepsilon^\bot = \Pi_{\mathrm{ran}(A)}^\Gamma(y) + \varepsilon^\bot$, hence \begin{align*} \lim_{t\to\infty} x(t) &= \argmin\{\|x-x_0\|_{\mathsf{C}_0}: x\in \mathbb R^n, Ax = \Pi_{\mathrm{ran}(A)}^\Gamma(y) \} + (A^T\Gamma^{-1}A)^-A^T\Gamma^{-1}\varepsilon^\bot. \end{align*} Since $A^T\Gamma^{-1}\varepsilon^\bot = 0$ by definition of the decomposition of $\varepsilon$, we can conclude. \end{proof} } \new{We close this section by deriving convergence rates of $x(t)$ to $x^\dagger$ in the case that there is no noise, i.e., $\varepsilon=0$. In the noisy case, \cref{prop:general_asymptotics} tells us that $x(t)$ does only converge to $x^\dagger$ up to a noise level. Hence, in this case one can only expect a \emph{semi-convergence} behavior which we will not study in this article. Instead, we report the following result for the noise-free case.} \begin{proposition}\label{prop:x_cvgc_rates} Assume that $y=Au^\dagger\in\mathrm{ran}(A)$ and let $x(t)$ and $x^\dagger$ be as in \cref{prop:general_asymptotics}. Then there exists a constant $C>0$ such that it holds \begin{align} \label{eq:rate_x} \norm{x(t)-x^\dagger} &\leq C\left(\frac{1}{\mathrm{gap}\cdot t}\right)^\frac{1}{\alpha},\quad\forall t\geq 0,\\ \label{eq:rate_fwd_x} \norm{Ax(t)-y}_\Gamma &\leq C\cdot \left({\frac{\mu_{\max}}{t}}\right)^\frac{1}{\alpha},\quad \forall t \geq 0. \end{align} Here $\mathrm{gap}:=\mu_k$ denotes the spectral gap and $\mu_{\max}:=\mu_1$ the largest eigenvalue of the matrix $\mathsf{C}_0A^T\Gamma^{-1}A$ . \end{proposition} \begin{proof} Diagonalizing and subtracting $x(t)$ (given by \labelcref{eq:ODE_general_sol_source_and_noise} for $\varepsilon=0$) and $x^\dagger$ (defined as in \cref{prop:general_asymptotics}) yields \begin{align*} x(t) - x^\dagger &= S({E(t)}^\frac{1}{\alpha}-E_\infty)S^{-1}(x_0-\xi)\\ &= S \operatorname{diag}\left(\frac{1}{({1+\alpha t\mu_1})^\frac{1}{\alpha}},\dots,\frac{1}{({1+2t\mu_k})^\frac{1}{\alpha}},0,\dots,0\right) S^{-1}(x_0-\xi), \end{align*} where $\mathrm{gap}:=\mu_k$ denotes the smallest non-zero eigenvalue of $A^T\Gamma^{-1}A$. Taking norms yields the convergence rate \labelcref{eq:rate_x}. Using this expression for $x(t)-x^\dagger$ and the diagonalization $A^T\Gamma^{-1}A = \mathsf{C}_0^{-1} SDS^{-1}$ we also find \begin{align*} &\phantom{=}A^T\Gamma^{-1}A(x(t)-x^\dagger)\\ &=\mathsf{C}_0^{-1}SD({E(t)}^\frac{1}{\alpha}-E_\infty)S^{-1}(x_0-\xi)\\ &=S \operatorname{diag}\left(\frac{\mu_1}{({1+\alpha t\mu_1})^\frac{1}{\alpha}},\dots,\frac{\mu_k}{({1+\alpha t\mu_k})^\frac{1}{\alpha}},0,\dots,0\right) S^{-1}(x_0-\xi). \end{align*} Multiplying this with $x(t)-x^\dagger$ and using $Ax^\dagger=y$ shows \labelcref{eq:rate_fwd_x}. \end{proof} \section{Ensemble and Residual Spreads of Deterministic EKI} \label{sec:spreads} In this section we study the ensemble and residual spreads (defined in \labelcref{eq:ensemble_spread,eq:residual_spread}) of the deterministic EKI \labelcref{eq:enkf_stoch} with $\Sigma=0$. \new{For this we utilize the system \labelcref{eq:ODE_cov_emp,eq:ODE_mean_emp} which describes the evolution of the empirical covariance $C(t)$ and mean $m(t)$, see \cref{thm:deterministic_EKI}.} \subsection{Convergence of the Spreads} \new{For convenience we repeat the definition of deviations $e^j$, residuals $r^j$, and residual mean $r$, given by \begin{align*} e^j(t) &:= u^j(t)-m(t),\\ r^j(t) &:=u^j(t) - u^\dagger,\\ r(t) &:= \frac{1}{J}\sum_{j=1}^J r^j(t) = m(t) - u^\dagger, \end{align*} and the time-dependent functions \labelcref{eq:ensemble_spread,eq:residual_spread,eq:fwd_ensemble_spread,eq:fwd_residual_spread}, which describe the ensemble collapse and residual convergence: \begin{alignat*}{2} &V_e(t) = \frac{1}{2J}\sum_{j=1}^J \norm{e^j(t)}^2, &&V_r(t) = \frac{1}{2J}\sum_{j=1}^J \norm{r^j(t)}^2,\\ &\mathfrak V_e(t) = \frac{1}{2J}\sum_{j=1}^J \norm{Ae^j(t)}_\Gamma^2,\quad &&\mathfrak V_r(t) = \frac{1}{2J}\sum_{j=1}^J \norm{Ar^j(t)}_\Gamma^2. \end{alignat*}} We recall from \cite{blomker2019well} that the ensemble spread $V_e(t)$ decreases monotonously with time. It does not necessarily converge to zero unless in the fully observed case of an invertible forward operator $A$, as we can only expect ensemble collapse along the components orthogonal to the kernel of $A$. To see the decrease of $V_e$ one computes \begin{align*} \dot{V}_e(t) &= \frac{1}{J}\sum_{j=1}^J \langle \dot e^j(t), e^j(t)\rangle = -\frac{1}{J}\sum_{j=1}^J \langle C(t)A^T\Gamma^{-1}A e^j(t), e^j(t)\rangle\\ &= -\frac{1}{J^2}\sum_{i,j=1}^J \langle e^i(t), e^j(t)\rangle \langle e^i(t), A^T\Gamma^{-1}A e^j(t)\rangle\leq 0, \end{align*} where Lemma~A.3. in~\cite{blomker2019well} ensures that the last term is non-negative. For the ensemble spread in observation space, given by \labelcref{eq:fwd_ensemble_spread}, one can even prove convergence to zero. It holds \begin{align*} \dot{\mathfrak V}_e(t) &= \frac{1}{J}\sum_{j=1}^J\langle A^T\Gamma^{-1}Ae^j(t),\dot e^j(t)\rangle = - \frac{1}{J}\sum_{j=1}^J\langle A^T\Gamma^{-1}Ae^j(t),C(t)A^T\Gamma^{-1}A e^j(t)\rangle \\ &=-\frac{1}{J^2}\sum_{j,k=1}^J \langle A^T\Gamma^{-1} Ae^j(t),e^k(t)\rangle^2 = -\frac{1}{J^2}\sum_{j,k=1}^J\langle Ae^ j(t),Ae^k(t)\rangle_\Gamma^2. \end{align*} We can bound this by removing all terms $j\neq k$ and applying Jensen's inequality for sums for the convex function $x\mapsto x^2$. \begin{align*} \dot{\mathfrak V}_e(t) &\leq -\frac{1}{J^2}\sum_{j=1}^J\norm{Ae^ j(t)}_\Gamma^4 \leq -\frac{1}{J}\left(\frac{1}{J}\sum_{j=1}^J\norm{Ae^ j(t)}_\Gamma^2\right)^2 = -\frac{4}{J}\mathfrak V_e(t)^2. \end{align*} An ordinary differential equation comparison principle then yields that \begin{align*} \mathfrak V_e(t) \leq \frac{1}{\frac{4}{J} t + \frac{1}{\mathfrak V_e(0)}}. \end{align*} Monotonous decrease is not true for the residual spread, whose derivative in time is \begin{align*} \dot{V}_r(t) &= \frac{1}{2J}\sum_{j=1}^J \langle \dot r^j(t), r^j(t)\rangle = -\frac{1}{2J}\sum_{j=1}^J \langle A^T\Gamma^{-1}A C(t) r^j(t), r^j(t)\rangle\\ &= -\frac{1}{2J^2}\sum_{i,j=1}^J \langle e^i(t), r^j(t)\rangle \langle e^i(t), A^T\Gamma^{-1}A r^j(t)\rangle, \end{align*} which does not carry a sign in general. \new{The residual spread does indeed not decay monotonously for deterministic EKI with noise-free data, which we study in the following.} Indeed, simple simulations (see \cref{fig:nonmonotonicity}) show that the residual norms can increase in time. There are two issues at play here: \begin{itemize} \item Correct choice of $u^\dagger$ in $r^j(t)=u^j(t)-u^\dagger$ and \item Skewness of the Euclidean norm with respect to the EKI dynamics. \end{itemize} First, the residuals $r^j$ and their mean $r$ are defined via a choice of ground truth parameter $u^\dagger$ such that $Au^\dagger = y$, \new{where $y$ is clean data.} While the degrees of freedom in this choice are irrelevant in the observation space, they are eminent in the parameter domain: For a given initial ensemble $\{u_0^j\}_{j=1}^J$, the mean $m(t)$ will converge to a well-defined limit $m_\infty := \lim_{t\to\infty}m(t)$ according to \cref{thm:deterministic_EKI} \new{and in the noise-free case it holds $m_\infty=m^\dagger$}. It is exactly this parameter we need to choose as a candidate for the reference parameter $u^\dagger$ in $r^j(t) := u^j(t)-u^\dagger$, which can be seen from the following reformulation of $V_r$. \begin{proposition} Let $u^\dagger\in\mathbb R^n$ such that $y=Au^\dagger$ and \new{let the residuals be defined by $r^j(t)=u^j(t)-u^\dagger$}. It holds \begin{align}\label{eq:reform_res_spread} V_r(t) = V_e(t) + \frac{1}{2}\norm{m(t)-u^\dagger}^2. \end{align} \end{proposition} \begin{proof} This can be seen as follows: \begin{align*} V_r(t) &= \frac{1}{2J}\sum_{j=1}^J\norm{r^j(t)}^2 =\frac{1}{2J}\sum_{j=1}^J\left(\norm{u^j(t)}^2-2\langle u^j(t),u^\dagger\rangle+\norm{u^\dagger}^2\right) \\ &=\frac{1}{2J}\sum_{j=1}^J\Big(\norm{u^j(t)}^2-2\langle u^j(t),m(t)\rangle+\norm{m(t)}^2 -2\langle u^j(t),u^\dagger-m(t)\rangle\\ &\hspace{5cm}+\norm{u^\dagger}^2-\norm{m(t)}^2\Big) \\ &=\frac{1}{2J}\sum_{j=1}^J\norm{u^j(t)-m(t)}^2 + \frac{1}{2}\norm{u^\dagger}^2-\langle m(t),u^\dagger-m(t)\rangle - \frac{1}{2}\norm{m(t)}^2 \\ &=V_e(t) + \frac{1}{2}\norm{m(t)-u^\dagger}^2. \end{align*} \end{proof} \new{We can see that $u^\dagger := m^\dagger$, defined as in \cref{thm:deterministic_EKI},} is the canonical choice because it guarantees \new{that the second term in the decomposition of $V_r(t)$ vanishes.} We obtain the following trivial corollary that for $u^\dagger=m^\dagger$ the ensemble spread and the residual spread converge to the same value (which is not zero, in general). \begin{corollary} Let $u^\dagger:=m^\dagger$. Then it holds \begin{align} \lim_{t\to\infty}V_e(t) = \lim_{t\to\infty}V_r(t). \end{align} \end{corollary} \begin{proof} The proof follows from \labelcref{eq:reform_res_spread} together with the fact that $m(t)\to m^\dagger$ \new{in the noiseless case according to \cref{thm:deterministic_EKI}.} \end{proof} However, even with this choice of $u^\dagger$, the residual spread $V_r$ can still \emph{fail to decrease} (see \cref{fig:nonmonotonicity} for an example). In contrast, the residual spread in observation space \labelcref{eq:fwd_residual_spread} does indeed decrease monotonously. Similar to above one can express $\mathfrak V_r(t)$ in terms of $\mathfrak V_e(t)$ and obtain monotonous convergence to zero. \begin{proposition} The residual spread in parameter space $\mathfrak V_r(t)$ admits the expression \begin{align}\label{eq:reform_fwd_res_spread} \mathfrak V_r(t) = \mathfrak V_e(t) + \frac{1}{2}\norm{A m(t) - y}_\Gamma^2. \end{align} Furthermore $t\mapsto\mathfrak V_r(t)$ is non-increasing and converges to zero with rate $1/t$. \end{proposition} \begin{proof} The proof of \labelcref{eq:reform_fwd_res_spread} work precisely as above for $V_r(t)$. Furthermore, it holds \new{% \begin{align*} \frac{\d}{\d t}\frac1 2 \|A m(t)-y\|_\Gamma^2 &= \langle A^T\Gamma^{-1}(A\mathsf{m}(t)-y),\dot{m}(t)\rangle\\ &=-\langle A^T\Gamma^{-1}(A m(t)-y),C(t)A^T\Gamma^{-1}(A m(t)-y)\rangle\\ &=-\frac{1}{J}\sum_{j=1}^J\langle A^T\Gamma^{-1}(A m(t)-y),e^j(t)\otimes e^j(t)A^T\Gamma^{-1}(A\mathsf{m}(t)-y)\rangle\\ &=-\frac{1}{J}\sum_{j=1}^J\langle A^T\Gamma^{-1}(A m(t)-y),e^j(t)\rangle^2 \leq 0. \end{align*} } Hence, using that $t\mapsto\mathfrak V_e(t)$ converges to zero monotonously with rate $1/t$ and that the same holds true for $t\mapsto\frac1 2 \|A m(t)-y\|_\Gamma^2$ (see~\labelcref{eq:rate_fwd_x} in the case $\alpha=2$ for the rate), we obtain the assertion. \end{proof} We emphasize again that monotone convergence of $t\mapsto \frac{1}{2}\|A m(t)-y\|_\Gamma^2$ does not mean that the quantities $\norm{m(t)-m^\dagger}$ or $V_r(t)$ decrease as well. First, the mapping of this quantity via $A$ only keeps track of the data-informed parameter dimensions, i.e., those orthogonal to the kernel of $A$. And secondly, even invertibility of $A$ still does not imply monotonicity of $\|m(t)-u^\dagger\| $ as the mapping $A$ can warp the coordinate system in such a way that this property is lost. This can be seen in an elementary example unrelated to the EKI: Consider the curve $x(t) = (\cos(t),\sin(t))$ for which $V(t) := \|x(t)\|^2$ is constant, i.e., monotonously non-increasing. On the other hand, with $A=\operatorname{diag}(2,1)$, the mapping $\tilde V(t) = \|Ax(t)\|^2$ is not monotonous. \begin{example}\label{ex:nonmon} As a concrete example for the non-monotonicity of the mean and the residual, we can consider the forward operator $A = \operatorname{diag}(100,1)$, observation $y = (0,0)^T$, and an initial ensemble with mean $m_0 = (100,100)^T$ and empirical covariance \begin{align*} C_0 = \begin{pmatrix} \phantom{-}25 & -24\\ -24 & \phantom{-}25 \end{pmatrix}, \end{align*} whose eigenvectors are $(-1,1)^T$ and $(1,1)^T$ with eigenvalues $49$ and $1$, respectively. \cref{fig:nonmonotonicity} shows the initial ensemble and the trajectories of the ensemble and its sample mean in the parameter space. Clearly, the sample mean and the whole ensemble move away from their final limit $(0,0)^T$ for quite some time until they finally ``change direction'' and converge towards their limit. The initial shearing of the ensemble combined with the strong weighting of the horizontal direction, which is encoded in the forward operator, leads to an initial movement of the ensemble along its principal axis to the top left. \begin{figure}[hbt] \centering {\includegraphics[width=0.6\textwidth,trim=2cm 2.5cm 5cm 2.5cm,clip]{figs/dynamics}}% \caption{From bottom right to top left: Trajectories of deterministic EKI (black curve is the mean $m(t)$) for $t\in[0,1]$. The Euclidean sphere (dotted) through $m_\infty$ demonstrates non-monotonicity of the mean.} \label{fig:nonmonotonicity} \end{figure} \end{example} Applied to the present setting, this means that the Euclidean norm is not the natural norm with respect to which we should view the dynamics of the ensemble. Hence, we need to either settle for \textit{non-monotonous convergence} of $\norm{m(t)-m^\dagger}$, or we need to pick a more \textit{problem-adapted norm}, as presented in the following. \subsection{Monotonicity in a Problem-adapted Norm} Now we see how to define a new norm on the parameter space with respect to which we can prove monotonous convergence of the residuals \new{of deterministic EKI with data in the range of the forward operator}. Recall the ordinary differential equation \labelcref{eq:ODE_mean_emp} for \new{the empirical mean $m(t)$ with data} $y=A\xi\in\mathrm{ran}(A)$: \begin{align*} \dot m(t) = -C(t)A^T\Gamma^{-1}(Am(t)-y) = -SD(t)S^{-1}(m(t) - \xi). \end{align*} By defining $\rho(t) = S^{-1}m(t)$, we obtain the ordinary differential equation \begin{align*} \dot\rho(t) = -D(t) (\rho(t) - S^{-1}\xi), \end{align*} which decouples into $n$ ordinary differential equations since $D(t)$ is a diagonal matrix. This allows us to prove the following Lyapunov type estimate. \begin{proposition}\label{prop:lyapunov} Let $S$ be such that $C_0 A^T\Gamma^{-1}A = SDS^{-1}$ as in \cref{thm:cov_dynamics}. Then \begin{align*} L(m) := \frac{1}{2}\|S^{-1}(m - \xi)\|^2 \end{align*} is a Lyapunov function for the dynamics of the empirical mean $m(t)$, meaning that $\frac{\mathrm d}{\mathrm d t}{L}(m(t)) \leq 0$. \end{proposition} \begin{proof} We observe $L(m(t))=\frac{1}{2}\norm{\rho(t) - S^{-1}\xi}^2$ and compute \begin{align*} \frac{\mathrm d}{\mathrm d t}L(m(t)) = \langle \dot\rho(t), \rho(t)-S^{-1}\xi\rangle = -\norm{\sqrt{D(t)}\rho(t)}^2\leq 0. \end{align*} \end{proof} \begin{remark} A couple of remarks regarding this Lypunov approach are in order. \begin{itemize} \item The strength of using the norm $\norm{S^{-1}\cdot}$ instead of the Euclidean norm, is that it captures exactly the correct notion of convergence of $x$ by respecting both the influence of the forward mapping $A$ and the initial ensemble $C_0$ \new{(which is either implicitly assumed to be an approximation of the prior, or explicitly chosen to have similar statistical properties). In particular, the ``pure'' Euclidean norm does not yield the right notion of tracking convergence, as \cref{ex:nonmon,fig:nonmonotonicity} show.} \item The proof that the dynamics behave monotonously in this norm did not use the explicit solution for $m(t)$ derived in \cref{thm:deterministic_EKI}. Therefore, we hope that this Lyapunov approach might work similarly in the stochastic setting where an explicit solution is not readily available. \item Since $S^{-1}$ is a regular matrix, the functional $L(m)$ is coercive which will turn out useful for showing existence of a limit of $m(t)$ as $t\to\infty$ in more general settings. \item ``Preconditioning'' with $S^{-1}$ also allows one to show that $C(t)A^T\Gamma^{-1}A$ has the same eigenvectors for all times, without using the explicit solution for $C(t)$. From $C_0 A^T\Gamma^{-1}A = SDS^{-1}$ we see that $\dot{C}(t)$ is diagonalized in the same way: \begin{align*} S^{-1} \dot{C}(t)A^T\Gamma^{-1}A S = -S^{-1} C(t)A^T\Gamma^{-1}A C(t)A^T\Gamma^{-1}AS, \end{align*} i.e., if we set $D(t):=S^{-1}C(t)A^T\Gamma^{-1}AS$, we obtain the very simple ordinary differential equation $\dot{D}(t) = -D^2(t)$. This proves that $D(t)$ stays diagonal for all $t\geq 0$ and we obtain the diagonalization $C(t)A^T\Gamma^{-1}A = S D(t) S^{-1}$, which we have already derived in \cref{thm:cov_dynamics} with other techniques. \end{itemize} \end{remark} \section{Spectral Decomposition of the Covariance} \label{sec:spectral} Recall that we are analysing the behavior of the EKI, with its covariance dynamics given by \labelcref{eq:ODE_cov}, which is \begin{align*} \dot{\mathsf{C}}(t) = -\alpha\mathsf{C}(t) A^T \Gamma^{-1} A \mathsf{C}(t), \quad \mathsf{C}(0) = \mathsf{C}_0. \end{align*} In the previous sections we have intensively used the diagonalization $\mathsf{C}(t)\mathsf{C}_0^{-1} = S E(t) S^{-1}$ from \cref{thm:cov_dynamics} to understand the \new{deterministic} dynamics of \new{different versions of} EKI. The eigenvectors of the matrix $\mathsf{C}(t)\mathsf{C}_0^{-1}$ do not change in time and their associated eigenvalues have an explicit expression, \new{as has been shown in \cref{sec:diagonalization}, and also in \cite{garbuno2020interacting}}. None of this is true for the covariance matrix $\mathsf{C}(t)$ itself and, in particular, its eigenvectors can change drastically in time. In this section, we derive a coupled system of ordinary differential equations which is solved by the eigenvalue and eigenvectors of $\mathsf{C}(t)$. \new{For this we \emph{do not} use the explicit solution for $\mathsf{C}(t)$. Instead we only utilize its governing dynamics \labelcref{eq:ODE_cov}, re-displayed above.} To this end, denote by $\lambda_1(t),\dots,\lambda_n(t)$ the eigenvalues of $\mathsf{C}(t)$, with eigenvectors $v_1(t),\ldots,v_n(t)$. Since $\mathsf{C}(t)$ is symmetric, all eigenvectors can be chosen orthonormal and we can assume the ordering $\lambda_1(t) \geq \dots \geq \lambda_n(t)$ of the eigenvalues. \begin{theorem}[Eigenvector Dynamics]\label{thm:eigenvectors} Let $\mathsf{C}(t)$ denote the solution of \labelcref{eq:ODE_cov} with initial condition $\mathsf{C}_0$. Denote the eigenvalues and normalized eigenvectors of $\mathsf{C}(t)$ by $\lambda_i(t)$ and $v_i(t)$ for $i=1,\dots,n$. Then it holds: \begin{itemize} \item Any set of eigenvalues $\lambda_i(t)$ with corresponding eigenvectors $v_i(t)$ for $i=1,\dots,n$, differentiable in time, solve the following differential algebraic system: \begin{subequations}\label{eq:DAE} \begin{align} \label{eq:ODE_eigvalues} \dot{\lambda}_i &= -\new{\alpha}\lambda_i^2 \norm{A v_i}_\Gamma^2,\\ \label{eq:ODE_eigvectors} \dot{v}_i &= \sum_{\substack{j\in\{1,\dots,n\}:\\\lambda_j \neq \lambda_i}} \frac{\new{\alpha}\lambda_i\lambda_j}{\lambda_j-\lambda_i} \langle A v_i, A v_j\rangle_\Gamma v_j,\\ \label{eq:AE_orth} 0 &= \lambda_i^2 \langle Av_i, Av_j\rangle_\Gamma,\quad\text{if }i\neq j\text{ but }\lambda_i=\lambda_j. \end{align} \end{subequations} \item The eigenvalue adhere to the following bounds \begin{align} \lambda_i(t) &\geq \frac{\lambda_i(0)}{\new{\alpha}\norm{\Gamma^{-1/2}A}^2t\lambda_i(0) + 1}, \qquad\forall i=1,\dots,n, \\ \lambda_{1}(t) &\geq \frac{\lambda_1(0)}{\new{\alpha}\norm{Av_1(0)}_\Gamma^2 t\lambda_1(0) + 1}, \\ \lambda_{n}(t) &\leq \frac{\lambda_n(0)}{\new{\alpha}\norm{Av_n(0)}_\Gamma^2t\lambda_n(0) + 1}. \end{align} \item The eigenvalues and eigenvectors have the following asymptotic behavior \begin{align} (\forall i = 1,\dots,n) \quad \lim_{t\to\infty}\lambda_i(t) = 0 \quad\text{or}\quad \lim_{t\to\infty} Av_i(t) = 0. \end{align} \end{itemize} \end{theorem} \begin{remark} The case of eigenvalues with $\lambda_i(t^\star) = \lambda_j(t^\star)$ but $\lambda_i(t)\neq \lambda_j(t)$ for $t\neq t^\star$ is a source of ambiguity for both the labelling of the eigenvalues and for the well-posedness of the system \labelcref{eq:DAE}. We disregard this case by considering the differential equations only for $t<t^\star$ and $t>t^\star$, and then completing by continuity. If $\lambda_i(t) = \lambda_j$ for all $t$ in a proper interval, then the dynamics of the eigenvectors has an intrinsic ambiguity which we remove by setting $\langle \dot v_i, v_j\rangle = 0$. \end{remark} \begin{proof} From the explicit solution \labelcref{eq:sol_cov} it follows that $\mathsf{C}(t)$ is symmetric and positive definite for all times and thus diagonalizable with orthonormal eigenvectors. We start with the defining equations for the eigenvectors and eigenvalues while enforcing normality: \begin{align*} \left(\mathsf{C}(t)-\lambda_i(t)E\right)v_i(t) &= 0,\\ \|v_i(t)\|^2 &= 1. \end{align*} By taking the time derivative in both equations (and dropping the explicit dependence on $t$ for brevity), we obtain \begin{align} (\mathsf{C}-\lambda_i E)\dot{v}_i &= \dot{\lambda}_i v_i - \dot{\mathsf{C}}v_i,\\ \langle\dot{v}_i,v_i\rangle &=0. \label{eq:increment_orth} \end{align} By using \labelcref{eq:ODE_cov} and $\mathsf{C} v_i = \lambda_i v_i$, this means that \begin{align*} \mathsf{C} \dot{v}_i - \lambda_i\dot{v}_i = \dot{\lambda}_i v_i + \new{\alpha}\mathsf{C} A^T\Gamma^{-1}A\mathsf{C} v_i = \dot{\lambda}_i v_i + \new{\alpha}\lambda_i \mathsf{C} A^T\Gamma^{-1}A v_i. \end{align*} Now we take the scalar product of both sides with $v_j$ and obtain \begin{align*} \langle\mathsf{C} \dot{v}_i,v_j\rangle - \lambda_i \langle\dot{v}_i,v_j\rangle = \dot{\lambda}_i \langle v_i,v_j\rangle + \new{\alpha}\lambda_i \langle\mathsf{C} A^T\Gamma^{-1}A v_i,v_j\rangle, \end{align*} which is equivalent to \begin{align}\label{eq:evecs_evals} (\lambda_j - \lambda_i)\langle\dot{v}_i,v_j\rangle = \dot{\lambda}_i \delta_{i,j} + \new{\alpha}\lambda_i \lambda_j \langle A v_i,A v_j\rangle_\Gamma, \end{align} where $\delta_{i,j}$ is $1$ if $i=j$ and $0$ otherwise. We have three cases to consider: Firstly, in the case $i=j$, we get \begin{align*} \dot{\lambda}_i = -\new{\alpha}\lambda_i^2 \|\Gamma^{-1/2}Av_i\|^2 = -2\lambda_i^2 \norm{Av_i}_\Gamma^2, \end{align*} which proves~\labelcref{eq:ODE_eigvalues}. Secondly, if $i\neq j$ and $v_i, v_j$ are two different eigenvectors for the same eigenvalue $\lambda_i = \lambda_j$, then \labelcref{eq:evecs_evals} implies \begin{align*} \lambda_i^2 \langle A v_i, A v_j\rangle_\Gamma = 0, \end{align*} which proves \labelcref{eq:AE_orth}. Thirdly, in the case $i\neq j$ and $\lambda_i\neq \lambda_j$ we conclude from \labelcref{eq:evecs_evals} that \begin{align*} \langle \dot{v}_i,v_j\rangle = \frac{\new{\alpha}\lambda_i\lambda_j}{\lambda_j-\lambda_i} \langle Av_i,Av_j\rangle_\Gamma. \end{align*} Using this equation together with \labelcref{eq:increment_orth} and orthonormality of the $v_i$, we can express $\dot{v}_i$ in the basis $\{v_j\}_j$ as \begin{align*} \dot{v}_i = \sum_{\lambda_j \neq \lambda_i} \frac{\new{\alpha}\lambda_i\lambda_j}{\lambda_j-\lambda_i} \langle A v_i,A v_j\rangle_\Gamma v_j. \end{align*} where we set $\langle\dot v_i, v_j\rangle = 0$ for $v_i,v_j$ in the same eigenspace for a joint eigenvalue $\lambda = \lambda_i=\lambda_j$. The lower bound decay rate on the eigenvalues follows from the fact that the $v_i$ have unit norm and we can bound $\norm{A v_i}_\Gamma^2 \leq \norm{\Gamma^{-1/2}A}^2$. Furthermore, we see that \begin{align*} \Gamma^{-1/2}A \dot{v}_i = \sum_{\lambda_j\neq\lambda_i} \frac{\new{\alpha}\lambda_i\lambda_j}{\lambda_j-\lambda_i} \langle A v_i,A v_j\rangle_\Gamma \Gamma^{-1/2}A v_j \end{align*} and thus \begin{align*} \frac{\d}{\d t}\frac{1}{2}\|A v_i\|_\Gamma^2 = \sum_{\lambda_j\neq \lambda_i} \frac{\new{\alpha}\lambda_i\lambda_j}{\lambda_j-\lambda_i} \langle A v_i,A v_j\rangle_\Gamma^2, \end{align*} from which we can similarly derive bounds for the cases $i=1$, i.e., $\frac{\d}{\d t}\|A v_i\|_\Gamma^2 \leq 0$ and $i=n$, i.e., $\frac{\d}{\d t}\|A v_i\|_\Gamma^2 \geq 0$. For the last pillar of \cref{thm:eigenvectors} we argue as follows: From \cref{thm:cov_dynamics}, we know that \begin{align*} A^T\Gamma^{-1}A\mathsf{C} (t) = S\operatorname{diag}\left(\frac{\mu_i}{1+\new{\alpha} t\mu_i}\right)_{i=1}^n S^{-1} \to 0,\quad t\to\infty. \end{align*} If we now multiply this from the right with an eigenvector $v_i(t)$ of $\mathsf{C} (t)$ corresponding to an eigenvalue $\lambda_i(t)$, we obtain \begin{align*} \lambda_i(t) A^T\Gamma^{-1}A v_i(t) = A^T\Gamma^{-1}A\mathsf{C} (t)v_i(t) \to 0,\quad t\to\infty. \end{align*} i.e., $\lambda_i(t)\to 0$ or $Av_i(t)\to 0$, as claimed. \end{proof} \begin{corollary} The function $t\mapsto \lambda_1(t)$ is convex. \end{corollary} \begin{proof} We can take another derivative in \labelcref{eq:ODE_eigvalues} and obtain \begin{align*} \ddot{\lambda}_i &= -2\new{\alpha}\lambda_i\dot{\lambda}_i\norm{Av_i}_\Gamma^2 - 2\new{\alpha}\lambda_i^2\frac{\mathrm d}{\mathrm d t}\frac{1}{2}\norm{Av_i}_\Gamma^2 \\ &= 2\new{\alpha}^2 \lambda_i^3\norm{Av_i}_\Gamma^4 - 2\new{\alpha}^2\lambda_i^3 \sum_{\lambda_j\neq\lambda_i}\frac{\lambda_j}{\lambda_j-\lambda_i}\langle Av_i,Av_j\rangle_\Gamma^2\\ &=2\new{\alpha}^2\lambda_i^3\left[\norm{Av_i}_\Gamma^4 - \sum_{\lambda_j\neq\lambda_i}\frac{\lambda_j}{\lambda_j-\lambda_i}\langle Av_i,Av_j\rangle_\Gamma^2\right]. \end{align*} Hence, for $i=1$ the second term is non-positive and one obtains $\ddot{\lambda}_i \geq 0$ which implies convexity. \end{proof} \section{Conclusions and Outlook} In this article we have provided a complete description of the deterministic dynamics of the Ensemble Kalman Inversion, based on the spectral decomposition of a preconditioned empirical covariance operator. We focused on deterministic EKI and mean-field EKI. In particular, we have derived the time-asymptotic behavior of the covariance and also studied asymptotic profiles, their second-order asymptotics. Then, we computed the explicit dynamics and convergence rates of particles and their empirical mean for noisy data, and showed consistency for vanishing noise. We applied these findings to the study of ensemble and residual spreads of deterministic EKI both in observation and parameter space, in particular, we constructed a counter example which shows that the residual spread in the Euclidean norm is not decreasing, in general. This inspired us to define a ``problem-adapted'' weighted norm, depending on the forward model and the prior, with respect to which the ensemble spread decreases monotonously. We concluded our studies with a spectral analysis of the ``pure'' covariance operator and derived the governing differential equations for its eigenvalues and eigenvectors. Our analysis shows that if one is interested in the Bayesian context, the stochastic version or an approximation of mean-field EKI might a better choice than deterministic EKI, although in the viewpoint of EKI as a derivative-free optimization method (e.g., in image reconstruction where the exact noise model is unknown), this is less relevant. All versions considered here share the same asymptotic behavior, albeit with different rates of convergence, and therefore are all basically equivalent in the zero-noise limit. The presented results about the convergence rates might be used to accelerate EKI using time re-parametrization. Furthermore, they might help to derive sharp quantitative estimates for the number of particles $J\in\mathbb N$ needed based on the noise level. \neww{ An important extension of our results is the analysis of the averaged stochastic equations \labelcref{eq:ODE_mean_av_emp,eq:ODE_cov_av_emp}. This is non-trivial since it requires a careful control of the covariance-like corrections terms and their behavior in terms of the ensemble size~$J$. } \new{Even more general, one can consider the much more challenging settings of purely stochastic EKI (i.e., where we need to analyse stochastic differential equations governing the evolution of the ensemble members), the EKI applied to a nonlinear forward problem (which means that the Bayesian posterior is not a Gaussian measure anymore), and of time-discretizations of the EKI used on problems in practice.} \printbibliography \clearpage \begin{appendix} \section{\new{Auxiliary derivations}} \begin{lemma}[Mean and covariance dynamics of deterministic EKI]\label{lem:derivation_ODEs} \new{Consider the particle dynamics \begin{equation*} \dot{u}^j(t) = -C(t) A^T\Gamma^{-1}(Au^j(t)-y) \end{equation*} with sample covariance $C(t) :=\frac{1}{J}\sum_{j=1}^J(u^j(t)-m(t)) \otimes (u^j(t)-m(t))$ and sample mean $m(t) := \frac{1}{J}\sum_{j=1}^J u^j(t)$. These two quantities are governed by the following differential equations: \begin{alignat}{2} \dot{m}(t) &= -C(t) A^T\Gamma^{-1}(Am(t)-y), \quad &&m(0) = m_0,\\ \dot{C}(t) &= -2C(t) A^T \Gamma^{-1} A C(t), \quad &&C(0) = C_0. \end{alignat} \begin{proof} The differential equation for the mean follows directly by summing the particle equation. The covariance dynamics can be derived as follows: \begin{align*} \dot C(t) &= \frac{1}{J}\sum_{j=1}^J(\dot u^j(t) - \dot m(t))\otimes (u^j(t) - m(t)) + (u^j(t) - m(t))\otimes (\dot u^j(t) - \dot m(t)) \\ &= -\frac{1}{J}\sum_{j=1}^J C(t)A^T\Gamma^{-1}A (u^j(t)-m(t))\otimes (u^j(t)-m(t)) \\ &- \frac{1}{J}\sum_{j=1}^J (u^j(t)-m(t))\otimes \left[C(t)A^T\Gamma^{-1}A (u^j(t)-m(t)) \right]\\ &= -C(t)A^T\Gamma^{-1}A\frac{1}{J}\sum_{j=1}^J (u^j(t)-m(t))\otimes (u^j(t)-m(t))\\ & - \frac{1}{J}\sum_{j=1}^J (u^j(t)-m(t))\otimes (u^j(t)-m(t)) \cdot A^T\Gamma^{-1}A C(t)\\ &=-2C(t)A^T\Gamma^{-1}AC(t) \end{align*} \end{proof}} \end{lemma} \begin{lemma}[Mean and covariance dynamics of EKI]\label{lem:derivation_SDEs} \neww{% We consider particles governed by \begin{align*} \dot{u}^j(t) = -C(t) A^T\Gamma^{-1}(Au^j(t)-y) + C(t)A^T\Gamma^{-1}\sqrt{\Sigma}\dot{\mathbf{W}}^j(t) \end{align*} with $\Sigma=\Gamma^{-1}$, sample covariance $C(t):=\frac{1}{J}\sum_{j=1}^J(u^j(t)-m(t)) \otimes (u^j(t)-m(t))$, and sample mean $m(t) := \frac{1}{J}\sum_{j=1}^J u^j(t)$. These two quantities are governed by the following stochastic differential equations: \begin{align} \dot m(t) &= -C(t)A^T\Gamma^{-1}(Am(t)-y) + C(t) A^T\Gamma^{-1/2}\dot{\overline\mathbf{W}}(t)\\ \dot C(t) &= -\frac{J+1}{J}C(t) A^T\Gamma^{-1}A C(t) \\\notag &+ \frac1 {J}\sum_{j=1}^J e^j(t)\otimes (\dot\mathbf{W}^j(t)-\dot{\overline\mathbf{W}}(t)) \Gamma^{-1/2}A C(t) \\\notag &+ \frac1 {J}\sum_{j=1}^J C(t)A^T\Gamma^{-1/2}(\dot\mathbf{W}^j(t)-\dot{\overline\mathbf{W}}(t))\otimes e^j(t) \end{align} with $\overline\mathbf{W}(t) = \frac 1 J\sum_{j=1}^J\mathbf{W}^j(t)$. In addition, the \emph{average empirical mean} $\mathbf{m}(t):= \mathbb E^\mathbf{W} m(t)$ and \emph{average empirical covariance} $\mathbf{C}(t) := \mathbb E^\mathbf{W} C(t)$ satisfy the following differential equations: \begin{align} \dot{\mathbf C}(t) &= -\frac{J+1}{J}\mathbf C(t) A^T \Gamma^{-1} A \mathbf C(t)\notag \\ &\qquad - \mathbb E^\mathbf{W}\left[(C(t)-\mathbf C(t))A^T\Gamma^{-1}A(C(t)-\mathbf C(t))\right],\\ \dot{\mathbf m}(t) &= -\mathbf C(t) A^T\Gamma^{-1}(A\mathbf m(t)-y) \notag \\ &\qquad -\mathbb E^\mathbf{W}\left[(C(t)-\mathbf C(t))A^T\Gamma^{-1}A(m(t)-\mathbf m(t))\right] \end{align} \begin{proof} The mean SDE is directly obtained by summing the particle dynamics. The SDE governing the covariance is computed as follows: By setting $e^j = u^j-m$ and using $C = \frac{1}{J}\sum_{j=1}^Je^j\otimes e^j$, It\=o's formula yields \begin{align*} \mathrm d(e^j\otimes e^j) &= \mathrm d e^j\otimes e^j + e^j\otimes \mathrm d e^j + \frac{1}{2}\cdot 2\cdot \mathrm d e^j\otimes \mathrm d e^j\\ &=-CA^T\Gamma^{-1}A(e^j\otimes e^j) \d t - (e^j\otimes e^j)A^T\Gamma^{-1}AC \d t\\ &+CA^T\Gamma^{-1}(\mathrm d\mathbf{W}^j-\mathrm d\overline \mathbf{W})\otimes e^j + e^j\otimes (\mathrm d\mathbf{W}^j-\mathrm d\overline \mathbf{W})\Gamma^{-1/2}AC \\ &+ CA^T\Gamma^{-1/2}(\mathrm d\mathbf{W}^j-\mathrm d\overline \mathbf{W})\otimes (\mathrm d\mathbf{W}^j-\mathrm d\overline \mathbf{W})\Gamma^{-1/2}AC. \end{align*} Now we use $(\mathrm d\mathbf{W}^j-\mathrm d\overline\mathbf{W})\otimes (\mathrm d\mathbf{W}^j-\mathrm d\overline\mathbf{W}) = \frac{J-1}{J}E$, where $E$ is the identity matrix, which is a simple application of It\=o calculus, or a consequence of Lemma A.1 in \cite{blomker2019well}. Then we can sum over $j$ in order to obtain \begin{align*} \mathrm d C(t) &= \frac{1}{J}\sum_{j=1}^J\mathrm d(e^j\otimes e^j)\\ &=\left(-2 + \frac{J-1}{J}\right)CA^T\Gamma^{-1}AC\d t \\ &+ \frac{1}{J}\sum_{j=1}^J CA^T\Gamma^{-1}(\mathrm d\mathbf{W}^j-\mathrm d\overline\mathbf{W})\otimes e^j\\ &+ \frac{1}{J}\sum_{j=1}^J e^j\otimes (\mathrm d\mathbf{W}^j-\mathrm d\overline\mathbf{W}) \Gamma^{-1}AC. \end{align*} We write this in integral form to obtain \begin{align*} C(t) -C(0) &=\left(-2 + \frac{J-1}{J}\right)\int_0^t C(s)A^T\Gamma^{-1}AC(s)\d s \\ &+ \frac{1}{J}\sum_{j=1}^J\int_0^t C(s)A^T\Gamma^{-1}(\mathrm d\mathbf{W}^j(s)-\mathrm d\overline\mathbf{W}(s))\otimes e^j\\ &+ \frac{1}{J}\sum_{j=1}^Je^j\otimes\int_0^t (\mathrm d\mathbf{W}^j(s)-\mathrm d\overline\mathbf{W}(s)) \Gamma^{-1}AC(s). \end{align*} By proving that the stochastic integrals are indeed martingales (as in \cite{blomker2019well}), we can drop them after taking the expectation, i.e. \begin{align*} \frac{\d}{\d t}\mathbb E C(t) &=\left(-2 + \frac{J-1}{J}\right) \mathbb E[C(t)A^T\Gamma^{-1}AC(t)] \\ &= -\frac{J+1}{J}\mathbb E[C(t)]A^T\Gamma^{-1}A\mathbb E[C(t)] \\ & -\frac{J+1}{J}\mathbb E[(C-\mathbb E[C])A^T\Gamma^{-1}A(C-\mathbb E[C])]. \end{align*} The ODE for $\mathbf{m}(t)=\mathbb E^\mathbf{W} m(t)$ is derived analogously. \end{proof} } \end{lemma} \neww{% Finally, we provide a formal derivation of the mean-field dynamics. \begin{lemma}[Mean and covariance dynamics for mean-field EKI]\label{lem:derivation_meanfield} Let $\rho(t,u)$ be a solution of \begin{align*} \partial_t \rho = \div\left(\rho~\mathfrak C(t)A^T\Gamma^{-1}(A u - y)\right) + \frac12 \operatorname{Tr}(D^2\rho~\mathfrak C(t)A^T\Gamma^{-1}A\mathfrak C(t)) \end{align*} and let $\mathfrak m(t)$ and $\mathfrak C(t)$ be defined by \labelcref{eq:meanfield_mean,eq:meanfield_cov}. These two quantities are governed by the following differential equations: \begin{align} \dot{\mathfrak{m}}(t) &= -\mathfrak{C}(t)A^T\Gamma^{-1}(A\mathfrak{m}(t)-y),\\ \dot{\mathfrak{C}}(t) &= -\mathfrak{C} (t) A^T\Gamma^{-1}A \mathfrak{C}(t). \end{align} \end{lemma} \begin{proof} The following computation is entirely formal, assuming that $\rho(t,u)$ is absolutely continuous with respect to the Lebesgue measure, and sufficiently smooth. Using the definition of $\mathfrak m(t)$ and expanding the trace operator, one computes \begin{align*} \dot{\mathfrak m}(t) &=\int u~\partial_t\rho(t,u)\d u\\ &=\int u \div(\rho(t,u)\mathfrak C(t)A^T\Gamma^{-1}(Au-y))\d u \\ &\qquad + \frac{1}{2}\sum_{i,j}\left(\mathfrak{C}(t)A^T\Gamma^{-1}A\mathfrak C(t)\right)_{ij}\int u \partial_i\partial_j \rho(t,u) \d u. \end{align*} Integrating the second term by parts, we see that it vanishes. Integrating the first term by parts yields \begin{align*} \dot{\mathfrak m}(t) &= - \mathfrak{C}(t)A^T\Gamma^{-1}A\int u~\rho(t,u)\d u + \mathfrak{C}(t)A^T\Gamma^{-1}y\int\rho(t,u)\d u \\ &= - \mathfrak{C}(t)A^T\Gamma^{-1}(A\mathfrak{m}(t)-y), \end{align*} where we used $\int\rho(t,u)\d u=1$. Using the dynamics of the mean we can derive the dynamics of the covariance as follows. We compute using the product rule \begin{align*} \dot{\mathfrak C}(t) &= 2\int (u-\mathfrak m(t))\otimes(\mathfrak C(t)A^T\Gamma^{-1}(A\mathfrak m(t)-y))\rho(t,u)\d u \\ &\qquad+ \int (u-\mathfrak m(t))\otimes(u-\mathfrak m(t))\div(\rho(t,u)\mathfrak C(t)A^T\Gamma^{-1}(Au-y))\d u \\ &\qquad+\mathfrak \sum_{i,j}\left(C(t)A^T\Gamma^{-1}A\mathfrak C(t)\right)_{ij}\int \frac{1}{2} (u-\mathfrak m(t))\otimes(u-\mathfrak m(t))\partial_i\partial_j\rho(t,u)\d u. \end{align*} Integrating the second term $(II)$ by parts shows \begin{align*} (II)=-2\int (u-\mathfrak m(t))\otimes\left(\mathfrak C(t)A^T\Gamma^{-1}(Au-y)\right)\rho(t,u)\d u. \end{align*} Similarly, the third term $(III)$ integrates to \begin{align*} (III)=\mathfrak C(t)A^T\Gamma^{-1}A\mathfrak C(t)\int \rho(t,u)\d u = \mathfrak C(t)A^T\Gamma^{-1}A\mathfrak C(t). \end{align*} Putting everything together one obtains \begin{align*} \dot{\mathfrak C}(t) &=-2\int(u-\mathfrak m(t))\otimes(\mathfrak C(t)A^T\Gamma^{-1}A(u-\mathfrak m(t)))\rho(t,u)\d u + \mathfrak C(t)A^T\Gamma^{-1}A\mathfrak C(t)\\ &=-2\int(u-\mathfrak m(t))\otimes(u-\mathfrak m(t))\rho(t,u)\d u \left(\mathfrak C(t)A^T\Gamma^{-1}A\right)^T + \mathfrak C(t)A^T\Gamma^{-1}A\mathfrak C(t)\\ &=-\mathfrak C(t)A^T\Gamma^{-1}A\mathfrak C(t). \end{align*} \end{proof} } \end{appendix} \end{document}

a8cebd99d157328860891d896446a979d5aee7db

\section{Introduction} \subsection{Boundedness of spectral projectors on Riemannian manifolds} \subsubsection{A general problem} Given a Riemannian manifold $M$ with Laplace-Beltrami operator $\Delta$, and for $\lambda \geq 1$, $0<\delta <1$, let $$ P_{\lambda,\delta} = P_{\lambda,\delta}^{\chi} = \chi \left( \frac{\sqrt{-\Delta} - \lambda}{\delta} \right). $$ where $\chi$ is a non-negative cutoff function supported in $[-1,1]$, equal to $1$ on $[-\frac{1}{2},\frac{1}{2}]$. A general question is to estimate $$ { \| P_{\lambda,\delta}^\chi \|_{L^2 \to L^p}, \qquad \mbox{where $p \in [2,\infty]$}}. $$ Using self-adjointness of $P_{\lambda,\delta}^\chi$ and a $TT^*$ argument, it follows that \begin{equation} \label{mallard} \| P^{\chi^2}_{\lambda,\delta} \|_{L^{p'} \to L^p} = \| P_{\lambda,\delta}^\chi \|_{L^{p'} \to L^2}^2 = \| P_{\lambda,\delta}^\chi \|_{L^{2} \to L^p}^2. \end{equation} Furthermore, given two cutoff functions $\chi$ and $\widetilde{\chi}$, the boundedness of $P_{\lambda,\delta}^{\chi}$ on $L^2$ implies the following: if $\| P_{\lambda,\delta}^{\chi} \|_{L^2 \to L^p}$ obeys, say, a polynomial bound of the type $\lambda^\alpha \delta^\beta$, so does $\| P_{\lambda,\delta}^{\widetilde \chi} \|_{L^2 \to L^p}$, with a different constant. Therefore, it will be equivalent to estimate either of the three quantities appearing in~\eqref{mallard}, and the result is essentially independent of the cutoff function, which might even be taken to be a sharp cutoff. Up to possibly logarithmic factors, this question is essentially equivalent to that of estimating the $L^2 \to L^p$ norm of the resolvent $R((x+iy)^2) = (\Delta + (x+iy)^2)^{-1}$; this is the point of view taken in Dos Santos Ferreira-Kenig-Salo~\cite{DKS} and Bourgain-Shao-Sogge-Yao \cite{BSSY}. Essentially, one can think of $R((x+iy)^2)$ as a variant of $\frac{1}{xy} P_{x,y}$. \subsubsection{The case of Euclidean space} \label{sec:euclidean-space} We will denote the Stein-Tomas exponent $$ p_{ST} = \frac{2(d+1)}{d-1}. $$ As will become clear, it often plays the role of a critical point when estimating the norm of $P_{\lambda,\delta}$. On $\mathbb{R}^d$ (with the Euclidean metric), there holds \begin{equation} \label{swallow} \| P_{\lambda,\delta} \|_{L^2 \to L^p} \lesssim \left\{ \begin{array}{ll} \lambda^{\sigma(p)/2} \delta^{1/2} & \mbox{if $p \geq p_{ST}$} \\ \lambda^{\frac{d-1}{2} \left( \frac{1}{2} - \frac{1}{p} \right)} \delta^{\frac{(d+1)}{2}\left( \frac{1}{2} - \frac{1}{p} \right)} & \mbox{if $2 \leq p \leq p_{ST}$}, \end{array} \right. \end{equation} where $$ \sigma(p) = d - 1 - \frac{2d}{p} \qquad \mbox{so that} \qquad \sigma(p_{ST}) = \frac{d-1}{d+1}; $$ see the appendix for a proof of the above bounds. For more general second order operators in the resolvent formulation, we refer to Kenig-Ruiz-Sogge~\cite{KRS}. Finally, the case of the hyperbolic space was recently treated by the first author and L\'eger~\cite{GL}. \subsubsection{The case of a compact manifold} On a compact manifold of dimension $d$, as was proved by Sogge~\cite{Sogge}, $$ \| P_{\lambda,1} \|_{L^2 \to L^p} \lesssim \left\{ \begin{array}{ll} \lambda^{\sigma(p)/2} & \mbox{if $p \geq p_{ST}$} \\ \lambda^{\frac{d-1}{2} \left( \frac{1}{2} - \frac{1}{p} \right)} & \mbox{if $2 \leq p \leq p_{ST}$}, \end{array} \right. $$ where $\sigma(p)$ is as above. For any given compact manifold, this estimate is optimal for $\delta = 1$. In the case of the sphere $ \mathbb{S}^d$ (or more generally of a Zoll manifold), it does not improve if $\delta$ decreases, since the eigenvalues of the sphere Laplacian are essentially distributed like squared integers. However, for ``most" manifolds, the estimates above are expected to improve as $\delta$ decreases. It is the aim of this article to examine this question in the case of the torus. If the manifold $M$ is negatively curved, then logarithmic improvements are possible over the allowed range of $\delta$, as in Bourgain-Shao-Sogge-Yao \cite{BSSY} and Blair-Sogge~\cite{BS}. The work of Sogge-Toth-Zelditch~\cite{STZ} shows that generic manifolds also allow improvements. \subsection{Spectral projectors on tori} \subsubsection{Formulating the problem} From now on, we focus on the case of tori given by the quotient $\mathbb{R}^d / (\mathbb{Z} e_1 + \dots + \mathbb{Z} e_1)$, where $e_1,\dots,e_d$ is a basis of $\mathbb{R}^d$, with the standard metric. This is equivalent to considering the operators $$ P_{\lambda,\delta} = \chi \left( \frac{\sqrt{-Q(\nabla)} - \lambda}{\delta} \right) \qquad \mbox{on} \;\; \mathbb{T}^d = \mathbb{R}^d / \mathbb{Z}^d, $$ where $\nabla$ is the standard gradient operator, and $Q$ is a positive definite quadratic form on $\mathbb{R}^d$, with coefficients $\beta_{ij}$: $$ Q(x) = \sum_{i=1}^d \beta_{ij} x^i x^j \qquad \implies \qquad Q(\nabla) = - \sum_{i=1}^d \beta_{ij} \partial_i \partial_j. $$ Dispensing with factors of $2\pi$, which can be absorbed in $Q$, the associated Fourier multiplier has the symbol $$ \chi \left( \frac{\sqrt{Q(k)} - \lambda}{\delta} \right). $$ \subsubsection{Known results for $p = \infty$: counting lattice points} \label{subsec1} Abusing notations by writing $P_{\lambda,\delta}(z)$ for the convolution kernel giving $P_{\lambda,\delta}$, we have the formula $$ P_{\lambda,\delta}(z) = \sum_n \chi \left( \frac{\sqrt{Q(n)} - \lambda}{\delta} \right) e^{2\pi i n \cdot z}. $$ It is easy to see that $$ \| P_{\lambda,\delta} \|_{L^1 \to L^\infty} = \| P_{\lambda,\delta}(z) \|_{L^\infty_{z}} = \sum_n \chi \left( \frac{\sqrt{Q(n)} - \lambda}{\delta} \right). $$ If $\chi = \mathbf{1}_{[-1,1]}$, this can be expressed as $$ \| P_{\lambda,\delta} \|_{L^1 \to L^\infty} = N(\lambda + \delta) - N(\lambda - \delta), $$ where $N(\lambda)$ is the counting function associated to the quadratic form $Q$: namely, it denotes the number of lattice points $n \in \mathbb{Z}^d$ such that $Q(n) < \lambda^2$. To leading order, $N(\lambda)$ equals $\operatorname{Vol}(B_1) \lambda^d$, where $\operatorname{Vol}(B_1)$ is the volume of the ellipsoid $\{Q(x) < 1\}.$ We denote the error term by $P(\lambda)$, thus: $$ N(\lambda) = \operatorname{Vol}(B_1) \lambda^d + P(\lambda). $$ For a general quadratic form $Q$, it was showed by Landau~\cite{Landau15} that $P(\lambda) = O( \lambda^{d - \frac{2d}{d+1}})$. \textcolor{black} {Consequently, we have \begin{align}\label{eqn:landau_shells} \| P_{\lambda,\delta} \|_{L^1 \to L^\infty} &\ll \delta\lambda^{d-1} &\text{for }\delta&> \lambda^{-\frac{d-1}{d+1}}. \end{align}} Landau's result, and hence the range for \(\delta\) in \eqref{eqn:landau_shells}, has been improved for every dimenson \(d\). Nonetheless \eqref{eqn:landau_shells} is a useful point of comparison since our approach is in a sense a refinement of a proof of Landau's theorem (see the comments after Theorem~\ref{thmcaps}). Regarding lower bounds for \(P\), when \(Q(x)=|x|_2^2\) one can show that \(P(\lambda_i)\gg \lambda_i^{d-2}\) for some sequence \(\lambda_i \to \infty\). The present state of the art is as follows: \begin{itemize} \item If \(d=2\) then estimating $P(\lambda)$ is a variation on the celebrated Gauss circle problem. One conjectures \(P(\lambda)=O_\epsilon(\lambda^{\frac{1}{2}+\epsilon})\), and the best known result is \(O(\lambda^{\frac{131}{208}}\log^{\frac{18627}{8320}}\lambda)\), see Huxley~\cite{Huxley03}. \item If \(d=3\) then one conjectures \(P(\lambda)=O_\epsilon(\lambda^{1+\epsilon})\), see Nowak~\cite[\S\S1.1-1.2]{Nowak2014}. We have \(O(\lambda^{\frac{231}{158}})\) by Guo \cite{Guo2012}. If moreover \(Q\) has rational coefficients then \(P(\lambda)=O(\lambda^{\frac{21}{16}})\) by Chamizo-Cristob\'al-Ubis~\cite{CCU}. \item If \(d=4\) then \(P(\lambda)= O(\lambda^2 \log^{\frac{2}{3}}\lambda)\) by Walfisz~\cite{walfisz1960}. The case \(Q(x)=|x|_2^2\) shows that up the log power this is best possible. \item If \(d>4\) then $P(\lambda) = O( \lambda^{d - 2})$, see Kr\"atzel \cite{Kraetzel00}. This is best possible if \(Q\) is a multiple of a form with rational coefficients, and if not then $P(\lambda) = o( \lambda^{d - 2})$ by G\"otze~\cite{Goetze}. \end{itemize} \subsubsection{Known results on standard tori: eigenfunctions of the Laplacian} \label{subsec2} It was conjectured by Bourgain~\cite{Bourgain} that an eigenfunction $f$ of the Laplacian on the standard torus with eigenvalue $\lambda^2$ satisfies $$ \| f \|_{L^p} \lesssim_\epsilon \lambda^{\frac{d-2}{2} - \frac{d}{p} + \epsilon} \| f \|_{L^2} \qquad \mbox{for $p \geq p^*$}, \;\;\;\mbox{where} \; p^* = \frac{2d}{d-2}, $$ which can be reformulated as $$ \| P_{\lambda,\frac{1}{\lambda}} \|_{L^2 \to L^p} \lesssim_\epsilon \lambda^{\frac{d-2}{2} - \frac{d}{p} + \epsilon} \qquad \mbox{for $p \geq p^*$.} $$ Progress towards this conjecture~\cite{Bourgain2, BourgainDemeter1, BourgainDemeter2} culminated in the work of Bourgain and Demeter on $\ell^2$-decoupling~\cite{BourgainDemeter3}, where the above conjecture is proved for $d \geq 4$ and $p \geq \frac{2(d-1)}{d-3}$. \subsubsection{Known results on standard tori: uniform resolvent bounds} \label{subsec3} It was proved in Dos Santos Ferreira-Kenig-Salo~\cite{DKS} that, for general compact manifolds, each $x,y \in \mathbb{R}$, and writing $ p^* = \frac{2d}{d-2}$, we have $$ \| (\Delta + (x+iy)^2)^{-1} \|_{L^{(p^*)'} \to L^{p^*}} \lesssim 1 \qquad \mbox{if $|y| \geq 1$}. $$ In terms of spectral projectors, this is equivalent (see Cuenin~\cite{Cuenin}) to the bound $$ \| P_{\lambda,\delta} \|_{L^{(p^*)'} \to L^{p^*}} \lesssim \lambda \delta, \qquad \mbox{if $|\delta| > 1$}. $$ It was also asked whether this bound could be extended to a broader range of $y$, or equivalently a broader range of \(\delta\). Bourgain-Shao-Sogge-Yao \cite{BSSY} showed that on the sphere the range above is optimal. In the case of the standard \(d\)-dimensional torus \(\mathbb R^d/\mathbb Z^d\), they could improve earlier results of Shen~\cite{Shen}. The results of Shen and Bourgain-Shao-Sogge-Yao were then sharpened by Hickman~\cite{Hickman}, who extended the range for the standard \(d\)-dimensional torus futher to $|\delta| > \lambda^{-\frac{1}{3} -\frac{d}{3(21d^2-d-24)} + \epsilon}$. \subsubsection{Known results in dimension 2} The classical estimate of Zygmund corresponds, in our language, to a sharp result for $d=2$, $p=4$, $\delta = \lambda^{-1}$. It was showed by Bourgain-Burq-Zworski~\cite{BBZ} that it can be extended to $\delta > \lambda^{-1}$. A striking feature of the estimates in~\cite{BBZ} is that they entail no subpolynomial loss ($\epsilon$ power in the exponent), which has important consequences for control theory in particular, as explained in that paper. \subsection{Conjecture and results} Based on two specific examples, developed in Section~\ref{lbac}, we conjecture that \begin{equation} \label{conj} \boxed{ \| P_{\lambda,\delta} \|_{L^{2} \to L^p} \lesssim (\lambda \delta)^{\frac{(d-1)}{2} \left( \frac{1}{2} - \frac{1}{p} \right)} + \lambda^{\sigma(p)/2} \delta^{1/2},} \end{equation} where $\delta > \lambda^{-1}$, and $\sigma(p) = d - 1 - \frac{2d}{p}$, for any fixed torus. We show there that this bound would be optimal, and describe when each term in the conjecture dominates. \bigskip The methods developed in the present paper give improvements on the range of validity of this conjecture. The precise statemenet is Theorem~\ref{thm:main} below. \textcolor{black}{As this is a rather cumbersome formula, we choose to state some simpler results in fairly natural cases of interest, namely $p<p_{ST}$, \(\delta\) large, and $d=3$. Here we have $p_{ST} = \frac{2(d+1)}{d-1}$; in sections~\ref{subsec2} and~\ref{subsec3} we saw that $p^* = \frac{2d}{d-2}$ has some special significance, so we will also state a result in this case.} \begin{thm}[The case $p<p_{ST}$] \label{thmpST} For any positive definite quadratic form $Q$, the conjecture~\eqref{conj} is verified, up to subpolynomial losses, if $\lambda>1$, \(\delta \geq \lambda^{-1}\) and $1<p<p_{ST}$. \end{thm} Here, subpolynomial losses means that the conjecture holds true with an additional $\lambda^\epsilon$ factor on the right-hand side, where the implicit constant depends on $\epsilon$, but $\epsilon$ can be chosen arbitrarily small. \textcolor{black}{ \begin{thm}[The case of large $\delta$] \label{thmsimple} For any positive definite quadratic form $Q$, the conjecture~\eqref{conj} is verified, up to subpolynomial losses, if $\lambda>1$, $p\geq p_{ST}$ and \[ \delta > \lambda^{-\frac{(d-1)p-dp_{ST}+2}{(d+1)p-dp_{ST}-2}} .\] \end{thm} } By substituting \(p=2d/(d-2)\) and performing a brief computation we obtain: \begin{cor}[The case $p = p^*$] \label{thmp*} \textcolor{black}{Let \(p^*=\frac{2d}{d-2}\).} For any positive definite quadratic form $Q$, the conjecture~\eqref{conj} is verified, up to subpolynomial losses, for $p=p^*$, if $\lambda>1$ and \textcolor{black}{\(\delta \geq \lambda^{-\frac{1}{2 d - 1}}\)}.\end{cor} \textcolor{black}{In Theorems~\ref{thmsimple} and Corollary~\ref{thmp*} we have aimed to provide simple statements, which are consequently somewhat weaker than Theorem~\ref{thm:main} below. For any particular \(d\) these last results can be improved by a short computation. We present the following as a representative example.} \begin{thm}[The case $d=3$] \label{thmd3} For any positive definite quadratic form $Q$, the conjecture~\eqref{conj} is verified, up to subpolynomial losses, if $d=3$, whenever $\lambda>1$, \textcolor{black}{\(\delta \geq \min\{\lambda^{-\frac{3p-8}{5p-8}}, \lambda^{-\frac{8-p}{5p-16}}\}\) and also \(\delta \geq\lambda^{-1/2}\)}. \end{thm} \textcolor{black}{In the proofs of the results above, the value \(\delta = \lambda^{-\frac{d-1}{d+1}}\) will emerge as playing a special role. In particular, to prove our conjecture in even a single case with \(p>p_{ST}\) and \(\delta \ll \lambda^{-\frac{d-1}{d+1}}\) needs a different approach. This threshold also appears in the classical result \eqref{eqn:landau_shells}, and more generally when counting lattice points in a \(\delta\)-thick shell around a manifold with curvature \(\sim \lambda^{-1}\) using for example Poisson summation. Substituting \(d=3\) into the last theorem does however yield the full range \(\delta > \lambda^{-\frac{d-1}{d+1}}\), as well as the the full range \(p>2\), in the following setting:} \begin{cor}If \(d=3\), then for any positive definite quadratic form $Q$ the conjecture holds for all \(\delta \geq \lambda^{-\frac{d-1}{d+1}} = \lambda^{-1/2}\) if \(p\leq p_{ST}+\frac{4}{7}\) \textcolor{black}{or \(p\geq p_{ST}+4\)}, and it holds for all \(p>2\) if $\delta> \lambda^{-2/5}$. \end{cor} The proof of the above results will combine a number theoretical argument, which allows one to count the number of caps in a spherical shell which contain many lattice points, with a harmonic analysis approach, relying in particular on the $\ell^2$ decoupling theorem of Bourgain and Demeter. \bigskip In order to understand better the statement of these theorems, it is helpful to spell out what they imply for each of the classical problems presented in sections~\ref{subsec1},~\ref{subsec2},~\ref{subsec3}. \begin{itemize} \item \textcolor{black}{For the problem of counting points in thin spherical shells (Subsection~\ref{subsec1}) we recover the bound \eqref{eqn:landau_shells} of Landau~\cite{Landau15}, see also the comments after Theorem~\ref{thmcaps}.} \item \textcolor{black}{The problem of bounding $L^p$ norms of eigenfunctions was prevously considered for rational tori, that is \(\mathbb R^d/A\mathbb Z^d\) where \(A\in \operatorname{GL}_d(\mathbb Q)\). Our results do not improve the bounds of Bourgain-Demeter~\cite{BourgainDemeter3} in this case. For generic tori, eigenfunction bounds are trivial; the natural analogue of bounding the $L^p$ norms of eigenfunctions is to bound the operator norm of $P_{\lambda,\frac{1}{\lambda}}$, and this question does not appear to have been considered before. For any torus, that is any \(\mathbb R^d/B\mathbb Z^d\) with \(B\in \operatorname{GL}_d(\mathbb R)\), we obtain from Theorem~\ref{thm:main} below the bound \begin{align*} \| P_{\lambda,\frac{1}{\lambda}} \|_{L^2 \to L^p} &\lesssim_\epsilon \lambda^\epsilon (\lambda^{\frac{d}{d+1}})^{\frac{1}{2}(1-\frac{2}{p})+\frac{d}{2}(1- \frac{p_{ST}}{p})-\sqrt{(1 - \frac{2}{p}) (1- \frac{p_{ST}}{p})}} & (p&\geq p_{ST}). \end{align*} } \item For the problem of proving uniform resolvent bounds (Subsection~\ref{subsec3}), it proves the desired estimate up to a subpolynomial loss $$\| P_{\lambda,\delta} \|_{L^{(p^*)'} \to L^{p^*}} \lesssim_\epsilon \lambda^{1+\epsilon} \delta, \qquad $$ if \( \delta \geq \lambda^{-1+\frac{2 d}{5 d - 4}}\), or if \(d=3\) and \(\delta > \lambda^{-1/2}\), improving over \textcolor{black}{Hickman's~\cite{Hickman} result that} $\lambda^{-\frac{1}{3} -\frac{d}{3(21d^2-d-24)}}$. \end{itemize} \subsection{Acknowledgments} The authors are grateful to the anonymous referee for a careful reading of their manuscript, and pointing out an error in an earlier version; they are also thankful to Yu Deng for insightful discussions at an early stage of this project. While working on this project SLRM was supported by DFG project number 255083470, and by a Leverhulme Early Career Fellowship. PG was supported by the NSF grant DMS-1501019, by the Simons collaborative grant on weak turbulence, and by the Center for Stability, Instability and Turbulence (NYUAD). \section{Notation}\label{sec:notation} Throughout, \textcolor{black}{$p_{ST} = \frac{2(d+1)}{d-1},$ and $\sigma(p) = d - 1 - \frac{2d}{p}$ will be as in section~\ref{sec:euclidean-space}, and $p^* = \frac{2d}{d-2}$ as in section~\ref{subsec2}.} We adopt the following normalizations for the Fourier series on $\mathbb{T}^d$ and Fourier transform on $\mathbb{R}^d$, respectively: \begin{align*} & f(x) = \sum_{k \in \mathbb{Z}^d} \widehat{f}_k e^{2\pi i k \cdot x}, \qquad \qquad \widehat{f}_k = \int_{\mathbb{T}^d} f(x) e^{-2\pi i k \cdot x} \,dx \\ & f(x) = \int_{\mathbb{R}^d} \widehat{f}(\xi) e^{2\pi ix \cdot \xi} \,dx, \qquad \qquad \widehat{f}(\xi) = \int_{\mathbb{R}^d} f(x) e^{-2\pi ix \cdot \xi} \,dx \end{align*} The Poisson summation formula is then given by $$ \sum_{n \in \mathbb{Z}^d} f(n) = \sum_{k \in \mathbb{Z}^d} \widehat{f}(k). $$ We write \((\vecs{1}{v}|\cdots|\vecs{k}{v})\) for the matrix with columns \(\vecs{i}{v}\). Given two quantities $A$ and $B$, we write $A \lesssim B$ or equivalently \(A = O(B)\) if there exists a constant $C$ such that $A \leq CB$, and $A \lesssim_{a,b,c} B$ if the constant $C$ is allowed to depend on $a,b,c$. We always allow \(C\) to depend on the dimension \(d\). In the following, it will often be the case that the implicit constant will depend on $\beta$, and on an arbitrarily small power of $\lambda$: $A \lesssim_{\beta,\epsilon} \lambda^\epsilon B$. When this is clear from the context, we simply write $A \lesssim \lambda^\epsilon B$. When we are assuming that the implicit constant is sufficiently small, we will write \(A\ll B\). If both \(A\lesssim B\) and \(B\lesssim A\) then we write \(A\sim B\). \section{Lower bounds and conjecture} \label{lbac} \subsection{The discrete Knapp example} \begin{lem} \label{llb1} For any $n \in \mathbb{N}$, there exists $\lambda \sim |n|$ such that if $\delta \in (0,1)$, $$ \| P_{\lambda,\delta} \|_{L^{2} \to L^p} \gtrsim (1 + \lambda \delta)^{\frac{(d-1)}{2} \left( \frac{1}{2} - \frac{1}{p} \right)}. $$ \end{lem} \begin{proof} Consider the ellipse $\{ \xi \in \mathbb{R}^d, \; Q(\xi) = 1\}$. Its normal vector is colinear to $e_d = (0,\dots,0,1)$ at the point $\xi_0$. We now dilate this ellipse by a factor $\lambda$ such that $\lambda \xi^d_0 = n \in \mathbb{N}$, smear it to a thickness $\delta$, and observe that, around the point $\lambda \xi_0$, it contains many lattice points of $\mathbb{Z}^d$. More precisely, the cuboid $C \subset \mathbb{R}^d$ defined by $$ C = \{ |\xi^i - \lambda \xi^i_0| < c \sqrt{\lambda \delta} \;\; \mbox{for $i = 1,\dots,d-1$} \quad \mbox{and} \quad |\xi^d - \lambda \xi^d_0| < c\delta \} $$ (where the constant $c$ is chosen to be sufficiently small) is such that $$ C \subset \left\{ \xi \in \mathbb{R}^d, \,\lambda-\frac{\delta}{2} < \sqrt{Q(\xi)} < \lambda + \frac{\delta}{2} \right\}. $$ Furthermore, for $\lambda\in\mathbb{Z}$, $C$ will contain $\sim (1 + \lambda \delta)^{\frac{d-1}{2}}$ points in $\mathbb{Z}^d$. Writing $\xi = (\xi',\xi^d)$ and $x=(x',x^d)$, let $$ f(x) = e^{2\pi i\lambda \xi_0^d x^d}\sum_{\xi' \in \mathbb{Z}^{d-1}} \phi \left( \frac{\xi'- \lambda \xi_0'}{\sqrt{\lambda \delta}} \right) e^{2\pi i \xi' \cdot x'}, $$ where $\phi \in \mathcal{C}_0^\infty (\mathbb{R}^{d-1})$ is such that $\widehat{\phi} \geq 0$ and $\operatorname{Supp} \widehat{\phi} \subset C$. By the Poisson summation formula, $f$ can be written $$ f(x) = e^{2\pi i\lambda x^d \xi_0^d} (\lambda \delta)^{\frac{d-1}{2}} \sum_{n \in \mathbb{Z}^{d-1}} \widehat{\phi}( \sqrt{\lambda \delta}(x'-n)) e^{2 \pi i \lambda \xi_0' \cdot (x'-n)}, $$ which implies $$ \| f \|_{L^p} \sim (1 + \lambda \delta)^{\frac{d-1}{2} \left( 1 - \frac{1}{p} \right)}. $$ Since $P_{\lambda,\delta} f = f$, we find that $$ \| P_{\lambda,\delta} \|_{L^{2} \to L^p} \geq \frac{\| f \|_{L^p}}{\| f \|_{L^{2}}} \sim (1 + \lambda \delta)^{\frac{(d-1)}{2} \left( \frac{1}{2} - \frac{1}{p} \right)}. $$ \end{proof} \subsection{The radial example} \begin{lem} \label{llb2} For any $n \in \mathbb{N}$ and $\delta \in (0,1)$, there exists $\lambda$ such that $|n-\lambda| \lesssim 1$ and $$ \| P_{\lambda,\delta} \|_{L^2 \to L^p} \gtrsim \lambda^{\sigma(p)/2} \sqrt \delta. $$ \end{lem} \begin{proof} For any $n \in \mathbb{N}$, there exists $\lambda$ with $|n-\lambda| \lesssim 1$, and such that the corona $$ \mathcal{C} = \{ \lambda -\frac{\delta}{2} < Q(x) < \lambda + \frac{\delta}{2} \} $$ contains $N \gtrsim \lambda^{d-1} \delta$ points in $\mathbb{Z}^d$. Define $$ f(x) = \sum_{\xi \in \mathcal{C} \cap \mathbb{Z}^{d-1}} e^{2\pi i \xi \cdot x}. $$ It is clear that $$ \| f \|_{L^\infty} = N \qquad \mbox{while} \qquad \| f \|_{L^2} = \sqrt{N}. $$ By Bernstein's inequality, for $p\geq 2$, $$ \| f \|_{L^p} \gtrsim \|f\|_{L^{\infty}} \lambda^{-\frac{d}{p}} \sim \lambda^{-\frac{d}{p}} N. $$ Therefore, $$ \| P_{\lambda,\delta} \|_{L^2 \to L^p} \geq \frac{\|f\|_{L^p}}{\|f\|_{L^2}} \gtrsim \lambda^{\sigma(p)/2} \sqrt \delta. $$ \end{proof} \subsection{The conjecture} Based on lemmas~\ref{llb1} and~\ref{llb2}, it is reasonable to conjecture that $$ \boxed{\| P_{\lambda,\delta} \|_{L^{2} \to L^p} \lesssim (\lambda \delta)^{\frac{(d-1)}{2} \left( \frac{1}{2} - \frac{1}{p} \right)} + \lambda^{\sigma(p)/2} \delta^{1/2}.} $$ The next question is: how small can $\delta$ be taken? In full generality, the limitation is $$ \delta \geq \frac{1}{\lambda}, $$ as this is best possible for rational tori; this will be the range we consider here. We now describe the different regimes involved in the above conjecture. \bigskip \noindent \underline{If $d=2$}, the conjecture can be formulated as \begin{itemize} \item $\| P_{\lambda,\delta} \|_{L^{2} \to L^p} \lesssim (\lambda \delta)^{\frac{1}{4} - \frac{1}{2p} }$ if $\left\{ \begin{array}{l} 2 \leq p \leq 6\; \mbox{and} \; \delta > \lambda^{-1} \\ \mbox{or} \; p \geq 6 \; \mbox{and} \; \lambda^{-1} < \delta < \lambda^{\frac{6-p}{2+p}} \end{array} \right.$. \item $\| P_{\lambda,\delta} \|_{L^{2} \to L^p} \lesssim \lambda^{\frac 1 2 - \frac{2}{p}} \delta$ if $p \geq 6$ and $\delta > \lambda^{\frac{6-p}{2+p}}$. \end{itemize} \bigskip \noindent \underline{If $d\geq 3$}, let $$ p_{ST} = \frac{2(d+1)}{d-1}, \qquad p^* = \frac{2d}{d-2}, \qquad \widetilde p = \frac{2(d-1)}{d-3} $$ and define $$ e(p) = \frac{d+1}{d-1} \cdot \frac{\frac{1}{p} - \frac{1}{p_{ST}}}{\frac{1}{p} - \frac{1}{\widetilde{p}}} $$ We have $$ p_{ST} < p^* < \widetilde{p}. $$ Keeping in mind that $\delta > \lambda^{-1}$, the above conjecture becomes \begin{itemize} \item $\| P_{\lambda,\delta} \|_{L^{2} \to L^p} \lesssim (\lambda \delta)^{\frac{(d-1)}{2} \left( \frac{1}{2} - \frac{1}{p} \right)}$ if $\left\{ \begin{array}{l} 2 \leq p \leq p_{ST} \\ \mbox{or} \; p_{ST} \leq p \leq p^* \; \mbox{and} \; \delta < \lambda^{e(p)} \end{array} \right.$ \item $\| P_{\lambda,\delta} \|_{L^{2} \to L^p} \lesssim \lambda^{\sigma(p)/2} \delta^{1/2}$ if $\displaystyle \left\{ \begin{array}{l} p_{ST} \leq p \leq p^* \; \mbox{and} \; \delta > \lambda^{e(p)} \\ \mbox{or} \; p \geq p^* \end{array} \right.$ \end{itemize} \section{Caps containing many points} We split the spherical shell $$ S_{\lambda,\delta} = \{ x \in \mathbb{R}^d, \, \left| \sqrt{Q(x)}-\lambda \right| < \delta \} $$ into a collection $\mathcal{C}$ of almost disjoint caps $\theta$: $$ S_{\lambda,\delta} = \bigcup_{\theta \in \mathcal{C}} \theta, $$ where each cap is of the form $$ \theta = \{ x \in \mathbb{R}^d, \, | x - x_\theta| < \sqrt{\lambda \delta} \} \cap S_{\lambda,\delta} \qquad \mbox{for some $x_\theta \in S_{\lambda,\delta}$}. $$ Each cap fits into a rectangular box with dimensions $\sim \delta \times \sqrt{\lambda \delta} \times \dots \times \sqrt{\lambda \delta}$. We call \(\vec{n}_\theta = \frac{x_\theta}{|\vec{x}_\theta|_2}\) the \emph{normal vector} to \(\theta\); observe that as \(\theta\) varies over caps, the normal vector \(\vec{n}_\theta\) varies over a \(\sqrt{\delta/\lambda}\)-spaced set. Denote $N_\theta$ for the number of points in $\mathbb{Z}^d \cap \theta$. On the one hand, it is clear that $N_\theta \lesssim (\sqrt{\lambda \delta})^{d-1}$. On the other hand, one expects that the average cap will contain a number of points comparable to its volume, in other words $N_\theta \sim (\sqrt{\lambda \delta})^{d-1} \delta$ (provided this quantity is $>1$, which occurs if $\delta > \lambda^{-\frac{d-1}{d+1}}$). This leads naturally to defining the following sets, which gather caps containing comparable numbers of points \begin{align*} & \mathcal{C}_0 = \{ \theta \in \mathcal{C}, \, N_\theta < (\sqrt{\lambda \delta})^{d-1} \delta \} \\ & \mathcal{C}_j = \{ \theta \in \mathcal{C}, \, (\sqrt{\lambda \delta})^{d-1} \delta 2^{j-1}<N_\theta\leq (\sqrt{\lambda \delta})^{d-1} \delta 2^{j}\} , \qquad \mbox{for $1\leq 2^j \lesssim \delta^{-1}$ }. \end{align*} \begin{thm} \label{thmcaps} There is a constant \(K>0\), depending only on \(d\), as follows. Let \( k \in \{ 1, \dotsc, d-1 \} \). If \( 2^{ j } >K \) and $(\sqrt{ \delta \lambda })^{ k } \delta 2^{ j } >K$ then \[ \# \mathcal{ C }_{ j } \lesssim ( 2^{ j / k } \delta )^{ - d }. \] \end{thm} \begin{rem} \textcolor{black}{There is an integer $k \in \{ 1, \dotsc, d-1 \}$ satisfying $(\sqrt{ \delta \lambda })^{ k } \delta 2^{ j } >K$ whenever $\delta > \lambda^{-\frac{d-1}{d+1}}$, which in practise we will assume whenever we apply the theorem above.} \end{rem} \textcolor{black}{ By summing over the caps in each \( \mathcal{C}_j \), we find that \( \#(S_{\lambda,\delta}\cap \mathbb Z^d) \ll \delta \lambda^{d-1} +\sqrt{\lambda/\delta}^{d-1} \), and in particular \( \#(S_{\lambda,\delta}\cap \mathbb Z^d) \ll \delta \lambda^{d-1}\) for \(\delta > \lambda^{-\frac{d-1}{d+1}}\). This recovers the classical result \eqref{eqn:landau_shells}. As explained in section~\ref{subsec1}, the range \(\delta > \lambda^{-\frac{d-1}{d+1}}\) has now been improved in every dimension. Thus Theorem~\ref{thmcaps} is certainly suboptimal if \(\delta \ll \lambda^{-\frac{d-1}{d+1}}\).} To further gauge the strength of this theorem we can compare it to the trivial bounds \begin{align} \# \mathcal{C}_0 &\lesssim ( \lambda \delta^{-1})^{\frac{d-1}{2}},\label{eqn:C-zero} \\ \# \mathcal{C}_j &=0\qquad( 2^j \gg \delta^{-1}). \label{eqn:C-max} \end{align} If $\delta = \lambda^{-\frac{d-1}{d+1}}$ then the theorem interpolates between these bounds. For it reduces, on the one hand, to \(\#\mathcal{C}_j \lesssim ( \lambda \delta^{-1})^{\frac{d-1}{2}}\) for \(2^j\sim 1\), and on the other, to $\# \mathcal{C}_j \lesssim (\delta 2^j)^{-d}$ for $\delta 2^j \gg (\sqrt{\lambda\delta})^{-1}$. If \(\delta > \lambda^{-\frac{d-1}{d+1}}\) then the theorem shows that for some constant \(C_d>0\), all but \(\frac{C_d}{\delta (\sqrt{\lambda\delta})^{d-1}}\%\) of the caps \(\theta\) satisfy \(N_\theta \lesssim\delta (\sqrt{\lambda\delta})^{d-1}\), and interpolates between this bound and \eqref{eqn:C-max}. Inspecting the proof, we could strengthen the former bound: if \(\delta > \lambda^{-\frac{d-1}{d+1}}\) then all but \(\frac{C_d}{\delta (\sqrt{\lambda\delta})^{d-1}}\%\) of the caps \(\theta\) contain a fundamental region for \(\mathbb Z^d\). If \(\delta < \lambda^{-\frac{d-1}{d+1}}\) then the theorem shows that all but \((C'_d\sqrt{\lambda}(\sqrt{\delta})^{\frac{d+1}{d-1}})\%\) of the caps \(\theta\) satisfy \(N_\theta \lesssim 1\), and interpolates between this bound and \eqref{eqn:C-zero}. Again by inspecting the proof we could strengthen first part: if \(\delta > \lambda^{-\frac{d-1}{d+1}}\) then all but \((C'_d\sqrt{\lambda}(\sqrt{\delta})^{\frac{d+1}{d-1}})\%\) of the caps \(\theta\) satisfy \(N_\theta \leq 1\). In this regime most caps should have \(N_\theta=0\). \begin{proof}[Proof of Theorem~\ref{thmcaps}] Let \(\theta\) be any cap. Let \(R_\theta\) be a rectangular box, centred at the origin, containing \(\theta-\theta\) and having dimensions $\sim \delta \times \sqrt{\lambda \delta} \times \dots \times \sqrt{\lambda \delta}$. Define a norm \(|\,\cdot\,|_\theta\) on \(\mathbb R^d\) by \[ |\vec{x}|_\theta = \inf\{r>0 : \vec{x}\in r R_\theta\}, \] so that \(R_\theta\) is the unit ball in this norm, with \begin{equation}\label{eqn:theta-norm} |\,\cdot\,|_2\lesssim \sqrt{\lambda\delta}|\,\cdot\,|_\theta. \end{equation} Because \(R_\theta\) is contained in a slab of the form \(\{\vec{x}\in\mathbb R^d:\vec{n}_\theta\cdot \vec{x}\lesssim \delta\}\), we also have \begin{equation}\label{eqn:theta-norm-normal} \vec{n}_\theta \cdot \vec{x} \lesssim \delta |\vec{x}|_\theta. \end{equation} Morally, the idea of the proof is to fix the lattice generated by \(R_\theta\cap \mathbb Z^d\), and count the number of caps with a given lattice. Carrying out this programme in a literal fashion seems possible but technically complex. We take advantage of a trick: we will construct a small integer vector \(\vec{v}\) which is orthogonal to all the integer vectors in \(R_\theta\cap \mathbb Z^d\) and approximately perpendicular to \(\vec{n}_\theta\), and it is this vector, rather than the lattice itself, which we will fix. The outline of the proof is as follows: after some further definitions we set out some basic results from the geometry of numbers in Step~1 below; we then construct \(\vec{v}\) in Step~2 and complete the proof in Step~3. Define \(r_\theta\) to be the dimension of the span (over \(\mathbb R\), say) of the vectors in \(R_\theta\cap \mathbb Z^d\). The hypotheses of the theorem imply that we cannot have \(r_\theta =d\), since then \(N_\theta\lesssim \delta (\sqrt{\lambda/\delta})^{d-1}\). \subsubsection*{Step 1} We show that there is a basis \(\vecs{1}{x},\dotsc,\vecs{d}{x}\) of \(\mathbb Z^d\) with \begin{equation}\label{eqn:step1} \delta (\sqrt{\delta\lambda})^{d-1} \sim \prod_{i=1}^{d} |\vecs{i}{x}|_\theta^{-1}, \qquad \#(\mathbb Z^d\cap R_\theta) \sim \prod_{i=1}^{r_\theta} |\vecs{i}{x}|_\theta^{-1}. \end{equation} This is more or less a standard result from the geometry of numbers. In this step only, let \(A\) be the matrix with \(AR_\theta=[-1,1]^d\), and for \(1\leq i \leq d\) let \(M_i\) be minimal such that there are \(i\) linearly independent vectors \(\vec{x}\in\mathbb Z^d\) with \(|\vec{x}|_\theta\leq M_i\), or equivalently there are \(i\) linearly independent vectors \(\vec{y}\in A\mathbb Z^d\) with \(|A\vec{y}|_\infty \leq M_i\). In particular \[ M_{r_\theta}\leq 1<M_{r_\theta+1}. \] By Theorem~V in \S VIII.4.3 of Cassels~\cite{casselsIntroduction} we have \begin{equation}\label{eqn:minkowski} \prod_{i=1}^d M_i \sim \det A \sim \frac{1}{\delta(\sqrt{\delta\lambda})^{d-1}}. \end{equation} Also by Lemma~2 in \S VIII.1.2 of Cassels there is a basis \(\vecs{1}{y},\dotsc,\vecs{d}{y}\) of \(A\mathbb Z^d\) such that \begin{equation}\label{eqn:nice-basis} M_{i}\leq|\vec{y}|_\infty < M_{i+1},\vec{y}\in A\mathbb Z^d \implies \vec{y} \in \mathbb Z \vecs{1}{y}+\dotsb+\mathbb Z \vecs{i}{y}. \end{equation} Let \(\vecs{i}{z}\) be the dual basis, so that \[ \vec{y} = (\vec{y}\cdot\vecs{1}{z} )\vecs{1}{y}+\dotsb +(\vec{y}\cdot\vecs{d}{z} )\vecs{d}{y} \] for all \(\vec{y}\in A\mathbb Z^d\). Note that we must have \( M_i\leq |\vecs{i}{y}|_\infty\) from the definition of \(M_i\). By the Corollary to Theorem VIII in \S VIII.5.2 of Cassels, we may choose the \(\vecs{i}{y}\) such that \begin{equation}\label{eqn:minkowski-basis} M_i\leq |\vecs{i}{y}|_\infty \leq \max\{1,i/2\} M_i, \qquad |\vecs{i}{z}| \leq (\tfrac{1}{2})^{n-1}(n!)^2/ |\vecs{i}{y}|_\infty. \end{equation} We now let \(\vecs{i}{x}=A^{-1}\vecs{i}{y}\) be our basis of \(\mathbb Z^d\). We then have \begin{equation}\label{eqn:basis-norms} |\vecs{i}{x}|_\theta =|\vecs{i}{y}|_\infty\sim M_i^{-1}, \end{equation} by \eqref{eqn:minkowski-basis}. Now \eqref{eqn:basis-norms} and \eqref{eqn:minkowski} together imply the first part of \eqref{eqn:step1}. It also follows from \eqref{eqn:minkowski-basis} that for some \(c>0\) depending only on \(d\) we have \begin{align*} |c_i|\leq c M_i^{-1}\;(1\leq i\leq r_\theta) &\implies |c_1\vecs{1}{x}+\dotsb+c_{r_\theta} \vecs{r_\theta}{x}|\leq 1, \intertext{and combining \eqref{eqn:nice-basis} with \eqref{eqn:minkowski-basis} shows that there is a constant \(C>0\) depending only on \(d\) such that} |c_1\vecs{1}{x}+\dotsb+c_d \vecs{d}{x}|\leq 1 &\implies |c_i|\leq CM_i^{-1}\;(1\leq i\leq r_\theta),\; c_i=0\;(i>r_\theta ). \end{align*} The last two displays yield \begin{gather*} \#(\mathbb Z^d\cap R_\theta) = \#([-1,1]^d\cap A\mathbb Z^d) \sim \frac{ 1}{\prod_{i=1}^{r_\theta} M_i} \sim\frac{ 1}{\prod_{i=1}^{r_\theta} |\vecs{i}{y}|_\infty}, \end{gather*} and the last part of \eqref{eqn:step1} follows by \eqref{eqn:basis-norms}. \subsubsection*{Step 2} We claim that if \(1\leq r_\theta<d\) and \(\vec{n}_\theta\) is the normal vector to \(\theta\) then there is \(\vec{v}\in\mathbb Z^d\setminus\{\vec{0}\}\) such that \[ |\vec{v}|_2 \lesssim \delta^{-1} \left( \frac{ \delta(\sqrt{\lambda\delta})^{d-1} }{ \#(\mathbb Z^d\cap R_\theta) } \right)^{ \frac{1}{d-r_\theta}, } \qquad \left| \vec{n}_\theta- \vec{v}/|\vec{v}|_2 \right| \lesssim (\sqrt{\lambda\delta})^{-1} |\vec{v}|_2 ^{-1} \left( \frac{ \delta(\sqrt{\lambda\delta})^{d-1} }{ \#(\mathbb Z^d\cap R_\theta) } \right)^{ \frac{1}{d-r_\theta} }. \] For the proof, let \(\vecs{i}{x}\) be as in Step~1. Recall the notation \((\vecs{1}{v}|\cdots|\vecs{k}{v})\) for the matrix with columns \(\vecs{i}{v}\). We let \[ \vec{v} = \vecs{1}{x}\wedge\dotsb\wedge \vecs{d-1}{x}. \] By \eqref{eqn:theta-norm-normal} we have \begin{equation}\label{eqn:dirn-of-normal-I} \bigg\lvert \vec{n}_\theta^T\left(\frac{\vecs{1}{x}}{|\vecs{1}{x}|_\theta} \,\middle|\, \cdots \,\middle|\, \frac{\vecs{d-1}{x}}{|\vecs{d-1}{x}|_\theta}\right) \bigg\rvert_\infty \lesssim \delta. \end{equation} Now every \(\vecs{i}{x}\) satisfies \( \frac{\lvert \vecs{1}{x}\rvert_\infty }{|\vecs{1}{x}|_\theta}\lesssim \sqrt{\lambda\delta}\) by \eqref{eqn:theta-norm}. Therefore \begin{equation}\label{eqn:dirn-of-normal-II} \left(\frac{\vecs{1}{x}}{|\vecs{1}{x}|_\theta} \,\middle|\, \cdots \,\middle|\, \frac{\vecs{d-1}{x}}{|\vecs{d-1}{x}|_\theta} \right) = U \begin{pmatrix} \operatorname{diag}(d_1,\dotsc,d_{d-1} \\O_{1\times (d-1) \end{pmatrix} V, \end{equation} where \(O\) is a matrix of zeroes, \(U,V\) are orthogonal, and \(d_i\lesssim \sqrt{\lambda\delta}\). It follows that \begin{equation}\label{eqn:size-of-v} |\vec{v}|_2 \sim \prod_{i=1}^{d-1} d_i |\vecs{i}{x}|_\theta \lesssim \delta^{-1}|\vecs{d}{x}|^{-1}_\theta. \end{equation} Now \eqref{eqn:dirn-of-normal-I} and \eqref{eqn:dirn-of-normal-II} together yield \[ \bigg\lvert\vec{n}_\theta^T U \left(\begin{smallmatrix} 1\vspace{-.5em}\\&\ddots\\&&1\\&&&0 \end{smallmatrix}\right)\bigg\rvert_\infty \lesssim \delta (\min_i |d_i|)^{-1}. \] For any vector with \(|\vec{m}|_2= 1\) we have \[ |m_d|-1 \lesssim \bigg\lvert\vec{m}^T \left(\begin{smallmatrix} 1\vspace{-.5em}\\&\ddots\\&&1\\&&&0 \end{smallmatrix}\right)\bigg\rvert_2^2, \] and so \[ \min\{ |\vec{m}-\vecs{d}{e}|_2, |\vec{m}+\vecs{d}{e}|_2 \} \lesssim\bigg\lvert\vec{m}^T \left(\begin{smallmatrix} 1\vspace{-.5em}\\&\ddots\\&&1\\&&&0 \end{smallmatrix}\right)\bigg\rvert_\infty. \] It follows that for some choice of sign, \[ |\vec{n}_\theta\pm U\vecs{d}{e}|_2 = |U^T\vec{n}_\theta\pm \vecs{d}{e}|_2 \lesssim \delta (\min_i |d_i|)^{-1}. \] By our definititon \(\vec{v}= \vecs{1}{x}\wedge\dotsb\wedge \vecs{d-1}{x}\) we have \begin{gather*} \vec{v}^T \left(\frac{\vecs{1}{x}}{|\vecs{1}{x}|_\theta} \,\middle|\, \cdots \,\middle|\, \frac{\vecs{d-1}{x}}{|\vecs{d-1}{x}|_\theta} \right) =(0,\dotsc,0), \end{gather*} so that \(U\vecs{d}{e}=\vec{v}/|\vec{v}|_2\). It follows that, possibly after replacing \(\vec{v}\) with \(-\vec{v}\) if necessary, we have \begin{equation} \label{eqn:approx_by_v} \left| \vec{n}_\theta- \vec{v}/|\vec{v}|_2 \right| \lesssim \frac{ \delta(\sqrt{\lambda\delta})^{d-2} \prod_{i=1}^{d-1} |\vecs{i}{x}|_\theta }{ |\vec{v}|_2 } \sim \frac{ 1 }{ \sqrt{\lambda\delta} |\vecs{d}{x}|_\theta |\vec{v}|_2 }, \end{equation} where the last part follows by Step~1. It remains to observe that, again by Step~1, we have \[ |\vecs{d}{x}|_\theta^{d-r_\theta} \geq \prod_{i=r_\theta+1}^{d} |\vecs{i}{x}|_\theta \sim \frac{\#(\mathbb Z^d\cap R_\theta)}{\delta (\sqrt{\delta\lambda})^{d-1}}, \] \textcolor{black}{and the result follows from this bound together with \eqref{eqn:size-of-v} and \eqref{eqn:approx_by_v}.} \subsubsection*{Step 3} By Step~2, the number of \(\theta \in \mathcal{C}\) with \(r_\theta=r\) and \(\#(\mathbb Z^d\cap R_\theta) \sim 2^J\delta (\sqrt{\delta\lambda})^{d-1}\) is \[ \lesssim \sum_{ \substack{\vec{v}\in \mathbb Z^d\setminus\{\vec{0}\}\\ |\vec{v}|_2 \leq \delta^{-1} 2^{ {-J/(d-r)} }}} \left(1+ \delta^{-1} |\vec{v}|_2 ^{-1} 2^{-J/(d-r) } \right)^{d-1} \lesssim \left( \delta 2^{J/(d-r)} \right)^{-d}. \] Let \(2^j\) be as in the theorem and suppose \(\theta\in \mathcal{C}_j\). Then \[\#(\mathbb Z^d\cap R_\theta)\geq N_\theta >2^{j-1}\delta (\sqrt{\lambda\delta})^{d-1},\] and it follows that \(\#(\mathbb Z^d\cap R_\theta)>{\color{black}{\frac{1}{2}}}K(\sqrt{\delta\lambda})^{d-k-1}\). \textcolor{black}{Recall \(R_\theta\cap \mathbb Z^d\) is contained in a box of size $\sim \delta \times \sqrt{\lambda \delta} \times \dots \times \sqrt{\lambda \delta}$, intersected with a linear space of dimension \(r_\theta<d\), and so \begin{equation*} \label{eqn:triv_upper_bound} \#(\mathbb Z^d\cap R_\theta) \ll \sqrt{\delta\lambda}^{r_\theta}. \end{equation*} Hence \(r_\theta\geq d-k\) since \(K\) may be taken arbitrarily large. So the number of possible \(\theta \) is \(\lesssim (\delta 2^{j/k})^{-d}\).} \end{proof} \section{Application of the $\ell^2$ decoupling theorem} We decompose $P_{\lambda,\delta}$ into projectors $P_{\lambda,\delta}^j$ which are supported on the union of caps in $\mathcal{C}_j$. To be more specific, we choose a partition of unity $(\chi_\theta)$ adapted to the caps defined in the previous section: $$ \operatorname{Supp} \chi_\theta \subset \theta \qquad \mbox{and} \qquad \sum_\theta \chi_\theta = 1 \quad \mbox{on $S_{\lambda, \delta}$}, $$ and let \begin{align*} & \chi_j = \sum_{\theta \in \mathcal{C}_j} \chi_\theta \\ & P_{\lambda,\delta}^j = P_{\lambda,\delta}\, \chi_j(D) \end{align*} (where $\chi_j(D)$ is the Fourier multiplier with symbol $\chi_j(k)$). Using the $\ell^2$ decoupling theorem of Bourgain and Demeter~\cite{BourgainDemeter3}, we can estimate the operator norm of $P_{\lambda,\delta}^j$ from $L^2$ to $L^{p}$: \begin{prop} \label{propj} For any $Q$, for any $\epsilon>0$, for $p \geq p_{ST}$, and for $\delta > \lambda^{-1}$, $$ \| P_{\lambda,\delta}^j \|_{L^{2} \to L^{p}} \lesssim_\epsilon \lambda^{\frac{\sigma(p)}{2} +\epsilon} \delta^{1/2} 2^{j \left( \frac{1}{2} - \frac{1}{p} \right)}. $$ \end{prop} \begin{proof} For simplicity in the notation, we only consider the case $Q = \operatorname{Id}$. Let $a_k$ be an arbitrary sequence in $\ell^2(\mathbb{Z}^d)$, or in other words the Fourier series associated to an arbitrary function in $L^2(\mathbb{T}^d)$. Changing variables to $X = \lambda x$ and $K = k/ \lambda$, and taking advantage of the periodicity of Fourier series, we get \begin{align*} & \left\| \sum_{k \in \mathbb{Z}^d} \chi_j(k) \chi\left( \frac{|k|-\lambda}{\delta} \right) a_k e^{2\pi i k\cdot x} \right\|_{L^p(\mathbb{T}^d)} \\ & \qquad \qquad \lesssim \left( \frac{\delta}{\lambda} \right)^{d/p} \left\| \phi \left( \frac{\delta X}{\lambda} \right) \sum_{K \in \mathbb{Z}^d/\lambda} \chi_j(\lambda K)\chi\left( \frac{|K|-1}{(\delta/\lambda)} \right) a_{\lambda K} e^{2\pi i K \cdot X} \right\|_{L^{p}(\mathbb{R}^d)}, \end{align*} where the cutoff function $\phi$ can be chosen to have compactly supported Fourier transform. As a result, the Fourier transform of the function on the right-hand side is supported on a $\delta/\lambda$-neighborhood of $\mathbb{S}^{d-1}$. Using the $\ell^2$ decoupling theorem of Bourgain and Demeter \footnote{The theorem, as stated in that paper, does not immediately apply to our setup. The following procedure can be applied: first, restrict to a coordinate patch on the sphere. Second, split our caps $\theta$ into a finite number of subcollections $\mathcal{S}_j$, with the following property: for each $j$, there exists a covering $\mathcal{P}_j$ as in Bourgain-Demeter such that any $\theta \in \mathcal{S}_j$ is contained in one element in $\mathcal{P}_j$. Third, sum over $j$ to obtain the result. Alternatively, one can resort to the version in Tao ~\cite{Tao}, Exercise 26.}, this is $$ \dots \lesssim_\epsilon \left( \frac{\delta}{\lambda} \right)^{\frac{d}{p}- \frac{d-1}{4} + \frac{d+1}{2p} - \epsilon} \left( \sum_{\theta \in \mathcal{C}_j} \left\| \phi \left( \frac{\delta X}{\lambda} \right) \sum_{K \in \mathbb{Z}^d / \lambda} \chi_\theta(\lambda K) \chi \left( \frac{|K|-1}{(\delta/\lambda)} \right) a_{\lambda K} e^{2\pi i K \cdot X}\right\|_{L^p(\mathbb{R}^d)}^2 \right)^{1/2} $$ (notice that $\theta/\lambda$ has dimensions $\sim \frac{\delta}{\lambda} \times \frac{\delta^{1/2}}{\lambda^{1/2}} \dots \times \frac{\delta^{1/2}}{\lambda^{1/2}}$). At this point, we use the inequality $$ \mbox{if $p \geq 2$}, \qquad \| f \|_{L^p(\mathbb{R}^d)} \lesssim \| f \|_{L^2} | \operatorname{Supp} \widehat{f} |^{\frac{1}{2} - \frac{1}{p}}, $$ which follows by applying successively the Hausdorff-Young and H\"older inequalities, and finally the Plancherel equality. We use this inequality for $f = \phi \left( \frac{\delta X}{\lambda} \right) \sum_{K} \chi \left( \frac{|K|-1}{(\delta/\lambda)} \right) \chi_\theta(\lambda K) a_{\lambda K} e^{2\pi i K \cdot X}$. Since $\theta \in \mathcal{C}_j$, its Fourier transform is supported on the union of at most $O((\delta \lambda)^{\frac{d-1}{2}}\delta 2^j)$ balls of radius $O( \delta / \lambda) $, giving $| \operatorname{Supp} \widehat{f} | \lesssim \delta^{\frac{3d+1}{2}} \lambda^{-\frac{d+1}{2}} 2^j$. Coming back to the quantity we want to bound, it is \begin{align*} & \lesssim \left( \frac{\delta}{\lambda} \right)^{\frac{d}{p}- \frac{d-1}{4} + \frac{d+1}{2p} - \epsilon} \left(\delta^{\frac{3d+1}{2}} \lambda^{-\frac{d+1}{2}} 2^j \right)^{\frac{1}{2} - \frac{1}{p}} \\ & \qquad \qquad \qquad \qquad \left( \sum_{\theta \in \mathcal{C}_j} \left\| \phi \left( \frac{\delta X}{\lambda} \right) \sum_{K \in \mathbb{Z}^d / \lambda} \chi_\theta(\lambda K) \chi \left( \frac{|K|-1}{(\delta/\lambda)} \right) a_{\lambda K} e^{2\pi i K \cdot X}\right\|_{L^2(\mathbb{R}^d)}^2 \right)^{1/2} \\ & \lesssim \left( \frac{\delta}{\lambda} \right)^{\frac{d}{p} - \frac{d-1}{4} + \frac{d+1}{2p} - \epsilon}\left( \delta^{\frac{3d+1}{2}} \lambda^{-\frac{d+1}{2}} 2^j\right)^{\frac{1}{2} - \frac{1}{p}} \left\| \phi \left( \frac{\delta X}{\lambda} \right) \sum_{K \in \mathbb{Z}^d/\lambda} \chi_j(\lambda K)\chi\left( \frac{|K|-1}{(\delta/\lambda)} \right) a_{\lambda K} e^{2\pi i K \cdot X} \right\|_{L^2(\mathbb{R}^d)}, \end{align*} where the last inequality is a consequence of almost orthogonality. Finally undoing the change of variables, this is \begin{align*} & \lesssim \left( \frac{\delta}{\lambda} \right)^{\frac{d}{p}-\frac{d}{2} - \frac{d-1}{4} + \frac{d+1}{2p} - \epsilon} \left( \delta^{\frac{3d+1}{2}} \lambda^{-\frac{d+1}{2}} 2^j \right)^{\frac{1}{2} - \frac{1}{p}} \left\| \sum_{k \in \mathbb{Z}^d} \chi_j(k) \chi\left( \frac{|k|-\lambda}{\delta} \right) a_k e^{2\pi i k\cdot x} \right\|_{L^2(\mathbb{T}^d)} \\ & \leq \left( \frac{\delta}{\lambda} \right)^{-\epsilon} \lambda^{\sigma(p)/2} \delta^{1/2} 2^{j \left( \frac{1}{2} - \frac{1}{p} \right)} \left\| \sum_{k \in \mathbb{Z}^d} a_k e^{2\pi i k\cdot x} \right\|_{L^2(\mathbb{T}^d)}. \end{align*} \end{proof} \section{Proof of the main theorems} \subsection{The case $p<p_{ST}$: proof of Theorem~\ref{thmpST}} Proposition~\ref{propj} gives the bounds $$ \| P_{\lambda,\delta} \|_{L^2 \to L^{p_{ST}}} \leq \sum_{j=0}^\infty \| P_{\lambda,\delta}^j \|_{L^2 \to L^{p_{ST}}} \lesssim \lambda^{\epsilon} (\delta \lambda)^{\frac{d-1}{2(d+1)}}. $$ Interpolating with the trivial $L^2 \to L^2$ bound, this gives the conjecture for $2 \leq p \leq p_{ST}$. \subsection{An exact but involved statement} The theorems in the introduction are deduced from the following result. \textcolor{black}{The first bound \eqref{eqn:full_bound} in this next theorem is precisely the result of interpolating between the bounds obtained above.} The last part of the theorem allows for the concise result in Theorem~\ref{thmsimple} but, as we will see in the proof, it is slightly weaker. \begin{thm}\label{thm:main} \textcolor{black}{ Assume $p_{ST} \leq p \leq \infty$ and write \begin{align}\label{eqn:excess-of-p} \alpha(p) &=1-\frac{2}{p}, & \beta(p) &= 1- \frac{p_{ST}}{p}. \end{align} Assume further that $\delta> \lambda^{-\frac{d-1}{d+1}}$. Then \begin{multline}\label{eqn:full_bound} \| P_{\lambda,\delta} \|_{L^2 \to L^p} \lesssim \lambda^{\sigma(p)/2} \delta^{1/2} + \lambda^\epsilon (\lambda\delta)^{\frac{d}{4}\alpha(p)} ( \lambda/\delta)^{\frac{d}{4}\beta(p)} \sum_{\substack{ 1\leq k\leq d-1 \\ (\delta \lambda)^{(k-1)/2} \delta<1 }} (\delta \lambda)^{-\frac{k}{4}\alpha(p)} ( \lambda/\delta)^{\frac{-1}{2k}\frac{d}{2}\beta(p)} \\ + \lambda^\epsilon (\lambda\delta)^{\frac{d}{4}\alpha({p})}(\lambda/\delta)^{-\frac{1}{4}\alpha({p})} \delta^{-\frac{d}{2}\beta({p})}. \end{multline} and it follows that \begin{multline*} \| P_{\lambda,\delta} \|_{L^2 \to L^p} \lesssim \lambda^{\sigma(p)/2} \delta^{1/2} + \lambda^\epsilon (\lambda\delta)^{\frac{d}{4}\alpha({p})}(\lambda/\delta)^{-\frac{1}{4}\alpha({p})} \delta^{-\frac{d}{2}\beta({p})} \\ + \lambda^\epsilon (\lambda\delta)^{\frac{d}{4}\alpha(p)} ( \lambda/\delta)^{\frac{d}{4}\beta(p)} e^{-\frac{1}{2}\sqrt{d\alpha(p) \beta(p)\log (\delta \lambda)\log ( \lambda/\delta)}}, \end{multline*} where the final term may be omitted if \(\delta > \lambda^{\frac{\alpha(p)-d\beta(p)}{\alpha(p)+d\beta(p)}}\). } \end{thm} In order to prove the theorem, we need a brief lemma. \begin{lem}\label{lem:optimalk} We introduce the notation $k_0$ for the optimal index in Theorem~\ref{thmcaps} given $\delta,\lambda, 2^j$. Namely, assume that $\delta> \lambda^{-\frac{d-1}{d+1}}$. Then, let $k_0 = k_0(\delta,\lambda,2^j)$ be the smallest $k\in\mathbb Z$ such that $$ (\delta \lambda)^{k/2} \delta 2^j >K. $$ If \(1< 2^j<K\delta^{-1}\) then \(k_0(\delta,\lambda,2^j)\in \{ 1 ,\dots,d-1\}\). \end{lem} \begin{proof} Let \(1< 2^j<K\delta^{-1}\), then \[ K(\delta \lambda)^{-k/2} \delta^{-1} <2^j \leq K(\delta \lambda)^{(1-k)/2} \delta^{-1} \iff k_0=k, \] and so it suffices to observe that as $\delta> \lambda^{-\frac{d-1}{d+1}}$ we have \((\delta \lambda)^{-(d-1)/2} \delta^{-1}< 1\). \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:main}] Throughout the proof, \(k_0\) will be as in Lemma~\ref{lem:optimalk}. \textcolor{black}{We begin by bounding $P^0_{\lambda,\delta}$, by interpolating between} \begin{align*} \| P^0_{\lambda,\delta} \|_{L^2 \to L^{p_{ST}}} \lesssim \lambda^{\frac{d-1}{2(d+1)}} \delta^{1/2}, \end{align*} which is a consequence of Proposition~\ref{propj}, and \begin{align*} \| P^0_{\lambda,\delta} \|_{L^2 \to L^\infty} \lesssim {\bigg\lvert\bigcup_{\theta \in \mathcal{C}_0} \mathbb Z^d\cap \theta\bigg\rvert^{1/2}} \leq \sqrt{|\mathcal{C}|(\sqrt{\lambda \delta})^{d-1} \delta } \lesssim (\lambda^{d-1} \delta)^{1/2}. \end{align*} Interpolating between these two estimates gives $$ \| P^0_{\lambda,\delta} \|_{L^2 \to L^p} \lesssim \lambda^{\sigma(p)/2} \delta^{1/2}. $$ If \(2^j\geq K\delta^{-1}\) then we can assume \(\mathcal{C}_j=\emptyset\) by \eqref{eqn:C-max}, by increasing the size of the constant \(K\) if neccessary. It follows that \(P^j_{\lambda,\delta}=0\) for such \(j\). Next let \(1< 2^j<K\delta^{-1}\). Now on the one hand, Proposition~\ref{propj} gives $$ \| P^j_{\lambda,\delta} \|_{L^2 \to L^{p_{ST}}} \lesssim \lambda^{\frac{1}{p_{ST}}} \delta^{1/2} 2^{\frac{j}{d+1}}. $$ \textcolor{black}{On the other hand}, bounding the number of points in $\cup_{\mathcal{C}_j} \theta$ through Lemma~\ref{lem:optimalk} and Theorem~\ref{thmcaps} gives $$ \| P^j_{\lambda,\delta}\|_{L^2 \to L^\infty} \lesssim \left( \# \mathcal{C}_j (\lambda \delta)^{\frac{d-1}{2}} \delta 2^j \right)^{1/2} \lesssim \left( (2^{j/k_0} \delta)^{-d} (\lambda \delta)^{\frac{d-1}{2}} \delta 2^j \right)^{1/2}. $$ Interpolating between the last two bounds gives \begin{align*} \| P^j_{\lambda,\delta}\|_{L^2 \to L^p} &\lesssim \lambda^{\frac{d-1}{4} - \frac{d-1}{2p}} \delta^{\frac{d+1}{4} - \frac{d+1}{2p}} 2^{\frac{j}{2} - \frac{j}{p}} (2^{j/k_0} \delta)^{-\frac{d}{2} + \frac{d(d+1)}{d-1} \frac{1}{p}} \\ &= (\lambda\delta)^{\frac{d}{4}(1- \frac{2}{p})}(\lambda/\delta)^{-\frac{1}{4}(1- \frac{2}{p})} 2^{\frac{j}{2}(1- \frac{2}{p})} (2^{j/k_0} \delta)^{-\frac{d}{2}(1- \frac{p_{ST}}{p})}. \end{align*} \textcolor{black}{From this point on it will simplify matters to write \(\alpha(p)=1-\frac{2}{p},\beta(p)=1-\frac{p_{ST}}{p}\) as in \eqref{eqn:excess-of-p}.} On summing over $j$, we obtain \begin{equation}\label{eqn:main_first_version} \| P_{\lambda,\delta} \|_{L^2 \to L^p} \lesssim \lambda^{\sigma(p)/2} \delta^{1/2} + (\lambda\delta)^{\frac{d}{4}\alpha({p})}(\lambda/\delta)^{-\frac{1}{4}\alpha({p})} \sum_{1< 2^j < K\delta^{-1}} 2^{\frac{j}{2}\alpha({p})} (2^{j/k_0}\delta )^{-\frac{d}{2}\beta({p})}. \iffalse \| P_{\lambda,\delta} \|_{L^2 \to L^p} \lesssim \lambda^{\sigma(p)/2} \delta^{1/2} + \lambda^{\frac{d-1}{4} - \frac{d-1}{2p}} \delta^{-\frac{d-1}{4} + \frac{(d+1)^2}{2(d-1)p}} \sum_{1< 2^j < K\delta^{-1}} 2^{\frac{j}{2} - \frac{j}{p}} (2^{j/k_0})^{-\frac{d}{2} + \frac{d(d+1)}{d-1} \frac{1}{p}}. \fi \end{equation}\textcolor{black}{ Recall that \(k_0\) is minimal such that \((\delta \lambda)^{-k/2} \delta^{-1} <2^j \), and that \(1\leq k_0\leq d-1\) by Lemma~\ref{lem:optimalk}. Hence \begin{align*} \sum_{1< 2^j < K\delta^{-1}} 2^{\frac{j}{2}\alpha({p})} (2^{j/k_0}\delta )^{-\frac{d}{2}\beta({p})} &\leq \sum_{1\leq k\leq d-1} \sum_{\substack{j\in\mathbb N \\ (\delta \lambda)^{-k/2} \delta^{-1} <2^j \\ (\delta \lambda)^{(1-k)/2} \delta^{-1} \geq 2^j \\ 1< 2^j < K\delta^{-1} }} 2^{\frac{j}{2}\alpha({p})} (2^{j/k} \delta)^{-\frac{d}{2}\beta({p})} \\ & \lesssim_\epsilon \sum_{1\leq k\leq d-1}\lambda^\epsilon \sum_{\substack{j \in\mathbb R \\ 2^j=\max\{1,(\delta \lambda)^{-\ell/2} \delta^{-1}\} \\\ell\in\{k-1,k\}}} 2^{\frac{j}{2}\alpha({p})} (2^{j/k} \delta)^{-\frac{d}{2}\beta({p})}, \intertext{and exchanging order of summation this is} & = \lambda^\epsilon \sum_{\substack{j \in\mathbb R \\ 2^j=\max\{1,(\delta \lambda)^{-\ell/2} \delta^{-1}\} \\0\leq \ell\leq d-1}} \sum_{\substack{ k\in\{\ell,\ell+1\} \\ 1\leq k\leq d-1}} 2^{\frac{j}{2}\alpha({p})} (2^{j/k} \delta)^{-\frac{d}{2}\beta({p})}. \end{align*} On recalling that \(p\geq p_{ST}\) and so \(\beta(p)\geq 0\), we see that when \(\ell\leq d-2\), we can obtain an upper bound for the inner sum in the last line above by substituting \(k= \ell+1\). Thus \begin{align*} \sum_{1< 2^j < K\delta^{-1}} % 2^{\frac{j}{2}\alpha({p})} (2^{j/k_0}\delta )^{-\frac{d}{2}\beta({p})} &\lesssim_{\epsilon} \lambda^\epsilon \sum_{\substack{j \in\mathbb R \\ 2^j=\max\{1,(\delta \lambda)^{-\ell/2} \delta^{-1}\} \\0\leq \ell\leq d-2}} 2^{\frac{j}{2}\alpha({p})} (2^{\frac{j}{\ell+1}} \delta)^{-\frac{d}{2}\beta({p})} \\ &\hphantom{{}\leq{}}+ \lambda^\epsilon \max\left\{1,(\delta \lambda)^{-(d-1)/2} \delta^{-1}\right\}^{\frac{1}{2}\alpha({p})-\frac{d}{2(d-1)}\beta({p})} \delta^{-\frac{d}{2}\beta({p})} \\ &= \lambda^\epsilon \sum_{\substack{ 1\leq k\leq d-1 \\ (\delta \lambda)^{(k-1)/2} \delta<1 }} \left((\delta \lambda)^{-(k-1)/2} \delta^{-1}\right)^{\frac{1}{2}\alpha({p})-\frac{d}{2k}\beta({p})} \delta^{-\frac{d}{2}\beta({p})} + \lambda^\epsilon \delta^{-\frac{d}{2}\beta({p})}, \end{align*} where to deal with the maximum we use the assumption \(\delta>\lambda^{-\frac{d-1}{d+1}}\) from the theorem.Upon writing \((\delta \lambda)^{-(k-1)/2} \delta^{-1} = (\delta\lambda)^{-k/2} (\lambda/\delta)^{1/2}\), it now follows by \eqref{eqn:main_first_version} that \begin{multline*} \| P_{\lambda,\delta} \|_{L^2 \to L^p} \lesssim \lambda^{\sigma(p)/2} \delta^{1/2} + \\ \lambda^\epsilon (\lambda\delta)^{\frac{d}{4}\alpha({p})}(\lambda/\delta)^{-\frac{1}{4}\alpha({p})} \sum_{\substack{ 1\leq k\leq d-1 \\ (\delta \lambda)^{(k-1)/2} \delta<1 }} (\delta\lambda)^{-\frac{k}{2}\big(\frac{1}{2}\alpha({p})-\frac{d}{2k}\beta({p})\big)-\frac{d}{4}\beta(p)} (\lambda/\delta)^{\frac{1}{2}\big(\frac{1}{2}\alpha({p})-\frac{d}{2k}\beta({p})\big)+\frac{d}{4}\beta(p)} \\ + \lambda^\epsilon (\lambda\delta)^{\frac{d}{4}\alpha({p})}(\lambda/\delta)^{-\frac{1}{4}\alpha({p})} \delta^{-\frac{d}{2}\beta({p})}. \end{multline*} This proves \eqref{eqn:full_bound}, and we now proceed to deduce the last part of the theorem. If \(A,B>0\) then \[\exp(-kA-B/k)\leq \exp(-2\sqrt{AB}),\] and moreover \[ \sum_{\substack{ 1\leq k\leq d-1 \\ k < k_1 }} \exp(-kA-B/k)\lesssim \exp(-k_1A-B/k_1) \qquad\text{if } k_1\leq \sqrt{B/A}. \] We apply this with \begin{align*} 4A&=\alpha(p)\log (\delta \lambda) ,& 4B&=d\beta(p)\log ( \lambda/\delta) & k_1 &= \log (\lambda/\delta) /\log (\lambda\delta), \end{align*} so that \( (\delta \lambda)^{(k_1-1)/2} \delta=1\). Noting that \[ \exp( -k_1A-B/k_1) = \exp\Big( - \tfrac{1}{4}\alpha(p)\log(\lambda/\delta) - \tfrac{d}{4}\beta(p) \log (\delta \lambda)\Big) = \delta^{-\frac{d}{2}\beta(p)}(\lambda/\delta)^{-\frac{d}{4} \beta(p)-\frac{1}{4}\alpha(p) } , \] we deduce from \eqref{eqn:full_bound} that \begin{multline*} \| P_{\lambda,\delta} \|_{L^2 \to L^p} \lesssim \lambda^{\sigma(p)/2} \delta^{1/2} + \lambda^\epsilon (\lambda\delta)^{\frac{d}{4}\alpha({p})}(\lambda/\delta)^{-\frac{1}{4}\alpha({p})} \delta^{-\frac{d}{2}\beta({p})} \\ + \lambda^\epsilon (\lambda\delta)^{\frac{d}{4}\alpha(p)} ( \lambda/\delta)^{\frac{d}{4}\beta(p)} e^{-\frac{1}{2}\sqrt{d\alpha(p) \beta(p)\log (\delta \lambda)\log ( \lambda/\delta)}}, \end{multline*} where the last term is omitted if \(k_1 <\sqrt{B/A}\). This is the remaining bound in Theorem~\ref{thm:main}. } \end{proof} \subsection{From Theorem~\ref{thm:main} to Theorems~\ref{thmsimple} and~\ref{thmd3}} \begin{proof}[Proof of Theorem~\ref{thmsimple}] \textcolor{black}{We use throughout the proof the notation \(\alpha(p)=1-\frac{2}{p},\beta(p)=1-\frac{p_{ST}}{p}\) from \eqref{eqn:excess-of-p}. As in theorem we assume \(\delta > \lambda^{\frac{\alpha(p)-d\beta(p)}{\alpha(p)+d\beta(p)}}\). We claim that \(\delta\geq \lambda^{-\frac{d-1}{d+1}}\), that is to say \[ \frac{1-\beta(p)\alpha(p)^{-1}d}{1-\beta(p)\alpha(p)^{-1}d}> \frac{1-d}{1+d}, \] which is true since \(\beta(p)\alpha(p)^{-1} < 1\) and \(\frac{1-dx}{1+dx}\) is an increasing function of \(x\).} The last bound in Theorem~\ref{thm:main} will now agree with Conjecture~\ref{conj} if \textcolor{black}{ \begin{align} (\lambda\delta)^{\frac{d}{4}\alpha({p})}(\lambda/\delta)^{-\frac{1}{4}\alpha({p})} \delta^{-\frac{d}{2}\beta({p})} &\leq\lambda^{\frac{d-1}{2}-\frac{d}{p}} \delta^{1/2}. \label{eqn:easy-case} \end{align} Grouping all the terms with a \(1/p\) and without a \(1/p\) in their exponent, the bound \eqref{eqn:easy-case} is exactly \[ \lambda^{ \frac{d+1}{2p} } \delta^{\frac{(d+1)^2}{2(d-1)p}} \leq \lambda^{\frac{d-1}{4}} \delta^{\frac{d+1}{4}}, \] which is to say \((\lambda \delta^{\frac{d+1}{d-1}})^{-\frac{d-1}{4}\beta(p)}\leq 1 \), and this is true since \(\delta\geq \lambda^{-\frac{d-1}{d+1}}\) and \(p>p_{ST}\).} \end{proof} \begin{proof}[Proof of Theorem~\ref{thmd3}] \textcolor{black}{As in the last proof, we use throughout the notation \(\alpha(p)=1-\frac{2}{p},\beta(p)=1-\frac{p_{ST}}{p}\) from \eqref{eqn:excess-of-p}, noting that since \(d=3\) we have \(p_{ST}=4\) and \(\beta(p)=1-\frac{4}{p}\).} If \(d=3\) the first bound in Theorem~\ref{thm:main} agrees with the conjecture when \begin{multline*} (\lambda\delta)^{\frac{3}{4}\alpha(p)} ( \lambda/\delta)^{\frac{3}{4}\beta(p)} (\delta \lambda)^{-\frac{1}{4}\alpha(p)} ( \lambda/\delta)^{-\frac{3}{4}\beta(p)} \\ + (\lambda\delta)^{\frac{3}{4}\alpha(p)} ( \lambda/\delta)^{\frac{3}{4}\beta(p)} (\delta \lambda)^{-\frac{2}{4}\alpha(p)} ( \lambda/\delta)^{-\frac{3}{8}\beta(p)} \leq (\lambda\delta)^{\frac{3}{4}\alpha(p)} (\lambda/\delta)^{\frac{1}{4}-\frac{3}{2p}} +(\lambda\delta)^{\frac{1}{2}\alpha(p)}, \end{multline*} that is \begin{equation*} (\lambda\delta)^{\alpha(p)} + (\lambda\delta)^{\frac{1}{4}\alpha(p)} ( \lambda/\delta)^{\frac{3}{8}\beta(p)} \leq (\lambda\delta)^{\frac{3}{4}\alpha(p)} (\lambda/\delta)^{\frac{1}{4}-\frac{3}{2p}} +(\lambda\delta)^{\frac{1}{2}\alpha(p)}. \end{equation*} The first term on the left-hand side is clearly bounded by the last term. The second term on the left-hand side is bounded by the right-hand side if \begin{equation*} ( \lambda/\delta)^{\frac{1}{8}} \leq (\lambda\delta)^{\frac{1}{2}\alpha(p)} \qquad\text{or}\qquad ( \lambda/\delta)^{\frac{3}{8}\beta(p)} \leq (\lambda\delta)^{\frac{1}{4}\alpha(p)}. \end{equation*} This is \(\delta \geq \min\{\lambda^{-\frac{3p-8}{5p-8}}\, \lambda^{-\frac{8-p}{5p-16}}\}\), and recalling our standing assumption \(\delta\geq \lambda^{-(d-1)/(d+1)}\) from Theorem~\ref{thm:main} this gives the result. \end{proof}

21da34fc89aff12f29a52813279bd989bf93cc58

\section{Introduction} Let $\mathcal{C}$ be a class of countable $\L$-structures on universe $\mathbb{N}$ and let $\mathcal{K}$ be a countable family of finitely generated $\mathcal{L}$-structures which has the {\bf Hereditary Property} (HP) and the {\bf Joint Embedding Property} (JEP): \begin{itemize} \item (HP) if $B\in \mathcal{K}$, and $A$ embeds in $B$ and is finitely generated, then $A\in \mathcal{K}$; \item (JEP) if $A,B\in \mathcal{K}$, then there is $C\in \mathcal{K}$ so that $A,B$ embed in $C$. \end{itemize} We say that $\mathcal{C}$ is {\bf approximable} by $\mathcal{K}$ if $\mathcal{C}$ consists precisely of all $\mathcal{L}$-structures $N$ on universe $\mathbb{N}$ with the properties: \begin{itemize} \item every finitely generated substructure $A$ of $N$ is in $\mathcal{K}$; \item $N$ is not finitely generated. \end{itemize} In this case, we write $\mathcal{K}=\mathrm{Age}(\mathcal{C})$ and $\mathcal{C}=\lim\mathcal{K}$. We say that $\mathcal{C}$ is {\bf approximable} if $\mathcal{C}=\lim\mathcal{K}$, for some $\mathcal{K}$ as above. Standard examples of approximable classes are classes consisting of all countably infinite graphs that omit all graphs which are included in some fixed countable collection $\mathcal{F}$ of ``forbidden" finite graphs. In what follows, we will always assume that $\mathcal{C}$ is approximable and that each structure in $\mathcal{C}$ has universe $\mathbb{N}$. We say that $\mathcal{C}$ has a {\bf universal element} $U\in \mathcal{C}$, if every $N\in\mathcal{C}$ embeds in $U$. We say that $\mathcal{C}$ has a {\bf generic element} $G\in \mathcal{C}$, if in the Polish space $\mathcal{C}$ (see Section \ref{S:E}) the following set is comeager: \[\{N\in \mathcal{C} \; \colon N \text{ is isomorphic to } G \}.\] Several approximable classes $\mathcal{C}$ admit a structure $M\in\mathcal{C}$ that is both universal and generic. For example, the Rado graph is both a universal and a generic element of the class of all countable graphs; see \cite{Rado,TZ}. More generally, if $\mathcal{K}=\mathrm{Age}(\mathcal{C})$ is a {\bf Fra\"iss\'e{} class} then $\mathcal{C}$ contains a canonical structure $M\in \mathcal{C}$---known as the Fra\"iss\'e{} limit of $\mathcal{K}$---which is both universal and generic; see \cite{TZ}. However, it is possible for an approximable class $\mathcal{C}$ to admit a universal $U$ that is not generic and a generic $G$ that is not universal; see Example \ref{E:1}. It is therefore natural to ask whether there is an example of an approximable class which admits a universal element but does not admit a generic element and similarly whether there is an approximable class which admits a generic element but does not admit a universal element. We settle this with the following theorem. \begin{theorem}\label{T:1} Let $\mathcal{C}$ be an approximable class of $\mathcal{L}$-structures. If $\mathcal{C}$ admits a universal element, then $\mathcal{C}$ admits a generic element. The converse is not true. \end{theorem} As a consequence of Theorem \ref{T:1}, we get several new examples of approximable classes of countable graphs which admit a generic element. Notice that, by the Lachlan-Woodrow classification theorem \cite{Lachlan}, the Fra\"iss\'e{} construction---in its original form---provides very few examples of approximable classes of graphs which admit a generic element. While we now know that a certain weak version of the Fra\"iss\'e{} construction is essentially the only way to produce generic elements~\cite{KKgames,K, KT, DiLiberti}, very few ``weak Fra\"iss\'e{} classes" of graphs have been explicitly constructed. On the other hand, there is a vast literature on approximable classes $\mathcal{C}$ of graphs which admit a universal element $U\in\mathcal{C}$; e.g., \cite{Ko,KMP, KP, DHV,CT,CSS}. By Theorem \ref{T:1}, these classes also admit a generic element $G\in\mathcal{C}$. We collect some examples in Section \ref{S:E}. Finally, there are several approximable classes $\mathcal{C}$ for which it has been established that they do not admit a universal element. A natural question stemming from Theorem \ref{T:1} is whether results about non-existence of universal elements can be strengthened to non-existence of generic elements. Here we strengthen a result of Hajnal and Pach on the non-existence of a universal countable $C_4$-free graph \cite{HP}. Recall that a graph $A$ is {\bf $C_4$-free} if there exists no injective homomorphism from $C_4$ to $A$, or equivalently, if $A$ does not contain a simple $4$-cycle. We show the following: \begin{theorem}\label{T:2} The class of all countable $C_4$-free graphs admits no generic element. \end{theorem} By Theorem \ref{T:1}, establishing that an approximable class $\mathcal{C}$ does not admit a generic element is apriori more difficult than establishing that $\mathcal{C}$ does not admit a universal element. The proof of Theorem \ref{T:2} demonstrates this in practice since, unlike with the arguments from \cite{HP}, its proof heavily relies on structure theorems for strongly regular graphs and $C_4$-free graphs of diameter $2$ from \cite{BB,DF}. \subsection*{Acknowledgments} We would like to thank A. Kruckman and W. Kubi\'s{} for their valuable feedback on an earlier draft of this paper. After the completion of this paper we were informed by W.Kubi\'s that Theorem \ref{T:1} has independently been established in \cite[Section 6]{KKgames}; see Remark \ref{R:1} below. However, the argument we provide here avoids the game-theoretic formalism of \cite{KKgames} and the counterexample we supply for the second statement of Theorem \ref{T:1} is different and interesting on its own right since it relates to the theory of bowtie free graphs. \section{Definitions and Examples}\label{S:E} We will be using the usual model theoretic notions of $\mathcal{L}$-structures, embeddings, etc.; see \cite{TZ}. In particular, an {\bf embedding} $f\colon (A,R^A)\to (B,R^B)$ of graphs is any injective function $f\colon A\to B$ with $(a,a')\in R^A \iff (f(a),f(a'))\in R^B$. An injective homomorphism, or a {\bf weak embedding} is any injective function $f\colon A\to B$ with $(a,a')\in R^A \implies (f(a),f(a'))\in R^B$. We warn the reader that, in the graph theoretic terminology from \cite{Ko,KMP, KP, DHV,CT,CSS}, the term embedding is used for what we call here weak embedding. Below we will often use the notation $N\leq M$ to indicate that the $\mathcal{L}$-structure $N$ is a substructure of the $\mathcal{L}$-structure $M$. That is, to indicate that the identity map induces an embedding of $N$ into $M$ Let $\mathcal{K}$ be a collection of $\mathcal{L}$-structures. We say that $\mathcal{K}$ has the {\bf Weak Amalgamation Property (WAP)} if for every $A\in\mathcal{K}$ there is an embedding $i\colon A\to\widehat{A}$ so that for every two embeddings $f\colon\widehat{A}\to B$ and $g\colon\widehat{ A}\to C$, with $B,C\in\mathcal{K}$, embeddings $f'\colon B\to D$ and $g'\colon C\to D$, with $D\in\mathcal{K}$, so that $f'\circ f \circ i = g' \circ g \circ i.$ We say that $\mathcal{K}$ has the {\bf Cofinal Amalgamation Property (CAP)} if we can replace $f'\circ f \circ i = g' \circ g \circ i$ with $f'\circ f = g' \circ g$ in the above definition. We say that $\mathcal{K}$ has the {\bf Amalgamation Property (AP)} if we can take $A=\widehat{A}$ in the above definition. Assume that $\mathcal{K}$ is a countable class of finitely generated $\mathcal{L}$-structures with HP and JEP. Let $\mathcal{C}=\mathrm{lim}(\mathcal{K})$ be the associated approximable class. As usual, we view $\mathcal{C}$ as a Polish space whose basic open sets are of the form: \[\mathcal{O}_A=\{N\in \mathcal{C} \colon \text{ the restriction of } N \text{ to } \mathrm{dom}(A) \text{ is equal to } A \},\] where $A$ ranges over all elements of $\mathcal{K}$ with $\mathrm{dom}(A)\subseteq \mathbb{N}$; see e.g., \cite{KT}. Since $\mathcal{C}$ is a Polish space, the collection of all comeager subsets of $\mathcal{C}$ forms a non-trivial $\sigma$-filter on $\mathcal{C}$; see \cite{Kechris}. We say $\mathcal{K}$ is a {\bf Fra\"iss\'e{} class} if $\mathcal{K}$ additionally has the amalgamation property. In this case, $\mathcal{C}$ contains a canonical structure $M\in\mathcal{C}$, known as the {\bf Fra\"iss\'e{} limit of $\mathcal{K}$}, which is both a universal and a generic element of $\mathcal{C}$; see \cite{TZ}. The full strength of the amalgamation property is not necessary for the existence of a generic element of $\mathcal{C}$. It turns out that an approximable class $\mathcal{C}=\mathrm{lim}(\mathcal{K})$ admits a generic element if and only if $\mathcal{K}$ is a {\bf weak Fra\"iss\'e{} class}, i.e., if $\mathcal{K}$ additionally has the weak amalgamation property; see \cite[Theorem 4.2.2]{K} or \cite[Theorem 2.5]{KT}. This was originally proved for certain special classes of partial automorphisms by Ivanov~\cite{Ivanov} and independently by Kechris and Rosendal~\cite{KR}. It was later extended to more general classes in \cite{K} and independently by several other authors in~\cite{KKgames, KT, DiLiberti}. Here we follow \cite{KT}, whose exposition is closer in spirit to this paper. The following example demonstrates that, in the absence of the full amalgamation property for $\mathcal{K}$, the associated approximable class $\mathcal{C}$ may admit a universal element $U$ which is not generic, as well as a generic element $G$ which is not universal. \begin{example}\label{E:1} Let $\mathcal{K}_{\mathrm{linear}}$ be the class of all finite graphs which are disjoint unions of linear graphs. By a linear graph we mean any graph which is isomorphic to the graph on domain $\{0,\ldots,n-1\}$ with $(k R \ell)\iff (|k-\ell|=1)$, for some $n\in\mathbb{N}$. It is easy to see that $\mathcal{K}_{\mathrm{linear}}$ has the cofinal amalgamation property and hence $\mathcal{C}_{\mathrm{linear}} = \lim \mathcal{K}_{\mathrm{linear}}$ admits a generic element $G$. A. Kruckman pointed out that $G$ is not universal since it does not embed any strict extension of $G$ which is in $\mathcal{C}_{\mathrm{linear}}$. It is also clear that $\mathcal{C}_{\mathrm{linear}}$ admits a universal which is not generic for $\mathcal{C}_{\mathrm{linear}}$. Namely, the graph $U$ consisting of countably infinite many disjoint copies of a bi-infinite line. \end{example} As an application of Theorem \ref{T:1}, we close this section by demonstrating that many well-studied classes $\mathcal{K}$ of finite graphs are actually weak Fra\"iss\'e{} classes. Notice that, by the Lachlan-Woodrow classification theorem \cite{Lachlan}, the only Fra\"iss\'e{} classes of finite graphs are: the class of all graphs, classes consisting of disjoint unions of complete graphs; the class of all $K_n$-free graphs; complements of the last two classes. Let $\mathcal{F}$ be a collection of finite graphs. We say that a graph {\bf $N$ omits graphs from $\mathcal{F}$}, if there exists no injective homomorphism $f\colon A\to N$ with $A\in \mathcal{F}$. Let $\mathcal{K}(\mathcal{F})$ and $\mathcal{C}(\mathcal{F})$ be all finite and all countable, respectively, graphs which omit graphs from $\mathcal{F}$. Below, $C_n$ and $K_n$ denote the cycle and the complete graph of size $n$, respectively. \begin{example}\label{E:2} The class $\mathcal{C}(\mathcal{F})$ does not admit a universal element if $\mathcal{F}$ consists of: \begin{enumerate}\setlength\itemsep{5pt}\vspace{2pt} \item the singleton $\{\bowtie\}$, where $\bowtie$ is the bowtie graph; see \cite{Ko}. \item the infinite collection $\{C_n,C_{n+1},\ldots\}$, for some $n>0$; see \cite{KMP}. \item the finite collection $\{C_3,C_5,\ldots,C_{2n+1}\}$, for some $n\geq 1$; see \cite{KMP} and \cite{Ko}. \item the singleton $\{P_n\}$, where $P_n$ is the length $n$ path, for some $n>0$; see \cite{KMP}. \item the singleton $\{N\}$, where $N$ is a near path, i.e., a finite tree consisting of a path with at most one additional edge adjoined; see \cite{CT}. \item the collection $\mathrm{top}(K_n)$ for some fixed $n\leq 4$, where $\mathrm{top}(K_n)$ consists of all finite graphs which are topologically equivalent to $K_n$, i.e., all finite graphs which are simplicial subdivisions of $K_n$; see \cite{DHV} and \cite{KP}. \item any finite collection of finite connected graphs that is homomorphism-closed, i.e., if $f\colon B\to A$ is a surjective graph homomorphism with with $B\in \mathcal{F}$ then there is a $C\in \mathcal{F}$ which weakly embeds in $A$; see \cite{CSS}. \end{enumerate} \end{example} \begin{corollary}\label{Cor:1} $\mathcal{K}(\mathcal{F})$ is a weak Fra\"iss\'e{} class, for all $\mathcal{F}$ from Example \ref{E:2}. \end{corollary} The fact that $\mathcal{K}(\{\bowtie\})$ has WAP was already known. In fact, in \cite{HN} it is shown that $\mathcal{K}(\{\bowtie\})$ has the CAP; see also \cite{S}. It turns out that the associated generic bowtie-free countable graph has several curious properties, both from a Ramsey-theoretic as well as from a model-theoretic standpoint; see \cite{HN}. We should point out that, even when it comes to general classes of $\mathcal{L}$-structures, we have very few (and rather artificial) examples of weak Fra\"iss\'e{} classes $\mathcal{K}$ which do not already satisfy the cofinal amalgamation property; see \cite{WnC}. It would be interesting if either of the examples (2)-(7) above could provide more natural examples. \begin{question} Does any $\mathcal{K}(\mathcal{F})$ in Corollary \ref{Cor:1} fail the cofinal amalgamation property? \end{question} \section{Universality vs Genericity} We first show that existence of a universal implies existence of a generic element. By \cite[Theorem 4.2.2]{K}, see also \cite[Theorem 2.5]{KT}, it suffices to show the following: \begin{theorem}\label{T:11} Let $\mathcal{C}$ be an approximable class of $\mathcal{L}$-structures and let $\mathcal{K}=\mathrm{Age}(\mathcal{C})$. If $\mathcal{C}$ admits a universal element then $\mathcal{K}$ has WAP. \end{theorem} \begin{proof} We prove the contrapositive. Assume that $\mathcal{K}$ does not have WAP and let $A\in\mathcal{K}$ so that for every embedding $i\colon A\to \widehat{A}$ with $\widehat{A}\in\mathcal{K}$ there are further extensions $B,C$ of $\widehat{A}$ which do not amalgamate over $A$. For every $\alpha\in 2^{\mathbb{N}}$ we will define a structure $N_{\alpha}\in\mathcal{C}$, together with an embedding $i_{\alpha}\colon A\to N_{\alpha}$, so that for all $\alpha, \beta\in 2^{\mathbb{N}}$ with $\alpha\neq \beta$ there are $B,C\in\mathcal{K}$ with $A\leq B\leq N_{\alpha}$ and $A\leq B\leq N_{\beta}$, so that $B$ and $C$ do not amalgamate over $A$. Assuming that this construction is possible, we conclude with the proof of the theorem as follows: assume that $\mathcal{C}$ admits a universal element $U$ and for every $\alpha\in 2^{\mathbb{N}}$ fix some embedding $f_{\alpha}\colon N_{\alpha}\to U$. Since $U$ is countable and $A$ is finitely generated, there is an uncountable $J\subseteq 2^{\mathbb{N}}$ so that for all $\alpha,\beta\in J$ we have that \[f_{\alpha} \circ i_{\alpha}=f_{\beta} \circ i_{\beta}.\] By assumption, we have $B,C\in\mathcal{K}$ with $A\leq B\leq N_{\alpha}, A\leq B\leq N_{\beta}$, so that $B,C$ do not amalgamate over $A$. Let $B'$ be the image of $B$ under $f_{\alpha}$ and let $C'$ be the image of $C$ under $f_{\beta}$. But the substructure $D$ of $M$ generated by both $B'$ and $C'$ is an amalgam of $B$ and $C$ over $A$. By the hereditary property of $\mathcal{K}$ and since $D$ is finitely generated, we have that $D\in\mathcal{K}$. This yields a contradiction. We may now turn to the construction of the family $(i_\alpha\colon \alpha\in 2^{\mathbb{N}})$. For every $t\in 2^{<\mathbb{N}}$ we will define some $A_t\in\mathcal{K}$ so that for every initial segment $s\preceq t$ we have that $A\leq A_{s}\lneq A_{t}$. Then we will define $N_{\alpha}$ as the increasing union of $(A_{\alpha|n}\colon n\in\mathbb{N})$. Finally, we will make sure that for every $s\in 2^{<\mathbb{N}}$ the extensions $A_{s^{\frown}0}$ and $A_{s^{\frown}1}$ of $A$ do not amalgamate over $A$. The construction is done inductively as follows. Set $A_{\emptyset}:=A$. Assume that $A_{s}\in\mathcal{K}$ has been defined for some $s\in 2^{<\mathbb{N}}$. Setting $\widehat{A}:=A_s$, we get extensions $B,C\in\mathcal{K}$ of $\widehat{A}$ which do not amalgamate over $A$. Set $A_{s^{\frown}0}:=B$ and $A_{s^{\frown}1}:=C$. \end{proof} \begin{rem}\label{R:1} In \cite[Corollary 6.3]{KKgames}, A. Krawczyk and W. Kubi\'s prove the following strengthening of Theorem \ref{T:11}: ``if there exists a family $\mathcal{U}\subseteq \mathcal{C}$ with $|\mathcal{U}|<2^{\aleph_0}$, so that for all $N\in\mathcal{C}$ there is $U\in\mathcal{U}$ with $N\leq U$, then $\mathcal{K}$ has WAP". A straightforward adaptation of the above proof can be used to establish the aforementioned strengthening. Indeed, if $\mathcal{U}$ is universal---as a family---for $\mathcal{C}$, with $|\mathcal{U}|<2^{\aleph_0}$, then one can first choose some uncountable $I\subseteq 2^{\mathbb{N}}$ and some $U\in\mathcal{U}$ so that $N_{\alpha}\leq U$, for all $\alpha\in I$. The rest of the proof is verbatim, modulo requiring that $J\subseteq I$. \end{rem} Next we show that the converse statement is not true. That is, there is some approximable $\mathcal{C}$ which admits a generic element but no universal element. Let $H$ be the \emph{Windmill graph} $Wd(3,3)$, i.e. the graph consisting of $3$ triangles all sharing exactly one common vertex $p$, as shown in the following figure: \begin{figure}[ht!] \begin{tikzpicture}[x=0.6pt,y=0.6pt,yscale=-1,xscale=1] \draw (0,75) node (1){$\bullet$} -- (150,75) node (2){$\bullet$}; \draw (37.50,10) node (3){$\bullet$} -- (115,140) node (4){$\bullet$} ; \draw (75,75) node (5){$\bullet$} -- (37.50,140) node (6){$\bullet$} ; \draw (110,10) node (7){$\bullet$} -- (75,75) ; \draw (37.50,10) -- (110,10) ; \draw (0,75) -- (37.50,140) ; \draw (150,75) -- (112.5,140) ; \draw (75,95) node (5){$p$} ; \end{tikzpicture} \caption{The graph $H$ together with its center $p$.}\label{F:Wheel} \end{figure} Let $\mathcal{K}(H)$ be the collection of all finite graphs which omit the graph $H$. That is, all graphs $A$ for which there is no injective homomorphism $f\colon H\to A$. In \cite{Ko} it is proved that $\mathcal{C}(H):=\lim \mathcal{K}(H)$ does not admit a universal element. \begin{theorem} The class $\mathcal{C}(H)$ admits a generic element. \end{theorem} \begin{proof} By \cite[Theorem 4.2.2]{K} it suffices to show $\mathcal{K}(H)$ has WAP. Let $A\in \mathcal{K}(H)$ and let $\bowtie$ be the bowtie graph, endowed with the following labeling: \begin{figure}[ht!] \begin{tikzpicture}[x=0.35pt,y=0.35pt,yscale=-1,xscale=1] \draw (150,75) node (1){$\bullet$} -- (0,0) node (2){$\bullet$} -- (0,150) node (3){$\bullet$} -- cycle; \draw (150,75) -- (300,150) node (4){$\bullet$} -- (300,0) node (5){$\bullet$}-- cycle ; \draw (150,105) node (1'){$v$}; \draw (-35,0) node (12){$v_2$}; \draw (-35,150) node (12){$v_1$}; \draw (335,0) node (12){$v_4$}; \draw (335,150) node (12){$v_3$}; \end{tikzpicture} \caption{}\label{F:bowtie} \end{figure} \begin{claim}\label{Claim1} There is an extension $A\leq E \in \mathcal{K}(H)$, of $A$ so that for every vertex $a$ of $A$ there is an injective homomorphism $j\colon\bowtie{}\to E$, with $j(v)=a$ \end{claim} \begin{proof}[Proof of Claim] Let $F\in \mathcal{K}(H)$ be any extension $A\leq F$ and fix some $a\in A$. We first define an extension $F\leq F(a)\in \mathcal{K}(H)$, for which there is an injective homomorphism $j\colon\bowtie{}\to F(a)$ with $j(v)=a$. If there is some injective homomorphism $j\colon\bowtie{}\to F$ with $j(v)=a$, then set $F(a):=F$. Otherwise, there is either no edge $b R b'$ in $A$, with $a R b$ and $a R b'$; or there is some such edge $bRb'$, so that for every edge $c R c'$ with $aR c$ and $aR c'$ we have either $c=b$ or $c'=b$. In the first case, we define the graph $F(a)$ as the extension of $A$ with a copy of $\bowtie$ above under the identification $v=a$. In the second case, we define the graph $F(a)$ as the result of introducing two new vertices $v_1,v_2$ to $A$ together with the edges $v_1 R v_2$, $v_1 R a$, $v_2 R a$. Let now $a_1,\ldots,a_n$ be an enumeration of the vertices of $A$. Define inductively $F_0:= A$ and $F_k:=F_{k-1}(a_k)$. It follows that $E:= F_n$ is the desired graph. \end{proof} Let $E$ be the graph from \ref{Claim1} and for every vertex $a$ of $A$ let $v_1[a],v_2[a],v[a],v_3[a],v_4[a]$ be the vertices of $E$ which are the images of the associated vertices of $\bowtie$ from Figure \ref{F:bowtie}, under some fixed injective homomorphism $\bowtie{} \to E$, with $v[a]=a$. \begin{claim}\label{Claim2} There is an extension $E\leq \widehat{A} \in \mathcal{K}(H)$ of $E$, so that for every vertex $a$ of $A$, and every $k\in \{1,2,3,4\}$, if there is a further extension $ \widehat{A}\leq F \in \mathcal{K}(H)$ containing a vertex $w$ with $w R v_k[a]$, $w R a$, and $w\not\in \{v_1[a],v_2[a], v_3[a],v_4[a]\}$, then there is a vertex $\hat{w}$ already in $\widehat{A}$, so that $\hat{w} R v_k[a]$, $\hat{w} R a$, and $\hat{w}\not\in \{v_1[a],v_2[a],v_3[a],v_4[a]\}$. \end{claim} \begin{proof}[Proof of Claim] The proof is by a simple induction, similar to the proof of Claim \ref{Claim1}, and it is left to the reader. \end{proof} Let now $B,C\in \mathcal{K}(H)$ be any two extensions $\widehat{A}\leq B,C$ of $\widehat{A}$. We will show that $B,C$ amalgamate over $A$ to some $D\in\mathcal{K}(H)$. Indeed, let $D$ be the {\bf free amalgam} $D:=B\sqcup_{A} C$ of $B$ and $C$ over $A$. That is, $D$ is the graph attained by taking the disjoint union of the graphs $B$ and $C$, and identifying the copy of each vertex $a$ of $A$ which lies in $B$, with the associated copy of $a$ which lies in $C$. We will show that $D\in \mathcal{K}(H)$. Assume towards contradiction that there is some injective homomorphism $j\colon H\to D$. Let $a:=j(p)$ be the image of the center $p$ of $H$ under $j$ and let $v_1[a],v_2[a],v[a]=a,v_3[a],v_4[a]$ be the vertices of some non-induced copy of $\bowtie$ in $\widehat{A}$. These exist by Claim \ref{Claim1}. By the structure of the free amalgam $D:=B\sqcup_{A} C$, it follows that $a$ lies in $A$: otherwise the copy of $H$ would already have to be entirely included either in $B$ or in $C$, contradicting that $B\in\mathcal{K}(H)$ or $C\in\mathcal{K}(H)$, respectively. The following fact is going to be used repeatedly: \begin{claim}\label{Claim3} If $(a,e_1,e_2)$ is a $3$-cycle in $D$, then $(a,e_1,e_2)$ is entirely included either in $B$ or in $C$. Moreover, $\{e_1,e_2\}\cap \{ v_1[a],v_2[a],v_3[a],v_4[a]\}\neq \emptyset$. \end{claim} \begin{proof}[Proof of Claim] The first statement follows from the structure of the free amalgam $D:=B\sqcup_{A} C$. But then, since $\widehat{A}\leq B,C$, if $\{e_1,e_2\}\cap \{ v_1[a],v_2[a],v_3[a],v_4[a]\}= \emptyset$, the $3$-cycles $(a,e_1,e_2)$, $(a,v_1[a],v_2[a])$, $(a,v_3[a],v_4[a])$ would form the three wings of a copy of $H$ in either $B$ or $C$. This would contradict $B\in\mathcal{K}(H)$ or $C\in\mathcal{K}(H)$. \end{proof} Let $(b_1,b_2,a), (c_1,c_2,a), (d_1,d_2,a)$ be the three cycles of length $3$ in $D$, which correspond to the three wings of $j(H)$. By the first statement of Claim \ref{Claim3} we may assume without loss of generality that $b_1,b_2\in B$ and $c_1,c_2,d_1,d_2\in C$. By the second statement of Claim \ref{Claim3} we may also assume that $b_1=v_1[a]$. It follows that $b_2\not\in \widehat{A}$, since otherwise, the wing $(b_1,b_2,a)$ would lie in $C$, contradicting that $C\in \mathcal{K}(H)$. By Claim \ref{Claim2} we can find some vertex $\hat{w}$ of $\widehat{A}$, so that $\hat{w} R v_1[a]$, $\hat{w} R a$, and $\hat{w}\not\in \{v_1[a],v_2[a],v_3[a],v_4[a]\}$. Notice that the three cycle $(b_1,\hat{w},a)$ lies entirely in $\widehat{A}$ and hence in $C$. Consider the map $j'\colon H\to C$ which is defined as $j$, with the only exception that $j'(j^{-1}(b_2))=w$. Clearly $j'$ is a homomorphism mapping the three wings of $H$ to the cycles $(b_1,\hat{w},a), (c_1,c_2,a), (d_1,d_2,a)$. We will show that $j'$ is also injective, which will conclude the proof by contradicting that $C\in \mathcal{K}(H)$. It suffices to show that $\hat{w}\not\in \{c_1,c_2,d_1,d_2\}$. First notice that one of $\{c_1,c_2,d_1,d_2\}$, say $c_2$, is not in $\widehat{A}$, since otherwise $j(H)$ would lie entirely in $B$, contradicting that $B\in \mathcal{K}(H)$. By Claim \ref{Claim3} and since $\{c_1,c_2\}$ is disjoint from $\{b_1\}=\{v_1[a]\}$ we have that $c_1\in \{ v_2[a],v_3[a],v_4[a]\}$. In particular we have that $\hat{w}\not\in \{c_1,c_2\}$. Notice now that if either $d_1$ or $d_2$ is not in $\widehat{A}$, then by repeating the same argument we would also have that $\hat{w}\not\in \{d_1,d_2\}$. So we may assume that the wing $(d_1,d_2,a)$ lies in $\widehat{A}$. By Claim \ref{Claim3}, we may assume without loss of generality that $d_1\in \{ v_2[a],v_3[a],v_4[a]\} \setminus \{c_1\}$. We have the following cases to consider: \smallskip{} \noindent{}{\bf Case 1.} Assume $d_1= v_2[a]$. In this case $\hat{w}\neq d_1$, since $\hat{w}\not\in \{v_1[a],v_2[a],v_3[a],v_4[a]\}$. Moreover, we have that $\hat{w}\neq d_2$, since otherwise we would have the $3$-cycle $(\hat{w},d_1,a)$ in $\widehat{A}$, which, together with the the $3$-cycles $(b_1,b_2,a)$ and $(v_3[a],v_4[a],a)$, would form the three wings of a copy of $H$ in $B$, contradicting that $B\in \mathcal{K}(H)$. \smallskip{} \noindent{}{\bf Case 2.} Assume $d_1= v_3[a]$. In this case $\hat{w}\neq d_1$, since $\hat{w}\not\in \{v_1[a],v_2[a],v_3[a],v_4[a]\}$. Assume towards contradiction that $\hat{w}=d_2$. We have the $3$-cycle $(\hat{w},d_1,a)$ in $\widehat{A}$. \smallskip{} \noindent{}{\it Case 2(i).} If $c_1=v_2[a]$, then $(c_1,c_2,a)$, together with the $3$-cycles $(v_3[a],v_4[a],a)$ and $(v_1[a],\hat{w},a)$, would form a copy of $H$ in $C$, contradicting that $C\in \mathcal{K}(H)$. \smallskip{} \noindent{}{\it Case 2(ii).} If $c_1=v_4[a]$, then $(\hat{w},d_1,a)$, together with the $3$-cycles $(c_1,c_2,a)$ and $(v_1[a],v_2[a],a)$, would form a copy of $H$ in $C$, contradicting that $C\in \mathcal{K}(H)$. \smallskip{} \noindent{}{\bf Case 3.} Assume $d_1= v_4[a]$. Same as Case 2. \smallskip{} This concludes the proof of $\hat{w}\not\in \{c_1,c_2,d_1,d_2\}$, yielding an injective homomorphism $j'\colon H\to C$, which contradicts that $C\in \mathcal{K}(H)$. Hence $D\in\mathcal{K}(H)$. \end{proof} We have established that, if $\mathcal{C}=\mathrm{lim}(\mathcal{K})$ has a universal element, then $\mathcal{K}$ satisfies WAP and that the converse is not true. This suggest the following question: \begin{question} Is there an amalgamation property for $\mathcal{K}$, strictly lying between AP and WAP, which characterizes when $\mathcal{C}=\mathrm{lim}(\mathcal{K})$ admits a universal element?\end{question} \section{$C_4$-free graphs have no generic element} In this section we prove Theorem \ref{T:2}. Let $\mathcal{K}(C_4)$ be the collection of all finite graphs which omit the graph $C_4$. That is, all graphs $A$ for which there is no injective homomorphism $f\colon C_4\to A$ from the $4$-cycle $C_4$. By \cite[Theorem 4.2.2]{K}, see also \cite[Theorem 2.5]{KT}, it suffices to show that $\mathcal{K}(C_4)$ does not have (WAP). We will need the following proposition. Let $v$ be a vertex of a graph $A$. We say that $v$ is {\bf dominating} if there is an edge between $v$ and any vertex $w$ of $A\setminus \{v\}$. \begin{proposition}\label{prop:Blokhuis} If $A\in \mathcal{K}(C_4)$ is of diameter $2$ without dominating vertex, then there is an edge in $A$ not contained in any $3$-cycle. \end{proposition} \begin{proof} Let $A$ be a $C_4$-free graph of diameter $2$ without dominating vertex. Then by \cite{BB} we are in one of two cases:\begin{enumerate} \item the graph $A$ is strongly regular, i.e. there are integers $k, \lambda, \mu$ such that every vertex has valency $k$, every two adjacent vertices have $\lambda$ common neighbours and every two non-adjacent vertices have $\mu$ common neighbours. Note that we have $\mu=1$. \item the graph $A$ has two valencies, and in this case there are edges not contained in a triangle (expressed in \cite{BB} as saying there are maximal cliques having $2$ elements). \end{enumerate} In case (1) above, we have $\lambda\neq 1$ by \cite{DF}, and so in this case $A$ contains no 3-cycles. Thus, any $C_4$-free graph of diameter 2 without dominating vertex has an edge not contained in a triangle. \end{proof} We can now turn to the proof of Theorem \ref{T:2}. \begin{proof}[Proof of Theorem \ref{T:2}] Let $A=C_5\in \mathcal{K}(C_4)$ be the pentagon and let $i\colon A\to \widehat{A}$ be any embedding with $\widehat{A}\in \mathcal{K}(C_4)$. We will define $f\colon \widehat{A}\to B$ and $g\colon \widehat{A}\to C$, with $B,C\in\mathcal{K}(C_4)$, which do not amalgamate over $A$. Let $E\in \mathcal{K}(C_4)$ be any extension $A\leq E$ of $A$ with $E\leq \widehat{A}$. A vertex $v\in E$ is {\bf determined over $A$} if it belongs to the smallest set of vertices $X\subseteq\mathrm{dom}(E)$ so that $\mathrm{dom}(A)\subseteq X$, satisfying: if $v\in \mathrm{dom}(E)$ and there are $x,y\in X$ with $E\models (v R x)$ and $E\models (v R y)$, then $v\in X$. \begin{claim}\label{C:3} If $f,g\colon E\to D$ are two embeddings with $D\in \mathcal{K}(C_4)$, so that $f(a)=g(a)$ for all $a\in \mathrm{dom}(A)$, then $f(v)=g(v)$ for every determined $v\in\mathrm{dom}(E)$. \end{claim} \begin{proof}[Proof of Claim \ref{C:3}] Assume that $f(x)=g(x)$ and $f(y)=g(y)$ for some $x,y\in\mathrm{dom}(E)$. If $E\models (v R x)$ and $E\models (v R y)$, then $f(v)=g(v)$, since otherwise we would have the $4$-cycle $f(x),f(v),f(y),g(v)$ in $E$. \end{proof} \begin{claim}\label{C:4} There is some $E\in \mathcal{K}(C_4)$ with $\widehat{A}\leq E$, and some geodesic path $x,x',y',y$ of length $3$ in $E$, consisting of vertices which are determined over $A_0$. \end{claim} \begin{proof}[Proof of Claim \ref{C:4}] Let $D$ be the subgraph of $\widehat{A}$ induced on the set of all vertices of $\widehat{A}$ which are determined over $A$. Notice that any path of length $\leq 3$ consisting of vertices which are determined over $A$ is geodesic in $D$ if and only if it is geodesic in $A$. Hence, if the diameter of $D$ is larger than $2$ then we take $E=\widehat{A}$. Otherwise, we may list $\{(x_i,y_i)\colon i\in I\}$ all edges $(x_i,y_i)$ of $D$ for which there is no $v\in \mathrm{dom}(D)$ so that $x_i,y_i,v$ is a $3$-cycle in $D$. Notice that any such vertex $v$ is determined over $A$ and hence $(x_i,y_i)$ is not contained in a $3$-cycle, even in the ambient graph $\widehat{A}$. We now extend $\widehat{A}$ by adding for every $i\in I$ a new vertex $v_i$ which we connect by an edge only with $x_i$ and with $y_i$. The resulting graph $E$ is clearly in $ \mathcal{K}(C_4)$ and the set of all vertices of $E$ which are determined over $A$ is $\mathrm{dom}(D)\cup\{v_i \colon i\in I\}$. The subgraph $D'$ of $E$ which is induced on the latter set of vertices has no edges which are not contained in a $3$-cycle. Moreover, it has no dominating vertex since it contains $A$. By Proposition \ref{prop:Blokhuis} there is a geodesic path $x,x',y',y$ of length $3$ in $D'$. This path remains geodesic in $E$. \end{proof} Let now $E\supseteq \widehat{A}$ and let $x,x',y',y\in \mathrm{dom}(E)$ as in Claim \ref{C:4}. We extend $E$ to $E'$ by first adding a new vertex $z$ which is connected only to $x$ and $y$. We then add a vertex $w_y$ which is connected only to $x$ and $z$ and a vertex $w_x$ which is connected only to $y$ and $z$. We can also assume without loss of generality that there are $v_x,v_y\in \mathrm{dom}(E)$ which are determined over $A$, so that $xRv_x R x'$ and $yRv_y R y'$. Indeed the construction in Claim \ref{C:4} guarantees that such $v_x,v_y$ exist. Notice that all vertices $v_x,v_y,w_x,w_y\in \mathrm{dom}(E')$ are determined over $A$. We define the extension $B$ of $E'$ by adding a new vertex $s$ that is connected only with $w_x$ and $v_x$ and a new vertex $t$ that is connected only with $w_y$ and $v_y$. We also connect $s$ and $t$ by an edge. We have that $B\in\mathcal{K}(C_4)$. This follows from the fact that $E'\models \neg (v_x R v_y)$ which holds since otherwise we would already have a $4$-cycle $x',y',v_y,v_x$ in $E$. The definition of the extension $C$ of $E'$ similar to $B$ with the only difference that instead of connecting $s$ and $t$ with an edge, we connect them using a new path of length $3$. \medskip{} \begin{figure}[ht!] \hspace*{-9em}{ \begin{tikzpicture}[x=0.68pt,y=0.75pt,yscale=-1,xscale=1] \path (0,300); \draw (203,200) node (1){$\bullet$} -- (245,60) node (2){$\bullet$} -- (358,60) node (3){$\bullet$} -- (400,200) node (4){$\bullet$} -- cycle ; \draw (224,131) node (5){$\bullet$} -- (258,199) node (6){$\bullet$} ; \draw (379,131) node (7){$\bullet$} -- (343,200) node (8){$\bullet$}; \draw (304,60) node (9){$\bullet$} -- (224,131); \draw (304,60) -- (379,131) ; \draw (245,60) .. controls (351,27) and (385,29) .. (450,60) node (10){$\bullet$} ; \draw (450,60) .. controls (449,114) and (414,176) .. (400,200) ; \draw (152,61) .. controls (192,31) and (298,30) .. (358,60) ; \draw (152,61) node (11){$\bullet$} .. controls (151,120) and (182,172) .. (200,200) ; \draw (298,65) node [anchor=north west][inner sep=0.75pt] {$z$}; \draw (385,122) node [anchor=north west][inner sep=0.75pt] {$y$}; \draw (208,122) node [anchor=north west][inner sep=0.75pt] {$x$}; \draw (260,202) node [anchor=north west][inner sep=0.75pt] {$x'$}; \draw (345,203) node [anchor=north west][inner sep=0.75pt] {$y'$}; \draw (200,205) node [anchor=north west][inner sep=0.75pt] {$v_{x}$}; \draw (394,205) node [anchor=north west][inner sep=0.75pt] {$v_{y}$}; \draw (361,53) node [anchor=north west][inner sep=0.75pt] {$w_{x}$}; \draw (220,53) node [anchor=north west][inner sep=0.75pt] {$w_{y}$}; \draw (140,47) node [anchor=north west][inner sep=0.75pt] {$s$}; \draw (454,44) node [anchor=north west][inner sep=0.75pt] {$t$}; \end{tikzpicture} } \vspace{-55pt} \caption{} \end{figure} We claim that $B,C$ do not amalgamate over $A$. Indeed, if $f'\colon B\to D$ and $g'\colon C\to D$ where embeddings with $f\mathord{\upharpoonright} A= g\mathord{\upharpoonright} A$, then by Claim \ref{C:4} we have that $f(s)=g(s)$ and $f(t)=g(t)$ as well. But this would yield a $4$-cycle in $D$. \end{proof} Cherlin and Komj\'{a}th \cite{CK} extended the results from \cite{HP} by showing that there is no universal element in the class $\mathcal{C}(C_k)$ of all countable $C_k$-free graphs, for all $k>3$. \begin{question} Does $\mathcal{C}(C_k)$ admit a generic element for some $k>4$? \end{question}

6b866cd261b35aab858c67a499039e96d9b20d76

\section{Introduction} In this paper we are concerned with the equilibrium problem stated as \begin{equation} \text{Find } x^* \in C \text{ such that } f(x^*,y) \geq 0 \text{ for all } y \in C, \tag{EP} \end{equation} in which $C$ is a nonempty closed convex subset in a Hilbert space $\mathcal{H}$ endowed with an inner product $\langle \cdot, \cdot \rangle$ and the induced norm $\| \cdot \|$, and $f: \mathcal{H} \times \mathcal{H} \to \mathbb{R} \cup \{+\infty\}$ is a bifunction such that $f(x, y) < +\infty$ for every $x, y \in C$. The inequality in Problem (EP) was first used in \cite{NI1955} for convex noncooperative game theory. The first result on solution existence of (EP) is due to K. Fan \cite{F1972}, where this problem was called a minimax inequality. The name equilibria was first used in \cite{MO1992}. After the appearance of the paper by Blum and Oettli \cite{BO1994}, the problem (EP) has attracted much attention of many authors and a lot of algorithms have been developed for solving the problem where the bifunction $f$ have monotonic properties. These algorithms are based upon different methods such as penalty and gap functions \cite{B1P2012, B2P2015, B5P2016, K2D2003, K4D2003, K3D2003, M2003, MO1992}, regularization \cite{AA2020, HM2011, LA2013, MQ2009, MQ2010}, extragradient methods \cite{B6P2017, HSM2020, K5S2005, QAM2012, SS2011, SS2015, SSA2011, TH2017, VM2019, VSN2012, YM2020}, splitting technique \cite{AH2017,DM2016,ML2018}. A comprehensive reference-list on algorithms for the equilibrium problem can be found in the interesting monograph \cite{BCPP2019}. An interesting of this problem is that, despite its simple formulation, it contains many problems such as optimization, reverse optimization, variational inequality, minimax, saddle point, Kakutani fixed point, Nash equilibrium problems, and some others as special cases (see the interesting monographs \cite{BCPP2019, KR2018} and the papers \cite{BO1994, MO1992}). In what follows we always suppose that $\phi: C \times C \to \mathbb{R}$ such that $\phi(x,\cdot)$ is convex for any $x \in C$, and $\varphi: C \to \mathbb{R}$ is a convex function on $C$. For continuity (resp. lower and upper continuity) of $\phi$ and $\varphi$ we means the continuity (reps. lower and upper continuity) with respect to the set $C$. Then we consider Problem (EP) with $f(x,y) :=\phi(x,y) + \varphi(y) - \varphi(x)$. In this case Problem (EP) becomes a mixed equilibrium problem of the form \begin{equation} \text{Find } x^* \in C \text{ such that } f(x^*,y):= \phi(x^*,y) + \varphi(y) - \varphi(x^*) \geq 0 \text{ for all } y \in C. \tag{MEP} \end{equation} By considering this mixed form one can employ special structures of each $\phi$ and $\varphi$ in subgradient splitting algorithms, where the bifunction $f(x,y)$ can be expressed by the sum of two bifunctions $f_1(x,y) + f_2(x,y)$ and the iterates are defined by taking the proximal mappings of each $f_1$ and $f_2$ separately, see \cite{AH2017,AA2020,DM2016,MQ2010,ML2018}. The first fixed point approach to equilibrium problem (EP) was first developed in 1972 by K. Fan in \cite{F1972}. There, by using the KKM lemma, it has been proved that if $C$ is compact and $f(x,\cdot)$ is quasiconvex on $C$, then Problem (EP) admits a solution under a certain continuity property of $f$. Note that in this {result of K. Fan, it does not require any monotonicity of the bifunction $f$. A direct proof using the Kakutani fixed point theorem for the solution existence of Problem (EP) is based upon the mapping $K$ defined by taking, for each $x\in C$, \begin{equation} K(x):= \text{argmin} \{ f(x, y): y \in C\}. \tag{P1} \end{equation} Clearly, if $f(x,x) = 0$ for every $x\in C$, then $x^*$ is a solution to (EP) if and only if it is a fixed point of $K$, i.e., $x^* \in K(x^*)$. Thus, if $C$ is convex, compact, $f(x,\cdot)$ is convex on $C$ and $K$ is upper semicontinuous on $C$, then by the well known Kakutani fixed point theorem, the mapping $K$ has a fixed point. It can be noticed that the mapping $K$ is set-valued in general. In order to avoid multivalues of $K$, an auxiliary principle has been used by defining the proximal mapping \begin{equation} B_{\lambda}(x) := \text{argmin} \left\{\lambda f(x,y) + \frac{1}{2} \langle y-x, G(y-x)\rangle \right\}, \tag{P2} \end{equation} where $\lambda > 0$ and $G$ is a self-adjoint positive linear bounded operator from $\mathcal{H}$ into itself. In the sequel, for simplicity of the presentation, we always suppose that $G$ is the identity operator. It is well known \cite{BV2004} that if $f(x,\cdot)$ is convex and subdifferentiable on $C$, Problem (P2) is uniquely solvable even for the case $C$ is not compact. Moreover, a point $x^* \in C$ is a solution of Problem (EP) if and only if $x^*$ is a fixed point of $ B_{\lambda}$ for any $\lambda > 0$. So the solution existence of (EP) can be proved by using the Brouwer fixed point theorem whenever $C$ is compact and $B_\lambda$ is continuous on $C$. These results suggest that one can apply the existing algorithms such as ones based on the Scaft pivoting method \cite{KM1975} for computing a fixed point of the mapping $B_{\lambda}$, thereby solving the equilibrium problem (EP). However, the computational results \cite{KM1975,TTM1978} show that the pivoting methods are efficient only for problems with moderate size. Since in the fixed point theory, iterative methods for computing a fixed point have been successfully applied to contractive, generalized contractive, and nonexpansive mappings, a natural question arises that under which conditions, the mapping $B_{\lambda}$ possesses certain contraction or generalized nonexpansiveness properties? This is a survey paper, but it also contains some new results on a fixed point approach to equilibrium problem (MEP). Namely, first we outline results on quasicontraction and contraction of the Moreau proximal mapping when the bifunction $f$ is strongly monotone and satisfies a certain Lipschitz-type condition. Next, in the case $f$ is not necessarily strongly monotone, but monotone, we present some results on approximate nonexpansiveness of the proximal mapping for monotone equilibrium problems satisfying a certain strongly Lipschitz-type condition. Finally, we show a result on quasinonexpansiveness of a composed proximal mapping defined by the equilibrium problem. This relationship allows that the equilibrium problem can be solved by the existing methods in the fixed point theory (see e.g. \cite{C2013, GD1997, I1974, M1998, TT2007, TX1993} and the references theirein). The paper is organized as follows. The next section contains preliminaries on the equilibrium problem under consideration and on generalized contractions in real Hilbert spaces. In Section \ref{ContractionGNP} we present some results on contraction, quasicontraction, nonexpansiveness, and approximate nonexpansiveness of the Moreau proximal mapping defined for the equilibrium problem. We close the paper by some conclusions in Section \ref{ConclusionSection}. \section{Preliminaries} The following definitions for a bifunction is commonly used in the literature, see e.g. \cite{BCPP2019}. \begin{defn}\label{Sect2Defn1} A bifunction $f:C\times C \to \mathbb{R}$ is said to be (i) strongly monotone with modulus $\gamma >0$ (shortly $\gamma$-strongly monotone) on $S \subseteq C$ if \begin{equation*} f(x,y) + f(y,x) \leq -\gamma \|x-y\|^2 \ \forall x,y \in S; \end{equation*} (ii) monotone on $S \subseteq C$ if \begin{equation*} f(x,y) + f(y,x) \leq 0 \ \forall x, y \in S; \end{equation*} (iii) strongly pseudomonotone on $S \subseteq C$ with modulus $\gamma > 0$ (shortly $\gamma$- strongly pseudomonotone) if \begin{equation*} f(x,y) \geq 0 \Rightarrow f(y,x) \leq -\gamma \|x-y\|^2 \ \forall x, y \in S; \end{equation*} (iv) pseudomonotone on $S \subseteq C$ if \begin{equation*} f(x,y) \geq 0 \Rightarrow f(y,x) \leq 0 \ \forall x, y \in S. \end{equation*} \end{defn} The notions on monotonicity properties of a bifunction are generalized ones for operators. In fact, it is easy to see that when $f(x,y) := \langle F(x), y-x\rangle +\varphi (y) - \varphi (x)$, then $f$ is $\gamma$-strongly monotone (resp. monotone, $\gamma$-strongly pseudomonotone, pseudomonotone) if and only if $F$ is $\gamma$-strongly monotone (resp. monotone, $\gamma$-strongly pseudomonotone, pseudomonotone). The following Lipschitz-type conditions has been introduced in \cite{M2003} and commonly used for Problem (EP). \begin{defn}\label{Sect2Defn2} A bifunction $f: C\times C \to \mathbb{R}$ is said to be of {\it Lipschitz-type} on $S \subseteq C$ if there exists constants $L_1, L_2 >0$ such that \begin{equation*} f(u,v) + f(v,w) \ge f(u, w) - L_1 \|u-v\|^2 - L_2\|v-w\|^2 \ \forall u, v, w \in S. \end{equation*} \end{defn} By taking $u=v=w$ we see that $f(u, u) \geq 0$, so if, in addition, $f$ is pseudomonotone, then $f(u,u) = 0$. The following concepts are well-known in the fixed point theory (see e.g. \cite{AOS2009}). \begin{defn}\label{Sect2Defn3} Let $T: \mathcal{H} \to C$. (i) $T$ is said to be {\it contractive} on $C$ if there exists $0 < \rho < 1$ such that \begin{equation*} \|T(x) - T(y)\| \leq \rho \|x-y\| \ \forall x, y \in C. \end{equation*} If $T$ satisfies this condition with $\rho = 1$, then it is said to be {\it nonexpansive}. It is said to be {\it quasicontractive} on $C$ if \begin{equation*} \|T(x) - T(y) \| \leq \rho \|x-y\| \ \forall x \in {\rm Fix}(T), y \in C, \end{equation*} where ${\rm Fix}(T)$ stands for the set of fixed points of $T$. If this condition holds for $\rho = 1$, then $T$ is said to be {\it quasinonexpansive}. (ii) $T$ is said to be {\it firmly nonexpansive on $C$} if \begin{equation*} \|T(x) - T(y) \|^2 \leq \|x-y\|^2- \|(I-T)(x) - (I-T)(y)\|^2 \ \forall x, y \in C. \end{equation*} (iii) $T$ is said to be {\it $\rho$-strongly converse monotone or $\rho$-cocoercive on $C$} with $\rho > 0$, if \begin{equation*} \langle T(x) -T(y), x-y\rangle \geq \rho \|T(x) -T(y)\|^2 \ \forall x, y \in C. \end{equation*} \end{defn} \section{Contraction and generalized nonexpansive properties of the proximal mapping}\label{ContractionGNP} Let $g: C \to \mathbb{R}$ be a convex function and $\lambda > 0$. The proximal mapping $P_{\lambda}$ with respect to $C, g, \lambda$ (shortly proximal mapping) is defined as follows (see e.g. \cite{RW1998}): \begin{equation*} P_{\lambda}(x) := \text{argmin}\left\{ \lambda g(y) + \frac{1}{2}\|y-x\|^2 : y \in C\right\}. \end{equation*} For the bifunction $f$ where $f(x,\cdot)$ is convex and finite on $C$, the proximal mapping $B_{\lambda}$ is defined by taking \begin{equation*} B_{\lambda}(x) := \text{argmin} \left\{ \lambda f(x,y) + \frac{1}{2} \|y-x\|^2 : y \in C\right\} \end{equation*} for each $x \in C$. Note that when either $f(x,\cdot)$ is continuous on $C$ or $C$ has an interior point, and $f(x,x) = 0$, then it is well known from \cite{BCPP2019} that $x^*$ is a fixed point of $B_{\lambda}$ if and only if it is a solution to Problem (EP). The following theorem says that when $f$ is strongly monotone and satisfies the Lipschitz-type on $C$, then one can choose a regularization parameter such that the proximal mapping is quasicontractive on $C$. For applying optimality condition for the problem defining the proximal mapping, we always assume that either $C$ has an interior point or, for any $x \in C$, $\phi(x,\cdot)$ is continuous with respect to $C$ at a point of $C$. \begin{thm}\label{T1} Suppose that $f$ is strongly monotone on $C$ with modulus $\tau$ and satisfies the Lipschitz-type condition with constants $L_1, L_2$ satisfying $L_1 + L_2 > \tau$. Then, the proximal mapping $B_{\lambda}$ is quasicontractive on $C$, namely \begin{equation*} \|B_{\lambda} (x) - x^*\| \le \sqrt{\alpha} \|B_{\lambda}(x)- x^*\| \ \forall x \in C, x^* \in {\rm Fix}(B_{\lambda}), \end{equation*} whenever $\lambda \in (0, \frac{1}{2L_2})$, where $\alpha := 1 -2\lambda (\tau-L_1) > 0$. \end{thm} \begin{proof} The following proof borrows some techniques from the one in \cite{QAM2012}. For simplicity of notation we let \begin{equation*} f_x(y): = \lambda f(x,y) + \frac{1}{2} \|y-x\|^2. \end{equation*} Since $\lambda > 0$ and $f(x, \cdot)$ is convex on $C$ by assumption, $f_x$ is strongly convex with modulus 1. As defined, $B_{\lambda}(x)$ is a minimizer of $f_x(\cdot)$ over the closed convex set $C$. Therefore, we have \begin{equation*} f_x(B_{\lambda}(x)) + \frac{1}{2} \|B_{\lambda}(x) - x^*\|^2 \le f_x(x^*), \end{equation*} that is \begin{equation*} \lambda f(x, B_{\lambda}(x)) + \frac{1}{2} \|B_{\lambda}(x) - x\|^2 + \frac{1}{2} \|B_{\lambda}(x) - x^*\|^2 \le \lambda f(x, x^*) + \frac{1}{2} \|x^* - x\|^2, \end{equation*} or equivalently \begin{equation}\label{1} \|B_{\lambda}(x) - x^*\|^2 \leq 2\lambda \left( f(x,x^*) - f(x, B_{\lambda}(x)) \right) + \|x-x^*\|^2 - \|B_{\lambda}(x) - x\|^2. \end{equation} Since $f$ is strongly monotone on $C$ with modulus $\tau$, it follows from (\ref{1}) that \begin{align} & \ \|B_{\lambda}(x) - x^*\|^2 \notag\\ \le & \ 2\lambda (-\tau \|x - x^*\|^2 - f(x^*, x) - f(x, B_{\lambda}(x))) + \|x-x^*\|^2 -\|B_{\lambda}(x) - x\|^2 \notag\\ \le & \ (1 - 2\lambda\tau) \|x - x^*\|^2 - 2\lambda \left(f(x^*, x) + f(x, B_{\lambda}(x)) \right) - \|B_{\lambda}(x) - x\|^2.\label{ev2} \end{align} Since $f$ satisfies Lipschitz-type condition, we have \begin{equation*} f(x^*, x) + f(x, B_{\lambda}(x)) \ge f(x^*, B_{\lambda}(x)) - L_1 \|x^* - x\|^2 - L_2 \|x - B_{\lambda}(x) \|^2. \end{equation*} Therefore, it follows from (\ref{ev2}) that \begin{align*} & \ \|B_{\lambda}(x) - x^*\|^2\\ \le & \ (1 - 2\lambda(\tau - L_1)) \|x - x^*\|^2 - (1 - 2\lambda L_2) \|x - B_{\lambda}(x)\|^2 - 2\lambda f(x^*, B_{\lambda}(x))\\ \le & \ (1 - 2\lambda(\tau - L_1)) \|x - x^*\|^2. \end{align*} The last inequality is due to the fact that $f(x^*, B_{\lambda}(x)) \ge 0$ and the assumption that $0 \le \lambda \le \frac{1}{2L_2}$. Since $\tau < L_1 + L_2$, we see that if $0< \lambda < \frac{1}{2L_2}$, then $1-2\lambda(\tau-L_1) < 1$ and hence $B_{\lambda}$ is quasicontractive. \end{proof} Theorem \ref{T1} allows that the contraction iterative method can be used for solving equilibrium problem (EP). A question arises here: under which condition, the proximal mapping is contractive? The following theorem in \cite{H2017} gives an answer for this question. For the statement of the theorem, first we recall from \cite{H2017} that a bifunction $f: C\times C \to \mathbb{R}$ is said to be {\it strongly Lipschitz-type} on $C$ if there exist $\alpha_i: C \times C \to C$, $\beta_i : C\to C$, $K_i, L_i > 0$ (with $i=1,\ldots,p$) such that \begin{equation*} f(x,y) + f(y,z) \ge f(x,z) + \sum_{j=1}^p \langle \alpha_i(x,y) , \beta_i(y-z) \rangle \ \forall x,y, z \in C, \end{equation*} where \begin{align*} \|\beta_i(x) - \beta_i(y)\| &\leq K_i\|x-y\| &&\forall x, y \in C, i = 1, \ldots, p,\\ \|\alpha_i(x,y)\| &\leq L_i\|x-y\| &&\forall x, y \in C, i = 1, \ldots, p,\\ \alpha_i(x,y) + \alpha_i(y,x) &= 0 &&\forall x, y \in C, i = 1, \ldots, p. \end{align*} As also remarked in \cite{H2017}, the following facts are not hard to see. (i) If $f$ is strongly Lipschitz-type on $C$, then it is Lipschitz-type on $C$ with both constants $\frac{1}{2} \sum_{i=1}^p K_i L_i$. (ii) If $f(x,y) = \langle F(x), y-x \rangle + \varphi(y) - \varphi(x)$, then $f$ is strongly Lipschitz-type if and only if $F$ is Lipschitz. \begin{thm}\label{T2} Let $C$ a be nonempty closed convex set, $f: C\times C\to \mathbb{R}$. Suppose that $f(x,\cdot)$ is lower semicontinuous, convex, $\gamma$-strongly monotone, and strongly Lipschitz-type on $C$. Then the proximal mapping $B_{\lambda}$ is contractive on $C$ whenever $\lambda \in (0, \frac{2\gamma}{M})$ with $M = \sum_{i=1}^p K_i L_i.$ \end{thm} In the case of mixed variational inequality, when $f(x,y) := \langle F(x), y-x \rangle + \varphi(y) - \varphi(x)$ with $F$ being Lipschitz continuous and strongly monotone, the proximal mapping is contractive (see. e.g. \cite{AMNS2005}). This result also follows from the above theorem due to the fact that $f(x,y) := \langle F(x), y-x\rangle + \varphi (y) - \varphi (x)$ is strongly Lipschitz-type on $C$ whenever $F$ is Lipschitz on $C$. In case that $F$ is co-coercive (strongly inverse monotone), the proximal mapping is nonexpansive on $C$ as stated in the following theorem. \begin{thm}\label{T3} Suppose that $f(x,y) := \langle F(x), y-x \rangle + \varphi(y) - \varphi(x)$ with $F$ being $\delta$-co-coercive (or strongly inverse monotone) on $C$. Then, whenever $0 < \lambda \leq \frac{1}{2\delta}$, the proximal mapping $B_{\lambda}$ is nonexpansive on $C$. \end{thm} \begin{proof} From the definition of $B_{\lambda}$, by using the optimization condition, it is easy to see that \begin{equation}\label{4} \|B_{\lambda}(x) - B_{\lambda}(y)\|^2 \leq \|x-y- \lambda ( F(x) - F(y) )\|^2 \ \forall x,y \in C. \end{equation} Since $F$ is co-coercive on $C$ with modulus $\delta$, we have \begin{equation*} \langle x-y, F(x) - F(y)\rangle \geq \delta\|F(x) - F(y)\|^2, \end{equation*} which implies \begin{equation*} \| x-y - \lambda (F(x) - F(y))\| \leq \|x-y\|. \end{equation*} Thus by (\ref{4}), we have \begin{equation*} \|B_{\lambda}(x) - B_{\lambda}(y) \| \leq \|x-y\|. \end{equation*} \end{proof} A question now may arise that is the proximal mapping $B_{\lambda}$ nonexpansive when $f$ is monotone? The following simple example gives a negative answer. Let us consider the linear variational inequality \begin{equation}\label{VI} \text{Find} \ x \in \mathbb{R}^2 \text{ such that } f(x, y) := \langle Ax, y-x \rangle \geq 0 \text{ for all } y \in \mathbb{R}^2,\tag{VI} \end{equation} where \begin{equation*} A = \begin{bmatrix} 0 & 1\\ -1 & 0 \end{bmatrix}. \end{equation*} For all $x, y \in \mathbb{R}^2$ we have \begin{align*} f(x, y) + f(y, x) &= \langle A(x-y), y - x \rangle\\ &= (x_2 - y_2)(y_1 - x_1) - (x_1 - y_1)(y_2 - x_2)\\ &= 0, \end{align*} therefore $f$ is monotone on $\mathbb{R}^2$. It is easy to see that $x^* = (0, 0)^t$ is a solution to the variational inequality (VI), since $f(x^*, y) = 0$ for all $y \in \mathbb{R}^2$. Furthermore, $x^*$ is the unique solution to (VI). Indeed, if $\overline{x} = (\overline{x}_1, \overline{x}_2)^t$ is a solution to (VI), then \begin{equation*} \langle A\overline{x}, y - \overline{x} \rangle \ge 0 \quad \forall y \in \mathbb{R}^2. \end{equation*} By taking $y = \overline{y} = (\overline{x}_1 - \overline{x}_2, \overline{x}_1 + \overline{x}_2)^t$, we have \begin{equation*} 0 \le \langle A\overline{x}, \overline{y} - \overline{x} \rangle = \left\langle \begin{bmatrix} \overline{x}_2 \\ -\overline{x}_1 \end{bmatrix}, \begin{bmatrix} -\overline{x}_2 \\ \overline{x}_1 \end{bmatrix} \right\rangle = -(\overline{x}_1^2 + \overline{x}_2^2) \le 0, \end{equation*} which implies $\overline{x} = (0, 0)^t = x^*$. Now we see that \begin{align*} & \lambda \langle Ax,y-x\rangle + \frac{1}{2}\|y-x\|^2 \\ = & \lambda \left\langle \begin{bmatrix} x_2 \\ -x_1 \end{bmatrix}, \begin{bmatrix} y_1 - x_1 \\ y_2 - x_2 \end{bmatrix} \right\rangle + \frac{1}{2} \left( (y_1 - x_1)^2 + (y_2 - x_2)^2 \right)\\ = & \lambda (x_2 y_1 - x_1 y_2) + \frac{1}{2} \left( (y_1 - x_1)^2 + (y_2 - x_2)^2 \right) \\ = & \frac{1}{2} \left(y_1^2 - 2(x_1 - \lambda x_2)y_1 + y_2^2 - 2(x_2 + \lambda x_1) y_2 + x_1^2 + x_2^2 \right)\\ = & \frac{1}{2} \left( (y_1 - x_1 + \lambda x_2)^2 + (y_2 - x_2 - \lambda x_1)^2 - \lambda^2 (x_1^2 + x_2^2) \right) \end{align*} which attains its minimum at $(y_1, y_2) = (x_1 - \lambda x_2, x_2 + \lambda x_1)$. We obtain the following explicit formula for the proximal mapping of (VI): \begin{equation*} B_{\lambda}(x) = \text{argmin} \{\lambda \langle Ax,y-x\rangle + \frac{1}{2}\|y-x\|^2 \mid y \in \mathbb{R}^2\} = \begin{bmatrix} x_1 - \lambda x_2 \\ x_2 + \lambda x_1 \end{bmatrix}. \end{equation*} Therefore, for any $\lambda > 0$ we have \begin{equation*} \|B_{\lambda}(x) - B_{\lambda}(x^*)\| = \left\| \begin{bmatrix} x_1 - \lambda x_2 \\ x_2 + \lambda x_1 \end{bmatrix} - \begin{bmatrix} 0 \\ 0 \end{bmatrix}\right\| = \sqrt{1 + \lambda^2} \sqrt{x_1^2 + x_2^2} > \|x - x^*\|, \end{equation*} which proves that $B_{\lambda}$ is not nonexpansive for any $\lambda > 0$. So in general the proximal mapping may not be nonexpansive even for the variational inequality when $f(x,y) = \langle F(x), y-x \rangle$ with $F$ being Lipschitz and monotone. However, it is well known from \cite{CH2005, TT2007} that if $f$ is monotone, $f(x, \cdot)$ is convex, lower semicontinuous, and $f(\cdot,y)$ is hemicontinuous, then the regularization proximal mapping $R_{\lambda}$ is defined everywhere, single valued, and firmly nonexpansive for any $\lambda > 0$. Here, for each $x \in C$, $R_{\lambda}(x)$ is defined as the unique solution of the strongly monotone equilibrium problem \begin{equation*} \text{Find } z \in C \text{ such that } f(z,y) + \frac{1}{2\lambda} \langle y-z, z-x \rangle \geq 0 \text{ for all } y\in C. \end{equation*} Moreover, the solution set of (EP) coincides with the fixed point set of the proximal mapping $R_{\lambda}$. The main difference between the proximal mapping and the regularization proximal mapping is that the former is defined as the unique solution of a strongly convex program, while the latter is defined by the unique solution of a strongly monotone equilibrium problem. We adopt the following definition. \begin{defn} For given $\epsilon > 0$ and $\lambda > 0$, the proximal mapping $B_{\lambda}$ is said to be {\it $\epsilon$-nonexpansive on $C$} if \begin{equation*} \|B_{\lambda}(x) - B_{\lambda}(y)\|^2 \leq (1 + \epsilon) \|x - y\| ^2 \ \forall x, y \in C. \end{equation*} \end{defn} The following theorem says that for monotone equilibrium problem, the proximal mapping is $\epsilon$-nonexpansive. \begin{thm}\label{T4} Suppose that the bifunction $\phi$ is monotone and satisfies the strongly Lipschitz-type condition on $C$. Then for any $\epsilon > 0$, there exists $\lambda > 0$ such that the proximal mapping $B_{\lambda}$ for Problem (MEP) is $\epsilon$-nonexpansive. \end{thm} \begin{proof} As before we see that if $\phi$ is monotone, strongly Lipschitz-type, then so is $f(x,y):= \phi(x,y) + \varphi(y)-\varphi (x)$ for any function $\varphi: C \to \mathbb{R}$. It is well known (see. e.g. \cite{M1965, RW1998}) that \begin{equation*} \langle B_{\lambda}(x) - x, B_{\lambda}(x) - z \rangle \leq \lambda [f(x, z) - f(x, B_{\lambda}(x) ] \ \forall x, z\in C. \end{equation*} Applying this inequality with $z:= B_{\lambda}(y)$ we obtain \begin{equation*} \langle B_{\lambda}(x)- x, B_{\lambda}(x)- B_{\lambda}(y) \rangle \leq \lambda [f(x, B_{\lambda}(y)) - f(x, B_{\lambda}(x))] \ \forall x, y\in C. \end{equation*} Similarly with $B_{\lambda}(y)$, we have \begin{equation*} \langle B_{\lambda}(y) - y, B_{\lambda}(y)- B_{\lambda}(x) \rangle \leq \lambda[f(y, B_{\lambda}(x)) - f(y, B_{\lambda}(y))] \ \forall x, y \in C. \end{equation*} Adding the two obtained inequalities we get \begin{align*} &\langle B_{\lambda}(x) - B_{\lambda}(y) + y - x , B_{\lambda}(x)- B_{\lambda}(y) \rangle \\ \leq \ &\lambda [f(x, B_{\lambda}(y)) - f(x, B_{\lambda}(x) + f(y, B_{\lambda}(x)) - f(y, B_{\lambda}(y))]. \end{align*} By simple arrangements we obtain \begin{align*} &\|B_{\lambda}(x) - B_{\lambda}(y)\|^2 \\ + \ &2\lambda \Big[ f(x, B_{\lambda}(y)) - f(x, B_{\lambda}(x) + f(y, B_{\lambda}(x)) - f(y, B_{\lambda}(y)) \Big]\\ \le \ &\| x-y\|^2 \end{align*} Now using the strongly Lipschit-type condition, by the same argument as in the proof of Theorem 3.7 in \cite{H2017} we arrive at the following inequality \begin{equation*} \|B_{\lambda}(x) - B_{\lambda}(y)\|^2 \leq (1+\lambda^2 M) \|x - y\|^2, \end{equation*} where $M = \sum_{j=1}^p K_i L_i$, with $K_j$ and $L_j$ being the Lipschitz constants defined in the strongly Litschitz-type. Hence, with $0 < \lambda^2 < \frac{\epsilon}{M}$, we obtain $\|B_{\lambda}(x) - B_{\lambda}(y)\|^2 \leq (1 +\epsilon )\|x-y\|^2$ for every $x, y \in C$. \end{proof} \begin{cor}\label{CH2005} Consider the mixed variational inequality \begin{equation*} \text{Find } x^* \in C \text{ such that } \langle F(x^*), y-x^*\rangle + \varphi(y) - \varphi(x^*) \geq 0 \text{ for all } y\in C. \tag{MVI} \end{equation*} Suppose that $F$ is monotone and Lipschitz on $C$. Then the proximal mapping $ B_{\lambda}$ defined by the bifunction $ \langle F(x), y-x\rangle + \varphi(y) -\varphi (x)$ is $\epsilon$-nonexpansive for any $\epsilon > 0$. \end{cor} \begin{proof} Since $F$ is monotone and Lipschitz on $C$, the bifunction $f(x,y) := \langle F(x), y-x\rangle + \varphi(y) -\varphi(x)$ is strongly Lipschitz and monotone. Thus the corollary follows directly from Theorem \ref{T4}. However one can prove this result simply as follows. From the definition of $B_{\lambda}$, by using the optimality condition for the problem defining $B_{\lambda}$ we can show \begin{equation}\label{ct3} \| B_{\lambda}(x) -B_{\lambda}(y)\| \leq \|x-y-\lambda( F(x) - F(y) )\| \ \forall x, y \in C. \end{equation} Then, from \begin{equation*} \|x-y-\lambda( F(x) - F(y) )\|^2 = \|x-y\|^2 -2\lambda\langle x-y, F(x) - F(y)\rangle +\lambda^2\|F(x) -F(y)\|^2 \end{equation*} by (\ref{ct3}) and monotonicity of $F$ we can write \begin{equation*} \|B_{\lambda}(x) -B_{\lambda}(y)\|^2 \leq \|x-y\|^2 + \lambda^2\|F(x)-F(y)\|^2, \end{equation*} from which, by Lipschitz continuity of $F$, it follows that \begin{equation*} \|B_{\lambda}(x) - B_{\lambda}(y)\|^2 \leq (1+L^2\lambda^2)\|x-y\|^2. \end{equation*} Hence the mapping $B_{\lambda}$ is $\epsilon$-nonexpansive on $C$ whenever $\lambda^2 L^2 \leq \epsilon$. \end{proof} Now a natural question may arise: how to modify the proximal mapping for monotone equilibrium problems such that it has a generalized nonexpansiveness property? In order to answer this question, let us define the mapping $T_{\lambda}$ from $C$ to itself by taking, for every $x\in C$, \begin{equation*} T_{\lambda}(x):= \text{argmin}\Big\{\lambda f(B_{\lambda}(x), y) + \displaystyle\frac{1}{2}\Vert y-x\Vert^2: y\in C \Big\} \end{equation*} where $\lambda$ is a fixed positive number. \begin{thm}\label{T5} {\rm (\cite{AM2018})}. Let $f: C\times C \to \mathbb{R}$ be a bifunction such that $f(x, \cdot)$ is subdifferentiable, pseudomonotone and Lipschitz-type on $C$. Suppose that the following conditions are satisfied: (A1) $f$ is jointly weakly continuous on $C \times C$ in the sense that, if $x, y\in C$ and $\{x_n\}, \{y_n\}\subset C$ converge weakly to $x$ and $y$, respectively, then $f(x_n, y_n) \to f(x, y)$ as $n \to \infty$. (A2) The solution set of Problem (EP) is nonempty.\\ Then the mapping $T_{\lambda}$ is quasi-nonexpansive on $C$ if $0<\lambda< \min\left\{\displaystyle\frac{1}{2L_1}, \displaystyle\frac{1}{2L_2}\right\}$. In addition, it is demiclosed at zero, in the sense that for every sequence $\{x_n\}$ contained in $C$ weakly converging to $x$ and $\|T(x_n) - x_n\| \to 0$, then $x\in Fix(T)$. \end{thm} By this theorem, the algorithms for finding a fixed point of quasi-nonexpansive mappings (see e.g. \cite{GD1997, I1974}) can be used for solving pseudomonotone equilibrium problems. A disadvantage of the composite proximal mapping $T_{\lambda}$ is that for evaluating it at a point, it requires solving two strongly convex programming problems. An open question is that how to define a nonexpansive or $\epsilon$-nonexpansive mapping with any $\epsilon >0$ for pseudomotone equilibrium problems, which requires solving only one strongly convex programs? As we have seen from the definition of the proximal mapping that when applying the iterative fixed point methods for solving mixed equilibrium problem (MEP), at an iterative point $x^k \in C$, we have to solve a strongly convex program of the form \begin{equation*} \min\left\{ f(x^k, y) : = \phi(x^k,y) + \varphi(y) - \varphi (x^k) + \frac{1}{2\lambda}\|y-x^k\|^2: y \in C\right\}. \eqno(P_k) \end{equation*} This problem can be solved by efficient algorithms of convex programming (see \cite{BV2004}). \section{Conclusions}\label{ConclusionSection} The mixed equilibrium problem (MEP) and the regularized Moreau proximal mapping $B_{\lambda}$ defined for it are equivalent in the following senses: (i) The solution set of (MEP) coincides with the fixed point set of $B_{\lambda}$ for any $\lambda > 0$. In addition: (ii) If (MEP) is strongly monotone and satisfies the Lipschitz-type condition, then one can choose $\lambda$ such that $B_{\lambda}$ is quasicontractive. (iii) If (MEP) is strongly monotone and satisfies the strongly Lipschitz-type condition, then one can choose $\lambda$ such that $B_{\lambda}$ is contractive. (iv) If (MEP) is monotone and satisfies the strongly Lipschitz-type, then one can choose $\lambda$ such that $B_{\lambda}$ is $\epsilon$-nonexpansive for any $\epsilon > 0$. (v) If (MEP) is pseudomonotne and satisfies the Lipschitz-type, then the composite proximal mapping is quasinonexpansive, and its fixed point-set coincides the solution-set of Problem (MEP). Applications to mixed variational inequality problems with Lipschitz cost operator have been presented. The following question seems to be interesting: How to extend these results for the equilibrium problem when the bifunction involved is quasiconvex with respect to its second variable? \section*{Acknowledgements} This research is funded by Vietnam National Foundation for Science and Technology Development (NAFOSTED) under grant number 101.01-2020.06.

95fdfacd0c5f10b80fb5483738aa3bf54cbfb2df

\section{Introduction} Interface engineering of oxide thin films is one of the great achievements of materials science, underpinning exotic physics\cite{Hwang2012} and giving rise to advanced computing and energy technologies.\cite{Martin2017} Because of strong interplay among lattice, charge, and spin degrees of freedom, even slight fluctuations in local order at oxide interfaces can greatly impact behaviors such as electronic\cite{Huang2018} and ionic conductivity.\cite{Fabbri2010} Order can be disrupted via numerous defect mechanisms, such as local structural perturbations\cite{Ismail-Beigi2017, Shamblin2016a} or the formation of cation and anion vacancies.\cite{Gunkel2020,Zhang2015} The desire to understand these mechanisms has motivated efforts to map defect formation pathways during oxide synthesis\cite{MacManus-Driscoll2020,Brahlek2018} and exposure to extreme environments.\cite{Spurgeon2020b,Zhang2018a,Beyerlein2013} Understanding oxides in extremes of temperature and irradiation is particularly important, since devices such as solid oxide fuel cells (SOFCs),\cite{Shamblin2016a,Tuller2011} durable nuclear waste forms,\cite{Ewing2004,Sickafus2000} and space-based electronics\cite{Cramer2016} must deliver reliable long-term performance under challenging conditions. At a more fundamental level, the nature of order-disorder processes represents a grand scientific question, with implications for the design of high entropy alloys\cite{George2019} and ceramics,\cite{Oses2020} strongly correlated quantum systems,\cite{Vepsalainen2020a} and more. To dictate order-disorder behavior, we must be able to visualize and direct defect formation processes at high spatial, chemical, and temporal resolution. The community has a long, successful history of using tailored ion irradiation to manipulate and induce defect populations in bulk metals\cite{Was2015,Odette2008} and oxides.\cite{Meldrum1998,Lumpkin2009,Sickafus2000,Shamblin2016,Lang2010} While recent studies\cite{Zhang2018a,Martinez2016,Beyerlein2015,Beyerlein2013} have showcased the promising properties of nanostructured systems, such as enhanced defect annihilation and radiation hardness, past work has focused almost exclusively on metals and bulk oxides.\cite{Zhang2018a} A handful of studies of model oxide interfaces have combined controlled ion irradiation and local characterization tools to identify unique behaviors such as anti-site defect buildup,\cite{Kreller2019} oxygen vacancy formation,\cite{Spurgeon2020} and orientation-/chemistry-dependent amorphization behavior.\cite{Kaspar2017,Aguiar2014,Aguiar2014a} These studies show that there is an interplay between defect formation energy, which depends on the interface configuration (e.g. strain, charge state, chemistry)\cite{Spurgeon2020,Dholabhai2014,Aguiar2014,Aguiar2014a,Zhuo2011} and defect kinetics, particularly defect mobility in different interface components.\cite{Zhuo2012a} However, most prior work has focused on static snapshots of these materials and has not possessed sufficient spatiotemporal resolution to probe the kinetics of local defect formation, particularly during the initial loss of crystallinity in oxide thin films. Moving beyond traditional static characterization approaches toward \textit{in situ} methods will allow us to capture the initial onset and evolution of local defects. \textit{In situ} (scanning) transmission electron microscopy ((S)TEM) has been widely employed by the radiation effects community to examine local defects in materials,\cite{Parrish2021,Lian2009,Birtcher2005} providing an invaluable window into irradiation-induced disorder, particularly in the low dose regime. These methods have long been used to examine bulk ceramics,\cite{Ye2011, Lian2009, Zhang2005a,Wang1998,Zinkle1992} revealing radiation-induced defect formation pathways and kinetics. Local (S)TEM probes are particularly well-suited to examining nanostructured oxide interfaces,\cite{Spurgeon2017a} whose non-equilibrium behavior can deviate greatly from bulk materials, but surprisingly little work has been done on complex oxide thin film systems in the context of radiation damage. More broadly, the microscopy community has also recognized the need for new data-driven approaches to characterization of transient processes, needed to detect and quantify salient features during high-speed imaging.\cite{Hattar2021,Spurgeon2020c,Taheri2016} Here we visualize the evolution of local disorder at a LaMnO$_3$ (LMO) / SrTiO$_3$ (STO) (001) perovskite oxide interface using ion irradiation coupled with \textit{in situ} high-resolution transmission electron microscopy (HRTEM) imaging at the I$^3$TEM irradiation facility at Sandia National Laboratories. We observe the initial onset of disorder and track its progression over continued irradiation. Using a Fourier filtering time series approach, we show that initial radiation damage is accommodated by the percolation of amorphous regions throughout the crystalline LMO matrix. In addition, we find evidence for a preserved crystalline interface region even at the highest fluence studied, whose origin we examine in the context of electronic structure calculations and past work. Taken together, these results demonstrate the power of high-resolution \textit{in situ} approaches to derive complex disordering pathways at oxide interfaces. More broadly, the data processing approaches we demonstrate may be applied to other dynamic studies of complex materials phase transitions. \section{Results and Discussion} We first consider the overall degradation of the system from the crystalline, as-grown condition to its disordered state. Figure \ref{ex_situ}.A shows the configuration of the electron and ion beams during irradiation, with the sample inclined 30$^{\circ}$ in the X tilt direction to minimize shadowing of the ion beam. Figure \ref{ex_situ}.B shows a cross-sectional STEM-HAADF image of the starting epitaxial 40 nm LMO film along the STO [100] crystallographic zone-axis prior to irradiation. In this mode, the directly-interpretable atomic number ($Z^{\sim1.7}$) contrast reveals a clear difference between the film and substrate, with minimal intermixing, excellent epitaxy, and no extended defects. Figure \ref{ex_situ}.C shows a HRTEM image of the same film in the I$^3$TEM system, with the interface rotated approximately 60$^{\circ}$ about the [100] zone (the normal direction to the image plane). Phase contrast in HRTEM is less directly interpretable than STEM-HAADF, but the inset fast Fourier transform (FFT) and clear lattice fringes confirm the crystalline starting condition. From this point, a series of four irradiations were performed in 1 hour increments with 2.8 MeV Au$^{4+}$ ions, as shown in Figures \ref{ex_situ}.D--G, with the sample inclined during irradiation and then tilted back to the zone for imaging. Each irradiation step corresponds to a fluence of $9.37 \times 10^{14}$ Au$^{4+}$ cm$^{-2}$. These figures show a gradual amorphization sequence, starting from the LMO film and extending to the STO substrate. A careful inspection of the images and associated FFTs reveals the emergence of local amorphous patches in the LMO at a fluence of $9.37 \times 10^{14}$ Au$^{4+}$ cm$^{-2}$ (Figure \ref{ex_situ}.D), with less apparent damage in the STO film. By a fluence of $1.87 \times 10^{15}$ Au$^{4+}$ cm$^{-2}$ (Figure \ref{ex_situ}.E), the patches have grown more extensive in the LMO, and the STO side also exhibits obvious disorder; this is accompanied by a decrease in the intensity of the FFT reflections. At a fluence of $2.81 \times 10^{15}$ Au$^{4+}$ cm$^{-2}$ (Figure \ref{ex_situ}.F), the LMO film is almost entirely amorphous, except for a thin, several nm-thick band at the interface. Finally, at a fluence of $3.75 \times 10^{15}$ Au$^{4+}$ cm$^{-2}$ (Figure \ref{ex_situ}.G), the LMO film appears similarly disordered to its state at a fluence of $2.81 \times 10^{15}$ Au$^{4+}$ cm$^{-2}$, but it is clearly thinner in the beam direction (as indicated by the larger vacuum region at the top left corner of the image), likely due to sputtering of the film under the ion bombardment. In addition, the STO layer is extensively damaged and the only remaining lattice fringes are from a 2--3 nm band at the film-substrate interface. There is almost no periodic signal left in the FFT, reflecting the loss of overall crystalline order. STEM-HAADF images of the $3.75 \times 10^{15}$ Au$^{4+}$ cm$^{-2}$ condition (Figure \ref{ex_situ}.H) show that some lattice fringes are still present in the bulk of the STO, but the bulk of the LMO is completely amorphous. Nonetheless, there is a pronounced and persistent crystalline band on the film side of the interface (marked by the arrow). We note that SRIM simulations, shown in Supporting Information Note 3, indicate a trace amount of Au ions may be retained in the TEM foil, which may partly influence the kinetics of the radiation response. However, a similar irradiation response has been observed in the La$_2$Ti$_2$O$_7$ / STO system,\cite{Spurgeon2020} where a 5--10 nm crystalline band persisted on the STO side of the interface. While the origins and mechanisms leading to the resistant interface behavior are yet not fully understood, we do note the presence of some cation intermixing in the irradiated sample (shown in Supporting Information Note 4), which may affect its stability. Overall, this points to the broader trend for perovskite oxide interface systems to remain crystalline under irradiation environments. \begin{figure*} \includegraphics[width=0.62\textwidth]{fig1.png} \caption{Evolution of local disorder with fluence. (A) Schematic of the irradiation geometry. (B--C) Colorized cross-sectional STEM-HAADF and HRTEM images of the as-grown film. (D--G) HRTEM images for fluences of $9.37 \times 10^{14}$, $1.87 \times 10^{15}$, $2.81 \times 10^{15}$, $3.75 \times 10^{15}$ Au$^{4+}$ cm$^{-2}$, respectively. (H) STEM-HAADF image of the $3.75 \times 10^{15}$ Au$^{4+}$ cm$^{-2}$ fluence sample. (Higher-resolution images are provided in Supporting Information Note 5). \label{ex_situ}} \end{figure*} We focus next on the initial stages of the irradiation process, probing the initial percolation of disorder and its buildup to the more global amorphization shown in Figure \ref{ex_situ}.H. For this experiment, we captured a high-speed time series with the sample inclined at 30$^{\circ}$ in the X tilt direction to maximize exposure to the ion beam. At this high tilt angle it is no longer possible to directly resolve atomic columns in the [100] zone-axis, and we instead image (100)-type lattice planes. However, the remarkable stability of the sample during the extended period of irradiation (40 minutes, total fluence $6.25 \times 10^{14}$ Au$^{4+}$ cm$^{-2}$) provided a unique window into the initial stages of disorder. We employ a temporal Fourier filtering approach, analogous to time-resolved geometric phase analysis (GPA),\cite{Hytch1998} processing the raw movie frames into maps of local crystallinity and lattice displacement. In addition, we quantify the total Bragg filter amplitude, which effectively measures the spatial abundance of the selected (100)-type lattice domains in any given frame. It is important to note that while the overall sample was quite stable, these measurements are very challenging and some periodic bending/rotation of the sample due to local heating did occur; since this results in large random drops in the Bragg filter amplitude, a trend line was fit to the Bragg filter amplitude only where crystalline signal is present (see Supporting Information Methods for details). These measurements are shown in Figure \ref{in_situ} and Supporting Information Movie S1. We see that there is initially a large contiguous block of crystalline region, as expected. By 5 minutes, some disorder emerges in the film, beginning at its center and extending laterally, but the Bragg amplitude is still around 95\% of the starting condition. Disorder is accompanied by significant local lattice displacement around the defective regions. At this stage, misfit dislocations emerge that run from the lower left to upper right corner of the frames. Between 9 and 19 minutes, the percolation extends from the film center to its surface and the substrate interface. At this point there is a substantial drop in the Bragg amplitude to nearly 80\%, accompanied by a still more complex pattern of lattice rotation and crystalline domains, some $< 10$ nm in size. The misfit dislocations appear to reconnect, forming repaired epitaxial domains, but there is a clear increase in disorder. Between 19 to 33 minutes the overall distribution remains steady near 80\%, but between 33 and 37 minutes the percolation continues and the crystalline domains become increasingly sparse and disconnected. In the final condition around 40 minutes, only about 60\% of the film is still similar to the starting crystalline condition and there is a large disconnected network of ordered regions. These results point to the highly non-uniform evolution of the interface during the initial stages of irradiation. \begin{figure*} \includegraphics[width=\textwidth]{fig2.png} \caption{Percolation of initial local lattice disorder. Top to bottom: Raw frames, crystallinity maps, displacement maps, and total Bragg filter amplitude as a function of frame time. The dark red line corresponds to an iteratively fitted time-average of the amplitude data and the inset shows the masked FFT reflections used for the analysis. Total combined fluence over 40 minutes is $6.25 \times 10^{14}$ Au$^{4+}$ cm$^{-2}$. \label{in_situ}} \end{figure*} To further explore this non-uniform interface response in later stages of irradiation, we have performed local position-averaged convergent beam electron diffraction (PACBED) and electron energy loss spectroscopy (STEM-EELS) measurements. The former is a mode of scanning nanodiffraction\cite{Ophus2019} that can quantify crystallinity at the nanoscale and has been used before to examine local disorder in oxides,\cite{Savitzky2020,Spurgeon2020,Janish2019} while the latter has been used to examine chemical states in both pristine and irradiated oxides.\cite{Spurgeon2020b} As shown in Figure \ref{eels}.A, at a fluence of $3.75 \times 10^{15}$ Au$^{4+}$ cm$^{-2}$ the bulk of the LMO film has amorphized, resulting in a diffuse, ring-like PACBED pattern. In contrast, PACBED from the uniform $\sim3$ nm interface band exhibits strong diffraction disks; there is also evidence for more weakly diffracting patches that extend further into the film. Similarly, the STO side of the interface also exhibits strong diffraction disks, despite some distributed damage, consistent with past observations.\cite{Zhang2005a,Jiang2012a} Correlative STEM-EELS chemical analysis of the interface, shown in Figure \ref{eels}.B, reveals the gradual breakup of the crystalline film upon transitioning from the interface into the bulk, resulting in three distinct regions (labeled 1--3). Region (1) consists of a $\sim 3$ nm region of LMO that is strongly diffracting and highly chemically ordered on both the La and Mn sublattices. EELS spectra from this region contain a characteristic pre-peak (a), main peak (b), and secondary peak (c) in the O $K$ edge fine structure, as well as a clear Mn $L_{2,3}$ white-line doublet, consistent with previous observations.\cite{Kaspar2019a,Varela2009} Moving away from the interface and into the LMO, we move into a more weakly diffracting $\sim2$ nm Region (2) consisting of poorly defined perovskite atomic order. This region exhibits a reduction in the O $K$ pre-peak feature and loss of definition between the main and secondary peaks. In addition, there is a clear 0.5 eV shift of the Mn edge to lower energy loss, pointing toward underlying reduction. This region is followed by third and final amorphous region (3) with no extended ordering of La and Mn; now the O $K$ edge pre-peak has disappeared and the main edge features have blurred into one from the increased randomization of the sublattice. There is a further 0.5 eV Mn chemical shift, reflecting further Mn reduction. The origin of this reduction is likely the loss of oxygen (oxygen vacancy formation) and electronic reconfiguration during the ion irradiation process.\cite{Aguiar2014,Spurgeon2020} \begin{figure*} \includegraphics[width=\textwidth]{fig3.png} \caption{Detail of the interface after irradiation. (A) Colorized cross-sectional STEM-HAADF image, with local PACBED patterns from the film, interface, and substrate, taken from the $3.75 \times 10^{15}$ Au$^{4+}$ cm$^{-2}$ fluence sample. (B) Colorized STEM-ADF and corresponding STEM-EELS Ti $L_{2,3}$, La $M_{4,5}$, Mn $L_{2,3}$, and composite maps taken from near the boxed region in (A). (C--D) Corresponding Mn $L_{2,3}$ and O $K$ spectra, respectively, from regions 1--3. \label{eels}} \end{figure*} In order to understand the energetics of oxygen vacancies at the LMO / STO interface, density functional theory (DFT) calculations have been performed, in which an oxygen vacancy has been introduced at different locations of the LMO / STO interface. These simulations used three different model interface terminations (i.e., two LMO / STO (001) and one LMO / STO (110)), as shown in Supporting Information Note 2. The two (100) interface models were built from the STO (100) substrate orientation, terminated either by SrO or TiO$_2$, yielding LMO / STO interface configurations with cation rows ordering as (Ti--Sr--Mn--La) and (Sr--Ti--La--Mn), respectively. As shown by the relative energy plots in Supporting Information Note 2, forming an oxygen vacancy on the LMO side of the interface is found more favorable than on the STO side. This behavior is likely due to the ability of Mn$^{4+}$ ions to reduce to Mn$^{3+}$ more readily than for Ti$^{4+}$ to reduce to Ti$^{3+}$. Supporting Information Note 2 also shows that creating an oxygen vacancy at the interface of the Ti-terminated LMO (100) / STO (100) configuration is less favorable than creating it at the Sr-terminated configuration. While the Ti-terminated interface provides a less favorable environment for oxygen vacancy creation compared to the Sr-terminated interface, the formation energy is still similar to that in STO. These results suggest that, in the LMO / STO system, the mechanisms for interface stabilization are not dominated by the oxygen vacancy creation. Nevertheless, these results overall show a net energy gain of 2 eV for vacancy formation on the LMO side of the interface relative to the STO side. This finding is in excellent agreement with the overall trend in the amorphization sequence that we observe experimentally. However, unlike in the case of LTO / STO,\cite{Spurgeon2020} no interfacial oxygen vacancy differences were observed to explain the retention of crystallinity at the interface, so the specific characteristics of the damage mechanism are likely different. \section{Conclusions} Using an \textit{in situ} imaging approach, we reveal the percolation of disorder in oxide thin film interfaces. Our results indicate the formation of a complex network of amorphization during the initial stages of irradiation, which progresses to global disorder over longer time scales. However, we also observe the preservation of a distinct crystalline interface region on the film side of the LMO / STO interface. Our calculations demonstrate a propensity for defect accumulation in the bulk of the LMO, which likely influences the course of the disordering processes. These results support the general trend toward retention of interfacial crystallinity, which appears to be unique to epitaxial oxides. Collectively, these results demonstrate the untapped potential of high-resolution spatiotemporal probes to survey the complex landscape of disorder and underscore the important role of interface engineering in mediating disordering processes. \section{Acknowledgements} This research was supported by the Nuclear Processing Science Initiative (NPSI) Laboratory Directed Research and Development (LDRD) at Pacific Northwest National Laboratory (PNNL). PNNL is a multiprogram national laboratory operated for the U.S. Department of Energy (DOE) by Battelle Memorial Institute under Contract No. DE-AC05-76RL0-1830. C.O. acknowledges support from the DOE Early Career Research Program. The STEM imaging shown was performed in the Radiological Microscopy Suite (RMS), located in the Radiochemical Processing Laboratory (RPL) at PNNL. Sample preparation was performed at the Environmental Molecular Sciences Laboratory (EMSL), a national scientific user facility sponsored by the Department of Energy's Office of Biological and Environmental Research and located at PNNL. \textit{In situ} ion irradiation work was performed at the Center for Integrated Nanotechnologies, an Office of Science User Facility operated for the U.S. DOE. Sandia National Laboratories is a multimission laboratory managed and operated by National Technology \& Engineering Solutions of Sandia, LLC, a wholly owned subsidiary of Honeywell International, Inc., for the U.S. DOE's National Nuclear Security Administration under contract DE-NA-0003525. Work at the Molecular Foundry was supported by the Office of Science, Office of Basic Energy Sciences, of the US DOE under Contract DE-AC02-05CH11231. The views expressed in the article do not necessarily represent the views of the U.S. DOE or the United States Government. \section{Competing Interests Statement} The authors declare no competing interests. \section{Data Availability Statement} The data used in this study are available from the authors upon reasonable request. \section{Supporting Information} The Supporting Information is available free of charge at X. Details of methods used, DFT calculations, SRIM calculations, sequence of the irradiation, and a movie of the irradiation are available. \section{Table of Contents Figure} \begin{figure*}[h] \includegraphics[width=0.7\textwidth]{toc_figure.jpg} \end{figure*} \clearpage \section*{Supporting Information References}} \begin{document} \section*{Supporting Information Note 1: Methods} \subsection{Thin Film Growth} Nominally 40 nm-thick epitaxial, single-crystal LaMnO$_3$ films were deposited on SrTiO$_3$ (001) substrates by oxygen-plasma-assisted molecular beam epitaxy (OPA-MBE), as described previously.\cite{Kaspar2019a} Briefly, deposition occurred at 625 $^{\circ}$C with oxygen supplied from a differentially pumped electron cyclotron resonance (ECR) microwave plasma source at a chamber pressure of 2.4 × 10$^{-6}$ Torr. La and Mn were supplied from effusion cells to realize an LMO growth rate of 45 s u.c.$\rm{}^{-1}$ ($\sim0.85$~\AA \, s$^{-1}$). After deposition, the films were cooled in the flow of activated oxygen from the ECR plasma source. \subsection{Transmission Electron Microscopy} Cross-sectional TEM samples were prepared using a FEI Helios NanoLab DualBeam Focused Ion Beam (FIB) microscope and a standard lift out procedure. Samples were mounted into a slot at the end of a lift out half grid to accommodate the ion beam geometry. We note that a slight wedge shape is inevitable, but the sample is nonetheless quite flat. High-angle annular dark field (STEM-HAADF) images of the as-grown and post-irradiated samples were collected using a probe-corrected JEOL GrandARM-300F STEM operating at 300 kV with a semi-convergence angle of 27.5 mrad and a collection angle of 82--186 mrad. Position-averaged convergent beam electron diffraction (PACBED) patterns were collected with a reduced convergence semi-angle of $\alpha \approx 4$ mrad to reduce diffraction disk overlap. STEM electron energy loss spectroscopy (STEM-EELS) measurements were conducted in DualEELS mode, with a at a 0.25 eV ch$^{-1}$ dispersion yielding an approximate $\sim0.75$ eV energy resolution. The spectra were subsequently corrected for zero loss energy drift, but no denoising was applied. The maps shown in the supplement were collected with $4 \times$ energy binning to improve signal-to-noise. \textit{In situ} ion irradiation was performed using the I$^3$TEM system at CINT, which consists of a JEOL 2100 LaB$_6$ TEM modified to incorporate an ion beam.\cite{Hattar2014} We utilized 2.8 MeV Au$^{4+}$ ions with a current density of $\sim$1.6 nA cm$^{-2}$. On-axis high-resolution transmission electron microscopy (HRTEM) imaging was performed prior to irradiation along the STO [100] zone-axis and the sample was subsequently inclined at 30$^{\circ}$ in the X tilt direction to minimize shadowing of the ion beam during each irradiation step. Irradiations were performed in 1 hour steps, corresponding to a fluence of $9.37 \times 10^{14}$ Au$^{4+}$ cm$^{-2}$ per step, with the electron beam on and illuminating the sample. We note that prior work\cite{Lian2009} has shown that electron beam illumination can influence the radiation response of the material, but we did not systematically explore this effect. After every hour, the sample was tilted back onto the STO [100] zone and additional HRTEM imaging was performed. This process was repeated for a total irradiation time of 4 hours, yielding a total fluence of $3.75 \times 10^{15}$ Au$^{4+}$ cm$^{-2}$. Movies of each irradiation step were taken at $1024 \times 1024$ px$^2$ resolution with a 3.56 frames per second framerate. \subsection{Post-Processing of \textit{In Situ} Data} Analysis of the \emph{in situ} HRTEM time series experiment was performed with custom Matlab codes. We used a similar filtering method to the Bragg filtering used for geometric phase analysis (GPA) in \citet{Hytch1998}. First, we measured the position of the lone Bragg peak pair visible throughout the time series, and computed a Bragg filter mask centered on one of these peaks using a 2D Gaussian function with a standard deviation equal to 0.0025 1 px$^{-1}$. The location of these Bragg peaks is shown inset in main text Figure 2 as red circles. Next, for each image in the time series, we computed the 2D Fourier transform of the image, multiplied it by the Bragg filter, and then took the inverse 2D Fourier transform. The resulting complex image was used in two ways; first, the amplitude was used to create a map of the spatial distribution where crystalline signal was present, shown for some images in the second row of main text Figure 2. The output minimum and maximum were scaled be equal to 0.004 and 0.008 of the normalized image intensity respectively. This was done to eliminate false positive signals, and to saturate the output value to 1 where crystalline signal was strongly visible. Next, we also used the phase of the complex output to map the lattice displacement, by dividing out the plane wave defined by the Bragg vector $\exp{ -2 \pi i \, {\bf q} \cdot {\bf r} }$, where ${\bf q}$ is the 2D Bragg vector and ${\bf r}$ is the spatial coordinate vector.\cite{Hytch1998} The resulting phase image divided by $2 \pi$ defines the local shift of the lattice, which we have plotted in the hue image channel in the third row of Figure 2. The amplitude mask was also applied to this row of images, to avoid plotting the lattice displacement in regions of pure noise (no crystalline signal present). The resulting hue images can show the present of dislocations and other lattice disconnections wherever the color field is discontinuous. Finally, the total 2D amplitude image signal normalized to 1 at time 0 is plotted at the bottom of main text Figure 2. A best-fit trend line is also plotted, which was iteratively fit to values $\geq 75\%$ of the local trend line value, and smoothed with a 401 frame long Gaussian distribution. The resulting trend line tracks the mean Bragg filter amplitude in the frames where the sample is on zone and crystalline signal is present. \subsection{SRIM Simulations} The design of the \textit{in situ} Au ion irradiation experiments was accomplished based on SRIM (Stopping and Range of Ions in Matter) simulations,\cite{Ziegler2010} where the specific gravities of LaMnO$_3$ and SrTiO$_3$ used are 6.52 and 5.12 g cm$^{-3}$, respectively. The full damage cascade mode was used in the simulation, where the threshold displacement energies of $E_d$ (La) = 80 eV, $E_d$ (Mn) = 70 eV and $E_d$ (O) = 45 eV in LaMnO$_3$ and $E_d$ (Sr) = 80 eV, $E_d$ (Ti) = 70 eV and $E_d$ (O) = 45 eV in SrTiO$_3$ were assumed.\cite{Zhang2008a} 2.8 MeV Au ions at an incidence angle of $60^{\circ}$ were selected for irradiation to provide a balanced combination of dose rate and retained Au percentage in a TEM foil estimated to be $\sim80$ nm in thickness. These data are described in Supporting Information Note 3. \subsection{Density Functional Theory} The relative formation energy of oxygen vacancy has been calculated within the density functional theory (DFT) framework, as implemented in the VASP code.\cite{Kresse1996} All the simulations used the PBEsol exchange correlation functional,\cite{Perdew2008} spin-polarization, and a $4 \times 4 \times 1$ Monkhorst-Pack\cite{Monkhorst1976} $k$-point mesh to sample the Brillouin zone. The cutoff energy for the projector augmented wave\cite{Blochl1994} pseudo-potential was 500 eV, with convergence criteria of $10^{-5}$ eV per cell for the energy and $10^{-3}$ eV \AA$^{-1}$ for the force components. Computational models of the LMO / STO interface used a periodic symmetric slab, such that LMO was sandwiched on both sides of a 4.5 unit cell of STO slab (i.e., LMO / STO / LMO). The lattice parameters were fixed and used the experimental lattice parameter for STO ($a = 3.905$~\AA). The simulation of the LMO / STO (001) interface used lateral lattice parameters of $a \sqrt{2} \times a \sqrt{2}$ and 76~\AA \, in the out-of-plane direction to avoid interactions between periodic images. Similarly, the simulation of the LMO / STO (110) interface used lateral lattice parameters of $a \sqrt{2} \times a \sqrt{2}$ and 93~\AA \, in the out-of-plane direction. In order to avoid artificial dipole effects, oxygen vacancies were symmetrically created on both sides from the center of the STO slab. \clearpage \section*{Supporting Information Note 2: Energy Cost of Oxygen Vacancy Formation} \begin{figure*}[h] \includegraphics[width=\textwidth]{figs1.png} \caption{Relative energy of oxygen vacancy creation across the LMO / STO interface for the (110) (a) and (001) (b) substrate orientations. For the (001) orientation, the Sr- and Ti-terminations are shown in left and right ball and stick models respectively. The La atoms are represented in yellow, O atoms in red, Sr atoms in grey, Mn atoms in purple, and Ti atoms in light blue. \label{si_dft}} \end{figure*} \clearpage \section*{Supporting Information Note 3: SRIM Simulations of Ion Damage} SRIM simulations were performed for an ion fluence of $1 \times 10^{16}$ Au$^{4+}$ cm$^{-2}$. The mean dose in the foil is taken at the approximate middle of the foil (40 nm). The mean doses in LaMnO$_3$ and SrTiO$_3$ are 54.9 and 44.6 dpa, as shown in Figures \ref{si_srim_lmo} and \ref{si_srim_sto}, respectively. The maximum Au concentration retained in the foil at the depth of 80 nm is 0.3 at. \% in LaMnO$_3$ and 0.2 at. \% in SrTiO$_3$. \begin{figure*}[h] \includegraphics[width=\textwidth]{figs_lmo.jpg} \caption{SRIM full damage cascade calculations for LaMnO$_3$. \label{si_srim_lmo}} \end{figure*} \begin{figure*}[h] \includegraphics[width=\textwidth]{figs_sto.jpg} \caption{SRIM full damage cascade calculations for SrTiO$_3$. \label{si_srim_sto}} \end{figure*} \clearpage \section*{Supporting Information Note 4: STEM-EELS Measurements of Intermixing} STEM-EELS measurements were conducted after irradiation to assess potential post-irradiation intermixing. As shown in Figure \ref{si_eels_intermixing}, there is a non-negligible amount of cation intermixing on both the $A$- and $B$-sites, raising the possibility that this may contribute to the enhanced irradiation stability of the film at the interface. \begin{figure*}[h] \includegraphics[width=0.5\textwidth]{figs_intermixing.png} \caption{Colorized STEM-ADF and corresponding STEM-EELS maps for the La $M_{4,5}$, Mn $L_{2,3}$, Sr $L_{2,3}$, and Ti $L_{2,3}$ edges. \label{si_eels_intermixing}} \end{figure*} \clearpage \section*{Supporting Information Note 5: Irradiation Sequence} The following images from main text Figure 1 are presented in greater detail to show the lattice structure and associated details during irradiation. \begin{figure*}[h] \includegraphics[width=\textwidth]{figs_seq1.jpg} \caption{Evolution of local disorder with fluence. Colorized cross-sectional STEM-HAADF image of the as-grown film.} \end{figure*} \begin{figure*}[h] \includegraphics[width=\textwidth]{figs_seq2.jpg} \caption{Evolution of local disorder with fluence. Colorized cross-sectional HRTEM image of the as-grown film.} \end{figure*} \begin{figure*}[h] \includegraphics[width=\textwidth]{figs_seq3.jpg} \caption{Evolution of local disorder with fluence. Colorized cross-sectional HRTEM image of the $9.37 \times 10^{14}$ Au$^{4+}$ cm$^{-2}$ irradiated film.} \end{figure*} \begin{figure*}[h] \includegraphics[width=\textwidth]{figs_seq4.jpg} \caption{Evolution of local disorder with fluence. Colorized cross-sectional HRTEM image of the $1.87 \times 10^{15}$ Au$^{4+}$ cm$^{-2}$ irradiated film.} \end{figure*} \begin{figure*}[h] \includegraphics[width=\textwidth]{figs_seq5.jpg} \caption{Evolution of local disorder with fluence. Colorized cross-sectional HRTEM image of the $2.81 \times 10^{15}$ Au$^{4+}$ cm$^{-2}$ irradiated film.} \end{figure*} \begin{figure*}[h] \includegraphics[width=\textwidth]{figs_seq6.jpg} \caption{Evolution of local disorder with fluence. Colorized cross-sectional HRTEM image of the $3.7 \times 10^{15}$ Au$^{4+}$ cm$^{-2}$ irradiated film.} \end{figure*} \begin{figure*}[h] \includegraphics[width=\textwidth]{figs_seq7.jpg} \caption{Evolution of local disorder with fluence. Colorized cross-sectional STEM-HAADF of the $3.7 \times 10^{15}$ Au$^{4+}$ cm$^{-2}$ irradiated film.} \end{figure*} \clearpage \section*{Supporting Information Note 6: \textit{In Situ} Movie} Supporting Movie 1 was created by processing the \textit{in situ} movie using Matlab codes. First, the movie dataset from frames 1 to 8400 were Fourier downsampled to $400 \times 400$ pixels, such that the primary lattice planes were just below the Nyquist sampling frequency. Next, a Tukey edge window was applied to each image to create smooth boundary conditions, and the relative translation between images was measured using cross correlation. All images were then zero-padded to an image size of $800 \times 800$ pixels, and the measured translations were applied to align all images. Finally, these images were Bragg filtered from one of the primary lattice diffraction spots, using a moving average of 11 frames to reduce flickering of the outputs. The output movie images were constructed in a hue-saturation-value ($HSV$) color space. The $H$ channel was set to the lattice displacement (complex angle of the Bragg filter output) with periodic wrapping of output colors. The $S$ channel was set to a scaled Bragg filter magnitude (absolute value of the Bragg filter output), such that the image is grayscale where no lattice planes are present, and fully color-saturated where the lattice plane signals are strongest. Finally the $V$ channel is set to the original image intensity, scaled for best visibility. The Fourier transform amplitude of each image is inset into the upper right corner. \clearpage

8ed04e5ba5d672076394c072a923c7f356e02097

\section{Application} \label{chap:application} \subsection{Training} In this section the entire FIML approach is applied. First, training data is generated by conducting field inversions on the S809 airfoil for angles of attack $\alpha = 6.2^\circ, 10.2^\circ$ and $20.1^\circ$. As discussed in the previous chapter, for each case the regularization parameter $\lambda$ is determined separately and the field inversions are conducted only on the fully resolved $1057 \times 145$ grid to ensure optimal training data. From the complete set of data, the samples in the freestream, where the turbulence model is not active, are discarded together with $99.5\%$ of the remaining samples where $\beta = 1$. This is done to reduce the frequency at which $\beta = 1$ appears in the training data as otherwise the machine learning model would simply learn to always predict $\beta = 1$. The remaining samples are split into a training set totaling approximately $38000$ samples and a testing set totaling approximately $4 000$ samples. Second, a machine learning model is trained to model $\beta$ according to Eq. (\ref{eq:ml-beta}). The neural network used in this case is a fully connected network with two layers with 200 neurons each. Dropout layers are applied after each hidden layer with a drop-out rate of $0.2$. The mean squared error is used as loss function and the network is trained for $500$ epochs using the Adam optimizer. The dimensionless flow features $\eta_i(\mathbf{U}, \tilde{\nu})$ used as input are \begin{equation} \eta_1 = \log (\frac{P}{D}), \eta_2 = \log(\chi), \eta_3 = \log(\frac{|S|}{|\Omega|}) \,\text{and}\, \eta_4 = \log(\frac{\mu_t |S|}{\tau_w}), \end{equation} similar to the features presented in \cite{holland}. The features represent the ratio of the production and the destruction term of the SA model, the non-dimensionalized SA transport variable, the ratio of the strain and vorticity magnitudes and the ratio of the local turbulent shear stress to the shear stress at the wall respectively. The logarithms are applied due to the fact that each features' values span multiple orders of magnitudes. \subsection{Numerical Results} Finally, RANS simulations with the ML-augmented turbulence model are conducted for multiple angles of attack from $\alpha = 0.0^\circ$ up to $\alpha = 20.1^\circ$. \begin{figure} \centering \begin{minipage}[t]{.5\textwidth} \def\textwidth {\textwidth} \def0.9818\textwidth {\textwidth} \input{figures/application_cl_v_aoa.tikz}% \end{minipage}% \hfill \begin{minipage}[t]{.5\textwidth} \def\textwidth {\textwidth} \def0.9818\textwidth {\textwidth} \input{figures/s809_augmented_cp.tikz}% \end{minipage}% \caption{Left: Lift coefficients for the S809 airfoil at Re = $2\times 10^6$. Right: $c_p$-distributions for $\alpha = 12.2^\circ$ } \label{fig:results_ml} \end{figure}% Figure \ref{fig:results_ml} compares results from the augmented RANS model, the original turbulence model, the field inversion and the reference data. For $\alpha \lessapprox 6.2^\circ$, the agreement between the measured lift and the predicted lift using the original RANS model is good, while it quickly worsens for larger $\alpha$ as soon as flow separation occurs. The lift predictions of the field inversion results agree much better, matching the reference lift with deviations of only $0.8\%$ ($\alpha = 6.2^\circ$), $2.3\%$ ($\alpha =10.2^\circ$) and $8\%$ ($\alpha = 20.1^\circ$). The predictions of the ML-augmented turbulence model remain unchanged in the linear area of the lift curve, which is intended since the results of the original turbulence model are already satisfying here. For higher $\alpha$, the augmentation improves the turbulence model noticeably as the lift is predicted far more accurately. The maximum deviation of the augmented model is considerably lower with $13\%$ (at $\alpha = 10.2^\circ$) as opposed to a maximum deviation of $> 50\%$ (at $\alpha = 20.1^\circ$) of the baseline model. \begin{figure} \input{figures/s809_fiml} \caption{% Left: $\beta$-field from field inversion and as predicted by the ML-model. Right: Mach number and stream lines from the baseline turbulence model and the ML-augmented model. } \label{fig:result_fiml} \end{figure} On the right in Figure \ref{fig:results_ml}, the improvement of the predictions due to the ML-augmentation is illustrated based on the $c_p$-distribution exemplary for $\alpha = 12.2^\circ$, for which the field inversion was done but not included in the training data for the ML model. While the $c_p$-distribution of the augmented model (dashed line) doesn't come as close to the experimental values as the field inversion result (dotted), it still represents a large improvement with respect to the baseline model (solid). Figure \ref{fig:result_fiml} compares the results for the $\beta$-field as computed during field inversion and as predicted by the ML-augmentation. While the inversion result is smoother in general and shows stronger modifications, the presumably most import area leading up to the point of flow separation shortly after the thickest airfoil section, is predicted similarly by the ML-model. In the right half of Figure \ref{fig:result_fiml}, the Ma-number and the streamlines of the solution are plotted for the original turbulence model and the ML-augmented model. While the original model shows only a very small region of flow separation, the ML-augmented model develops a much larger separation bubble as expected. \section{Introduction} Turbulent flows often play a key role in engineering and in science, hence it is important to be able to predict these flows accurately by numerical simulation. Methods which can resolve turbulent scales like LES and DNS are often still too expensive to be used regularly. Hence, Reynolds-Averaged Navier Stokes (RANS) simulations, which model the influence of turbulence on the mean flow, are the standard method for daily use. One difficulty however is that RANS simulations often fail to predict flow separation accurately e.g. in high-lift flows. Over the last decades, many approaches aiming to improve turbulence modeling have been tried with varying degrees of success, such as extending existing models, for example with rotation corrections, or using more complex turbulence models, e.g. two equation models or Reynolds-stress models. One approach that has gained increasing interest in the recent years is the Field Inversion and Machine Learning (FIML) approach \cite{fiml}. For this approach, it is argued that RANS models have always been largely empirical, and hence, with the increasing maturity of machine learning techniques, they can further be improved using machine learning models trained on datasets derived from high-fidelity simulations like DNS and LES or from experiments. In this paper, an implementation of the FIML approach based on the negative Spalart-Allmaras turbulence model \cite{SAneg} in the DLR TAU-Code is described. In particular, findings on the influence of the regularization, grid resolutions and regions of the computational domain where the field inversion is active on the field inversion results are presented. Finally, the turbulence model is augmented with a machine learning model and improved results computed with the augmented model are shown. \section{Methodology} \label{chap:methodolgy} \subsection{Field Inversion and Machine Learning} The basic idea of the FIML approach is to introduce a discrepancy term $\beta$ into the turbulence model, here the negative Spalart-Allmaras model: \begin{align} \frac{D \tilde{\nu}}{D t} &= \beta(\mathbf{U}, \tilde{\nu}) P(\mathbf{U}, \tilde{\nu}) - D(\mathbf{U}, \tilde{\nu}) + T(\mathbf{U}, \tilde{\nu}) \label{eq:extended_turb_model} \end{align} Here, $P, D$ and $T$ denote the turbulent production, destruction and transport terms respectively and $\mathbf{U}$ and $\tilde{\nu}$ denote the flow state and the Spalart-Allmaras transport variable. The correction term $\beta$ is an unknown, spatially varying variable which is assumed to be a function of $\mathbf{U}$ and $\tilde{\nu}$. Since $\beta$ is unknown, it must be modeled as well. The modeling is done in two separate steps: During a \textit{field inversion} step training data is generated which is then used in a \textit{machine learning} step to train a model relating $\beta$ and $(\mathbf{U},\tilde{\nu})$. \paragraph{Field Inversion.} The training data needed for machine learning consists of pairs of $\beta$ and the corresponding flow state $(\mathbf{U}, \tilde{\nu})$, one pair per control volume. These pairs are obtained from high-fidelity reference data $d_\text{ref}$ by solving the inverse problem \begin{equation} d_\text{RANS}(\beta) = d_\text{ref}, \label{eq:inverse_problem} \end{equation} where $d_\text{RANS}$ is the RANS solution depending on $\beta$, for $\beta$. Solving Eq. (\ref{eq:inverse_problem}) is an optimization problem, for which a cost function, here \begin{equation} \mathcal{I} = \underbrace{\frac{1}{2} \sum_{i}^{N_i} \sum_{j}^{N_{j,i}} \left[ d_{i, \text{ref}}^j - d_{i, \text{RANS}}^j(\beta, \mathbf{U}, \tilde{\nu}) \right]^2}_{\mathcal{I}_1} + \lambda \underbrace{\frac{1}{2} \sum_{k}^{N_k} \left( \beta_k - 1 \right)^2}_{\mathcal{I}_2} \label{eq:cost_fn} \end{equation} needs to be minimized. The first term of the cost function, $\mathcal{I}_1$, measures the deviation between $d_\text{ref}$ and $d_\text{RANS}$ and is calculated as the squared difference of $d_\text{ref}$ and $d_\text{RANS}$, summed over all $N_{j,i}$ cells where reference data is given and over all $N_i$ reference quantities. Reference quantities can be field variables, such as velocities $u$, surface variables, such as the skin friction $c_f$, or integral values such as the lift coefficient $c_l$. The second term, $\mathcal{I}_2$, measures and penalizes the magnitude of the turbulence model modification, acting as a Tikhonov regularization. Regularization is needed as the optimization problem typically is ill-posed due to noise in the reference data and the degrees of freedom, i.e. number of cells where $\beta$ must be optimized, being much higher than the number of cells where reference data is given. The regularization parameter $\lambda$ is selected according to the \textit{L-Curve} criterion \cite{hansen}. Because of the non-linearities in the RANS equations which are contained in $\mathcal{I}$, minimization of $\mathcal{I}$ with respect to $\beta$ must be done iteratively. Hence, a gradient descent method is employed for which the gradient can be computed efficiently using the adjoint method. \paragraph{Machine Learning.} The field inversion step returns discrete values of $\beta$ at the grid nodes, which can't be transferred directly to simulations on different geometries or different flow conditions. Hence, a relation between $\beta$ and $(\mathbf{U}, \tilde{\nu})$ must be found first. Therefore, dimensionless flow features $\eta_i(\mathbf{U}, \tilde{\nu})$ are derived and then machine learning is used to find a model \begin{equation} f_\beta: \eta_0(\mathbf{U}, \tilde{\nu}), \dots, \eta_n \mapsto \beta \label{eq:ml-beta} \end{equation} approximating $\beta$. In particular, neural networks are used for this regression task. Due to the limited space here, the interested reader is referred to dedicated publications such as \cite{deeplearning} for a detailed description of the principles of neural networks. \subsection{Implementation} The Field Inversion and Machine Learning approach has been implemented using the DLR TAU code \cite{TAU}, a highly optimized, parallel, state of the art CFD solver for unstructured grids, using and extending its adjoint capabilities. TensorFlow \cite{tensorflow} is used for training and evaluating the machine learning model. The negative Spalart-Allmaras (SA-neg) \cite{SAneg} model is used as turbulence model due to its focus on aeronautical boundary layer flow. \section{Sensitivity Analysis} \label{chap:numerical_results} For the following sections, field inversions are conducted on the S809 airfoil. This airfoil was developed at the National Renewable Energy Laboratory (NREL) for wind turbines, and pressure distributions from wind tunnel measurements are available for a wide range of angles of attack and for multiple Reynolds numbers \cite{s809}. At an angle of attack of $\alpha \approx 6^\circ$, flow separation starts to develop at the trailing edge and the predictions of the RANS simulations become unreliable, hence making this case suitable for the FIML approach, where it has been used successfully before \cite{fiml}. The field inversions in this section are based on a case with an angle of attack of $\alpha=12.2^\circ$, a Mach number of Ma = 0.1 and a Reynolds number of Re = $2 \times 10^6$ with the experimentally measured surface distribution of $c_p$ from \cite{s809} as reference. \subsection{Regularization} As described in Chapter \ref{chap:methodolgy}, the present optimization problem is ill-posed and therefore requires regularization. For an optimal choice of the parameter $\lambda$, the \textit{L-Curve} criterion \cite{hansen} is considered. Hence, field inversions are conducted for a range of values for $\lambda$ and the resulting cost function terms $\mathcal{I}_1$, $\mathcal{I}_2$ are plotted against each other in the log-log plot in Figure \ref{fig:s809_reg}, on the left. The optimal $\lambda_\text{opt}$ is found at the curve's inflection point, indicating the point beyond which $\mathcal{I}_1$ does not decrease significantly anymore but the magnitude of the model modification represented by $\mathcal{I}_2$ increases fast. For the current case, the optimal value for the regularization parameter is $\lambda = \num{5e-4}$ as indicated by the arrow in Figure \ref{fig:s809_reg}. \begin{figure} \begin{minipage}[t]{.45\textwidth} \def\textwidth {\textwidth} \def0.9818\textwidth {1.2\textwidth} \input{figures/s809_regularization_l-curve.tikz} \end{minipage}% \hfill \begin{minipage}[t]{.55\textwidth} \def\textwidth {\textwidth} \def0.9818\textwidth {0.9818\textwidth} \input{figures/s809_regularization_cp.tikz} \end{minipage} \caption{Left: The L-Curve for the field inversion results for the S809. The optimal $\lambda$ is marked by the arrow. Right: $c_p$-distributions from field inversions for different $\lambda$.} \label{fig:s809_reg} \end{figure} The right side of Figure \ref{fig:s809_reg} shows the $c_p$ distributions of the baseline turbulence model, the field inversion results for $\lambda = \lambda_\text{opt}$, a $\lambda < \lambda_\text{opt}$ and a $\lambda > \lambda_\text{opt}$ and the reference data. The reference data as well as the field inversion results exhibit a pressure plateau beginning at $x/c \approx 0.5$, indicating flow separation at the trailing edge which is not predicted by the baseline model. For $\lambda = \lambda_\text{opt} = \num{5e-4}$, the $c_p$-distribution (solid line) is close to the reference data and the value of cost function part $\mathcal{I}_1$ has decreased to $5.16\%$ of its initial value. For stronger regularization, here e.g. $\lambda = \num{2e-3}$, the $c_p$-distribution (dash-dotted line) deviates further from the reference data with a decrease of cost function $\mathcal{I} _1$ to $20.1\%$ of its initial value. A pressure plateau is still developed, but is predicted to begin further downstream compared to the results for $\lambda = \lambda_\text{opt}$ and the reference data. For weaker regularization, here e.g. $\lambda = \num{1e-6}$, the $c_p$-distribution (dashed line) matches the reference data very closely with a decrease of cost function $\mathcal{I}_1$ down to $3.62\%$ of its initial value. However, overfitting can be observed: In the pressure plateau, the reference data include slight fluctuations, to which the inversion result with the weak regularization tries to fit as well. These fluctuations most probably stem from imprecisions during the process of injecting the reference data and should hence not be incorporated. \begin{figure}[ht] \centering \input{figures/figure_regularization} \caption{Resulting fields for $\beta$ for different values of the regularization parameter $\lambda$. The black contour marks $\beta = 1$.} \label{fig:s809_reg_beta} \end{figure} Resulting $\beta$-fields for all three cases are shown in Figure \ref{fig:s809_reg_beta}. While the $\beta$-field for the stronger regularization ($\lambda = \num{2e-3}$) appears to be equivalent to a downscaled version of the $\beta$-field for $\lambda = \lambda_\text{opt}$, the $\beta$-field for the weak regularization ($\lambda = \num{1e-6}$) shows signs of overfitting by exhibiting more pronounced and additional extrema. These findings confirm the relevance of choosing an optimum $\lambda$ and the L-Curve criterion as a method to do so. \subsection{Grid Resolution} In the next step, the influence of the grid resolution on the field inversion result is investigated. Therefore, field inversions are conducted on six different grids at the same flow conditions as before. The grids were obtained from \cite{s809-grids} and range from resolutions of $353 \times 49$ grid points to $2113 \times 289$ grid points, with $y^+_\text{max}$ ranging from $4.185$ to $0.799$. The outer boundary of the computational domain has a distance of $1000c$ from the airfoil, where $c$ is the airfoil chord length. Grid convergence with the baseline turbulence model is reached for the $1057 \times 145$ grid. For each grid, the optimal regularization parameter $\lambda_\text{opt}$ was determined separately as described in the previous section. The corresponding results for $\beta$ are plotted in Figure \ref{fig:s809_grid_beta}. On all grids, the field inversion process was successful and the cost function could be reduced adequately. In general, the magnitude of the correction $\beta$ appears to increase with the grid resolution. For the $1057 \times 145$ and the $1409 \times 193$ grid, almost no differences in magnitude for the $\beta$-field can be made out in the region of flow separation. For the $2113 \times 289$ grid, additional maxima can be seen, qualitatively similar to those appearing due to over-fitting as found in the chapter before. \begin{figure}% \input{figures/figure_grid} \caption{Field inversion results $\beta$ for different grids.} \label{fig:s809_grid_beta} \end{figure} Since the field inversion shows good results on all grids, including under-resolved grids, it must be assumed that the obtained $\beta$-fields on the under-resolved grids also account for discretization errors. Currently, we are interested only in improving the turbulence model itself and hence use only results from converged grids, but it can be reasoned that a FIML approach could be used to improve simulations on under-resolved grids in general as well. \subsection{Area of Optimization} Finally, the influence of the correction term $\beta$ depending on the area where it is active is investigated. As seen on the left in Figure \ref{fig:s809_regions}, three areas where $\beta$ has changed noticeably can be distinguished, marked with $A$, $B$ and $C$. In region $A$, at the upper surface close to the leading edge, turbulent production was increased by increasing $\beta$. In region $B$, spanning almost the entire upper surface except for the leading edge, turbulent production was reduced by decreasing $\beta$. Region $C$, which encompasses the two spots corresponding to the vortices downstream of the trailing edge, again sees a reduction in turbulent production. \begin{figure} \input{figures/figure_regions} \caption{ Left: The three areas of interest marked by $A$, $B$ and $C$. Right: Results after activating $\beta$ in the different regions. } \label{fig:s809_regions} \end{figure} The influence of the correction term depending on the area where it is active is investigated by running RANS simulations with the augmented model starting with the full $\beta$-field as seen in Figure \ref{fig:s809_regions} and turning off the correction in the different areas one by one by setting $\beta$ to $1$. The different tries are declared as $R_{ABC}$, where $A$, $B$ and $C$ are either $1$ or $0$, depending on whether the correction is active in the respective region or not. The results are shown on the right in Figure \ref{fig:s809_regions}. Turning the correction term off in all regions ($R_{000}$) corresponds to the baseline turbulence model, hence the cost function part $\mathcal{I}_1$ remains unchanged at $100\%$ of its initial value. Activating the correction term in regions $A$ and $C$ decreases $\mathcal{I}_1$ only slightly to $98.0\%\,(R_{100}), 95.8\%\,(R_{001})$ and $93.7\%\,(R_{101})$ of its initial value. The correction term in region $B$ has by far the largest influence by reducing $\mathcal{I}_1$ to $4.53\%\,(R_{010})$ of the initial value alone and to $4.35\%\,(R_{110}), 3.88\%\,(R_{011})$ and $3.84\%\,(R_{111})$ when active in conjunction with regions $A$ and $C$. This behaviour was expected however since region $B$ covers the largest area and especially the area where flow separation occurs. \section{Conclusion} \label{chap:conclusion} The field inversion and machine learning approach was successfully reproduced based on the DLR TAU-code and the negative Spalart-Allmaras model. The influence of the regularization parameter, the grid resolution and the regions where $\beta$ is optimized during field inversion on the field inversion results were investigated based on the flow around the S809 airfoil with flow separation at the trailing edge. For the regularization, a method to determine the optimal regularization parameter was demonstrated and the influence a suboptimal choice can have was discussed. It was shown and discussed that the field inversion can also compensate for spatial discretizaton, returning correct results on coarse grids as well. Different regions where the optimization is active were detected and their influence on the end result was investigated. Under consideration of these information, a machine learning model was trained using data from field inversions of the S809 at multiple angles of attack. The RANS turbulence model was then augmented by the machine learning model and improved results were shown on the S809 airfoil. A future step in the context of the FIML approach is to include inversion results from additional test cases with different geometries and at different flow conditions in the training data, which is expected to be essential to be able to provide a robust and reliable augmentation.

515f197ac9dfa18ce55d10595f689625b243d0ad

\section{Introduction} Ultracold molecules have emerged in recent years as a versatile platform for studies of complex quantum phenomena~\cite{CarrNJP09,QuemenerCR12,BohnScience17}. The rich internal molecular structure and intermolecular interactions have been employed in studies of quantum many-body physics, allowing for the realization of many-body Hamiltionians of yet unexplored complexity~\cite{BaierScience16,GrossScience17}. The controllability of molecular collisions with external magnetic or electric fields, along with precise control over molecular quantum states, have enabled research on ultracold controlled chemical reactions~\cite{OspelkausScience10,TomzaPRL15,HuScience19}. Furthermore, the complexity of molecular structure provides novel possibilities for precision tests of fundamental physics, which include tests of fundamental symmetries, searches for spatiotemporal variations of fundamental constants, tests of quantum electrodynamics, and tests of general relativity, among others~\cite{DeMilleScience17,SafronovaRMP18}. Ultracold molecules containing alkaline-earth-type atoms are promising candidates for high-precision measurement experiments~\cite{McDonaldNature16} and emerging quantum technologies~\cite{KondovNP19}, while alkaline-earth-type atoms have already served as important building blocks of high-precision physics~\cite{DereviankoRMP11}. For example, optical lattice clocks based on the $^1S_0 \rightarrow{}^3P_0$ transition in alkaline-earth-type atoms have played a substantial role in establishing current time and frequency standards~\cite{KatoriPRL03,LudlowScience08,LemkePRL09,LeTargatNatCom13}. Here, the main focus has been put on optical lattice clocks based on strontium~\cite{KatoriPRL03,LudlowScience08,LeTargatNatCom13}, ytterbium~\cite{PoliPRA08, LemkePRL09}, and mercury~\cite{HachisuPRL08,McFerranPRL12,YamanakaPRL15} atoms; however, recent proposals have brought attention to two other suitable candidates, zinc and cadmium atoms~\cite{OvsiannikovPRA16,DzubaJPhysB19,YamaguchiPRA19,PorsevPRA20}. Optical lattice clocks based on group-IIB atoms, such as Zn, Cd, and Hg, have been shown to exhibit reduced susceptibility to the black body radiation (BBR) as compared to Sr- and Yb-based clocks~\cite{HachisuPRL08,McFerranPRL12,SafronovaPRA13,YamanakaPRL15,OvsiannikovPRA16,YamaguchiPRA19}. With BBR being the major factor limiting the accuracy of atomic clocks, Zn and Cd atoms serve as promising alternatives to the currently operational Sr and Yb clocks~\cite{DzubaJPhysB19,PorsevPRA20}. In addition, optical lattice clocks based on alkaline-earth-type atoms are excellent systems for quantum simulations of many-body physics~\cite{KolkowitzNature17,GobanNature18}. Finally, optical clock transitions in divalent atoms have been suggested as a tool to explore potential variations in the fine-structure constant~\cite{AngstmannPRA04, HachisuPRL08} or establishing constraints on the value of the electron's electric dipole moment (EDM)~\cite{HachisuPRL08}. The use of ultracold molecules based on alkaline-earth-type atoms provides further enhancement of sensitivities to the variations of fundamental constants or EDM effects~\cite{SafronovaRMP18}. In this context, one potentially interesting class of molecules is heteronuclear molecules composed of a closed-shell alkaline-earth-like atom interacting with an open-shell atom, such as an alkali-metal atom~\cite{ZuchowskiPRL10} or a halide~\cite{KozlovPRL06}. Such molecules have been proposed to be useful for measuring the variations in the proton-to-electron mass ratio~\cite{KajitaPRA11} and suggested as appealing candidates for searches of the electron's EDM~\cite{KozlovPRL06,MeyerPRA09,PrasannaaPRL15,SunagaPRA18,VermaPRL20}. Moreover, homonuclear dimers of alkaline-earth-type atoms also show prospects for precise measurements of the proton-to-electron mass ratio~\cite{ZelevinskyPRL08,KotochigovaPRA09}, while heteronuclear $^2\Sigma$-symmetry molecules have been proposed as quantum simulators with prospects for creating topologically ordered states~\cite{MicheliNatPhys06}. $^2\Sigma$-state molecules can be formed from ultracold mixtures of closed-shell and open-shell atoms, following recent experimental advances in studies of Yb+Rb~\cite{NemitzPRA09}, Sr+Rb~\cite{BarbeNP18}, Yb+Li~\cite{GreenPRX20}, and Yb+Cs~\cite{WilsonPRA21} combinations. In this work, we propose the formation of ultracold heteronuclear molecules composed of a transition-metal zinc or cadmium atom interacting with an alkali-metal or alkaline-earth-metal atom. The electronic structure of homonuclear dimers of group-IIB atoms, such as Zn$_2$, Cd$_2$, and Hg$_2$, has been the subject of theoretical studies~\cite{PetersonTCA05,PahlTCA11,WeiJCP13,UrbanczykIRPC17}, while heteronuclear molecules composed of a Hg atom interacting with an alkali-metal atom have been investigated both theoretically~\cite{ThielJCP03,SunagaPRA19} and experimentally~\cite{WitkowskiOE17,WitkowskiPRA18,BorkowskiPRA17}. To the best of our knowledge, molecules composed of a Zn or Cd atom interacting with an alkali-metal or alkaline-earth-metal atom have not yet been investigated in the literature. Transition-metal zinc and cadmium atoms, compared to the alkaline-earth-metal atoms, possess a richer structure of excited electronic states due to the possibility of electron excitations from the $d$ subshell. While the ground-state electronic structure of such zinc- or cadmium-containing molecules resembles the electronic structure of alkaline-earth-metal or alkali-metal--alkaline-earth-metal molecules, the richer electronic structure of constituent atoms would have its reflection in a more complex structure of excited electronic states, which may find application in precision measurements~\cite{SafronovaRMP18}. Additionally, zinc or cadmium atom interacting with other atoms may form weakly bound van der Waals molecules that may potentially be used as precise probes of new gravity-like forces~\cite{SalumbidesPRD13,BorkowskiSciRep19}. The ongoing progress in laser cooling and trapping of cadmium atoms~\cite{XuPRA04,BrickmanPRA07,KanedaOL16,YamaguchiPRA19} further motivates our investigation. Here, we theoretically investigate the ground-state properties of diatomic molecules composed of either a Zn or Cd atom interacting with an alkali-metal (Li, Na, K, Rb, Cs, Fr) or alkaline-earth-metal (Be, Mg, Ca, Sr, Ba, Ra) atom. We use state-of-the-art electronic structure methods to calculate the potential-energy curves (PECs) and spectroscopic constants for the investigated molecules. We predict that the considered molecules in the ground electronic state are weakly bound van der Waals complexes, which are chemically reactive. They possess rather small permanent electric dipole moments, despite Zn and Cd atoms having electronegativity significantly larger than that of alkali-metal and alkaline-earth-metal atoms. In this way, the present study extends the range of species available for ultracold molecular studies. This paper is constructed as follows. Section~\ref{sec:methods} introduces the \textit{ab initio} electronic structure methods employed in our calculations. Section~\ref{sec:results} presents and analyzes the obtained numerical data, including the potential-energy curves and electric properties of the investigated molecules. Finally, Sec.~\ref{sec:summary} summarizes our paper. \section{Computational details} \label{sec:methods} In order to calculate the potential-energy curves within the Born-Oppenheimer approximation, we employ the closed-shell and the spin-restricted open-shell coupled-cluster methods restricted to single, double, and noniterative triple excitations [CCSD(T)]. Next, we include the full iterative triple-excitation correction, $\Delta$T, calculated with the use of the coupled-cluster method restricted to single, double, and full triple excitations (CCSDT). We obtain the counterpoise-corrected interaction energies within the supermolecule approach~\cite{BoysMolPhys70}. We use the small-core scalar-relativistic energy-consistent pseudopotentials from the Stuttgart/K\"oln library, ECP$n$MDF, to describe $n$ inner-shell electrons of studied transition-metal atoms, and heavier alkali-metal and alkaline-earth-metal atoms (ECP10MDF for Zn, K, and Ca; ECP28MDF for Cd, Rb, and Sr; ECP46MDF for Cs and Ba; and ECP78MDF for Fr and Ra)~\cite{FiggenCP05, LimJCP05,LimJCP06}. This approach treats only the electrons from the two outermost shells of a given atom explicitly [i.e.,~$3s^23p^63d^{10}4s^2$ from Zn, $4s^24p^64d^{10}5s^2$ from Cd, $(n-1)s^2(n-1)p^6{n}s^1$ from alkali-metal, and $(n-1)s^2(n-1)p^6{n}s^2$ from alkaline-earth-metal atoms], and hence, it allows us to use larger basis sets for more accurate molecular calculations. We correlate all remaining electrons. For the presented computations at the CCSD(T) level of theory, we employ the corresponding correlation-consistent polarized weighed core-valence quintuple-$\zeta$ quality basis sets (aug-cc-pwCV5Z-PP~\cite{PetersonTCA05,HillJCP17} with ECP and aug-cc-pwCV5Z~\cite{PrascherTCA11} for Li, Na, Be, and Mg) augmented by the set of the $[3s3p2d2f1g]$ bond functions. To account for the full triple-excitation correction ($\Delta$T), we perform electronic structure calculations at the CCSDT level of theory with the use of valence-only triple-$\zeta$ quality basis sets (aug-cc-pVTZ for Li, Na, Be, and Mg atoms, and aug-cc-pVTZ-PP for the remaining atoms). Additionally, for two representative systems, an open-shell RbZn molecule and a closed-shell SrZn molecule, we carry out convergence tests to analyze the accuracy of the obtained interaction energies and confirm the optimal method and basis sets for the remaining calculations. To this end, we compute interaction energies using the CCSD(T) method and aug-cc-pwCV$n$Z-PP basis sets with $n =$ D, T, Q, 5. We use these basis sets to extrapolate the interaction energies to the complete-basis-set (CBS) limit and show that adding a bond function (BF) to aug-cc-pwCV5Z-PP basis sets allows us to reproduce the CBS limit accurately. Next, we obtain the full iterative triple-excitation correction ($\Delta$T), given as a difference between interaction energies calculated at the CCSDT and CCSD(T) levels of theory, in smaller basis sets (aug-cc-pV$n$Z-PP, with $n =$ D, T, Q, and aug-cc-pwCVDZ-PP). Analogously, we estimate the magnitude of noniterative and iterative quadruple excitations [$\Delta$(Q) and $\Delta$Q] using the CCSDT(Q) and CCSDTQ methods, respectively, with the aug-cc-pVDZ-PP and aug-cc-pVTZ-PP basis sets. For completeness, we also compare the PECs obtained within the coupled-cluster method with the ones calculated using the multireference configuration interaction method restricted to single and double excitations (MRCISD). The permanent electric dipole moments and static electric dipole polarizabilities are calculated using the finite-field method at the CCSD(T)/aug-cc-pwCV5Z level of theory. The $z$ axis is chosen along the internuclear axis and oriented from a Zn or Cd atom to an alkali-metal or alkaline-earth-metal atom. All electronic structure calculations are performed using the MOLPRO package of \textit{ab initio} programs~\cite{MOLPRO-WIREs,MOLPRO}. The full triple and quadruple contributions are computed using the MRCC code embedded in MOLPRO~\cite{MRCC}. Vibrational eigenstates are calculated numerically by employing the exact diagonalization of the nuclear motion Hamiltonian within the discrete variable representation (DVR) on the nonequidistant grid~\cite{DVR}. Atomic masses of the most abundant isotopes are assumed. \section{Results and discussion} \label{sec:results} \subsection{Potential-energy curves} \begin{figure*}[!tb] \begin{center} \includegraphics[width=0.95\linewidth]{fig1} \end{center} \caption{Potential-energy curves of (a) the $AM$Zn molecules in the $X^2\Sigma^+$ electronic state, (b) the $AEM$Zn molecules in the $X^1\Sigma^+$ electronic state, (c) the $AM$Cd molecules in the $X^2\Sigma^+$ electronic state, and (d) the $AEM$Cd molecules in the $X^1\Sigma^+$ electronic state.} \label{fig:potentials} \end{figure*} \begin{table*}[!tb] \caption{Spectroscopic characteristics of the $AM$Zn and $AM$Cd molecules in the $X\,^2\Sigma^+$ electronic state and $AEM$Zn and $AEM$Cd molecules in the $X\,^1\Sigma^+$ electronic state: equilibrium bond length $R_e$, well depth $D_e$, harmonic constant $\omega_e$, first anharmonicity constant $\omega_e x_e$, number of bound vibrational states $N_\nu$, rotational constant $B_e$, permanent electric dipole moment $d_e$, parallel and perpendicular components of the static electric dipole polarizability $\alpha^\parallel_e$ and $\alpha^\perp_e$, and long-range dispersion-interaction coefficient $C_6$. The results for the Zn$_2$, Cd$_2$, and ZnCd molecules are also presented.\label{tab:characteristics}} \begin{ruledtabular} \begin{tabular}{lccccccccccc} Molecule & State & $R_e$ (bohr) & $D_e$ (cm$^{-1}$) & $\omega_e$ (cm$^{-1}$) & $\omega_e x_e$ (cm$^{-1}$) & $N_\nu$ & $B_e$ (cm$^{-1}$) & $d_e$ (D) & $\alpha^\parallel_e$ (a.u.) & $\alpha^\perp_e$ (a.u.) & $C_6$ (a.u.)~\cite{QiaoJCP12} \\ \hline LiZn & $X\,^2\Sigma^+$ & 5.67 & 827 & 131 & 6.84 & 17 & 0.296 & 0.30 & 375 & 164 & 541 \\ NaZn & $X\,^2\Sigma^+$ & 6.61 & 474 & 50.9& 1.51 & 24 & 0.081 & 0.20 & 315 & 173 & 597\\ KZn & $X\,^2\Sigma^+$ & 7.57 & 395 & 34.8 & 0.35 & 28 & 0.043 & 0.18 & 457 & 291 & 837 \\ RbZn & $X\,^2\Sigma^+$ & 7.93 & 368 & 26.6& 0.45 & 34 & 0.026 & 0.16 & 481 & 321 & 959 \\ CsZn & $X\,^2\Sigma^+$ & 8.33 & 354 & 23.4& 0.62 & 37 & 0.020 & 0.12 & 558 & 390 & 1129 \\ FrZn & $X\,^2\Sigma^+$ & 8.45 & 339 & 21.5& 0.41 & 39 & 0.017 & 0.11 & 467 & 331& 1056\\ BeZn & $X\,^1\Sigma^+$ & 6.91 & 218 & 39.9& 1.88 & 11 & 0.159 &-0.03 & 101 & 69 & 270 \\ MgZn & $X\,^1\Sigma^+$ & 7.39 & 280 & 33.2& 0.46 & 18 & 0.063 & -0.003 & 148 & 99 & 450\\ CaZn & $X\,^1\Sigma^+$ & 7.60 & 404 & 33.4& 0.30 & 28 & 0.042 & -0.08 & 271 & 175 & 771\\ SrZn & $X\,^1\Sigma^+$ & 7.92 & 415 & 27.4& 0.23 & 35 & 0.026 & -0.07 & 321 & 215 & 916\\ BaZn & $X\,^1\Sigma^+$ & 8.17 & 446 & 25.7& 0.70 & 40 & 0.021 & -0.12 & 417 & 287 & 1138 \\ RaZn & $X\,^1\Sigma^+$ & 8.49 & 396 & 22.2& 0.21 & 41 & 0.017 & -0.02 & 373 & 266 & 1044 \\ LiCd & $X\,^2\Sigma^+$ & 5.80 & 988 & 134& 5.41 & 19 & 0.270 & 0.54 & 409 & 165 & 708 \\ NaCd & $X\,^2\Sigma^+$ & 6.66 & 596 & 53.2& 1.19 & 28 & 0.071 & 0.40 & 350 & 175 & 783 \\ KCd & $X\,^2\Sigma^+$ & 7.58 & 515 & 36.4& 1.50 & 34 & 0.036 & 0.44 & 511 & 288 & 1090 \\ RbCd & $X\,^2\Sigma^+$ & 7.92 & 482 & 26.6& 0.30& 44 & 0.020 & 0.42 & 537 & 318 & 1251 \\ CsCd & $X\,^2\Sigma^+$ & 8.32 & 467 & 22.7& 0.33 & 50 & 0.014 & 0.40 & 624 & 383 & 1470 \\ FrCd & $X\,^2\Sigma^+$ & 8.43 & 443 & 20.1& 0.31 & 54 & 0.011 & 0.34 & 519 & 328 & 1381 \\ BeCd & $X\,^1\Sigma^+$ & 6.95 & 255 & 42.2& 1.59 & 12 & 0.149 & -0.03 & 115 & 75 & 365 \\ MgCd & $X\,^1\Sigma^+$ & 7.44 & 329 & 33.8& 0.48 & 21 & 0.055 & 0.02 & 164 & 105 & 605 \\ CaCd & $X\,^1\Sigma^+$ & 7.62 & 496 & 34.5& 1.38 & 33 & 0.035 & -0.02 & 299 & 179 & 1023 \\ SrCd & $X\,^1\Sigma^+$ & 7.93 & 513 & 26.7& 0.12 & 44 & 0.019 & 0.01 & 352 & 219 & 1212 \\ BaCd & $X\,^1\Sigma^+$ & 8.17 & 563 & 24.7& 0.10 & 53 & 0.014 & -0.02 & 456 & 290 & 1499\\ RaCd & $X\,^1\Sigma^+$ & 8.49 & 494 & 20.4& 0.20 & 55 & 0.011 & 0.08 & 406 & 269 & 1335 \\ Zn$_2$ & $X\,^1\Sigma^+_g$ & 7.23 & 231 & 23.4& 0.62& 22 & 0.036 & 0 & 97 & 70 & 359 \\ ZnCd & $X\,^1\Sigma^+$ & 7.28 & 275 & 22.6 & 0.53 & 27 & 0.028 & 0.01 & 110 & 76 & 495 \\ Cd$_2$ & $X\,^1\Sigma^+_g$ & 7.32 & 330 & 21.1 & 0.24 & 35 & 0.020 & 0 & 124 & 83 & 686 \\ \end{tabular} \end{ruledtabular} \end{table*} We consider interactions between a zinc or cadmium atom and an alkali-metal $AM$ ($AM$ = Li, Na, K, Rb, Cs, Fr) or alkaline-earth-metal $AEM$ ($AEM$ = Be, Mg, Ca, Sr, Ba, Ra) atom in their electronic ground states. The ground-state Zn and Cd atoms, as well as alkaline-earth-metal atoms, are described with the $^1 S_0$ electronic term, while the alkali-metal atoms are described with the $^2 S_{1/2}$ electronic term. This yields the $^2\Sigma^+$ molecular electronic states for the ground-state open-shell molecules composed of a Zn or Cd atom and an alkali-metal atom and the $^1 \Sigma^+$ molecular electronic states for the ground-state closed-shell molecules composed of a Zn or Cd atom and an alkaline-earth-metal atom. For the above molecules, we compute the potential-energy curves and provide spectroscopic characteristics: the equilibrium bond lengths $R_e$, potential-well depths $D_e$, harmonic constants $\omega_e$, first anharmonicity constants $\omega_e x_e$, numbers of bound vibrational states $N_\nu$, and rotational constants $B_e$. We also report the permanent electric dipole moments $d_e$ and parallel and perpendicular components of the static electric dipole polarizabilities, $\alpha^\parallel_e$ and $\alpha^\perp_e$, respectively, at equilibrium distances. The computed curves are presented in Fig.~\ref{fig:potentials}, and obtained characteristics are collected in Table~\ref{tab:characteristics}. We estimate the number of bound vibrational states $N_\nu$ with the use of the DVR method, in which we employ the computed PECs, describing the short-range part of the interaction, smoothly connected with the long-range part of the interaction, $-C_6/R^6$, where the dispersion-interaction coefficients $C_6$ are taken from Ref.~\cite{QiaoJCP12} and presented in Table~\ref{tab:characteristics} for completeness. We also provide the results for the ground-state homonuclear Zn$_2$ and Cd$_2$ dimers described with the $^1 \Sigma_g^+$ molecular term and the ZnCd molecule in the $^1 \Sigma^+$ electronic ground state. Figure~\ref{fig:potentials} presents the potential-energy curves of the $AM$Zn and $AM$Cd molecules in the $X\,^2\Sigma^+$ electronic ground state, and $AEM$Zn and $AEM$Cd molecules in the $X\,^1\Sigma^+$ electronic ground state calculated at the CCSD(T)+$\Delta$T level of theory. All PECs exhibit a smooth behavior with well-defined minima. For the $AM$Zn and $AM$Cd molecules, the well depths systematically decrease with the increasing atomic number of the alkali-metal atoms (a deviation from the trend is observed for radium-containing molecules), while for the $AEM$Zn and $AEM$Cd molecules, the well depths systematically increase with the increasing atomic number of the alkaline-earth-metal atoms. The opposite trends can be explained by different characters of bonding within molecules containing alkali-metal and alkaline-earth atoms: the open-shell $AM$Zn and $AM$Cd molecules are bound chemically (with a bond order of $\frac{1}{2}$), while the closed-shell $AEM$Zn and $AEM$Cd molecules are bound solely by the dispersion forces. All considered molecules are of van der Waals character, with moderate equilibrium distances and well depths not exceeding 1000$\,$cm$^{-1}$. We also notice that molecules containing cadmium are more strongly bound than molecules containing zinc due to the larger polarizability of the cadmium atom. The well depths of the $AM$Zn molecules in the ground $X\,^2\Sigma^+$ electronic state range from 827$\,$cm$^{-1}$ for LiZn to 339$\,$cm$^{-1}$ for FrZn, systematically decreasing with the atomic number of the alkali-metal atom $AM$. The equilibrium distances range from 5.67$\,$bohrs for LiZn to 8.45$\,$bohrs for FrZn, systematically increasing with the atomic number of $AM$. We observe the same trend for the $AM$Cd molecules in the ground $X\,^2\Sigma^+$ electronic state, whose well depths range from 988$\,$cm$^{-1}$ for LiCd to 433$\,$cm$^{-1}$ for FrCd, and equilibrium distances increase from 5.80$\,$bohrs for LiCd to 8.43$\,$bohrs for FrCd. The computed number of vibrational levels increases with the reduced mass of the molecule, from 17 and 19 for LiZn and LiCd to 39 and 54 for FrZn and FrCd, respectively. For the $AEM$Zn molecules in the ground $X\,^1\Sigma^+$ electronic state, the well depth systematically increases from 218$\,$cm$^{-1}$ for BeZn to 446$\,$cm$^{-1}$ for BaZn and slightly drops to 396$\,$cm$^{-1}$ for RaZn. $AEM$Cd molecules in the ground $X\,^1\Sigma^+$ electronic state are characterized by well depths which also systematically increase with the atomic number of $AEM$, ranging from 255$\,$cm$^{-1}$ for BeCd to 563$\,$cm$^{-1}$ for BaCd and 494$\,$cm$^{-1}$ for RaCd. The equilibrium distance systematically increases from 6.91$\,$bohrs for BeZn to 8.49$\,$bohrs for RaZn and from 6.95$\,$bohrs for BeCd to 8.49$\,$bohrs for RaCd. The estimated number of vibrational levels amounts to 11 and 12 for BeZn and BeCd and increases with the reduced mass of the molecule up to 41 and 55 for RaZn and RaCd, respectively. The observed trends in the studied molecules are similar to those reported for analogous alkali-metal--alkaline-earth-metal and alkaline-earth-metal molecules~\cite{GueroutPRA10,HeavenCPL11,PototschnigPCCP16}. However, the potential-well depths are smaller, and equilibrium distances are larger in the present case due to smaller polarizabilities of the Zn and Cd atoms than those of alkaline-earth-metal atoms. \begin{figure}[tb!] \begin{center} \includegraphics[width=1\linewidth]{fig2} \end{center} \caption{Potential-energy curves of the Zn$_2$ and Cd$_2$ molecules in the $X\,^1\Sigma^+_g$ electronic state and the ZnCd molecule in the $X\,^1\Sigma^+$ electronic state.} \label{fig:potentials_ZnCd} \end{figure} For completeness, we also provide results for the homonuclear Zn$_2$ and Cd$_2$ and heteronuclear ZnCd molecules in their ground $X\,^1\Sigma^+_g$ and $X\,^1\Sigma^+$ electronic states, respectively. The calculated spectroscopic characteristics are collected in Table~\ref{tab:characteristics}. Figure~\ref{fig:potentials_ZnCd} presents the PECs, which were calculated at the CCSD(T)+$\Delta$T+$\Delta$(Q) level of theory. The potential-well depths amount to 231, 275, and 330$\,$cm$^{-1}$ for Zn$_2$, ZnCd, and Cd$_2$, respectively, and the respective equilibrium distances are 7.23, 7.28, and$\,$7.32 bohrs. The estimated number of vibrational levels is equal to 22 for Zn$_2$, 27 for ZnCd, and 35 for Cd$_2$. The values of the well depths $D_e$ and harmonic constants $\omega_e$ obtained for homonuclear dimers are in good agreement with the results of previous theoretical calculations and spectroscopic measurements, which are compared in Table~\ref{tab:Zn2Cd2}. Like previous theoretical works, we observe discrepancies between the calculated equilibrium distances and their experimental values, especially for Zn$_2$. \begin{table}[h!] \caption{Spectroscopic constants of the Zn$_2$ and Cd$_2$ molecules in the $X\,^1\Sigma^+_g$ electronic state: Comparison with previous studies.} \label{tab:Zn2Cd2} \begin{ruledtabular} \begin{tabular}{llccc} Molecule & Source & $R_e\,$(bohr) & $D_e\,$(cm$^{-1}$) & $\omega_e$ (cm$^{-1}$)\\ \hline Zn$_2$ & This work & 7.23 & 231 & 23.4 \\ & Theory~\cite{PetersonTCA05} & 7.27 & 226 & 23.9 \\ & Theory~\cite{PahlTCA11} & 7.23 & 226 & 24.0 \\ & Theory~\cite{WeiJCP13} & 7.32 & 242 & 25.65 \\ & Experiment~\cite{StrojeckiCP06} & 7.92 & 242 & 25.9 \\ Cd$_2$ & This work & 7.32 & 330 & 21.1 \\ & Theory~\cite{PetersonTCA05} & 7.36 & 325 & 20.2 \\ & Theory~\cite{PahlTCA11} & 7.32 & 319 & 21.3 \\ & Theory~\cite{WeiJCP13} & 7.75 & 328 & 21.5 \\ & Experiment~\cite{StrojeckiCPL10} & 7.14$\pm$0.06 & 328$\pm$3 & 21.4$\pm$0.2 \\ \end{tabular} \end{ruledtabular} \end{table} \subsection{Permanent electric dipole moments and static electric dipole polarizabilities} The permanent electric dipole moments of the $AM$Zn and $AM$Cd molecules in the $X\,^2\Sigma^+$ electronic state and the $AEM$Zn and $AEM$Cd molecules in the $X\,^1\Sigma^+$ electronic states as functions of the internuclear distance are presented in Fig.~\ref{fig:dipoles}. The values of permanent electric dipole moments at equilibrium distances are collected in Table~\ref{tab:characteristics}. They govern the strength of the intermolecular dipolar interaction and coupling with an external static electric field. \begin{figure*}[!tb] \begin{center} \includegraphics[width=0.95\linewidth]{fig3} \end{center} \caption{Permanent electric dipole moments of (a) the $AM$Zn molecules in the $X^2\Sigma^+$ electronic state, (b) the $AEM$Zn molecules in the $X^1\Sigma^+$ electronic state, (c) the $AM$Cd molecules in the $X^2\Sigma^+$ electronic state, and (d) the $AEM$Cd molecules in the $X^1\Sigma^+$ electronic state. The points mark the permanent electric dipole moments at equilibrium distances.} \label{fig:dipoles} \end{figure*} The $AM$Zn and $AM$Cd molecules in the $X\,^2\Sigma^+$ electronic state have small permanent EDMs, not exceeding 0.54 debye at equilibrium distances. The values of the EDMs at equilibrium distances systematically decrease with the increasing atomic number of the alkali-metal atom, ranging from 0.30 debye for LiZn to 0.11 debye for FrZn and from 0.54 for LiCd to 0.34 for FrCd. The permanent EDMs of the $AEM$Zn and $AEM$Cd molecules in the $X\,^1\Sigma^+$ electronic state take even smaller values, not exceeding 0.12 debye at equilibrium distances. The equilibrium-distance EDMs take values ranging from 0.003 debye for MgZn to 0.12 debye for BaZn and from 0.01 for SrCd to 0.08 for RaCd, with no distinct systematics of atomic number dependence. In contrast to the $AM$Zn molecules, the permanent electric dipole moments of the $AEM$Zn molecules point from the alkaline-earth-metal atom to the zinc atom. For BeCd, CaCd, and BaCd molecules, the dipoles are oriented from the alkaline-earth-metal atom to the cadmium atom, while for MgCd, SrCd, and RaCd, the dipoles are oriented from the cadmium atom to the alkaline-earth-metal atom. The permanent electric dipole moments of all the studied molecules take very small values despite relatively large electronegativity differences between involved atoms. The electronegativity by the Pauling scale of the Zn (1.65) and Cd (1.69) atoms is almost two times larger than that of the alkali-metal (0.79-0.98) and alkaline-earth-metal (0.89-1.57) atoms~\cite{Pauling1960}. For such large differences, significant permanent EDMs, larger than those for analogous alkali-metal--alkaline-earth-metal and alkaline-earth-metal molecules~\cite{GueroutPRA10,HeavenCPL11,PototschnigPCCP16}, could be expected (similar to what was recently reported for molecules containing Cu and Ag atoms~\cite{SmialkowskiPRA21}). Unfortunately, the present values are significantly smaller, partially as a result of weak interatomic interactions and van der Waals nature of the studied molecules and partially because around minima the permanent EDMs cross zero and change sign. At large distances, the permanent EDMs present an expected systematic dependence on the electronegativity differences. We also calculate the parallel, $\alpha^\parallel_e$, and perpendicular, $\alpha^\perp_e$, components of the static electric dipole polarizability tensor, which play an important role in the evaluation of intermolecular interactions and coupling of molecular rovibrational dynamics with a laser field. The values of the parallel and perpendicular components of static electric dipole polarizabilities at equilibrium distances are collected in Table~\ref{tab:characteristics}. \subsection{Convergence and accuracy analysis} In order to investigate the uncertainty of the present molecular electronic structure calculations, we first examine whether the employed \textit{ab initio} methods describe the atomic properties accurately. To this end, we employ the CCSD(T) method to calculate the atomic static electric dipole polarizabilities and ionization potentials of the Zn and Cd atoms using the aug-cc-pwCV5Z-PP basis sets. The obtained atomic polarizabilities of the zinc and cadmium atoms amount to 37.7~and 45.8$\,$a.u.~and agree well with the recommended combined experimental-theoretical values of 38.7$\pm$0.3~and 46$\pm$2$\,$a.u.~\cite{SchwerdtfegerMP19}, respectively. The calculated ionization potentials of 75848$\,$cm$^{-1}$ for Zn and 72526$\,$cm$^{-1}$ for Cd are also in good agreement with respective experimental values of 75769$\,$cm$^{-1}$~\cite{SugarJCP95} and 72540$\,$cm$^{-1}$~\cite{ShenstoneOSA49}. The atomic polarizabilities and ionization potentials of the alkali-metal and alkaline-earth-metal atoms obtained with the used theoretical methods are also in good agreement with experimental data, as confirmed in Ref.~\cite{SmialkowskiPRA21}. Next, we analyze the convergence of the interatomic interaction energy with the size of the employed basis sets. Figure~\ref{fig:conv_basis} presents the potential-energy curves for the representative RbZn and SrZn molecules in the $X\,^2\Sigma^+$ and $X\,^1\Sigma^+$ electronic states, respectively, obtained with different basis sets. An inspection of Fig.~\ref{fig:conv_basis} allows us to conclude that the inclusion of inner-shell electron correlation is crucial for an accurate description of the interatomic interactions and the core-core and core-valence contributions are significant, especially for the $AM$Zn and $AM$Cd molecules. The addition of a bond function to the aug-cc-pwCV5Z-PP basis set allows describing the complete-basis-set-limit energy accurately. \begin{figure}[!tb] \begin{center} \includegraphics[width=1\linewidth]{fig4} \end{center} \caption{Potential-energy curves of (a)~the RbZn molecule in the $X\,^2\Sigma^+$ electronic state and (b)~the SrZn molecule in the $X\,^1\Sigma^+$ electronic state, computed with the CCSD(T) method, using different-sized Gaussian basis sets. CBS limit energy for the aug-cc-pwCV$n$Z basis sets is also presented.} \label{fig:conv_basis} \end{figure} \begin{figure}[!tb] \begin{center} \includegraphics[width=1\linewidth]{fig5} \end{center} \caption{Potential energy curves of (a)~the RbZn molecule in the $X\,^2\Sigma^+$ electronic state and (b)~the SrZn molecule in the $X\,^1\Sigma^+$ electronic state computed at different levels of theory: RHF, MRCISD, MRCISD+Q, MP2, CCSD, CCSD(T), CCSD(T)+$\Delta$T, and CCSD(T)+$\Delta$T+$\Delta$Q. See the text for details.} \label{fig:conv_method} \end{figure} Finally, we analyze the convergence of the interatomic interaction energy with the quality of employed wave-function representation. Figure~\ref{fig:conv_method} presents the potential-energy curves for the representative RbZn and SrZn molecules in the $X\,^2\Sigma^+$ and $X\,^1\Sigma^+$ electronic states, respectively, calculated at different levels of theory: spin-restricted Hartree-Fock RHF, MRCISD, MRCISD+Q, second-order M\o{}ller-Plesset perturbation theory MP2, CCSD, CCSD(T), CCSD(T)+$\Delta$T, and CCSD(T)+$\Delta$T+$\Delta$Q. The aug-cc-pwCV5Z-PP+BF basis set is used in the CCSD(T) calculations, aug-cc-pVTZ-PP is used in the calculation of the full triple correction $\Delta$T, and aug-cc-pVDZ-PP is used in the calculation of the full quadruple correction $\Delta$Q. In the coupled-cluster calculations, the inclusion of higher-order excitations significantly improves the description of the interaction energies. The obtained potential-well depths for the RbZn and SrZn molecules are equal to, respectively, 352 and 391$\,$cm$^{-1}$ at the CCSD(T) level, 368 and 415$\,$cm$^{-1}$ at the CCSD(T)+$\Delta$T level, and 372 and 435$\,$cm$^{-1}$ at the CCSD(T)+$\Delta$T+$\Delta$Q level. Hence, the full triple correction increases the well depth by about 4\% for RbZn and about 6\% for SrZn, while the full quadruple correction introduces a further 1\% well-depth increase for RbZn and about 5\% for SrZn. Moreover, for the Zn$_2$, ZnCd, and Cd$_2$ dimers, the perturbative quadruple correction $\Delta$(Q) (calculated with the use of the aug-cc-pVTZ-PP basis set) increases the potential-well depths obtained at the CCSD(T)+$\Delta$T level by 5\% to 7\%. Therefore, an accurate description of the interatomic interactions between closed-shell transition-metal atoms, like zinc and cadmium, and alkali-metal or alkaline-earth-metal atoms has to take into consideration higher-order electron correlation, in contrast to alkali-metal dimers with two valence electrons~\cite{GronowskiPRA20}. The triple-excitation contribution is significant for both open-shell $AM$Zn and $AM$Cd and closed-shell $AEM$Zn and $AEM$Cd molecules, while the quadruple-excitations contribution is particularly large for closed-shell $AEM$Zn and $AEM$Cd molecules. It is in agreement with the fact that the CCSDT method for the $AM$Zn and $AM$Cd molecules with three valence electrons already provides the description of valence electrons at the full configuration interaction level, while the $AEM$Zn and $AEM$Cd molecules with four valence electrons require the CCSDTQ method for the same. We also find that the values of the full-iterative triple and full-iterative quadruple corrections systematically decrease with the increasing size of the basis sets, leading us to conclude that the calculated corrections may be slightly overestimated. Overall, we estimate that the uncertainty of our calculations should be at most 5\% for molecules containing alkali-metal atoms and slightly more for molecules containing alkaline-earth atoms. In the MRCISD calculations, the full-valence active space is used. The well depths of the potential-energy curves obtained at the MRCISD level amount to 22.8$\,$cm$^{-1}$ for RbZn and 51.5$\,$cm$^{-1}$ for SrZn. The addition of the Davidson correction, MRCISD+Q, yields deeper PECs, with the well depths of 103 and 175$\,$cm$^{-1}$, respectively, for RbZn and SrZn, yet the results are still not comparable to those obtained with the coupled-cluster method. We also calculate the PECs within the MRCISD method with $4p$ Zn orbitals and $5p$ Rb/Sr orbitals included in the active space; however, this approach leads to almost identical results. Interestingly, those PECs do not differ significantly from the ones obtained at the CISD level, meaning that the ground electronic states of RbZn and SrZn are well described by a single reference, and higher-order excitations need to be taken into account to reproduce the coupled-cluster results. We also see that the energies obtained within the second-order M{\o}ller-Plesset perturbation theory (MP2) are overestimated, particularly for the SrZn molecule. \subsection{Chemical reactions} We use the calculated potential-well depths $D_e$ to assess the stability of the investigated molecules against atom-exchange chemical reactions~\cite{ZuchowskiPRA10,TomzaPRA13,SmialkowskiPRA20}. A ground-state heteronuclear molecule $AB$ can undergo an atom-exchange chemical reaction, \begin{equation} AB + AB \rightarrow A_2 + B_2, \end{equation} provided that the sum of the dissociation energies $D_0$ of the products, $A_2$ and $B_2$, is larger than or equal to the sum of the dissociation energies of the reactants, $2 AB$, \begin{equation} D_0(A_2) + D_0(B_2) \ge 2 D_0(AB). \end{equation} The dissociation energy $D_0$ is related to the potential-well depth $D_e$, $D_0 \approx D_e - \frac{1}{2} \omega_e$. All the studied $AM$Zn, $AM$Cd, $AEM$Zn, and $AEM$Cd molecules in the ground rovibrational levels of their ground electronic states are chemically unstable and atom-exchange reactions are energetically possible, \begin{equation} \begin{split} 2 AM\mathrm{Zn}(X^2\Sigma^+) &\to \mathrm{Zn}_2 (X^1\Sigma^+_g) + AM_2 (X^1\Sigma^+_g) \,,\\ 2 AM\mathrm{Cd}(X^2\Sigma^+) &\to \mathrm{Cd}_2 (X^1\Sigma^+_g) + AM_2 (X^1\Sigma^+_g) \,,\\ 2 AEM\mathrm{Zn}(X^1\Sigma^+) &\to \mathrm{Zn}_2 (X^1\Sigma^+) + AEM_2 (X^1\Sigma^+_g) \,,\\ 2 AEM\mathrm{Cd}(X^1\Sigma^+) &\to \mathrm{Cd}_2 (X^1\Sigma^+_g) + AEM_2 (X^1\Sigma^+_g) \,, \end{split} \end{equation} because the well depths of alkali-metal $AM_2$ and alkaline-earth-metal $AEM_2$ dimers~\cite{ZuchowskiPRA10} are significantly larger than those of the present molecules. The atom-exchange chemical reactions for the $AM$Zn and $AM$Cd molecules could potentially be suppressed by restricting the collision dynamic to the high-spin potential-energy surfaces by fully spin polarizing the molecules in an external magnetic field~\cite{TomzaPRA13}. Unfortunately, only channels leading to alkali-metal dimers in the $X^3\Sigma^+_u$ electronic state are closed, while ones leading to Cd$_2$ and Zn$_2$ in the $X^3\Sigma^+_u$ electronic state are open, \begin{equation} \begin{split} 2 AM\mathrm{Zn}(X^2\Sigma^+) & \not\to \mathrm{Zn}_2 (X^1\Sigma^+_g) + AM_2 (a^3\Sigma^+_u) \,,\\ 2 AM\mathrm{Cd}(X^2\Sigma^+) & \not\to \mathrm{Cd}_2 (X^1\Sigma^+_g) + AM_2 (a^3\Sigma^+_u) \,.\\ 2 AM\mathrm{Zn}(X^2\Sigma^+) & \to \mathrm{Zn}_2 (a^3\Sigma^+_u) + AM_2 (X^1\Sigma^+_g) \,,\\ 2 AM\mathrm{Cd}(X^2\Sigma^+) & \to \mathrm{Cd}_2 (a^3\Sigma^+_u) + AM_2 (X^1\Sigma^+_g) \,.\\ \end{split} \end{equation} Finally, most likely, the trimer formation reactions are another path of chemical losses for all considered molecules~\cite{ZuchowskiPRA10,TomzaPRA13,SmialkowskiPRA20}, \begin{equation} AB + AB \to A_2B + B, \end{equation} but their detailed study is out of the scope of this work. \section{Summary and conclusions} \label{sec:summary} Motivated by the recent progress in laser cooling and trapping of cadmium atoms~\cite{XuPRA04,BrickmanPRA07,KanedaOL16,YamaguchiPRA19} and experimental realizations of ultracold mixtures of closed-shell and open-shell atoms~\cite{NemitzPRA09,BarbeNP18,GreenPRX20,WilsonPRA21}, in this paper, we brought attention to diatomic molecules composed of a closed-shell zinc or cadmium atom interacting with an alkali-metal (Li, Na, K, Rb, Cs, Fr) or alkaline-earth-metal (Be, Mg, Ca, Sr, Ba, Ra) atom. Such molecules are potential candidates for ultracold quantum physics and chemistry experiments, ranging from controlled chemical reactions to precision measurements. To this end, we have carried out state-of-the-art \textit{ab initio} calculations of the potential energy curves, permanent electric dipole moments, and spectroscopic constants for the molecules in their electronic ground states. We have used the \textit{ab initio} electronic structure coupled-cluster method with single, double, and triple excitations combined with large Gaussian basis sets and small-core relativistic energy-consistent pseudopotentials for heavier atoms. We have predicted that the studied molecules in the ground electronic state are weakly bound van der Waals complexes. We have also found that they have rather small permanent electric dipole moments, despite Zn and Cd atoms having electronegativity significantly larger than that of alkali-metal and alkaline-earth-metal atoms. Finally, we have concluded that they are chemically reactive, and for applications other than studies of ultracold chemical reactions, they should be segregated in an optical lattice, or shielding strategies should be employed~\cite{Anderegg2021}. Full potential energy curves and permanent electric dipole moments as a function of interatomic distance in a numerical form are collected in the Supplemental Material~\footnote{See Supplemental Material at http://link.aps.org/supplemental/XXXX for the calculated potential energy curves and permanent electric dipole moments in a numerical form.} \begin{acknowledgments} We acknowledge the financial support from the National Science Centre of Poland (Grant No. 2016/23/B/ST4/03231) and the Foundation for Polish Science within the First Team program co-financed by the European Union under the European Regional Development Fund. The computational part of this research was partially supported by the PL-Grid Infrastructure. \end{acknowledgments}

75a7640e77b0e34c7d1f3633ecca965ce95ec608

\section{Introduction} Wireless networked control systems (WNCSs) are composed of spatially distributed sensors, actuators, and controllers communicating through wireless networks \cite{Park2018}. Despite their success in industrial monitoring applications, existing wireless sensor-actuator network technologies face significant challenges in supporting control systems due to their lack of real-time performance and dynamic wireless conditions in industrial plants \cite{Lu2016}. A key challenge in WNCSs design is the channel modeling in an industrial environment because of its inherent complexity \cite{Lu2016} \cite{Alhen2019}. These communication channels are frequently subject to time-varying fading and interference, which may lead to packet losses. From the automatic control perspective, an example of WNCS consists of a nonlinear process with intermittent control packets due to the lossy communication channel described by the following equations: \begin{equation}\label{eq:mjs} \begin{cases} y\left( k+1 \right)=f \left(y(k), u_a(k)\right), k \in \mathbb{N} \\ u_a(k) = \nu(k)u(k) \\ y(0) = y_0 \in \mathbb{R}^{n_y} \end{cases} \end{equation} where $y(k) \in \mathbb{R}^{n_y}$ is the output of the system and $u(k)\in \mathbb{R}^{n_u}$ is the input signal. $\left\{\nu(k)\right\}_{k\in\mathbb{N}}$ is a discrete-time Boolean process modeling the packet delivery of the control signals: if the packet is correctly delivered then $\nu(k)=1$, otherwise if it is lost then the actuator does nothing $\nu(k)=0$. The control packet error probability (PEP) $\mathbb{P}(\nu(k)=0)$, depends on the signal-to-interference-plus-noise-ratio (SINR) of the communication $\Gamma(k)$, i.e., \begin{equation}\label{eq:Bern} \mathbb{P}(\nu(k)=0) = g\left( \Gamma (k)\right), \end{equation} where $g:\mathbb{R} \to \left[0,1\right]$ is a deterministic function defined by the communication standard. We assume full state observation with no measurement noise, and no observation packet loss. In this work, we consider a channel model based on WirelessHART \cite{Chen2010}, a on-the-market wireless communication standard specifically designed for process automation. We assume that the value of the SINR $\Gamma(k)$ is measurable and is sent to the controller via an acknowledgment (ACK). However, this ACK is available only after the current decision on the control input to apply has been made and sent through the link, since the actual success of the transmission is not known in advance. Hence, at each time $k$ the controller receives measurements of $y(k)$ and $\Gamma(k-1)$, as depicted in Fig.~\ref{fig:WNCS}. \begin{figure}[h!] \begin{center} \centerline{\includegraphics[width=0.65\columnwidth]{figures/WNCS.png}} \caption{Wireless networked control system} \label{fig:WNCS} \end{center} \end{figure} \vspace{-20pt} Since $\Gamma(k)$ is represented by a generic stochastic process, the obtained model may not be computationally tractable when it is derived for wireless communications in industrial environment (see Sec.~\ref{sec:channel}), especially if the objective is to apply optimal control algorithms (e.g. model predictive control -- MPC). For this reason, a preliminary investigation of channel model abstraction is fundamental to design a controller. In the WNCS literature the packet dropouts have been modeled either as stochastic or deterministic phenomena. For what concerns stochastic models, a vast amount of research assumes memoryless packet drops, so that dropouts are realizations of a Bernoulli process. In \cite{Schenato2007}, the packet delivery process is modeled as a Bernoulli random process, then an optimal controller is derived. In \cite{Lun2019} a Markov chain model is used to derive an accurate abstraction of the WirelessHART channel, and it is proven that such model allows to characterize the stability of a WNCS for the scenarios where the simple Bernoulli-like channel models (which cannot model packet bursts) fail. Markov models are a powerful tool for modeling stochastic random processes. They are general enough to model with high accuracy a large variety of processes and are relatively simple, allowing us to compute analytically many important parameters of the process which are very difficult to calculate for other models \cite{Turin1998}. Hidden Markov models (\cite{Bilmes1998b} \cite{Lawrance1989}) have been exploited to learn channels models in \cite{Alhen2019} and \cite{Salamatian2001}. In \cite{Sadghi2008}, the authors expose the benefit of exploiting finite-state Markov chains to model the behavior of wireless fading channels. In \cite{Lun2020}, the authors derive a Markov chain that estimates the packet error probability (PEP) of an industrial wireless protocol, and then propose an optimal stochastic controller for linear systems. Inspired by the challenges in \cite{Lun2020} related to the derivation of a physics-based Markov chain abstraction of the wireless channel, the main contribution of this paper is a novel data-driven methodology, based on regression trees \cite{Breiman2017classification}, to identify such Markov chain abstraction. More precisely, we propose a novel methodology to model the PEP $\nu(k)$ on the basis of a Markov chain $\theta(k)$: each state of $\theta(k)$ is associated to a partition of $\mathbb{R}^{{n_y}+1}$ consisting of rectangular sets $\{R_i\}_{i=1}^{\ell}$, each representing the range of possible values of $\Gamma(k) \in \mathbb R$ and $y(k) \in \mathbb{R}^{{n_y}}$ at time $k$. We construct the transition probability matrix (TPM) of $\theta(k)$ as follows: \begin{align}\label{eqMCidentification}\small p(j \mid i) &\doteq \mathbb{P}(\theta(k+1)=j \mid \theta(k)=i)\notag \\ &= \mathbb{P}(\theta(k+1)=j \mid (\Gamma(k), y(k)) \in R_i)\notag \\ &= h(\Gamma(k), y(k)), \end{align} where identifying from historical data the function $h:\mathbb{R}^{{n_y}+1} \to \left[0,1\right]$, which depends on the current measurements $\Gamma(k)$ and $y(k)$, is the objective of this paper. Note that, given any two states $i,j$ of the Markov chain, $p(j\mid i)$ also depends on the plant output: indeed, $y(k)$ may for example determine the distances between the transmitter, the receiver and the interferer, and, as illustrated in Sec.~\ref{sec:channel}, this strongly affects the dynamics of $\Gamma(k)$. Finally, the PEP can be easily computed using \eqref{eq:Bern}, which in this paper is based on a point-to-point WirelessHART communication based on the IEEE 802.15.4 standard. Nevertheless, we remark that our data-driven methodology is independent on the transmission technology and can be replicated for any communication protocol if the corresponding wireless channel can be effectively modeled by a Markov chain. In summary, in an industrial environment the derivation of a Markov model based on the wireless communication physics can be prohibitive as it requires a complete knowledge of both the communication dynamics parameters and of the disturbances/interferers. Our methodology has 3 main advantages: (1) it only exploits historical data, hence does not require any a-priori knowledge of the system and channel parameters. Moreover, most of physics based approaches cannot handle time-varying parameters, as the computational complexity of the obtained abstraction would be intractable, while with our methodology we consider a dependency between the parameters of the communication system and the dynamics of the plant which is still tractable in terms of computational complexity; (2) as a byproduct of leveraging our techniques in \cite{SmarraNAHS2020}, and beyond the main contribution of identifying the Markov chain abstraction of $\Gamma(k)$, we also construct a switching auto-regressive exogenous (SARX) model for the nonlinear plant $f$ in \eqref{eq:mjs}, i.e. our methodology does not require any linearity assumption on the plant's dynamics; (3) the obtained identified models, both for $\Gamma(k)$ and $f$, can be combined obtaining a Markov jump system, which can be directly used to setup a classical MPC problem that can be solved very efficiently, i.e. using quadratic programming (QP). \section{Channel modeling} \label{sec:channel} In this section we first describe the channel model under consideration, then we present the Markovian model for the wireless link developed in \cite{Lun2020}, and finally we introduce an analytical model that will be useful to experiment the data-driven approach we propose in this paper. \subsubsection{WirelessHART physical model} We analyze the industrial environment described in Fig. \ref{fig:WNCS}, wherein the wireless communication is affected by interference. We study a point to point transmission based on the WirelessHART protocol, i.e. the IEEE 802.15.4-2006 defined in \cite{2006}, interfered by another WirelessHART transmission. The proposed model considers the effect of path loss, the shadow fading, and the residual power fluctuations left by the power control. The effect of multipath fading is supposed to be compensated by the aforementioned power control. We denote with $i\!=\!0$ the reference Transmitter-Receiver (Tx-Rx) link, and with $i\!=\!1$ the link between the interferer Tx and the reference Rx. The effect of the path loss model is defined in \cite{2006}. For a system with bandwidth $W\!=\!2.4$GHz the path loss coefficient of the link between transmitter $i$ and the reference receiver (i.e., a mobile plant in our case) is $\alpha_i(k) \!=\! 10^{-\frac{\varsigma(d_i(k))}{10}}$, where \begin{gather}\label{eq:path} \varsigma(d_i(k)) = \begin{cases} 40.2+20\log_{10}(d_i(k)), \text{ if } d_i(k) \leq 8, \\ 58.5 + 33 \log_{10}(\frac{d_i(k)}{8}), \text{ otherwise;} \end{cases} \end{gather} and $d_i(k)$ is the length of the link $i$ at time instant $k$, i.e.~a distance in meters that may e.g.~depend on the position of the plant (see the inverted pendulum on a cart in Sec.~\ref{sec:NR}). The shadow fading is modeled following \cite{Goldsmith2005} by assuming a log-normal model for each link $i$, which introduces a multiplicative factor $e^{\beta_i(k)}$, where $\beta_i(k)$ is a zero-mean Gaussian process with variance $\sigma_{\!\beta_i}^{2}$ and auto-covariance function $c_{\beta_i}(\tau)$, with $\tau$ being a time lapse between two consecutive (time-driven) control packets. We remark that $c_{\beta_i}(\tau)$ may also depend on the state of the plant (e.g. the speed of a cart) thus exhibiting a time-varying behavior. For each link $i$ the residual power control error (PCE) is also modeled as a log-normal process, $e^{\xi_i(k)}$, where $\xi_i(k)$ is a zero-mean Gaussian process with variance $\sigma_{e_i}^2$ and auto-covariance $c_{\xi_i}(\tau)$. By considering the characteristics of the offset quadrature phase-shift keying (OQPSK) with direct-sequence spread spectrum (DSSS) modulation, as specified in \cite{2006}, we can derive the power value of the SINR, $\gamma(k)$, at time $k$ as in \cite{Lun2020}: \begin{equation}\label{eq:sinr} \gamma(k) = \sqrt{\frac{P_0(k)\alpha_0^2(k)e^{\beta_0(k)+\xi_0(k)}}{\frac{N_0}{4}+\frac{8}{3G}P_1(k)\alpha_1^2(k)e^{\beta_1(k)+\xi_1(k)}}}, \end{equation} where $P_0$ and $P_1$ are respectively the reference user transmission power and the interferer transmission power, $N_0$ is the noise spectral density, and $G=WT_s$ is the processing gain, with $T_s$ being the symbol time. In the rest of the work we denote as $\Gamma(k)=10\log_{10}(\gamma(k))$ the SINR in decibel. In WirelessHART the forward error correction is not implemented, thus even one erroneous bit leads to a corrupted WirelessHART data packet. For this reason, the packet error probability $R_p$ is related to the SINR as follows \begin{align} \label{eq:Rp} R_b(\gamma(k))&=\frac{1}{30} \sum_{\iota=2}^{16}(-1)^{\iota} \binom{16}{\iota}\exp\left(20\cdot \gamma(k)\frac{1-\iota}{\iota}\right)\nonumber \\ R_p(\gamma(k))&=1-\left(1-R_b\left(\gamma\left(k\right)\right)\right)^{l_f}, \end{align} where $l_f$ is the number of bits in a frame, and $R_b$ is the bit error rate computed according to \cite{2006}. It is worth remarking that the distance Tx-Rx $d_i(k)$ influences the SINR $\gamma(k)$ that, in turn, influences the packet error rate. This will be useful for the discussion of the simulation results in Sec.~\ref{sec:NR}. The main issue with the above channel model is that, despite the ability to describe the SINR, in the majority of the control applications equation \eqref{eq:sinr} is intractable, both in the continuous-time and discrete-time form. For this reason, in \cite{Lun2020}, a Markov chain model to derive a discrete-time abstraction of \eqref{eq:sinr} has been proposed. We recall such technique in the following subsection as a comparison for the method we propose in this paper. \subsubsection{Finite-state Markov chain}\label{ssec:FSMC} It is straightforward to see that \eqref{eq:sinr} can be expressed as a weighted linear combination of correlated log-normal processes: hence there is no exact explicit closed form expression of its distribution. We can use the moment matching technique \cite{fischione2007} to approximate the probability distribution of \eqref{eq:sinr} by a log-normal process, thus presenting $\Gamma(k)$ as a Gaussian process with mean $\mu_{\Gamma}(k)$, variance $\sigma_{\Gamma}^2(k)$ and auto-covariance $c_{\Gamma}(\tau)$, as detailed in \cite{Lun2020}. Clearly, due to the dependence of the SINR on the state of a mobile plant through, e.g. the parameters $d_i(k)$ (which define the path loss coefficients $\alpha_i(k)$ via \eqref{eq:path}), the moment matching approximation should be done for each relevant value of the aforementioned parameters. In the rest of this subsection we will focus on a Gaussian process $\hat{\Gamma}(k)$, which is a moment matching approximation of the SINR with arbitrary values of the parameters in \eqref{eq:sinr} corresponding to any given state of the plant, i.e. for $\alpha_i(k)\!=\!\hat{\alpha}_i$, $\beta_i(k)=\hat{\beta}_i(k)$, and $P_i(k)\!=\!\hat{P}_i$ in \eqref{eq:sinr}, with $i\!=\!0,1$, we have that $\hat{\Gamma}(k)\approx 10\log_{10}(\gamma(k))$. At this point, to obtain a finite-state Markov channel abstraction, the first step is to divide the range of $\hat{\Gamma}(k)$ into several consecutive regions, each associated with a certain representative PEP. Specifically, a region $r$ of the values of the SINR is mapped into a state $\boldmath{s_r}$ of the related Markov chain, and it is delimited by two thresholds $\zeta_r$ and $\zeta_{r+1}$ belonging to the set of extended reals. These SINR thresholds are determined by the chosen partitioning method \cite{Sadghi2008}. In this paper, we rely on a well-known equiprobable partitioning, where the thresholds are selected in such a way that the steady state probabilities of being in any state are equal. Then, the steady state probability $\boldsymbol{p_r}$ of a state $\boldmath{s_r}$ is defined as the probability that the value of $\hat{\Gamma}(k)$ is between the two thresholds of the region, and it is given by \begin{equation} \boldsymbol{p_r} = \int_{\zeta_r}^{\zeta_{r+1}} \phi_{\mathcal{N}}(\zeta;\mu_{\hat{\Gamma}},\sigma_{\hat{\Gamma}}^2)d\zeta, \end{equation} where $\mu_{\hat{\Gamma}}$ and $\sigma_{\hat{\Gamma}}^2$ are respectively the mean and variance of $\hat{\Gamma}(k)$, and $\phi_{\mathcal{N}}(\zeta;\mu,\sigma^2)$ is the probability density function (PDF) of a Gaussian random variable (GRV) $\mathcal{N}(\mu,\sigma^2)$. The Packet Delivery (PDP) associated to the same state within the respective region is given by \begin{equation} 1-\nu_M^{(r)} = \frac{1}{\boldsymbol{p_{r}}} \int_{\zeta_r}^{\zeta_{r+1}}R_p (10^{\frac{\zeta}{10}})\phi_{\mathcal{N}}(\zeta;\mu_{\hat{\Gamma}},\sigma_{\hat{\Gamma}}^2)d\zeta. \end{equation} Finally, the channel state transition probabilities are derived from integrating the joint PDF of the SINR $\hat{\Gamma}(k)$ over two consecutive packet transmissions and over the desired regions, $r$ and $q$, as \begin{equation*} p(q\mid r)\!=\!\frac{1}{\boldsymbol{p_r}}\!\! \int_{\zeta_{r}}^{\zeta_{r+1}}\!\!\!\!\int_{\zeta_{q}}^{\zeta_{q+1}}\!\!\!\!\! \varphi_{\mathcal{N}}\!\left(\varsigma_{k-1},\varsigma_{k};\mu_{\hat{\Gamma}},\sigma_{\hat{\Gamma}}^2,c_{\hat{\Gamma}}(\tau)\!\right)\!d\varsigma_{k-1}d\varsigma_{k}, \end{equation*} where $\varphi_{\mathcal{N}}(\varsigma_{k-1},\varsigma_{k};\mu_{\hat{\Gamma}},\sigma_{\hat{\Gamma}}^2,c_{\hat{\Gamma}}(\tau))$ is the two-dimensional PDF of the Gaussian process {$\hat{\Gamma}(k)$}, as detailed in \cite{Lun2020}. However, one issue related to the above technique is that in real (industrial) cases some of the channel parameters required for the above modeling are time-varying (due to their dependence on the state of the plant) and often only partially known. For these reasons, we propose a new approach to model the SINR, and thus the PEP. In this respect, a promising direction is the exploitation of historical data of the communication channel to identify a model via system identification and machine learning techniques. \subsubsection{Auto-regressive model for channel simulation}\label{sec:AR} The application of data-driven methodologies requires the existence of a dataset containing trajectories of the SINR: in this paper we run Monte Carlo simulations of the wireless transmission model \eqref{eq:sinr} and then apply our techniques to such trajectories. To this aim, in this section we illustrate the approach presented in \cite{Baddour2005}, where the study of auto-regressive stochastic models for computer simulation of fading channels is addressed, to derive discrete-time trajectories of the process described in equation \eqref{eq:sinr}. In particular, let us consider a discrete-time Gaussian process $\left\{z(k)\right\}_{k\in \mathbb{N}}$ with auto-correlation function (ACF) $R_{zz}(n)$. We can derive an auto-regressive (AR) model of the following form that is able to generate trajectories of the process: \begin{equation} z(k) = -\sum\nolimits_{n=1}^{p}a_nz(k-n)+w(k), \end{equation} where $w(k)$ is a zero-mean white Gaussian noise process. The AR model parameters consist of coefficients $\{a_1,\ldots, a_p\}$ and variance $\sigma_p^2$ of the driving noise $w(k)$. To estimate the coefficients $a_j$, $j = 1,\ldots,p$, once the ACF $R_{zz}$ is fixed from $\beta_i$ and $\xi_i$, $i = 0,1$, the relationship between $R_{zz}$ and $a_j$ is given as follows \cite{Kay1988}: \begin{equation}\label{eq:Rzz} \small R_{zz}(n) = \begin{cases} -\sum_{m=1}^{p}a_mR_{zz}(n-m), & n \geq 1 \\ -\sum_{m=1}^{p}a_mR_{zz}(-m) + \sigma_p^2, & k=0 \end{cases} \end{equation} % % Finally, coefficients $a_j$ can be determined solving the set of Yule-Walker equations that can be easily derived from \eqref{eq:Rzz} (we refer the reader to \cite{Kay1988} for further details). The generated AR process has the following auto-correlation function: \begin{equation} \hat{R}_{zz}(n) = \begin{cases} R_{zz}(n), & 0\leq n \leq p \\ -\sum_{m=1}^{p}a_m \hat{R}_{zz}[n-m], & n>p \end{cases} \end{equation} The simulated process has the attractive property that its sampled ACF perfectly matches the desired sequence of ACF up to lag $p$. Therefore, since we know the ACF of both the residual power control error $\xi_i(k)$ and the shadowing correlation $\beta_i(k)$ we can exploit the AR model to generate sequences of $\xi_i(k)$ and $\beta_i(k)$ and then apply \eqref{eq:sinr} to obtain the discrete-time trajectories of $\Gamma(k)$. \section{The Classification And Regression Tree (CART) algorithm}\label{ssec:BackgroundRT} The aim of this section is to provide a short description of the classification and regression trees (CART) algorithm \cite{Breiman2017classification}, in order to provide the basic notions to present our method in Sec.~\ref{sec:RTMC} that identifies a Markov chain wireless channel abstraction. In a supervised framework, we consider a predictor dataset $\mathcal{P} = \{\lambda(k)\}_{k=1}^{D}$ and a response dataset $\mathcal{R}=\{\rho(k)\}_{k=1}^{D}$ of $D$ samples each, where $\rho(k) \in \mathbb{R}$ is called response variable and $\lambda(k) \in \mathbb{R}^n$ is called predictor variable. The final goal of CART is to identify a function $\mathcal{T}$ to estimate $\hat \rho(k)=\mathcal{T}(\lambda(k))$. In the specific case of the CART algorithm \cite{Breiman2017classification}, the dataset is partitioned into a set of hyper-rectangles $R_1,\ldots,R_\ell$, corresponding to the $\ell$ leaves of the tree. Then, $\hat \rho(k)$ is estimated in each leaf $\tau_i$ using a constant $c_{\tau_i}$ given by the average of the samples in the partition. Without any loss of generality we restrict our attention to recursive binary partition. Due to space limitations, we only briefly recall the partitioning algorithm of CART, and refer the reader to \cite{Breiman2017classification} for more details. The CART algorithm creates the partition using a greedy algorithm to optimize the split variables and split points: starting with the whole dataset, consider a split variable $j$ over the $n$ available and a split point $s$, and define the pair of half-planes as $R_L(j,s) =\{\lambda(k)\mid \lambda_j(k) < s\}$ and $R_R(j,s) = \{\lambda(k) \mid \lambda_j(k) \geq s\}$. Then, CART solves the following optimization problem to find the optimal $j$ and $s$ \begin{equation}\label{eq:minCART} \scriptsize \min_{j,s} \left[\min_{c_L} \sum_{\lambda(k) \in R_L(j,s)} (\rho(k)-c_L)^2 + \min_{c_R} \sum_{\lambda(k) \in R_R(j,s)} (\rho(k)-c_R)^2\right], \end{equation} \normalsize and for any choice of $j$ and $s$ the inner minimization is solved by $c_L = \text{ave}(\rho(k) \mid \lambda(k) \in R_L(j,s))$ and $c_R = \text{ave}(\rho(k) \mid \lambda(k) \in R_R(j,s))$, where $\text{ave}(\cdot)$ is the arithmetic mean of the output samples. In other words, the optimal $j^*$ and $s^*$ minimize the sum of the quadratic prediction errors of the left and right partitions induced by the split variable and split point. For each splitting variable, the determination of the split point $s$ can be done very quickly and hence, by scanning through all of the inputs, the determination of the best pair $(j, s)$ is feasible. Once the best split is found the dataset is partitioned into the two resulting regions, then the splitting procedure is repeated on each of the two regions. The process is repeated on all of the resulting regions until a stopping criterion is applied, e.g. tree size is a tuning parameter chosen to avoid overfitting and variance phenomena. In the rest of this work we denote with $\mathcal{T}$ the regression tree, with $\tau_i$ the $i^{th}$ leaf of $\mathcal{T}$, with $|\mathcal{T}|$ the number of leaves of $\mathcal{T}$, with $|\tau_i|$ the number of samples in $\tau_i$, with $c_{\tau_i} = \text{ave}(\rho(k)| \lambda(k) \in \tau_i)$ the prediction of leaf $\tau_i$ and, with a slight abuse of notation, with $\mathcal{T}(\lambda)$ the prediction of the regression tree, i.e. \begin{equation}\label{eq:RTpred}\small \mathcal{T} (\lambda) = \sum_{\tau_i \in \mathcal{T}} \sum_{ \lambda(k) \in \tau_i} \frac{\rho(k)}{|\tau_i|} I \left\{\lambda \in \tau_i\right\} = \sum_{\tau_i \in \mathcal{T}} c_{\tau_i} I \left\{\lambda \in \tau_i\right\}, \end{equation} where $I \left\{\lambda \in \tau_i \right\}$ is the indicator function, which is equal to 1 if $\lambda \in \tau_i$ and 0 otherwise. \section{Switching ARX Identification}\label{ssec:BackgroundSARX} As discussed above, in this paper we leverage the techniques in \cite{SmarraNAHS2020} to construct a switching auto-regressive exogenous (SARX) model for the nonlinear plant $f$ in \eqref{eq:mjs} that can be directly used to setup a MPC problem, as will be done in Sec.~\ref{sec:NR}. In particular, starting from a dataset $\mathcal{D}=\{(y(k),u(k))\}_{k=1}^{D}$ of $D$ samples collected from the measurements of a physical system, respectively consisting of outputs $y(k)\in\mathbb{R}^{n_y}$ and inputs $u(k)\in\mathbb{R}^{n_u}$, we will derive for each $j=0,\ldots,N-1$ a model as follows: \footnotesize \begin{align}\label{eqIdentifiedModelN} \nonumber x(k+j+1) = A_{\sigma_j(x(k))} x(k+j)+ B_{\sigma_j(x(k))} u(k+j) + F_{\sigma_j(x(k))}, \end{align} \normalsize \noindent with $\sigma_j:\mathbb{R}^{n_x}\rightarrow\mathcal{M}\subset\mathbb{N}$ the switching signal, $x(k)\doteq\left[y^\top(k)\ \cdots\ y^\top(k-\delta_y)\ u^\top(k-1)\ \cdots\ u^\top(k-\delta_u)\right]^\top\in\mathbb{R}^{n_x}$ the state consisting of the regressive terms of the inputs and the outputs, $\delta_y,\delta_u\geq 0$, and $n_x = (\delta_y + 1)n_y + n_u \delta_u$. % % % % \section{Markov model based on regression trees}\label{sec:RTMC} In industrial environments, the derivation of Markov models based on the physics of wireless communication can be prohibitive as it requires complete knowledge of both the communication dynamics parameters and the disturbances/interferers. Furthermore, the presence of time-varying parameters can increase the complexity of the obtained model. We propose a novel methodology to derive a Markov model of the PDP based on the WNCS measurements. The approach exploits the transmissions' historical data to deal with the circumstances illustrated above. In particular, we handle the case wherein there is a dependency between the communication system's and the plant measurable outputs. The main idea is to derive a Markovian model $\left\{\theta(k) \in \Theta \right\}_{k \in \mathbb{N}}$ abstracting the SINR stochastic process $\Gamma(k)$, with TPM $P$ as in \eqref{eqMCidentification}, and associate to each state $i\in \Theta$ a PDP. The learning procedure consists of three steps: (1) we grow a regression tree $\mathcal{T}$ with predictor variables the current SINR and the plant measurements, i.e. $\left(\Gamma(k), x(k) \right)$, and with response variable the SINR at the next time step, i.e. $\Gamma(k+1)$. Since the leaves of $\mathcal{T}$ form a partition of the predictor space, we define the state-space of $\theta(k)$ associating to each element of the partition $\{R_{\tau_i}\}_{i=1}^{\ell_\mathcal{T}}$ a state of the Markov chain, i.e. $\Theta \doteq \left\{ i \in \mathbb{N}: \tau_i \in \mathcal{T} \right\}$; (2) we grow a regression tree $\Pi$ with predictor variables the current SINR, i.e. $\Gamma(k)$, and with response variable the prediction of the SINR at the next time step obtained using the regression tree $\mathcal{T}$, i.e. $\mathcal{T}(\Gamma(k),x(k))$. We will combine $\Pi$ and $\mathcal{T}$ to identify the TPM of $\left\{\theta(k) \right\}_{k \in \mathbb{N}}$; (3) we associate to each state of $\left\{\theta(k) \right\}_{k \in \mathbb{N}}$ a PDP. \subsubsection{Step 1} We grow a regression tree $\mathcal{T}$ using as predictor dataset $\left\{\Gamma(k), x(k) \right\}_{k=1}^{D} = \mathcal{P}_\mathcal{T} \subset \mathbb{R}^{n_x+1}$, consisting of the current SINR and the plant measurable outputs, and as response dataset $\left\{ \Gamma(k+1) \right\}_{k=1}^{D} = \mathcal{R}_\mathcal{T} \subset \mathbb{R}$, consisting of the SINR at the next time step. Let $\mathcal{T}(\left( \Gamma(k), x(k)\right))$, with $\mathcal{T}: \mathcal{P}_\mathcal{T} \to \mathcal{R}_\mathcal{T}$, be the prediction of $\mathcal{T}$ given the current measurements. As mentioned above, we associate to each element of the partition $\{R_{\tau_i}\}_{i=1}^{\ell_\mathcal{T}}$ a state of the Markov chain $\left\{\theta(k) \in \Theta \right\}_{k \in \mathbb{N}}$, i.e. $\Theta \doteq \left\{ i \in \mathbb{N}: \tau_i \in \mathcal{T} \right\}$. Note that $\forall \left(\Gamma(k), x(k)\right) \in \mathbb{R}^{n_x+1}, \exists! \ \tau_i \in \mathcal{T}: \left(\Gamma(k), x(k)\right) \in R_{\tau_i}$, i.e. the current measurement $\left(\Gamma(k), x(k)\right)$ is deterministically associated to one, and only one, discrete state of $\theta(k)$. As each leaf $\tau_i$ contains a random subset of the SINR sequence, we assume independent and identically distributed samples belonging to the same partition element. We fit a GRV $G_{\tau_i} \sim \mathcal{N}(\mu_i,\sigma_i^2)$ to model the current level of SINR in each leaf $\tau_i \in \mathcal{T}$ of the tree: \begin{gather}\small \begin{aligned} &\mu_{i} =\frac{1}{|\tau_i|}\sum_{ \left( \Gamma(k), x(k)\right) \in R_{\tau_i}} \Gamma(k), \\ &\sigma_{i}^2 = \frac{1}{|\tau_i|-1} \sum_{ \left( \Gamma(k), x(k)\right) \in R_{\tau_i}} (\Gamma(k)-\mu_{i})^2. \end{aligned} \end{gather} In Sec.~\ref{sec:pep} we will exploit $\left\{G_{\tau_i} \right\}_{\tau_i \in \mathcal{T}}$ to estimate the PEP in each state of $\left\{\theta(k)\right\}_{k\in \mathbb{N}}$. Following the same reasoning, we fit a GRV $G_{\tau_i}^+ \sim \mathcal{N}(\mu_{i^+},\sigma_{i^+}^2)$ to model the next-step level of SINR for each leaf $\tau_i \in \mathcal{T}$ of the tree: \begin{gather}\small\label{eq:cp} \begin{aligned} \mu_{i^+} &=\frac{1}{|\tau_i|}\sum_{ \left( \Gamma(k), x(k)\right) \in R_{\tau_i}} \Gamma(k+1), \\ \sigma_{i^+}^2 &= \frac{1}{|\tau_i|-1} \sum_{ \left( \Gamma(k), x(k)\right) \in R_{\tau_i}} (\Gamma(k+1) - \mu_{i^+})^2. \end{aligned} \end{gather} In Sec.~\ref{sec:TPM} we will exploit $\left\{G_{\tau_i}^+ \right\}_{\tau_i \in \mathcal{T}}$ to identify the TPM of $\left\{\theta(k)\right\}_{k\in \mathbb{N}}$. We remark that the TPM of $\left\{\theta(k)\right\}_{k\in \mathbb{N}}$ could be identified using only $\mathcal{T}$, based on the number of samples that at time $k$ stay in a leaf $\tau_i\in \mathcal{T}$, and at time $k+1$ jump to a leaf $\tau_j \in \mathcal{T}$, i.e. $p(j \mid i) = \tilde{n}(\tau_i,\tau_j) \cdot |\tau_i|^{-1}$ where \small \begin{align}\label{eq:naive} \tilde{n}(i,j) \doteq &|\left\{ \left( \Gamma(k_\epsilon), x(k_\epsilon)\right) \in \mathcal{P}_\mathcal{T} : \right. \\ & \left. \left( \Gamma(k_\epsilon), x(k_\epsilon)\right) \in R_{\tau_i}, \left( \Gamma(k_\epsilon+1), x(k_\epsilon+1)\right) \in R_{\tau_j} \right\}|.\notag \end{align} \normalsize The lack of this approach is that the tree $\mathcal{T}$ partitions the dataset to minimize the prediction error of a deterministic estimate of $\Gamma(k+1)$, instead of minimizing the prediction error with respect to the estimate $\mathbb{E} \left[\mathcal{T}(\Gamma(k),x(k)) \mid \Gamma(k) \right]$. We overcome this issue in the next section growing an additional regression tree $\Pi$, and combining $\mathcal{T}$ and $\Pi$ to identify the TPM to minimize the error with respect to $\mathbb{E} \left[\mathcal{T}(\Gamma(k),x(k)) \mid \Gamma(k) \right]$. \subsubsection{Step 2}\label{sec:TPM} The idea is to define a partition $\left\{ R_{\pi_r}\right\}_r^{|\Pi|}$ of $\mathbb{R}$ minimizing the prediction error with respect to the estimate $\mathbb{E} \left[\mathcal{T}(\Gamma(k),x(k)) \mid \Gamma(k) \right]$ , and exploit the Markov property to split the identification process of the TPM of $\left\{\theta(k) \right\}_{k\in \mathbb{N}}$ in two steps: \begin{gather} \small \label{eq:tpm} \begin{aligned} &\mathbb{P}(\theta(k+1)=j \mid \theta(k)=i) \\& \doteq \mathbb{P}((\Gamma(k+1), x(k+1)) \in R_{\tau_j} \mid (\Gamma(k),x(k)) \in R_{\tau_i}) \\ &=\sum_{\pi_r \in \Pi} \{ \mathbb{P}((\Gamma(k+1), x(k+1)) \in R_{\tau_j} \mid \Gamma(k+1) \in R_{\pi_r}) \\ &\mathbb{P}(\Gamma(k+1) \in R_{\pi_r} \mid (\Gamma(k),x(k)) \in R_{\tau_i}) \}. \end{aligned} \end{gather} \begin{figure}[h!] \begin{center} \centerline{\includegraphics[width=0.9\columnwidth]{figures/trees.png}} \caption{TPM identification via regression trees} \label{figtrees} \end{center} \end{figure} To estimate $\mathbb{P}((\Gamma(k+1), x(k+1)) \in R_{\tau_j} \mid \Gamma(k+1) \in R_{\pi_r})$ we define a new predictor dataset $\{\Gamma(k) \}_{k=1}^{D} = \mathcal{P}_\Pi \subset \mathbb{R}$ and a new response dataset $\{\mathcal{T}(\Gamma(k),x(k)) \}_{k=1}^{D} = \mathcal{R}_\Pi\subset \mathbb{R}$. Then, we derive a new regression tree $\Pi$ applying the CART algorithm to identify the function $\Pi: \mathcal{P}_\Pi \to \mathcal{R}_\Pi$. Let $ \Pi(\Gamma(k))$ be the optimal estimation given by the tree $\Pi$. Define \begin{gather}\label{eq:tp} \scriptsize \begin{aligned} \mathbb{P}((\Gamma(k+1), x(k+1)) \in R_{\tau_j} \mid \Gamma(k+1) \in R_{\pi_r}) \doteq p(\tau_j\mid\pi_r) = \frac{n(\pi_r,\tau_j)}{|\pi_r|} \end{aligned} \end{gather} where $n(\pi_r,\tau_j)$ is the number of samples that belong both the leaf $\pi_r \in \Pi$ and the leaf $\tau_j \in \mathcal{T}$, i.e \scriptsize \begin{equation}\label{eq:n} n(\pi_r,\tau_j) \doteq |\{( \Gamma(k_\epsilon), x(k_\epsilon)) \in \mathcal{P}_\mathcal{T} : \Gamma(k_\epsilon) \in R_{\pi_r}, ( \Gamma(k_\epsilon), x(k_\epsilon)) \in R_{\tau_j} \}|. \end{equation} \normalsize Notice that in contrast to the definition in equation \eqref{eq:naive}, equation \eqref{eq:n} does not involve the time evolution. For this reason, eq. \eqref{eq:n} is useful to estimate the probabilities $\mathbb{P}((\Gamma(k+1), x(k+1)) \in R_{\tau_j} \mid \Gamma(k+1) \in R_{\pi_r})$. \begin{proposition}\label{prop:p} Let us define the transition probabilities $p(\tau_j\mid\pi_r)$ as in Equation \eqref{eq:tp}, then the CART algorithm creates the partition induced by $\left\{ R_{\pi_r} \right\}_{\pi_r \in \Pi}$ optimally estimating the conditional expectation of the prediction in each leaf $\pi_r\in\Pi$, i.e. $c_{\pi_r} =\mathbb{E} \left[\mathcal{T}(\Gamma(k),x(k)) \mid \Gamma(k) \in R_{\pi_r} \right], \forall \pi_r \in \Pi$. \end{proposition} \scriptsize \begin{proof} \begin{align} &c_{\pi_r} = \sum_{\Gamma(k_\epsilon) \in R_{\pi_r}} \frac{\mathcal{T}(\Gamma(k_\epsilon), x(k_\epsilon))}{|\pi_r|}\\ &= \sum_{\Gamma(k_\epsilon) \in R_{\pi_r}} \frac{1}{|\pi_r|} \sum_{\substack{\left(\Gamma(k_\epsilon'), x(k_\epsilon') \right)\in R_{\tau_j} \\ \left(\Gamma(k_\epsilon), x(k_\epsilon) \right) \in R_{\tau_j} }} \frac{\Gamma(k_\epsilon'+1)}{|\tau_j|}\\ &= \sum_{\substack{\Gamma(k_\epsilon) \in R_{\pi_r}\\ \left(\Gamma(k_\epsilon), x(k_\epsilon) \right) \in R_{\tau_j} }} \frac{1}{|\pi_r|} \sum_{\substack{\left(\Gamma(k_\epsilon'), x(k_\epsilon') \right)\in R_{\tau_j} }} \frac{\Gamma(k_\epsilon'+1)}{|\tau_j|}\\ &= \sum_{\substack{\Gamma(k_\epsilon) \in R_{\pi_r}\\ \left(\Gamma(k_\epsilon), x(k_\epsilon) \right) \in R_{\tau_j} }} \frac{c_{\tau_j}}{|\pi_r|}=\sum_{\tau_j \in \mathcal{T}} \frac{c_{\tau_j} \cdot n(\pi_r, \tau_j)}{|\pi_r|} =\sum_{\tau_j \in \mathcal{T}} p(\tau_j \mid \pi_r)c_{\tau_j} \\ &\simeq \sum_{\tau_j \in \mathcal{T}} p(\tau_j \mid \pi_r) \mathbb{E}\left[\mathcal{T}(\Gamma(k),x(k)) \mid \left(\Gamma(k), x(k) \right) \in R_{\tau_j} \right] \label{eq:proof1}\\ &= \mathbb{E}\left[\mathcal{T}(\Gamma(k),x(k)) \mid \Gamma(k) \in R_{\pi_r} \right] \end{align} \end{proof} \normalsize In \eqref{eq:proof1} we assume that the dataset consists of independently drawn observations, and that the number of samples in each region of the tree is large enough to neglect the Standard Error of the sample Mean (SEM): as a consequence, the expectation can be assumed approximately equal to the sample mean. In conclusion, running the CART algorithm on our extended dataset derives transition probabilities that minimize the square of the error between the samples of the dataset $\left\{\mathcal{P}_\mathcal{T} , \mathcal{R}_\mathcal{T} \right\}$ and the corresponding conditional expectation of the predictive model of $\mathcal{T}$. To estimate $\mathbb{P}(\Gamma(k+1) \in R_{\pi_r} \mid (\Gamma(k),x(k)) \in R_{\tau_i})=p(\pi_r\mid\tau_i)$ we exploit the set of GRVs defined in Equation \eqref{eq:cp} and the partition induced by the tree $\Pi$: \begin{equation}\label{eq:tp2} p(\pi_r\mid\tau_i) = \mathbb{P} \left( G_{\tau_i}^+ \in R_{\pi_r} \right) =\int_{R_{\pi_r}} \phi_{\mathcal{N}}(\zeta;\mu_{i^+},\sigma_{i^+}^2) d\zeta \end{equation} where $\phi_{\mathcal{N}}(\zeta;\mu,\sigma^2)$ is the PDF of the GRV $\mathcal{N}(\mu,\sigma^2)$. The TPM $P = [p(j\mid i)]_{i,j}\in \mathbb{R}^{|\mathcal{T}|\times |\mathcal{T}|}$ is defined by $p(j\mid i) = \mathbb{P}(\theta(k+1)=j\mid\theta(k)=i)$ $ \forall \tau_i,\tau_j \in \mathcal{T}$, as in Equation \eqref{eq:tpm}. \subsubsection{Step 3}\label{sec:pep} The variable of interest in control applications is the PDP. The IEEE-802.15.4 provides the estimation of the PER given the SINR, see Equation \eqref{eq:Rp}. Let $\boldsymbol{\nu} = (\nu_1, \ldots, \nu_{|\mathcal{T}|})$ associate a PDP for each channel operating mode as follows: \begin{equation}\label{eq:pep} \scriptsize 1-\nu_i = E[R_p(10^{G_{\tau_i}/10}) \mid \theta(k) = i] = \int_{\mathbb{R}} R_p(10^{\zeta/10}) \phi_{\mathcal{N}}(\zeta;\mu_{c_i},\sigma_{c_i}^2)d\zeta. \end{equation} \normalsize Algorithms \ref{algPEPIde} and \ref{algPDPPre} summarize respectively the learning methodology and the procedure to estimate the packet delivery probability (PDP) over a time horizon of length $N$. \begin{algorithm}[ht!] \small \caption{Learning algorithm (off-line)} \label{algPEPIde} \begin{algorithmic}[1] \State \textsc{Input: $\left\{\Gamma(k), x(k) \right\}_{k=1}^{D}$, Output: $\mathcal{T}$, $P$, $\boldsymbol{\nu}$} \Function{Learn PEP}{} \State \textsc{Define $\mathcal{P}_\mathcal{T} = \left\{\Gamma(k), x(k) \right\}_{k=1}^{D} , \mathcal{R}_\mathcal{T}= \left\{ \Gamma(k+1) \right\}_{k=1}^{D} $;} \State \textsc{Build $\mathcal{T}$ using $\mathcal{P}_\mathcal{T} $ to predict $\mathcal{R}_\mathcal{T} $;} \State \textsc{Define $\mathcal{P}_\Pi =\left\{\Gamma(k)\right\}_{k=1}^{D} , \mathcal{R}_\Pi = \left\{\Pi(\left(\Gamma(k), x(k )\right)) \right\}_{k=1}^{D}$;} \State \textsc{Build $\Pi$ using $\mathcal{P}_\Pi $ to predict $\mathcal{R}_\Pi $;} \State \textsc{Build $P$ based on \eqref{eq:tpm};} \State \textsc{Compute $\boldsymbol{\nu}$ using \eqref{eq:pep};} \State \Return $\mathcal{T}$, $P$, $\boldsymbol{\nu}$ \EndFunction \end{algorithmic} \normalsize \end{algorithm} \begin{algorithm}[ht!] \small \caption{PDP estimation (run-time)} \label{algPDPPre} \begin{algorithmic}[1] \State \textsc{Input: $\mathcal{T}$, $P$, $\boldsymbol{\nu}$, $\left( \Gamma(k), x(k) \right)$, $N$, Output: $\boldsymbol{\hat\nu}$ } \Function{Predict PDP}{} \State Let $i \in \mathbb N$ be the integer such that $\left(\Gamma(k), x(k) \right) \in \tau_i$; \ForAll{$j=1,\ldots,N$} \State $\boldsymbol{\hat\nu}(j) = \boldsymbol{\nu}P^j(i,:)$, where $P^j(i,:)$ is the $i$-th row of $P^j$. \EndFor \State \Return $\boldsymbol{\hat\nu}$ \EndFunction \end{algorithmic} \normalsize \end{algorithm} \section{Case study} \label{sec:NR} We consider a WNCS consisting of an inverted pendulum on a cart remotely controlled over a WirelessHART link as illustrated in Sec.~\ref{sec:channel}. In the numerical simulations, we model the plant using the nonlinear discrete-time model of the inverted pendulum and we model the packet delivery process as in equation \eqref{eq:Rp} through the SINR trajectories obtained from the AR model in Sec.~\ref{sec:AR}. We compare two implementations of Stochastic Model Predictive Control (S-MPC) \cite{Bernardini2011}: one based on the physics based channel model of Sec.~\ref{ssec:FSMC} as in \cite{Lun2020}, and one based on the data-driven channel model of Sec.~\ref{sec:RTMC}. \begin{figure}[h!] \begin{center} \centerline{\includegraphics[width=0.65\columnwidth]{figures/Inverted_Pendulum.png}} \caption{Inverted pendulum on cart.} \label{figIPC} \end{center} \end{figure} \subsubsection{Plant model} We consider the following nonlinear discrete-time model $y(k+1) = y(k) + f(x(k),u_a(k))T_u$, where $y \in \mathbb{R}^4$, $u \in \mathbb{R}$ and \begin{equation} \small \begin{aligned} f_1(y,u) &= y_2 \\ f_2(y,u) &= \frac{mL^2}{D} \left[ -mg\cos(y_3)\sin(y_3)+ mLy_4^2\sin(y_3) \right] \\ &\qquad + \frac{mL^2}{D} \left(u-\delta y_2\right) \\f_3(y,u) &= y_4 \\ f_4(y,u) &= \frac{mL}{D}\left[ (m+M)g \sin(y_3)-\cos(y_3) mLy_4^2\sin(y_3)\right]\\ &\qquad - \frac{mL\cos(y_3)}{D}\left(u-\delta y_2\right) \\ D &= mL^2\left[ M+m\left( 1- \cos(y_3)^2\right) \right], \end{aligned} \end{equation} where $y_1$ is the cart position, $y_2$ is the velocity, $y_3$ is the pendulum angle, $y_4$ is the angular velocity, $m$ is the pendulum mass, $M$ is the cart mass, $L$ is the length of the pendulum arm, $g$ is the gravitational acceleration, $\delta$ is a friction damping on the cart, $u$ is a control force applied to the cart and $T_u$ is the sampling time. We analyze a case wherein the model's dynamics influence the channel behavior. More in detail, we consider that a time-varying distance between the plant and the controller, $d_0$ in equation \eqref{eq:path}, depends on the cart's position, $y_1$, i.e. $d_0(k) = y_1(k) + \bar{d}_0$, $ \bar{d}_0 \in \mathbb{R}^+$. The other channel parameters are assumed constants. As consequence, the intermittent control packets can be modeled as follows: \begin{gather} \begin{aligned} u_a(k) = \nu(k)u(k), \quad \nu(k) \sim B(1, 1-R_b(\gamma(k)) ) \end{aligned} \end{gather} where $u(k)$ is the input computed by the controller, $B(n,p)$ is the Bernoullian distribution with a time-varying parameter. Notice that $y_1(k)$ influences $\gamma(k)$ via equation $\eqref{eq:path}$. \subsubsection{Numerical results} The data-driven model derived in Sec.~\ref{ssec:BackgroundSARX} can be used to formalize the following: \begin{problem}\label{pbMPC-SA} \textit{Stochastic Model Predictive Control} \begin{equation*} \small \begin{aligned} & \underset{u_k}{\mathrm{min}} & & \mathbb{E} \left[ e_{k+N}^\top Q e_{k+N} + \sum_{j=0}^{N-1}e_{k+j+1}^\top Q e_{k+j+1} + u^\top_{k+j} R u_{k+j} \right] \\ & \mathrm{s.t.\ } & & x_{k+j+1} = A'_{i_j}x_{k} + \sum_{\alpha = 0}^{j}{B'_{i_j,\alpha}\hat{\nu}_{k+\alpha}} u_{k+\alpha} + F'_{i_j},\\ & & & u_{k+j} \in \mathcal{U}, \mathbb{E} \left[x_{k+j+1} \right] \in \mathcal{O},\\ & & &\mathbb{E} \left[x_{k+N} \right] \in \mathcal{O}_N, x_{k} = x(k), j=0\ldots,N-1 , \\ \end{aligned} \end{equation*} \end{problem} \noindent where $e_k = x_k -x_k^* $ is the difference between the current state of the plant and the target, $\hat{\nu}_{k+\alpha}$ is the PDP prediction given the current measurements, and $\mathcal{O}, \mathcal{U}$ are polyhedra that specify the variables constraints. At each time step the optimal inputs $u^*_k,\ldots,u^*_{k+N-1}$ are computed using QP, and only the first one is applied {to the system}, i.e. $u(k) = u^*_k$. The plant dynamics used in the MPC solution are identified using the methodology summarized in Sec.~\ref{ssec:BackgroundSARX} and using a dataset consisting of 100 simulations, each one with time duration of 6 seconds, of the input and output of the plant with no packet losses. At any time $k$ we can use such model and the measurement of $x_k = x(k)$ to determine the switching sequence $i_0, \ldots, i_{N-1}$ and hence the matrices $A_{i_{j-1},i_{j}}, B_{i_{j-1},i_{j}}, F_{i_{j-1},i_{j}}$. The solution of Problem \ref{pbMPC-SA} also requires knowledge of the initial state of $\hat{\nu}_k$ and of its TPM $P$ \cite{Bernardini2011}: we derive two models of $\hat{\nu}_k$, respectively using the methodology in \cite{Lun2020} and in of Sec.~\ref{sec:RTMC}: for both the data-driven and physics-based channel Markov models, we set the number of channel operating modes equal to 9. The cost function in Problem \ref{pbMPC-SA} models a tradeoff between penalizing deviations from the desired trajectory $x_k^*$ and minimizing the control effort. We define as control performance metric the cumulative cost of Problem \ref{pbMPC-SA} to compare the performance using the physics based channel model of Sec.~\ref{ssec:FSMC} as in \cite{Lun2020} and the data-driven channel model of Sec.~\ref{sec:RTMC}. \begin{table} \centering \setlength{\extrarowheight}{5pt} \begin{tabular*}{0.95\columnwidth}{lcc} \toprule \textbf{Parameter} & \textbf{Value} & \textbf{Unit} \\ \midrule Cart mass M & 0.5 & kg \\ Pendulum mass m & 0.2 & kg \\ Distance from the pivot to the mass center L & 0.3 & m \\ Friction coefficient of the cart d & 0.1 & N$\cdot$s/m \\ Sampling time $T_u$ &0.001 &s \\ \bottomrule \end{tabular*} \caption{Parameter values inverted pendulum} \captionsetup{justification=centering} \label{tab:plant} \end{table} \begin{table} \centering \setlength{\extrarowheight}{5pt} \begin{tabular*}{0.95\columnwidth}{lcc} \toprule \textbf{Parameter} & \textbf{Value} & \textbf{Unit} \\ \midrule symbol rate 1/$T_s$ & 62.5 & ksymb/s \\ channel bandwidth W & 2 & MHz \\ users speed $v_0$, $v_1$ & 5.37 & m/s \\ shadowing decay distance $dc_0$, $dc_1$ & 9 & m \\ shadowing standard dev. $\sigma_{\beta_0}$, $\sigma_{\beta_1}$ & 2 & dB \\ power control error standard dev. $\sigma_{e_0}$, $\sigma_{e_1}$ & 1.5 & dB \\ power control error decorr. time $\tau_{\xi_0}$ and $\tau_{\xi_1}$ & 1.5 & dBm \\ reference user tx. power $P_0$ & 0 & dBm \\ interferer tx. power $P_1$ & 10 & dBm \\ distance reference tx-rx pair $d_1$ & 10 & m \\ \bottomrule \end{tabular*} \caption{Parameter values channel} \captionsetup{justification=centering} \label{tab:whart} \end{table} Tables \ref{tab:plant} and \ref{tab:whart} show the plant and channel parameters, respectively. The minimum update period $T_u = 0.1 s$ of the WirelessHART standard is too slow for several control applications, and makes the wireless link uncorrelated at the packet level. Thus, in view of showing the impact of our algorithms as a methodological enabler for the development of mobile network technologies that support much higher update rates, we consider $T_u$ = 0.001s. Moreover, to emphasise the impact and improvements of stochastic vs deterministic MPC, we set $\bar{d}_0 = 14$ to consider a scenario based on significant packet loss rates, where it is evident that deterministic MPC cannot guarantee acceptable performance while stochastic MPC does. \begin{figure}[htb] \centering \subfigure{\label{figState1}\includegraphics[width=0.9\columnwidth]{figures/State1.eps}} \subfigure{\label{figState3}\includegraphics[width=0.9\columnwidth]{figures/State3.eps}} \caption{Controlled states in the closed-loop simulation.} \label{figStates} \end{figure} \begin{figure}[h!] \begin{center} \centerline{\includegraphics[width=0.9\columnwidth]{figures/CumulativeCost.eps}} \caption{Cumulative cost of the closed-loop simulation.} \label{figCumCost} \end{center} \end{figure} \begin{figure}[h!] \begin{center} \centerline{\includegraphics[width=0.9\columnwidth]{figures/PER.eps}} \caption{PER of the closed-loop simulation.} \label{figPer} \end{center} \end{figure} To statistically validate the control performance we ran Monte Carlo simulations generating 500 admissible trajectories, each with $6000$ samples (corresponding to 6s). For all the simulations, the initial state is $x(0) = \left[2, 0, \pi, 0\right]^\top$ and the target state is $x^* = \left[5, 0, \pi, 0\right]^\top$. Figures \ref{figStates}, \ref{figCumCost} and \ref{figPer} illustrate simulation results. For each figure the averaged performance is displayed in solid line, and the $95.4\%$ confidence interval is represented with a shaded area. The performance of the controllers based on physics-based and data-driven models is very close, with the advantage of the data-driven approach that no a-priori knowledge of channel model and parameters is required. \section{Conclusions} \label{sec:Con} This paper provides a novel technique to learn Markov models representing fading wireless channels. We consider a validation scenario consisting of a WNCS that exploits a WirelessHART radio link to send the optimal control inputs generated by a Stochastic MPC, and show that the control performances of our data-driven approach and of a physics-based approach based on a stationary finite-state Markov chain are extremely close: this implies that in practical applications, when assuming perfect knowledge of the channel model and parameters is not possible, the methodology presented in this paper is a valid and very effective alternative. In future work we plan to validate our techniques in an experimental setup and to consider more general communication scenarios. \bibliographystyle{IEEEtran}

5053d049101f3a72cc652a058a0527d7b4261e1e

\section{INTRODUCTION} \IEEEPARstart{W}{ith} 5G being currently under deployment, wireless communication engineers are very close to have harnessed and improved the whole chain of wireless systems. Indeed, during the past generations, modulation and coding techniques have been largely optimized to maximize the data rates, MIMO ideas have been introduced to multiply capacities, and multiple bands have been aggregated to maximize communication bandwidths\cite{Foschini1998,paulraj2003introduction,Tarokh1998,Alamouti1998}. With 5G, even more bricks are being introduced, with both millimeter wave (mmWave) frequencies offering ultrawide bandwidths and Massive MIMO antennas bringing in spatial multiplexing, to offer very high data rates to the end users\cite{Marzetta2013, Larsson2014,Lu2014,Bjornson2015,Bjornson2016}. It is to be noted that these new performances almost always come at the price of increased hardware complexities and higher energy consumption. Looking at it from this perspective, the horizon may seem somehow technologically cluttered. Yet there is a missing part of paramount importance that has been kept aside: the channel itself. This is why the idea of smart environments, that can adapt themselves for superior wireless communications in real time, has attracted much attention in the context of 6G. Being able to modify the channel at will, with energy-sober technologies, could indeed pave the way to greener and potentially limitless wireless communications. The concept of smart environments was first introduced in \cite{Subrt1,Subrt2}, where the authors proposed to optimize the usage of various access points inside a building, by controlling the electromagnetic reflectiveness of its walls, in order to distribute the signal in an optimal way in the building according to the needs. A bit later, the idea of using electronically reconfigurable surfaces for improved wireless communications was proposed in \cite{Kaina2014,Dupre2015}, where the fields on the surface were locally controlled in order to focus on a given antenna. Some follow up works were published over the years, such as \cite{X.Tan,Welkie,Hougne}, but the topic remained of rather little interest to the community, until a series of papers were published in late 2018 and 2019 \cite{Hougne,Liaskos_7,Liaskos_8,Renzo2019,Basar2019}. From there, the idea of smart environments using electronically reconfigurable surfaces, coined Reconfigurable Intelligent Surfaces (RIS) emerged as a credible major evolution for 6G and has attracted an enormous and growing interest in the wireless communications community. Associated to RIS there are numerous research topics ranging from energy efficient wireless communication\cite{8741198} to channel modeling \cite{DiRenzo2020,Garcia2020}, RIS based signal modulation and encoding \cite{Basar2020}, MIMO channel estimation and beamforming \cite{Nadeem2020,He2020,Park2020}, telecommunication performance evaluation \cite{Badiu2020}, mathematical model and optimization method for wavefront shaping with RIS \cite{Wu2020,Zappone2020,Huang2019} and even stochastic analysis approach \cite{DiRenzo2019}. Yet of the numerous works proposed, a very large part is concerned with theoretical and mathematical approaches, and very few deal with experimental demonstrations of RIS. This lack of practical demonstrations is even worse in the context of millimeter waves, where RIS first use case may well be found for instance as passive access point extenders. In this paper we propose an implementation of a low complexity RIS at mmWave and show its potential for wireless communications. To do so, we demonstrate an electronically tunable binary phase metasurface, able to control the reflections of electromagnetic waves in the 30GHz range, at a half wavelength pitch and with a $\pi$-phase shift on two polarizations independently. We first start by explaining the design of the unit cell, which is based on a patch reflector capacitively coupled to a parasitic resonator controlled by a PIN diode, as proposed in \cite{Kaina:14}. In contrast to other reconfigurable metasurfaces, which are often built from a patch with a PIN diode connected to the RF ground~\cite{5765450,7448838,7902179}, in our design we control the reflection phase by means of a capacitively-coupled to a patch parasitic resonator loaded with a PIN diode. Although this approach is more challenging at mm-Waves, it provides us more degrees of freedom for optimizing the design to significantly increase the instantaneous bandwidth and to reduce power dissipation in active elements. We study the properties and performance of the unit cell using commercially available FEM software. Then, we fabricate the metasurface unit cell, characterize its properties using standard laboratory equipment, and verify that we can obtain a $\pi$-phase shift when changing the state of the PIN diode (forward or biased polarized). We show dissipation levels of the unit cell below 2dB, for both polarizations of the unit cell. We finally provide a description of the final assembled RIS, that contains $20\times20$ unit cells on a $10$~cm~$\times$~$10$~cm surface, with a total of 800 PIN diodes controlled by an FPGA board. In the next part of the manuscript, we derive an analytical model of binary phase RIS. This model is then used to investigate two physical scenarios of a wave impinging the RIS namely (i) a spherical wave and (ii) a plane wave. These scenarios correspond to the two practical uses of the binary phase RIS in the near field (reflectarray configuration) and in the far field (access point extender). This analytical investigation underlines the peculiarities of binary phase shifting metasurfaces when dealing either with spherical or planes waves, and explain their mechanism. Finally, in the last part of the paper, we propose to experimentally explore both physical situations. In a first experiment, we use the RIS, that we have designed, in a near field configuration, corresponding to a typical reflectarray antenna. In this context, we demonstrate experimentally that the RIS can be used to beamform mmWaves very efficiently, achieving directivity close to 30dBi, and reaching angles as high as $60^\circ$ as predicted by the analytical model. The last experiment leverages the RIS in a far field configuration and provides an experimental proof of wireless transmission of mmWaves in a scenario where two horn antennas connected to a VNA don’t have any line of sight. We demonstrate how the RIS can be used to create such a line of sight, hence providing a 25dB gain over the received energy. \section{ Binary reconfigurable intelligent surface} \subsection{Unit cell design} The single unit cell (or pixel) of the designed metasurface (\ref{sec1:fig1}) is represented by a conducting square patch (made of copper) placed on a low loss grounded dielectric substrate Meteorwave 8000 ($\varepsilon=$ 3.3 @ 10~GHz, $tan(\delta)=$ 0.0016 @ 10~GHz)\cite{meteorwave}. At the resonance frequency of the patch the incident wave strongly interacts with the unit cell leading to enhanced power dissipation and significant phase advance of the reflected wave. The unit cell is designed to provide the patch resonance $f_0$ to be approximately around 27.5 GHz. \begin{figure} \centerline{\includegraphics[width=3.5in]{fig_simu_pixel.pdf}} \caption{\label{sec1:fig1} a) Unit cell of the designed reflect array. $M$ - main patch, $P_H$ and $P_V$ are horizontal and vertical parasitic resonators, respectively. When studied experimentally the pixel is placed inside the WR-34 hollow waveguide. b) Simulated reflection from the metasurface unit cell. Solid line represents OFF-state and doted line represents ON-state of the pixel. Vertical dashed lines show the operating frequency range of the pixel.} \end{figure} In order to provide the proper operation of the binary reflect array, two conditions should be fulfilled: the amplitude of the reflected wave should be maximum and the reflection phase should be able to switch in 2 states with a $\pi$-phase difference. There are many ways to change the phase of the reflected wave and most of them are based on changing the electrical length of the resonator. The simplest way to do it is to shunt the patch resonator onto the ground using an active element i.e. PIN diode. Nevertheless this leads first to a strong dissipation growth due to increased current through the PIN diode. Moreover this system becomes quite sensitive to the fabrication intolerances of the diodes. Due to this reason here we use a different method to control reflection phase. In order to provide a $\pi$-phase shift, the design of the pixel is complemented with a parasitic resonator ($P_H$ and $P_V$ on Fig.\ref{sec1:fig1}a for both corresponding horizontal and vertical polarizations of the E-vector of incident wave). The resonance frequency of the parasitic resonator lies close to the resonance of the patch. That leads to appearance of strong coupling between two resonators and corresponding anti-crossing behavior when coupled resonances repulse and the mutual resonance shifting occurs. While changing the electrical length of the parasitic resonator we change its resonance frequency and can tune the mutual coupling between resonant modes of the parasitic resonator and the patch. This in turn allows to change the resonance frequency of the patch and the phase of the reflected wave at a given frequency. This method was first suggested and verified in \cite{Kaina:14}. In order to provide control of the electrical length of the parasitic resonator, it is split into two parts connected with a PIN diode (MACOM-000907). To make the unit cell more compact the parasitic resonator is partially placed inside the preliminary carved patch resonator. The diode state is switched via applied voltage or current, controlled through the biasing network. The network is isolated from the RF-part of the unit cell through the lumped inductive elements (chokes). To increase the structure symmetry, one of the chokes of the parasitic resonator which is closest to patch is placed in the middle of the resonator and is connected to the patch. The patch is then connected with the other pole of the control network through the common patch choke, which is placed on the right side of the patch (see Fig.~\ref{sec1:fig1}.a). Electrical connection to the biasing network between top and bottom layers is provided through the vias. All the control electronics is placed on the back side of the multilayer PCB. When we apply a 5V biasing voltage, the state of the PIN diode is changed from isolating (we call it \textit{OFF-state}) to conducting (\textit{ON-state}). The electrical length of the parasitic resonator changes correspondingly providing shifting of its own resonance frequency and consequently of the resonance frequency of the main patch. Fig.~\ref{sec1:fig1}.b) shows simulated reflection coefficient from the infinite periodic structure composed of the described pixels. The simulation was performed with the time domain solver of the CST Studio Suite software. The unit-cell is simulated in the rectangular waveguide (WR-34) as illustrated by Fig.~\ref{sec1:fig1} a). A single-mode rectangular waveguide port is set two wavelength (at 28.5~GHz) away from the top of the unit cell. However, the simulation is performed in a wide frequency range and at the lowest frequency the distance to the port is approximately one wavelength. The port is de-embedded having the reference plane at the top of the unit cell in order to subtract the propagation phase. In the OFF-state the resonance of the patch (27.5~GHz, in Fig.~\ref{sec1:fig1}.b)) lies to the left of the resonance of the parasitic resonator (30~GHz at Fig.~\ref{sec1:fig1}.b)). In the ON-state the electrical length of the parasitic resonator increases and the parasitic resonance is shifted to the lower frequency (22.5~GHz). The coupled patch resonance is correspondingly pushed to the higher frequency (29.5~GHz). The operating frequency range of the unit cells is defined from the acceptable level of the reflection coefficient amplitude and difference of the reflection coefficient phases between ON- and OFF-states. To operate, a RIS is required to modify the phase of the reflected wave in comparison to the incident wave and not to dissipate all the incident power. The maximum phase difference possible is 180 degrees, but any value close to it is enough. For the designed pixel shown in Fig.\ref{sec1:fig1}.a), the operating frequency range lies in the frequency band from 27.5~GHZ to 29.5~GHz. In this range, marked with the two vertical dashed lines in Fig.\ref{sec1:fig1}.b), independently of the pixel's state, the phase difference is always larger than 160 degrees. Meanwhile, the reflection amplitude level never goes under the -5~dB level, meaning that more than 20\% of power is reflected. Although these characteristics are not perfect and can be further improved, we demonstrate in what follows that they are acceptable for the target applications. \subsection{Unit cell measurements} \begin{figure} \centerline{\includegraphics[width=3.5in]{fig_exp_pixel.pdf}} \caption{\label{sec1:fig3}a) Pixel fabricated for the waveguide characterization. The big copper area around is intended to form electrical contact with the waveguide flange. b) Magnified photo of the fabricated pixel. The vias around provide isolation of the field inside waveguide. $M$, $P_H$ and $P_V$ are main patch, horizontal and vertical parasitic resonator respectively. Dots and triangles tag respectively chokes and pin diodes. c) The setup of the experiment. d) Measured reflection from the single pixel inside the rectangular WR-34 waveguide. Solid line represents OFF-state and doted line represents ON-state of the pixel. Vertical dashed lines correspond to frequency where actual $\pi$-phase shift occurred showing the operating frequency range of the pixel.} \end{figure} To experimentally verify the properties of the designed pixel a set of pixels with slightly variable dimensions of resonators was fabricated. The WR-34 rectangular waveguide was used to characterize reflection properties of the single pixel. The pixel size was adjusted to fit the waveguide aperture (W$\times$L = 0.34$\times$0.17 inches). In order to avoid leakage of the field through the pixel substrate and provide isolation of the field inside the waveguide, a periodic via structure was formed at the edge of the pixel acting as an artificial electric wall. The bias network vias are connected with the pins on the back side of the pixel for the power supply connection. In order to place electronics on the back side of pixel there is a FR-4 substrate placed under the radiofrequency grounding layer (RF-ground). That is, single pixels consists of 2 layers of substrate separated by the ground layer. The pixel is tightly attached to the flange of the WR-34 waveguide in order to provide proper electrical connection with the waveguide. The waveguide (which is also the waveguide to coaxial adapter) is connected to a Rohde~\&~Schwarz ZVA-40 vector network analyzer . In order to provide reflection state control, the electrical pins of the diodes, placed on the back side of the pixel, were connected to the constant voltage power source. The OFF-state of the diode was provided by the 0-voltage applied to the anode of the diode. The ON-state was provided by the -5V voltage applied to the cathode of the diode. The measured reflection properties of the single pixel inside a rectangular waveguide is shown on Fig.~\ref{sec1:fig3}. We see that we can indeed control the phase difference of the reflection coefficient by changing the position of the main patch resonance coupled with the resonance of parasitic resonator. In the 2~GHz frequency range between 27.5~GHz and 29.5~GHz the phase difference between ON and OFF sates is between $144^\circ$ and $180^\circ$. The reflection amplitude in this range doesn't drop below -6~dB and average value in the frequency range bet wen 2 states is equal to -3~dB reaching -1~dB for some of the frequencies. It is worth noting that during the operation of the RIS, the average number of pixels in ON and OFF state are approximately equal. The difference between the simulated and the experimental results can be explained by the following reasons. First there is the way we simulate the properties of the lumped components (PIN diodes, chokes). The equivalent RLC-model is used to describe the diodes and chokes properties and the parameters of the model may differ from the properties of real components. Secondly, in the simulation model, we don't consider the effect of the FR4 layer placed behind the RF-ground. \subsection{Reconfigurable intelligent surface fabrication} \begin{figure} \centerline{\includegraphics[width=\columnwidth]{fig_metasurface.pdf}} \caption{\label{sec1:fig:metasurface} a) Front and b) back side of the fabricated $10~\times~10~\textrm{cm}^2$ RIS. c) RIS control board with different interfaces. } \end{figure} After the validation of the properties of 30 different pixels inside the waveguide the geometry which provides lower dissipation and better phase shift is selected. The selected pixel is used for fabricating the metasurface acting as a a reconfigurable intelligent surface (RIS). We fabricate a $10~\times~10~\textrm{cm}^2$ RIS -- approximately $10\times10 \lambda$ ($\lambda$ is a wavelengths at 30~GHz) -- made up of 400 unit cells periodically placed every $\lambda/2=0.5$~cm on a rectangular lattice of $20\times20$ pixel as shown by Fig.~\ref{sec1:fig:metasurface}.a). Each unit cells having to control independently both components of the reflected EM field, an overall of 400 diodes to control vertical polarisation and 400 diodes to control horizontal polarization are used. More technically speaking, our RIS consists in a 6 layer PCB. The first layer is made from a low loss substrate, namely the METEORWAVE 8000 from AGC, and supports the 400 unit cells (see front view in Fig~\ref{sec1:fig:metasurface}.a)). The 5 remaining layers are made from FR4 substrate. The last layer shown in Fig~\ref{sec1:fig:metasurface}.b), notably supports the electronic components that allow us to control the states of the diodes of the pixels on the first layer. This control is operated by a $10 \times 10$ matrix of shift registers, each of them managing the states of 8 PIN diodes associated to the horizontal and vertical polarization states of 4 pixels. In order to provide control of each pixel of the RIS, a FPGA control board has been designed and programmed (see Fig~\ref{sec1:fig:metasurface}.c)). This control board has 4 ports to connect 4 independent RIS. The control board is connected to the desktop through LAN or USB interface and is controlled with a Python or MATLAB code. The developed software allows to independently change the state of each diode on the RIS providing the required reflection properties across the whole RIS. PIN diodes are thus connected to the control board through the network of shift registers. We should point out that the assembly composed by the metasurface and the FPGA control board shown in Fig~\ref{sec1:fig:metasurface} is a very low power consuming RIS. Indeed, it consumes in average no more than 8 Watts, with a potential peak power of 16 Watts if all diodes are in ON-state. This power consumption can even be more reduced by tuning of the bias voltage applied to diode. In practice, in more recent version of our RIS, we have been able to reduce this power consumption by a factor of two. A last interesting propriety of our RIS is its switching state rate. While remotely controlling from a desktop, the latter lies in a kHz range. If the pattern of the PIN diode states is stored locally on the board, update rate reaches nearly 100 kHz. Current maximum speed is function of the shift register chain organization. In the following, we will know see how to use this RIS for beamforming applications. \section{Analytical model of binary phase RIS} \begin{figure}[b] \centerline{\includegraphics[width=2.5in]{schematics.pdf}} \caption{Schematics of the mutual positions of the RIS, receiving (Rx) and transmitting (Tx) horn antennas in the spherical coordinates system. \label{sec3:fig1}} \end{figure} \begin{figure*}[h] \centerline{\includegraphics[width=7in]{NF_FF_FF.pdf}} \caption{(a) Illustration of a spherical wave impinging the RIS and a plane wave being reflected. (b)--(d) Results of the analytical model for the far-field radiation pattern created by the RIS for different configurations defined by steering directions: $\theta_{Rx}=30^\circ, 45^\circ$ and $60^\circ$ are shown ($\varphi_{Rx}$ is $0^\circ$). The Tx antenna is at $D_{Tx}=145$ mm, $\theta_{Tx}=45^\circ$ and $\varphi_{Tx}=270^\circ$. (e) Illustration of a plane wave impinging the RIS and a plane wave being reflected. (f)--(h) Results of the analytical model for far-field radiation patterns created by the RIS for different configurations defined by steering directions: $\theta_{Rx}= 30^\circ, 45^\circ$ and $60^\circ$ are shown ($\varphi_{Rx}$ is $0^\circ$). \label{sec3:fig2}} \end{figure*} In this section we derive an analytical model of binary phase RIS. This model is then used to investigate different physical scenarios of a wave impinging the RIS, underlining the peculiarities of binary phase shifting metasurfaces and explaining their mechanism. The reflection of the electromagnetic wave impinging the RIS can be controlled at will by judiciously switching PIN diodes. As it was shown in the previous section, states of PIN diodes determine the phase of the local reflection coefficient along the RIS. The distribution of the phase $\phi_{nm}$, where $n$ and $m$ number pixels in the array, is established according to the following equation~\cite{Hum2014_review}: \begin{equation} \label{eq:phase_grad} \phi_{nm} = \phi_r(x_n,y_m) - \phi_i(x_n,y_m), \end{equation} where $\phi_i(x_n,y_m)$ and $\phi_r(x_n,y_m)$ are the phase distributions created by the incident and (desired) reflected waves at the position $x_n$, $y_m$ of the $nm$th pixel. The incident wave is created by the transmitting (Tx) antenna illuminating the RIS. Since PIN diodes allow one to switch a pixel between two phase-states ($0$ and $\pi$) for each polarization, the phase $\phi_{nm}$ is set to $0$ when $-\pi/2\leq \phi_r(x_n,y_m) - \phi_i(x_n,y_m) \leq \pi/2$ and to $\pi$ otherwise. The radiation pattern $E(\theta,\varphi)$ created by the RIS set to a given configuration can be approximated as the radiation pattern created by the array of patch-like antennas and reads ~\cite{yang2016programmable} \begin{eqnarray} \label{eq:rad_pattern} &&E_r(\theta,\varphi) = \cos(\theta) \sum_{n,m = 1}^N \Gamma_{nm} E_i(x_n,y_m) \cos(\theta_{nm})\nonumber\\ &&\times\exp(-jk\sin(\theta)[x_n\cos(\varphi)+y_m\sin(\varphi)]), \end{eqnarray} where $\cos(\theta)$ represents the radiation pattern of each individual antenna in the array, $\Gamma_{nm}$ is the local reflection coefficient from the $nm$th pixel, $E_i(x_n,y_m)$ is the electric field of the incident wave at the $nm$th pixel, $\theta_{nm}$ is the $\theta$-angle at which $nm$th pixel sees the Tx antenna, see Fig.~\ref{sec3:fig1}. In what follows we focus on two practically important examples when: (i) Near-field configuration, \textit{i.e} the Tx antenna is placed at a distance to the RIS comparable to its size $D_m$ and (ii) Far-field configuration, \textit{i.e} the Tx antenna is far away from the RIS such that the incident wave can be considered as a plane wave. The receiving (Rx) horn antenna is always assumed to be far away from the RIS, i.e. $D_{Rx}\gg D_m$. \subsection{Tx antenna is close to RIS: Near-field configuration} The first scenario resembles a classical reflectarray antenna configuration. The incident wave radiated by the Tx (horn) antenna is modeled as a spherical wave $\exp(-j k \sqrt{x^2+y^2+z^2})/\sqrt{x^2+y^2+z^2}$, although more elaborated models can be used to take into account such parameters of the antenna as its directivity. The reflected wave is set to be a plane wave $\exp(-jk\textbf{r}\textbf{n})$, which propagates towards the Rx antenna in the direction $\theta_{Rx}$, $\varphi_{Rx}$ defined by the vector $\textbf{n}$. Under this configuration Eq.~\eqref{eq:phase_grad} takes the following form: \begin{eqnarray} \label{eq:phase_NF_FF} &&\phi_{nm} = - k\sin(\theta_{Rx})[x_n\cos(\varphi_{Rx}) + y_m\sin(\varphi_{Rx})] \nonumber\\ &&+ k \sqrt{(x_n-x_{Tx})^2+(y_n-y_{Tx})^2+z_{Tx}^2}, \end{eqnarray} where $z_{Tx}=D_{Tx}\cos(\theta_{Tx})$, $x_{Tx}=D_{Tx}\sin(\theta_{Tx})\cos(\varphi_{Tx})$, $y_{Tx}=D_{Tx}\sin(\theta_{Tx})\sin(\varphi_{Tx})$. Fig~\ref{sec3:fig2} (a)--(d) demonstrates different numerical examples of beam-forming with $10\lambda\times 10\lambda$ large RIS. The configuration of the RIS is set according to Eq.~\eqref{eq:phase_NF_FF} and the resulting radiation pattern is calculated with Eq.~\eqref{eq:rad_pattern}. The colorbars in panels of Fig~\ref{sec3:fig2}.b)--d) indicate the directivity in each particular example and help to understand how well the incident wave is steered by RIS in the direction of a receiving antenna. The directivity is maximum in Fig.~\ref{sec3:fig2}.b) corresponding to $\theta_{Rx}=30^\circ$ steering angle and reaching $27.9$ dBi value. When increasing the steering angle to $\theta_{Rx}=60^\circ$, the directivity drops to $25.2$ dBi. These values can be compared to the maximum possible directivity of approximately $30.5$ dBi, which is created by the uniform aperture of the same area as the RIS ($A=100\lambda^2$) at $\theta_{Rx}=0^\circ$ and calculated as $4\pi A/\lambda^2$ \cite{balanis2016antenna}. The directivity of the RIS decreases from the maximum value due to the following factors: (i) the steering angle decreases the physical aperture of RIS, (ii) only two phase-states are available for a binary RIS, which increases the level of undesired secondary lobes, (iii) the incident spherical wave does not illuminate the RIS uniformly. Out of these three factors, only the last one is responsible for the reduced directivity in the near-field configuration. Indeed, the aperture size sets a fundamental limit on the directivity and only by increasing the RIS area this limit can be increased. Even though the considered RIS is binary, the level of secondary lobes, in the near-field configuration, remains low even for such large steering angles as $60^\circ$. However, the near field configuration also implies that a RIS cannot be effectively illuminated and this decreases the aperture efficiency and thus the effective size of the physical aperture. \subsection{Tx antenna is far from RIS: Far-field configuration} In the second case, the Tx antenna is assumed to be further away from the RIS such that $D_{Tx}\gg D_m$, when the impinging wave can be well approximated as a plane wave. The phase profile along the RIS is thus calculated by means of the following formula \begin{eqnarray} \label{eq:phase_FF_FF} &&\phi_{nm} = - k\sin(\theta_{Rx})[x_n\cos(\varphi_{Rx}) + y_m\sin(\varphi_{Rx})] \nonumber\\ &&+ k\sin(\theta_{Tx})[x_n\cos(\varphi_{Tx}) + y_m\sin(\varphi_{Tx})]. \end{eqnarray} It leads to a periodic phase profile which cannot be well-approximated by only two phase states. The binary RIS put in a periodic configuration does not allow one to control the far-field radiation pattern completely and a very strong undesired lobe appears in the addition to the main lobe. This effect is well demonstrated by Figs.~\ref{sec3:fig2}.e)--h) where a normally incident plane wave is split in two symmetrical beams according to periods of set configurations. It is curious to note that although the strong secondary lobe reduces the directivity, the values calculated in the three examples are very close to those in the corresponding near-field configurations (see Figs.~\ref{sec3:fig2}.b)--d)). It can be explained by the fact that the plane-wave illumination creates the maximum aperture efficiency of $1$, that cannot be achieved with a Tx antenna close to a RIS. To conclude this section, the analytical model allows one to find out in advance a configuration of RIS for a given functionality and estimate its efficiency in terms of the directivity. More elaborate algorithm can be implemented on the basis of Eq.~\eqref{eq:rad_pattern} to control with RIS multiple or shaped beams~\cite{yang2016programmable}. An analytical configuration can also be a starting point for an experimental optimization procedure with an implemented feedback mechanism as the one proposed in the following experimental section. \section{Experimental results} \begin{figure}[t] \centerline{\includegraphics[width=\columnwidth]{fig_reflect4.pdf}} \caption{\label{sec4:figreflect_setup} a) Photograph of the experimental reflectarray setup with our $10$~cm by $10$~cm RIS. b) Experimental setup for measuring the far field radiation pattern of a).} \end{figure} \begin{figure}[t] \centerline{\includegraphics[width=3.5in]{fig_reflect_spectrum.pdf}} \caption{ a) Experimentally measured frequency response of the reflectarray setup of Fig~\ref{sec4:figreflect_setup} in the steering directions $\theta_{Rx}=0^\circ$ (continuous curves), $\theta_{Rx}=30^\circ$ (dashed curves), $\theta_{Rx}=45^\circ$ (dashed-dotted curves) and $theta_{Rx}=60^\circ$ (dotted curve) ($\varphi_{Rx}$ is $0^\circ$) for the initial RIS configuration (top figure) and for the optimized configurations (bottom figure) . b) Experimentally measured normalized far-field radiation patterns for the same steering angles and for the optimized configurations . \label{sec4:fig2}} \end{figure} In this section we investigate experimentally the two possible physical situations mentioned in the above section and corresponding to RIS placed respectively in the near field (Fig~\ref{sec3:fig2}.a)) or in the far field (Fig~\ref{sec3:fig2}.e)) of the Tx antenna. \subsection{Reflectarray setup: RIS in Near-field configuration} To study the first situation, a horn antenna is placed $145$ mm away from our mmWave RIS and tilted $45^\textrm{\degree}$ in odrer to reduce shading by the feed horn. The overall set-up is mounted on a mmWave transparent plastic holder as shown in Fig~\ref{sec4:figreflect_setup}.a). In order to measure the radiation pattern of the thus-obtained reflectarray set-up, the latter is fixed on an Elevation-Azimuth rotation stage and installed in an anechoic chamber in front of a fixed horn Tx antenna 1 as depicted in Fig~\ref{sec4:figreflect_setup}.b). Then, for different relative angular position ($\varphi_0,\theta_0$) between the reflectarray setup and the fixed RX horn antenna 1, starting from an initial configuration of the RIS corresponding with all pixel in on state, we optimize the configuration of the latter in order to increase the amplitude of the scattering parameter measured by the VNA at $27.5 $~GHz between the Rx horn antenna 2 on the reflectarray setup and the Tx antenna. The optimization scheme we use is a straightforward iterative one comprising measuring the transmission amplitude $\abs{S_{12}}$ for each pixel for its four possible combinations of PIN diodes states and and selecting the state of the pixel where $\abs{S_{12}}$ is maximized. This operation is repeated for all pixels. This optimization scheme is inspired from one commonly used in adaptive optics \cite{Bridges74} and has been successfully implemented during the last decade in numerous wavefront shaping applications in complex media dealing equally with optical \cite{Vellekoop2007,Mosk2012} , electromagnetic waves \cite{Kaina2014,Hougne} and even acoustic waves \cite{Ma2018}. Then, after optimization, we measure the far field radiation pattern of the optimized configuration. In Fig~\ref{sec4:fig2}.a), the top and bottom figure respectively correspond to the frequency responses for the initial configuration and the optimized one for four different azimuth angles $\varphi_0=0^\textrm{\degree},30^\textrm{\degree},45^\textrm{\degree}$ and $60^\textrm{\degree}$ and a constant elevation angle $\theta_0=0^\textrm{\degree}$. For each angles, the optimized transmissions are at the optimized frequency about $30$~dB above the initial ones. Fig~\ref{sec4:fig2}.~b) shows four radiation patterns associated with optimized configuration for elevation angle $\theta_0=0^\textrm{\degree}$ and azimuth angles $\varphi_0=0^\textrm{\degree}$ (continuous line) , $\varphi_0=30^\textrm{\degree}$ (dashed line), $\varphi_0=45^\textrm{\degree}$ (dashed-dotted line) and $\varphi_0=60^\textrm{\degree}$ (dotted line). This figure is in good agreement with the numerical model shown above and demonstrates experimentally the ability of our millimeter-wave RIS to steer a beam even at $60^\textrm{\degree}$. At that point, we would like to draw the reader's attention to the fact that, regardless of the steering angle, the RIS configuration obtained through the optimization process is very similar to the one we would predict with the numerical model describes in the previous section. However the optimized configuration generally leads to a field which is about 2.5dB higher in absolute value in the direction of the main beam. This is due to the fact that optimization process is able to deal with imperfections in the manufacturing of the RIS (not all pixels are perfectly equivalent), and probably also with a slight misalignment in the experimental setup. We can also note that the first sidelobe levels are about $-13$~dB as expected for rectangular aperture. And finally, due to the decrease in the effective size of the geometric aperture as the optimization angles increase, the beamwidth and gain naturally increases and decreases respectively for higher optimization angles. Furthermore, with the setup of Fig~\ref{sec4:figreflect_setup}, we have checked that the beam can been steered in a range of $120^\textrm{\degree}$ with a precision of $0.2^\textrm{\degree}$. \subsection{ RIS in access-point extender configuration } \begin{figure}[h] \centerline{\includegraphics[width=3.5in]{fig_far_2_far_all.pdf}} \caption{ (a) Experimental setup emulating a wireless transmission of mmWaves between two horn antennas connected to a VNA. Line of sight between the receiver and the transmitter is suppress by a right angle barrier. A tilted reflector face the corner of the barrier. In scenario I, the reflector is a metallic plate which is replace by a RIS in scenario II. b) Comparison of transmission parameter measured in scenario I (light green dotted curve) and after optimization of the RIS configuration in scenario II (dark green continuous curve ) \label{sec4:fig3}} \end{figure} The second physical situation we are experimentally addressing is the RIS placed in the far-field of both Rx and Tx antenna as illustrated in Fig~\ref{sec3:fig2}.e). In order to play with the RIS in a far field configuration, we propose the wireless communication scenario at mmWave depicted in Fig;~\ref{sec4:fig3}.a). The latter consists in a wireless transmission of mmWaves between two horn antennas connected to a VNA while we avoid any line of sight between the receiver and the transmitter by placing each antenna at both extremity of a right angle barrier (emulating for example the intersection between two corridors). Each arm of the barrier is $1$~m long. The overall setup is installed outside the anechoic chamber to be closer to an actual wireless communication scenario. Strictly speaking, we are not in the far-field of the RIS in antenna terms (by the factor of 2), but the Tx and Rx antenna do not have a line of sight and are one order of magnitude further away from the RIS in comparison to the previous experiment. However, we can still consider the RIS to be in the far-field zone of a Rx or a Tx antenna if the field radiated by either of these antennas can be well-approximated by a plane wave at the position of the RIS. As we use in our experiments simple horn antennas (having around 17 dBi gain), the far-field zone for these antennas is around 15 cm. To demonstrate the advantage of using a RIS in this kind of problem, we measure the transmission between the antennas in two different situations. In the situation I, a tilted metallic plate is facing the corner of the barrier. The light green dotted curve in Fig.~\ref{sec4:fig3}.b) corresponds to the obtained transmission coefficient in a frequency range of $2$~GHz around $27.5$~GHz. In this situation the transmission between the antennas is very low. In the situation II, the metallic plate is replaced by a RIS which is also tilted. The relative position between the horn antennas 1 and 2 and the RIS not being evaluated, we use the same iterative optimization scheme describe above to find an optimized configuration that increase the mean value of the transmission coefficient $\langle \abs{S_{12}(f) }\rangle_f`$ in a frequency window of $250$~MHz around $27.5$~GHz. The frequency response associated to the optimized configuration is plotted in Fig.~\ref{sec4:fig3}.b (dark green continuous curve). Around the optimized frequency windows, we measure a $25$~dB gain over the received energy. We show also that the energy gain is valid well beyond the $250$~MHz frequency window where we did the optimization. This is consistent with the bandwidth of the metasurface. Thus, we have proposed an experimental proof of wireless transmission of mmWaves where a normally absent line of sight is restored thanks to our RIS acting as a highly efficient and and low-power consumption access point extender. It is worth noting that here a $25$~dB gain is obtained merely on cost of the power dissipated by the PIN diodes, namely 8 Watts in average for the whole metasurface (reduced to 4 Watts in our last version of metasurface). As a remark, in this experiment we used a metal plate as a reference, but RIS in a random configuration can also be used as a benchmark. However, the result will depend on the particular random configuration and an averaging over a large ensemble of configurations has to be performed. \begin{table}[tb] \centering \caption{RIS specifications table.} \label{tab1} \begin{tabular}{|c|c|} \hline Aperture size & $10$ cm $\times$ $10$ cm\\ \hline Number of elements & 400 \\ \hline Polarization & dual linear \\ \hline Element's spacing & $\lambda/2$ \\ \hline Operating frequency range & $[27.5$ GHz, $29.5$GHz$]$\\ \hline Instantaneous bandwidth & 500 MHz\\ \hline Directivity & $22.5$ dBi\\ \hline Scan range (El., Az.) & $\pm 60^\circ$\\ \hline Switching rate & 100 kHz\\ \hline \end{tabular} \end{table} \section{Conclusion} In this paper, we introduced a design of the binary RIS operating in a transmitting spectrum of the Ka-band, key performance parameters can be found in Tab.~\ref{tab1}. The design of the single pixel of the RIS provides low level of reflection wave dissipation, near-$\pi$ value of the phase difference and wide instantaneous bandwidth. We showed that the RIS can be used in two different configurations: as a general reflection array illuminated by a closely positioned horn antenna and as highly efficient and low power consumption access point extender. In the last case, the RIS provides an amplification of the signal between transmitter and receiver in a scenario where originally there is no line of sight between them. In the future work we plan to demonstrate the functionality of the developed RIS as a part of communication system using Software-Defined-Radio, especially regarding robustness of the link with respect to the power noise potentially introduced by the RIS. \bibliographystyle{ieeetr}

77b9ee12b9758c8f5159cfdfc54d20ec5235f448

\section{Introduction} Machine Learning (ML) plays a crucial part in a wide variety of applications. These applications are usually data-intensive and use ML algorithms for building a mathematical model from the training data for decision making. The training data used for building the model could come from controlled and uncontrolled data sources. Any imperfection, intentional or unintentional, in the data could lead to model corruption, and thus producing unexpected results post deployment. The introduction of intentional imperfection in the training data is commonly known as \emph{data poisoning}, which is a type of adversarial attack \cite{xiao2015feature} that degrades the performance of the trained classifier. As the adversary could manipulate the training data to subvert the learning process of the model or manipulate the test data to evade the model prediction, therefore it becomes imperative to secure the ML model against such attacks. Two of the major attacking strategies are \emph{evasion} and \emph{poisoning} attacks. In an evasion attack, the test data is manipulated to evade the classifier's decision boundary. In a recent work, Labaca et al \cite{labaca2019poster} demonstrated the evasion attack on malware classifiers, wherein they generated valid adversarial malware samples against convolutional neural networks using gradient descent technique. In contrast to this, in a data poisoning attack, the training data is manipulated to achieve the attacker's objective. For example, Nelson et al \cite{nelson2008exploiting} have shown that even a 1\% perturbation in training data makes the spam-detection system useless. Biggio et al \cite{biggio2012poisoning} described data poisoning attack as a causative attack in which training data is injected with especially crafted malicious data-points. An attacker can mount this attack either by directly modifying the training dataset (insider threat) \cite{homoliak2019insight} or by submitting malicious samples to the ML model \cite{biggio2012poisoning}. In past, anti-virus vendors have been blamed for injecting poisoned samples into VirusTotal\footnote{ https://www.virustotal.com/} for degrading the performance of the competing products \cite{biggio2018wild}. The machine learning models which use training data from unconstrained and unmonitored source have high susceptibility to data poisoning attacks. The twitter bot, \emph{Microsoft Tay}, is one such illustrious example, where a crowd, aged between 18 to 24, corrupted it to post inflammatory and offensive tweets within 16 hours after its deployment \cite{price2016microsoft}. Similarly, Lam et al \cite{lam2004shilling} studied the shilling attack in recommender systems wherein the untrustworthy \emph{crowd} could influence the system in recommending low-quality items to an unsuspecting user. Data poisoning attack is a serious threat to a ML model as it could result in security and privacy issues, monetary losses and other serious implications \cite{biggio2012poisoning,mei2015using,zhao2017efficient,koh2018stronger,zhang2019online}. The impact of data poisoning attack could be seen on a variety of applications, such as spam filtering \cite{nelson2008exploiting}, malware detection \cite{labaca2019poster}, recommender system \cite{lam2004shilling} and sentiment analysis \cite{newell2014practicality}. For example, Lowd and Meek \cite{lowd2005good} have shown that it is possible to fool a spam classifier by carefully composing the message body of the email. Similarly, it has been shown that a \emph{malicious producer} could change the outcome of a recommender system by inserting fake product reviews, such that their product is often recommended to customers~\cite{chen2019data}. The existing literature on data poisoning primarily focuses on offline setting, wherein the entire data is available for training of the classifier and for mounting data poisoning attack \cite{biggio2012poisoning,mei2015using,zhao2017efficient}. However, these studies do not address the adversarial threat against an online learner. For example, consider an e-commerce application where the user generated data arrives sequentially. As the user is in control of data generation, it is relatively easy for a malicious user to manipulate the data sample before it reaches to the learner. Thus, making the online learner an easy target for data poisoning attack. Moreover, the adversary has the advantage of observing the impact of previous data-item she poisoned and accordingly adjusting her attack strategy~\cite{zhang2019online}. One of the well-known defense strategies against data poisoning attack is \emph{data sanitization} \cite{cretu2008casting}, where the suspicious data is filtered out before it reaches to the training process \cite{paudice2018label,chan2018data}. Steinhardt et al \cite{steinhardt2017certified} described an approximate upper bound on the maximum test error a defender (who sanitizes the data before training) can suffer under attack. In their work, Koh et al \cite{koh2018stronger} explored the data poisoning attacks against the common data sanitization defenses. They concluded that data sanitization defense tends to fail easily and discussed different strategies for stronger defenses. The other defense strategies against data poisoning prescribe training of model from reliable dataset \cite{nelson2008exploiting,paudice2018detection}, robust learning of the models in presence of corrupted data \cite{diakonikolas2018sever,zhu2019generalized}. However, defense against the data poisoning attack in online learning is not well-studied in the existing literature \cite{steinhardt2017certified,collingedefending,wang2019investigation}. \emph{Slab} defense \cite{steinhardt2017certified} is an effective data sanitization defense against many attacks \cite{wang2019investigation}. In this work, we propose an influence-based defense strategy in addition to the slab defense \cite{steinhardt2017certified} to minimize the impact of the poisoned data points on the learner's model. Influence function \cite{cook1982residuals} is a well-studied technique in robust statistics. In the recent past, machine learning researchers have used it to improve the model reliability \cite{schulam2019can}, to improve the model fairness \cite{wang2019repairing}, and to measure the group effects in the model prediction \cite{koh2019accuracy}. Also, the effect of a training point on a model prediction is efficiently estimated using the influence function without repeating the training process \cite{koh2017understanding}. In this paper, we propose a method for defending against data poisoning attacks on online learning. Our contributions are as follows: \begin{itemize} \item We propose an influence-based defense against data poisoning attacks on online learning. We formulate the defense algorithm in two steps. First, we use the slab defense \cite{steinhardt2017certified} for initial data sanitization. Then, we apply our influence-based approach for minimizing the degradation caused by the poisoned data. \item We empirically investigate the performance of the proposed defense under multiple attacks and across multiple datasets. \end{itemize} The rest of the paper is organized as follows. Section \ref{sec:relatedworks} presents a brief survey of the related work. The threat model and system assumptions are discussed in Section \ref{sec:threat_model}. The necessary prerequisites are discussed in Section \ref{sec:prelim}, while Section \ref{sec:Influence_Defense} discusses our proposed Influence based defense for online learning. In Section \ref{sec:experiments}, we present the experiment results and observations. Finally, we conclude the paper in Section \ref{sec:conclusion}. \section{Related Work} \label{sec:relatedworks} Data poisoning attacks are a serious threat to machine learning applications. There are extensive studies that have proposed attacks and defenses in the past. Biggio et al. \cite{biggio2012poisoning} have discussed data poisoning attack on Support Vector Machines (SVM). Mei et al. \cite{mei2015using} presented data input poisoning attack as a bilevel optimization problem and showed its effectiveness on SVM, logistic regression and linear regression. Koh et al. \cite{koh2017understanding} studied the use of influence functions for tracing a model's prediction through the learning algorithm and back to its training data. They further used this method for mounting a data input poisoning attack. Similarly, defending a machine learning model against such attacks is also an active and extensively studied area. Paudice et al. \cite{paudice2018detection} presented a defense mechanism to mitigate the effect of the poisoning attacks based on outlier detection. In a different work, Paudice et al. \cite{paudice2018label} have described a label flipping attack, a special case of data poisoning, and methods to mitigate such attacks. Most of the research in this area considers offline learning, and less work has been reported on the attacks and defenses for the online setting. Wang and Chaudhuri \cite{wang2018data} presented an attack strategy, formulated as an optimization problem, for semi-online and fully-online settings. For their proposed attack, they considered a white-box adversary, where the attacker knows the entire data stream, the model class, the training algorithm and the deployed defenses. More recently, Wang et al. \cite{wang2019investigation} have studied the various defenses that can help mitigating the effect of data input poisoning on online learning. Their study showed that the \emph{Slab} defense is an effective method, and weaker threat models can result in fairly powerful attacks. Zhang et al. \cite{zhang2019online} formulated an online data poisoning attack as a stochastic optimal control problem where the attacker has no knowledge about the future training data and the distribution of data. Collinge et al. \cite{collingedefending} discussed an attack strategy against online learning classifiers in their work. They also discussed a defense strategy that analyzes the impact of a data point on the learning process and rejects suspicious inputs if their impact on the model is more than a threshold. An extra level of defense is proposed by training machine learning classifiers with different learning rates. Outlier detection is a common defense strategy, wherein the objective is to remove the training points that substantially deviate from the normal \cite{hodge2004survey}. Such approaches are well suited for removing typical noises in the data, but not for removing certain well-crafted adversarial noises. Also, these approaches cannot remove all the poisoned points from the data stream. Steinhardt et al \cite{steinhardt2017certified} in their work presented certified data sanitization defenses, where they considered two classes of defense: fixed defense and data-dependent defense with distinct outlier detection rules. Paudice et al \cite{paudice2018detection} introduced a defense strategy based on K-Nearest Neighbor (K-NN) that requires a trusted dataset. The requirement of a trusted dataset makes their approach unrealistic and impractical. Moreover, approaches that rely on a trusted dataset are vulnerable to the adversaries who can tamper it. Recent works have proposed attacks and defenses using techniques from robust statistics \cite{koh2017understanding}. Kang et al \cite{kang2019testing} studied methods for evaluating the defense strategies against unforeseen attacks. They have empirically shown that the robustness of a model based on the evaluation against single type of attack is not enough to provide information about the model robustness. Koh et al \cite{koh2018stronger} studied the effectiveness of data-dependent defenses ($L2$ and $Slab$) against certain attacks, for example, influence, min-max and KKT. They observed that the data-dependent defenses ($L2$ and $Slab$) perform better against min-max attack with MNIST and Dogfish datasets, but not with Enron and IMDB datasets. Zhu et al. \cite{zhu2019generalized} proposed a method that uses a generalization of minimum distance functionals, which projects the poisoned data distribution onto a given set of well-behaved distributions. In this work, we combine the $Slab$ defense with the influence function \cite{koh2017understanding} for an efficient defense strategy. In rest of the paper we will refer our proposed approach as influence based approach. \section{Threat Model and Assumptions}\label{sec:threat_model} Attackers can be broadly classified into three types based on their capability and knowledge - offline, semi-online and fully-online \cite{wang2018data,wang2019investigation}. The objective of all three attacker models is to degrade the performance of the learner's model. \textbf{\emph{Offline Attacker.}} An offline attacker has access to the training data and can add poisoning data instances to the clean training data and she is oblivious to the streaming nature of the input. The attacker uses a method similar to \cite{biggio2012poisoning} for generating the poisoned data points. From an attacker's perspective, both the attack and the learning process happen in an offline fashion. \textbf{\emph{Semi-online Attacker.}} A semi-online attacker can add or update the data instances at any position in the clean data stream. The resultant poisoned data is then used by the victim for training their ML model. Goal of the semi-online attacker is same as that of the offline attacker, that is, to degrade the performance of the final model. The attacker knows the entire clean data stream like an offline attacker, but the model is updated in an online manner. \textbf{\emph{Fully-online Attacker.}} A fully-online attacker knows the data stream up to the current time step and does not know about the future data instances. Hence, the fully-online attacker differs from the semi-online attacker as the latter has access to the entire data-stream. Moreover, the attacker can add or update the data points at pre-specified position only. The poisoned data resulted at specific time steps are then used by the learner for training. The attacker's objective is similar to an offline or semi-online attacker, that is, to degrade the victim model's performance, but over the entire time horizon. \textbf{\emph{Assumptions.}} For fair comparison with the baseline, we consider semi-online and fully-online attackers with attacker's capability similar to that of Wang et al. \cite{wang2019investigation}. We assume that the attacker's objective is to reduce the accuracy of the victim's model. In a semi-online attack, the entire clean data stream is known to the attacker and she can add a predetermined number of poisoned data points at any position in the stream. On the contrary, the capacity of the online attacker at a time step is limited to the data points received up to that time step. We assume a white-box setting, wherein the attacker know about the learner's model, the training algorithm, hyper-parameters and any defense deployed. \section{Preliminaries}\label{sec:prelim} We develop a defense mechanism against data poisoning targeted at online binary classification task. In this section, we introduce data poisoning attacks against binary classifiers for online learning and then define the data poisoning attacker. Further, we define the data poisoning defender and discuss common defense strategies. \subsection{Data Poisoning Attacks}\label{sec:data_poisoning} Let us consider $\mathcal{D}$ = $\{(\bm{x}_{i},y_{i})\}_{i=1}^{n}$, where $\bm{x}_{i} \in \mathbb{R}^{d}$ is the $i^{th}$ input instance with $d$ dimensions and $y_{i} \in \{-1, +1\}$ is the corresponding binary classification label. A linear classifier can be learned using the objective function, \begin{equation} \begin{aligned} {\underset{f \in \mathcal{H}}{min}} \; \sum_{i=1}^{n} L(y_{i},f(\bm{x}_{i})) + \Omega(\theta) \label{eq:ObjFun} \end{aligned} \end{equation} where $f(\bm{x}_{i}) = \bm{\theta}^{T}\bm{x}_{i} + b$ is the decision function learned by minimizing the objective function (Equation \ref{eq:ObjFun}) on the data $\mathcal{D}$, $\Omega(\theta) = \frac{c}{2}||\theta||^{2}$ is a regularization parameter and $c$ is a constant, $\mathcal{H}$ is the hypothesis space, and $L$ is the loss function. For a given test instance $\bm{x}_{i}$, $\hat{y}_{i}= sgn(f(\bm{x}_{i}))$ is the predicted label. Additionally, we formalize the data poisoning attack as follows: \begin{definition Data poisoning attacker takes as input a feasible data set $\mathcal{D}$, an initial parameter of the model $\theta_{0}$, the budget $K$, and outputs a poisoned version of the data samples $\mathcal{D}_{p}$ with at most $K$ samples are inserted or updated. The model parameter $\theta^{*}$ learned thereafter results in maximum error rate on test data $\mathcal{D}_{test}$. \end{definition} \subsection{Online Learning} In an online setting, arrival order of the data instances matter. Let the sequentially arrived data instances be $\{(x_{0},y_{0}), \dots, (x_{t},y_{t})\}$ and the model parameter $\theta_{t}$ at time $t$ is updated iteratively to $\theta_{t+1}$ based on the instance $(x_{t},y_{t})$. Online gradient descent (OGD) is used for updating $\theta_{t}$ at time $t$ with $\eta_{t}$ as the learning rate, $\theta_{0}$ as the initial model parameter \cite{wang2018data}. Therefore, the update at time $t$ is computed as: \begin{equation}\label{eq:OGD} \theta_{t+1} = \theta_{t} - \eta_{t} (\nabla L (\theta_{t},(x_{t},y_{t})) + \nabla \Omega(\theta_{t})) \end{equation} where $L$ is a convex loss function. Also, regularization parameter $\Omega(\theta) = \frac{c}{2}||\theta||^{2}$, where $c$ is a constant. \subsection{Data Poisoning Attack on Online Learning} An optimization problem from an attacker's perspective for a data stream is formulated in \cite{wang2018data}. Let $\mathcal{D}$ be a data stream and ($x_t, y_t$) be the data instance at time $t$. As the only difference between fully-online and semi-online is when attacker's objective is evaluated, therefore we discuss only the semi-online for brevity. Also, based on the evaluation (fully or semi), poisoned data is added either at the end or at any location in the stream. For a prediction task for an input $x \in \mathcal{X}$ to an output $y \in \mathcal{Y}$, let $\mathcal{F}$ be a feasible set, such that $\mathcal{F} \subseteq \mathcal{X} \times \mathcal{Y}$. The feasible set $\mathcal{F}$ is of bounded diameter on which each example ($x_t, y_t$) is projected on arrival. This is an assumed defense used by the learner for excluding trivial attacks that tries to modify the learning process by adding outlier examples with very high norm. The learning process only considers data points for training that are in $\mathcal{F}$. As we are considering binary classification, therefore $\mathcal{Y} = \{-1, +1\}$ \cite{steinhardt2017certified}, and the feasible set is $\mathcal{F} = [-1,1]^d \times \{-1, 1\}$ where $d$ is the overall number of features \cite{biggio2012poisoning}. According to \cite{wang2018data}, under this defense an attacker would poison data points that are in the feasible set. Let us consider a semi-online setting where $\mathcal{D}_{train}$ be the input data stream and $T$ denotes the length of the data stream, then the attacker's optimal strategy is the solution to the following optimization problem: \begin{equation} \begin{aligned} & \underset{\mathcal{D} \in \mathcal{F}^{T}}{max} \; g(\theta_{T}) \\ & \text{subject to:} \; | \mathcal{D} - \mathcal{D}_{train} | \le K, \\ & \theta_{t} = \theta_{t-1} - \sum_{\tau=0}^{t-1} \eta_{\tau}\;(\nabla L (\theta_{\tau}, \mathcal{D}_{\tau}) + \nabla \Omega(\theta_{\tau})), \; 1\le t \le T \end{aligned} \end{equation} where, $\mathcal{D} - \mathcal{D}_{train}$ is the difference between the two data streams, $g$ is the attacker's objective function. The attacker can poison at most $K$ instances in the input data stream and $\theta_{t}$ updated at time $t$ using the Online Gradient Descent (OGD) method. The objective function $g(\theta_t)$ depends on a specific weight vector $w$, that is the learned classifier and, the attack setting, for example, in fully-online setting, the objective function $g(\theta_{T})$ becomes $\sum_{t=0}^{T} g(\theta_{t})$. The attack can be done in two ways: (i) the attacker either manipulates the data in the input training data, or (ii) crafts the poisoned data and inserts it into the clean training data. \subsubsection{Targeted Attack Strategies} \label{sec:target} In our work, we evaluate the proposed defense strategy against the following data poisoning attacks: \textbf{\emph{Simplistic Attack} \cite{wang2018data,wang2019investigation}}. Let $\theta_{0}$ be the initial model and $\theta^{*}$ be the attacker's target. In this attack, some data points are appended to the clean data stream $\mathcal{D}_{c}$ which results in $\mathcal{D}_{p}$, further $\mathcal{D}_{p}$ is used for training the model $\theta^{p}$. The attack is considered to be successful when $\mathcal{D}_{p}$ satisfies the following three properties: \begin{enumerate} \item $|| \theta^{*} - \theta^{p}|| \le \epsilon$ where, $\epsilon$ is the tolerance parameter, \item at most $K$ data points are inserted into $\mathcal{D}_{c}$, \item the data points lie in the feasible set $\mathcal{F}$, where $\mathcal{F} \subseteq \mathcal{X} \times \mathcal{Y}$. \end{enumerate} The attack projects the generated data sample onto $\mathcal{F}$ when it falls out of the feasible region. \textbf{\emph{Online attack \cite{collingedefending}}}. The attacker's objective is to change the state of the existing model by injecting the poisoned data points into the stream, thus degrading the system's performance. Let $g$ be the attacker's objective function at the iteration $t$ that is evaluated on the target dataset $D_c$. Also let attacker inject a poisoned data point $(x_t, y_t) \in \mathcal{F}$, such that: \begin{equation} \textbf{$x_t$}^{*} \in \underset{\textbf{$x_t$} \in \mathcal{F}}{argmax} \; g(\mathcal{D}_{c}, \theta_{t+1}) \label{eq:onlineattack} \end{equation} where, $g$ is evaluated at iteration $t+1$ with the parameters of the classifier. For the learning rate $\eta$ and loss function $\mathcal{L}$, defined by the defender, the parameters are computed using the OGD as $\theta_{t+1} = \theta_t - \eta \bigtriangledown_{\theta_t} \mathcal{L}(x_t, \theta_t)$. Equation \ref{eq:onlineattack} can be solved using gradient ascent: \begin{equation} x_{t_{n+1}} = x_{t_n} + \alpha \bigtriangledown_{x_{t_n}} g(D_c, \theta_{t+1}) \end{equation} where, $\alpha$ is the learning rate of the attacker in Equation \ref{eq:onlineattack}. The attack is mounted by first training the learning algorithm using the whole training set for some fixed number of iterations. After that, for each poisoning point a random subset of the training data is used, along with the poisoned points created in the previous iterations. The attack algorithm uses gradient descent for one epoch for updating the state of the learning algorithm and computing the poisoned data points \cite{collingedefending}. \subsection{Defenses against Data Poisoning Attacks} \emph{Data Sanitization} \cite{cretu2008casting} is a common defense against the data poisoning attacks, wherein the anomalous training points are removed before training the model. The objective is to remove the instances that are very different from the clean data instances. The existing defenses differ from each other on the basis of how they define an instance to be anomalous \cite{koh2018stronger}. Some of the defenses related to our work are \emph{L2}, \emph{slab} and \emph{loss}. The \emph{L2} defense discards points that are far away from the corresponding centroid, whereas, the \emph{slab} defense first projects points onto the line between the two class centroids and then discards points that are far away from the corresponding centroid. Similarly, the \emph{loss} defense discards data instances that are not well fit by the model on the full dataset. \section{Influence-based Defense for Online Learning}\label{sec:Influence_Defense} Koh et al \cite{koh2017understanding} show that the \emph{influence functions} give us an efficient approximation on how the model's predictions change if we did not have the training data point $\mathcal{D}_{i}$. As per their notations, let $L(\mathcal{D}, \theta)$ be the loss and $\frac{1}{n}\sum_{i=1}^{n} L(\mathcal{D}_{i}, \theta)$ be the empirical risk, which can be calculated by averaging the loss function on the training set. Also, let $\mathcal{D}_{i}$ be the training data point and $\theta$ be the model parameter. The empirical risk minimizer is given by $ \hat{\theta} \stackrel{\text { def }}{=} \arg \min _{\theta \in \Theta} \frac{1}{n} \sum_{i=1}^{n} L\left(\mathcal{D}_{i}, \theta\right)$ with an assumption that $L$ is twice-differentiable and strictly convex in $\theta$. \begin{algorithm}[ht] \caption{Influence-based Defense Algorithm} \label{algo:Inf_def} \begin{algorithmic}[1] \renewcommand{\algorithmicrequire}{\textbf{Input:}} \renewcommand{\algorithmicensure}{\textbf{Parameter:}} \Require {Poisoned Data Stream $D$ = $\{(x,y)\}_{1}^{n}$, Pre-train data $D_{pre}$, initial model $\theta_{0}$ learned over pre-train data} \Ensure {Gradient descent step size $\eta_{def}$, influence window size $w_{inf}$} \renewcommand{\algorithmicensure}{\textbf{Output:}} \Ensure {Model parameter $\theta_{n}$} \State Compute $\mathcal{F}_{slab}(D_{pre})$ on pre-train data using Equation \ref{eq:slab} \For{$x \in D$} \Comment{Doing for each $x$ from $D$ at time $t$} \State Filter $x_{t}$ based on $\mathcal{F}_{slab}$ \If{$x_{t}$ lies into feasible set $\mathcal{F}_{slab}(D_{pre})$} \State Pre-compute the Hessian using Equation \ref{eq:hessian} \State Pre-compute gradient loss using Equation \ref{eq:grad_loss} \State $Inf(x_{t}) \leftarrow$ Compute the Influence using Equation \ref{eq:influence} \State $Inf_{thres} \leftarrow$ Average of previous $w_{inf}$ influences \If {$Inf(x_{t}) \ge Inf_{thres}$} \State $x_{t}^{*}$ = $\underset{x}{\textbf{\emph{minimize}}} \; Inf(x)$ using Equation \ref{eq:grad_influence} \If{$Inf(x_{t}) \ge Inf(x_{t}^{*})$} \State $x_{t} \leftarrow x_{t}^{*}$ \EndIf \EndIf \State $\theta_{t} \leftarrow $ Update $\theta_{t-1}$ using $x_{t}$ \Else \State continue without updating the $\theta_{t}$ \EndIf \EndFor \State \textbf{return} $\theta_n$ \end{algorithmic} \end{algorithm} Given a training point $\mathcal{D}_{i}$, the change in parameter can be defined formally as $\hat{\theta}_{-\mathcal{D}_{i}}-\hat{\theta}$ where, \begin{equation*} \hat{\theta}_{-\mathcal{D}_{i}} \stackrel{\text{ def }}{=} \arg \min _{\theta \in \Theta} \sum_{\substack{j=1 \dots n \\ \mathcal{D}_{j} \neq \mathcal{D}_{i}}} L\left(\mathcal{D}_{j}, \theta\right) \end{equation*} The idea is to compute the parameter change if $\mathcal{D}_{i}$ were upweighted by some small $\epsilon$, that is, $\mathcal{D}_{i}$ is multiplied by $1+\epsilon$. The influence of upweighting $\mathcal{D}_{i}$ on the parameters $\hat{\theta}$ is given by: \begin{equation}\label{eq:influence} \mathcal{I}_{(\mathcal{D}_i)} \overset{def}{=} \frac{d \hat{\theta}_{\epsilon,\mathcal{D}_i}}{d\epsilon} \Bigr|_{\epsilon = 0} = -{H_{\hat{\theta}}^{-1}}\; \nabla_{\theta} {L}(\mathcal{D}_i, \hat{\theta}), \end{equation} where $H$ is the Hessian given by, \begin{equation}\label{eq:hessian} H_{\hat{\theta}} \stackrel{\text { def }}{=} \nabla^{2} R(\mathcal{D},\hat{\theta})=\frac{1}{n} \sum_{i=1}^{n} \nabla_{\theta}^{2} L(\mathcal{D}_{i}, \hat{\theta}) \end{equation} \begin{equation}\label{eq:RTheta} R(\mathcal{D},\theta) \stackrel{\text { def }}{=} \frac{1}{n} \sum_{i=1}^{n} L\left(\mathcal{D}_{i}, \theta\right) \end{equation} \begin{equation}\label{eq:thetaCap} \hat{\theta}_{\epsilon, \mathcal{D}_i}=\arg \min _{\theta \in \Theta}\{R(\mathcal{D}, \theta)+\epsilon L(\mathcal{D}_i, \theta)\} \end{equation} Given a training point $\mathcal{D}_{i} = (\bm{x},y)$ $\in \mathcal{D}$, we first run a data-sanitization defense \cite{steinhardt2017certified} called $slab$ wherein $\mu_{y}$, $\mu_{-y}$ are class centroids and $s_{y}$ is threshold, \begin{equation}\label{eq:slab} \mathcal{F}_{\text {slab }} \stackrel{\text { def }}{=}\left\{(\bm{x}, y):\left|\left\langle \bm{x}-\mu_{y}, \mu_{y}-\mu_{-y}\right\rangle\right| \leq s_{y}\right\} \end{equation} Our objective is to reduce the degradation caused by the poisoned instances present in the feasible set. We achieve that by using influence functions (refer Equation \ref{eq:influence}). Consider a generalized linear model function $h(.)$, parameterized by $\theta$. For example, the logistic regression classifier $$h_\theta(\bm{x})=\sigma(z)=\frac{1}{1+e^{-\theta^{T}\bm{x}}} = Pr(y=1 |~\bm{x}, \theta)$$ where, $z = \theta^{T}\bm{x}$. Let the loss function be $$L(\theta)=\sum_{i=1}^n -\Big( y_i\log\sigma(z_i)+(1-y_i)\log(1-\sigma(z_i))\Big)$$ Our focus is on the online setting where the weight vector is updated as in Equation \ref{eq:OGD}. The gradient of loss is given as \begin{equation}\label{eq:grad_loss} \begin{aligned} {\nabla}L(\mathcal{D}_{i},\theta) \; \; &\; = \; \frac{\partial L(\mathcal{D}_i,\theta)}{\partial \theta^T} \\ & \; = \; -y_i\bm{x}_i(1-\sigma(z_i))+(1-y_i)\bm{x}_i\sigma(z_i) \\ & \; =\; \bm{x}_i(\sigma(z_i)-y_i) \end{aligned} \end{equation} In order to achieve our objective, we first scalarize the influence function by taking the euclidean norm of it. If $\mathcal{S}$ is the scalarizing function, then $$ \nabla \mathcal{S}(\mathcal{I}_{(\bm{x})})=\mathcal{J}( \mathcal{I}_{(\bm{x})})^{\top} \nabla \mathcal{S}(a) \; \Bigr|_{a = \mathcal{I}_{(\bm{x})}} $$ where $\mathcal{J}$ is the jacobian matrix. $\nabla \mathcal{S}(a)$ is the gradient of $\mathcal{S}$ at the point $a = \mathcal{I}_{(\bm{x})} $. We then minimize the influence of training instances by applying gradient descent and perturbing the points \begin{equation}\label{eq:min_inf} \bm{x}_{t}^{*} = \underset{\bm{x}}{\textbf{\emph{minimize}}} \; \mathcal{I}_{(\bm{x})} \end{equation} \begin{equation}\label{eq:grad_influence} \nabla \mathcal{I}_{(\bm{x})}= - \nabla H^{-1}\left(\nabla_{\theta} L(\bm{x}, \theta)\right) \end{equation} \begin{equation}\label{eq:hessian_inf} H = {\nabla}^{2} L(\theta)=\sum_{i=1}^{n} \bm{x}_{i} \bm{x}_{i}^{T} \sigma\left(z_{i}\right)\left(1-\sigma\left(z_{i}\right)\right) \end{equation} where hessian $H$ is a constant with respect to the incoming point. $\mathcal{I}_{(\bm{x})}$ decreases fastest if one goes from x in the direction of the negative gradient of $\mathcal{I}_{(\bm{x})}$ at $\bm{x}$ i.e., -$\nabla \mathcal{I}_{(\bm{x})}$. We may lose some information because we perturb the clean data as well. We formalize the influence-based defense as follows. \emph{Defender takes as input the data instances, then filters out the suspicious poisoning points and \textbf{minimizes the influence} of the remaining data points such that the learned model parameter $\hat{\theta}$ has the lowest error on a test data set $\mathcal{D}_{test}$}. The complete procedure is explained in the Algorithm \ref{algo:Inf_def}. \subsection{Effect of Slab and Influence Based Defense on Clean Data} To study the effect of Slab and Influence based approaches, we generated synthetic data using numpy's\footnote{https://numpy.org/} random package with different seed values for pre-train (size:50,seed:0), train (size:100,seed:1), validation (size:50,seed:2) and test data (size:50,seed:3). Figure \ref{fig:trainingdata} shows the plot for training data, where the decision boundary is computed on the training data. Figure \ref{fig:radius} and Figure \ref{fig:slab}, show the plot of training data with slab radius and training data after slab based sanitization. The decision boundary in Figure \ref{fig:slab} is computed on the sanitized data. Figure \ref{fig:influence} shows the plot of data points that changes after applying the influence-based method on sanitized data. The decision boundary in Figure \ref{fig:influence} is computed using both the changed and unchanged data after applying influence-based defense. The accuracy of the classifier on the training data is observed to be 76\%. After applying the slab defense the accuracy changes to 68\%, however the accuracy reduces to 66\% after applying influence based defense. This drop in accuracy is due to `change' of some of the clean data points with high influence score. \begin{figure}[ht] \centering \begin{subfigure}[b]{0.48\textwidth} \centering \includegraphics[width=1\linewidth]{plot_of_training_data} \caption{} \label{fig:trainingdata} \end{subfigure \begin{subfigure}[b]{0.48\textwidth} \centering \includegraphics[width=1\linewidth]{plot_of_training_data_with_slab_radius} \caption{} \label{fig:radius} \end{subfigure} \begin{subfigure}[b]{0.48\textwidth} \centering \includegraphics[width=1\linewidth]{plot_of_sanitized_train_data_with_slabs} \caption{} \label{fig:slab} \end{subfigure \begin{subfigure}[b]{0.48\textwidth} \centering \includegraphics[width=1\linewidth]{plot_of_training_data_after_influence} \caption{} \label{fig:influence} \end{subfigure} \caption{Comparison of Slab and Influence based methods on synthetic data. Minimizing the influence sometimes affects the clean data points, which leads to information loss.} \label{fig:poisoning} \end{figure} \begin{table}[ht] \centering \caption{Datasets used for comparing Slab and Influence based method.} \label{tab:datasets} \begin{tabular}{@{}lccccc@{}} \toprule \textbf{\#} & \textbf{Dataset} & \textbf{Features} & \textbf{Pre-Train} & \textbf{Training} & \textbf{Test} \\ \midrule \textbf{D1} & Australian & 14 & 200 & 300 & 150 \\ \textbf{D2} & Banknote & 4 & 200 & 400 & 572 \\ \textbf{D3} & MNIST 1v7 & 50 & 8000 & 1000 & 2163 \\ \textbf{D4} & Spambase & 57 & 2000 & 1000 & 1519 \\ \textbf{D5} & UCI Breast Cancer & 9 & 100 & 400 & 100 \\ \textbf{D6} & Fashion MNIST (Bag versus Sandal) & 50 & 8000 & 1000 & 1000 \\ \bottomrule \end{tabular} \end{table} \section{Experiments}\label{sec:experiments} In this section we will briefly discuss the datasets that we have used, the test environment, experimental results and observations. \textbf{\emph{Hardware and Software Configuration.}} We have carried out all our experiments on a machine having Intel Core i7-8550U CPU with 4 cores, 8 logical processors and a base speed of 2.0 GHz. The system has 16 GB DDR4 RAM and 500 GB HD with 230 GB free space. Also, the test system has L1, L2 and L3 cache of size 256 KB, 1 MB and 8 MB respectively. Our proposed defense is implemented using Python 3.6 with scikit-learn\footnote{https://scikit-learn.org/stable/supervised\_learning.html} implementation of the classifiers, and tested on 64-bit Windows 10 enterprise edition and Ubuntu 16.04 LTS version. \begin{table}[ht] \centering \caption{Comparison of results on six datasets for 10\% poisoning budget and with different learning rates. Results, where one method is better than the other, are in bold. The datasets \textbf{D1} to \textbf{D6} are discussed in Table \ref{tab:datasets}. \textbf{Slab (\textit{S})} and \textbf{Influence-based (\textit{I})} defenses are compared against the Simplistic and Online attacks.} \label{tab:results} \begin{tabular}{@{}cccccccccc@{}} \toprule & & \multicolumn{2}{c}{\textbf{LR: 0.01}} & \multicolumn{2}{c}{\textbf{LR: 0.05}} & \multicolumn{2}{c}{\textbf{LR: 0.09}} & \multicolumn{2}{c}{\textbf{LR: Optimal}} \\ \midrule & & \textit{Simplistic} & \textit{Online} & \textit{Simplistic} & \textit{Online} & \textit{Simplistic} & \textit{Online} & \textit{Simplistic} & \textit{Online} \\ \multirow{2}{*}{\textbf{D1}} & \textit{S} & 0.8736 & 0.8736 & 0.8736 & 0.8526 & 0.8789 & 0.7947 & 0.8578 & 0.5736 \\ & \textit{I} & 0.8736 & 0.8736 & 0.8736 & \textbf{0.8578} & 0.8789 & \textbf{0.8052} & 0.8578 & \textbf{0.6211} \\ \multirow{2}{*}{\textbf{D2}} & \textit{S} & 0.8391 & 0.8671 & \textbf{0.8287} & 0.8933 & \textbf{0.7639} & \textbf{0.9161} & 0.4423 & 0.9493 \\ & \textit{I} & 0.8391 & 0.8671 & 0.806 & \textbf{0.8951} & 0.5804 & 0.8951 & 0.4423 & \textbf{0.9755} \\ \multirow{2}{*}{\textbf{D3}} & \textit{S} & 0.9773 & 0.9884 & \textbf{0.9242} & \textbf{0.9917} & \textbf{0.8335} & 0.9801 & 0.6643 & 0.5839 \\ & \textit{I} & \textbf{0.9787} & \textbf{0.9889} & 0.8974 & 0.9847 & 0.8220 & \textbf{0.9810} & \textbf{0.7753} & \textbf{0.9223} \\ \multirow{2}{*}{\textbf{D4}} & \textit{S} & 0.8933 & 0.8531 & 0.6893 & 0.7544 & 0.6372 & \textbf{0.7281} & 0.6254 & \textbf{0.7182} \\ & \textit{I} & \textbf{0.8940} & 0.8531 & \textbf{0.6899} & \textbf{0.7551} & \textbf{0.6458} & 0.7228 & \textbf{0.6445} & 0.6886 \\ \multirow{2}{*}{\textbf{D5}} & \textit{S} & 0.92 & 0.88 & 0.94 & 0.53 & 0.94 & 0.52 & 0.88 & 0.49 \\ & \textit{I} & 0.92 & 0.88 & 0.94 & 0.53 & 0.94 & \textbf{0.68} & 0.88 & \textbf{0.71} \\ \multirow{2}{*}{\textbf{D6}} & \textit{S} & 0.986 & 0.922 & 0.985 & \textbf{0.767} & 0.988 & 0.697 & 0.988 & 0.981 \\ & \textit{I} & 0.986 & \textbf{0.926} & \textbf{0.988} & 0.733 & 0.988 & \textbf{0.737} & 0.988 & 0.981 \\ \cmidrule(l){2-10} \end{tabular} \end{table} \begin{figure}[ht] \centering \begin{subfigure}[b]{0.33\textwidth} \begin{tikzpicture} \begin{axis}[ymin=0.6, ymax=1, xtick={0.01,0.05,0.09,0.13}, xticklabels={$LR_{1}$,$LR_2$,$LR_3$,$LR_4$}, ytick={0.6, 0.7, 0.8,0.9,1.0}, ymajorgrids=true, grid style=dashed, width=4.8cm, height=4.8cm] \addplot[color=red,mark=square,line width=1pt] coordinates {(0.01,0.8736)(0.05,0.8736)(0.09,0.8578)(0.13,0.7263)}; \addplot[color=black,mark=triangle,line width=1pt] coordinates {(0.01,0.8736)(0.05,0.8736)(0.09,0.8789)(0.13,0.8578)}; \end{axis} \end{tikzpicture} \caption{Australian} \end{subfigure \begin{subfigure}[b]{0.33\textwidth} \begin{tikzpicture} \begin{axis}[ymin=0.4, ymax=1, xtick={0.01,0.05,0.09,0.13}, xticklabels={$LR_{1}$,$LR_2$,$LR_3$,$LR_4$}, ytick={0.4,0.5,0.6, 0.7,0.8,0.9,1.0}, ymajorgrids=true, grid style=dashed, width=4.8cm, height=4.8cm] \addplot[color=red,mark=square,line width=1pt] coordinates {(0.01,0.8531)(0.05,0.8549)(0.09,0.7622)(0.13,0.4423)}; \addplot[color=black,mark=triangle,line width=1pt] coordinates {(0.01,0.8391)(0.05,0.806)(0.09,0.5804)(0.13,0.4423)}; \end{axis} \end{tikzpicture} \caption{Banknote} \end{subfigure \begin{subfigure}[b]{0.33\textwidth} \begin{tikzpicture} \begin{axis}[ymin=0.6, ymax=1, xtick={0.01,0.05,0.09,0.13}, xticklabels={$LR_{1}$,$LR_2$,$LR_3$,$LR_4$}, ytick={0.6, 0.7,0.8,0.9,1.0}, ymajorgrids=true, grid style=dashed, width=4.8cm, height=4.8cm] \addplot[color=red,mark=square,line width=1pt] coordinates {(0.01,0.9778)(0.05,0.9288)(0.09,0.8340)(0.13,0.6532)}; \addplot[color=black,mark=triangle,line width=1pt] coordinates {(0.01,0.9787)(0.05,0.8974)(0.09,0.8220)(0.13,0.7753)}; \end{axis} \end{tikzpicture} \caption{MNIST} \end{subfigure} \begin{subfigure}[b]{0.33\textwidth} \begin{tikzpicture} \begin{axis}[ymin=0.6, ymax=1, xtick={0.01,0.05,0.09,0.13}, xticklabels={$LR_{1}$,$LR_2$,$LR_3$,$LR_4$}, ytick={0.6, 0.7,0.8,0.9,1.0}, ymajorgrids=true, grid style=dashed, width=4.8cm, height=4.8cm] \addplot[color=red,mark=square,line width=1pt] coordinates {(0.01,0.9019)(0.05,0.688)(0.09,0.6379)(0.13,0.6629)}; \addplot[color=black,mark=triangle,line width=1pt] coordinates {(0.01,0.8940)(0.05,0.6899)(0.09,0.6458)(0.13,0.6445)}; \end{axis} \end{tikzpicture} \caption{Spambase} \end{subfigure \begin{subfigure}[b]{0.33\textwidth} \begin{tikzpicture} \begin{axis}[ymin=0.8, ymax=1, xtick={0.01,0.05,0.09,0.13}, xticklabels={$LR_{1}$,$LR_2$,$LR_3$,$LR_4$}, ytick={0.8,0.9,1.0}, ymajorgrids=true, grid style=dashed, width=4.8cm, height=4.8cm] \addplot[color=red,mark=square,line width=1pt] coordinates {(0.01,0.92)(0.05,0.95)(0.09,0.91)(0.13,0.87)}; \addplot[color=black,mark=triangle,line width=1pt] coordinates {(0.01,0.92)(0.05,0.94)(0.09,0.94)(0.13,0.88)}; \end{axis} \end{tikzpicture} \caption{UCI Breast Cancer} \end{subfigure \begin{subfigure}[b]{0.33\textwidth} \begin{tikzpicture} \begin{axis}[ymin=0.98, ymax=1, xtick={0.01,0.05,0.09,0.13}, xticklabels={$LR_{1}$,$LR_2$,$LR_3$,$LR_4$}, ytick={0.98,1.0}, ymajorgrids=true, grid style=dashed, width=4.6cm, height=4.8cm] \addplot[color=red,mark=square,line width=1pt] coordinates {(0.01,0.985)(0.05,0.985)(0.09,0.988)(0.13,0.989)}; \addplot[color=black,mark=triangle,line width=1pt] coordinates {(0.01,0.986)(0.05,0.988)(0.09,0.988)(0.13,0.988)}; \end{axis} \end{tikzpicture} \caption{Fashion MNIST} \end{subfigure} \caption{Accuracy versus Learning Rate comparison for classifier with and without defense for \textbf{simplistic attack}. The \ref{L2} refers to classifier with Influence based defense and \ref{L1} refers to classifier without defense. $LR_{1}$,$LR_2$,$LR_3$ and $LR_4$ refers to 0.01, 0.05, 0.09 and optimal learning rates.} \label{fig:simplistic} \end{figure} \begin{figure}[ht] \centering \begin{subfigure}[b]{0.33\textwidth} \begin{tikzpicture} \begin{axis}[ymin=0.4, ymax=1, xtick={0.01,0.05,0.09,0.13}, xticklabels={$LR_{1}$,$LR_2$,$LR_3$,$LR_4$}, ytick={0.4,0.5,0.6, 0.7, 0.8,0.9,1.0}, ymajorgrids=true, grid style=dashed, width=4.8cm, height=4.8cm] \addplot[color=red,mark=square,line width=1pt] coordinates {(0.01,0.8473)(0.05,0.5526)(0.09,0.5263)(0.13,0.6421)}; \label{L1} \addplot[color=black,mark=triangle,line width=1pt] coordinates {(0.01,0.8736)(0.05,0.8578)(0.09,0.8052)(0.13,0.6210)};\label{L2} \end{axis} \end{tikzpicture} \caption{Australian} \end{subfigure \begin{subfigure}[b]{0.33\textwidth} \begin{tikzpicture} \begin{axis}[ymin=0.5, ymax=1, xtick={0.01,0.05,0.09,0.13}, xticklabels={$LR_{1}$,$LR_2$,$LR_3$,$LR_4$}, ytick={0.5,0.6, 0.7, 0.8,0.9,1.0}, ymajorgrids=true, grid style=dashed, width=4.8cm, height=4.8cm] \addplot[color=red,mark=square,line width=1pt] coordinates {(0.01,0.7727)(0.05,0.6573)(0.09,0.5769)(0.13,0.5576)}; \addplot[color=black,mark=triangle,line width=1pt] coordinates {(0.01,0.8671)(0.05,0.8951)(0.09,0.8951)(0.13,0.9755)}; \end{axis} \end{tikzpicture} \caption{Banknote} \end{subfigure \begin{subfigure}[b]{0.33\textwidth} \begin{tikzpicture} \begin{axis}[ymin=0.4, ymax=1, xtick={0.01,0.05,0.09,0.13}, xticklabels={$LR_{1}$,$LR_2$,$LR_3$,$LR_4$}, ytick={0.4,0.5,0.6, 0.7, 0.8,0.9,1.0}, ymajorgrids=true, grid style=dashed, width=4.8cm, height=4.8cm] \addplot[color=red,mark=square,line width=1pt] coordinates {(0.01,0.9870)(0.05,0.4877)(0.09,0.4845)(0.13,0.4845)}; \addplot[color=black,mark=triangle,line width=1pt] coordinates {(0.01,0.9889)(0.05,0.9847)(0.09,0.9810)(0.13,0.9223)}; \end{axis} \end{tikzpicture} \caption{MNIST} \end{subfigure} \begin{subfigure}[b]{0.33\textwidth} \begin{tikzpicture} \begin{axis}[ymin=0.6, ymax=1, xtick={0.01,0.05,0.09,0.13}, xticklabels={$LR_{1}$,$LR_2$,$LR_3$,$LR_4$}, ytick={0.6, 0.7, 0.8,0.9,1.0}, ymajorgrids=true, grid style=dashed, width=4.8cm, height=4.8cm] \addplot[color=red,mark=square,line width=1pt] coordinates {(0.01,0.8077)(0.05,0.7564)(0.09,0.7175)(0.13,0.6787)}; \addplot[color=black,mark=triangle,line width=1pt] coordinates {(0.01,0.8531)(0.05,0.7551)(0.09,0.7228)(0.13,0.6886)}; \end{axis} \end{tikzpicture} \caption{Spambase} \end{subfigure \begin{subfigure}[b]{0.33\textwidth} \begin{tikzpicture} \begin{axis}[ymin=0.3, ymax=1, xtick={0.01,0.05,0.09,0.13}, xticklabels={$LR_{1}$,$LR_2$,$LR_3$,$LR_4$}, ytick={0.3,0.4,0.5,0.6, 0.7, 0.8,0.9,1.0}, ymajorgrids=true, grid style=dashed, width=4.8cm, height=4.8cm] \addplot[color=red,mark=square,line width=1pt] coordinates {(0.01,0.6)(0.05,0.44)(0.09,0.38)(0.13,0.43)}; \addplot[color=black,mark=triangle,line width=1pt] coordinates {(0.01,0.88)(0.05,0.53)(0.09,0.68)(0.13,0.71)}; \end{axis} \end{tikzpicture} \caption{UCI Breast Cancer} \end{subfigure \begin{subfigure}[b]{0.33\textwidth} \begin{tikzpicture} \begin{axis}[ymin=0.3, ymax=1, xtick={0.01,0.05,0.09,0.13}, xticklabels={$LR_{1}$,$LR_2$,$LR_3$,$LR_4$}, ytick={0.3,0.4,0.5,0.6, 0.7, 0.8,0.9,1.0}, ymajorgrids=true, grid style=dashed, width=4.8cm, height=4.8cm] \addplot[color=red,mark=square,line width=1pt] coordinates {(0.01,0.424)(0.05,0.45)(0.09,0.501)(0.13,0.455)}; \addplot[color=black,mark=triangle,line width=1pt] coordinates {(0.01,0.926)(0.05,0.733)(0.09,0.737)(0.13,0.981)}; \end{axis} \end{tikzpicture} \caption{Fashion MNIST} \end{subfigure} \caption{Accuracy versus Learning Rate comparison for classifier with and without defense for \textbf{online attack}. The \ref{L2} refers to classifier with Influence based defense and \ref{L1} refers to classifier without defense. $LR_{1}$,$LR_2$,$LR_3$ and $LR_4$ refers to 0.01, 0.05, 0.09 and optimal learning rates.} \label{fig:online} \end{figure} \textbf{\emph{Datasets.}} We consider 6 datasets for our experiments. Table \ref{tab:datasets} shows the list of the datasets, number of features used for training the model and the distribution of the data for initialization, training and testing. The poisoning budget computation is relative to the training data. We validate our results on six datasets, three of which are common in \cite{wang2019investigation}. \textbf{\emph{Baseline Attacks.}} We consider two existing poisoning attacks on online learning for evaluating our defense strategy: a) \emph{simplistic attack} \cite{wang2018data}, and b) \emph{online attack} \cite{collingedefending} (refer to Section \ref{sec:target}). We have used the constant learning rates of 0.01 and 0.05 as per \cite{wang2019investigation} and an additional constant learning rate of 0.09. Moreover, we have also used the optimal learning rate provided by the scikit-learn to assess the performance of Slab and our influence based method. \textbf{\emph{Defenses.}} Our defense is in conjunction with the slab defense. The influence window size ($w_{inf}$) in our defense algorithm is empirically found using grid search. The poisoned points are generated by the adversary using the baseline attacks. The defender first uses slab defense for data sanitization in the incoming data stream. She then minimizes the influence of the data point based on its impact on the model. The defense mechanism is effective when the classifier has higher score than the classifier under baseline attacks. \textbf{\emph{Results and Discussion.}} Table \ref{tab:results} shows the effectiveness of Slab (\textit{S}) and Influence based (\textit{I}) defenses against Simplistic and Online attacks. We have performed experiments on six different datasets .We have considered three constant learning rates of 0.01, 0.05 and 0.09 respectively. In addition, we have also compared the two defenses for an optimal learning rate. It can be observed that on an average the performance of slab and influence based methods are nearly same for constant learning rate. On the contrary, for an optimal learning rate, the influence based method has far better result than slab for most of the datasets. For example, Influence defense is more than 9\% accurate than Slab for simplistic attack and more than 30\% accurate than online attack for MNIST dataset. As per our observation, the simplistic attack is apparently more powerful than the online attack. This is due to the fact that simplistic attack inserts poisoned examples which lie in the feasible set $\mathcal{F}$, whereas the online attack of \cite{collingedefending} does not ensure that. We found that the number of poisoned examples kept after the Slab defense is almost equal to the attacker's budget against the simplistic attack. However, comparatively few poisoned examples (in some cases zero) are kept against the online attack. Comparatively the low accuracy of influence based method in some cases could be because of the change in the influence of clean samples. Also, the percentage of clean data samples that remain after slab defense for datasets \textbf{D1} to \textbf{D6} are 43.33, 76.5, 60.8, 69.8, 97.75, 90.90 respectively, for both the attacks. The percentage difference is due to the data-dependent nature of the slab defense \cite{wang2019investigation}. Table \ref{tab:results_5_budget} and Table \ref{tab:results_15_budget} show comparative results for 5\% and 15\% poisoning budget (Section \ref{sec:appresult}). Figure \ref{fig:simplistic} shows the comparison of accuracy of the classifier with and without influence based defense for simplistic attack for varying learning rates. Similarly, Figure \ref{fig:online} shows the change in classifier's accuracy with and without influence defense for the online attack. The poison budget is kept 10\% and learning rate is changed from fixed (0.01, 0.05, 0.09) to optimal. One common observation is that the accuracy of the classifier degrades more with optimal learning rate and in absence of a defense. For online attack of \cite{collingedefending}, our proposed works well to improve the accuracy in most cases. On the contrary, for simplistic attack our defense has relatively lower performance as the poisoned data points are added in the feasible set only, that is irrespective of their influence. \section{Conclusion}\label{sec:conclusion} In this paper, we formulated a defense algorithm that compliments the slab defense with the influence function such that the degradation caused by the poisoned data is minimized. We studied the performance of our defense against different attacks and across multiple datasets. Further, we validated our experiments with simplistic and online attacks of \cite{wang2019investigation,collingedefending}. We have also demonstrated the performance degradation of the classifier with and without defense. One of the trade-offs from the defender's side is that the objective function to minimize the influence sometimes affects the clean data points, which leads to information loss. \bibliographystyle{splncs04}

735fb94cc3e8697d7a204cb27ee2e83ca6148159

\section{Introduction} Positron Emission Tomography - Computed Tomography (PET/CT) scanning is an effective imaging tool for lymphoma segmentation with application to clinical diagnosis and radiotherapy planning. The standardized uptake value (SUV) for PET images is widely used to locate and segment lymphomas thanks to its high sensitivity and specificity to the metabolic activity of tumor. CT images are usually used jointly with PET images because of their anatomical feature representation capability. Although a lot of progress has been made in computer-aided lymphoma segmentation, the segmentation of whole-body lymphomas is still challenging. (Fig.~\ref{fig1} shows an example of lymphoma patient. There is great variation in intensity distribution, shape, type and number of lymphomas). The methods can be classified into three main categories: SUV-threshold-based \cite{ilyas2018defining}, region-growing-based \cite{hu2019detection} and Convolutional Neural Network (CNN)-based \cite{li2019densex} methods. For PET images, it is common to segment lymphomas with a set of fixed SUV thresholds. This method is fast but lacks of flexibility in boundary delineation and requires domain knowledge to locate the region of interest. Region-growing-based methods have been proposed to optimize boundary delineation by taking texture and shape information into account. However, those methods still need clinicians to locate the seeds for region growing \cite{onoma2014segmentation}. \begin{figure} \includegraphics[width=\textwidth]{fig1.pdf} \caption{Examples of patient with lymphomas. The first and second rows show, respectively PET and CT, slices of one patient in axial, sagittal and coronal views. The lymphomas are marked in red.} \label{fig1} \end{figure} CNN-based segmentation methods have recently achieved great success. The UNet architecture \cite{ronnebergerconvolutional} has become the most popular medical image segmentation model. Driven by different tasks and datasets, many extended and optimized variants of UNet have been proposed, such as VNet \cite{milletari2016v}, nnUNet \cite{isensee2018nnu} and SegResNet \cite{myronenko20183d}. In \cite{li2019densex}, Li et al. propose a SegResNet-based lymphoma segmentation model with a two-flow architecture (segmentation and reconstruction flows). In \cite{blanc2020fully}, Blanc-Durand et al. propose a nnUNet-based lymphoma segmentation network. Because of low resolution and contrast due to limitations of medical imaging technology, PET/CT image segmentation results are tainted with uncertainty. Belief function (BF) theory \cite{shafer1976mathematical}\cite{denoeux20b}, also known as Dempster-Shafer theory, is a formal theory for information modeling, evidence combination and decision-making under uncertainty. In this paper, we propose a 3D PET/CT diffuse large b—cell lymphoma segmentation model based on BF theory and deep learning. The proposed deep neural network architecture is composed of a UNet module for feature extraction and a BF module for decision with uncertainty quantification. End-to-end learning is achieved by minimizing a two-part loss function allowing us to increase the Dice score while decrease the uncertainty. The model will first be described in Section \ref{sec:model} and experimental results will be reported in Section \ref{sec:results}. \section{Methods} \label{sec:model} \subsection{Network Architecture} Fig.~\ref{fig2} shows the global lymphoma segmentation architecture (ES-UNet). It is composed of (1) an encoder-decoder feature extraction module (UNet), and (2) an evidential segmentation (ES) module comprising a distance activation layer, a basic belief assignment layer and a mass fusion layer. Details about the ES module will be given in Section \ref{subsec:ES}. Two loss terms are used for optimizing the training process: the \emph{Dice loss}, which quantifies the segmentation accuracy and the \emph{uncertainty loss}, which quantifies the segmentation uncertainty. These loss functions will be described in Section \ref{subsec:loss}. A ``slim UNet'' with $(8, 16, 32, 64, 128)$ convolution filters was implemented to reduce computation cost and avoid overfitting. \begin{figure} \includegraphics[width=\textwidth]{fig2.pdf} \caption{Global lymphoma segmentation model (ES-UNet).} \label{fig2} \end{figure} \subsection{Evidential segmentation module} \label{subsec:ES} A probabilistic network with a softmax output layer may assign voxels a high probability of belonging to one class while the segmentation uncertainty is actually very high because, e.g., the voxel is located close to the fuzzy boundary between the tumor region and other tissues. Based on the evidential neural network model introduced in \cite{denoeux2000neural} and using an approach similar to that recently described in \cite{tong21a}, we propose a BF theory-based ES module to quantify the uncertainty about the class of each voxel by a Dempster-Shafer mass function. The main idea of the ES module is to assign a mass to each of the $K$ classes and to the whole set of classes $\Omega$, based on the distance between the feature vector of each voxel and $I$ prototype centers. For a given voxel $x$, each prototype $p_{i}$ is considered as a piece of evidence, the reliability of which decreases with the Euclidean distance $d_{i}$ between $x$ and $p_{i}$. Each prototype $p_{i}$ is assumed to have a membership degree $u_{ik}$ to each class $\omega_k$ with the constraint $\sum_{k=1}^{K}u_{ik}=1$. The mass function induced by prototype $p_{i}$ is \begin{subequations} \label{eq:1} \begin{align} m_i(\{\omega_{k}\})&=\alpha _{i}u _{ik}\exp(-\gamma _{i}d_{i}^{2}), \quad k=1,\ldots,K\\ m_{i}(\Omega )&=1-\alpha _{i}\exp(-\gamma _{i}d_{i}^{2}), \end{align} \end{subequations} The mass functions induced by the $I$ prototypes are then combined by Dempster's rule \cite{shafer1976mathematical} \begin{equation} m=\bigoplus _{i=1}^{I}m_{i}. \label{eq:3} \end{equation} In our case, $\Omega=\{a, b\}$ and $K=2$. The ES module outputs for each voxel three mass values: two masses corresponding to lymphoma ($\{a\}$) and background ($\{b\}$), and an additional mass corresponding to ignorance ($\Omega$). For the voxels that are easy to classify into lymphoma or background, the mass values $m(\{a\})$ or $m(\{b\})$ are high and the mass $m(\{a, b\})$ is low. A high mass $m(\{a, b\})$ signals a lack of information to make a reliable decision. \subsection{Loss function based on accuracy and uncertainty for segmentation} \label{subsec:loss} In general, a good segmentation system is expected to make few segmentation errors while providing as informative outputs as possible. Since we quantify uncertainty by the ``ignorance class'' via the evidential network, we propose to minimize a loss function defined as the sum of two terms: a Dice loss $\textsf{loss}_{d}$ that measures the discrepancy between the ground truth and segmentation outputs, and an uncertainty loss $\textsf{loss}_u$ that measures the uncertainty of the segmentation outputs. More precisely, the Dice loss is defined as \begin{equation} \textsf{loss}_{d}=1-\frac{2 \sum_{n=1}^{N} S_n G_n}{ \sum_{n=1}^{N} S_n+ \sum_{n=1}^{N} G_n}, \label{eq:4} \end{equation} where $N$ is the number of voxels in the image volume, $S$ is the segmentation outputs of our model and $G$ is the ground truth. The uncertainty loss is defined as \begin{equation} \textsf{loss}_{u}={\frac {1}{N}}\sum _{n=1}^{N} [m_n(\Omega)]^2, \label{eq:5} \end{equation} where $m_n$ is the mass function computed for voxel $n$. The total loss function is then \begin{equation} \textsf{loss}=\textsf{loss}_{d}+\textsf{loss}_{u} +\lambda \left \| \alpha \right \|_1, \label{eq:6} \end{equation} where $\lambda$ is the regularization coefficient for parameter vector $\alpha=(\alpha_1,\ldots,\alpha_I)$ with $\alpha_i$ defined in \eqref{eq:1}. The regularization term allows us to decrease the influence of unimportant prototype centers and avoid overfitting. \section{Experimental results} \label{sec:results} \subsection{Experimental settings} The dataset contains images from 173 patients who were diagnosed with large b-cell lymphoma and underwent PET/CT examination. The study was approved as a retrospective study by the Henri Becquerel Center Institutional Review Board. The lymphomas in mask images were delineated manually by experts and considered as ground truth. The size and spatial resolution of PET and CT images and the corresponding mask images vary due to different imaging machines and operations, which makes it difficult to transfer the data into a deep neural model directly. For prepossessing, we resized PET, CT and mask images to the same size $256\times256 \times128$. For normalization, we set (shift, scale) values to (0, 0.1) and (1000, 1/2000) for, respectively PET and CT images. We randomly selected 80\% of the data for training, 10\% for validation and 10\% for testing. Dice score, sensitivity, specificity, precision and F1 score were used to evaluate the segmentation performance. We first computed the five indices for each test patient and then averaged these indices over the patients. During training, PET and CT images were concatenated as a two-channel input. The prototype vectors and membership degrees where initialized randomly with uniform distributions, and the vectors \textsf{$\alpha$} and \textsf{$\gamma$} where initialized with constants 0.5 and 0.01. The number of prototypes was set to 20. The learning rate was set to $10^{-3}$ during training and the model was trained with 50 epochs using the Adam optimization algorithm. The regularization coefficient $\lambda$ in \eqref{eq:6} was set to $10^{-5}$. All methods were implemented in Python with a PyTorch-based, medical image framework MONAI and were trained and tested on a desktop with a 2.20GHz Intel(R) Xeon(R) CPU E5-2698 v4 and a Tesla V100-SXM2 graphics card with 32 GB GPU memory. \subsection{Results and discussion} The quantitative results are shown in Table~\ref{tab1}. Our model outperforms the baseline model UNet as well as the other state-of-the-art methods. In particular, our model outperforms the best model SegResNet by, respectively, 1.9\%, 2.4\%, 1.4\% in Dice score, Sensitivity and F1 score. It should be noted that the state-of-the-art models were trained with 100 epochs on our dataset because they are slower to converge during training. Fig.~\ref{fig3} displays the learning curves of the training loss and validation Dice score for UNet and ES-UNet, showing the advantage of ES-UNet in terms not only of segmentation accuracy, but also of convergence speed. Fig.~\ref{fig4} shows the segmentation and uncertainty maps at different steps during the training of ES-UNet. Our model quantifies the uncertainty of ambiguous pixels instead of classifying them unambiguously into a single class. The uncertainty decreases during the learning process thanks to the minimization of the uncertainty loss term. \begin{table} \caption{Performance comparison with the baseline methods on the test set.} \centering \label{tab1} \begin{tabular}{|l|l|l|l|l|l|} \hline Models &Dice score &Sensitivity& Specificity & Precision & F1 score\\ \hline ES-UNet (our model) & \textbf{0.830} & 0.923 & 0.908 &0.912 & \textbf{0.915}\\ UNet \cite{ronnebergerconvolutional} & 0.769& 0.798 & \textbf{0.963} &0.890&0.833\\ nnUNet \cite{isensee2018nnu} & 0.702 & \textbf{0.950} & 0.499 &0.758 & 0.807\\ VNet \cite{milletari2016v}&0.802&0.882 & 0.904 & 0.916&0.909\\ SegResNet \cite{myronenko20183d}& 0.811 & 0.899 & 0.942 & \textbf{0.925} & 0.901\\ \hline \end{tabular} \end{table} \begin{figure} \centering \includegraphics[width=\textwidth]{fig3.pdf} \caption{Training process visualization: training loss (left) and validation Dice score (right).} \label{fig3} \end{figure} \begin{figure} \centering \includegraphics[width=\textwidth]{fig4.pdf} \caption{Uncertainty maps obtained during the training. For one map, the pixels classified to background, lymphoma, ignorance are marked in purple, yellow and iridescent, respectively.} \label{fig4} \end{figure} Fig.~\ref{fig5} shows an example of segmentation results obtained by ES-UNet. Our model can locate and segment most of the lymphomas. The segmentation results were found credible and were confirmed by experts. \begin{figure} \centering \includegraphics[width=\textwidth]{fig5.pdf} \caption{Segmentation results of ES-UNet. From left to right: ground truth, segmentation results, difference map between the ground truth, and segmented lymphomas.} \label{fig5} \end{figure} \section{Conclusion} An evidential segmentation framework (ES-UNet) for segmentation of lymphomas from 3D PET/CT with uncertainty quantification has been introduced. The proposed architecture is based on the concatenation of a UNet and an evidential segmentation layer, making it possible to compute output mass functions for each voxel. The training is performed by minimizing a two-part loss function composed of a Dice loss and an uncertainty loss, with the effect of increasing the Dice score while decreasing the uncertainty. Qualitative and quantitative evaluations show promising results when compared to the baseline model UNet as well as the state-of-the-art methods. Future research will tackle multi-modality medical image fusion with BF theory. \bibliographystyle{splncs04}

887ce1517266acbd7961c11a68af3a94db560a28

\section{Introduction} \label{SEC:Introduction} Understanding the transverse structure of hadrons is an important step towards the three-dimensional imaging of hadrons. One of the key quantities that characterizes such transverse structure is the transverse-momentum-dependent parton distribution functions (TMDPDFs), which are a natural generalization of collinear PDFs to incorporate the transverse momentum of partons in the hadron, and provide crucial inputs for describing multi-scale, noninclusive observables at high-energy colliders such as the LHC~\cite{Lin:2020rut}. Currently, our knowledge of TMDPDFs mainly comes from studies of Drell-Yan and semi-inclusive deep-inelastic scattering processes where the transverse momenta of final state particles are measured. QCD factorization theorems allow to relating the relevant experimental observables to TMDPDFs via perturbatively calculable kernels, and thus provide the theoretical basis for extracting TMDPDFs from such observables. In the past, there have been various TMDPDF fittings in the literature~\cite{Bacchetta:2017gcc,Scimemi:2017etj,Bertone:2019nxa,Scimemi:2019cmh,Bacchetta:2019sam,Bacchetta:2020gko}. However, calculating TMDPDFs from first principles has been a challenge, because they are nonperturbative quantities defined in terms of light-cone correlations. Early lattice efforts have been focused on extracting certain information of TMDPDFs by studying ratios of suitable correlators~\cite{Hagler:2009mb,Musch:2010ka,Musch:2011er,Engelhardt:2015xja,Yoon:2017qzo}, whereas the full distribution also becomes accessible due to the proposal of large momentum effective theory (LaMET)~\cite{Ji:2013dva,Ji:2014gla,Ji:2020ect} which provides, in principle, a general recipe to calculate light-front (LF) correlations from lattice QCD. In the past few years, there has been rapid progress~\cite{Ji:2014hxa,Ji:2018hvs,Ji:2019ewn,Ji:2020jeb,Zhang:2020dbb,Ebert:2018gzl,Ebert:2019okf,Ebert:2019tvc,Shanahan:2020zxr,Shanahan:2019zcq,Ebert:2020gxr,Vladimirov:2020ofp} on how to extract the quark TMDPDFs from appropriately defined quasi-LF correlations involving staple-shaped Wilson line operators. A viable matching between the quasi-TMDPDF and TMDPDF, with a proper Euclidean construction of soft function for the former, has been established, although either for not fully renormalized quasi-TMDPDFs~\cite{Ji:2019ewn} or in a scheme~\cite{Ebert:2019tvc} that introduces undesired nonperturbative effects at large longitudinal distances. In addition, there have been exploratory lattice studies on the soft function~\cite{Zhang:2020dbb} and the Collins-Soper kernel~\cite{Shanahan:2020zxr,Schlemmer:2021aij} controlling the rapidity evolution of the TMDPDFs, as well as on the potential operator mixings under lattice regularization~\cite{Shanahan:2019zcq,Green:2020xco}. Another important quantity that encompasses information on the transverse structure of hadrons is the TMD wave functions (TMDWFs) or LF wave functions, from which one can actually obtain all parton densities. They are defined by the same staple-shaped Wilson line operators, and thus the lattice computation follows a similar strategy as that for the TMDPDFs~\cite{Ji:2020ect}. The quasi-TMDWF also enters the calculation of soft function through the TMD factorization of a light-meson form factor at large momentum transfer~\cite{Ji:2020ect,Zhang:2020dbb}. In this work, we perform a systematic analysis of the mixing pattern of staple-shaped Wilson line operators under lattice regularization using symmetry considerations. Similar analysis has been performed for straight Wilson line operators defining the quark quasi-PDFs in Ref.~\cite{Chen:2017mie}, where the authors analyzed the transformation properties of straight Wilson line operators with various Dirac structures and found the same mixing observed in one-loop lattice perturbation theory calculations for Wilson fermions~\cite{Constantinou:2017sej}. The lattice perturbation theory studies have also been extended to quark quasi-TMDPDFs in Ref.~\cite{Constantinou:2019vyb}, revealing certain mixings among operators with different Dirac structures (see also Ref.~\cite{Green:2020xco}). However, a systematic analysis of the operator mixing pattern from symmetry considerations is still missing. Here we generalize the discussion of Ref.~\cite{Chen:2017mie} to staple-shaped Wilson line operators. The results show mixings that are not present in one-loop lattice perturbation theory calculations. We also discuss the renormalization and matching of quasi-TMDPDFs and -TMDWFs in a scheme where no extra non-perturbative effects are introduced at large distances in the renormalization stage, in the same spirit as the hybrid renormalization~\cite{Ji:2020brr} proposed recently for the quasi-PDFs. The rest of the paper is organized as follows: In Sec.~\ref{SEC:qTMD}, we give a brief overview of the quasi-TMDPDFs and -TMDWFs in LaMET, both are defined in terms of staple-shaped Wilson line operators along spatial directions. We then discuss in Sec.~\ref{SEC:mixing} the transformation properties of such operators and their mixing pattern under lattice regularization. In Sec.~\ref{SEC:renmat} we discuss the renormalization and matching of quasi-TMDPDFs and -TMDWFs in a scheme following the spirit of hybrid renormalization and give the relevant one-loop matching kernel. Finally we conclude in Sec.~\ref{SEC:conclusion}. \section{Quasi-TMDPDFs and -TMDWFs in LaMET} \label{SEC:qTMD} Let us begin with the definition of quasi-TMDPDFs in LaMET with Euclidean metric in four-dimensions~\cite{Ji:2014hxa,Ji:2018hvs,Ebert:2019okf,Ji:2019ewn} \begin{align}\label{eq:quasi_TMD} & \tilde f(z ,b_\perp,\mu,P^z) \\ &=\! \lim_{L \rightarrow \infty} \frac{\langle PS| \bar \psi\big(\frac{\vec z+\vec{b}_\perp}{2}\big)\Gamma{\overline W}(\vec z, \vec{b}_\perp;\vec L)\psi\big(-\!\frac{\vec z+\vec{b}_\perp}{2}\big) |PS\rangle}{\sqrt{Z_E(2L,b_\perp,\mu)}} \ , \nonumber \end{align} where we have chosen a symmetric setup to simplify the analysis. $P=(P^0,0,0,P^z)$ is the hadron momentum and $S$ denotes its spin, $\vec L\equiv L n_z$, $\vec z=z n_z$ with $ n_z=(0,0,0,1)$ being a unit four-vector along the spatial $z$ direction, and $\vec{b}_\perp=(0,b_1,b_2,0)$. The staple-shaped Wilson line takes the following form \begin{align}\label{eq:staplez} {\overline W}(\vec z, \vec{b}_\perp;L)&={W_z^\dagger\Big(\vec L+\frac{\vec b_\perp}{2}; \frac{\vec z}{2}-\vec L\Big) W_{\perp}\Big(\vec L -\frac{{\vec b}_\perp}{2};\vec b_\perp\Big)\nonumber\\ &\times W_{z}\Big(-\frac{\vec z+\vec{b}_\perp}{2};\vec L+\frac{\vec z}{2}\Big), \nonumber\\ W_{i}(\eta;L)&= {\cal P}{\rm exp}\Big[-ig\int_{0}^{L} dt\, {n}_i\cdot A(\eta^\mu+t n_i^\mu)\Big], \end{align} for an illustration see Fig.~\ref{fig:stapleWL}. $\Gamma$ denotes a Dirac matrix. $\sqrt{Z_E(2L,b_\perp,\mu,0)}$ is the square root of the vacuum expectation value of a flat rectangular Euclidean Wilson-loop along the $n_z$ direction with length $2L$ and width $b_\perp$: \begin{align}\label{eq:Z_E} Z_E(2L,b_\perp,\mu)&=\frac{1}{N_c}{\rm Tr}\langle 0|{W_\perp^\dagger(-\vec\xi_-;-b_\perp)W_z^\dagger(\vec\xi_+;-2L)}\nonumber\\ &{\times W_{\perp}(\vec \xi_-; b_\perp) W_z(-\vec \xi_+;2L)}|0\rangle \, , \end{align} where \begin{align} \vec \xi_\pm = L \vec n_z \pm \frac{\vec b_\perp}{2}\, . \end{align} In contrast to the usual TMDPDF which contains lightlike separations between quark fields, the quasi-TMDPDF defined above involves spatial separations only. However, the same lightcone physics is projected out when the hadron momentum becomes infinite, as one can unboost the hadron at large momentum and apply the boost operator to the spatial correlator in Eq.~(\ref{eq:quasi_TMD}), yielding the same LF correlator defining the TMDPDF~\cite{Ji:2019ewn,Ji:2020ect}. This is similar to shifting from Schr\"odinger picture to Heisenberg picture in quantum mechanics. Note that the LF correlator in TMDPDF leads to rapidity divergences which require a proper regulator. Given the finite hadron momentum, the quasi-TMDPDF can be viewed in a sense as the definition of TMDPDF with the hadron momentum as a rapidity regulator~\cite{Ji:2019ewn}. \begin{figure}[tbp] \includegraphics[width=0.25\textwidth]{stapleWL} \caption{Staple-shaped gauge link used to define the quasi-TMDPDF and -TMDWF.} \label{fig:stapleWL} \end{figure} In the above definition, also the length of the longitudinal link is kept finite to regulate the pinch-pole singularity associated with infinitely long Wilson lines~\cite{Ji:2018hvs}. Such link length dependence drops out in the ratio of Eq.~(\ref{eq:quasi_TMD}) so that the final result has a proper $L\to\infty$ limit. The introduction of $Z_E$ also removes additional contributions arising from the transverse gauge link. From Eq.~(\ref{eq:quasi_TMD}), the momentum space density is given by the following Fourier transform \begin{equation} \label{eq:TMD-mom} \tilde f(x, k_\perp,\mu,\zeta_z) = \int\frac{d\lambda d^2\vec b_\perp}{(2\pi)^3}e^{ix\lambda+i\vec k_\perp\cdot \vec b_\perp}\tilde f(\lambda ,b_\perp,\mu,P^z)\, , \end{equation} with $\lambda=z P^z$ being the quasi-LF distance, and $\zeta_z=(2xP^z)^2$ is the Collins-Soper scale. The thus defined quasi-TMDPDF depends on two scales, $\mu$ and $\zeta_z$. The dependence on $\mu$ is controlled by the renormalization group equation~\cite{Collins:1981uk,Ji:2004wu} \begin{equation}\label{eq:RG_TMD} \mu^2\frac{d}{d\mu^2}\ln \tilde f(x, b_\perp,\mu,\zeta_z)=\gamma_F(\alpha_S(\mu)), \end{equation} where $\alpha_S=g^2/(4\pi)$, and $\gamma_F$ is most easily obtained from the anomalous dimension of the quark field in the axial gauge $A^z=0$. In the auxiliary field language~\cite{Dorn:1986dt,Ji:2017oey,Green:2017xeu}, a straight segment of Wilson line can be replaced by the two-point function of an auxiliary heavy quark field, $\gamma_F$ then represents the anomalous dimension of the auxiliary heavy-light quark current. The Wilson line cusp anomalous dimension does not enter because it has been canceled between the numerator and denominator in Eq.~(\ref{eq:quasi_TMD}). The $\zeta_z$ dependence characterizes how the quasi-TMDPDF changes with momentum or rapidity, and the evolution is controlled by the Collins-Soper equation~\cite{Collins:1981uk,Ji:2014hxa} \begin{equation}\label{eq:CS_TMD} P^z\frac{d}{dP^z}\ln \tilde f(x, b_\perp, \mu, \zeta_z)=K(b_\perp,\mu)+G(\zeta_z,\mu), \end{equation} where $K(b_\perp,\mu)$ is the Collins-Soper kernel that is independent of the rapidity regularization, while $G(\zeta_z,\mu)$ is a perturbative term existing only in the off-light-cone regularization scheme, its explicit expression at one-loop can be found in Ref.~\cite{Ji:2019ewn}. Analogously, one can define the quasi-TMDWF with the same staple-shaped Wilson line operator, but now between the vaccum and a hadron state~\cite{Ji:2020ect} \begin{align}\label{eq:qTMDWF} &\tilde\psi(z,b_\perp,\mu,P^z)\\ &=\! \lim_{L \rightarrow \infty} \frac{\langle 0| \bar \psi\big(\frac{\vec z+\vec{b}_\perp}{2}\big)\Gamma{\overline W}(\vec z, \vec{b}_\perp;\vec L)\psi\big(-\!\frac{\vec z+\vec{b}_\perp}{2}\big) |PS\rangle}{\sqrt{Z_E(2L,b_\perp,\mu)}}. \nonumber \end{align} Its scale dependence is controlled by evolution equations similar to Eqs.~(\ref{eq:RG_TMD}) and (\ref{eq:CS_TMD}). \section{Mixing pattern of staple-shaped Wilson line operators on the lattice} \label{SEC:mixing} To calculate the TMDPDFs or TMDWFs, we need to calculate the coordinate space correlation functions defined above on the lattice. A discretized lattice has less symmetry than the continuum, and thus more operator mixings can appear. Moreover, chiral symmetry might be broken after the fermion fields are discretized, leading to additional operator mixings. Nevertheless, the lattice action exhibits important discrete symmetries: parity, time reversal and charge conjugation. Investigating the transformation properties of relevant operators under these symmetries helps to unravel potential mixings that can occur. Such an analysis has been done for straight Wilson line operators defining the quasi-PDFs in Ref.~\cite{Chen:2017mie}. In this section, we extend it to staple-shaped Wilson line operators relevant for the quasi-TMDPDFs and -TMDWFs. \subsection{${\cal P}$, ${\cal T}$, ${\cal C}$ and axial transformations} For the convenience of the reader, we briefly summarize in this subsection the transformation properties of fields under parity (${\cal P}$), time-reversal (${\cal T}$), charge conjugation (${\cal C}$), and the axial transformation. We follow the convention of Ref.~\cite{Chen:2017mie} with the Euclidean spacetime coordinates $(x, y, z, \tau) = (1, 2, 3, 4)$. Dirac matrices are chosen to be Hermitian: $\gamma_{\mu}^{\dagger}=\gamma_{\mu}$, and $\gamma_5=\gamma_1\gamma_2\gamma_3\gamma_4$. Since there is no distinction between time and space in Euclidean space, the parity transformation in the $\mu$-direction, denoted by ${\cal P}_{\mu}$ with $\mu\in\{1,2,3,4\}$, can be defined with respect to any direction \begin{eqnarray} \psi(x)&\xrightarrow[]{{\cal P}_{\mu}}& \psi(x)^{{\cal P}_{\mu}}=\gamma_{\mu}\psi(\mathbb{P}_{\mu}(x)),\\ \overline{\psi}(x)&\xrightarrow[]{{\cal P}_{\mu}}& \overline{\psi}(x)^{{\cal P}_{\mu}}=\overline{\psi}(\mathbb{P}_{\mu}(x))\gamma_{\mu},\\ U_{\nu\not=\mu}(x)&\xrightarrow[]{{\cal P}_{\mu}}& {U_{\nu}(x)^{{\cal P}_{\mu}} =U_{{-\nu}}^{\dagger}(\mathbb{P}_{\mu}(x) ,\\ U_{\mu}(x)&\xrightarrow[]{{\cal P}_{\mu}}& U_{\mu}(x)^{{\cal P}_{\mu}}=U_{\mu}(\mathbb{P}_{\mu}(x)), \end{eqnarray} where $\mathbb{P}_{\mu}(x)$ is the vector $x$ with sign flipped except for the component in the $\mu$-direction. In other words, it is the parity transformation in the $x_\mu$ direction. $U_\mu(x)$ denotes a generic Wilson line along the $\mu$ direction with the starting point at $x$. Analogously, the time reversal transformation ${\cal T}_{\mu}$ can also be generalized in any direction in Euclidean space \begin{eqnarray} \psi(x)&\xrightarrow[]{{\cal T}_{\mu}}& \psi(x)^{{\cal T}_{\mu}}=\gamma_{\mu}\gamma_5\psi(\mathbb{T}_{\mu}(x)),\\ \overline{\psi}(x)&\xrightarrow[]{{\cal T}_{\mu}}& \overline{\psi}(x)^{{\cal T}_{\mu}}= \overline{\psi}(\mathbb{T}_{\mu}(x))\gamma_5\gamma_{\mu},\\ U_{\mu}(x)&\xrightarrow[]{{\cal T}_{\mu}}& U_{\mu}(x)^{{\cal T}_{\mu}}=U_{{-}\mu}^{\dagger}(\mathbb{T}_{\mu}({x} ),\\ U_{\nu\not=\mu}(x)&\xrightarrow[]{{\cal T}_{\mu}}& U_{{\nu}}(x)^{{\cal T}_{\mu}} =U_{{\nu}}(\mathbb{T}_{\mu}(x)), \end{eqnarray} where $\mathbb{T}_{\mu}(x)$ is the vector $x$ with sign flipped only in the $\mu$-direction. Charge conjugation ${\cal C}$ transforms particles into their antiparticle counterparts. Under charge conjugation, one has \begin{eqnarray} \psi(x)&\xrightarrow[]{\cal C}& \psi(x)^{\cal C}=C^{-1}\overline{\psi}(x)^{\top},\\ \overline{\psi}(x)&\xrightarrow[]{\cal C}& \overline{\psi}(x)^{\cal C}=-\psi(x)^{\top}C,\\ U_{\mu}(x)&\xrightarrow[]{\cal C}& U_{\mu}(x)^{\cal C}=U_{\mu}(x)^{\ast}=(U_{\mu}^{\dagger}(x))^{\top}, \end{eqnarray} with $\top$ denoting the transpose operation, and \begin{eqnarray} C\gamma_{\mu}C^{-1}=-\gamma_{\mu}^{\top}, \qquad C\gamma_5C^{-1}=\gamma_5^{\top}. \end{eqnarray} The continuous axial rotation $\cal A$ of the fermion field reads \begin{align} \psi(x)&\xrightarrow[]{\cal{A}}\psi'(x)=e^{i\alpha\gamma_5}\psi(x),\nonumber\\ \overline{\psi}(x)&\xrightarrow[]{\cal{A}}\overline{\psi}'(x) =\overline{\psi}(x)e^{i\alpha\gamma_5}. \label{eq:chitransf} \end{align} \subsection{Operator mixings} Based on transformation properties of the fields listed above, we can investigate the transformation under discrete symmetries of the following nonlocal operators involving a staple-shaped Wilson line \begin{align}\label{eq:O_Gamma} O_\Gamma(z, \vec b_\perp, L)&=\bar \psi\big(\frac{\vec z+\vec{b}_\perp}{2}\big)\Gamma{\overline W}(\vec z, \vec{b}_\perp;L)\psi\big(-\!\frac{\vec z+\vec{b}_\perp}{2}\big). \end{align} \iffalse \begin{widetext} \centering \hspace{4em}\begin{table}[t] \begin{tabular}{ccccccccc} \hline\hline & $\Gamma={\bf 1}_{1234}$ & $\gamma_{i,1234}$ & $\gamma_{3,1234}$ & $\gamma_{5,1234}$ & $i\gamma_i\gamma_{5,1234}$ & $i\gamma_3\gamma_{5,1234}$ & $\sigma_{i3,1234}$ & $\epsilon_{ijk}\sigma_{jk,1234}$\\ \hline ${\cal P}_3$ & EEEE & OOOO & EEEE & OOOO & EEEE & OOOO & OOOO & EEEE\\ ${\cal P}_{l\not=3}$ & EEOO & EEOO$_{(l=i)}$ & OOEE & OOEE & OOEE$_{(l=i)}$ & EEOO & OOEE$_{(l=i)}$ & EEOO$_{(l=i)}$\\ & & OOEE$_{(l\not=i)}$ & & & EEOO$_{(l\not=i)}$ & & EEOO$_{(l\not=i)}$ & OOEE$_{(l\not=i)}$\\ ${\cal T}_3$ & EEOO & EEOO & OOEE & OOEE & OOEE & EEOO & OOEE & EEOO\\ ${\cal T}_{l\not=3}$ & EEEE & OOOO$_{(l=i)}$ & EEEE & OOOO & EEEE$_{(l=i)}$ & OOOO & OOOO$_{(l=i)}$ & EEEE$_{(l=i)}$\\ & & EEEE$_{(l\not=i)}$ & & & OOOO$_{(l\not=i)}$ & & EEEE$_{(l\not=i)}$ & OOOO$_{(l\not=i)}$\\ ${\cal C}$ & EOOE & OEEO & OEEO & EOOE & EOOE & EOOE & OEEO & OEEO\\ ${\cal A}$ & V & I & I & V & I & I & V & V\\ \hline\hline \end{tabular} \caption{Transformation properties of the staple-shaped Wilson line operators $O_{\Gamma}^\alpha(z, \vec b_\perp, L)$, $i,j,k\not=3$. Other notations follow those of Ref.~\cite{Chen:2017mie}.} \label{tab:CPTAtransf} \end{table} \end{widetext} \fi Given that the hadron is to be boosted along the $z$-direction, we treat the $z$-direction differently from other directions, as was done in the case of straight Wilson line operators, and categorize the Dirac structure as follows \begin{eqnarray} \Gamma\in\{{\bf{1}}, ~\gamma_i, ~\gamma_3, ~\gamma_5, ~i\gamma_i\gamma_5, ~i\gamma_3\gamma_5, ~\sigma_{i3}, ~\epsilon_{ijk}\sigma_{jk}\}, \end{eqnarray} where $i, j, k\not=3$. From the field transformation properties in the previous subsection, one can work out with some effort the transformation properties of $O_{\Gamma}(z, \vec b_\perp, L)$ under discrete symmetries. \begin{align} O_{\Gamma}(z, \vec b_\perp, L)&\xrightarrow[]{{\cal P}_{i\not=3}} O_{\gamma_i\Gamma\gamma_i}(-z,{-}\vec b_\perp^{{(-)^i}}, -L), \nonumber\\ O_{\Gamma}(z,\vec b_\perp, L)&\xrightarrow[]{{\cal P}_3} O_{\gamma_3\Gamma\gamma_3}(z,-\vec b_\perp,L),\nonumber\\ O_{\Gamma}(z,\vec b_\perp, L)&\xrightarrow[]{{\cal T}_{i\not=3}} O_{\gamma_5\gamma_i\Gamma\gamma_i\gamma_5}(z,\vec b_\perp^{{(-)^i}}, L),\nonumber\\ O_{\Gamma}(z,\vec b_\perp, L)&\xrightarrow[]{{\cal T}_3} O_{\gamma_5\gamma_3\Gamma\gamma_3\gamma_5}(-z,\vec b_\perp, -L), \end{align} {where $\vec b_{\perp,j}^{(-)^i}\equiv (-1)^{\delta_{ij}} \vec b_{\perp,j}$ with $j=1,2$ labeling the component of the transverse vector $\vec b_\perp$. Here no summation is implied over index $j$.} Under $\cal C$, one has \begin{equation} O_{\Gamma}(z,\vec b_\perp, L) \xrightarrow[]{\cal C} O_{(C\Gamma C^{-1})^T}(-z,-\vec b_\perp, L). \end{equation} {We can, therefore, define the following combinations that are eigenstates under discrete symmetries \begin{align} {\cal O}^n_\Gamma(z, \vec b_1, L) & = \big\{\big[\big(O_\Gamma(z,\vec b, L)+s_{n0} O_\Gamma(z,b_1,-b_2, L)\big)\notag\\ &\,\,+s_{n1}\left(b_1\mapsto -b_1\right)\big]+s_{n2}\big[z\mapsto -z]\big\}\notag\\ &\quad+s_{n3}\big\{ L\mapsto -L \big\}\, , \end{align} where $s_{ij}=(-1)^{\floor{i/2^j}}$. ${\cal O}^n_\Gamma (z,\vec b,L)$ with $n=1,2,\cdots,16$ is a linear combination of 16 independent operators constructed from freely choosing the sign in front of each argument $z,b_1,b_2,L$. Here $\floor{x}$ is the floor function. {$O_{\Gamma}^n$ therefore form an operator basis with definite ${\cal C}, {\cal P}$, and ${\cal T}$ properties, and only operators with the same ${\cal C}, {\cal P}, {\cal T}$ eigenvalues can mix, making the mixing pattern manifest. Nevertheless, the $O_\Gamma^n$ basis given above is much more complicated than the original one without $-z$, $-b_\perp$, and $-L$ dependence. Thus, in the following we present the results for the operators $O_n$ defined in Eq.~\eqref{eq:O_Gamma} only, although our analysis is mainly based on the operator basis $O_{\Gamma}^n$. For the operators in~\eqref{eq:O_Gamma}, Lorentz covariance helps to identify what mixings are forbidden. } For example, the scalar operator $O_1$ does not mix with $O_{\gamma_5}$ and $O_{\gamma_3\gamma_5}$ for the TMDPDF with a staple-shape Wilson line. The same is true for $O_{\gamma_\mu}$ and $O_{\gamma_5}$, $O_{\gamma_3}$ and $O_{\gamma_3\gamma_5}$, and etc. These patterns are consistent with the observation in lattice computations \cite{Shanahan:2019zcq,Zhang:2020dbb} taking into account the external off-shell quark state. However, it is worth pointing out that the operator basis $O_\Gamma$'s are not complete by themselves, as they are not operators of definite twist and operators of higher twist mix with operators of higher Fock states with additional elementary fields due to QCD equations of motion. In fact, a proper treatment for the mixings between operators observed from one-loop calculation in~\cite{Ebert:2019tvc} requires the introduction of higher Fock states. This is already evident from operators with straight Wilson line at twist-3 level, see Ref.~\cite{Braun:2021aon} and references therein. Lorentz symmetry allows certain mixings to be identified as mixings with higher Fock states which are statistically suppressed in general. This explains the smallness of mixings between certain Dirac structures observed in~\cite{Shanahan:2019zcq}. A thorough investigation including operators of higher Fock state is beyond the scope of the present paper, and left to future work. On the other hand, we find that $O_\Gamma$ mixes in general with $O_{\gamma_3\Gamma}$ for arbitrary Dirac structure $\Gamma$, provided that a fermion action that does not preserve chiral symmetry is used. This is in contrast to operators with straight Wilson line, where such mixings do not occur for a specific set of $\Gamma$, e.g., $\Gamma=\gamma_4$ in the unpolarized case. The above mixing pattern contains the mixings observed in lattice perturbation theory calculations ~\cite{Constantinou:2019vyb}, but also contains mixings that are not present in such calculations. Note that the definition of quasi-TMDPDFs or -TMDWFs also involves a factor of $\sqrt {Z_E}$, but it does not change the mixing pattern discussed above, as it is a common factor for operators with all Dirac structures and depends only on the length $b_\perp$ and $L$. \section{Renormalization and matching of quasi-TMDPDFs and -TMDWFs in a simple scheme} \label{SEC:renmat} In this section, we discuss the renormalization and matching of quasi-TMDPDFs and -TMDWFs in a simple scheme that does not introduce extra non-perturbative effects at large distances, following the same spirit as the hybrid renormalization scheme introduced in Ref.~\cite{Ji:2020brr}. Using the auxiliary field formalism~\cite{Dorn:1986dt}, the straight Wilson line operators have been shown to be multiplicatively renormalized~\cite{Ji:2017oey,Green:2017xeu,Ishikawa:2017faj}. The same can be shown to be true for the staple-shaped operators, with the renormalization factors eliminating the power divergences associated with the Wilson line, the cusp divergences as well as the endpoint divergences~\cite{Ebert:2019tvc}. Thus, the renormalization of the quasi-TMDPDFs and quasi-TMDWFs can be carried out in analogy with that of the quasi-PDFs, and for the latter a commonly used renormalization scheme is the regularization-independent momentum subtraction (RI/MOM) scheme (or its variation $\rm{RI'/MOM}$)~\cite{Constantinou:2017sej,Stewart:2017tvs,Alexandrou:2017huk,Chen:2017mzz} or the ratio scheme~\cite{Radyushkin:2017cyf,Orginos:2017kos,Braun:2018brg,Li:2020xml}. These schemes have the advantage of avoiding certain discretization effects at short distances. The $\rm{RI'/MOM}$ renormalization for the quasi-TMDPDFs has been discussed in the literature~\cite{Constantinou:2019vyb,Ebert:2019tvc,Shanahan:2019zcq}. However, as pointed out in Ref.~\cite{Ji:2020brr}, both the RI/MOM and the ratio schemes introduce undesired non-perturbative effects at large $z$ in the renormalization stage, and thus become unreliable at large distances. This can be clearly seen in a recent analysis of data at multiple lattice spacings in Ref.~\cite{Huo:2021rpe}. In contrast, the Wilson line mass subtraction scheme~\cite{Chen:2016fxx,Ishikawa:2017faj} does not have this issue. Based on this, an alternative renormalization strategy, the hybrid renormalization, has been proposed in Ref.~\cite{Ji:2020brr}, which utilizes the advantages of different schemes at short and long distances. Another issue with the RI/MOM scheme is that it involves off-shell external states, which bring in a lot of complications when going to higher orders in perturbation theory~\cite{Chen:2020ody} or dealing with gauge particles such as gluons~\cite{Zhang:2018diq,Wang:2019tgg}. This can be avoided if one chooses physical matrix elements for renormalization. The discussion above indicates that for the quasi-TMDPDFs or -TMDWFs, one shall also switch to a more reliable renormalization scheme such as the hybrid scheme. Fortunately, the lattice study of quasi-TMDPDFs/-TMDWFs is focused on non-perturbative or large $b_\perp$ region (for small $b_\perp$ the TMDPDFs can be studied through a factorization into integrated PDFs, similar factorization is expected also for the TMDWFs). Thus, even at small longitudinal separation $z$ one does not need to worry about the discretization effects plaguing the quasi-PDFs. As a result, we can perform the renormalization in a simple manner by removing the logarithmic and linear divergences separately, in the same spirit as the Wilson line mass subtraction scheme~\cite{Chen:2016fxx,Ishikawa:2017faj}. Bearing in mind the mixing discussed in the previous section, we can write down the following renormalization \begin{align}\label{eq:renorm} \begin{pmatrix} {\bar O}^B_{\Gamma} \\ {\bar O}^B_{\Gamma'} \end{pmatrix} &= {\cal Z} \begin{pmatrix} {\bar O}_{\Gamma}^R \\ {\bar O}_{\Gamma'}^R \end{pmatrix} = \begin{pmatrix} {\cal Z}_{\Gamma\Gamma} & {\cal Z}_{\Gamma\Gamma'}\\ {\cal Z}_{\Gamma'\Gamma} & {\cal Z}_{\Gamma'\Gamma'} \end{pmatrix} \begin{pmatrix} {\bar O}_{\Gamma}^R \\ {\bar O}_{\Gamma'}^R \end{pmatrix} +{\rm{h.f.}} \nonumber\\ &= Z\begin{pmatrix} {Z}_{\Gamma\Gamma} & {Z}_{\Gamma\Gamma'}\\ {Z}_{\Gamma'\Gamma} & {Z}_{\Gamma'\Gamma'} \end{pmatrix} \begin{pmatrix} {\bar O}_{\Gamma}^R \\ {\bar O}_{\Gamma'}^R \end{pmatrix}+{\rm{h.f.}}\ , \end{align} where {${\rm h.f.}$ stands for higher Fock states which are an integral part of a complete operator basis. They are generated by terms that are proportional to the external quark momenta should an external quark state in momentum space is used in loop calculations. The investigation of their contributions is beyond the scope of this paper and left for future work.} The superscript $B$ and $R$ denotes bare and renormalized operators, respectively, $\Gamma'=\gamma_3\Gamma$, and we assume the factor $\sqrt {Z_E}$ has been divided in all $\bar O_\Gamma$'s, and ignore the arguments for notational simplicity. In the second row, we have separated an overall renormalization factor $Z$ because the multiplcative renormalization is independent of the Dirac structure involved in the operator~\cite{Ji:2017oey,Musch:2010ka}. Moreover, the linear and cusp divergences associated with the Wilson line are canceled by $\sqrt {Z_E}$, thus one only needs to renormalize the remaining logarithmic divergences at the endpoints of the operator. If the operator mixing is absent, there is an easy way to achieve this. One can calculate the straight line operator matrix elements at two different distances $z_0$ and ${z_0}/2$ (with $z_0$ being in the perturbative region) and form the ratio which effectively removes the factorized linear divergence ${\rm e}^{\delta m |z|}$ from the self-energy of the Wilson-line while still retains the required logarithmic divergence, \begin{align}\label{eq:renormZ} Z&=\frac{\tilde h_\Gamma^2(z_0/2, P^z)}{\tilde h_\Gamma(z_0, P^z)}\, , \end{align} where $\tilde h_\Gamma$ can be chosen, e.g., as the zero-momentum hadron matrix element used in the ratio scheme or the off-shell quark matrix element used in the RI/MOM scheme, and $P^z$ is not necessarily the same as the momentum used in calculating the quasi-TMDPDF matrix element. For example, in the unpolarized case, one can use \begin{align} &\tilde h_{\gamma^t}(z_0,P^z=0)\\ &=\frac{1}{2P^t}\langle P^z=0|\bar\psi(z_0)\gamma^t W_z(z_0,0)\psi(0)|P^z=0\rangle, \nonumber \end{align} where $P$ denotes the hardon momentum, or \begin{align} &\tilde h_{\gamma^t}(z_0,p^z=0)\nonumber\\ &=\left.\frac{\sum_s\langle p,s|O_{\gamma^t}(z_0)|p,s\rangle}{\sum_s\langle p,s|O_{\gamma^t}(z_0)|p,s\rangle_{\rm tree}}\right|_{\tiny{p^2=-\mu_R^2,\,p^z=0}}\, , \end{align} with $p,s$ being the off-shell quark momentum and spin, respectively. Note that although the RI/MOM matrix element exhibits a non-universal linear divergence behavior depending on the lattice action used~\cite{Huo:2021rpe}, it is still allowed here because in $Z$ all such linear divergences cancel out by construction. In this way, what is left in $Z$ is just the endpoint renormalization factors associated with some perturbative corrections. However, in the presence of operator mixing, one needs to determine the mixing matrix, which requires calculating certain quasi-TMDPDF matrix elements. One can choose, for example, the hadron matrix element of the quasi-TMDPDF operator or the RI/MOM renormalization factor, but at perturbative $z$ and $b_\perp$, so that extra non-perturbative effects are avoided at the renormalization stage. The first option of determining the mixing matrix entries is to follow the calculation of renormalization factors in the presence of mixing in the RI/MOM scheme~\cite{Constantinou:2017sej}. Since all $z$- and $b_\perp$-dependent UV divergences have been canceled in $\bar O_\Gamma$, we prefer to choose renormalization conditions such that $Z_{ij}$ are independent of $z$ and $b_\perp$. In other words, we can still apply the RI/MOM renormalization conditions similar to those in~\cite{Constantinou:2019vyb,Shanahan:2019zcq}, but only at a given perturbative $z$ and $b_\perp$, and use the results for the renormalization of correlators at all distances. In other words, we may require \begin{align}\label{eq:RIMOMnew} &Z_q^{-1}(p) {\cal {\bar Z}}_{\Gamma\Gamma'}{\rm Tr}[\Lambda_{{\bar O}_\Gamma}(p)\Gamma']|_{\tiny{z=z_0,b_\perp=b_{\perp 0},p^\mu=p_0^\mu}} \nonumber\\ &={\rm Tr}[\Lambda_{{\bar O}_\Gamma}^{\rm tree}(p)\Gamma']|_{\tiny{z=z_0,b_\perp=b_{\perp 0},p^\mu=p_0^\mu}}, \end{align} where ${\cal {\bar Z}}={\cal Z}^{-1}$, $z_0, b_{\perp0}$ are chosen within the perburbative region, $p_0^\mu=(p_0,0,0,0)$, $\Lambda_{{\bar O}_\Gamma}$ is the amputated Green's function of the operator ${\bar O}_\Gamma$ in an off-shell quark state, and the superscript ``tree" denotes its tree-level value. $Z_q$ is the quark wave function renormalization factor determined as \begin{equation} Z_q(p)=\frac{1}{12}{\rm Tr}[S^{-1}(p)S^{\rm tree}(p)], \end{equation} with $S(p)$ and $S^{\rm tree}(p)$ denoting the quark propagator and its tree-level value, respectively. From Eq.~(\ref{eq:renorm}) and the renormalization factors calculated in Eq.~(\ref{eq:RIMOMnew}), one obtains the renormalized quasi-TMDPDF in the RI/MOM scheme \begin{align} \hspace{-1em}{\tilde f}_R(\Gamma,z, b_\perp)&= {{\cal {\bar Z}}}_{\Gamma\Gamma}\tilde f_B(\Gamma,z,b_\perp)+{{\cal {\bar Z}}}_{\Gamma\Gamma'}\tilde f_B(\Gamma',z,b_\perp), \end{align} which can be converted to the $\overline{\rm MS}$ scheme by applying a conversion factor \begin{align} {\tilde f}_R^{\overline{\rm MS}}(\Gamma,z,b_\perp)={\tilde C}{\tilde f}_R(\Gamma,z,b_\perp). \end{align} The conversion factor in general can take a diagonal form~\cite{Constantinou:2019vyb} or a non-diagonal form~\cite{Ebert:2019tvc}, depending on the choice of projectors. Here we choose to define a diagonal conversion factor which reads for $\Gamma=\gamma^\mu$ \begin{align} \tilde C_\Gamma&=1+\Big[\frac{1}{6}{\cal V}_\Gamma^{\mu\mu}(\mu,p_0)-Z_q^{(1)}\Big], \end{align} where $V^{\mu\mu}_\Gamma$ has been calculated in Ref.~\cite{Ebert:2019tvc}, and no summation is implied over the index $\mu$. Similar renormalization conditions can be constructed if the quasi-TMDPDF matrix elements of hadrons are used. There are different ways to choose such conditions. We choose to determine the renormalization factors by requiring that the renormalized quasi-TMDPDF matrix element be equal to the tree-level value at given perturbative $z$ and $b_\perp$. To avoid potential collinear divergences in the perturbative result and thus in the renormalization factor, we can set $z=0$ or $P^z=0$. In the following, we illustrate this procedure by taking $\Gamma=\gamma^\mu$ (with $\Gamma'=\gamma_3\gamma^\mu$) as an example, we require that the renormalization factors satisfy \begin{align} &\langle P|{\bar O}_\Gamma^B|P\rangle_{\rm tree}=\big[{\cal {\bar Z}}_{\Gamma\Gamma}\langle P|{\bar O}_\Gamma^B|P\rangle\nonumber\\ &+{\cal {\bar Z}}_{\Gamma\Gamma'}\langle P|{\bar O}_{\Gamma'}^B|P\rangle\big]\big|_{z=z_0,b_\perp=b_{\perp 0},P^\mu=p_0^\mu}, \end{align} where $P_0^\mu=(p^0,0,0,p^z)$. The corresponding conversion factor can be computed with dimensional regularization in the continuum, which also takes a diagonal form. For $z_0=0$, we have \begin{align} {\tilde C}_\Gamma=1-\frac{\alpha_S C_F}{2\pi}\Big(\frac{1}{2}L_b^2-\frac{3}{2}L_b+L_b L_z+\frac{1}{2}L_z^2-L_z+\frac{3}{2}\Big), \end{align} with $L_b=\ln(b_\perp^2\mu^2 e^{2\gamma_E}/4)$, $L_z=\ln(4p_z^2/\mu^2)$. While for $p^z=0$, we have \begin{align} {\tilde C}_\Gamma=1+\frac{\alpha_S C_F}{4\pi}\Big(L_b+4\ln\frac{b_\perp^2+z^2}{b_\perp^2}-4\frac{z}{b_\perp}\arctan\frac{z}{b_\perp}+1\Big). \end{align} With the renormalization and conversion factors above, we can convert the renormalized quasi-TMDPDF to the $\overline{\rm MS}$ scheme, and then match it to the TMDPDF. Following Refs.~\cite{Ji:2019ewn,Ji:2020ect}, the connection between the $\overline{\rm MS}$ scheme quasi-TMDPDF $\tilde f$ and TMDPDF $f^{\rm TMD}$ takes the following form, \begin{align}\label{eq:quasiTMDmatch} &f^{\rm TMD}(x,b_\perp,\mu,\zeta)=C\left(\frac{\zeta_z}{\mu^2}\right)\\ &\times e^{-\frac{1}{2}\ln (\frac{\zeta_z}{\zeta})K(b_\perp,\mu)} {\tilde f}(x,b_\perp,\mu,\zeta_z)S_r^{\frac{1}{2}}(b_\perp,\mu)+...\ ,\nonumber \end{align} where $C(\zeta_z/\mu^2)$ is a perturbative matching kernel and its explicit expression to $O(\alpha_S)$ can be found in Ref.~\cite{Ji:2020ect}. The exponential term contains the Collins-Soper evolution kernel which can be computed from the ratio of $\tilde f$'s at different rapidity scales. $\zeta$ corresponds to the Collins-Soper scale characterizd by the full hadron momentum. $S_r$ is the so-called reduced soft factor. The omitted terms are power corrections of order ${\cal O}\left(\Lambda_{\rm QCD}^2/\zeta_z,M^2/(P^z)^2,1/(b^2_\perp\zeta_z) \right)$. Similar matching relations have also been discussed in Refs.~\cite{Constantinou:2019vyb,Ebert:2019tvc}. For completeness, let us also summarize here how the remaining terms in Eq.~(\ref{eq:quasiTMDmatch}) can be calculated. The Collins-Soper kernel can be extracted by forming the ratio of quasi-TMDPDFs at two different $\zeta_z$'s~\cite{Ebert:2019tvc} \begin{align} \frac{\tilde f_R(x, b_\perp, \mu, \zeta_{z})}{\tilde f_R(x, b_\perp, \mu, \zeta'_{z})}=\frac{C(\frac{\zeta'_z}{\mu^2})}{C(\frac{\zeta_z}{\mu^2})}\Big(\frac{\zeta_z}{\zeta'_z}\Big)^{K(b_\perp,\mu)/2}\, . \end{align} As for the soft function $S_r(b_\perp,\mu)$, it is calculable from the form factor of a pseudoscalar light-meson state with quark content $\bar\psi\eta$ defined as~\cite{Ji:2020ect,Zhang:2020dbb} \begin{align} F(b_\perp, P, P', \mu)=\langle P'| \bar\eta(\vec b_\perp)\Gamma'\eta(\vec b_\perp)\bar\psi(0)\Gamma\psi(0)|P\rangle, \end{align} where $P,P'$ are two large momenta approaching opposite light-like directions. Making use of the quasi-TMDWFs $\tilde\psi_{\bar q q}$ in Eq.~\eqref{eq:qTMDWF} with light quark state $q$, the form factor takes the following factorized form, \begin{align} &F(b_\perp, P, P', \mu)=\nonumber\\ &\int dx\, dx'\, H(x, x'){\tilde\psi}^\dagger_{\bar q q}(x', b_\perp){\tilde\psi}_{\bar q q}(x, b_\perp)S_r(b_\perp, \mu), \end{align} where the perturbative matching kernel up to one-loop has been given in Ref.~\cite{Ji:2020ect}. The non-perturbative renormalization and conversion factors also apply to the quasi-TMDWFs. After converting to the $\overline{\rm MS}$ scheme quasi-TMDWFs, one can use the matching derived in Ref.~\cite{Ji:2020ect} to obtain the TMDWFs, where the matching relation takes the same form as Eq.~(\ref{eq:quasiTMDmatch}) with a different matching kernel, whose result at one-loop has also been given in Ref.~\cite{Ji:2020ect}. \section{Conclusion and outlook} \label{SEC:conclusion} In this paper, we have investigated the mixing patterns of staple-shaped Wilson line operators defining the quasi-TMDPDFs and -TMDWFs under lattice regularization using symmetry considerations. Our analysis shows that for non-chiral fermions the mixing with other Dirac structures is generally allowed, except for certain specific cases such as $O_1 \mathrel{\ooalign{$\leftrightarrow$\cr\hidewidth$/$\hidewidth}} O_{\gamma_5}$, etc. There is no Dirac structure, however, that is \textit{completely} free from mixing with other structures. {To be more specific, all $\Gamma$s mix with $\gamma^3\Gamma$.} This is in contrast with the case of straight Wilson line operators, where for certain choice of $\Gamma$, no mixing occurs. {We emphasize that we have excluded operator mixings with higher Fock states, which in itself is not self-consistent as the evolution of higher-twist two-particle TMDPDFs are not autonomous. A complete treatment is, however, beyond the scope of this paper and calls for further investigations.} We have also discussed the renormalization of quasi-TMDPDFs and -TMDWFs in a simple scheme that does not introduce extra non-perturbative effects at large distances, and presented the relevant one-loop matching. The results will facilitate the numerical calculation of TMDPDFs and TMDWFs on the lattice. It is worth pointing out that the operator mixing analysis presented here is based on transformation properties of the relevant operators under discrete symmetries, chiral symmetry and Euclidean symmetry only. For a more thorough analysis of the operator mixing pattern, it might be convenient to start from the auxiliary field formalism and replace the nonlocal operators with local ones in the framework of lattice field theory, and study the mixing of the former from that of the latter. This could be important not only for the staple-shaped Wilson line operators but also for the ones with a straight line, given that a different linear divergence behavior has been observed on lattice in the RI/MOM matrix elements from that of hadron matrix elements, which needs to be understood. \vspace{2em} \acknowledgments We thank Yizhuang Liu, Wei Wang, Yibo Yang and Yong Zhao for valuable discussions. YJ is supported by the DFG grant SFB TRR 257. JHZ is supported in part by National Natural Science Foundation of China under grant No. 11975051, No. 12061131006, and by the Fundamental Research Funds for the Central Universities. SZ is supported by Jefferson Science Associates, LLC under U.S. DOE Contract \#DE-AC05-06OR23177 and by U.S. DOE Grant \#DE-FG02-97ER41028. RZ is supported in part by National Natural Science Foundation of China under grant No. 12075124. \bibliographystyle{apsrev}

75090a93613fdae651e6ec9bb7dae87f204d3cbc

\section{Introduction} \label{sec:intro} Interpretability has long been a desirable property of machine learning (ML) models. With the success of complex models like neural networks, and their expanding reach into high-stakes and decision-critical applications, explaining ML models' predictions has become even more important. Interpretability can enable model debugging and lead to more robust ML systems, support knowledge discovery, and boost trust~\cite{HHB20}. It can also help to mitigate unfairness by surfacing undesirable model behavior~\cite{TCHL18, dodge19explaining}, lead to increased accountability by enabling auditing~\cite{SB18}, and enable ML practitioners to better communicate model behavior to stakeholders~\cite{BVV+18,HHB20}.\looseness=-1 There are two primary techniques for achieving interpretability of ML models. The first is to train transparent, or \emph{glass-box}, models that are intended to be inherently interpretable, such as decision trees \citep{quinlan1986induction} and sets \citep{lakkaraju2016interpretable}, simple point systems~\cite{zeng2017interpretable,jung2020simple}, and generalized additive models~\cite{hastie1990generalized,Caruana2015-qf}. Although some researchers have argued that glass-box models should always be used in high-stakes scenarios~\cite{rudin2018please}, complex \emph{black-box} models, such as neural networks, random forests, and ensemble methods, are very widely used in practice. As a result, other ML researchers have gravitated towards interpretability methods that generate post-hoc \emph{local} explanations for individual predictions produced by~such models~\citep[e.g.,][]{simonyan2013deep, selvaraju2017grad, ribeiro2016why, lundberg2017unified, alvarez-melis2017causal}.\looseness=-1 Local explanations aim to answer the question of why a model $\mathcal{M}$ predicted a particular output $y$ for some input $x$. There are many ways of operationalizing this abstract question, but most methods \edit{do so by addressing} the proxy question of how much the value of each input feature $x_i$ contributed to the prediction $y$. Thus, in practice, the explanations generated by many such methods consist of importance scores indicating the positive or negative relevance of each input feature $x_i$. Although the way these scores are computed varies from method to method, most start from an axiomatic or algorithmic derivation of some notion of feature importance, and only later investigate whether the resulting explanations are useful to humans. Some methods forgo this last step altogether, relying exclusively on intrinsic evaluation of mathematical properties of explanations, such as robustness~or faithfulness to~the~underlying model~\citep{alvarez-melis2018robustness, alvarez-melis2018towards, jacovi2020towards}.\looseness=-1 Interpretability, however, is fundamentally a human-centered concept. In \edit{light} of this, we put human needs at the center of both the design and evaluation of interpretability methods. Our work builds upon and weaves together two literatures that study the relationship between humans and explanations. First, researchers in philosophy, cognitive science, and the social sciences have long studied what it means to explain, and how humans do it \citep[e.g.,][and references therein]{pitt1988theories,miller2019explanation}. Second, a recent line of work within the human--computer interaction community has focused on how humans understand and utilize interpretability tools (i.e., software implementations of interpretability methods)~\cite{LD11,BLL12,BSO15, hohman2019gamut,LCH+19,AvdW+20,PG+21}, where a common finding is that practitioners misunderstand, over-trust,~and~misuse~these tools \citep[e.g.,][]{kaur2020interpreting}.\looseness=-1 Inspired by \citet{miller2019explanation}, we start by surveying the literature on the nature of explanation, revealing recurring characteristics of human explanation that are often missing from interpretability methods. We distill these characteristics into design principles that we argue human-centered machine-generated explanations should satisfy. We then realize our design principles \edit{using} the concept of \textit{weight of evidence} (WoE) from information theory \citep{good1985weight}, which has recently been advocated for by \citet{spiegelhalter2018neurips}, but, to the best of our knowledge, has yet to be investigated in the context of interpretability. We demonstrate that WoE can be adapted to handle high-dimensional, multi-class settings, yielding a suitable theoretical foundation for interpretability. We provide a general, customizable meta-algorithm to generate explanations for black-box models. We also show experimentally that WoE can be estimated from finite samples and is robust to small perturbations of the inputs.\looseness=-1 Evaluation of interpretability methods is notoriously difficult ~\citep{doshi-velez2017towards,kaur2020interpreting}. Although recent work has focused on abstract, intrinsic metrics such as robustness or faithfulness to the underlying model~\citep{alvarez-melis2018robustness}, considerably less attention has been given to understanding how the resulting explanations are used in practice. This discrepancy between the intended use of the explanations\,---\,by a human, for a specific goal such as auditing, debugging, or building trust in a model\,---\,and their experimental evaluation\,---\,\edit{typically performed using} abstract\edit{, intrinsic} metrics, in generic settings\,---\,hampers understanding of the benefits~and failure points of different \edit{interpretability} methods.\looseness=-1 We build on a recent thread of work~\cite[e.g.,][]{LCH+19,nourani2019effects,li2020understanding,kaur2020interpreting,VW21,PG+21}\,---\,including several recent papers from the human computation community\,---\,that argues that evaluations should be grounded in concrete use cases and should put humans at the center, taking into account not only how they use interpretability tools, but how well they understand the principles behind them. We carry out an artifact-based interview study with ten ML practitioners to investigate their use of a tool implementing our meta-algorithm in the context of a practical task. Qualitative themes from this study suggest that most participants successfully used the tool to answer questions, despite struggling with background concepts like prior class probabilities. Although the study was designed to identify preferences for different tool modalities, participants often used all of them and requested the option to switch between them interactively. Our results additionally highlight the importance of providing well-designed tutorials for interpretability tools\,---\,even for experienced ML practitioners\,---\,which are often overlooked in the literature \edit{on interpretability methods}, and which we argue should be an integral part of any interpretability tool.\looseness=-1 \section{Human-Centered Design Principles}\label{sec:design_principles} What it means to explain and how humans do it have long been studied in philosophy, cognitive science, and the social sciences. We draw on this literature to propose human-centered design principles for interpretability methods. \citet{hempel1948studies} and \citet{vanfraassen1988pragmatic} define an explanation as consisting of two main pieces: the \textit{explanandum}, a description of the phenomenon to be explained, and the \textit{explanans}, the facts or propositions that explain the phenomenon, which may rely on relevant aspects of context. As is often done colloquially, we will refer to the explanans as the \emph{explanation}. Different ways of formalizing the explanation have given rise to various theories, ranging from logical deterministic propositions \citep{hempel1948studies} to probabilistic ones \citep{salmon1971statistical, vanfraassen1988pragmatic}. An excellent historical overview can be~found in the surveys by~\citet{pitt1988theories} and~\citet{miller2019explanation}. In the context of local explanations for predictions made by ML models, the phenomenon to be explained is why a model $\mathcal{M}$ predicted output $y$ for input $x$. This why-question can be operationalized in different ways. The facts used to explain this phenomenon may include information about the input features, the model parameters, the data used to train the model, or the manner in which the model was trained. Although the nature of explanation is far from settled, recurring themes emerge across disciplines. At the core of the theories by \citet{vanfraassen1988pragmatic} and \citet{lipton1990contrastive} is the hypothesis that humans tend to explain in contrastive terms (e.g., ``a fever is more consistent with the flu than with a cold''), with explanations that are both factual and counterfactual (e.g., ``had the patient had chest pressure too, the diagnosis would instead have been bronchitis''). Yet, the explanations produced by most current interpretability methods refer only to why the input $x$ points to a single hypothesis (i.e., the prediction $y$) rather than ruling out all alternatives.\footnote{Exceptions include recent work advocating for contrastive or counterfactual explanations~\citep{wachter2017counterfactual, miller2019contrastive, vanderwaa2018contrastive}, partly inspired by contrast sets \citep{azevedo2010rules, bay1999detecting, webb2003detecting, novak2009supervised}.} In light of this, we propose our first two design principles: \begin{enumerate}[topsep=2pt, leftmargin=1.5em,noitemsep \item \textbf{Explanations should be contrastive}, i.e., explicate why the model predicted $y$ instead of alternative $y'$. \item \textbf{Explanations should be exhaustive}, i.e., provide a justification for why every alternative $y'$ was not predicted. \end{enumerate} Another theme, featured prominently by \citet{hempel1962deductive}, is that human explanations decompose into simple components. In other words, humans usually explain using multiple simple accumulative statements, each addressing a few aspects of the evidence (e.g., ``a fever rules out a cold in favor of bronchitis or pneumonia; among these, chills suggest the latter''). Each component is intended to be understood without further decomposition. Again, this contrasts with current interpretability methods that explain in one shot, for example, by providing importance scores for all features simultaneously. Our next two design principles are therefore:\looseness=-1 \begin{enumerate}[topsep=2pt, leftmargin=1.5em,noitemsep \setcounter{enumi}{2} \item \textbf{Explanations should be modular and compositional}, breaking up predictions into simple components \item \textbf{Explanations should rely on easily-understandable quantities}, so that each component is understandable. \end{enumerate} Another recurring theme is minimality. In a survey of over 250 papers, \citet{miller2019explanation} argued that it is important, but underappreciated in ML, that only the most relevant facts be included in explanations. Thus, our final principle is:\looseness=-1 \begin{enumerate}[topsep=2pt, leftmargin=1.5em,noitemsep \setcounter{enumi}{4} \item \textbf{Explanations should be parsimonious}, i.e., only the most relevant facts should be provided as components. \end{enumerate} These design principles are not exhaustive; each could be refined or generalized, and other principles could be derived from the same literature. However, we posit that these principles provide a reasonable starting point because they capture some of the most apparent discrepancies between human and machine-generated explanations. More generally, these principles point to a broader theme of human explanations as a \textit{process} rather than (only) a \textit{product} \citep{miller2019explanation, lombrozo2012explanation}. Therefore, these principles work to shift interpretability methods from the latter towards the former. \section{Explaining with the Weight of Evidence}\label{sec:woe} The set of design principles proposed in the previous section outlines a framework for human-centered interpretability in ML. In this section, we show how this framework can be operationalized by means of the \emph{weight of evidence}, a simple but powerful concept from information theory. We operationalize the question of why model $\mathcal{M}$ predicted output $y$ for input $x$ in terms of how much \emph{evidence} each input feature $x_i$ (or feature group) provides in favor of $y$ relative to alternatives. An explanation based on this question adheres to our design principles because it is based on a familiar concept (evidence) that is grounded in common language, it naturally evokes a contrastive statement (evidence \textit{for} or \textit{against} something), and, as we explain below, it can be formalized using simple quantities that admit modularity.\looseness=-1 \label{sec:background} \subsection{Weight of Evidence: Foundations}\label{sec:woe_foundations} The weight of evidence (WoE) is a well-studied probabilistic approach for analyzing variable importance that traces its origins back to \citet{pierce1878probability}, but was popularized by \citet{good1950probability, good1968corroboration, good1985weight}, whose definition and notation we follow here. Given a hypothesis and some evidence, the WoE seeks to answer the following question: \textit{"How much does the~evidence speak in favor of or against the hypothesis?"}\looseness=-1 The \edit{WoE} is usually \edit{defined} for some evidence $e$, a hypothesis $h$, and its logical complement $\overline{h}$. For example, in a simple binary classification setting\edit{,} $e=(X_1, \dots, X_n)$, $h: Y=1$, and $\overline{h}: Y=0$. The WoE of $e$ in favor of $h$ is the log-odds ratio between $h$ conditioned on $e$ and $h$ marginally:\looseness=-1 \begin{equation}\label{eq:woe_def} \text{woe}(h : e) \triangleq \log \frac{O( h \mid e) }{O(h)}, \end{equation} where $O(\cdot)$ denotes the odds of a hypothesis, i.e., \begin{equation} O(h) \triangleq \frac{P(h)}{P(\overline{h})}\qquad \text{and} \qquad O( h \mid e) \triangleq\frac{P(h \mid e)}{P(\overline{h} \mid e)}. \end{equation} Using Bayes' rule, $\text{woe}(h : e)$ can also be defined as \begin{equation}\label{eq:woe_alter_def} \text{woe}(h : e) \triangleq \log \frac{P(e \mid h)}{P(e\mid \overline{h})}. \end{equation} These two equivalent definitions provide complementary views of the WoE: the \textit{hypothesis-odds} and \textit{evidence-likelihood} interpretations. Using Equation~\eqref{eq:woe_def}, $\text{woe}(h:e) > 0$ indicates that the odds of $h$ are higher under $e$ than marginally. Equivalently, using Equation~\eqref{eq:woe_alter_def}, it indicates that the likelihood of $e$ is larger when conditioning on $h$ than on its complement. In other words, the evidence \textit{speaks in favor of} hypothesis $h$. Analogously, if $\text{woe}(h:e) < 0 $ we would say that the evidence \textit{speaks against} $h$. The quantities in Equations~\eqref{eq:woe_def} and \eqref{eq:woe_alter_def} are contrastive (cf.~Principle 1)\,---\,that is, defined in terms of ratios.\looseness=-1 As a concrete example, suppose that a doctor wants to know whether a patient's symptoms indicate the presence of a certain disease, say, the flu. Denote $e=$ ``\textit{the patient has a fever},'' $h=$ ``\textit{the patient has the flu},'' and $\bar{h}=$ ``\textit{the patient doesn't have the flu}.'' The doctor might know that for a patient, the odds of having the flu roughly double once the patient's fever is taken into account (i.e., the hypothesis-odds interpretation), which corresponds to $\text{woe}(h:e) \approx \log 2$. Alternatively, using the evidence-likelihood interpretation, the doctor could conclude that a patient is twice as likely to have a fever if they have the flu compared to when they do not. Note that neither interpretation tells us anything about the base rate of the flu.\looseness=-1 The WoE generalizes beyond these simple scenarios. For example, it can be conditioned on additional information $c$: \[ \text{woe}(h : e \mid c) \triangleq \log \frac{P( e \mid h, c) }{P(e \mid \overline{h}, c)} . \] It can also contrast $h$ to an arbitrary alternative hypothesis $h'$ instead of $\bar{h}$ (e.g., evidence in favor of \edit{the flu} and against \edit{a cold}): $\text{woe}(h/h': e) \triangleq \text{woe}(h: e \mid h \lor h')$. Thus, we can, in general, talk about the strength of evidence in favor of $h$ and against $h'$ provided by $e$ (perhaps conditioned on $c$).\looseness=-1 When the evidence is decomposable into multiple parts\,---\,that is, when $e= \bigcup_{i=1}^n e_i$\,---\,the WoE is also decomposable: \begin{equation}\label{eq:woe_multievidence} \text{woe}(h/h': e) = \sum_{i=1}^n \log \frac{P(e_i \mid e_{i-1}, \dots, e_1, h)}{P(e_i \mid e_{i-1}, \dots, e_1, h')} . \end{equation} This is crucial to defining an extension of the WoE to high-dimensional inputs that adheres to Principle 3 (modularity). A further appealing aspect of the WoE is its immediate connection to Bayes' rule through the following identity:\looseness=-1 \begin{equation}\label{eq:woe_prior_posterior} \underbrace{\log \frac{P(h \mid e)}{P(h' \mid e)}}_{\text{Posterior log odds}} =\underbrace{\log\frac{P(h)}{P(h')}}_{\text{Prior log odds}} + \underbrace{\log \frac{P(e \mid h)}{P(e \mid h')}}_{\text{Weight of evidence}}. \end{equation} In other words, the WoE can be understood as an adjustment \edit{to} the prior log odds caused by observing the evidence. \edit{In~a~simple binary classification setting, this amounts to} \begin{equation*} \log \frac{P(Y\!=\!1 \mid\! X)}{P(Y\!=\!0 \mid \! X)} = \log\frac{P(Y\!=\!1)}{P(Y\!=\!0)} + \text{woe}(Y=1 : X), \end{equation*} which shows that a positive (respectively, negative) WoE implies that the posterior log odds of $Y=1$ versus~$Y=0$ are higher (lower) than the prior log odds, indicating that the evidence~makes $Y=1$ more (less) likely than it was \textit{a priori}. Equation~\eqref{eq:woe_prior_posterior} shows that the WoE is modular (cf.~Principle 3) in another important way: it disentangles prior class probabilities and input likelihoods. This is important because of the \textit{base rate fallacy} studied in the behavioral science literature \citep{tversky1974judgment, bar-hillel1980base, eddy1982probabilistic, koehler1996base}. This cognitive bias, prevalent even among domain experts, is characterized by a frequent misinterpretation of posterior probabilities, primarily caused by a neglect of base rates (i.e., prior probabilities). Despite this, many interpretability methods do not explicitly display prior probabilities, and even when they do, they focus on explaining posterior probabilities, which invariably entangle~information about priors and the input being explained.\looseness=-1 Additionally, the units in which the WoE is expressed (log-odds ratios) are arguably easily understandable (cf.~Principle 4). There is evidence from the cognitive-neuroscience literature that log odds are a natural unit in human cognition. For example\edit{,} degrees of confidence expressed by humans are proportional to log odds \citep{peirce1885small}\edit{,} people are less biased when responding in log odds that in linear scales \citep{phillips1966conservatism}, and there exist plausible neurological hypotheses for encoding of log odds in the human brain \citep{gold2001neural, gold2002banburisms}. We refer the reader to \citet{Zhang2012ubiquitous} for a meta-analysis of these various studies of log odds.\looseness=-1 We provide \edit{additional} properties of the WoE, along with an axiomatic derivation\edit{,} in the appendix, \edit{which can be found in the longer version of this paper, available online.}\footnote{\url{https://arxiv.org/abs/2104.13299}}\looseness=-1 \subsection{Composite Hypotheses and Evidence}\label{sec:woe_extension} Traditionally, the WoE has been mostly used in simple settings, such as a single binary output and only a few input features. Its use in the more complex settings typically considered in modern ML therefore poses new challenges.\looseness=-1 The first such challenge is that in multi-class \edit{settings}, there is flexibility in choosing the hypotheses $h$ and $h'$ to contrast. The obvious choice of letting $h$ correspond to the predicted class $y^*$ and $h'$ its complement is unlikely to yield useful explanations when the number of classes is large (e.g., explaining the evidence in favor of one disease against one hundred thousand others). Following Principle 3 (modularity), and taking inspiration from Hempel's model \citeyearpar{hempel1962deductive} and the view of explanation as a process \citep{lombrozo2012explanation, miller2019explanation}, we address this by casting explanation as a sequential procedure, whereby a subset of \edit{the} possible classes is ruled out at each step. For example, in medical diagnosis, we might first explain why bacterial diseases were ruled out in favor of viral ones, and then explain why a specific viral disease was predicted instead of the others. In general, for a classification problem over labels $\mathsf{Y}=\{1,\dots,k\}$, we will consider a (given or constructed) nested partitioning of $\mathsf{Y}$ into a sequence of $T$-many subsets $\mathsf{U}_i$ of classes such that $\{y^*\} \triangleq \mathsf{U}_T \subset \mathsf{U}_{T-1} \subset \dots \subset \mathsf{U}_{0} \triangleq \mathsf{Y}$. As \edit{we} show in \edit{Figure 4} in the appendix, this partition implies a sequence of pairs of hypotheses $(h_t,h_t')=(y\!\in\!\mathsf{U}_t,y\!\in\!\mathsf{U}_{t-1}\!\setminus\!\mathsf{U}_t)$.\looseness=-1 A second challenge arises when the the number of input features is large. For very high-dimensional inputs (such as images or detailed health records), providing a WoE value for each feature will rarely be informative. Again, imagine our hypothetical doctor having to simultaneously analyze the relevance of thousands of symptoms. For such cases, we propose aggregating the input features into feature groups (e.g., super-pixels for images or groups of related symptoms for medical diagnosis). Formally, for an input $X$ of dimension $n$, we partition the feature indices into $m$ disjoint subsets, with $\mathsf{S}_1 \cup \dots \cup \mathsf{S}_m = \{1,\dots,n\}$. Equation~\eqref{eq:woe_multievidence} (or, equivalently, the chain rule of probability) allows for arbitrary groupings, so for any such partition we can compute \vspace{-0.1cm} \begin{equation}\label{eq:atom_woe} \!\text{woe}( h/h'\!:\!X) =\!\!\sum_{i=1}^m \underbrace{\log \frac{P(X_{\mathsf{S}_i}\!\mid\!X_{\mathsf{S}_{i-1}}, \tightdots, X_{\mathsf{S}_1}, h)}{P(X_{\mathsf{S}_i}\!\mid\!X_{\mathsf{S}_{i-1}}, \tightdots, X_{\mathsf{S}_1}, h')}}_{= \text{woe}(h/h': X_{\mathsf{S}_i}\mid X_{\mathsf{S}_{i-1}}, \dots, X_{\mathsf{S}_1})} \end{equation} \vspace{-0.1cm} where $X_{\mathsf{S}_i} = \{X_j\}_{j \in \mathsf{S}_i}$ is the $i$th feature group, or ``atom.'' \begin{algorithm}[t!] \caption{WoE meta-algorithm for complex models}\label{algo:greedy_woe} \begin{algorithmic}[1] \STATE {\bfseries Input:} Instance $X \in \mathbb{R}^n$, prediction $y^* \in \{1,\dots, k\}$ \STATE {\bfseries Parameters:} Features $\mathcal{A} = \{\{1\}, \dots,\{n\}\} $ or feature groups $\mathcal{A} = \{S_1, \dots, S_m\}$ \STATE Initialize $\mathsf{U}_0 \gets \{1,\dots, k\}$ \STATE $t \gets 0$\; \WHILE{$|\mathsf{U}_t| > 1$} \STATE $t \gets t+1$\; \STATE $\mathsf{U}_{t} \gets \textsc{SelectHypothesis}(\mathsf{U}_{t-1}, y^*)$\; \STATE $\overline{\mathsf{U}}_{t} \gets \mathsf{U}_{t-1} \setminus \mathsf{U}_{t}$ \COMMENT{relative complement} \STATE $\pi(\mathsf{U}_{t}) \gets \log \tfrac{P( y \in \mathsf{U}_{t})}{P(y \in \overline{\mathsf{U}}_{t})}$ \COMMENT{prior log odds} \FOR{$i=1,\dots, |\mathcal{A}|$} \STATE $\omega_i^t \!\gets\! \text{woe}(y\!\in\!\mathsf{U}_{t}/y\!\in\!\overline{\mathsf{U}}_{t} \!:\! X_{\mathcal{A}_i} \!\mid \! X_{\mathcal{A}_{i-1}}, \!\dots,\! X_{\mathcal{A}_{1}}) $\; \ENDFOR \STATE $\Omega_t \gets \sum_{i=1}^{|\mathcal{A}|}\omega_i^t$ \STATE $\textsc{DisplayExplanation}(\mathsf{U}_{t}, \overline{\mathsf{U}}_{t}, \mathcal{A}, \pi(\mathsf{U}_t), \{\omega_i^t\}_i, \Omega_t)$\; \ENDWHILE \end{algorithmic} \end{algorithm} \begin{figure*}[htp!] \centering \includegraphics[width=0.5\linewidth]{figs/example_multistep_expl_step2.png}% \includegraphics[width=0.5\linewidth]{figs/example_multistep_expl_step3.png}\\ \caption{Example two-step explanation produced by our \edit{method} for a \edit{model that} predict\edit{ed} the \edit{flu for some input}. The first step \edit{explains}, using \edit{the} weight of evidence, why the model favors \edit{viral diseases} (instead of \edit{bacterial ones}) for this \edit{input}. Then, the second step explains why the model predict\edit{ed the flu instead of} the remaining possible classes (\edit{i.e.,} other \edit{viral} diseases).}\label{fig:example_expl} \end{figure*} \subsection{A Meta-Algorithm for WoE Explanations}\label{sec:woe_algo} Using these extensions of the WoE, we propose a meta-algorithm for generating explanations for complex classifiers (Algorithm~\ref{algo:greedy_woe}). Given a model, an input, and a prediction, the algorithm generates an explanation for the prediction sequentially by producing WoE values for progressively smaller nested hypotheses. Specifically, at every step $t$, a subset of classes $\mathsf{U}_t \subset \mathsf{U}_{t-1}$ is selected and the remaining classes $\overline{\mathsf{U}}_{t} = \mathsf{U}_{t-1} \setminus \mathsf{U}_t$ \edit{are} ruled out. The user is shown a comparison of hypotheses $h_t: y \in \mathsf{U}_t$ and $h_t': y \in \overline{\mathsf{U}}_{t}$ consisting of both their prior log odds $\pi(\mathsf{U}_{t})$ \edit{(line 9)} and the WoE in favor of $h_t$ and against $h_t'$. WoE values are computed sequentially with each atom $\mathcal{A}_i$ (either an individual input feature or a group of features) as the evidence (line 11) and these values are summed to obtain the total WoE using the additive property (line 13). These values are presented to the user, and the process continues until all classes except the prediction $y^*$ have been ruled out (cf.~Principles 2--3).\looseness=-1 Left unspecified in this meta-algorithm are four key choices that are application-dependent and require further discussion. First is the question of how to define the $\textsc{SelectHypothesis}$ method to progressively partition the classes (line 7). If there is an inherent natural partitioning (e.g., ``viral'' versus ``bacterial,'' as in the example discussed previously), then $\textsc{SelectHypothesis}$ simply amounts to retrieving the largest subset in the partition containing the prediction $y^*$. For the general case, we propose selecting~the~hypothesis that maximizes a WoE-based objective: \[ \mathsf{U}_{t}\! \gets \!{\displaystyle \argmax_{\mathsf{U} \subset \mathsf{U}_{t-1}; y^*\in \mathsf{U}}} \text{woe}(y \in \mathsf{U}\medspace/\medspace(y \in \mathsf{U}_{t-1}\!\setminus\!\mathsf{U}) : X) \!-\! R(\mathsf{U}), \] where $R$ is a cardinality-based regularizer. $R$ should be chosen to penalize sets that are too small (which would yield granular explanations with many steps, in opposition to Principle 5) or too large (which would yield coarse explanations, to the detriment of Principle 3). Although the choice of $R$ should ideally be informed by the application and the user, a sensible generic choice is $R(\mathsf{U}) \propto \left||\mathsf{U}|-\frac{1}{2}|\mathsf{U}_{t-1}| \right|^p$, normalized so that $R(\mathsf{U})\in [0,1]$. Using this regularizer, Algorithm 1 approximately splits the remaining classes in half at every step, yielding roughly $O(\log k)$ steps in total.\looseness=-1 Second, it should be noted that lines 10--12 in Algorithm~\ref{algo:greedy_woe} implicitly assume an ordering of the atoms, and that this ordering might affect the WoE values. In some applications, there might be a conditional independence structure known a priori that could inform the choice of atoms and their ordering (e.g., those simplifying the conditioning in line 11 the most). If not, the ordering can again be chosen randomly or based on the sorted per-atom conditional WoE values.\looseness=-1 Third, computing the per-atom WoE (line 11) requires the conditional likelihoods $P(X_{\mathcal{A}_i} | X_{\mathcal{A}_{i-1}}, \dots, X_{\mathcal{A}_1}, Y)$. Ideally, the \edit{model} would compute these likelihoods internally. If, instead, it computes only marginal feature likelihoods $P(X_{\mathcal{A}_i} | Y)$, we can use a na\"ive Bayes (NB) approximation\,---\,that is, use these in place of the conditional likelihoods in Equation~\eqref{eq:atom_woe}. If the model is a black box or does not compute \edit{conditional} likelihoods internally,~then~these must be estimated \edit{as we explain below}. \looseness=-1 Finally, there is the question of how to implement $\textsc{DisplayExplanation}$. When the number of atoms is large, the WoE values for only the most salient atoms can be displayed (cf.~Principle 5)\,---\,e.g., those with absolute WoE larger than a given threshold $\tau$; \citet{good1985weight} suggests $\tau= 2$ as a rule of thumb. Otherwise, all per-atom WoE values can be displayed along with the total WoE and prior log odds. An example two-step explanation produced by our method (on fabricated data from our user study tutorial) is shown in Figure~\ref{fig:example_expl}. The sorting and color coding of the features by their WoE value\edit{s} makes it apparent which of these contribute the most evidence in favor or against the selected class (or set of classes), and the labeling along the x-axis provides guidelines for context. The visualization suggests the additive nature of these values (i.e., that stacking blue bars and subtracting red ones yields the total WoE). The bottom panel, a graphical representation of Equation~\eqref{eq:woe_prior_posterior}, disentangles the model's estimated prior class odds (which \textit{a priori} weakly favor \edit{the} flu and other viral diseases), from its total WoE (very strong when contrasting \edit{viral and bacterial~diseases}, less so for the \edit{flu versus other viral diseases}).\looseness=-1 \subsection{WoE Estimation for Black-box Models} As \edit{we} noted \edit{above}, computing \edit{the per-atom} WoE \edit{(line 11 in Algorithm 1)} requires evaluating the conditional likelihoods $P(X_{\mathcal{A}_i} | X_{\mathcal{A}_{i-1}}, \dots, X_{\mathcal{A}_1}, Y)$. In many practical settings, \edit{including those in which the model is a black box}, these \edit{conditional likelihoods are not computed internally}, so they \edit{must} be estimated. In such \edit{settings}, we propose \edit{fitting} a conditional likelihood \edit{estimation} model \edit{using the model's predictions $\hat{y}=\mathcal{M}(x)$ (not the true labels $y$) as a preliminary step. This conditional likelihood estimation model can then be called on demand when computing the per-atom WoE}.\looseness=-1 In some settings\edit{,} it \edit{may} be \edit{possible} to fit a full (i.e., conditioned on \textit{both} the class $Y$ and all previous atoms \edit{$A_i-1, \ldots, A_1$) conditional likelihood estimation model}, for example, via kernel or spectral density estimation methods when \edit{working with} low-dimensional data, or \edit{via} autoregressive or recurrent neural networks \edit{when working with} text or time-series data. For more complex \edit{types of data}, such as images, methods based on normalizing flows and neural autoregressive models \citep[e.g.,][]{rezende2015variational, papamakarios2017masked} are likely \edit{to be} more appropriate. \edit{In settings where fitting a full conditional likelihood estimation model is infeasible}, an NB approximation can be used \edit{to estimate} class- (but not atom-) conditional likelihoods\edit{, for example,} via a Gaussian NB classifier.\looseness=-1 We emphasize that fitting \edit{a conditional} likelihood estimation model\,---\,the main computational bottleneck \edit{of} our method\,---\,must be done only once, potentially offline. This is in contrast \edit{to} perturbation-based interpretability methods~like LIME that fit \edit{a new} model for every prediction.\looseness=-1 We \edit{assess} the quality of finite-sample WoE estimation experimentally in the \edit{``}Quantitative Experiments\edit{''} section. \looseness=-1 \subsection{Relation to LIME and SHAP} When viewed from a probabilistic perspective, most post-hoc interpretability methods revolve around a model's predictive posterior\,---\,that is, they seek explanations that deconstruct $P(Y=y^*\mid X)$ in various ways. For example, LIME \citep{ribeiro2016why} seeks to approximate $f(x) = P(Y \mid X=x)$ in the vicinity of $x_0$ through a simpler, interpretable surrogate model $\tilde{f}(x)$. Similarly, SHAP \citep{lundberg2017unified} quantifies variable importance by analyzing the effect on the posterior of ``dropping'' variables $X_i$ from the input $X$. In contrast, the WoE focuses\,---\,directly, in the case of the evidence-likelihood interpretation and indirectly in the case of the hypothesis-odds interpretation\,---\,on the conditional likelihood $P(X = x \mid Y)$. In other words, for a given input $x$, a WoE explanation is based on the probability assigned by the model to $x$ (or a subset of its features) given, e.g., $Y=y^*$.\looseness=-1 When viewed in this way, the relationship between WoE explanations and other post-hoc interpretability methods like LIME and SHAP is akin to the relationship between generative and discriminative models. Indeed, as is the case for some pairs of generative and discriminative models (e.g., na\"{i}ve Bayes and logistic regression), these different interpretability methods also turn out to be equivalent\,---\,two sides of the same coin\,---\,for some simple classifiers, as \edit{we} show in the appendix for logistic regression. However, this is not generally the case. Moreover, even when the explanations generated by different interpretability methods qualitatively agree (i.e., the same features are highlighted as being important), the specific interpretations of the explanations will differ. Indeed, the WoE uses a different operationalization of the notion of feature importance, in turn entailing different units of explanation: log likelihoods and log\edit{-}odds ratios in the case of the WoE and linear attribution scores for~posterior probabilities in the case of LIME and SHAP.\looseness=-1 \section{Quantitative Experiments} \label{sec:quant} Here, we assess the quality of finite-sample WoE estimation and the robustness of the WoE to perturbations of the inputs.\looseness=-1 \begin{figure} \centering \includegraphics[width=\linewidth, trim={0 1cm 0 0}, clip]{figs/mse_n_dim.png} \includegraphics[width=\linewidth, trim={0 0.25cm 0 0}, clip]{figs/ndcg_n_dim.png} \caption{Quality of WoE estimation. Top: MSE. Bottom: NDCG, ranging from $0$ (worst) to $1$ (perfect) ranking quality.} \label{fig:recovery_plot} \end{figure} \begin{figure*}[ht!] \centering \includegraphics[width=0.9\linewidth, trim={0 1cm 0 0}]{figs/lip_combined_datasets.png} \caption{Explanation robustness across benchmark datasets. Extreme values away from $L=1$ (dashed line) are undesirable.} \label{fig:robustness_plot} \end{figure*} \subsection{Quality of Finite-Sample WoE Estimation}\label{sec:woe_approx_experiments} Our first experiment evaluates the quality of WoE estimates from finite samples. As \edit{we} explained above, such estimates are needed if the model is a black box or does not compute likelihoods \edit{internally}. For evaluation purposes, we consider a model that, by construction, computes all the quantities required for exact WoE computation, but treat it as black box\,---\,that is, its internal WoE computation will be used only for evaluation, and is not available to \edit{our method}. Instead, we separately fit a \edit{conditional} likelihood estimation model by querying the model for a small number of inputs, and use this estimation model to compute WoE values at explanation time. We control for model \edit{mis}specification by having both the model and our \edit{conditional} likelihood estimation model use the \edit{NB} assumption. Specifically, we use a smoothed Gaussian \edit{NB} (GNB) classifier. This allows us to focus on the quality of WoE estimation from finite~samples, but does not address model misspecification.\looseness=-1 First, we generate a dataset of a given dimension. We train a \edit{model} on a subset of this dataset of size $N_{\text{train}}=1000$ and fit the \edit{conditional} likelihood estimation model on a separate subset of size $N_{\text{fit}}$, which we vary. For every test \edit{input} $x_i$ ($N_{\text{test}}=10$), we compute true WoE values for each input feature using the model's prior and posterior probabilities, and then compute estimated WoE values according to the estimation model in the appendix. We compare these using two metrics: mean squared error and normalized discounted cumulative gain (NDCG), a measure of ranking quality that might be relevant to practitioners, applied to the relative ranking of input features by their (true or estimated) WoE.\footnote{The NDCG is only defined for positive values, so we compute it separately for positive and negative values and average them.}\looseness=-1 Figure~\ref{fig:recovery_plot} shows these metrics as a function of input dimension and sample size $N_{\text{fit}}$. As expected, estimation quality improves with the number of samples used for fitting, and degrades gracefully as the input dimension increases. These results suggest that the WoE can be accurately estimated\,---\,even in relatively high dimensions\,---\,from finite samples, \edit{although} we caution that these results may look different for other models, and we do not measure \edit{model misspecification} error due to the NB approximation.\looseness=-1 \subsection{Robustness of WoE}\label{sec:woe_robustness_experiments} Previous work has argued that interpretability methods should be robust in the sense that the explanations they provide should not vary dramatically when the input whose prediction is being explained changes by a small amount. To investigate the robustness of our method, we follow the set-up of~\citet{alvarez-melis2018robustness}. Letting $\mathcal{E}(\cdot)$ be a function that maps feature vectors $x\in\mathbb{R}^n$ to explanation vectors (e.g., importance scores) $e \in \mathbb{R}^n$, we quantify its robustness around $x_0$ through its local Lipschitz constant: \begin{equation}\label{eq:lipschitz} L(x_0) = \max_{x_j \in \mathcal{B}_{\varepsilon}(x_0)} \frac{\| \mathcal{E}(x_j) - \mathcal{E}(x_0)\|}{\|x_j - x_0\|} , \end{equation} where $ \mathcal{B}_{\varepsilon}(x_0) = \{ x \medspace|\medspace \|x - x_0\| \leq \varepsilon\}$. Intuitively, $L(x_0)$ quantifies the largest relative change in importance scores in a small neighborhood around $x_0$. Extreme values are usually undesirable, as they indicate explanations that are either too sensitive (large $L$) or not responsive enough (very small $L$) to changes in the input features. In most settings, values below $1$ but bounded away from $0$ are preferable.\looseness=-1 Concretely, we first train a GNB classifier\edit{.} Then, for any input $x$ we use the classifier to generate a prediction $y$, and input both of these to our method to generate $\mathcal{E}(x)$, a vector of WoE values for each feature $x_i$. Since computing the robustness metric \eqref{eq:lipschitz} involves maximization, we estimate this quantity from finite samples using Bayesian optimization, making repeated calls to $\mathcal{E}(\cdot)$. We focus on standard benchmark classification datasets from the UCI repository \citep{dua2017uci}, and compare \edit{our method to} LIME \citep{ribeiro2016why} and SHAP \citep{lundberg2017unified}. Figure~\ref{fig:robustness_plot} shows \edit{the} results \edit{for} ten repetitions with different random seeds for each \edit{dataset and} interpretability \edit{method} pair. The red dashed line indicates the bound $L(x)=1$. \edit{For} all but one dataset, WoE explanations are\edit{,} on average\edit{,} as close or closer to the ideal Lipschitz robustness as \edit{explanations} generated by LIME and SHAP.\looseness=-1 \section{User Study}\label{sec:user_study} Throughout this paper, we have argued that interpretability is fundamentally a human-centered concept and that the evaluation of interpretability methods should therefore focus on the needs of humans, exploring how they use interpretability tools, as well as their understanding of the concepts that underlie them. In this section, we present a user study that we carried out to assess the usefulness of our method in the types of scenarios in which it would be used in practice. Such studies are commonly used in the HCI community to distinguish between designers' intended use of a tool and users' mental models~\cite{gibson1977theory,norman2013design}.\looseness=-1 \subsection{Study Design} We conducted an artifact-based interview study with 10 ML practitioners to evaluate the use of our interpretability \edit{method, implemented in a simple} tool (the artifact)\edit{,} in a controlled setting. In such qualitative studies, the goal is to \edit{ensure sufficient interaction} time and nuanced data collection \edit{for each} participant, which is typically only feasible for relatively small sample sizes~\cite{hudson2014concepts,olsen2007evaluating,turner2006determining}. \edit{Our} study followed a think-aloud protocol\edit{, in which participants were asked to verbalize their thought processes as they used the tool to perform specific tasks, in order to help us identify specific} concepts and \edit{functionalities} that \edit{might} be confusing. We place\edit{d} participants in a controlled setting \edit{(}rather than observe them using the tool on their own models\edit{)} because of challenges including data access, inconsistencies in the types of data analyzed by the \edit{participants}, and \edit{the potential} difficulty \edit{of} establishing patterns \edit{across settings}.\looseness=-1 Our study consisted of two main parts: a tutorial \edit{intended} to introduce the concepts and \edit{functionalities} needed to use our tool, followed by the main study in which participants answered questions about a \edit{pre-trained ML} model using the tool. We also conducted pre-study interviews to establish participants' background\edit{s} in ML and post-study interviews in which participants reflected on their experience\edit{s} with the tool and how they might use it in their ML pipeline\edit{s}. The study design was approved by our internal \edit{institutional review board}. \edit{Excerpts (screenshots) from the} Jupyter notebooks \edit{that we used in} the tutorial and \edit{in the} main study \edit{are in the appendix; the complete notebooks are available online.}\footnote{\url{http://github.com/dmelis/interpretwoe}}\looseness=-1 \subsubsection{Tutorial}\label{sec:tutorial} \edit{\citet{kaur2020interpreting} found} that practitioners often use interpretability tools without fully understanding them, highlighting the importance of \edit{providing well-designed} tutorials and other accompanying documentation. For our user study, we designed a tutorial to introduce the \edit{concepts and functionalities needed to use our tool}, to evaluate \edit{participants'} understanding of these concepts, and to check whether \edit{their} responses in the main study were \edit{likely} based on a sound understanding\,---\,without being too time-consuming or tedious. After several iterations and pilot studies, we converged on an approximately 40-minute\edit{-long} tutorial based on a Jupyter notebook \edit{containing} equations, text, and images. This tutorial covered \edit{log-odds ratios}, weight of evidence, feature group for high-dimensional inputs, and sequential explanations for multi-class settings.\looseness=-1 \subsubsection{Main Study} The goal of the main study was to assess participants' understanding of \edit{the WoE} and \edit{to investigate their} use of \edit{our interpretability} tool in the context of a realistic ML task. Participants were given a Jupyter notebook that included a dataset, an ML model trained using \edit{the} dataset, and our tool\edit{. They} were \edit{then} asked to answer several questions with the help of outputs (\edit{i.e.,} visualizations) from \edit{our} tool.\looseness=-1 The ML model was a random forest classifier trained \edit{using} the Online News popularity dataset~\citep{fernandes2015proactive}\edit{, which} consists of 39,797 news articles\edit{. Each article is} represented \edit{using} 59 features \edit{that} captur\edit{e} metadata about the article, \edit{such as its} length, \edit{any} links, and \edit{its sentiment} polarity. We trained the model to predict the category that the article was published under (\edit{e.g.,} ``Lifestyle'' or ``Business''), creating a six-class classification \edit{task}. We chose this dataset and \edit{this task} because the domain is understandable without expert knowledge or prior experience, the number of classes is large enough to \edit{permit} meaningful sequential explanations, and there are enough features to make explanations based on feature groups sufficiently different from \edit{explanations} based on individual features. The main study was \edit{itself} divided into two parts, \edit{which were designed to let us observe the use} of the two \edit{extensions} of \edit{the} WoE described in \edit{the ``Composite Hypotheses and Evidence'' section}: feature \edit{groups} and sequential explanations. In the first part, participants were given the option to view \edit{explanations} based \edit{on feature} groups \edit{or} explanations \edit{based on individual features}, and were asked questions that could be answered using either \edit{type} of \edit{explanation} (e.g., ``What aspects of the news article contributed the most to this prediction?''). \edit{This part of the study} was \edit{intended} to \edit{surface participants'} preferences. In the second part, participants were given the option to generate one-shot or sequential explanations, \edit{and} were asked questions that could only be precisely answered using sequential explanations (e.g., ``Why didn't the model predict [subset of classes]?''). \edit{This part of the study} was \edit{intended} to \edit{assess} whether participants could successfully use sequential explanations.\looseness=-1 \subsubsection{Participants} \edit{Potential} participants were recruited via email. To be considered for the study, they were asked to complete a survey about their \edit{ML background and their} experience with interpretability tools. Of 41 \edit{survey} respondents, we randomly selected 10 \edit{to participate in the study}. All \edit{participants} were ML practitioners (e.g., data scientists) with 1--20 years of experience. On average, \edit{participants} rated the role of ML in their jobs as 6.7 and their experience with interpretability tools as 3.2, both on a scale of 0 (``not at all'') \edit{to} 7 (``extremely''). \edit{P}articipants \edit{also rated} their familiarity with concepts from probability \edit{relevant to the WoE on a scale of 0 to 7}. Their average ratings were 2.7 for posterior \edit{class probabilities}, 3.3 for log likelihood\edit{s}, 3.2 for \edit{log-odds ratios}, and 0.9 for \edit{the WoE. On average, p}articipants took 1.7 hours to complete the study\edit{. Each participant was} compensated with a \$40 Amazon gift card.\looseness=-1 \subsubsection{Methods} Participant\edit{s'} open-ended answers were scored by comparing them \edit{to} an answer key prepared in advance by two of the \edit{authors}. \edit{Answers that correctly identified key aspects (e.g., a feature with a large positive WoE value, pushing the model toward a particular prediction) were treated as correct even if the participants' specific language did not exactly match the language in the answer key.} To examine patterns of tool use, \edit{the} usability of \edit{our} tool, \edit{participants' interpretability} needs, and participants' general impressions, we analyzed automatically generated audio transcripts for high-level themes using inductive thematic analysis~\cite{braun2012thematic} and affinity diagramming. \looseness=-1 \subsection{Results} \edit{T}he results \edit{from our} user study \edit{divide naturally} into three categories: participants' \edit{understanding of relevant concepts}, tool \edit{usability} and \edit{participants' preferences}, and \edit{general needs} for interpretability tools. First, the pre-study interview\edit{s} and answers to the checkpoint questions in the tutorial provide\edit{d} insight into participants' understanding of concepts relevant \edit{to the WoE}. Second, participants' approach\edit{es} to the questions \edit{in} the \edit{main study and} the\edit{ir} \edit{patterns of tool use enabled us to examine} the usability of our tool. Finally, \edit{via} the \edit{main} study and \edit{the} post-study interviews, we \edit{were} able to \edit{uncover participants' general interpretability needs and} additional criteria (\edit{beyond} our design principles) \edit{to consider when designing and} evaluating interpretability tools.\looseness=-1 \subsubsection{Understanding of Relevant Concepts} Our analysis \edit{showed} that most participants (7/10) struggled to understand and use prior class probabilities in the tutorial. The section on this topic was time-consuming: \edit{on average}, participants spent a third of their tutorial time on this section. Eventually, most participants either ignored the prior class probabilities or used them incorrectly, supporting the base rate fallacy. Nonetheless, participants were able to use \edit{the} WoE to \edit{correctly} answer questions in the main study for which prior class probabilities were relevant. This raises the possibility that \edit{although} they struggled with the abstract concept, they were able to use the information indirectly (e.g., via displayed class probabilities). This \edit{finding} is \edit{consistent} with \edit{those of \citet{kaur2020interpreting}}, who \edit{showed} that data scientists struggle to explain the \edit{concepts underlying} the explanations produced by \edit{generalized additive models~\cite{hastie1990generalized,Caruana2015-qf}} and SHAP \edit{\citep{lundberg2017unified}, even though they} still find these tools useful.\looseness=-1 Although participants generally understood the concept of WoE, some confused negative WoE values with negative values for input features, thus finding it challenging to make sense of the explanations. As a result, two participants provided incorrect answers for \edit{the} questions in the main study. \subsubsection{Tool Usability and \edit{Participants' Preferences}} Participants had no overwhelming preference between \edit{explanations based on feature groups and explanations based on individual features}. Indeed, they noted that the two levels of granularity provide complementary information, and switching between the two options was a clear pattern across all participants. \edit{Although} feature groups provide a high-level overview, making it ``easier to manage [reading the plot]...[and the] direction of analysis is a lot clearer'' (P8), \edit{explanations based on individual features} help participants in ``looking into more details in general...to know exactly which feature it was [that was responsible for a prediction]'' (P10). We observed some differences in behavior based on participants' roles and expertise, though of course these are inconclusive with our small sample size. Participants with more ML experience tended to rely on feature-level plots, while those with customer-facing jobs more often provided high-level answers based on feature groups, noting that feature~groups~``provide customer-friendly explanations'' (P8).\looseness=-1 Participants found sequential explanations to be a helpful breakdown of a larger explanation into parts. P3 noted these were like a ``story of how the prediction was made.'' Sequential explanations prompted more detailed answers to our questions and most participants (7/10) accurately answered questions in \edit{the second} part of the \edit{main} study using sequential explanations. They explained that the type of questions\,---\,which required understanding how each of the classes were ruled out\,---\,could not be answered via one-shot explanations. P8 commented, ``I find this to be quite helpful... I guess without this breaking it down to this point I wouldn't have thought twice really about this [input feature group] being a [differentiating] factor between the two [output classes]...I think that this would be a nice like final understanding [of the predicted class], this goes a lot deeper than I probably could have gone just looking at that without the tool. So I think it was very helpful in that case.'' \looseness=-1 Although most participants were happy with the level of detail presented, some participants with more ML experience expressed a desire for deeper understanding of how the explanations were generated. They understood the underlying concepts, but were wary of anything that appeared automated, including the breakdown of class comparisons in sequential explanations (which was automated) and the feature groups (which were actually manually generated).\looseness=-1 \edit{Even though} participants said that the tool helped them understand the model's predictions, not all of them envisioned the tool being added to their ML pipelines. Participants with significant prior ML experience already had established ways of ensuring that model predictions are reasonable, but recognized other exciting use cases for the tool, such as communicating complex predictions to less experienced end users. Particularly for high-risk domains, visualizations from the tool could help users probe odd predictions. P7 noted, ``With my focus on medical data, I do see the need in working with~a customer...there this [tool] would be a must-have. My team, we are engaged with customers and we have to educate them fast... So for me model interpretability there comes very close side by side with fairness.''\looseness=-1 \subsubsection{General Needs for Interpretability Tools} Most interpretability tools, including ours, rely on tutorials \edit{and other accompanying} documentation to provide an introduction to the tool's concepts and functionalities. All participants appreciated the information presented in our tutorial: ``I can't imagine doing this [study] without the tutorial. I generally know a lot more about these concepts now'' (P5). The tutorial seemed to impact participants' overall accuracy in answering the \edit{questions in the} main study\,---\,those who spent longer on the tutorial tended to provide more accurate and more thoughtful answers. This manifested as longer time spent on exploring the tool in the tutorial and ensuring that their answers to the checkpoint questions were accurate and thorough. Even participants with less ML experience provided accurate answers when they devoted time to the tutorial. Participants appreciated the example in the tutorial and were able to generalize from \edit{this} example to \edit{the} questions the main study: ``The tutorial...helps you start in the right place. I went back to the example in the tutorial to~[determine how to] answer questions in the study'' (P6).\looseness=-1 Finally, participants expressed a desire to be able to more easily switch between different options (e.g., input features versus feature groups) rather than re-running code. Interactivity was consequently the most commonly requested functionality in the post-study interview. This is in line with prior work on human-centered design principles for ML~\cite{amershi2019guidelines,hohman2019gamut,WB19}.\looseness=-1 \section{Limitations} \edit{All} interpretability \edit{methods}, \edit{including ours}, involve various design choices and assumptions (both implicit and explicit), many of which give rise to potential limitations. \edit{First}, the \edit{concept} of interpretability is notoriously ambiguous, and unlike \edit{supervised ML tasks}, there is no ground truth to \edit{use for evaluation}, even for proxy \edit{concepts like} feature importance. \edit{As a result}, different \edit{interpretability methods} assume different notions of interpretability, propose different quantities to operationalize them, and (when needed) rely on different \edit{techniques} to estimate the\edit{m. In turn,} these choices \edit{mean} that no interpretability method \edit{will ever} be universally ideal. Moreover, summarizing the behavior of complex models comes at a price~\cite{rudin2018please}\,---\,that is, the explanations are partial, only hold in a small neighborhood \citep{ribeiro2016why}, or make strong assumptions about the data \citep{lundberg2017unified}. As a result, explanations generated by one interpretability method \edit{seldom} strictly \edit{dominate explanations generated by} another\edit{.} Furthermore, different explanations might reveal information about different aspects of the underlying model's behavior.\looseness=-1 Under this perspective, the design choices \edit{and assumptions involved in} our interpretability method necessarily limit its scope and applicability. Starting from our \edit{decision to distill characteristics} of human explanation in\edit{to human-centered design principles, our method} assum\edit{es} that \edit{human characteristics} are desirable for \edit{machine-generated explanations. And, although the specific characteristics} that we \edit{focus on yield a coherent} set of design principles, \edit{these principles are} not exhaustive or universal. Some \edit{may} not be~necessary in all \edit{settings,} and all are open to refinement.\looseness=-1 \edit{By relying on the concept of} weight of evidence \edit{(WoE) from information theory, our} method inherits many of its strengths and limitations. Concretely, there are three main settings in which there is a clear case for using explanations based on the WoE: \edit{1)} when the underlying model is generative, \edit{2) when the underlying model is} log linear, and \edit{3) when the underlying model is} a multi-class classifier. We provide a detailed discussion of these three settings in the appendix. In terms of limitations, the WoE requires access to \edit{the} conditional likelihoods \edit{$P(X_{\mathcal{A}_i} | X_{\mathcal{A}_{i-1}}, \dots, X_{\mathcal{A}_1}, Y)$}, which limits its use to settings in which these are accessible or can be accurately estimated from finite samples\edit{. Estimating densities for more complex types of data, such as images,} is an active area of research, and \edit{although it may be possible to integrate} new advances \edit{into our method, its} applicability to \edit{such types of data} is \edit{currently} limited. Other important design choices \edit{involved in our method} include the \edit{technique} for partitioning \edit{the} classes \edit{in multi-class settings to yield} sequential explanations and the \edit{technique for partitioning input} features into feature \edit{groups}. Although we \edit{chose} generic solutions to these \edit{challenges, there are other techniques, which may be more appropriate in some settings,} which \edit{will} invariably lead to different explanations. We \edit{defer a} thorough \edit{investigation} of \edit{these choices} for future work.\looseness=-1 The main limitations of our user study \edit{are the} number of participants, \edit{the type of participants,} and the extent to which the study conditions \edit{mimic a realistic setting}. We \edit{chose to conduct an} artifact-based interview \edit{study to ensure sufficient interaction time and nuanced data collection for each participant, but this limited} the number of participants \edit{that we could consider, thereby precluding} statistical analys\edit{e}s. Our participants were also limited to \edit{ML} practitioners. \edit{Following previous} work \cite[e.g.,][]{kaur2020interpreting}, we \edit{chose to focus on ML practitioners} because \edit{they} are frequent users of interpretability tools in the wild. \edit{Finally, although} we tried to \edit{design the study so as to mimic} a realistic setting, we cannot be sure that this experience was representative of participants' day-to-day \edit{experiences} (e.g., working with their own datasets). Ideally, \edit{we would have} run a longitudinal field study with multiple types \edit{of participants to enable us to} observ\edit{e participants'} tool use over time as \edit{they} gain expertise in using it. However, this would \edit{have} require\edit{d} additional resources (e.g., to support multiple types of data) \edit{and} was \edit{therefore infeasible}. Instead, our \edit{user study} serves as an initial evaluation \edit{of our} interpetability \edit{method}.\looseness=-1 \section{Discussion} \edit{In this paper, we take inspiration from the study of human explanation, drawing on the literature on human explanation in philosophy, cognitive science, and the social sciences to propose a list of design principles for machine-generated explanations that are meaningful to humans. We develop a method for generating explanations that adhere to these principles using the concept of weight of evidence from information theory. We show that this method can be adapted to meet the needs of modern ML\,---\,that is, high-dimensional, multi-class settings\,---\,and that the explanations can be estimated accurately from finite samples, are robust to small perturbations~of the inputs, and are usable by ML practitioners.}\looseness=-1 \edit{This paper opens several avenues for future work. Adapting modern density estimation methods for complex types of data, such as images, might hold the key to wider applicability of our method. Regarding evaluation, an immediate next step would be to carry out a follow-up user study to investigate various design choices, such as the technique for partitioning the classes in multi-class settings. Ideally, future user studies should involve participants' own models and should rely on questions that attempt to uncover insights that are relevant to their day-to-day experiences.}\looseness=-1 \edit{The findings from our user study offer important lessons that we believe are generally applicable to other interpretability tools. Chief among these is the importance of user-friendly and engaging tutorials that provide users with the necessary understanding of the tool and its intended usage, and users' desire for flexibility in tools. These results underscore the importance of putting human needs at the center of the design and evaluation of interpretability methods. The human computation community is uniquely situated to drive this work as it requires interdisciplinary expertise in both ML and HCI, fields that are central to HCOMP.}\looseness=-1 \edit{In the spirit of developing AI responsibly, we believe that papers proposing new interpretability methods should also provide a discussion, as we have done in the previous section, of not only those settings for which the proposed method is suitable, but also those settings that fall outside its scope. In addition, we recommend that authors should explicitly describe the notion of interpretability (or explanation) that they aim to operationalize, allowing readers to situate the contributions in relation to other interpretability~methods and to understand their scope and applicability.}\looseness=-1

398f475137643a3182b97cab1bee9819d504a1d6

\section{Introduction} \label{sec:introduction} Numerical techniques for studying marine propellers have seen immense advances in the past decades. For example, the lifting surface procedures based on vortex lattice and panel methods have been widely used in propeller design. A representative program in this category is the PSF code developed at MIT in the 1980's \cite{greeley-1982}. The vortex lattice and panel methods have very quick turnaround time, and can predict propeller loads rather accurately when combined with empirical vortex and wake models. They are, however, incapable of revealing the fine details of a flow field for in-depth analysis of flow physics such as flow instability and acoustics. Since the 1990's, the Reynolds-averaged Navier–Stokes (RANS) methods have gained popularity in simulating propeller flows. The RANS methods do provide richer flow field information than the inviscid vortex lattice and panel methods. But the averaging nature of a RANS (even an unsteady RANS) method still smooths out a lot of flow details, especially the small and intermediate instantaneous eddies. In more recent years, the increasing high-performance computing power has allowed eddy-resolving techniques, such as detached-eddy simulation (DES) and large-eddy simulation (LES), to have been performed on marine propellers in the scale of millions or even trillions of grid elements. For instance, \citet{muscari-2013} and \citet{di-2014} employed DES to study the vortex dynamics of a propeller in different flow conditions. \citet{verma-2012} and \citet{jang-2013} used LES to investigate the flow around a reverse rotating marine propeller. \citet{balaras-2015} applied an immersed boundary based LES technique to explore the flow around a propeller with and without an upstream appendage. \citet{kumar-2017} systematically studied the wake instability of a propeller using LES. However, all the aforementioned DES and LES approaches are of low-order accuracies (at most second-order) that could potentially introduce large numerical dissipation and dispersion to a flow field. On the other hand, high-order (third and above) methods, especially polynomial based high-order methods, are seeing rapid development in recent years. High-order methods show many advantages over the traditional low-order ones. For example, for the same number of degrees of freedom, a high-order method produces solutions with much smaller numerical error than a low-order method does. Furthermore, a high-order method can employ high-order unstructured curved meshes that approximate curved geometries much more accurately than the linear meshes for a low-order method. The most popular high-order methods include the discontinuous Galerkin (DG) method \cite{reed-1973, cockburn-2011}, the spectral element method \cite{patera-1984, karniadakis-2005, kopriva-2009}, the spectral volume method \cite{wang-2002a, wang-2002b}, the spectral difference (SD) method \cite{kopriva-1996a, kopriva-1998, liu-2006, wang-2007a,jameson-2010,balan-2012}, etc. Among these methods, the SD method solves equations in differential form directly, and is one of the most efficient high-order methods. Recently, the ideas of collocating solution and flux points of the SD method and correcting fluxes using higher-degree polynomials have led to an even more efficient high-order method --- the flux reconstruction (FR) method \cite{huynh-2007, huynh-2009}, also known as the correction procedure via reconstruction (CPR) \cite{wang-2009}. Besides its better efficiency, by choosing different correction polynomials, the FR method can recover many existing high-order schemes such as DG and SD, and can even produce new schemes that had never been reported before. The stability of the FR method has been proved in \cite{jameson-2012}. The most recent developments on the FR method are summarized in \cite{wang-2016}. To the authors' knowledge, there is still no reported high-order eddy-resolving simulation of marine propellers. One important reason is the severe challenge on how to treat the complex rotational geometry of a propeller in a high-order method without deteriorating the method's accuracies in both space and time. To tackle this challenge, \citet{zhang-2015a, zhang-2015b} developed a high-order sliding-mesh method for the SD and FR methods by introducing the concept of curved dynamic mortar elements. A parallelization approach was proposed for this method in \cite{zhang-2016a}, and the extension to three-dimensions was achieved in \cite{zhang-2016c}. More recently, this method was further extended to sliding interfaces with general nonuniform meshes \cite{zhang-2018}. An updated version that is high-order in time, arbitrarily high-order in space, provably conservative, and provably outflow preservative has also been established \cite{zhang-2020a}. In the present work, we apply this method to implicit LES (without using an explicit sub-grid-scale model) \cite{parsani-2010, uranga-2011, beck-2014, bull-2015, vermeire-2016} of the loads and flow fields of a real marine propeller at various working conditions. This is the first time a high-order method being applied to simulate a marine propeller. The rest of this paper is organized as follows. Section \ref{sec:methods} gives a brief description of the flow equations and the numerical methods. Section \ref{sec:setup} details the simulation setup. Simulation results and discussions are reported in Section \ref{sec:results}. Finally, Section \ref{sec:summary} concludes this paper. \section{Numerical methods} \label{sec:methods} \subsection{The physical equations} We numerically solve the following three-dimensional unsteady Navier-Stokes equations in a conservative form, \begin{equation} \frac{\partial \mathbf{Q}} {\partial t} + \frac{\partial \mathbf{F}} {\partial x} + \frac{\partial \mathbf{G}} {\partial y} + \frac{\partial \mathbf{H}} {\partial z} = \boldsymbol{0}, \label{eq:physical} \end{equation} where $\mathbf{Q}$ is the vector of conservative variables, $\mathbf{F}$, $\mathbf{G}$ and $\mathbf{H}$ are the flux vectors in each coordinate direction. These terms have the following expressions, \begin{align} \mathbf{Q} &= [\rho ~ \rho u ~ \rho v ~ \rho w ~ E]^\mathsf{T}, \label{eq:Q} \\ \mathbf{F} &= \mathbf{F}_{\text{inv}}(\mathbf{Q}) + \mathbf{F}_{\text{vis}}(\mathbf{Q},\nabla \mathbf{Q}), \label{eq:F}\\ \mathbf{G} &= \mathbf{G}_{\text{inv}}(\mathbf{Q}) + \mathbf{G}_{\text{vis}}(\mathbf{Q},\nabla \mathbf{Q}), \label{eq:G}\\ \mathbf{H} &= \mathbf{H}_{\text{inv}}(\mathbf{Q}) + \mathbf{H}_{\text{vis}}(\mathbf{Q},\nabla \mathbf{Q}), \label{eq:H} \end{align} where $\rho$ is fluid density, $u$, $v$ and $w$ are the velocity components, $E$ is the total energy per volume defined as $E = p/(\gamma-1) + \frac{1}{2}\rho(u^2+v^2+w^2)$, $p$ is pressure, $\gamma$ is the ratio of specific heats and is set to 1.4. The fluxes have been split into inviscid and viscous parts. The inviscid fluxes are only functions of the conservative variables and have the following expressions, \begin{equation} \mathbf{F}_{\text{inv}} = \left[ \begin{array}{c} \rho u \\ \rho u^2 + p \\ \rho uv \\ \rho uw \\ u(E+p) \end{array} \right], \ \mathbf{G}_{\text{inv}} = \left[ \begin{array}{c} \rho v \\ \rho uv \\ \rho v^2 + p \\ \rho vw \\ v(E+p) \end{array} \right],\ \mathbf{H}_{\text{inv}} = \left[ \begin{array}{c} \rho w \\ \rho uw \\ \rho vw \\ \rho w^2 + p \\ w(E+p) \end{array} \right]. \label{eq:FGHinv} \end{equation} The viscous fluxes are functions of the conservative variables and their gradients. The expressions are \begin{equation} \mathbf{F}_{\text{vis}} = -\left[ \begin{array}{c} 0 \\ \tau_{xx} \\ \tau_{yx} \\ \tau_{zx} \\ u\tau_{xx}+v\tau_{yx}+w\tau_{zx}+ \kappa T_x \end{array} \right], \end{equation} \begin{equation} \mathbf{G}_{\text{vis}} = -\left[ \begin{array}{c} 0 \\ \tau_{xy} \\ \tau_{yy} \\ \tau_{zy} \\ u\tau_{xy}+v\tau_{yy}+w\tau_{zy}+\kappa T_y \end{array} \right], \end{equation} \begin{equation} \mathbf{H}_{\text{vis}} = -\left[ \begin{array}{c} 0 \\ \tau_{xz} \\ \tau_{yz} \\ \tau_{zz} \\ u\tau_{xz}+v\tau_{yz}+w\tau_{zz}+\kappa T_y \end{array} \right], \end{equation} where $\tau_{ij}$ is shear stress tensor which is related to velocity gradients as $\tau_{ij} = \mu (u_{i,j}+u_{j,i}) + \lambda\delta_{ij}u_{k,k}$, $\mu$ is dynamic viscosity, $\lambda=-2/3\mu$ based on Stokes' hypothesis, $\delta_{ij}$ is the Kronecker delta, $\kappa$ is thermal conductivity, $T$ is temperature that is related to density and pressure through the ideal gas law $p=\rho \mathcal{R} T$, where $\mathcal{R}$ is the gas constant. \subsection{The computational equations} As will be discussed later, we map each moving grid element from the physical space to a stationary standard cubic element in the computational space where the equations are solved. Assume the mapping is: $t=\tau$, $x=x(\tau,\xi,\eta,\zeta)$, $y=y(\tau,\xi,\eta,\zeta)$ and $z=z(\tau,\xi,\eta,\zeta)$, where $(\tau,\xi,\eta,\zeta)$ are the computational time and space. It can be shown that the flow equations will take the following conservative form in the computational space, \begin{equation} \frac{\partial \widetilde{\mathbf{Q}}} {\partial t} + \frac{\partial \widetilde{\mathbf{F}}} {\partial \xi} + \frac{\partial \widetilde{\mathbf{G}}} {\partial \eta} + \frac{\partial \widetilde{\mathbf{H}}} {\partial \zeta} = \boldsymbol{0}, \label{eq:computational} \end{equation} and the computational variable and fluxes are related to the physical ones through \begin{equation} \left[ \begin{array}{c} \widetilde{\mathbf{Q}} \\ \widetilde{\mathbf{F}} \\ \widetilde{\mathbf{G}} \\ \widetilde{\mathbf{H}} \end{array} \right] = |\mathcal{J}|\mathcal{J}^{-1} \left[ \begin{array}{c} \mathbf{Q} \vphantom{\widetilde{\mathbf{Q}}} \\ \mathbf{F} \vphantom{\widetilde{\mathbf{F}}} \\ \mathbf{G} \vphantom{\widetilde{\mathbf{G}}} \\ \mathbf{H} \vphantom{\widetilde{\mathbf{H}}} \end{array} \right], \label{eq:transform_relation} \end{equation} where $|\mathcal{J}|$ is determinant of the Jacobian matrix and $\mathcal{J}^{-1}$ is the inverse Jacobian matrix, and their expressions are \begin{gather} |\mathcal{J}| = \left| \frac{\partial(t, x,y,z)}{\partial(\tau, \xi,\eta,\zeta)} \right| = \left| \begin{array}{cccc} 1 & 0 & 0 & 0 \\ x_t & x_{\xi} & x_{\eta} & x_{\zeta} \\ y_t & y_{\xi} & y_{\eta} & y_{\zeta} \\ z_t & z_{\xi} & z_{\eta} & z_{\zeta} \end{array} \right|, \\ \mathcal{J}^{-1} = \frac{\partial(\tau, \xi, \eta, \zeta)}{\partial(t,x,y,z)} = \left[ \begin{array}{cccc} 1 & 0 & 0 & 0 \\ \xi_t & \xi_x & \xi_y & \xi_z \\ \eta_t & \eta_x & \eta_y & \eta_z \\ \zeta_t & \zeta_x & \zeta_y & \zeta_z \\ \end{array} \right]. \end{gather} Besides the flow equations, the Geometric Conservation Law (GCL) \cite{thomas-1979} also needs to be numerically satisfied to ensure free-stream preservation on moving grids. The GCL equations and the steps for solving them are described in our previous papers \cite{zhang-2015b,zhang-2016c}. \subsection{The flux reconstruction method} The meshes in this work consist of hexahedral elements only. The first step of the FR method is to map each hexahedral element to a standard cubic element of unit size, i.e., $0\leq \xi, \eta, \zeta \leq 1$. This can be done using the following iso-parametric mapping, \begin{equation} \left[ \begin{array}{c} x \\ y \\ z \end{array} \right] = \sum_{i=1}^{K} M_i(\xi,\eta,\zeta) \left[ \begin{array}{c} x_i(t) \\ y_i(t) \\ z_i(t) \end{array} \right], \label{eq:mapping} \end{equation} where $K$ is the number of nodes of a hexahedral element, $M_i$ (detailed expressions can be found in, for example, \cite{bathe-2006}) is the shape function of the $i$-th node, and $(x_i,y_i,z_i)$ are the coordinates of the $i$-th node. Next, solution points (SPs, denoted by $X_s$) and flux points (FPs, denoted by $X_f$) are defined along each coordinate direction in the standard element. Fig. \ref{fig:spfp} shows a schematic of the distribution of the SPs and FPs in the $\xi$-$\eta$ plane for a fourth-order FR method. Generally, for an $N$-th order FR scheme, there are $N$ SPs and FPs in each direction, where the SPs are in the interior and the FPs are on the boundaries of the standard element. The SPs and FPs are chosen as the Legendre points in this work. \begin{figure}[H] \centering \includegraphics[width=2in]{./figures/spfp.pdf} \caption{Schematic of solution points (circles) and flux points (squares) in the $\xi\textnormal{-}\eta$ plane for a fourth-order FR method.} \label{fig:spfp} \end{figure} Subsequently, Lagrange interpolation bases are defined at each SP. For example, at the $i$-th SP we have \begin{equation} h_i(X) = \prod_{s=1,s\neq i}^{N}\left(\frac{X-X_s}{X_i-X_s}\right). \label{eq:hbasis} \end{equation} The resulting bases also form a basis of polynomials of degrees less than or equal to $N-1$, i.e., $\boldsymbol{\mathsf{P}}_{N\!-\!1}$. These interpolation bases allow the construction of solution and flux polynomials inside each element through tensor products. For example, \begin{align} \widetilde{\mathbf{Q}}(\xi,\eta,\zeta) &= \sum_{k=1}^{N} \sum_{j=1}^{N} \sum_{i=1}^{N} \widetilde{\mathbf{Q}}_{ijk} h_i(\xi) h_j(\eta) h_k(\zeta), \label{eq:Qt} \\ \widetilde{\mathbf{F}}(\xi,\eta,\zeta) &= \sum_{k=1}^{N} \sum_{j=1}^{N} \sum_{i=1}^{N} \widetilde{\mathbf{F}}_{ijk} h_i(\xi) h_j(\eta) h_k(\zeta), \label{eq:Ft} \end{align} where the subscript $ijk$ denotes the discrete value at the $ijk$-th SP. All these polynomials are in $\boldsymbol{\mathsf{P}}_{N\!-\!1,N\!-\!1,N\!-\!1}$. The above solution and flux polynomials are continuous within each element, but discontinuous across cell boundaries. Therefore, common values need to be defined on cell boundaries. There are various ways to calculate these common values. In this work, the common solution is calculated as the average of the discontinuous values from the two sides of a boundary; the common inviscid fluxes are computed using the Rusanov Riemann solver\cite{rusanov-1961}; the common viscous fluxes are computed from the common solutions and common gradients. There is one more issue: after taking the first-order spatial derivatives in the governing equations, the three flux terms become elements in $\boldsymbol{\mathsf{P}}_{N\!-\!2,N\!-\!1,N\!-\!1}$, $\boldsymbol{\mathsf{P}}_{N\!-\!1,N\!-\!2,N\!-\!1}$, and $\boldsymbol{\mathsf{P}}_{N\!-\!1,N\!-\!1,N\!-\!2}$, respectively, and are inconsistent with the solution term. To fix this issue, the original fluxes need to be reconstructed. This is done by using correction functions that are polynomials of degree no less than $N$. Taking the flux in the $\xi$ direction as an example, the reconstructed flux polynomials is \begin{equation} \widehat{\mathbf{F}}(\xi,\eta,\zeta) = \widetilde{\mathbf{F}}(\xi,\eta,\zeta) + \big[\widetilde{\mathbf{F}}^\text{com}(0,\eta,\zeta) - \widetilde{\mathbf{F}}(0,\eta,\zeta)\big] \cdot g_\text{\tiny L}(\xi) + [\widetilde{\mathbf{F}}^\text{com}(1,\eta,\zeta) - \widetilde{\mathbf{F}}(1,\eta,\zeta)] \cdot g_\text{\tiny R}(\xi) \end{equation} where $\widetilde{\mathbf{F}}(\xi,\eta,\zeta)$ is from (\ref{eq:Ft}); $\widetilde{\mathbf{F}}^{\text{com}}$ represents the common flux on a cell boundary; $g_\text{\tiny L}$ and $g_\text{\tiny R}$ are the left and right correction functions, and are required to at least satisfy \begin{equation} \begin{alignedat}{2} g_\text{\tiny L}(0) &= 1, \quad g_\text{\tiny L}(1) &&= 0, \\ g_\text{\tiny R}(0) &= 0, \quad g_\text{\tiny R}(1) &&= 1, \end{alignedat} \end{equation} which ensures that \begin{equation} \widehat{\mathbf{F}}(0,\eta,\zeta) = \widetilde{\mathbf{F}}^\text{com}(0,\eta,\zeta), \quad \widehat{\mathbf{F}}(1,\eta,\zeta) = \widetilde{\mathbf{F}}^\text{com}(1,\eta,\zeta), \end{equation} i.e., the reconstructed fluxes still take the common values at cell interfaces. In this work, we have employed the $g_{\text{\tiny DG}}$ correction function \cite{huynh-2007}. The other two fluxes are reconstructed in the same way. Finally, the governing equations can be written in the following residual form, \begin{equation} \left. \frac{\partial \widetilde{\mathbf{Q}}} {\partial t} \right|_{ijk} = - \left[ \frac{\partial \widehat{\mathbf{F}}}{\partial\xi} + \frac{\partial \widehat{\mathbf{G}}}{\partial\eta} + \frac{\partial \widehat{\mathbf{H}}}{\partial\zeta} \right]_{ijk} = \mathfrak{R}_{ijk}, \quad i,j,k=1,2,\cdots,N, \label{eq:resid} \end{equation} where $\mathfrak{R}_{ijk}$ is the residual at the $(i,j,k)$-th SP. This system can be time marched using either explicit or implicit temporal schemes. \subsection{A sliding-mesh method} In three-dimensions, we consider two types of sliding interfaces as depicted in Fig. \ref{fig:sliding_mesh}: one is annular, and the other is cylindrical. To simplify the explanation, we have required the mesh to only unmatch in the azimuthal direction but match in the radial (for annular sliding) or axial (for cylindrical sliding) direction. We also require equal mesh size in the azimuthal direction. These restrictions are imposed for explanation purposes only and can be easily lifted in practice. More details can be found in our previous papers \cite{zhang-2015a,zhang-2015b,zhang-2016a,zhang-2016c,zhang-2018,zhang-2020a}. \begin{figure}[H] \centering \includegraphics[width=1.6in]{./figures/mortar_msh1.pdf} \qquad\qquad \includegraphics[width=1.6in]{./figures/mortar_msh2.pdf} \caption{Two types of sliding meshes: left, annular sliding; right, cylindrical sliding.} \label{fig:sliding_mesh} \end{figure} Since the SPs on the two sides of a nonconforming sliding interface do not match, we aim to find the best-possible common values of a variable on the two sides of a sliding interface. This can be achieved by least-squares projections using mortar elements as the intermediate medium. A mortar element is formed by the overlapping region of two cell faces. Taking the cylindrical sliding interface as an example, based on the assumptions we have made, a cell face has two mortar elements as sketched in Fig. \ref{fig:mortar_map}. The first step is to map a cell face and the mortars to standard ones as shown in the same figure, using, for example, iso-parametric mapping or transfinite mapping. \begin{figure}[H] \centering \includegraphics[width=4.5in]{./figures/mortar_map.pdf} \caption{Map curved cell face and mortars to straight ones.} \label{fig:mortar_map} \end{figure} Let the $(\xi',\eta')$ denote the mortar space, then the computational and the mortar spaces are related as \begin{equation} \xi =o + s \xi', ~~~ \eta = \eta', \label{eq:xz} \end{equation} where $0\le \xi, \eta, \xi', \eta' \le 1$, and $o$ and $s$ are the offset and scaling of a mortar with respect to a cell face. Let $\phi$ represent the variable of interest, and obviously it can be represented by the following polynomials on a cell face $\Omega$ and on the left side of a mortar $\Xi$, \begin{align} \phi^{\Omega}(\xi,\eta) &= \sum_{j=1}^{N} \sum_{i=1}^{N} \phi_{ij}^{\Omega} h_i(\xi) h_j(\eta), \label{eq:phi_face} \\ \phi^{\Xi,L}(\xi',\eta') &= \sum_{j=1}^{N} \sum_{i=1}^{N} \phi_{ij}^{\Xi,L} h_i(\xi') h_j(\eta'), \label{eq:phi_mortar} \end{align} where $\phi_{ij}^{\Omega}$ and $\phi_{ij}^{\Xi,L}$ are the discrete values at the $(i,j)$-th SP on $\Omega$ and the left side of $\Xi$, respectively. The $(\phi_{ij}^{\Xi,L})$'s are unknown, and can be obtained through the following projection (refer to Fig. \ref{fig:mortar_prj}(a)), \begin{equation} \int_0^1 \int_0^1 (\phi^{\Xi,L}(\xi',\eta') - \phi^{\Omega}(\xi,\eta)) h_\alpha(\xi')h_\beta(\eta') \text{d}\xi' \text{d}\eta' = 0,~ \forall\, \alpha,\beta=1,2,...,N. \label{eq:prj1a} \end{equation} Considering the relations in (\ref{eq:xz}), it can be shown that the above two-dimensional projection is equivalent to the following one-dimensional one, \begin{equation} \int_0^1 (\phi^{\Xi,L}(\xi',X_j) - \phi^{\Omega}(\xi,X_j)) h_\alpha(\xi') \text{d} \xi' = 0,~ \forall\, \alpha=1,2,...,N, \label{eq:prj1b} \end{equation} where $X_j$ is the coordinate of the $j$-th SP. Evaluating the above equation for all the $\alpha$'s, we will get a system of equations about $\phi_{1:N,j}^{\Xi,L}$. Repeating this process for every $j$, we will obtain every $\phi_{ij}^{\Xi,L}$. The values on the right side of a mortar, i.e., the $(\phi_{ij}^{\Xi,R})$'s, can be obtained in the same way. \begin{figure}[H] \centering \includegraphics[width=4.5in]{./figures/mortar_prj.pdf} \caption{Projection between face and mortar: (a) from face to the left side of a mortar, (b) from two mortars back to a face.} \label{fig:mortar_prj} \end{figure} After that, we are able to compute a common value on the mortar, e.g., through averaging or Riemann solver. Let us denote this common value as $\Phi^\Xi$. We then project this common variable back to a cell face from mortars as demonstrated in Fig. \ref{fig:mortar_prj}(b). And the projection is \begin{equation} \begin{split} \int_{\xi=0}^{\xi=o_2} & \int_{\eta=0}^{\eta=1} (\Phi^{\Omega}(\xi,\eta) - \Phi^{\Xi_1}(\xi',\eta')) h_\alpha(\xi) h_\beta(\eta) \text{d}\xi \text{d}\eta \\ &+ \int_{\xi=o_2}^{\xi=1} \int_{\eta=0}^{\eta=1} (\Phi^{\Omega}(\xi,\eta) - \Phi^{\Xi_2}(\xi',\eta')) h_\alpha(\xi) h_\beta(\eta) \text{d}\xi \text{d}\eta = 0, ~ \forall\, \alpha,\beta=1,2,...,N, \label{eq:prj2a} \end{split} \end{equation} where $\Phi^\Omega$ represents the polynomial of the unknown common variable on face $\Omega$. Similarly, this projection is equivalent to the following one-dimensional projection, \begin{equation} \int_{0}^{o_2} (\Phi^{\Omega}(\xi,X_j) - \Phi^{\Xi_1}(\xi',X_j)) h_\alpha(\xi) \text{d}\xi + \int_{o_2}^{1} (\Phi^{\Omega}(\xi,X_j) - \Phi^{\Xi_2}(\xi',X_j)) h_\alpha(\xi) \text{d}\xi = 0, ~ \forall\, \alpha =1,2,...,N, \label{eq:prj2b} \end{equation} from which $\Phi_{1:N,j}^{\Omega}$ are obtained, and then also every $\Phi_{ij}^{\Omega}$ by repeating this process for $j$. \section{Simulation setup} \label{sec:setup} The propeller studied in this work is the DTMB 4119 model designed at the David Taylor Model Basin \cite{denny-1968,jessup-1984,jessup-1989}. Its parameters are summarized in Tab. \ref{tab:propellerparam}, and the geometry is given in Tab. \ref{tab:propellergeo}, where $r$ represents radial position, $R=D/2$ is propeller radius, $c$ is blade section chord length, $P$ is section pitch, $\phi_P$ is section pitch angle, $t$ is section thickness, and $f$ is section camber. \begin{table}[H] \setlength{\tabcolsep}{2mm} \centering \begin{tabular}{>{\raggedright}m{5.0cm} >{\centering\arraybackslash}m{3cm}} \hline parameter & value \\ \hline Number of blades $Z$ & 3 \\ Diameter $D$ [m] & 0.305 \\ Hub diameter ratio $D_h/D$ & $0.2$ \\ Design advance ratio $J$ & 0.833 \\ Rotation & Right handed \\ Section thickness & NACA66 modified \\ Section mean line & NACA, $a=0.08$ \\ \hline \end{tabular} \caption{Design parameters of the DTMB 4119 propeller.} \label{tab:propellerparam} \end{table} \begin{table}[H] \setlength{\tabcolsep}{3mm} \centering \begin{tabular}{m{7mm} m{1cm} m{1cm} m{9mm} m{1.2cm} m{1.2cm}} \hline $r/R$ & ~~$c/D$ & $P/D$ & $\phi_P$ [$^\circ$] & ~~~$t/c$ & ~~~$f/c$ \\ \hline 0.2 & 0.3200 & 1.105 & 60.38 & 0.20550 & 0.01429 \\ 0.3 & 0.3625 & 1.102 & 49.47 & 0.15530 & 0.02318 \\ 0.4 & 0.4048 & 1.098 & 41.15 & 0.11800 & 0.02303 \\ 0.5 & 0.4392 & 1.093 & 34.84 & 0.09016 & 0.02182 \\ 0.6 & 0.4610 & 1.088 & 29.99 & 0.06960 & 0.02072 \\ 0.7 & 0.4622 & 1.084 & 26.24 & 0.05418 & 0.02003 \\ 0.8 & 0.4347 & 1.081 & 23.28 & 0.04206 & 0.01967 \\ 0.9 & 0.3613 & 1.079 & 20.88 & 0.03321 & 0,01817 \\ 0.95 & 0.2775 & 1.077 & 19.84 & 0.03228 & 0.01631 \\ 1.0 & 0.0 & 1.075 & 18.89 & 0.03160 & 0.01175 \\ \hline \end{tabular} \caption{Geometry of the DTMB 4119 propeller.} \label{tab:propellergeo} \end{table} When the geometry is given, a propeller flow is governed by two nondimensional parameters: advance ratio and Reynolds number. The advance ratio is defined as \begin{equation} J = \frac{U_\infty}{n D}, \end{equation} where $U_\infty$ is the incoming flow speed, $n$ is propeller's revolution per second (RPS), and $D$ is propeller diameter. The following Reynolds number is usually adopted in experimental studies of marine propellers \begin{equation} Re_c = \frac{c_{0.7}\, U_{0.7}}{\nu} = \frac{c_{0.7} \sqrt{U_\infty^2 + (2\pi 0.7 R n)^2}}{\nu}, \end{equation} where $c_{0.7}$ and $U_{0.7} = \sqrt{U_\infty^2 + (2\pi 0.7 R n)^2}$ are the chord length and the relative speed, respectively, at $r/R=0.7$, and $\nu$ is fluid kinematic viscosity. For numerical simulations, it is more convenient to define the Reynolds number as, \begin{equation} Re_D = \frac{D \, U_\infty}{\nu}. \end{equation} It can be shown that these two Reynolds numbers are conveniently convertible through the following relation, \begin{equation} Re_D = \frac{Re_c}{\frac{c_{0.7}}{D} \sqrt{1 + \left(\frac{0.7\pi}{J}\right)^2}}. \end{equation} Depending on $c_{0.7}/D$ and $J$, $Re_D$ could be either larger or smaller than $Re_c$, but usually not much different. The geometry is visualized in Fig. \ref{fig:geo_ellip}. As a typical screw-type propeller, DTMB 4119 consists of four components: the shaft, the blades, the hub, and the fairwater. All components, except the shaft, rotate at an angular speed $\omega$. Hub and fairwater are usually not part of propeller design. In this work, the hub has a length $L_h=0.5D$, with the three blades installed evenly along the circumferential direction of the mid-hub. The fairwater in the figure is a $1:2$ ellipsoid, but cylindrical and hemispherical fairwaters are also employed in a later section to study their effects. \begin{figure}[H] \centering \includegraphics[scale=0.9]{./figures/ellip_geo.pdf} \caption{Two views of the DTMB 4119 propeller: left, side view; right, back view (pressure side).} \label{fig:geo_ellip} \end{figure} The overall computational domain is cylindrical as shown in Fig. \ref{fig:ellip_dom}. It has a length of $15D$ in the streamwise (i.e., $x$) direction and a diameter of $12D$. The resulting blockage ratio of this domain with respect to the propeller is $0.69\%$, which is small enough to guarantee negligible confinement effects according to the study in \cite{wilson-1994}. The propeller locates $3D$ downstream from the inlet, with the shaft extends all the way to the inlet. The blades and the hub are enclosed in a small sliding disk region whose radius and thickness are $0.75D$ and $0.5D$, respectively. A global view of the mesh (with a $1/4$ cutout to expose the propeller) is shown on the right of Fig. \ref{fig:ellip_dom}, where the propeller and the sliding interfaces are colored in red. \begin{figure}[H] \centering \includegraphics[width=2.7in]{./figures/ellip_domain.pdf} \qquad \includegraphics[width=2.7in]{./figures/ellip_msh_global.pdf} \caption{Overall computational domain (left) and global view of the mesh (right).} \label{fig:ellip_dom} \end{figure} The overall mesh consists of about $235,000$ quadratic curved hexahedral elements, of which about $36,000$ are within the small sliding disk region. The mesh is refined around the propeller as well as in the wake region. The first layer of the off wall elements on each blade surface has a height of approximately $0.015D$, and the first off wall solution point is about $0.0007D$ (for the fifth-order scheme) away from the walls. Two local views of the meshes on the sliding interfaces and the blade surfaces are shown in Fig. \ref{fig:ellip_msh}. We see that the high-order curved mesh captures the curvatures of the blade surfaces very well. \begin{figure}[H] \centering \includegraphics[width=2.5in]{./figures/ellip_msh_local1.pdf} \qquad \includegraphics[width=2.5in]{./figures/ellip_msh_local2.pdf} \caption{Close views of the mesh on the sliding interfaces (left) and on the propeller surfaces (right).} \label{fig:ellip_msh} \end{figure} We treat the inlet as a Dirichlet boundary, the outer cylindrical surface and the outlet as characteristic farfields that allow waves and flow to leave without reflection \cite{jameson-1983}, and all solid surfaces as no-slip adiabatic walls. The incoming freestream flow has a low Mach number of $Ma_\infty=0.05$ so that compressibility effects are small. The Reynolds number is $Re_D=5.59\times 10^5$, which is equivalent to $Re_c=7.3\times 10^5$. Various advance coefficients are studied but with a focus on the design value (i.e., $J=0.833$). The nondimensional angular speed of the propeller is $\omega^*=\omega D/U_\infty = 2\pi n D/U_\infty = 2\pi/J$, and the rotation period is $t U_\infty/D=J$. The simulations are performed in two ways: to collect time-averaged statistics, the outer subdomain is fixed, only the inner sliding region rotates at $\omega$, and a velocity boundary condition is applied on the fairwater surface; to collect phase-averaged statistics, the whole domain rotates at angular speed $\omega$, and velocity boundary condition is applied to the shaft to make it stationary. A four-stage third-order SSP-RK scheme \cite{spiteri-2002} with a nondimensional time step size of $\Delta t U_\infty/D = 2.5\times 10^{-5}$ is adopted for the time marching. The propeller therefore rotates about $0.01$ degree per time step at the design condition. For spatial discretization, it is known that high-order simulation of turbulent flow may experience instabilities due to aliasing errors \cite{jameson-2012}. We observed such instabilities on the fifth- and above orders. To overcome this issue, we have employed the filter reported in \cite{fischer-2001} (with strength $\alpha=0.05$) to stabilize the simulations. Meanwhile, we compared the design-condition results from the fourth-, the fifth-, and the sixth-order schemes to ensure sufficient resolution. It was observed that the mean loads from the fourth- and the fifth-order schemes have small differences (around $3\%$), but the fifth-order scheme resolves the flow structures with more details. On the other hand, both the mean loads and the flow fields from the fifth- and the sixth-order schemes are almost indistinguishable, which indicates that the fifth-order is the optimum choice considering both accuracy and cost. For this reason, the fifth-order scheme has been used for the simulations in what follows. It is also worth mentioning that all simulations were run for a nondimensional time of $tU_\infty/D=100$, and phase- and time-averaging were performed on the last $85$ time units, which represents approximately 102 revolutions at the design condition. \section{Results and discussion} \label{sec:results} \subsection{Propeller loads} \label{sec:loads} The loads on a propeller are measured by the thrust and torque coefficients defined as below, \begin{equation} K_T = \frac{T}{\rho n^2 D^4} \qquad\text{and}\qquad K_Q = \frac{Q}{\rho n^2 D^5}, \end{equation} where $T$ and $Q$, respectively, represent the force and the torque exerted by the fluid on a propeller in the axial direction. The efficiency of a propeller is defined as \begin{equation} \eta = \frac{TU_\infty}{2\pi n Q} = \frac{J}{2\pi} \cdot \frac{K_T}{K_Q}, \end{equation} where the variables have the same meanings as previously explained. We first compare the loads predicted by the simulation with those measured in a previous experiment \cite{jessup-1989} at various advance ratios. As can be seen from Fig. \ref{fig:loads}, the present numerical results agree very well with the experimental values under all working conditions: overloading condition (when $J$ is small), near-design condition, and underloading condition (when $J$ is large). The maximum difference is observed on the efficiency curve, which is around $5\%$. The present simulation predicts the highest efficiency around $J=0.9$, which is close to the design condition that is $J=0.833$ (see Tab. \ref{tab:propellerparam}). The experiment, however, shows an optimum performance slightly above $J=1.0$, which is further away from the design condition. Meanwhile, in the same figure, we also compare the present results with the latest (also the best available) results on the same propeller from a low-order simulation \cite{sezen-2019} using the commercial software STAR-CCM+. It is evident that the low-order solver predicts the efficiency well only in a narrow range of working conditions where $J$ is small. When $J$ increases, the low-order prediction becomes much worse. As will be shown in the next section that when $J$ increases, the flow vortices become weaker, which are more vulnerable to numerical dissipations. Since the high-order method introduces much smaller numerical dissipations, it predicts the loads very accurately under all conditions. In contrast, the large numerical dissipations of the low-order method have completely demolished the predictions when $J$ is large. \begin{figure}[H] \centering \includegraphics[width=4.5in]{./figures/ellip_loads.pdf} \caption{The mean blade loads of DTMB 4119 at different working conditions.} \label{fig:loads} \end{figure} Since a propeller works around the design condition most of the time, therefore a detailed look into this condition is presented in what follows. The instantaneous $K_T$ and $K_Q$ of the blades at this condition are plotted in Fig. \ref{fig:ktkq} from $tU_\infty/D=15$ to $45$ for about $36$ revolutions. The two coefficients are seen to fluctuate at small amplitudes about their means chaotically due to the turbulent nature of the flow. The mean values (averaged for about 102 revolutions) are compared with the design \cite{denny-1968} and the experimental \cite{jessup-1984,jessup-1989} values in Tab. \ref{tab:ktkq_mean}, where the difference is defined as (simulation/experiment$-1$)$\times 100\%$. The design was based on potential flow theories. The two experiments were performed in open water at slightly different Reynolds numbers. The present Reynolds number is chosen to exactly match that of \cite{jessup-1984}, but it is obvious that the Reynolds number effects are small at this level. Overall, we see very good agreements between the simulation and the experiments as well as the design. \begin{figure}[H] \centering \includegraphics[width=5.8in]{./figures/ellip_KtKq.pdf} \caption{Instantaneous loads on the blades of DTMB 4119 at design condition.} \label{fig:ktkq} \end{figure} \begin{table}[H] \setlength{\tabcolsep}{2mm} \centering \begin{tabular}{>{\centering}m{1.8cm} >{\centering}m{1.9cm} >{\centering}m{1.8cm} >{\raggedleft}m{2.7cm} >{\raggedleft}m{2.7cm} >{\centering\arraybackslash}m{2.0cm}} \hline ~ & $Re_D~(\times 10^5)$ & $Re_c~(\times 10^5)$ & $\overline{K}_T$ ~~~(diff.) & $\overline{K}_Q$ ~~~(diff.) & $\overline{\eta}$ \\ \hline present & 5.59 & 7.3 & 0.1514 \hphantom{(-$2.3\%$)} & 0.0274 \hphantom{(-$2.3\%$)} & 0.7326 \\ design \cite{denny-1968} & - & - & 0.1540 (-$1.7\%$) & 0.0290 (-$5.5\%$) & 0.7040 \\ ~\! exp.~ \cite{jessup-1984} & 5.59 & 7.3 & 0.1500 (~$0.9\%$) & 0.0285 (-$3.9\%$) & 0.6978 \\ ~\! exp.~ \cite{jessup-1989} & 7.66 & 10.0 & 0.1460 (~$3.7\%$) & 0.0280 (-$2.1\%$) & 0.6913 \\ \hline \end{tabular} \caption{Mean loads on the blades of DTMB 4119 at design condition.} \label{tab:ktkq_mean} \end{table} The simulation allows us to study the loads on different parts of the propeller. We have summarized the results in Tab. \ref{tab:ktkq_mean2}, where ``mean'' denotes the time-averaged values, ``r.m.s'' denotes the root-mean-square of the unsteady components, and the subscripts ``$p$'' and ``$v$'' denote pressure and viscous contribution, respectively. Before proceeding further, it is worth noting that the thrust on the fairwater needs to be treated carefully. Unlike the blades which are two-sided and closed, the fairwater is ``one-sided'', i.e., it has no direct upstream counterpart to balance the pressure force on its outer surface. In real applications, this force will be partially compensated by the force on an upstream surface with the same cross-sectional area (for example, part of the hull), resulting in a much smaller net force. The flow in the upstream region is usually at reduced speed (or even at stagnation) with higher pressure than that of the freestream. Thus, an underestimated pressure force on this upstream surface is $p_\infty(\pi D_h^2/4)$, where $D_h$ is the hub diameter. We then subtract this force from that on the fairwater to approximate the net thrust on the fairwater. From Tab. \ref{tab:ktkq_mean2}, we see that only the blades experience a thrust (positive $K_T$), while the hub and the fairwater experience drags (negative $K_T$). On the other hand, all three parts experience positive torques. For the blades, pressure contribution dominates the loads; the {r.m.s.} values are at least three magnitudes smaller than the means, suggesting quasi-steady loads. The $K_T$ and $K_Q$ of the hub are four magnitudes smaller than those of the blades, which indicates that the hub has negligible contribution/effect to the overall performance of the propeller. Furthermore, viscous contribution dominates the hub loads, which is consistent with the fact that the hub has no projection in the axial direction and thus pressure has no way to contribute. In fact, pressure should ideally have zero contribution to the hub loads, and the present very small pressure contribution implies very small geometric imperfection of the hub, which obviously has benefited from the high-order curved representation of the geometry. For the fairwater, the drag is about $2.1\%$ of the thrust on the blades, and pressure dominates; the torque is negligibly small, and viscosity dominates due to the geometric symmetry. Overall, we conclude that the thrust and the torque of the blades, and the drag of the fairwater are the main factors that affect the propeller's performance. \begin{table}[H] \setlength{\tabcolsep}{2mm} \centering \begin{tabular}{>{\centering}m{1.4cm} m{1.0cm} >{\centering}m{1.3cm} >{\centering}m{1.3cm} >{\centering}m{1.3cm} >{\centering}m{1.3cm} >{\centering}m{1.3cm} >{\centering\arraybackslash}m{1.3cm}} \hline & & $K_T$ & $K_{T,p}$ & $K_{T,v}$ & $K_Q$ & $K_{Q,p}$ & $K_{Q,v}$ \\ \hline \multirow{2}{*}{blades} & mean & ~0.1514 & ~0.1519 & -4.9E-4 & 0.0274 & 0.0272\hphantom{2} & 2.4E-4 \\ & r.m.s. & ~3.6E-4 & ~3.6E-4 & ~5.6E-7 & 7.0E-5 & 7.0E-5\hphantom{2} & 2.6E-8 \\ \hline \multirow{2}{*}{hub} & mean & -8.5E-5 & -6.7E-7 & -8.5E-5 & 3.9E-6 & 3.1E-8\hphantom{2} & 3.9E-6 \\ & r.m.s. & ~2.1E-7 & ~4.5E-8 & ~2.0E-7 & 2.2E-8 & 5.1E-9\hphantom{2} & 2.2E-8 \\ \hline \multirow{2}{*}{fairwater} & mean & -3.2E-3 & -3.2E-3 & -2.4E-5 & 8.8E-7 & 6.9E-10 & 8.8E-7 \\ & r.m.s. & ~1.2E-4 & ~1.2E-4 & ~4.8E-7 & 9.9E-8 & 9.3E-8\hphantom{2} & 3.2E-8 \\ \hline \end{tabular} \caption{Loads on different parts of DTMB 4119 with an ellipsoidal fairwater at design condition.} \label{tab:ktkq_mean2} \end{table} The time and frequency scales of the major loads can be identified through the autocorrelation and the power spectral density (PSD) curves in Fig. \ref{fig:spectra}. It is worth mentioning that the torque and the thrust of a blade have almost identical characteristics, the curves for the torque are therefore not repeated here. The autocorrelation is defined as $\rho(\tau)=R(\tau)/R(0)$, where $R(\tau)=\left<T(t)T(t+\tau)\right>$ is the autocovariance with a time lag $\tau$, and ``$\left<~\right>$'' denotes ensemble average. We calculate the PSD using Welch's method with a $50\%$ overlapping Hanning window and then average the results over 102 revolutions. The narrow main lobe of the autocorrelation of the blade in Fig. \ref{fig:spectra} implies small integral time scales of the unsteadiness around the blade. It also agrees with the corresponding high-frequency peak around $fD/U_\infty \approx 7.9$ on the PSD curve. In contrast, the autocorrelation of the fairwater has a much wider main lobe, which indicates much larger time scales of the dominant unsteadiness around the fairwater. The corresponding PSD curve is rather broadband and is dominated by very low-frequency components. A comparison of the two PSD curves reveals that the fairwater experiences more unsteadiness at very high-frequencies (e.g., $fD/U_\infty > 20$) than the blade does. These time and frequency scales are directly related to flow structures which will be discussed in the next section. \begin{figure}[H] \centering \includegraphics[width=5.4in]{./figures/ellip_spectra.pdf} \caption{Autocorrelation and PSD of the blade thrust and fairwater drag.} \label{fig:spectra} \end{figure} \subsection{Flow fields} The flow fields of a propeller have very distinct flow structures that are of crucial importance to the propeller's performance. This has made studying the formation, mutual interaction, and stability of these flow structures a constant research topic for decades. In this section, we report the details of the flow fields of DTMB 4119, including the vortices, the velocity field, and the pressure field. \subsubsection{Vortices} \label{sec:vortices} The vortical structures in a flow field can be well visualized by isosurfaces of Q-criterion \cite{hunt-1988}. The Q-criterion (denoted by $Q_{cr}$) is defined as the second invariant of the velocity gradient tensor, i.e., $Q_{cr}=(\Omega_{ij}\Omega_{ij} - S_{ij}S_{ij})/2$, where $\Omega_{ij}=(u_{i,j}-u_{j,i})/2$ and $S_{ij}=(u_{i,j}+u_{j,i})/2$ are the antisymmetric and the symmetric component, respectively, of the velocity gradient tensor. When the Reynolds number is given, the only parameter that determines a propeller's flow field is the advance ratio $J$. Figure \ref{fig:ellip_instant_Qcrs} shows the instantaneous vortical structures as $J$ decreases from $1.1$ to $0.4$. Note that the nondimensional rotational speed is related to the advance ratio as $\omega^*\!= 2\pi/J$. Thus, a decreasing $J$ is equivalent to an increasing $\omega^*$. We notice two dramatic changes in the flow field as $J$ decreases: the increase of vortex strength and the occurrence of flow instabilities. At $J=1.1$ and $1.0$, the vortices are so weak that they are quickly dissipated by the wake flow. At $J=0.9$ and $0.8$, the vortices become strong enough to sustain for a long distance in the wake, and a hub vortex is also well established. In addition, up to this point the flow remains stable. Obvious instability occurs when $J$ decreases to $0.7$, and the instability is caused by mutual interactions between two tip vortices around $x/D=4.8$. At $J=0.6$, the instability is still caused by mutual tip vortex interactions, but the occurrence moves upstream to $x/D=3.4$. The occurrence further moves upstream to $x/D=3.2$ and $2.7$, for $J=0.5$ and $0.4$, respectively. However, the cause of the instability becomes more complicated. At $J=0.5$, it seems the instability not only comes from the mutual interaction between the tip vortices, but also the interaction between tip and hub vortices. Finally, at $J=0.4$, it looks like the trailing edge vortices have become strong enough to be the leading cause of the instability. It was conjectured in \cite{kumar-2017} that blade trailing edge vortices are an important source of flow instabilities. Based on our observations here, this is only possible when $J$ is small enough (i.e., propeller is at very high relative rotational speed) and when blade trailing edge vortices are strong enough. \begin{figure}[H] \centering \includegraphics[width=\textwidth]{./figures/ellip_instant_Qcr.pdf} \caption{Isosurfaces of $Q_{cr}D^2/U_\infty^2=40$ at different working conditions.} \label{fig:ellip_instant_Qcrs} \end{figure} We already saw that the flow fields can be very different at different working conditions. In the rest of this paper, we focus on the design condition only. Figure \ref{fig:ellip_instant_Qcr} shows an instantaneous view of the flow structures at this condition. It is seen that the tip vortices are very much equally spaced along the axial direction, with the distance between two successive vortices being approximately $0.36D$, which is about one-third the tip pitch (see Tab. \ref{tab:propellergeo}). Meanwhile, the surface velocity contours reveal that a tip vortex has lower streamwise speed on the outer surface, and higher speed on the inner surface. This indicates that a tip vortex not only revolves helically about the propeller's axes, but also about its own core at the same time. The topology of the root vortices are not very obvious from this instantaneous flow field due to the many small turbulent structures. These high-frequency small flow structures are closer to the fairwater than to the blades, and thus have contributed more to the load unsteadiness of the fairwater than to the blades, which agrees with our previous observation on the PSD curves in Fig. \ref{fig:spectra}. These small structures, however, do not dominate the load unsteadiness, which is likely because of their isotropy that leads to mutual cancellation of the effects. While the tip vortices break up about $5D$ downstream from the propeller, the hub vortex stays strong and does not break up even at the outlet (i.e., $12D$ downstream from the propeller; complete picture not shown here due to limited space). \begin{figure}[H] \centering \includegraphics[width=5.2in]{./figures/ellip_instant_Qcr1.pdf} \\ \includegraphics[width=5.2in]{./figures/ellip_instant_Qcr2.pdf} \caption{Isosurface of instantaneous Q-criterion $Q_{cr}D^2/U_\infty^2=40$ at design condition.} \label{fig:ellip_instant_Qcr} \end{figure} The phase-averaged isosurfaces of Q-criterion are shown in Fig. \ref{fig:ellip_phase_Qcr}, and they reveal the major flow structures, especially the root vortices, more evidently. The phase-averaged hub vortex is still seen to vary along the axial direction. In fact, it is the variations of these big structures that dominate the load unsteadiness of the fairwater. Similarly, the load unsteadiness of the blades is likely dominated by the unsteadiness of the tip and the trailing edge vortices. \begin{figure}[H] \centering \includegraphics[width=5.2in]{./figures/ellip_phase_Qcr1.pdf} \\ \includegraphics[width=5.2in]{./figures/ellip_phase_Qcr2.pdf} \caption{Isosurface of phase-averaged Q-criterion $\left<Q_{cr}\right>D^2/U_\infty^2=40$ at design condition.} \label{fig:ellip_phase_Qcr} \end{figure} The FR method is a discontinuous-type of method, and reconstruction must be performed to make the solutions and fluxes globally continuous. This rule also applies to the statistics, which are only element-wise continuous unless reconstructed. However, reconstructing the statistics will impose extra computational and memory cost to the simulation. For this reason, this process is not performed in this work, which results in non-smoothness across cell boundaries as can be seen from Fig. \ref{fig:ellip_phase_Qcr}, especially in the vicinity of the propeller where flow changes rapidly. Nevertheless, this simplification should not alter any of the conclusions here. The time-averaged flow field is shown in Fig. \ref{fig:ellip_time_Qcr}. It is worth noting that time-averaging is impossible for the sliding region due to the movement of the propeller, and the flow in this region is thus not shown in the figure. Over time, the tip vortices form a slightly converging-diverging ``duct'' in space. The root vortices, because of the very small instantaneous turbulent structures, are very difficult to converge in time. Nevertheless, they still have a tube-like shape over time in space. Unlike the instantaneous and the phase averaged ones, the time-averaged the hub vortex is very symmetric about the axis, and almost sees no deviation from the axial direction. This clearly demonstrates that time-averaging not only helps remove most of the small unsteadiness, but also the large ones, from the flow field. \begin{figure}[H] \centering \includegraphics[width=5.2in]{./figures/ellip_time_Qcr1.pdf} \\ \includegraphics[width=5.2in]{./figures/ellip_time_Qcr2.pdf} \\ \includegraphics[width=5.2in]{./figures/ellip_time_Qcr3.pdf} \caption{Isosurface of time-averaged Q-criterion $\overline{Q}_{cr}D^2/U_\infty^2=8$ at design condition.} \label{fig:ellip_time_Qcr} \end{figure} The Q-criterion isosurfaces are able to reveal the most coherent vortical flow structures. They are, however, inefficient to expose the weak ones like the trailing edge vortices at the design condition. Additionally, the Q-criterion cannot reveal the sign of a vortex. For this reason, we have plotted the streamwise vorticity contours in Figs. \ref{fig:ellip_phase_vort} and \ref{fig:ellip_phase_vort_yz} to fill these gaps. We can clearly see the footprints of the trailing edge vortices (TEVs) in Fig. \ref{fig:ellip_phase_vort}. One end of each TEV connects to a tip vortex, and the other end connects to the hub or root vortex. As the flow goes downstream, the TEVs tilt more and more towards downstream, which obviously complies with the wake velocity distribution. The signs of the vortices reveal that the tip vortices rotate in the opposite direction to all other vortices as well as the propeller. \begin{figure}[H] \centering \includegraphics[width=5.2in]{./figures/ellip_phase_vort.pdf} \caption{Contours of phase-averaged streamwise vorticity in the central $x$-$y$ plane.} \label{fig:ellip_phase_vort} \end{figure} The vorticity contours in Fig. \ref{fig:ellip_phase_vort_yz} uncover how the vortices develop in the azimuthal and radial directions as the flow travels downstream. The evolution of the TEVs is the most prominent: they are elongated and bent in the azimuthal direction and finally impinge onto the tip vortices. However, because of flow dissipation and the weak strength of the TEVs, these interactions do not destabilize the flow. Again, the change fo the TEVs is closely related to the velocity distribution that will be discussed in the next section. \begin{figure}[H] \centering \includegraphics[width=5.8in]{./figures/ellip_phase_vort_yz.pdf} \caption{Contours of phase-averaged streamwise vorticity in $y$-$z$ planes at different streamwise locations.} \label{fig:ellip_phase_vort_yz} \end{figure} \subsubsection{Velocity field} \label{sec:velocity} We already noticed that the outer and inner surfaces of a tip vortex have different speeds. The tip vortices thus should be well bounded by velocity isosurfaces. To confirm this, we plot two isosurfaces of the phase-averaged streamwise velocity in Fig. \ref{fig:ellip_uiso1}. It is seen that the isosurfaces of $\left<u\right> /U_\infty = 0.9$ and $1.4$ follow the trajectories of the tip vortices very well. Meanwhile, the increasing gap between the two isosurfaces also agrees with the decreasing strength of the tip vortices as the flow moves downstream. \begin{figure}[H] \centering \includegraphics[width=5.2in]{./figures/ellip_phase_uiso1.pdf} \caption{Isosurfaces of phase-averaged streamwise velocity $\left<u\right>/U_\infty=0.9$ (gray) and $1.4$ (yellow).} \label{fig:ellip_uiso1} \end{figure} A closer look of the velocity isosurfaces in the very vicinity of the blades also reveals the formation of the tip vortices. In Fig. 20, we are looking towards downstream at the suction side in (a1) and (a2), and towards upstream at the pressure side in (b1) and (b2). From (a1), it is obvious that each leading edge (LE) decelerates the incoming flow, resulting in a strand of low-speed flow along the LE and finally sheds off around the tip. From (a2), each trailing edge (TE) accelerate the flow and sheds off a strand of high-speed flow slightly below the tip. When these two strands of flow meet and be convected downstream, a helical tip-vortex system is generated. \begin{figure}[H] \centering \includegraphics[width=6.2in]{./figures/ellip_phase_uiso2.pdf} \caption{Isosurfaces of $\left<u\right>/U_\infty=0.9$ (gray) and $1.4$ (yellow): (a1,a2), suction side; (b1,b2), pressure side.} \label{fig:ellip_uiso2} \end{figure} Figure \ref{fig:ellip_u_xy} shows the phase-averaged streamwise velocity contours in the central $x$-$y$ plane (i.e., $z=0$). It is seen that the blade wake is overall accelerated, while the fairwater wake is mostly at reduced speed. The flow immediately downstream of the fairwater has very low speed, suggesting that a bubble is likely formed in this region. The tip vortices show up as local velocity min-max pairs along the outer edge of the slipstream. The footprints of other vortices, such as the TEVs and the hub vortex, are also visible on the velocity contours. \begin{figure}[H] \centering \includegraphics[width=5.2in]{./figures/ellip_phase_u.pdf} \caption{Contours of phase-averaged streamwise velocity in the central $x$-$y$ plane.} \label{fig:ellip_u_xy} \end{figure} The Cartesian velocity can be decomposed into three components: streamwise, radial, and azimuthal, denoted by $u$, $v_r$, and $v_\theta$, respectively. Figure \ref{fig:ellip_uvw_xy} shows the time-averaged contours of these components. We see that the overall slipstream has a converging-diverging shape and is well contained in the propeller's swept area, i.e., $r \le 0.5D$ (note that $r=|y|$ in the central plane). The radial velocity is small almost everywhere, except in the very vicinity of the fairwater. The azimuthal speed is large only in the fairwater wake and is induced by the strong hub vortex. The two low-speed strips on the azimuthal speed contours around $r/D=0.1$ and $0.4 \le x/D \le 1.2$ are footprints of the root vortices. \begin{figure}[H] \centering \includegraphics[width=5.2in]{./figures/ellip_time_u.pdf} \\ \includegraphics[width=5.2in]{./figures/ellip_time_v.pdf} \\ \includegraphics[width=5.2in]{./figures/ellip_time_w.pdf} \caption{Contours of time-averaged streamwise, radial and azimuthal velocity components in the central $x$-$y$ plane.} \label{fig:ellip_uvw_xy} \end{figure} More detailed velocity profiles are shown in Fig. \ref{fig:ellip_time_uvw1d}. We see almost no induced velocity outside the slipstream (i.e., $r/D>0.5$) on all profiles, which suggests that the propeller introduces very little disturbance to the flow outside its swept area. The streamwise velocity $\overline{u}$ reaches its maximum values in the region $0.2 < r/D < 0.4$, where the flow is accelerated by more than $30\%$ over $U_\infty$. Outside this region, $\overline{u}$ quickly decreases to $U_\infty$ around $r/D=0.5$, and also decreases to the hub-vortex-core speed at $r/D=0$. This velocity distribution leads to the TEV deformation that is observed in Fig. \ref{fig:ellip_phase_vort}. The streamwise hub-vortex-core speed is close to zero in the near wake region (e.g., $x/D=0.46$), and then consistently increases towards downstream. At $x/D=5.0$, it is already slightly over $U_\infty$. The radial velocity $\overline{v}_r$ is only noticeable in the near wake, for example, at $x/D=0.46$ and $0.5$, and then quickly drops to very small values as the flow travels downstream. The azimuthal velocity $\overline{v}_\theta$ is mostly induced by the hub vortex. It has almost the same profile at different locations, except in two small regions: one around $r/D=0.1$ that is affected by the root vortices, and the other around $r/D=0.42$ that is affected by the tip vortices. The profiles of $\overline{v}_\theta$ are very close to that of a typical Rankine vortex, and is responsible for the TEV deformation in Fig. \ref{fig:ellip_phase_vort_yz}. The maximum value of $\overline{v}_\theta$ is about $0.62U_\infty$ in the near field, and $0.36U_\infty$ in the intermediate wake. These are very large values, and clearly indicate how strong the hub vortex is. \begin{figure}[H] \centering \includegraphics[height=2.65in]{./figures/ellip_time_u1d.pdf} \includegraphics[height=2.65in]{./figures/ellip_time_vw1d.pdf} \caption{Time-averaged velocity profiles at different streamwise locations in the central plane.} \label{fig:ellip_time_uvw1d} \end{figure} As a further validation of the simulation, in Fig. \ref{fig:ellip_time_vcompare} we compare the velocity profiles at $x/R=0.951$ with a water tunnel measurement from \cite{jessup-1989}. Overall, very good agreements are seen between the simulation and the experiment. For example, in the region $0.3 <r/R<1.2$, the maximum difference on $\overline{u}$ is only around $2\%$. However, we do see large discrepancies in the region that is close to $r/R=0.2$. This is because of the setup difference between the experiment and the simulation. In the experiment, the propeller shaft is actually at downstream, making it a stationary wall surface at $(x/R,r/R)=(0.951,0.2)$. In contrast, for the simulation, the shaft is at upstream, and it is a flow region at the same location. \begin{figure}[H] \centering \includegraphics[width=4.2in]{./figures/ellip_time_vcompare.pdf} \caption{Comparison of mean velocity profiles at $x/R=0.951$ with experiment.} \label{fig:ellip_time_vcompare} \end{figure} Based on the Kutta-Joukowski theorem, for an inviscid flow the ``lift'' on a unit span of a body (such as a propeller blade) is proportional to the circulation. Following the derivation in \cite{kerwin-1982,wang-1985}, the circulation around a blade section at $r$, denoted by $\Gamma(r)$, is related to the circumferential speed of the slipstream as \begin{equation} \Gamma(r) \approx - 2\pi r v_\theta(r)/Z, \end{equation} where $Z$ is the number of blades of a propeller. Applying the above relation to the mean flow, we can define the following nondimensional circulation for a blade, \begin{equation} G(r) = -\frac{1}{Z} \frac{r}{R} \frac{\overline{v}_\theta(r)}{U_\infty} \approx \frac{\Gamma(r)}{2\pi R U_\infty}. \end{equation} This variable can be employed to measure the load distribution on a blade. Moreover, $G$ is conservative for inviscid flows, and should be roughly conservative for high Reynolds number flows (where viscous effects are small). Figure \ref{fig:ellip_time_circu} shows the circulation profiles at different streamwise locations. The curves in (a) and (b) start around $r/R=0.2$ because of the presence of the fairwater at these two locations. An experimental measurement from \cite{jessup-1989} is also shown in (c), which agrees well with the simulation result. Overall, we see that all profiles have very similar shapes and amplitudes, which confirms that the circulation is indeed roughly conserved. Nevertheless, viscous effects are still evident since the local narrow peaks are gradually smoothed out as the flow travels downstream. These curves also signify that the load is mostly concentrated around the mid-section (i.e., $r/R=0.5$) of each blade. The large peak around $r/R=0.9$ in the near filed indicates that the propeller is also heavily loaded around the tips, which is consistent with the strong tip vortices that we have observed in the flow field. \begin{figure}[H] \centering \includegraphics[width=5.8in]{./figures/ellip_time_circu.pdf} \caption{Flow circulation at different streamwise locations.} \label{fig:ellip_time_circu} \end{figure} \subsubsection{Pressure field} Figure \ref{fig:ellip_phase_p} shows the phase-averaged pressure field in the central $x$-$y$ plane. Comparing with the isosurfaces of Q-criterion in Fig. \ref{fig:ellip_phase_Qcr}, we see that the tip vortices show up as local pressure minima along the edge of the slipstream. Meanwhile, the fairwater wake, especially the hub vortex, is a very low pressure region, and is responsible for the large drag on the fairwater (see Tab. \ref{tab:ktkq_mean2}). \begin{figure}[H] \centering \includegraphics[width=5.2in]{./figures/ellip_phase_p.pdf} \caption{Contours of phase-averaged pressure in the central $x$-$y$ plane.} \label{fig:ellip_phase_p} \end{figure} Of great importance is the pressure distribution around the blades, which directly affects the thrust and torque on the propeller. To see the pressure effects on the thrust more clearly, we have plotted in Fig. \ref{fig:ellip_p_xy} several slices in the $x$-$y$ plane through the top blade at different spanwise (i.e., $z$) locations. From (a) to (h), we are moving from the leading edge to the trailing edge of the blade (refer to Fig. \ref{fig:geo_ellip}) along the $z$ direction. The suction side is on the left, and the pressure side is on the right. As expected, we see that most part of the suction side is in a low pressure region, and the pressure side is in a high pressure region. When we go from the leading edge to the trailing edge, the size of the low pressure region first increases and then decreases, with the maximum size around the mid-span, i.e., $z=0$. In contrast, the size of the high pressure region first decreases, and then increases. These pressure distributions apparently suggest that the thrust load is more concentrated on the trailing portion ($z>0$) of the blade. \begin{figure}[H] \centering \includegraphics[width=6.4in]{./figures/ellip_phase_p_xy.pdf} \caption{Contours of phase-averaged pressure in different $x$-$y$ planes through the top blade.} \label{fig:ellip_p_xy} \end{figure} Similarly, the contribution from different parts of a blade to the torque can be visualized through the pressure distribution in different $y$-$z$ planes through the blades as shown in Fig. \ref{fig:ellip_p_yz}. Again, from (a) to (h) we are moving from the leading edge to the trailing edge (refer to Fig. \ref{fig:geo_ellip}), but along the $x$ direction this time. Taking the top blade for example, the suction side is on the left, and the pressure side is on the right. From this perspective, the suction side is almost always in a low pressure region. In (a)-(d), a large portion of the pressure side actually has low surface pressure (although \begin{figure}[!hbt] \centering \includegraphics[width=6.4in]{./figures/ellip_phase_p_yz.pdf} \caption{Contours of phase-averaged pressure in different $y$-$z$ planes through the blades.} \label{fig:ellip_p_yz} \end{figure} there is a large high-pressure ``bubble'', but it is detached from the blade surface). In contrast, the pressure side has increased pressure in (e)-(h). This pressure distribution results in higher torque on the trailing portion of the blade. We also notice that the tips of the cross-sections have the largest pressure difference most of the time. Considering the fact that the tips also has the largest arms in the cross-sections, the torque is therefore also very heavily loaded around the edge of each blade. The time-averaged pressure field in the central $x$-$y$ plane is shown in Fig. \ref{fig:ellip_time_p}, and a series of pressure profiles are given in Fig. \ref{fig:ellip_time_p1d}. From the contours, it is seen that on average the propeller generates an obvious high-pressure region in the near field $x/D<1.0$ (inside and outside of the slipstream). Other than this region, the propeller's effects on the pressure field are mostly contained within the slipstream. The hub vortex represents the pressure minima of the whole flow field. From the profiles, we notice three local pressure minima around $r/D=0$, $r/D \approx 0.12$ (only in the very near field), and $r/D\approx 0.42$. They actually correspond to the hub vortex, the root vortices, and the tip vortices, respectively. The pressure recovery in the blade wake ($0.1<r/D<0.5$) is evident as the flow goes downstream. In contrast, we do not see consistent pressure recovery in the fairwater wake ($r/D<0.1$) due to the strong hub vortex. \begin{figure}[H] \centering \includegraphics[width=5.2in]{./figures/ellip_time_p.pdf} \caption{Contours of time-averaged pressure in the central $x$-$y$ plane.} \label{fig:ellip_time_p} \end{figure} \begin{figure}[H] \centering \includegraphics[width=3.6in]{./figures/ellip_time_p1d.pdf} \caption{Profiles of time-averaged pressure at different streamwise locations in the central $x$-$y$ plane.} \label{fig:ellip_time_p1d} \end{figure} \subsection{Fairwater effects} The fairwater is usually not considered in propeller design. A user has the freedom to choose a fairwater based on their preference or the availability of parts. The quantitative effects of fairwater shape have rarely been studied. In this section, we briefly study two more fairwater shapes: cylindrical and hemispherical, to compare with the ellipsoidal one from the previous sections. For a fair comparison, we require the three fairwaters to have the same surface area so that they contact with the same amount of fluids. This means that the fairwaters will have different lengths. The 1:2 ellipsoidal fairwater has a length of $0.2D$ as shown in Fig. \ref{fig:geo_ellip}. The geometries and sizes of the other two are shown in Fig. \ref{fig:fairwater_geo}. As marked in the figure, the cylindrical fairwater has a length of $0.121D$, and the hemispherical one has a length of $0.171D$ (a hemisphere of radius $0.1D$ sitting on top of a cylinder whose height is $0.071D$). Overall, the shape is more elongated (streamlined) as the shape changes from cylindrical to hemispherical and then to ellipsoidal. \begin{figure}[H] \centering \includegraphics[width=4.2in]{./figures/fairwater_geo.pdf} \caption{DTMB 4119 with cylindrical fairwater (left) and hemispherical fairwater (right).} \label{fig:fairwater_geo} \end{figure} The loads for the above two configurations are summarized in Tabs. \ref{fig:ktkq_flat} and \ref{fig:ktkq_hemi}, respectively. The overall loads on the hubs and the torques on the fairwaters are once again negligibly small. Comparing with the ellipsoidal configuration (see Tab. \ref{tab:ktkq_mean2}), we notice that the blades in both configurations here have smaller thrust and torque. However, the efficiencies of the blades (excluding fairwater contributions) in all three configurations stay almost unaffected as shown in the first row of Tab. \ref{tab:efficiency}, where the efficiency differences are around $0.3\%$. The drags on the fairwaters, on the other hand, are dramatically different for the three configurations. The cylindrical fairwater has the largest drag, followed by the hemispherical one, and then the ellipsoidal one. The overall efficiencies (with fairwater contributions included) are also summarized in Tab. \ref{tab:efficiency}. We see that the cylindrical, the hemispherical, and the ellipsoidal fairwaters reduce the propeller's overall efficiency by $3.3\%$, $2.9\%$, and $2.1\%$, respectively. We need to emphasize that these numbers are not small for a propulsion system, and we also need to repeat that these numbers are underestimated based on the assumptions that we made in Sec. \ref{sec:loads}. Since pressure contribution dominates the drag on a fairwater, it is worth looking into the pressure fields to see how the fairwaters affect the pressure distributions. \begin{table}[H] \setlength{\tabcolsep}{2mm} \centering \begin{tabular}{>{\centering}m{1.4cm} m{1.0cm} >{\centering}m{1.3cm} >{\centering}m{1.3cm} >{\centering}m{1.3cm} >{\centering}m{1.3cm} >{\centering}m{1.3cm} >{\centering\arraybackslash}m{1.3cm}} \hline & & $K_T$ & $K_{T,p}$ & $K_{T,v}$ & $K_Q$ & $K_{Q,p}$ & $K_{Q,v}$ \\ \hline \multirow{2}{*}{blades} & mean & ~0.1503 & ~0.1507 & -4.8E-4 & 0.0271 & 0.0268\hphantom{2} & 2.4E-4 \\ & r.m.s. & ~1.4E-4 & ~1.4E-4 & ~1.1E-7 & 2.8E-5 & 2.8E-5\hphantom{2} & 5.8E-9 \\ \hline \multirow{2}{*}{hub} & mean & -9.8E-5 & -8.5E-7 & -9.7E-5 & 4.0E-6 & 2.9E-8\hphantom{2} & 3.9E-6 \\ & r.m.s. & ~1.1E-7 & ~3.9E-8 & ~9.4E-8 & 1.1E-8 & 2.2E-9\hphantom{2} & 1.0E-8 \\ \hline \multirow{2}{*}{fairwater} & mean & -4.9E-3 & -4.9E-3 & -2.9E-5 & 1.5E-6 & 1.3E-11 & 1.5E-6 \\ & r.m.s. & ~6.0E-5 & ~6.0E-5 & ~1.2E-7 & 1.9E-8 & 7.9E-12 & 1.9E-8 \\ \hline \end{tabular} \caption{Loads on different parts of DTMB 4119 with a cylindrical fairwater.} \label{fig:ktkq_flat} \end{table} \begin{table}[H] \setlength{\tabcolsep}{2mm} \centering \begin{tabular}{>{\centering}m{1.4cm} m{1.0cm} >{\centering}m{1.3cm} >{\centering}m{1.3cm} >{\centering}m{1.3cm} >{\centering}m{1.3cm} >{\centering}m{1.3cm} >{\centering\arraybackslash}m{1.3cm}} \hline & & $K_T$ & $K_{T,p}$ & $K_{T,v}$ & $K_Q$ & $K_{Q,p}$ & $K_{Q,v}$ \\ \hline \multirow{2}{*}{blades} & mean & ~0.1502 & ~0.1507 & -4.8E-4 & 0.0271 & 0.0269\hphantom{2} & 2.4E-4 \\ & r.m.s. & ~3.2E-4 & ~3.2E-4 & ~5.6E-7 & 6.2E-5 & 6.2E-5\hphantom{2} & 3.0E-8 \\ \hline \multirow{2}{*}{hub} & mean & -9.1E-5 & -9.2E-7 & -9.0E-5 & 3.9E-6 & 2.9E-8\hphantom{2} & 3.9E-6 \\ & r.m.s. & ~1.8E-7 & ~5.2E-8 & ~1.6E-7 & 2.9E-8 & 3.4E-9\hphantom{2} & 2.9E-8 \\ \hline \multirow{2}{*}{fairwater} & mean & -4.4E-3 & -4.4E-3 & -3.9E-5 & 1.3E-6 & 3.5E-11 & 1.3E-6 \\ & r.m.s. & ~1.6E-4 & ~1.6E-4 & ~2.8E-7 & 4.0E-8 & 5.8E-10 & 4.0E-8 \\ \hline \end{tabular} \caption{Loads on different parts of DTMB 4119 with a hemispherical fairwater.} \label{fig:ktkq_hemi} \end{table} \begin{table}[H] \setlength{\tabcolsep}{2mm} \centering \begin{tabular}{>{\centering}m{4.5cm} >{\centering}m{2.2cm} >{\centering}m{2.2cm} >{\centering\arraybackslash}m{2.2cm}} \hline & cylindrical & hemispherical & ellipsoidal \\ \hline blades~ efficiency (excl. FW) & 0.7353 & 0.7348 & 0.7326 \\ overall efficiency (incl. FW) & 0.7113 & 0.7133 & 0.7171 \\ efficiency loss ~~~ (from FW) & -3.3\% & -2.9\% & -2.1\% \\ \hline \end{tabular} \caption{Blade efficiency and overall efficiency for different fairwater (FW) configurations.} \label{tab:efficiency} \end{table} The time-averaged pressure fields for the above two configurations are shown in Fig. \ref{fig:flat_hemi_p}. Comparing with the ellipsoidal configuration in Fig. \ref{fig:ellip_time_p}, we see that the three configurations overall have very similar pressure distribution in the wake, which explains why the blade efficiencies are not affected much. The differences are mostly limited to the region immediately downstream of the fairwater. The cylindrical fairwater generates a very large low-pressure region at its end. The hemispherical fairwater has a low-pressure region of similar size to that of the ellipsoidal one. However, at the junction of the hemispherical fairwater and the hub, there is another small but very low-pressure region due to the geometric change, which leads to an overall larger drag for this configuration than the ellipsoidal one. \begin{figure}[H] \centering \includegraphics[width=5.2in]{./figures/flat_time_p.pdf} \\ \includegraphics[width=5.2in]{./figures/hemi_time_p.pdf} \caption{Time-averaged pressure in the central $x$-$y$ plane of DTMB 4119 with cylindrical and hemispherical fairwaters.} \label{fig:flat_hemi_p} \end{figure} \section{Summary} \label{sec:summary} The first high-order eddy-resolving simulation of a marine propeller has been successfully performed in this work using a recently developed sliding-mesh method. This method combines the flux reconstruction framework and a new dynamic curved mortar approach to deal with the complex rotating geometry of a propeller without sacrificing the high-order accuracy at all. Even on a very coarse mesh with less than one-fourth million cells, the method predicts the propeller loads very accurately under a wide range of working conditions, and also captures the flow structures with a lot of details. Moreover, this method allows both phase and time averaging on the same set of grid, and thus can provide more information about a flow field. Through visualization of vortical flow structures, it is revealed that when the advance coefficient $J$ decreases, the strengths of the major vortices grow and flow instability gradually develops. The instability first comes from tip-tip vortex interaction, then tip-tip as well as tip-hub vortex interactions, and finally the trailing-edge vortices become strong enough and start playing an important rule when $J$ is sufficiently small. At the design condition, the sources of each tip vortex are identified through velocity isosurfaces to be a strand of decelerated flow from the leading edge and a strand of accelerated flow from the trailing edge of each blade. A comparison between the present high-order simulation and a previous low-order one on the same propeller has clearly demonstrated the low-dissipation advantage of the high-order method, which has allowed accurate prediction of the loads under all working conditions. In contrast, the high-dissipation of the low-order method completely failed the mission for large $J$ that generates weak flow vortices. Detailed load analysis at the design condition has revealed that the major loads are the blade thrust and torque as well as the fairwater drag, and pressure contribution dominates these loads. The pressure field and the circulation distribution show that the blade loads concentrate more on the trailing portion as well as the radial mid-section of each blade. By studying three fairwaters of different shapes, it is found that these fairwaters do not have obvious effects on the blade performance in the present setups. They, however, do dramatically change the pressure distribution on their surfaces, resulting in different induced drags and different performance degradation to the overall propulsion system. More specifically, we see an efficiency loss of at least $3.3\%$, $2.9\%$, and $2.1\%$, from the cylindrical, the hemispherical, and the ellipsoidal fairwater, respectively. It remains to be investigated whether there is an optimum fairwater shape that can minimize the efficiency loss. \section*{Acknowledgment} This work was supported by: the Office of Naval Research (Grant N00014-18-1-2265), the George Washington University, and Clarkson University. The computing resources were provided by the Navy DoD Supercomputing Resource Center through the High Performance Computing Modernization Program. The authors would like to thank Dr. Thad Michael from the Naval Surface Warfare Center Carderock Division for providing the propeller geometry and for helpful discussions. Thanks are also due to Dr. Zihua Qiu for helping on part of the mesh generation. \section*{Dedication} This paper is dedicated to Prof. Antony Jameson for celebrating his 85th birthday and his immense and continuing contributions to the many aspects of computational fluid dynamics. \newpage \biboptions{sort&compress} \bibliographystyle{elsarticle}

6d58d79d21796abbd085fa08bd720f3be70aa0d1

\section{Acknowledgments} The research presented in this paper is supported in part by the DiCyPS project and by grants from the Obel Family Foundation and VILLUM FONDEN. We also thank the \glsfirst{osm} contributors who have made this work possible. \section{Definition of AGG}\label{app:aggregation-baseline-paper5} The distribution derivation process when no historical records are available is not clear from the literature, but Hu et al.~\citep{hu2017enabling} suggests that they use the speed limit as a deterministic (rather than probabilistic) travel speed estimate. A direct application of this approach to our setting results in a Gaussian with the speed limit as the mean and (near-)zero variance. However, such a low variance is unrealistic and severely decreases the travel speed distribution modeling performance of AGG. To achieve a fair comparison, we therefore do the following. Given a road segment $e_i$, AGG outputs the mean $\mu_i$ and the standard deviation $\sigma_i$: \begin{equation}\label{eq:aggregation-baseline-outputs} \begin{split} \mu_i &= \begin{cases} M_{\tilde{T}_i} & \text{if } |\tilde{T}_i| \geq k \\ 0.79\cdot\text{SL}(e_i) & \text{otherwise} \\ \end{cases} \\ \sigma_i &= \begin{cases} S_{\tilde{T}_i} & \text{if } |\tilde{T}_i| \geq k \text{ and } |\tilde{T}_i| > 1 \\ 0.07\cdot\mu_i & \text{otherwise} \end{cases} \end{split} \end{equation} Here, $M_{\tilde{T}_i}$ and $S_{\tilde{T}_i}$ are the arithmetic mean and the standard deviation of historical records $\tilde{T}_i$, respectively. The function $\text{SL}$ returns the speed limit for its argument road segment $e_i$. The factors $0.79$ and $0.07$ used in \cref{eq:aggregation-baseline-outputs} to derive a mean and standard deviation when the number of historical records is insufficient are chosen based on our knowledge of the domain and the dataset, and takes into account that vehicles tend to travel at speeds below the speed limit on urban roads~\citep{yang2013using}, which occur occur in the data set used in our study. As an example, if the speed limit is $50$ km/h then drivers drive at around $40$ km/h on average, and $99.7\%$ of drivers are expected to drive below the speed limit (as a consequence of the empirical rule). From our experience, this scenario is quite realistic, and using of $79\%$ (rather than $100\%$) of the speed limit as the distribution mean yielded substantial performance improvements in terms of travel time point estimation in preliminary experiments. \subsection{Speed Limit Derivation} The function invocation $\text{SL}(e_i)$ in \cref{eq:aggregation-baseline-outputs} returns the speed limit of road segment $e_i$ if it exists in our dataset. However, as noted in \cref{sec:data-set-paper5}, the dataset used in our study does not contain a speed limit for all road segments. When no speed limit is given, $\text{SL}(e_i)$ instead returns a speed limit derived from road segment $e_i$'s attributes using a \gls{osm} heuristic\footnote{\url{https://wiki.openstreetmap.org/wiki/OSM_tags_for_routing/Maxspeed\#Default_speed_limits}}. Since our data is from Denmark, we use the \gls{osm} speed limit heuristic for Denmark. It is as follows. \begin{enumerate} \item If the road category of a road segment is motorway then assign a speed limit of $130$. \item If the road category is trunk then assign a speed limit $80$. \item If the road category is neither motorway or trunk, but the road segment is within a city, then assign a speed limit of $50$. \item Otherwise, assign a speed limit of $80$. \end{enumerate} A road segment is considered to be in a city if either the source intersection or the target intersection of the road segments is in a city. \subsection{Record Selection}\label{app:record-selection} The AGG baseline relies on a record selection strategy to find historical records $\tilde{T}_i$ to compute the sample mean and sample standard deviation used in \cref{eq:aggregation-baseline-outputs}. The algorithm used for record selection is presented in \cref{alg:record-selection}. The algorithm takes as input a set of training trajectories $\mathcal{TR}$, the route $p$ for which the travel time is in the process of being estimated, the road segment $e_i \in p$ for which historical records $\tilde{T}_i$ are currently being collected, and the arrival time $\tau_i$ at road segment $e_i$. In addition, the algorithm takes as input two parameters: an integer contextual relevance parameter, $c$, and a temporal relevance parameter, $\delta$, in some unit of time. Higher values of $c$ returns fewer, but more relevant historical records. Higher values of $\delta$ returns more, but less relevant historical records. \begin{algorithm} \caption{Record Selection Algorithm\label{alg:record-selection}} \begin{algorithmic}[1]\raggedright \Function{RecordSelection}{$\mathcal{TR}$, $p=(e_1, \dots, e_q)$; $e_i$, $\tau_i$, $c$, $\delta$} \State{$\tilde{T}_i \gets \emptyset$} \State{$C_i \gets (e_{i-c}, \dots, e_{i-1}, e_i, e_{i+1}, \dots, e_{i+c})$}\label{line:record-selection-context-i} \State{$I_i \gets [\tau_i - \frac{\delta}{2}; \tau_i + \frac{\delta}{2}]$} \ForEach{trajectory $\mathit{TR} \in \mathcal{TR}$} \State{\Let{$\mathit{TR} = \big((e'_1, \tau_1, t_1), \dots, (e'_n, \tau_n, t_n)\big)$}} \For{$j = 1$ to $n$} \State{$C_j \gets (e'_{j-c}, \dots, e'_{j-1}, e'_j, e'_{j+1}, \dots, e'_{j+c})$} \If{$C_i = C_j \land \tau_j \in I_i \land t_j \neq \varnothing$} \State{$\tilde{T}_i \gets \tilde{T}_i \cup \{t_j \}$} \EndIf \EndFor \EndFor \State{\Return{$\tilde{T}_i$}} \EndFunction \end{algorithmic} \end{algorithm} \cref{alg:record-selection} constructs the set of historical records $\tilde{T}_i$ as follows. The algorithm scans the set of trajectories in the loop in {Lines~$5$--$10$} for historical records. For each trajectory, the algorithm scans each traversal in the trajectory for historical records in the loop in {Lines~$7$--$10$}. A historical record refers strictly to the recorded travel time or travel speed of the traversal. To be selected, a historical record must satisfy the three conditions in Line~9. First, it must be \emph{contextually relevant} s.t.\ the contexts of road segments $e_i$ and $e'_j$ are identical, i.e., $C_i = C_j$. Here, context refers to the preceding and succeeding road segments of a road segment in a trajectory or a route. Second, the historical record must be \emph{temporally relevant}, i.e., occur at a similar time of week as $\tau_i$, defined by the interval $I_i$ (Line~4). Finally, the historical record is added to $\tilde{T}_i$ if $t_j \neq \varnothing$, i.e., if the historical record is derived from \gls{gps} data and is not created by the map-matching algorithm. \section{Definition of GRU}\label{app:gru} We express GRU in the \gls{aha} framework as the prior function $f$. Let $p=(e_1, \dots, e_n)$ be the input route starting at time $\tau_1$. For each road segment $e_i$ in $p$, \gls{gru} computes the following. \begin{equation}\label{eq:gru} \begin{split} \mathbf{x}_i &= \mathbf{e}_i \oplus \boldsymbol\tau_i \\ \mathbf{z}_i &= \text{GRU}(\mathbf{x}_t, \mathbf{z}_{i-1}) \\ \mathbf{h}_i &= \mathbf{z}_i \oplus \mathbf{x}_i \\ f(\mathbf{e}_i, \boldsymbol\tau_i; \psi) &= \text{PriorFunctionLayer}(\mathbf{h}_i), \end{split} \end{equation} where $z_0$ is a $d$-dimensional vector of zeros. For $i > 1$, we compute $\tau_i$ by incrementing $\tau_{i-1}$ by the expected time to traverse road segment $e_{i-1}$, i.e., we increment $\tau_{i-1}$ by $\frac{l_{i-1}}{\mu_{i-1, m}}$ where $l_{i-1}$ is the length of road segment $e_i$ and $\mu_{i-1, m}$ is the expected travel speed when traversing road segment $e_i$ (calculated using \cref{eq:posterior-parameters}). The prior function layer is described in \cref{alg:hybrid-output-layer}. During training, we use the time $\tau_i$ recorded during the input training trajectory if $\tau_i \neq \varnothing$. As shown in \cref{eq:gru}, \gls{gru} takes as input the $32$-dimensional vector $\mathbf{x}_i$, a concatenation of the $16$-dimensional vector representations of road segment $e_i$ and $\tau_i$. We explain how $\boldsymbol\tau_i$ is constructed in \cref{app:representation-of-time}. Vector $\mathbf{x}_i$ is passed to a \gls{gru} cell that outputs a $32$-dimensional vector $\mathbf{z}_i$. The \gls{gru} cell is recurrent and therefore takes as input vector $\mathbf{z}_{i-1}$, the output of the \gls{gru} cell at the previous road segment of the route. Preliminary experiments indicated that the use of a \emph{skip connection} is beneficial to the \gls{gru} baseline, i.e., a connection from an earlier layer to later layer with at least one layer in-between. The skip connection is captured in the computation of $\mathbf{h}_i$ in \cref{eq:gru} where the output of the \gls{gru} cell $\mathbf{z}_i$ is concatenated with the input vector $\mathbf{x}_i$. Finally, \gls{gru} applies the prior function layer described in \cref{alg:hybrid-output-layer} to vector $\mathbf{h}_i$ to output the prior hyperparameters. Note that $\tilde{T}_i = \emptyset$ when computing the posterior predictive $\Pr(\hat{t}_i=t_i \mid \tilde{T}_i, \theta_i)$ of the \gls{gru} baseline, e.g., when computing \gls{nll} (cf.\ \cref{eq:nll}) during training or evaluation. \subsection{Representation of Time}\label{app:representation-of-time} The representation of time $\boldsymbol\tau_i$ used by the \gls{gru} algorithm in \cref{eq:gru} is a $16$-dimensional feature vector representation of the time of week, where $8$ dimensions are used to represent the time of day and $8$ dimensions are used to represent the day of the week. Formally, we learn a time-of-week vector $ \boldsymbol\tau = \boldsymbol\tau_{\mathit{tod}} \oplus \boldsymbol\tau_{\mathit{dow}} $ for time $\tau$, where $\boldsymbol\tau_{\mathit{tod}} \in \mathbb{R}^8$ represents the time of day, $\boldsymbol\tau_{\mathit{dow}} \in \mathbb{R}^8$ represents the day of the week, and $\oplus$ denotes vector concatenation. Preliminary experiments indicated that our results are robust to changes to the dimensionality of this vector representation. To represent time of day in our experiments, we divide the time of day into $96$ $15$-minute intervals $\mathcal{I} = \{I_1, \dots, I_{96} \}$ s.t.\ $I_1 = \interval{0:00}{0:15}$, $I_2 = \interval{0:15}{0:30}$, and so forth. Then, we one-hot encode the time of day $\tau_{\mathit{tod}}$ into a $96$-dimensional vector $\boldsymbol\tau'_{tod} = \begin{bmatrix} \mathbbm{1}[\tau_{\mathit{tod}} \in I_1] & \dots & \mathbbm{1}[\tau_{\mathit{tod}} \in I_{96}] \end{bmatrix}$, where $\mathbbm{1}$ is the indicator function. Finally, we multiply the one-hot encoding $\boldsymbol\tau'_{\mathit{tod}}$ by a trainable matrix $\mathbf{W}_{\mathit{tod}} \in \mathbb{R}^{96 \times 8}$ to achieve the time of day representation $\boldsymbol\tau_{\mathit{tod}} = \boldsymbol\tau'_{\mathit{tod}}\mathbf{W}_{\mathit{tod}}$ with dimensionality $8$. We represent the day of week in a manner similar to the time of day, but with $7$ dimensions in the one-hot encoding $\boldsymbol\tau'_{\mathit{dow}}$---one for each day of the week---and multiply $\boldsymbol\tau'_{\mathit{dow}}$ by a trainable weight matrix $\mathbf{W}_{\mathit{dow}} \in \mathbb{R}^{7 \times 8}$ to get an $8$-dimensional vector representation $\boldsymbol\tau_{\mathit{dow}}$ of the day of the week. \section{Discussion} \begin{itemize} \todo{Mention in method and possibly under experiments. Point in AGHA section that this is apparent from, e.g., the computation of the mean.} \item We have not tested with state-of-the-art methods \item Doesn't matter, our approach is theoretically guaranteed to be no worse than either approach. \item Function-fitting: Express extremely high confidence in the prior \item Aggregation-based: Express extremely low confidence in the prior \end{itemize} \section{Conclusions and Future Work}\label{sec:epilogue-paper5} We have presented \gls{aha}, a novel framework that provides a Unifying Approach to Travel time and speed Estimation. \gls{aha} unifies function-fitting and aggregation-based approaches to travel time and speed estimation to leverage the generalizability of function-fitting approaches with the accuracy of aggregation-based approaches. By virtue of being a Bayesian framework, \gls{aha} is able to switch smoothly between its constituent function-fitting and aggregation-based components depending on data availability. In our empirical study, we found that \gls{aha} can improve the accuracies of travel speed distribution and travel time estimation by {$40$--$64\%$} and {$3$--$23\%$}, respectively, compared to using function fitting or aggregation alone. These improvements result from the superior generalizability relative to both the function-fitting approach and the aggregation-based approach in our study, while maintaining superior or similar accuracy relative to the aggregation-based approach across all data availability scenarios in our dataset. \gls{aha} has a number of other benefits in addition to estimation performance improvements. First, the framework implicitly regularizes its function-fitting component to handle issues of data imbalance and reduce model bias due to its Bayesian nature. Second, \gls{aha} models are less reliant on the structural quality of both the mapping function, i.e., the neural network architecture in our empirical study, and input feature vector representations since they also use aggregation during estimation. This property of \gls{aha} can reduce the typically substantial resources required for feature engineering and neural network architecture design when using neural networks for function-fitting. Future directions for \gls{aha} include exploring more complex models of travel time and speed distributions. This may necessitate more sophisticated techniques than the conjugate Bayesian analysis~\citep{murphy2007conjugate} used in \gls{agha}, e.g., variational inference techniques~\citep{blei2017variational}. In addition, investigating further synergistic effects of function-fitting and aggregation within the \gls{aha} framework is of interest. For instance, in our empirical study, the internal state of the recurrent \gls{gru} cell used in the \gls{aha} models is unaffected by the computation of the posterior. This makes it more difficult for \gls{gru} cells to leverage correlations in travel speed between adjacent road segments in a route for better estimation accuracy. \COMMENT{ have studied a rather simple realization of the \gls{aha} which assumes that travel time and speed distributions are Gaussian. However, in practice, such distributions are often highly complex and do not follow any standard distributions~\citep{pace} and they are typically multimodal due to, e.g., traffic lights in intersections where some travelers arrive when the intersection is open in their direction and others when it is closed. \gls{aha} is designed to complement existing function-fitting and aggregation-based approaches s.t.\ they can be integrated into the framework. However, in our empirical study we have only integrated simple examples of function-fitting and aggregation-based approaches, which are not in themselves state-of-the-art. An interesting future direction is therefore to investigate the performance of \gls{aha} using the context of state-of-the-art function-fitting and aggregation-based approaches. We expect that this will uncover further synergistic effects between the function-fitting and aggregation-based approaches. Finally, we have formulated \gls{aha} from a segment- and route-based perspective, but, as we discuss in \cref{sec:related-work-paper5}, we expect that the \gls{aha} framework can be readily applied to tasks with other types of data instances, e.g., origin-destination pairs, a measuring station, or a loop detector. We expect that applying the \gls{aha} framework to other tasks will reveal task-specific improvements to be made. \todo{Mulighed for future work - Twin networks: Matching the future for sequence generation} } \section{Empirical Study}\label{sec:experiments-paper5} We evaluate \gls{aha} on the task of trajectory travel time estimation. In particular, we are interested in evaluating \gls{aha}'s capability for improving estimation of the travel speed distributions of road segments traversed during a trip over function-fitting and aggregation-based approaches, but also \gls{aha}'s capability for improving point estimates of travel times. In addition, we investigate the behavior of \gls{aha} under varying degrees of data availability and different choices of parameters. \subsection{Dataset}\label{sec:data-set-paper5} For our experiments, we use a dataset of $336\,253$ trajectories from Aalborg Municipality in Denmark that occurs between January 1st 2012 and December 31st 2014~\citep{trajectorydata}. The trajectories have been map-matched to the road network of Aalborg Municipality extracted from \gls{osm}~\citep{osm} with $16\,294$ intersections and $35\,947$ road segments. See \cref{sec:road-network-modeling-paper5} and \cref{sec:trajectory-modeling-paper5} for details on road network trajectory modeling, respectively. See~\citep{trajectorydata} for details on the trajectory data and map-matching process. We use the $148\,746$ trajectories from the period January 1st 2012 to June 31st 2013 for training and set aside the $72\,693$ trajectories from July 1st 2013 to December 31st 2013 for validation. We use the remaining $114\,028$ trajectories from January 1st 2014 to December 31st 2014 for testing. To characterize each road segment in a trajectory, we use a set of $16$ features derived from \gls{osm} data and data from the Danish business authority~\citep{rfnlong}. These road segment feature representations are sparse, containing information about just four attributes: road segment length, road segment category (e.g., motorway), and the kind of zone (city, rural, or summer cottage) the source and target intersections of the road segment are in. The sparsity in the feature representation makes function-fitting difficult~\citep{workshop,rfnlong}. In addition, $19\,510$ ($54\%$) of the road segments are annotated with a speed limit derived from \gls{osm} and municipal data~\citep{rfnlong}. For further details, see \citep{rfnlong}. \COMMENT{ \subsection{Representation of Time}\label{sec:representation-of-time} Travel speed estimation is a time-dependent estimation task. We therefore learn a $16$-dimensional feature vector representation of the time of week, where $8$ dimensions are used to represent the time of day and $8$ dimensions are used to represent the day of the week. To represent the time of day, we divide the time of day into $96$ $15$-minute intervals and learn an $8$-dimensional vector for each time interval. Preliminary experiments indicated that our results are robust to changes to the dimensionality of this feature vector representation. We refer to \cref{app:representation-of-time} for details on the time of week vector representation. } \subsection{Objective Function} We optimize for travel speed distribution modeling performance using the per-trajectory mean \gls{nll}. The \gls{nll} of a trajectory $\mathit{TR}$ is \begin{equation}\label{eq:nll} \text{NLL}(\mathit{TR}) = \sum_{(e_i, \tau_i, t_i) \in \mathit{TR}, t_i \neq \varnothing} \text{sNLL}(e_i, \tau_i, t_i) \end{equation} where $\text{sNLL}(e_i, \tau_i, t_i) = -\ln \Pr(t_{i} \mid \theta)$ and $\theta$ are model parameters. In the case of \gls{aha}, $\Pr(t_{i} \mid \theta)$ is the Gaussian posterior predictive (see \cref{eq:gaussian-posterior-predictive}). The \gls{nll} directly measures the likelihood of a trajectory occurring by considering how well an algorithm models the travel speed distributions of its constituent road segments. A good model achieves a high likelihood across all trajectories, resulting in a low \gls{nll} score. \subsection{Algorithms}\label{sec:exp-algos} In our empirical study, we combine a function-fitting baseline and an aggregation-based baseline in a unified approach using the \gls{aha} framework and compare their separate performance with their unified performance. The output of all algorithms is the density function $\Pr(\hat{t}_i=t_i \mid \theta)$ used in \cref{eq:nll} where $\theta$ are model parameters specific to each algorithm. All algorithms are implemented in Python~3\footnote{https://www.python.org/} and trained using the MXNet deep learning framework\footnote{\url{https://mxnet.incubator.apache.org}}. We have made the implementation of all algorithms publicly available\footnote{To be released upon acceptance.}. \subsubsection{AGG} Existing aggregation-based approaches~\citep{pace,hu2017enabling,dai2016path,yuan2011tdrive} are very similar and we use the AGG baseline to represent these approaches in our empirical study. Firstly, when estimating the travel speed distribution of a road segment $e$ at time $\tau$, these approaches aggregate historical records from trajectories where the road segment is traversed at a similar time within some interval of size $\delta$. Secondly, rather than modeling uncertainty about the hyperparameters of the distribution model like \gls{aha}, they set a threshold $k$ for the minimal number of historical records considered sufficient. If the number of historical records is insufficient, an estimate is derived from the speed limit~\citep{pace,hu2017enabling,dai2016path,yuan2011tdrive}. We represent existing aggregation-based approaches using a single baseline algorithm AGG with the two features of existing aggregation-based approaches. For a fair comparison with \gls{agha}, AGG also models travel speed distributions as Gaussian distributions rather than histograms~\citep{pace,hu2017enabling,dai2016path} or a mean value~\citep{yuan2011tdrive}. See \cref{app:aggregation-baseline-paper5} for a detailed description of the AGG baseline. \paragraph{Record Selection} AGG's performance is strongly dependent on the record selection strategy used. In general, aggregation-based approaches must balance \emph{record relevance} with \emph{record availability} The selection strategy used by AGG considers two kinds of relevance: contextual relevance and temporal relevance. The contextual relevance hyperparameter $c$ is an integer that adjusts the contextual relevance where the context is the $c$ preceding and succeeding road segments in the trajectory from which a historical records originates. Only historical travel speed records from the training trajectories with the same context are selected for aggregation. Thus, increasing $c$ increases contextual relevance. The temporal relevance parameter $\delta$ is given in some unit of time and adjusts how inclusive the record selection strategy is w.r.t.\ historical records from different times of week. For instance, when estimating for time $\tau$, contextually-relevant historical records occurring in the time interval $[\tau_i - \frac{\delta}{2}; \tau_i + \frac{\delta}{2}]$ are selected. The record selection algorithm is described fully in \cref{app:record-selection}. \subsubsection{GRU} Recurrent neural networks are a popular neural network architecture among function-fitting literature. We therefore use a simple recurrent neural network, \emph{GRU}, as the function-fitting approach in our study, which features a GRU cell~\citep{gru} with a skip connection. The GRU model uses the prior function layer described in \cref{alg:hybrid-output-layer} as the final layer. GRU is fully specified in \cref{app:gru}. Unlike AGG, recurrent models can model correlations in segment travel speeds within a trajectory. For instance, if a vehicle drives onto a motorway segment `A' from a motorway approach `B' with a lower speed limit than the motorway, some time is spent on acceleration to cruising speed. On the other hand, if the vehicle drove onto motorway segment `A' from an adjacent motorway segment `C' less time, if any, is spent on acceleration assuming similar traffic conditions. \subsubsection{\gls{aha}} We unify the AGG and GRU baselines using the \gls{aha} framework following the instructions in \cref{sec:integration}. In brief, we use the function-fitting approach, GRU, to estimate the prior and the record selection strategy of the aggregation-based approach, AGG, to compute the posterior. As mentioned in \cref{sec:hybrid-objective}, we train the generative \gls{aha} model using a discriminative objective. To evaluate the this decision, we consider both a discriminative and a generative variant of \gls{aha}. UniTE-DIS is UniTE as described in \cref{sec:aha-paper5} where the posterior predictive is optimized directly during training. This is considered an \emph{end-to-end} approach from a machine learning perspective. UniTE-GEN instead operates in two step. First, UniTE-GEN outputs only the prior predictive at training time and then computes the posterior predictive only at test time. A benefit of UniTE-GEN is that it can be applied to already training function-fitting approaches. In our empirical study, we always reuse a GRU model for UniTE-GEN in a one-to-one fashion. That is, whenever we train and evaluate a GRU model, we also evaluate the corresponding UniTE-GEN model. \subsection{Evaluation Metrics} We use \gls{nll} (see \cref{eq:nll}) to evaluate travel speed distribution estimation of each algorithm in our study. In addition, we evaluate each algorithm's travel time point estimation performance using \gls{mae}, a commonly used measure for this purpose. Finally, to make our results more easily comparable to those of other papers using different methods and datasets, we also measure the \gls{mape} of the algorithms used in our empirical study, another commonly used measure in the traffic travel time and speed estimation literature. We compute a travel time point estimate for a trajectory following route $p = (e_1, \dots, e_n)$ as $ \sum_{i=1}^n \frac{l_i}{\mathbb{E}[d_i]} = \sum_{i=1}^q \frac{l_i}{\hat{\mu_{i}}} $ where $l_i$ is the length of road segment $e_i$ and $\hat{\mu}_i = \mu_{i, m}$ is the expected travel speed computed using \cref{eq:posterior-parameters}. Recall that $m=0$ for GRU. The \gls{mae} is the mean absolute error of the estimated and actual travel time point estimate. For a trajectory $\mathit{TR}$, the absolute error is $ \text{AE} = |\hat{y}_{\mathit{TR}} - y_{\mathit{TR}}| $ where $\hat{y}_{\mathit{TR}}$ and $y_{\mathit{TR}}$ is the travel time point estimate and ground truth, respectively, of trajectory $T$. The \gls{mape} is the mean absolute percentage error of the estimated and actual travel time point estimate. For a trajectory $\mathit{TR}$, the absolute percentage error is $ \text{APE} = \frac{|\hat{y}_{\mathit{TR}} - y_{\mathit{TR}}|}{y_{\mathit{TR}}}. $ \subsection{Training and Hyperparameter Selection} The GRU and UniTE-DIS models are trained by to minimizing the \gls{nll} in \cref{eq:nll} across all trajectories in the training using the ADAM optimizer~\citep{adam}. Each GRU model in our study is reused in a UniTE-GEN model as the function-fitting component. Based on preliminary experiments, we train each GRU and UniTE-DIS model for $10$ epochs with a batch size of $128$ trajectories. The training trajectories are divided uniformly at random into $1163$ batches and are reused across all epochs. The batches are shuffled before each epoch s.t.\ the are in random order. For both UniTE-DIS and GRU, we selected the learning rate $\lambda$ by training a GRU model using for each learning rate $\lambda \in \{0.1, 0.01, 0.001, 0.0001\}$. We use the learning rate $\lambda = 0.001$ with the best validation \gls{nll}. For AGG, we selected aggregation threshold parameter $k$, contextual relevance parameter $c$, and the temporal relevance parameters $\delta$ using a grid search over values $k \in \{ 1, 2, 4, 8\}$, $c \in \{0, 1, 2, 4\}$ and $\delta \in \{15, 30, 60, 120 \}$. We selected the hyperparameter configuration $k=1$, $c=0$, and $\delta=120$ with the best validation \gls{nll}. The parameter $a$ used in the prior function layer (see \cref{alg:hybrid-output-layer}) serves a similar role as the aggregation threshold $k$ used in AGG. We therefore set $a=1$ based on the best value $k=1$ for AGG. We choose hyperparameters $c$ and $\delta$ using a grid search over the same values as for AGG. For UniTE-DIS the best hyperparameters are $c=1$ and $\delta=120$, and $c=4$ and $\delta=15$ for UniTE-GEN. \subsection{Performance Evaluation}\label{sec:results} \todo{Update GRU performance increases} We repeat our experiments ten times for each algorithm and report their mean performance with standard deviations in \cref{tab:results-overview}. However, one of the GRU models did not finish properly. The numbers for GRU and UniTE-GEN in \cref{tab:results-overview} are therefore based on nine runs. Since the AGG baseline is deterministic, no standard deviations are associated with its performance figures. \begin{table} \caption{Algorithm performance on the test trajectories.\label{tab:results-overview}} \centering \begin{tabular}{llll} \toprule \emph{Algorithm} & \emph{\gls{nll}} & \emph{\gls{mae} (s)} & \emph{\gls{mape} ($\%$)} \\ \midrule UniTE-DIS & $\mathbf{26.83} \boldsymbol\pm \mathbf{0.52}$ & $\mathbf{75.13} \boldsymbol\pm \mathbf{0.70}$ & $\mathbf{17.48} \boldsymbol\pm \mathbf{0.15}$ \\ UniTE-GEN & $42.99 \pm 7.84$ & $83.35 \pm 1.66$ & $20.86 \pm 0.42$ \\ GRU & $44.47 \pm 1.00$ & $97.38 \pm 2.77$ & $24.82 \pm 0.81$ \\ AGG & $75.51 \pm 0.00$ & $77.09 \pm 0.00$ & $18.53 \pm 0.00$ \\ \bottomrule \end{tabular} \todo{Update GRU and UniTE-GEN when final run finishes.} \end{table} \paragraph{Distribution Modeling} As shown in \cref{tab:results-overview}, UniTE-DIS is the best-performing algorithm on average for both travel speed distribution modeling and travel time point estimation. \todo{Update GRU performance difference} On average, UniTE-DIS outperforms the GRU and AGG baselines by $39.68\%$ and $64.48\%$, respectively, on distribution modeling performance in terms of \gls{nll}. UniTE-GEN also outperforms the baselines on distribution modeling, although not as substantially as UniTE-DIS, with improvements of $43.07\%$ and $3.33\%$ over AGG and GRU, respectively. In addition, the results vary substantially between runs with a standard deviation of $7.48$. Given these variations in \gls{nll} in combination with the low number of runs of UniTE-GEN, it is unclear whether UniTE-GEN outperforms GRU in general. A detailed analysis of the results suggests that UniTE-GEN implicitly sacrifices estimation accuracy of the distribution mean for estimation accuracy of the distribution variance, a pattern not found in UniTE-DIS or the GRU baseline. Specifically, when comparing any two runs of UniTE-DIS and the GRU baseline, the run with the best estimation of the distribution typically also has the distribution variance estimate. In addition, the GRU model that provides the prior hyperparameters has been pre-trained on the data and therefore also to estimate travel speed distribution. However, as a consequence of the rules for computing the posterior hyperparameters in \cref{eq:posterior-parameters}, the degrees of freedom of the estimated $t$-distribution of the posterior predictive increases with the number of historical records used. This results in a reduced spread of the distribution, particularly for road segments with relatively few historical records that do not capture the travel speed distribution. \paragraph{Travel Time Point Estimation} Interestingly, the ranking of the algorithms w.r.t.\ travel time point estimation is different from the ranking w.r.t.\ distribution modeling, as shown in \cref{tab:results-overview}. The differences between the algorithms are also smaller, with the GRU baseline being a notable exception. UniTE-DIS remains best with average performance increases of $2.54\%$, $9.86$, and $22.85\%$ over AGG, UniTE-GEN, and GRU, respectively, in terms of \gls{mae}. Unlike for the task of distribution modeling, UniTE-GEN provides a substantial performance improvement of $14.40\%$ over the GRU baseline for travel time point estimation, with a smaller standard deviation, i.e., $1.66$ vs. $2.77$ for \gls{mae}. \paragraph{Summary} UniTE-DIS performs {$39.68$--$64.48\%$} and {$2.54$--$22.85\%$} better on distribution modeling and travel time point estimation, respectively, than the function-fitting and aggregation-based baselines. In addition, UniTE-DIS outperformed UniTE-GEN by {$9.86$--$37.60\%$} across all measures and does not suffer from the mean-variance estimation accuracy trade-off we observed in the UniTE-GEN model, leading to large variances in distribution modeling performance. These findings show that, while perhaps unconventional, training UniTE models to optimize a discriminative objective rather than a conventional generative objective, is worthwhile. \subsection{The Generalizability-Accuracy Trade-Off} One of our primary goals for \gls{aha} is that \gls{aha} models are flexible s.t.\ they can exploit the generalizability of the function-fitting component in data-sparse situations and can exploit the accuracy of aggregation-based methods in data-abundant situations. To investigate this capability, we collect the sNLL of different road segments during evaluation of the test trajectories. Then, we group them by the number of historical records available according to the ground-truth arrival times at the segments and compute the mean sNLL for each of these groups using the least restrictive record selection strategy (AGG's). The number of historical records used by each algorithm may differ from this number, depending on the restrictiveness of the record selection strategy and the accuracy of the expected arrival time at the road segments, but they are very strongly correlated. The relationship between mean sNLL and the number of historical records available is shown in \cref{fig:robustness}. \begin{figure} \centering \begin{center} \scalebox{0.85}{ \hspace*{-0.525cm} \input{illustrations/robustness_zoomed_in.pgf} } \end{center} \vspace*{-0.35cm} \caption{The relationship between mean sNLL for road segment with different number of historical records available during evaluation on the test trajectories. \label{fig:robustness}} \end{figure} \COMMENT{ \begin{figure*} \centering \begin{subfigure}{0.495\textwidth} \hspace*{-0.525cm} \input{illustrations/robustness_zoomed_out.pgf} \caption{\label{fig:robustness-zoomed-out}} \end{subfigure} \begin{subfigure}{0.495\textwidth} \hspace*{-0.275cm} \input{illustrations/robustness_zoomed_in.pgf} \caption{\label{fig:robustness-zoomed-in}} \end{subfigure} \caption{Mean test sNLL for road segment with different numbers of historical records available.\label{fig:robustness}} \end{figure*} } \paragraph{Analysis} \cref{fig:robustness} shows that AGG has substantially worse performance than the other algorithms when few historical records are available, but that it overtakes GRU when $8$ historical records available. Until about $35$ historical records are available, UniTE-DIS has the best performance, but from this point on, AGG has similar performance. UniTE-DIS and AGG maintain quite similar performance when $80$ to $250$ historical records are available. In this interval, the historical records have high quality, and there are sufficiently many. The AGG baseline therefore overtakes UniTE-DIS and UniTE-GEN in terms of performance because it relies solely on the data and does not make use of a prior. In addition, the record selection strategies of UniTE-DIS and UniTE-GEN are more restrictive, causing them to use, respectively, $34\%$ and $91\%$ less data on average than does AGG. The adverse effect of using a prior in this interval is particularly strong for UniTE-GEN.\@ We expect this is (a) because it has the by far most restrictive record selection strategy and (b) because it expresses much higher certainty in the prior than UniTE-DIS does.\@ When more than $250$ historical records are available, UniTE-DIS and UniTE-GEN achieve similar performance to AGG since the influence of the prior becomes less significant. \paragraph{Summary} Both UniTE variants exhibit better generalizability than AGG when few historical records are available and achieve similar accuracy when many historical records are available, where UniTE-DIS achieves superior generalizability and accuracy compared to UniTE-GEN.\@ Between these two extremes, AGG is superior to the UniTE variants due to the availability of sufficient high-quality data. These data conditions are particularly favorable for AGG since it does not make use of a prior that may be inaccurate. \subsection{Regularizing Properties} An important property of the \gls{aha} framework is the implicit regularization w.r.t.\ data imbalances that results from the definition of the posterior. In particular, the posterior (cf.\ \cref{eq:posterior-parameters}) implicitly regularizes the prior function since the influence of the prior function is inversely proportional to the number of historical records. When training UniTE-DIS, we expect that this implicit regularization encourages the GRU architecture used as the prior function to output prior travel speed distributions that are accurate when only few historical records are available. However, we do not expect this effect in GRU (or, equivalently, in UniTE-GEN) since historical records are not used at training time. To study the regularization properties, we compare the segment-wise mean \gls{nll}, s\gls{nll}, of UniTE-DIS and GRU on the training set. Let $e$ be a road segment that occurs $n$ times in the set of training trajectories. Then, the sNLL of $e$ is $\frac{1}{n}\sum_{i=1}^n - \ln p_i$, where $p_i = \Pr(t_i \mid \mathbf{e}, \tau_i, \tilde{T}_i = \emptyset; \theta)$ is the value of the density function of the prior predictive at the $i$th occurrence. Here, $\tau_i$ is the time of the occurrence, and $t_i$ is the ground truth travel speed. \begin{figure} \begin{center} \scalebox{0.85}{ \hspace*{-0.475cm} \input{illustrations/performance_over_frequency.pgf} } \end{center} \vspace*{-0.35cm} \caption{Moving average (sample window size $\mathbf{250}$) of the zero-centered training sNLL of the prior predictive of UniTE-DIS and GRU at different road segment frequencies.\label{fig:regularization-analysis}} \end{figure} \cref{fig:regularization-analysis} plots the sNLL of all road segments as a function of their frequency. For ease of comparison, we have centered the values around $0$ and use a moving average with a sample window size of $250$. Thus, values above $0$ indicate below average performance and values below $0$ indicates above than average performance within the sample window. As expected, the prior predictive of UniTE-DIS favors low-frequency road segments more than GRU and GRU favors high-frequency road segments more than UniTE-DIS. Although not visible in \cref{fig:regularization-analysis}, the point of diversion occurs at a road segment frequency of $n=38$ corresponding to the $56$th percentile. \christian{Hvis det er det vi skal se, hvorfor så ikke logaritmisk x-akse?} This discrepancy continue to increase as the road segment frequency increases. These findings suggest that the implicit regularization of the \gls{aha} framework contributes to the superior performance, shown in \cref{tab:results-overview}, of UniTE-DIS. \subsection{Data Efficiency} We investigate how the performance of the algorithms changes depending on the training data available by giving them part of the training data, while keeping the number of iterations (i.e., backpropagations) constant. For the sake of brevity, we show only the results in terms of test \gls{nll} in \cref{fig:data-influence}, but the patterns when using \gls{mae} and \gls{mape} are similar. \begin{figure} \centering \vspace*{-0.65cm} \begin{subfigure}{\columnwidth} \begin{center} \scalebox{0.85}{ \hspace*{-0.5cm} \input{illustrations/data_influence_nll_big.pgf} } \end{center} \end{subfigure} \begin{subfigure}{\columnwidth} \vspace*{-0.35cm} \begin{center} \scalebox{0.85}{ \hspace*{-0.5cm} \input{illustrations/data_influence_nll_small.pgf} } \end{center} \end{subfigure} \vspace*{-0.35cm} \caption{Algorithm travel speed distribution modeling performance for different data subsets.\label{fig:data-influence}} \end{figure} \paragraph{AGG and GRU} As shown in \cref{fig:data-influence}, the performance of AGG is highly dependent on the size of the training set, whereas GRU is comparatively good at generalizing when using few training trajectories. As discussed in \cref{sec:introduction-paper5}, function-fitting approaches such as GRU are good at generalization, and aggregation-based approaches such as AGG are not. The results are therefore as expected. However, it is notable that the performance of GRU is near-constant (within six standard deviations) and does not improve as more data becomes available. We expect the primary cause to be the lack of high quality features available: using four attributes represented as $16$ features means that the feature space can be observed almost completely through few trajectories. \paragraph{UniTE-GEN} \cref{fig:data-influence} shows that UniTE-GEN tends to nearly-match or outperform it's pre-trained GRU component. The results also suggest that UniTE-GEN tends to scale better with training data availability. However, the value of the UniTE framework when used in a generative manner is highly dependent on the pre-trained GRU model it uses. As shown in the figure, this dependence results in a large performance variance that is consistent with the results in \cref{tab:results-overview}. However, when measuring travel time point estimation using \gls{mae} or \gls{mape} (not shown), UniTE-GEN is strictly superior to its pre-trained function-fitting GRU component for all data set sizes and the performance difference increases proportionally to data set size. \paragraph{UniTE-DIS} UniTE-DIS outperforms the other algorithms for all data set sizes. Unlike GRU, UniTE-DIS scales with data availability, although not as aggressively as AGG. The results show that UniTE-DIS has high data efficiency and can substantially outperform a purely function-fitting and a purely aggregation-based approach even at very small data set sizes. For instance, when using ca. $10\%$ of the training trajectories, UniTE-DIS outperforms AGG and GRU by $728\%$ and $20\%$, respectively. \cref{fig:data-influence} suggests that the performance of UniTE-DIS deteriorates beyond $80\%$ of the training trajectories. Given that the differences are small, we expect that this is due to the stochastic nature of the training process, but it may also be due to differences in the distributions of the training and test sets. If the latter is the case, regularization techniques may be used during training to enhance generalizability. However, since the performance differences are within six standard deviations, we cannot conclude which is the case. \paragraph{Summary} UniTE-DIS exhibits superior data efficiency compared to GRU and AGG and achieves superior performance for all data set sizes considered in our study and achieves superior performance at all data set sizes considered in our study. And Unlike the AGG baseline, UniTE-DIS exhibits good generalizability for small data set sizes. And unlike the GRU baseline, the performance of UniTE-DIS improves proportionally to size of the data set. UniTE-GEN exhibits the same behavior as UniTE-DIS when measuring travel time point estimation performance, albeit with strictly worse performance at all data set sizes. However, when measuring distribution modeling performance, the potential performance increase of using UniTE-GEN on a pre-trained GRU model is highly dependent on the particular GRU model. \subsection{Record Selection Strategies} The value of the \gls{aha} framework depends on the record selection strategy. In particular, there is a trade-off between data availability and data quality. If the historical records are irrelevant, the posterior predictive may perform worse than the prior predictive. However, if no historical records are available, \gls{aha} offers no benefits (but also no drawbacks). In addition, we hypothesize that optimal record selection strategies for purely aggregation-based approaches differs from optimal selection strategies for \gls{aha}. We therefore investigate how different values of the temporal relevance hyperparameter $\delta$ and the contextual relevance hyperparameter $c$ influence AGG, as well as UniTE-DIS and UniTe-GEN. Recall that large $c$ values indicates high relevance and high $\delta$ indicates low relevance. \begin{figure*}[ht] \centering \begin{subfigure}{\textwidth} \centering \scalebox{0.85}{ \input{illustrations/record_selection_legend.pgf}} \vspace*{-0.1cm} \end{subfigure} \begin{subfigure}{0.33\textwidth} \begin{center} \hspace*{-0.55cm} \scalebox{0.85}{ \input{illustrations/record_selection_dis.pgf}} \vspace*{-0.55cm} \caption{UniTE-DIS} \end{center} \end{subfigure} \begin{subfigure}{0.33\textwidth} \begin{center} \hspace*{-0.55cm} \scalebox{0.85}{ \input{illustrations/record_selection_gen.pgf}} \vspace*{-0.55cm} \caption{UniTE-GEN} \end{center} \end{subfigure} \begin{subfigure}{0.33\textwidth} \begin{center} \hspace*{-0.55cm} \scalebox{0.85}{ \input{illustrations/record_selection_agg.pgf}} \vspace*{-0.55cm} \caption{AGG} \end{center} \end{subfigure} \caption{The distribution modeling performance on the validation set of (a) UniTE-DIS, (b) UniTE-GEN, and (c) AGG for different values of parameters $c$ and $\delta$ that regulate contextual and temporal relevance of the retrieved historical records.\label{fig:record-selection}} \end{figure*} \paragraph{Analysis} \cref{fig:record-selection} shows the validation \gls{nll} found during hyperparameter selection to select the best combination of $c$ and $\delta$ values for AGG, UniTE-DIS, and UniTE-GEN.\@ As shown in the figure, the optimal values differ across the algorithms. In addition, they follow quite different patterns. AGG benefits from high data availability even at the expense of data relevance. This is not surprising given the inaccurate heuristic used to estimate the travel speed distributions when too few records are available. UniTE-GEN follows the opposite pattern and prefers high data relevance. This is likely caused by the variance-reducing effects of UniTE-GEN discussed in \cref{sec:results}, leading to too narrow travel speed distributions. If few records are available, this effect is not as pronounced. Finally, UniTE-DIS is somewhere in-between, preferring a moderate contextual relevance with $c=1$. Temporal relevance is less important, with $\delta=60$ and $\delta=120$ yielding nearly equal performance, particularly when $c=1$. \paragraph{Summary} We find that optimal the record selection strategies differ across the \gls{aha} and AGG. In particular, our results suggest that optimal strategies for \gls{aha} are more selective than those for aggregation-based approaches. We attribute this to the difference in the quality of the `default' mechanism used to find travel speed distributions when few or no historical records are available. \subsection{Processing and Storage Complexity} Computational efficiency has not been the focus of our implementation efforts for UniTE or the baselines AGG and GRU. However, we find it prudent to discuss the empirical efficiency of our approach as well as the theoretical complexity of \gls{aha} \subsubsection{Empirical Performance} Our implementation can train a UniTE-DIS model with approximately two seconds of processing time per trajectory using less than $15$ GB of RAM on a non-dedicated server with $2.3$ GHz processors. For comparison, GRU takes approximately one and a half seconds of processing time per trajectory to train. Once trained, both UniTE-DIS and UniTE-GEN can estimate the road segment travel speed distributions along a single trajectory in approximately $0.15$ seconds of processing time per trajectory. \subsubsection{Complexity Analysis} Estimation of road segment travel speed distributions along a single trajectory $\mathit{TR} = (\mathit{tr}_1, \dots, \mathit{tr}_n)$ through \gls{aha} requires $n$ historical record look-ups and $n$ computations of the posterior hyperparameters. In the continuous time case, which we consider in our empirical study, look-ups can be performed efficiently by preprocessing the $N$ segment traversals across all trajectories. The traversals may be stored in a list or an array, requiring $O(N)$ space, and can be sorted in $O(N\log N)$ time. This enables the use of binary search to perform a look-up, requiring $O(\log N)$ time, but the search can be accelerated considerably in practice using indexing. For instance, in our implementation, we use indexes to find historical records of a particular road segment from a particular day of the week in constant time and perform binary search only on this subset of historical records. In general, each posterior hyperparameter computation can be done in $O(g(m))$ time, where $m$ is the number of historical records. The function $g$ depends on the particular instance of \gls{aha}, and in the case of \gls{agha}, computation of the posterior parameters takes $O(m)$ time. After preprocessing, the whole estimation procedure thus takes $O(n(\log N + g(m)))$ time in general and $O(n(\log N + m))$ for \gls{agha} with storage complexity $O(N)$ for a trajectory of length $n$. Discrete time is commonly used in the literature, where time is partitioned into a set of fixed-size intervals $I$. In this case, the estimation of the travel time or speed distributions along a trajectory can be reduced to $O(n|E||I|)$ time by computing and caching the posterior hyperparameters for each road segment and interval pair in $E \times I$ since, typically, $|E||I| << \log N + g(m)$. \section{Hybrid Driving Speed Estimation} Model-based approaches aim to learn the model parameters $\theta$ that best explain the relationship between the unknown future driving speed $d_i$ when traversing route $p_i$ and the feature vector representation $\mathbf{p}_i$ of route $p_i$. Formally, the model parameters $\theta$ are chosen to maximize the conditional likelihood \begin{equation}\label{eq:model-based} \prod_{i=1}^n \Pr(d_i \mid \mathbf{p}_i; \theta) \end{equation} based on a set of $n$ training trajectories. The probability density $\Pr(d_i \mid \mathbf{p}_i; \theta)$ may for instance be computed using forward propagation through a neural network with input $\mathbf{p}_i$ and network weights $\theta$. By sharing model parameters across training trajectories during optimization, model-based approaches can generalize to any unseen routes. Aggregation-based approaches aim to learn model parameters $\theta_i$ for each route in the set of training trajectories. In other words, given a set of $n$ training trajectories, instance-based approaches solve $n$ optimization problems. For the $i$th training trajectory, $\theta_i$ is chosen s.t.\ it is the maximum a posteriori probability (MAP) estimate of \begin{equation}\label{eq:aggregation-based} \Pr(\theta_i \mid \tilde{D}_i) \end{equation} where $\tilde{D}_i$ is a set of historical driving speeds. The driving speeds in $\tilde{D}_i$ are typically collected from other trajectories traversing route $p_i$. By using distinct model parameters for each route in the set of training trajectories, aggregation-based approaches can learn a more accurate representation of the driving speed distribution of $d_i$ provided they have enough data. Unfortunately, aggregation-based approaches rely on simple heuristics to provide reasonable estimates for unseen routes and are prone to overfit when only a few historical driving speeds are available. In this work, we present a hybrid approach that leverages the generalizability of model-based driving speed estimation with the accuracy of aggregation-based approaches. Given a set of $n$ training trajectories, we aim to learn model parameters $\theta$ s.t.\ $\theta$ maximizes the conditional likelihood \begin{equation} \prod_{i=1}^n \Pr(d_i \mid \tilde{D}_i) \end{equation} where \begin{equation} \begin{split} \prod_{i=1}^n \Pr(d_i \mid \tilde{D}_i) &= \prod_{i=1}^n \int\Pr(d_i \mid \theta_i, \tilde{D}_i) \Pr(\theta_i \mid \tilde{D}_i; f(\mathbf{p}_i, \theta)) \mathit{d\theta_i} \\ &= \prod_{i=1}^n \int\Pr(d_i \mid \theta_i, \tilde{D}_i) \frac{\Pr(\theta_i; f(\mathbf{p}_i, \theta)) \Pr(D \mid \theta_i)}{\Pr(D)} \mathit{d\theta_i} \end{split} \end{equation} In essence, we parameterize the computation of the Compared to \cref{eq:} In essence, we have parameterized the di Here, $\Theta$ is a function of the \section{Problem Formulation} In this work, we aim to combine the model-based approach and instance-based approach into a hybrid approach to predict time-varying driving speeds when traversing edges and routes. Model-based approach typically learn a discriminative model of the form \begin{equation} \Pr(d \mid \mathbf{p}, \theta) \end{equation} where $d$ is the driving speed when traversing route $p$ (i.e., a path or an edge) at time $\tau$, and $\theta$ are the model parameters, e.g., the weights in a neural network. The vector $\mathbf{p}$ is a time-varying $d$-dimensional feature vector of route $p$. The model parameters $\theta$ may be learned using maximum likelihood estimation by finding the parameters $\theta$ that maximizes the likelihood \begin{equation} \Pr(D \mid \theta) = \prod_{i=1}^n \Pr(d_i \mid \mathbf{p}_i, \theta) \end{equation} where $d_i$ is driving speed when traversing route $p_i$. The driving speeds $D = \{d_1, \dots, d_n\}$ are assumed to be independent and identically distributed. In this work, we aim to condition the likelihood not only on the model parameters $\theta$ but also on set of relevant instances $\tilde{\mathcal{D}}=\{\tilde{D}_1, \dots, \tilde{D}_n \}$. Formally, we use maximum likelihood estimation to find the parameters $\theta$ that maximizes the likelihood \begin{equation}\label{eq:prob-form-hybrid} \Pr(D \mid \theta, \tilde{\mathcal{D}}) = \prod_{i=1}^n \Pr(d_i \mid \mathbf{p}_i, \theta, \tilde{D}_i) \end{equation} where $\tilde{D}_i \subseteq D \setminus \{d_i\} $ is a (possibly empty) set of driving speed samples that are relevant to the prediction of driving speed $d_i$ and serve as extra input data. \todo{The evidence being subset of the data is unusual and should probably be commented on.} For instance, $\tilde{D}_i$ may be driving speed samples from route $p_i$ within a time interval around time $\tau$. We discuss possible ways to select $\tilde{D}_i$ in \cref{sec:data-selection}. \todo{Maybe mixture density networks are related and could help the reader.} \subsection{Gaussian Mixture with Component Normal-Gamma Priors} In this work, we assume that the distribution of driving speeds follows a $K$-component Gaussian mixture model, i.e., that \begin{equation} \Pr(d \mid \mathbf{p}, \theta, \tilde{D}) = \\ \sum_{k=1}^K w_{m, k} \Pr(d \mid Z=k, \mathbf{p}, \theta, \tilde{D}) \end{equation} where $w_{m, k} = \Pr(Z=k \mid \mathbf{p}, \theta, \tilde{D})$ is the weight of the $k$th component after observing relevant driving samples $\tilde{D} = \{\tilde{d}_1, \dots, \tilde{d}_m\}$. A component in the mixture can represent, e.g., samples where the driver that had to stop at an intersection, and another component may represent those who could travel through the intersection uninhibited. We assume that this probability distribution of driving speed $d_i$ follows \begin{figure*} \centering \input{illustrations/conceptual_framework.tikz.tex} \caption{\label{fig:conceptual-framework}} \end{figure*} Let $E = \{ d_1, \dots, d_n \}$ be the evidence to be inserted. Let $\boldsymbol\mu = [\mu_1 \dots \mu_K]$ and $\boldsymbol{\lambda} = [\lambda_1 \dots \lambda_K]$ denote component means and component precisions, respectively, of a $K$-component Gaussian mixture. Let $\mathbf{w} = [w_1 \dots w_K]$ denote component weights s.t.\ $w_k = \Pr(C_k)$ is the prior probability of the $k$th component. We want to compute the posterior predictive distribution of a Gaussian mixture. \begin{equation} \begin{split} \Pr(d \mid E=\{d_1, \dots, d_n\}) = \sum_{k=1}^K \Pr(C_k \mid E) \Pr(d \mid C_k, E) \\ \end{split} \end{equation} where $\Pr(d \mid C_k, E)$ is the posterior predictive distribution of the $k$th Gaussian component in the distribution. Note, that in the case of a single-component Gaussian, $\Pr(d \mid C_k, E) = \Pr(d \mid E)$ which is a student's $t$-distribution as described in \cref{sec:parameter-estimation-gaussian}. \cref{sec:parameter-estimation-gaussian} describes how to compute the posterior predictive of a Gaussian from which all the samples in $E$ originates. Updating all the components based on $E$ results in a convergence to a single Gaussian, thus losing expressibility of the mixture. The challenge is therefore to partition $E$ into $K$ partitions, one for each component, s.t.\ the $k$th component can be updated based on the $k$th. Partitioning $E$ is non-trivial since there is uncertainty as to which component a sample $d_i \in E$ originates from. Thankfully, this problem has been studied in the context of fitting Gaussian Mixture Models using an Expectation-Maximization approach~\citep{murphy-mixtures}. Here, each component $C_k$ is given a \emph{responsibility score} $r_{k,i}$ for each sample $d_i$ s.t.\ \begin{equation} \begin{split} r_{k, i} &= \Pr(C_k \mid d_i) \\ &= \frac{\Pr(d_i, C_k)}{\Pr(d_i)} \\ &= \frac{\Pr(d_i \mid C_k)\Pr(C_k)}{\sum_{l=1}^K \Pr(d_i \mid C_l)\Pr(C_l)}. \end{split} \end{equation} In other words, the responsibility score $r_{k, i}$ is the probability that driving speed sample $d_i \in E$ originates from component $C_k$. When updating a component using the rules in \cref{eq:parameter-update}, we use the responsibility scores to compute \emph{responsible sample size} $n_k = \sum_{i=1}^n r_{k, i}$, \emph{responsible sample mean} $\mu_{E, k} = \frac{\sum_{i=1}^n r_{k, i}d_i}{n_k}$, and \emph{responsible sample variance} $\sigma^2_{E, k} = \frac{\sum_{i=1}^n r_{k, i} {(d_i - \mu_{E, k})}^2}{n_k}$ to replace the sample size $n$, sample mean $\mu_E$, and sample variance $\sigma^2_E$, respectively. From this perspective, each driving speed sample $d_i \in E$ is soft-assigned to a component where the confidence in the assignment is given by $r_{k,i}$. Effectively, each component is therefore updated on different partitions of the data and thereby avoids component convergence towards the sample single Gaussian. In addition, substituting $n_k$, $\mu_{E, k}$, and $\sigma^2_{E, k}$ for $n$, $\mu_{E}$ and $\sigma^2_E$, respectively, has no effect in the single component case. In summary, we compute the posterior predictive of a Gaussian mixture, i.e., \begin{equation} \begin{split} \Pr(d \mid E=\{d_1, \dots, d_n\}) &= \sum_{k=1}^K \Pr(C_k \mid E) \Pr(d \mid C_k, E) \\ &= \sum_{k=1}^K w_{n, k} T_{2\alpha_{n, k}}(d \mid \mu_{n, k}, \frac{\beta_{n, k}(\kappa_{n, k} + 1)}{\alpha_{n, k}\kappa_{n, k}}) \end{split} \end{equation} where \begin{equation} \begin{split} w_{n, k} &= \Pr(C_k \mid E) \\ &= \frac{\Pr(E \mid C_k)\Pr(C_k)} {\Pr(E)} \\ &= \frac{\Pr(C_k) \prod_{i=1}^n \Pr(d_i \mid C_k)} {\prod_{i=1}^n \sum_{l=1}^K \Pr(d_i \mid C_l) \Pr(C_l)} \\ &= \frac{w_{0, k} \prod_{i=1}^n T_{2\alpha_{0, k}}(d_i \mid \mu_{0, k}, \frac{\beta_{0, k}(\kappa_{0, k} + 1)}{\alpha_{0, k}\kappa_{0, k}})} {\prod_{i=1}^n \sum_{l=1}^K T_{2\alpha_{0, k}}(d_i \mid \mu_{0, l}, \frac{\beta_{0, l}(\kappa_{0, l} + 1)}{\alpha_{0, l}\kappa_{0, l}}) w_{0, l}} \\ \mu_{n, k} &= \frac{\kappa_{0,k} \mu_{0,k} + n_k \mu_{E, k}} {\kappa_{0,k} + n_k} \\ \kappa_{n, k} &= \kappa_{0,k} + n_k \\ \alpha_{n, k} &= \alpha_{0,k} + \frac{n_k}{2} \\ \beta_{n, k} &= \beta_{0,k} + \frac{1}{2} n_k \sigma^2_{E,k} + \frac{1}{2}\frac{\kappa_{0,k} n_k {(\mu_{E, k} - \mu_{0, k})}^2} {\kappa_{0,k} + n_k} \end{split} \end{equation} \COMMENT{ First, we consider the single-component case and model the evidence $E = \{d_1, \dots, d_n \}$ as a random variable where the probability of sampling a driving speed $d_i \in E$ is uniform, i.e., $\Pr(E=d_i) = \frac{1}{n}$. In this case, we can rewrite the rules for posterior parameters s.t.\ they are expressed in terms of random variable $E$, i.e., \begin{equation}\label{eq:reformulated-parameter-update} \begin{split} \mu_n =& \frac{\kappa_0 \mu_0 + n \mu_E} {\kappa_0 + n} \\ \kappa_n =& \kappa_0 + n \\ \alpha_n =& \alpha_0 + \frac{n}{2}\\ \beta_n =& \beta_0 + \frac{1}{2}n_E\sigma^2_E + \frac{1}{2}\frac{\kappa_0 n {(\mu_E - \mu_0)}^2} {\kappa_0 + n} \end{split} \end{equation} where $E = \{d_1, \dots, d_n\}$ is the driving speed samples given as evidence, $\mu_E = \mathbb{E}[E] = \sum_{i=1}{n} \Pr(E=d_i) d_i$ is the sample mean, and $\sigma^2_E = Var(E) = \sum_{i=1}^n \Pr(E=d_i){(d_i - \mu_E)}^2$ is the biased sample variance. Next, we extend the rules in \cref{eq:reformulated-parameter-update} to the multi-component case. To update a Gaussian component $C_k$ in this case, we must model the uncertainty of the evidence, represented by random variable $E$, originates from $C_k$. To do this, we replace random variable $E$ in \cref{eq:reformulated-parameter-update} with conditional random variable $E \mid C_k$ with probability distribution \begin{equation} \Pr(E=d_i \mid C_k) = \frac{\Pr(C_k \mid D=d_i)}{\sum_{j=1}^n \Pr(C_k \mid D=d_j)} \end{equation} where \begin{equation*} \begin{split} \Pr(C_k \mid D=d_i) &= \frac{\Pr(D=d_i, C_k)}{\Pr(D=d_i)} \\ &= \frac{\Pr(D=d_i \mid C_k)\Pr(C_k)}{\sum_{l=1}^K \Pr(D=d_i \mid C_k)} \end{split} \end{equation*} The probability $\Pr(E=d_i \mid C_k)$ denotes the probability of sampling driving speed sample $d_i \in E$ from random variable $E$ given that the samples are generated by component $C_k$. In other words, the probability of samples $d_i \in E$ is no longer uniform, but proportional to their probability density w.r.t.\ the component being updated. In summary, the posterior parameters of the $k$th Gaussian component of the mixture is \begin{equation}\label{eq:component-parameter-update} \begin{split} \mu_{n, k} =& \frac{\kappa_{0,k} \mu_{0,k} + n \mu_{E \mid C_k}} {\kappa_{0,k} + n} \\ \kappa_{n, k} =& \kappa_{0,k} + n \\ \alpha_{n, k} =& \alpha_{0,k} + \frac{n}{2} \\ \beta_{n, k} =& \beta_{0,k} + \frac{1}{2} n_E \sigma^2_{E \mid C_k} + \frac{1}{2}\frac{\kappa_{0,k} n {(\mu_{E \mid C_k} - \mu_{0, k})}^2} {\kappa_{0,k} + n} \end{split} \end{equation} where \begin{equation*} \begin{split} \mu_{E \mid C_k} &= \sum_{i=1}^n{\Pr(E=d_i \mid C_k) d_i}\text{, and} \\ \sigma^2_{E \mid C_k} &= \sum_{i=1}^n{\Pr(E=d_i \mid C_k)} {(d_i - \mu_{E \mid C_k})}^2. \end{split} \end{equation*} } \section{Interesting Properties} Interesting regularization: Model will have little influence on loss in cases where there is a lot of data, encouraging better performance in areas with little data. These areas are precisely where a model-based approach is required. \section{Introduction}\label{sec:introduction-paper5} Estimation of travel time or speed is central to many intelligent transportation applications~\citep{realtime-speed-pred} such as trajectory analysis~\citep{stuttgart}, annotating road segments with travel times~\citep{yang2013using,zheng2013time}, and traffic forecasting~\citep{yu2017spatio}. This often concerns routes in a road network, with road segments being special cases. For clarity of presentation, we thus assume that estimation is done for routes in the remainder of the paper, but our work is also relevant to other kinds of data, e.g., location-based travel speed forecasting using loop detectors~\citep{cui2019traffic}. Existing approaches to travel time and speed estimation can be categorised as either function-fitting% ~\citep{wei2020spatial,lu2020st,ge2020global,zhang2020novel,zhang2020graph,zhang2020network, yin2020attention,lee2020predicting,guo2019attention,cui2019traffic,yu2017spatio,htte,stad,zheng2013time,fu2020estimation,barnes2020bustr,lan2019travel,wu2019deepeta,shen2019tcl,hu2019stochastic,hu2020stochastic,taoyang2019deepist,xi2019path,compacteta,rade2018wedge,zheng2018learning,wang2018will,wang2014travel,yang2013using,workshop} or aggregation-based% ~\citep{pace,hu2017enabling,dai2016path,yuan2011tdrive} approaches. Function-fitting approaches fit a function $f$ with parameters $\psi$ that maps feature vector representations of routes, to travel time and speed estimates. The parameters $\psi$ are found by using input-output pairs from historical data and minimizing the discrepancy between the mapping's output and the expected output. Aggregation-based approaches use historical travel time or speed data to compute corresponding at estimates of the travel speed or time for routes, often for different time-of-day intervals. Estimates are typically given as histograms~\citep{pace}, e.g., as parameters that denote heights of bins in equi-width histograms. \begin{figure}[t] \begin{tikzpicture} \hspace*{-0.05cm} \node[inner sep=0pt] (russell) at (0,0) {\includegraphics[height=4.25cm,trim=13cm 0cm 6cm 0cm, clip]{illustrations/intro-figure-segment.png}}; \node[inner sep=0pt] (whitehead) at (4,0) {\includegraphics[height=4.25cm, trim=0cm 0.42cm 0cm 0.42cm, clip]{illustrations/intro-figure-plot.png}}; \end{tikzpicture} \caption{ A road segment (left) and its ground-truth and estimated travel-speed distributions (right). \label{fig:compelling-example-paper5}} \end{figure} Function-fitting and aggregation-based approaches represent different trade-offs between estimation generalizability and accuracy. This trade-off is illustrated in \cref{fig:compelling-example-paper5}. As shown in the figure, aggregation-based approaches can provide the most accurate estimates given a sufficiently representative sample of size $n=189$. However, when $n=10$ it underestimates the mean and variance since the data sample is not representative and when $n=0$ it relies on a heuristic. The heuristic overestimates the mean considerably and also underestimates the variance. Unlike aggregation-based approaches, function-fitting approaches are unaffected by the representativeness and availability of the data at estimation time. Instead, it extrapolates the travel speed distribution from other road segments with similar feature representations. Thus, function-fitting approaches are highly reliant on the availability of high quality feature representations which are typically not easily accessible~\citep{workshop,rfnlong}. In this case, it estimates the mean well, but overestimates the variance. As illustrated in \cref{fig:compelling-example-paper5}, function-fitting and aggregation-based approaches perform well in different situations. However, choosing which approach is preferable in a particular situation is difficult in practice since the representability of the samples and the performance of the function-fitting approach are not known a priori. The sample representability is particularly important when only few examples are available, which is typically the case in practice. For instance, in the data set used in our empirical study $42.61\%$ of road segments have been traversed ten times or less over an 18 month period. The paper's overall contribution is \glsfirst{aha}, a Bayesian travel time and speed estimation framework that integrates (and complements) function-fitting and aggregation-based approaches to leverage the strengths of both. In addition, we present an instance of the \gls{aha} framework, \gls{agha}, that models travel time or speed as a Gaussian variable with unknown mean and variance. The use of conjugate priors in \gls{agha} allows for efficient computation of the posterior, easy implementation, and makes \gls{agha} applicable to neural network learning using standard deep learning frameworks. Finally, we investigate the capabilities of the \gls{aha} framework in an empirical study using \gls{agha}. \todo{update improvement numbers} The study shows that \gls{aha} can achieve $12.93\%\text{-}32.18\%$ and $8.34\%\text{-}18.55\%$ better performance in terms of travel speed distribution modeling and travel time point estimation, respectively, compared to using a function-fitting or aggregation-based approach alone. In addition, \gls{aha} can achieve better generalizability than function-fitting approaches while maintaining similar or better accuracy to aggregation-based approaches regardless of data availability. The remainder of the paper is structured as follows. \cref{sec:preliminaries-paper5} provides the necessary background on function-fitting and aggregation-based approaches, as well as graph modeling of road networks. \cref{sec:aha-paper5} presents the \gls{aha} framework and describe how \gls{aha} can unify existing function-fitting and aggregation-based approaches. \cref{sec:aha-gaussian-paper5} presents the Gaussian instance of the \gls{aha} framework, \gls{agha}. \cref{sec:experiments-paper5} reports on the empirical study. \cref{sec:related-work-paper5} reviews related work, and \cref{sec:epilogue-paper5} concludes and offers directions for future research. \COMMENT{ The mapping function $f$ in function-fitting approaches can generalize to unseen routes, since only a feature vector representation of the route needs to be available. However, achieving this generalizability requires substantial efforts in choosing a good structure for the mapping function $f$~\citep{rfnlong}, a good feature vector representation of the inputs~\citep{workshop,rne-review}, and a strategy for handling imbalances of, e.g., road segment popularity, in the data~\citep{workshop}. Finally, the mapping function is imperfect in practice: it may perform well in general, but poorly for particular routes. Aggregation-based approaches rely only on the sufficient availability of relevant and representative data when estimating the travel speed or time of, e.g., a route. Unlike function-fitting approaches, aggregation-based approaches use data directly at estimation time without an intermediary mapping function. If sufficient data is available, they can therefore achieve highly accurate point or distribution estimates with substantial less effort of implementation than function-fitting approaches. However, such data is typically not available for all routes at all times of day~\citep{yang2013using}, in which case aggregation-based approaches default to simple but inaccurate heuristics, resulting in poor generalizability. The generalizability-accuracy trade-off between function-fitting and aggregation-based approaches is illustrated in \cref{fig:compelling-example-paper5} using experimental results from our empirical study described in \cref{sec:experiments-paper5}. The aggregation-based approach achieves the best fit when observing $n=189$ samples, but when $n=10$ underestimates the mean and variance since the data sample is not representative. When $n=0$, the aggregation-based approach defaults to its heuristic which in this case overestimates the mean considerable and underestimates the variance even more. The function-fitting approach makes no use of data at estimation time and is unaffected by the representativeness of any available data. Instead, it extrapolates the travel speed distribution from other road segments with similar feature representations. In this case, it achieves a good estimation of the mean, but overestimates the variance. } \COMMENT{ First, we present \glsfirst{aha}, a Bayesian framework that aims to unify (and complement) existing function-fitting and aggregation-based approaches. In brief, \emph{\gls{aha}} integrates function-fitting approaches and aggregation-based approaches using Bayesian statistics where a function-fitting component provides a prior estimate of the travel time or travel speed, and, given a set of historical records, a posterior estimate is computed using an aggregation-based component. This allows \gls{aha} to always provide a reasonable estimate using the function-fitting component when no or few historical records are available. In addition, \gls{aha} gradually rely more on the aggregation-based component as the number of historical records increase, thus improving the accuracy for specific routes, when historical data is abundant. Second, we present a simple instance of the \gls{aha} framework, \emph{\gls{agha}}, which allows us to explore the capabilities of \gls{aha} analytically and empirically. As its name implies, \gls{agha} models travel time and speed distributions as Gaussian distributions with unknown means and variances. The use of conjugate priors for the mean and variance allows a differentiable closed-form formula for computing the posterior predictive that is easy to both implement and interpret In addition, this formula allows efficient training of the function-fitting component jointly with the aggregation-based component in an end-to-end manner using existing function-fitting approaches. The theoretical benefits of such end-to-end training is an implicit regularization of the function-fitting component s.t.\ the function-fitting component is optimized for situations where few historical records are available. Third, we investigate the potential of the \gls{aha} framework using the Gaussian instance on the task of trajectory travel time point estimation. Given a vehicle trajectory, we use a recurrent \gls{aha} model to predict the travel speed distribution of individual road segments along the route and use these distributions to compute the expected travel time of the trajectory. Our experiments show that using a hybrid approach achieves $12.93\%\text{-}32.18\%$ and $8.34\%\text{-}18.55\%$ better performance in terms of travel speed distribution modeling and travel time point estimation, respectively, compared to using a function-fitting or aggregation-based approach alone. The experiments suggests that these performance improvements are due to the implicit regularization of the hybrid objective function, but also that the aggregation-based approach can capture information that the function-fitting approach cannot \todo{make sure this is mentioned in the experiments section} due to lack of quality feature representation or sufficient modeling capability. \todo{something about the applicability of the framework to a wide variety of other tasks, e.g., using loop detectors} In summary our contributions are as follows. We present \gls{aha}, a travel time and speed estimation framework that leverages the generalizability of function-fitting approaches, but can achieve the accuracy of aggregation-based approaches. In addition, we present an instance of the \gls{aha} framework, \gls{agha}, that models travel time or travel speed as a Gaussian variable with unknown mean and variance. The use of conjugate prior allows for efficient computation of the posterior and easy implementation. Finally, we investigate the capabilities of the \gls{aha} framework in an empirical study using \gls{agha}. We find that \gls{aha} is able to achieve substantially better performance than using a function-fitting or aggregation-based approach alone. } \section*{Appendices} \begin{appendices} \crefalias{section}{appendix} \input{algorithm_descriptions.tex} \end{appendices} \clearpage \printbibliography \end{document} \section*{Appendices} } \section*{Appendices} \begin{appendices} \crefalias{section}{appendix} \input{representation_of_time.tex} \input{record_selection.tex} \end{appendices} \clearpage \printbibliography \end{document} \section{Introduction} Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse a arcu quis arcu malesuada ultricies vitae in felis. Curabitur porta lacus at felis viverra hendrerit in non eros. Nam tempus tincidunt metus vitae fermentum. Donec sed risus felis. Cras luctus massa elementum, semper urna vel, efficitur ipsum. Morbi at tellus libero. Praesent imperdiet, lacus nec varius placerat, est ex eleifend justo, a vulputate leo massa consectetur nunc. Donec posuere in mi ut tempus. Pellentesque sem odio, faucibus non mi in, laoreet maximus arcu. In hac habitasse platea dictumst. Nunc euismod neque eu urna accumsan, vitae vehicula metus tincidunt. Maecenas congue tortor nec varius pellentesque. Pellentesque bibendum libero ac dignissim euismod. Aliquam justo ante, pretium vel mollis sed, consectetur accumsan nibh. Nulla sit amet sollicitudin est. \section{Core Structural Elements} Nulla placerat feugiat augue, id blandit urna pretium nec. Nulla velit sem, tempor vel mauris ut, porta commodo quam. Donec lectus erat, sodales eu mauris eu, fringilla vestibulum nisl. Morbi viverra tellus id lorem faucibus cursus. Quisque et orci in est faucibus semper vel a turpis. Vivamus posuere sed ligula et. \subsection{Figures} Aliquam justo ante, pretium vel mollis sed, consectetur accumsan nibh. Nulla sit amet sollicitudin est. Etiam ullamcorper diam a sapien lacinia faucibus. Duis vulputate, nisl nec tincidunt volutpat, erat orci eleifend diam, eget semper risus est eget nisl. Donec non odio id neque pharetra ultrices sit amet id purus. Nulla non dictum tellus, id ullamcorper libero. Curabitur vitae nulla dapibus, ornare dolor in, efficitur enim. Cras fermentum facilisis elit vitae egestas. Nam vulputate est non tellus efficitur pharetra. Vestibulum ligula est, varius in suscipit vel, porttitor id massa. Nulla placerat feugiat augue, id blandit urna pretium nec. Nulla velit sem, tempor vel mauris ut, porta commodo quam \autoref{fig:duck}. \begin{table*}[t] \caption{A double column table.} \label{tab:commands} \begin{tabular}{ccl} \toprule A Wide Command Column & A Random Number & Comments\\ \midrule \verb|\tabular| & 100& The content of a table \\ \verb|\table| & 300 & For floating tables within a single column\\ \verb|\table*| & 400 & For wider floating tables that span two columns\\ \bottomrule \end{tabular} \end{table*} \subsection{Tables} Curabitur vitae nulla dapibus, ornare dolor in, efficitur enim. Cras fermentum facilisis elit vitae egestas. Mauris porta, neque non rutrum efficitur, odio odio faucibus tortor, vitae imperdiet metus quam vitae eros. Proin porta dictum accumsan \autoref{tab:commands}. Duis cursus maximus facilisis. Integer euismod, purus et condimentum suscipit, augue turpis euismod libero, ac porttitor tellus neque eu enim. Nam vulputate est non tellus efficitur pharetra. Aenean molestie tristique venenatis. Nam congue pulvinar vehicula. Duis lacinia mollis purus, ac aliquet arcu dignissim ac \autoref{tab:freq}. \begin{table}[hb \caption{Frequency of Special Characters} \label{tab:freq} \begin{tabular}{ccl} \toprule Non-English or Math & Frequency & Comments\\ \midrule \O & 1 in 1000& For Swedish names\\ $\pi$ & 1 in 5 & Common in math\\ \$ & 4 in 5 & Used in business\\ $\Psi^2_1$ & 1 in 40\,000 & Unexplained usage\\ \bottomrule \end{tabular} \end{table} Nulla sit amet enim tortor. Ut non felis lectus. Aenean quis felis faucibus, efficitur magna vitae. Curabitur ut mauris vel augue tempor suscipit eget eget lacus. Sed pulvinar lobortis dictum. Aliquam dapibus a velit. \subsection{Listings and Styles} Aenean malesuada fringilla felis, vel hendrerit enim feugiat et. Proin dictum ante nec tortor bibendum viverra. Curabitur non nibh ut mauris egestas ultrices consequat non odio. \begin{itemize} \item Duis lacinia mollis purus, ac aliquet arcu dignissim ac. Vivamus accumsan sollicitudin dui, sed porta sem consequat. \item Curabitur ut mauris vel augue tempor suscipit eget eget lacus. Sed pulvinar lobortis dictum. Aliquam dapibus a velit. \item Curabitur vitae nulla dapibus, ornare dolor in, efficitur enim. \end{itemize} Ut sagittis, massa nec rhoncus dignissim, urna ipsum vestibulum odio, ac dapibus massa lorem a dui. Nulla sit amet enim tortor. Ut non felis lectus. Aenean quis felis faucibus, efficitur magna vitae. \begin{enumerate} \item Duis lacinia mollis purus, ac aliquet arcu dignissim ac. Vivamus accumsan sollicitudin dui, sed porta sem consequat. \item Curabitur ut mauris vel augue tempor suscipit eget eget lacus. Sed pulvinar lobortis dictum. Aliquam dapibus a velit. \item Curabitur vitae nulla dapibus, ornare dolor in, efficitur enim. \end{enumerate} Cras fermentum facilisis elit vitae egestas. Mauris porta, neque non rutrum efficitur, odio odio faucibus tortor, vitae imperdiet metus quam vitae eros. Proin porta dictum accumsan. Aliquam dapibus a velit. Curabitur vitae nulla dapibus, ornare dolor in, efficitur enim. Ut maximus mi id arcu ultricies feugiat. Phasellus facilisis purus ac ipsum varius bibendum. \subsection{Math and Equations} Curabitur vitae nulla dapibus, ornare dolor in, efficitur enim. Cras fermentum facilisis elit vitae egestas. Nam vulputate est non tellus efficitur pharetra. Vestibulum ligula est, varius in suscipit vel, porttitor id massa. Cras facilisis suscipit orci, ac tincidunt erat. \begin{equation} \lim_{n\rightarrow \infty}x=0 \end{equation} Sed pulvinar lobortis dictum. Aliquam dapibus a velit porttitor ultrices. Ut maximus mi id arcu ultricies feugiat. Phasellus facilisis purus ac ipsum varius bibendum. Aenean a quam at massa efficitur tincidunt facilisis sit amet felis. \begin{displaymath} \sum_{i=0}^{\infty} x + 1 \end{displaymath} Suspendisse molestie ultricies tincidunt. Praesent metus ex, tempus quis gravida nec, consequat id arcu. Donec maximus fermentum nulla quis maximus. \begin{equation} \sum_{i=0}^{\infty}x_i=\int_{0}^{\pi+2} f \end{equation} Curabitur vitae nulla dapibus, ornare dolor in, efficitur enim. Cras fermentum facilisis elit vitae egestas. Nam vulputate est non tellus efficitur pharetra. Vestibulum ligula est, varius in suscipit vel, porttitor id massa. Cras facilisis suscipit orci, ac tincidunt erat. \section{Citations} Some examples of references. A paginated journal article~\cite{Abril07}, an enumerated journal article~\cite{Cohen07}, a reference to an entire issue~\cite{JCohen96}, a monograph (whole book) ~\cite{Kosiur01}, a monograph/whole book in a series (see 2a in spec. document)~\cite{Harel79}, a divisible-book such as an anthology or compilation~\cite{Editor00} followed by the same example, however we only output the series if the volume number is given~\cite{Editor00a} (so Editor00a's series should NOT be present since it has no vol. no.), a chapter in a divisible book~\cite{Spector90}, a chapter in a divisible book in a series~\cite{Douglass98}, a multi-volume work as book~\cite{Knuth97}, an article in a proceedings (of a conference, symposium, workshop for example) (paginated proceedings article)~\cite{Andler79}, a proceedings article with all possible elements~\cite{Smith10}, an example of an enumerated proceedings article~\cite{VanGundy07}, an informally published work~\cite{Harel78}, a doctoral dissertation~\cite{Clarkson85}, a master's thesis~\cite{anisi03}, an finally two online documents or world wide web resources~\cite{Thornburg01, Ablamowicz07}. \begin{acks} This work was supported by the [...] Research Fund of [...] (Number [...]). Additional funding was provided by [...] and [...]. We also thank [...] for contributing [...]. \end{acks} \bibliographystyle{ACM-Reference-Format} \section{A Unified Approach}\label{sec:aha-paper5} We now present the proposed \gls{aha} framework. \subsection{Framework} The primary goal of \gls{aha} is to unify a function-fitting component with an aggregation-based component s.t.\ we can seamlessly and smoothly switch between the two to leverage their respective strengths. The two components should be integrated s.t.\ \gls{aha} relies on the function-fitting component when no historical data is available or if the data is not representative, e.g., due to low availability. Conversely, we want \gls{aha} to rely on the aggregation-based component when historical data abounds. To achieve this, \gls{aha} adopts a Bayesian foundation that provides a solid theoretical basis for unifying function fitting and aggregation. \begin{figure}[h] \centering \input{illustrations/conceptual_framework.tikz.tex} \caption{The \gls{aha} framework illustrated using plate notation.\label{fig:conceptual-framework}} \end{figure} A conceptual model of \gls{aha} is illustrated in \cref{fig:conceptual-framework}. In brief, we assume that the travel time or speed distribution of a route $p_i$ at time $\tau_i$ follows a distribution with uncertain hyperparameters $\theta_i$. The prior distribution of $\theta_i$ is computed as $\Pr(\theta_i \mid f(\mathbf{p}_i, \boldsymbol\tau_i; \psi))$ using a \emph{prior function} $f$ with function parameters $\psi$ that are shared across all routes. The prior function $f$ represents the function-fitting component of \gls{aha}, and allows \gls{aha} to estimate prior distributions for routes even if no historical data is available at estimation time. This estimate is based on the travel time or travel speed distributions of similar routes at similar times. However, if historical records $\tilde{T}_i = \{\tilde{t}_{i, 1}, \dots, \tilde{t}_{i, m}\}$ are available, a posterior travel time or speed distribution $\Pr(t_i \mid \tilde{T}_i;\psi)$ for route $p_i$ at time $\tau_i$ is computed. The computation of the posterior represents the aggregation-based component in \gls{aha}. \subsection{The \gls{aha} Objective}\label{sec:hybrid-objective} We now present the objective function used to train models within the \gls{aha} framework. \subsubsection{Objective Function} \cref{fig:conceptual-framework} depicts a generative model, i.e., a model that specifies how to generate the new records from the distribution $\Pr(\hat{t}_i \mid \theta_i)$~\citep{murphy2012machine}. Generative models are usually trained by selecting parameters $\theta_i$ that maximize the joint likelihood~\citep{murphy2012machine}. However, training a generative model by maximizing the conditional likelihood is guaranteed to yield better estimations when the true travel time or travel speed distribution is different from the distribution family assumed by the model~\citep{salojarvi2005discriminative}. This is generally the case in practice, where, e.g., travel time distributions are highly complex~\citep{pace}. In this work, we are interested in predictive performance and therefore maximize the conditional likelihood \begin{equation}\label{eq:hybrid-objective} \Pr(t_i \mid \tilde{T}_i, \theta_i) = \Pr(t_i \mid \tilde{T}_i, \mathbf{p}_i, \boldsymbol\tau_i; \psi) \end{equation} across $n$ training trajectories where $t_i$ is the ground truth travel time or travel speed observed in the $i$th trajectory when traversing route $p_i$ at time $\tau_i$. The posterior predictive $\Pr(\hat{t}_i \mid \tilde{T}_i, \theta_i)$ equals the prior predictive $\Pr(\hat{t}_i \mid \theta_i)$ if no historical records are available as evidence, i.e., if $\tilde{T}_i = \emptyset$. \subsubsection{Regularizing Properties}\label{sec:regularizing-properties} The \gls{aha} framework inherently addresses the issues of data imbalance issues of function-fitting approaches where they tend to fit best to frequently occurring types of, e.g., road segments. Because the \gls{aha} framework is Bayesian, the \gls{aha} objective in \cref{eq:hybrid-objective} is implicitly regularized s.t.\ the performance of the function-fitting component represented by prior function $f$ is inversely proportional to the number of historical records available. In other words, the function-fitting component is trained to perform well in data-sparse situations. The posterior predictive in \cref{eq:hybrid-objective}, i.e., \begin{equation*} \Pr(\hat{t}_i \mid \tilde{T}_i, \theta_i) = \int_{\theta_i} \Pr(\hat{t}_i \mid \theta_i)\Pr(\theta_i \mid \tilde{T}_i) \mathit{d\theta_i}\text{,} \end{equation*} depends on the posterior distribution \begin{equation}\label{eq:regularization-properties} \Pr(\theta_i \mid \tilde{T}_i) \propto \Pr(\theta_i) \prod_{j=1}^m \Pr(\tilde{t}_{i,j} \mid \theta_i). \end{equation} As \cref{eq:regularization-properties} shows, the importance of the prior distribution $\Pr(\theta_i)$ on the posterior distribution, and thus the posterior predictive, is inversely proportional to the number of historical records. As a consequence, the influence of the function-fitting component on the \gls{aha} objective in \cref{eq:hybrid-objective} is largest when there are no historical records at all, and it gradually becomes less important if more historical records are available. Thus, by maximizing the conditional likelihood in \cref{eq:hybrid-objective}, the function-fitting component is trained to perform well in data-sparse situations. This ensures that no explicit regularization of the function-fitting component, e.g., oversampling~\citep{workshop}, is required to handle data imbalance issues. \subsection{Relation to Existing Approaches} \gls{aha} can be viewed as a hybrid of function-fitting and aggregation-based approaches. \subsubsection{Hybrid Characteristics} The \gls{aha} framework and it corresponding objective in \cref{eq:hybrid-objective} are hybrid in the sense that, like the function-fitting approaches, it fits a function $f$ that maps input feature vector representations to a prior travel time or travel speed estimate by minimizing the discrepancy between the output of the mapping and the expected output. However, like the aggregation-based approaches, \gls{aha} can also use historical records directly in the estimation process, i.e., without a typically imperfect intermediary mapping function, to adjust the prior estimate of its function-fitting component by computing the posterior. This adjustment allows \gls{aha}, like aggregation-based approaches, to approximate a travel time or travel speed distribution at arbitrary precision given sufficient data and an appropriate choice of the distribution family for $\Pr(\hat{t}_i \mid \tilde{T}_i, \mathbf{p}_i; \theta)$. A very appealing property of \gls{aha} is that from a modeling capability perspective, \gls{aha} models are capable of being at least as powerful as either their function-fitting or aggregation-based components. Specifically, a \gls{aha} model can match the performance of its function-fitting component at arbitrary precision by expressing very high confidence in the prior. Similarly, a \gls{aha} model can match the performance of its aggregation-based component at arbitrary precision by expressing very low confidence in the prior. \subsubsection{Integration with Existing Approaches}\label{sec:integration} An important feature of the \gls{aha} framework is that it is complimentary and integrable with existing approaches to travel time and speed estimation. Within the framework, existing function-fitting approaches provide the structure of the prior function $f$ used to estimate the prior hyperparameters. We expect that most function-fitting approaches can be integrated with the \gls{aha} framework with only minor modifications to the output layer and the objective function, depending on the choice of distribution for $\hat{t}$. Next, existing aggregation-based approaches are primarily concerned with the selection of historical records for aggregation. Aggregation-based approaches thus provide record selection strategies to construct the set of historical records $\tilde{T}_i$ in \cref{eq:hybrid-objective}. \subsubsection{Remarks on Training} Unlike aggregation-based approaches, \gls{aha} maximizes the conditional likelihood, and it is desirable that a function-fitting component (represented by prior function $f$) compensates for an aggregation-based component when insufficient historical data is available. In the extreme case, no data may be available. To simulate this situation during training, we recommend excluding the ground truth travel time or travel speed from the set of historical records, i.e., we recommend that $t_i \notin \tilde{T}_i$ in \cref{eq:hybrid-objective}. By excluding the ground truth travel time or travel speed from the set of historical records during the training, the function-fitting component must always contribute information missing from the set of historical records $\tilde{T}_i$ during training. Specifically, it must contribute information about the ground truth travel time or speed $t_i$ to optimize the objective in \cref{eq:hybrid-objective}. \section{{Gaussian UniTE}}\label{sec:aha-gaussian-paper5} \gls{aha} is a framework for which many instantiations are possible. To study the prospects of \gls{aha} analytically and empirically, we present one such instantiation, \glsfirst{agha}, that is both easy to implement and integrate with existing approaches. As the name suggests, \gls{agha} assumes that $\hat{t}_i \mid \theta_i$ follows a Gaussian distribution with parameters $\theta_i =(\mu_i, \lambda_i)$. The Gaussian assumption combined with the use of conjugate priors over the uncertain mean $\mu_i$ and precision $\lambda_i$ allows the generally difficult-to-compute posterior predictive in \cref{eq:hybrid-objective} to be computed efficiently and in closed-form~\citep{murphy2007conjugate}. In addition, the closed-form computation of the posterior predictive is also differentiable. This enables the use of gradient-based optimization techniques that are commonly used in function-fitting approaches based on neural networks. Note that \gls{aha} is far more general than \gls{agha} which is just one possible instantiation of the \gls{aha} framework. Differentiability of the posterior predictive is a convenient property of \gls{agha} but the \gls{aha} framework is not restricted to gradient-based optimization. \subsection{Prior}\label{sec:agha-prior} Let $\Pr(\hat{t}_i \mid \mu_i, \lambda_i)$ denote the likelihood of a Gaussian distribution with mean $\mu_i$ and variance $\sigma_i^2 = \lambda^{-\frac{1}{2}}$. We adapt the work of \citet{murphy2007conjugate} to our setting, and estimate $\mu_i$ and $\lambda_i$ using the normal-gamma prior \begin{multline}\label{eq:conjugate-priors} \Pr(\mu_i, \lambda_i \mid \mathbf{p}_i; \psi) = \mathit{NG}(\mu_i, \lambda_i \mid \mu_{i,0}, \kappa_{i,0}, \alpha_{i,0}, \beta_{i, 0}) = \\ \mathcal{N}(\mu_i \mid \mu_{i, 0}, \frac{1}{\kappa_{i,0} \lambda_{i, 0}})\mathit{Ga}(\lambda_i \mid \alpha_{i, 0}, \beta_{i, 0}), \end{multline} where $f(\mathbf{p}_i; \psi) = \begin{bmatrix}\mu_{i,0} & \kappa_{i,0} & \alpha_{i,0} & \beta_{i, 0}\end{bmatrix}$. Here, $\kappa_{i, 0} > 0$, shape parameter $\alpha_0 > 0$, and rate parameter $\beta_0 > 0$. In other words, the parameters of the prior, or the \emph{prior hyperparameters}, are output by the function $f$. \subsection{Posterior}\label{sec:agha-posterior} After observing a sample of $m$ historical records $\tilde{T}_i = \{\tilde{t}_{i, 1}, \dots, \tilde{t}_{i, m}\}$, beliefs about $\mu$ and $\lambda$ may change. Formally, the posterior distribution over $\mu_i$ and $\lambda_i$ is given as ~\citep{murphy2007conjugate} \begin{multline}\label{eq:conjugate-posteriors} \Pr(\mu_i, \lambda_i \mid \mathbf{p}_i, \tilde{T}_i; \theta) \mathit{NG}(\mu_i, \lambda_i \mid \mu_{i,m}, \kappa_{i,m}, \alpha_{i,m}, \beta_{i, m}) = \\ \mathcal{N}(\mu_i \mid \mu_{i, m}, \frac{1}{\kappa_{i,m} \lambda_{i, m}})\mathit{Ga}(\lambda_i \mid \alpha_{i, m}, \beta_{i, m}), \end{multline} with \emph{posterior hyperparameters} \begin{equation}\label{eq:posterior-parameters} \begin{split} \mu_{i, m} =& \frac{\kappa_{i, 0}\mu_{i, 0} + mM_{\tilde{T}_i}} {\kappa_{i, 0} + m} \\ \kappa_{i, m} =& \kappa_{i, 0} + m \\ \alpha_{i, m} =& \alpha_{i, 0} + \frac{m}{2}\\ \beta_{i, m} =& \beta_{i, 0} + \frac{1}{2}mS^2_{\tilde{T}_i} + \frac{1}{2}\frac{\kappa_{i, 0}m{(M_{\tilde{T}_i} - \mu_{i, 0})}^2} {\kappa_{i, 0} + m}, \end{split} \end{equation} where $f(\mathbf{p}_i; \theta) = \begin{bmatrix}\mu_{i,0} & \kappa_{i,0} & \alpha_{i,0} & \beta_{i, 0}\end{bmatrix}$, $M_{\tilde{T}_i} = \frac{\sum_{j=1}^n \tilde{T}_{i, j}}{m}$ is the sample mean, and $S^2_{\tilde{T}_i} = \frac{\sum_{j=1}^m{(\tilde{t}_{i, j} - M_{\tilde{T}_i})}^2}{m}$ is the biased sample variance. Note that if there is no data, i.e., $m = 0$, then the posterior hyperparameters are equal to the prior hyperparameters. The regularizing properties of the \gls{aha} framework discussed in \cref{sec:regularizing-properties} are reflected in the formulas for the posterior hyperparameters in \cref{eq:posterior-parameters}. For instance, the posterior mean $\mu_{i, m}$ is a weighted mean of the prior mean $\mu_{i, 0}$ and the mean of the historical records $M_{\tilde{T}_i}$ where $\mu_{i, 0}$ has weight $\kappa_{i, 0}$ and $M_{\tilde{T}_i}$ has weight $m$. Thus, the influence of the prior mean $\mu_{i, 0}$ on the posterior mean $m_{i, m}$ diminishes as $m$ increases. The remaining posterior hyperparameters follow the same pattern. \subsection{Posterior Predictive} It follows from the posterior in \cref{eq:conjugate-posteriors}, that the posterior predictive $\Pr(\hat{t}_i \mid \mathbf{p}_i \tilde{T}_i; \theta)$---which we seek to optimize in the objective function in \cref{eq:hybrid-objective}---follows a student's $t$-distribution $t_{\nu_i}(\hat{t}_{i} \mid \hat{\mu}_i, \hat{\sigma}_i)$ with $\nu_i = 2\alpha_{i, m}$ degrees of freedom, location $\hat{\mu}_i = \mu_{i, m}$, and scale $\hat{\sigma}_i = \sqrt{\frac{\beta_{i, m}(\kappa_{i, m} + 1)}{\alpha_{i, m}\kappa_{i, m}}}$~\citep{murphy2007conjugate}, and with probability density function \begin{multline}\label{eq:gaussian-posterior-predictive} h(t_i \mid \nu_i, \hat{\mu}_i, \hat{\sigma}_i) = \frac{\Gamma(\frac{\nu_i +1}{2})} {\Gamma(\frac{\nu_i}{2}) \sqrt{\nu_i \mathrm{\pi}} \hat{\sigma}_i } {\Bigg( 1 + \frac{1}{\nu_i} {\Big( \frac{t_i - \hat{\mu}_i}{\hat{\sigma}_i} \Big)}^2 \Bigg)}^{-\frac{\nu_i+1}{2}}. \end{multline} \subsection{A Prior Function Layer} To illustrate how to use \gls{agha} with neural networks, we present a prior function layer in \cref{alg:hybrid-output-layer} that outputs the prior hyperparameters in \gls{agha}. The prior function layer is intended to be used as the final layer of a neural network s.t.\ the neural network models the prior function $f(\mathbf{p}_i, \boldsymbol\tau_i; \psi)$ in \cref{fig:conceptual-framework}, where $\psi$ are neural network weights. \begin{algorithm} \caption{Forward Propagation through the Prior Function Layer\label{alg:hybrid-output-layer}} \begin{algorithmic}[1]\raggedright \Function{PriorFunctionLayer}{$\mathbf{x}_i$} \State{$x_i \gets h(\mathbf{p}_i, \boldsymbol\tau_i; \psi_h)$} \State{$\mathbf{h}_1 \gets \mathbf{W} \cdot \mathbf{p}_i$} \State{\Let{$\mathbf{h}_1 = \begin{bmatrix} h_{1,1} & h_{1,2} & h_{1,3} & h_{1,4}\end{bmatrix}$}} \State{$\mu_{i, 0} \gets h_{1,1}$} \State{$\kappa_{i, 0} \gets \textsc{ELU}_{a}(h_{1,2}) + a + \epsilon$} \State{$\alpha_{i, 0} \gets \lvert h_{1,3} \rvert + \epsilon$} \State{$\beta_{i, 0} \gets \lvert h_{1,4} \rvert + \epsilon$} \State{\Return{$\begin{bmatrix} \mu_{i,0} & \kappa_{i, 0} & \alpha_{i,0} & \beta_{i, 0}\end{bmatrix}$}} \EndFunction \end{algorithmic} \end{algorithm} The prior function layer in \cref{alg:hybrid-output-layer} takes as input a feature vector $\mathbf{x}_i$. In the context of neural networks, $\mathbf{x}_i$ may be the result of a function $h$ s.t.\ $h(\mathbf{p}_i, \boldsymbol\tau_i) = \mathbf{x}_i$ where $h$ represents forward propagation of vectors $\mathbf{p}_i$ and $\boldsymbol\tau_i$ through multiple layers. In lines 3--4, $\mathbf{x}_i$ is projected to a four-dimensional vector $\mathbf{h}_1$, one for each of the prior hyperparameters, using a learnable weight matrix $\mathbf{W}$. Recall from \cref{sec:agha-prior} that the prior hyperparameters are constrained s.t.\ $\kappa_{i, 0} > 0$, $\alpha_0 > 0$, and $\beta_0 > 0$. These constraints are enforced in lines 6--8. The values $h_{1,3}$ and $h_{1,4}$ are interpreted as the prior hyperparameters $\alpha_{i,0}$ and $\beta_{i,0}$ and are constrained by taking their absolute values and adding a small non-zero positive constant $\epsilon$ to ensure that they are greater than zero. We chose this way of enforcing non-negativity due to its simplicity. We initially constrained the value $h_{1,2}$, interpreted as the prior hyperparameter $\kappa_{i, 0}$, in the same way as $h_{1, 3}$ and $h_{1, 4}$. However, as discussed in \cref{sec:agha-posterior}, $\kappa_{i, 0}$ represents the confidence in the prior, i.e., the output of the function-fitting component. Experiments showed that, since the function-fitting component performs poorly in the initial stages of training, the value of $\kappa_{i, 0}$ will be very low. To alleviate this problem, we use the expression in line $6$ of \cref{alg:hybrid-output-layer} instead, which makes use of the \gls{elu}~\citep{elu} function \begin{equation}\label{eq:elu} \text{ELU}_a(x) = \begin{cases} x & x > 0 \\ a(\mathrm{e}^x-1) & x \leq 0,, \end{cases} \end{equation} where $a > 0$. Because \cref{eq:elu} has a minimum value of $-a$, we can enforce non-negativity of $\kappa_{i, 0}$ by adding $a$ and $\epsilon$, as shown in line $6$ of \cref{alg:hybrid-output-layer}. This expression makes the value of $\kappa_{i, 0}$ less sensitive to changes that decrease its value, thus discouraging decreases of $\kappa_{i, 0}$ during early stages of training that needs to be corrected in later stages. Hyperparameter $a$ regulates this effect, s.t.\ the effect is inversely proportional to $a$. In addition, the value of $h_{1,2}$ is initially very close to zero in a neural network setting. Using $\kappa_{i, 0} = |h_{1,2}| + \epsilon$ would therefore result in a $\kappa_{i, 0}$ value close to zero indicating, an unreasonably low confidence in the model. The constraint measure used in line~6 in \cref{alg:hybrid-output-layer} instead ensures that the initial value of $\kappa_{i,0}$ is close to $a$. Preliminary experiments showed performance improvements when enforcing non-negativity of $\kappa_{0, i}$ in this way, as opposed, taking the absolute value and adding a small constant, but they showed improvements when non-negativity of $\alpha_{i, 0}$ and $\beta_{i, 0}$ were enforced in the same way. \section{Notes} \subsection{Observations} \begin{enumerate} \item Driving speeds of adjacent segments tend to be correlated, but not necessarily similar \item In a vehicle trajectory, the driving speed when traversing a road segment is influenced by how the vehicle arrives and departs from the segment. This includes, e.g., acceleration due to a difference in speed limits or traffic conditions, the angle of a turn, or the presence of a traffic light at the connecting segments. \item Vehicle trajectories are often incomplete such that short road segments are more likely to travel time information than long road segments. \item Only a few paths will have sufficient trajectory data to adequately estimate the travel time distribution of a path. Even individual edges are likely to lack sufficient trajectory data. However, given enough data, the data may provide a very rich source of information that captures non-obvious information of an edge or path. For instance, the type of drivers that generally drive there. \item Travel time when traversing a road segment is highly time-dependent. \end{enumerate} \subsection{Proposal 1: Observation Update a Prior Estimated by a Neural Network} Estimates a $k$-component Gaussian Mixture for each road segment in a route. Each mixture can subsequently be updated based on any relevant trajectory data, e.g., all historical trajectories that have taken the route. Finally, the travel time distribution of the route is computed by summing over the mixtures of each road segment in the route (this step is very expensive, $k^{|R|}$, where $|R|$ is the length of the route. Possible Monte Carlo approximation, but maybe not so satisfying. Also not clear how to make a joint per-segment travel time loss and per-trajectory loss). This idea is computationally feasible with point estimates, but requires assuming segment driving speeds are Gaussian (which they often are not). \subsection{Proposal 2: Treat Observations as a Particular Attribute} \section{Hybrid Travel Time Estimation} We model the driving speed $D$ of a road segment $e$ as a random variable variable with probability distribution $\Pr(D \mid \theta)$ where $\theta$ is the set of distribution parameters. Our framework estimates $\theta$ using a model $f$ and adjusts this estimate based on a set driving speed samples $\{d_1, \dots, d_n\}$ from $D$. \begin{figure}[h] \centering \input{illustrations/conceptual_framework.tikz.tex} \caption{Overview of our method in plate notation.\label{fig:update-overview}} \end{figure} \cref{fig:update-overview} illustrates our framework using plate notation. In the figure, a model $f$ takes as input an edge $e$ and generates set of prior parameters $\theta_0$ without considering any driving speed samples. These prior parameters parameterize the prior probability $\Pr(D \mid \theta_0)$. In practice, the model $f$ is unlikely to be perfect. Therefore, any available driving speed samples $E=\{d_1, \dots, d_n\}$ are inserted as evidence, as illustrated in \cref{fig:update-overview}. The resulting posterior probability $\Pr(D \mid E, \theta_0) = \Pr(D \mid \theta_n)$ is used as the hybrid model output. We refer to $\theta_n$ as posterior parameters. The framework we propose is hybrid in the sense that it can use a model $f$ to predict prior parameters for the distribution of the driving speed $D$, but it can also use driving speed samples observed during training to adjust these parameters. We have only talked of the model $f$ of being dependent on the road segment $e$. In practice $f$ may also depend on other factors, such as the time of day or the information about the driver. \subsection{Estimating Driving Speed Distributions} We model a driving speed distribution $\Pr(D \mid \theta)$ as a uni-variate Gaussian mixture with $K$ components. In a Gaussian mixture, the probability of a driving speed $d$ is a weighted mean of the probability of $d$ originating from each component, i.e., \begin{equation*} \Pr(D=d \mid \theta) = \sum_{k=1}^K w_k \Pr(D=d \mid \theta_k) \end{equation*} where $\sum_{k=1}^K w_k = 1$ and $\theta_k \in \theta$ is the set of parameters for the $k$th component. In practice, the mean and variance of the driving speed is unknown, since driving speed data is incomplete and may also be noisy. Consequently, the mean and variance of each Gaussian component is also unknown. \subsubsection{Bayesian Parameter Estimation of a Gaussian}\label{sec:parameter-estimation-gaussian} We first consider the case of estimating the parameters of a single-component mixture, i.e., a single uni-variate Gaussian distribution with unknown mean and variance. For this case, there exists a well-known approach based on Bayesian inference in conjunction with conjugate priors, i.e., priors of the form $\Pr(\theta)$ that belong to the same distribution family as their posterior $\Pr(\theta \mid E)$. The use of conjugate priors allows closed-form derivation of the posterior distribution parameters which can be computed efficiently and are differentiable. Differentiability is desirable since makes it possible to train the model $f$ to be jointly with the estimation of the posterior parameters. Let $\Pr(D \mid \mu, \lambda)$ denote a Gaussian driving speed distribution with mean $\mu$ and variance $\sigma^2$. We use the following normal-gamma prior to estimate $\mu$ and $\lambda$. \begin{equation}\label{eq:conjugate-priors} \mathit{NG}(\mu, \lambda \mid \mu_0, \kappa_0, \alpha_0, \beta_0) = \mathcal{N}(\mu \mid \mu_0, \frac{1}{\kappa_0\lambda_0}) \mathit{Ga}(\lambda \mid \alpha_0, \beta_0). \end{equation} Note, that, for ease of notation, we use the precision $\lambda = \frac{1}{\sigma^2}$, shape parameter $\alpha_0$, rate parameter $\beta_0$ to parameterize the normal-gamma prior distribution. Also note that \cref{eq:conjugate-priors} is only defined for $\kappa_0 > 0$, $\alpha_0 > 0$, and $\beta_0 > 0$. In \cref{eq:conjugate-priors}, model $f$ is required to output the prior parameters $\mu_0$, $\kappa_0$, $\alpha_0$, and $\beta_0$. The prior parameters represent prior beliefs about the mean $\mu$ and precision $\lambda$ of a Gaussian distribution. After observing driving speed samples $E = \{d_1, \dots, d_n\}$ (for $n \geq 0 $), confidence in the values of $\mu$ and $\lambda$ will change. Formally, the prior parameters are updated as follows. \begin{equation}\label{eq:parameter-update} \begin{split} \mu_n =& \frac{\kappa_0\mu_0 + n\bar{d}} {\kappa_0 + n} \\ \kappa_n =& \kappa_0 + n \\ \alpha_n =& \alpha_0 + \frac{n}{2}\\ \beta_n =& \beta_0 + \frac{1}{2}nS^2 + \frac{1}{2}\frac{\kappa_0n{(\bar{d} - \mu_0)}^2} {\kappa_0 + n} \end{split} \end{equation} where $\bar{d} = \frac{\sum_{i=2}^n d_i}{n}$ is the sample mean and $S^2 = \frac{\sum_i^n{(d_i - \bar{d})}^2}{n}$ is the biased sample variance. Note that if there is no data, i.e., $n = 0$, then none of the prior parameters are updated in \cref{eq:parameter-update}. After observing driving speed samples $E = (d_1, \dots, d_n)$, the driving speed $D$ follows a student's $t$-distribution, s.t., the probability of driving speed $d$ is \begin{equation} \begin{split} \Pr(D=d \mid \mu_n, \kappa_n, \alpha_n, \beta_n) & = t_{\nu}(d \mid \hat{\mu}, \hat{\sigma}) \\ & = \frac{\Gamma(\frac{\nu +1}{2})} {\Gamma(\frac{\nu}{2}) \sqrt{\nu \pi} \hat{\sigma} } {\Bigg( 1 + \frac{1}{\nu} {\Big( \frac{d - \hat{\mu}}{\hat{\sigma}} \Big)}^2 \Bigg)}^{-\frac{\nu+1}{2}}. \end{split} \end{equation} where $\Gamma$ is the gamma function and $\nu$, $\hat{\mu}$, and $\hat{\sigma}$ are, respectively, the degrees of freedom, the center, and the scale of the distribution. \subsubsection{Bayesian Parameter Estimation of a Gaussian Mixture} It is well-established in the literature that driving speeds do not follow any simple distribution, such as a Gaussian. We therefore extend the single-component approach described in \cref{sec:parameter-estimation-gaussian} to the Gaussian mixture case. In this case, the model $f$ is required to output prior parameters $w_{0, k}, k$, $\mu_{0, k}$, $\lambda_{0, k}$, $\alpha_{0, k}$, and $\beta_{0, k}$ for each component $C_k$ in the $K$-component mixture s.t.\ the probability density function of the mixture is \begin{equation} \Pr(D=d \mid \theta_0) = \sum_{k=1}^K \Pr(C_k \mid w_{0, k}) \Pr(D=d \mid \mu_{0, k}, \lambda_{0, k}, \alpha_{0, k}, \beta_{0, k}) \end{equation} where $\Pr(C_k \mid w_{0, k}) = w_{0, k}$. After observing $n$ driving speed observations $E = \{d_1, \dots, d_n\}$ as evidence, we are interested in the posterior distribution \begin{equation} \begin{split} \Pr(D=d \mid E, \theta_0) &= \Pr(D=d \mid \theta_n) \\ &= \sum_{k=1}^K \Pr(C_k \mid w_{n, k}) \Pr(D=d \mid\mu_{n, k}, \lambda_{n, k}, \alpha_{n, k}, \beta_{n, k}) \end{split} \end{equation} where the posterior parameters are $w_{n, k}$, $\mu_{n, k}$, $\lambda_{n, k}$, $\alpha_{n, k}$, and $\beta_{n, k}$. In brief, we introduce an update rule for the component weight of a Gaussian component $C_k$ in the mixture. In addition, we modify the update rules in \cref{eq:parameter-update} s.t.\ the effect of a driving speed sample $d$ on the update of the prior parameters of a component $C_k$ is proportional to the probability $\Pr(C_k \mid d)$ that $d$ is generated by that component. This ensures that the component means and variances do not all converge towards the sample mean and sample variance when there are is more than one component. Given evidence $E = \{d_1, \dots, d_n\}$, we compute the posterior parameters of each Gaussian component $C_k$ in the mixture as follows. \begin{equation}\label{eq:mixture-component-update} \begin{split} w_n &= \frac{\kappa_0 w_0 + nw_d} {\kappa_0 + n} \\ \mu_n &= \frac{\kappa_0 \mu_0 + n\bar{d}} {\kappa_0 + n} \\ \kappa_n &= \kappa_0 + n \\ \alpha_n &= \alpha_0 + \frac{n}{2} \\ \beta_n &= \beta_0 + \frac{1}{2} n S^2 + \frac{1}{2}\frac{\kappa_0 n {(\bar{d} - \mu_0)}^2} {\kappa_0 + n} \end{split} \end{equation} where \begin{equation*} \begin{split} nw_d &= \Pr(C_k) = \sum_{i=1}^n\frac{\Pr(C_k \mid d_i)}{\sum_{j=1}^K\Pr(C_j \mid d_i)} \\ \bar{d} &= \frac{\sum_{i=1}^n \Pr(C_k \mid d_i) x_i}{\sum_{j=1}^n \Pr(C_k \mid d_j)}, \text{and} \\ S^2 &= \frac{\sum_{i=1}^n{\Pr(C_k \mid d_i)(d_i - \bar{d})}^2}{\sum_{j=1}^n \Pr(C_k \mid d_j)}. \end{split} \end{equation*} Here, $w_d$ represents the component weight of component $C_k$ when estimated from the evidence $E$, and $p$ is a pseudo count representing the expected number of driving speed samples to have originated from component $C_k$. Similarly, $\bar{d}_k$ and $S^2_k$ is the expected sample mean and expected biased sample variance of the samples originating from component $C_k$. The primary differences between the Gaussian update rules in \cref{eq:parameter-update} and the Gaussian component update rules in \cref{eq:mixture-component-update} is that \begin{enumerate} \item an update rule for the component weights in the mixture is introduced, and \item that the influence of the given samples in the update of a component $C_k$ is proportional to the number of expected samples originating from $C_k$. \end{enumerate} In other words, the samples given as evidence do not increase certainty about the mean $\mu$ and the variance $\sigma^2$ of component $C_k$ substantially if the samples are unlikely to have originated from $C_k$. \subsubsection{Joint Model Training} TODO \begin{itemize} \item Prior and posterior loss \item Initialization \item Shared $kappa_0$ \item Regularization \end{itemize} \subsubsection{Sample Selection} TODO \section{Preliminaries}\label{sec:preliminaries-paper5} We now provide the necessary background on data modeling and existing approaches to travel time and speed estimation. \subsection{Data Modeling} For clarity, we present \gls{aha} as well as existing function-fitting and aggregation-based approaches for routes, i.e., where travel times and travel speeds are estimated for routes in in a road network with road segments as a special case. \gls{aha} is applicable to other kinds of data as well, to be discussed in \cref{sec:related-work-paper5}. \subsubsection{Road Network Modeling}\label{sec:road-network-modeling-paper5} A road network is modeled as a directed graph $G=(V, E)$ where a vertex $v \in V$ represents an intersection or the end of a road, and an edge $e \in E$ represents a road segment. A route is a connected path $p = (e_1, \dots, e_n)$ where $e_i \in E$ for $1 \leq i \leq n$. In addition, a route can be mapped to a $d$-dimensional feature vector describing its characteristics using the mapping function $\phi$. For brevity, we use the notation $\mathbf{p}$ to refer to the feature vector representation $\phi(p)$ of a route $p$. If $p$ consists of one edge, i.e., $p=e$, we use the notation $\mathbf{e}$ instead. \subsubsection{Trajectory Modeling}\label{sec:trajectory-modeling-paper5} Vehicle trajectories are sequences of time-stamped \gls{gps} locations, but can be map-matched to a road network modeled as a directed graph (as described in \cref{sec:road-network-modeling-paper5}). Each map-matched trajectory is a sequence $\mathit{TR}=(\mathit{tr}_1, \dots, \mathit{tr}_n)$ where $\mathit{tr}_i=(e_i, \tau_i, t_i)$ is a triple consisting of a road segment $e_i \in E$, an arrival time $\tau_i$ corresponding to the timestamp of the first recorded \gls{gps} location on road segment $e_i$, and $t_i$ the travel time or travel speed recorded during the traversal of segment $e_i$. In some cases, the traversal of a road segment in a trip is inferred by the map-matching algorithm due to a lack of \gls{gps} data. In such cases, $t_i = \varnothing$ and $\tau_i = \varnothing$. \subsection{Existing Approaches} We now describe function-fitting and aggregation-based approaches to travel time or speed estimation for routes. \subsubsection{Function-Fitting Approaches} Function-fitting approaches% ~\citep{wei2020spatial,lu2020st,ge2020global,zhang2020novel,zhang2020graph,zhang2020network, yin2020attention,lee2020predicting,guo2019attention,cui2019traffic,yu2017spatio,htte,stad,zheng2013time,fu2020estimation,barnes2020bustr,lan2019travel,wu2019deepeta,shen2019tcl,hu2019stochastic,hu2020stochastic,taoyang2019deepist,xi2019path,compacteta,rade2018wedge,zheng2018learning,wang2018will,wang2014travel,yang2013using,workshop} assume that the relationship between the unknown future travel time or travel speed $\hat{t}_i$ when traversing route $p_i$ at time $\tau_i$ can be modeled by a function $f$ s.t. $\Pr(\hat{t}_i \mid p_i, \tau_i; \psi) = f(\mathbf{p}_i, \boldsymbol\tau_i; \psi)$ where $\psi$ is the function parameters of $f$, $\mathbf{p}_i$ is the feature vector representation $\mathbf{p}_i$ of route $p_i$, and $\boldsymbol\tau_i$ is the vector representation of time $\tau_i$. As an example of a function $f$, if the probability density $\Pr(\hat{t}_i \mid p_i, \tau_i; \psi)$ is a uni-variate Gaussian, then a possible choice of $f$ is $f(\mathbf{p}_i, \boldsymbol\tau_i; \psi) = \mathcal{N}(\hat{t}_i \mid \boldsymbol\psi_{\mu}(\mathbf{p}_i \oplus \boldsymbol\tau_i), \boldsymbol\psi_{\sigma^2}(\mathbf{p}_i \oplus \boldsymbol\tau_i))$ where $\oplus$ denotes vector concatenation and the parameters $\psi=\{\boldsymbol{\psi}_{\mu}, \boldsymbol{\psi}_{\sigma^2}\}$ consists of two real vectors. In this case, the parameters $\boldsymbol{\psi}_{\mu}$ and $\boldsymbol{\psi}_{\sigma^2}$ that are used to compute the mean and variance of the Gaussian, respectively. To fit a function $f$ to a set of $n$ training trajectories, function-fitting approaches learn model parameters $\theta$ that maximize the conditional likelihood $ \prod_{i=1}^n \Pr(\hat{t}_i \mid p_i, \tau_i; \psi) = f(\mathbf{p}_i, \boldsymbol\tau_i; \psi). $ By sharing model parameters $\psi$ across training trajectories during optimization, function-fitting approaches can generalize to unseen routes. However, in practice, generalization is imperfect for non-trivial travel time and speed estimation tasks. \subsubsection{Aggregation-Based Approaches} Aggregation-based approaches% ~\citep{pace,hu2017enabling,dai2016path,yuan2011tdrive} aim to learn model parameters $\theta_i$ for each route at time $\tau_i$ using a set of training trajectories. In other words, given a set of $n$ training trajectories, these approaches solve $n$ optimization problems. Unlike function-fitting approaches, aggregation-based approaches do not optimize the posterior predictive directly. Instead, the model parameters $\theta_i$ are chosen s.t.\ they are the maximum a posteriori probability (MAP) estimate of $ \Pr(\theta_i \mid \tau_i, \tilde{T}_i) $, where $\tilde{T}_i$ is a set of historical travel speed or time records that are typically collected from historical trajectories traversing route $p_i$ during some interval based on time $\tau_i$, e.g., the same time-of-week interval as $\tau_i$. By using distinct model parameters for each route set of training trajectories, aggregation-based approaches can learn a more accurate representation of the travel time or travel speed distribution of $\hat{t}_i$ provided sufficient historical records are available. However, in practice, it is unlikely that there is sufficient data for all routes at all times of day~\citep{yang2013using,tan2020cycle,wei2020a}. In such cases, aggregation-based approaches rely on simplistic heuristics to provide reasonable estimates for unseen routes, and they are prone to overfitting when only a few historical records are available. \COMMENT{ \subsection{Road Network and Trajectory Modeling} We model a road network as a directed graph $G=(V, E)$ \todo{road network} \todo{routes, road segment, trajectory, historical records, features}} \COMMENT{ \subsection{Bayesian Parameter Estimation of a Gaussian}\label{sec:parameter-estimation-gaussian} The parameters of a univariate Gaussian distribution can be estimated using Bayesian inference in conjunction with conjugate priors, i.e., priors of the form $\Pr(\theta)$ that belong to the same distribution family as their posterior $\Pr(\theta \mid x)$. Here, $\theta$ is the parameters of the distribution parameters and $x$ is observed data. The use of conjugate priors allows closed-form derivation of the posterior distribution parameters which can be computed efficiently and are differentiable. For the purposes of this paper, we are interested in the case where both the mean and the variance of the Gaussian distribution is unknown. Let $\Pr(X \mid \mu, \lambda)$ denote the likelihood of a Gaussian distribution with mean $\mu$ and variance $\sigma^2 = \lambda^{-\frac{1}{2}}$. For ease of notation, we use $\lambda$, also known as \emph{the precision}, to parameterize the distribution. We use the following normal-gamma prior to estimate $\mu$ and $\lambda$. \begin{equation}\label{eq:conjugate-priors} \mathit{NG}(\mu, \lambda \mid \mu_0, \kappa_0, \alpha_0, \beta_0) = \mathcal{N}(\mu \mid \mu_0, \frac{1}{\kappa_0\lambda_0}) \mathit{Ga}(\lambda \mid \alpha_0, \beta_0). \end{equation} Here, $\kappa_0 > 0$, shape parameter $\alpha_0 > 0$, and rate parameter $\beta_0 > 0$. In \cref{eq:conjugate-priors}, the prior parameters $\mu_0$, $\kappa_0$, $\alpha_0$, and $\beta_0$ represent prior beliefs about the mean $\mu$ and precision $\lambda$ of a Gaussian distribution. After observing $n$ data points $X = \{X_1, \dots, X_n\}$ (for $n \geq 0 $), beliefs about $\mu$ and $\lambda$ may change. Formally, the prior parameters are updated as follows. \begin{equation}\label{eq:parameter-update} \begin{split} \mu_n =& \frac{\kappa_0\mu_0 + n\bar{X}} {\kappa_0 + n} \\ \kappa_n =& \kappa_0 + n \\ \alpha_n =& \alpha_0 + \frac{n}{2}\\ \beta_n =& \beta_0 + \frac{1}{2}nS^2_X + \frac{1}{2}\frac{\kappa_0n{(\bar{X} - \mu_0)}^2} {\kappa_0 + n} \end{split} \end{equation} where $\bar{X} = \frac{\sum_{i=1}^n X_i}{n}$ is the sample mean and $S^2_X = \frac{\sum_i^n{(X_i - \bar{X})}^2}{n}$ is the biased sample variance. Note that if there is no data, i.e., $n = 0$, then none of the prior parameters are updated in \cref{eq:parameter-update}. After observing data $X = (X_1, \dots, X_n)$ (for $n \geq 0$), the likelihood $\Pr(X_{n+1} \mid \mu, \lambda)$ of a new data point $X_{n+1}$ follows a student's $t$-distribution $t_{\nu}(X_{n+1} \mid \hat{\mu}, \hat{\sigma})$ with $\nu = 2\alpha_n$ degrees of freedom, location $\hat{\mu} = \mu_n$, and scale $\hat{\sigma} = \frac{\beta_n(\kappa_n + 1)}{\alpha_n\kappa_n}$. The probability density function of $t_{\nu}(\hat{\mu}, \hat{\sigma})$ is \begin{equation} f(X_i \mid \nu, \hat{\mu}, \hat{\sigma}) = \frac{\Gamma(\frac{\nu +1}{2})} {\Gamma(\frac{\nu}{2}) \sqrt{\nu \pi} \hat{\sigma} } {\Bigg( 1 + \frac{1}{\nu} {\Big( \frac{X_i - \hat{\mu}}{\hat{\sigma}} \Big)}^2 \Bigg)}^{-\frac{\nu+1}{2}} \end{equation} } \section{Record Selection Algorithm}\label{app:record-selection-algorithm} The AGG algorithm and the \gls{aha} variants used in our empirical study (see \cref{sec:exp-algos}) rely on a record selection strategy to find historical records to compute posterior hyperparameters. The algorithm used for record selection is presented in \cref{alg:record-selection}. The record selection algorithm takes as input training trajectories $\mathcal{T}$, the route $p$ for which travel time is in the process of being estimated, the road segment $e_i \in p$ for which historical records $\tilde{T}_i$ are currently being selected, and the arrival time $\tau_i$ at road segment $e_i$. In addition, the algorithm takes as input two parameters: an integer contextual relevance parameter, $c$, and a temporal relevance parameter, $\delta$, in some unit of time. Recall, from \cref{sec:exp-algos} that the contextual and temporal relevance hyperparameters $c$ and $\delta$ adjust the record selection strategy's emphasis on record relevance and record availability. For instance, higher values of $\delta$ typically result in more historical records being selected, but they are less indicative of the ground truth travel speed distribution. Similarly, lower values of $c$ also typically result in more historical records being selected, but they are less contextually relevant where context refers to the preceding and succeeding road segments in the trajectory from which the historical record originated. \begin{algorithm} \caption{Record Selection\label{alg:record-selection}} \begin{algorithmic}[1]\raggedright \Function{RecordSelection}{$\mathcal{T}$, $p=(e_1, \dots, e_q)$; $e_i$, $\tau_i$ $c$, $\delta$} \State{$\tilde{T}_i \gets \emptyset$} \State{$C_i \gets (e_{i-c}, \dots, e_{i-1}, e_i, e_{i+1}, \dots, e_{i+c})$}\label{line:record-selection-context-i} \ForEach{trajectory $T \in \mathcal{T}$} \State{\Let{$T = \big((e'_1, \tau_1, t_1), \dots, (e'_n, \tau_n, t_n)\big)$}} \For{$j = 1$ to $n$} \State{$C_j \gets (e'_{j-c}, \dots, e'_{j-1}, e'_j, e'_{j+1}, \dots, e'_{j+c})$} \If{$C_i = C_j$} \If{$\tau_i - \frac{\delta}{2} \leq \tau_j \leq \tau_i + \frac{\delta}{2}$} \If{$t_i \neq \varnothing$} \State{$\tilde{T}_i \gets \tilde{T}_i \cup \{t_i \}$} \EndIf \EndIf \EndIf \EndFor \EndFor \State{\Return{$\tilde{T}_i$}} \EndFunction \end{algorithmic} \end{algorithm} Given the necessary inputs, the record selection algorithm in \cref{alg:record-selection} proceeds to construct the set of historical records $\tilde{T}_i$ as follows. First, it stores the context of road segment $e_i$ in variable $C_i$ on Line~$3$. Then, the algorithm scans the set of training trajectories in the loop on {Lines~$4$-$11$} for historical records. For each trajectory, the algorithm scans each traversal in the trajectory for historical records in the loop on {Lines~$6$-$11$}. A historical record refers strictly to the recorded travel time or travel speed of the traversal. To be selected, a historical record must satisfy two conditions. First, it must be \emph{contextually relevant}, i.e.,from a traversal of road segment $e_i$ and have the same $c$ preceding and succeeding road segments in trajectory $T$ as $e_i$ in route $p$. This check is performed by comparing the contexts of the traversal $C_j$ with that of the road segment $C_i$ on {Lines~7-8} of \cref{alg:record-selection}. Second, the historical record must be \emph{temporally relevant}, i.e., occur at a similar time of week as $\tau_i$ This check is performed on Line~9. Finally, the historical record is added to $\tilde{T}_i$ on Line~$11$ if $t_i \neq \varnothing$, i.e., that the historical record is derived from \gls{gps} data rather than the map-matching algorithm. During training, $t_i$ is excluded from $\tilde{T}_i$ if it originates from the trajectory currently being forward propagated. \section{Related Work}\label{sec:related-work-paper5} As discussed in \cref{sec:introduction-paper5}, approaches for travel time and speed estimation can be categorised broadly as either function-fitting% ~\citep{wei2020spatial,lu2020st,ge2020global,zhang2020novel,zhang2020graph,zhang2020network, yin2020attention,lee2020predicting,guo2019attention,cui2019traffic,yu2017spatio,htte,stad,zheng2013time,fu2020estimation,barnes2020bustr,lan2019travel,wu2019deepeta,shen2019tcl,hu2019stochastic,hu2020stochastic,taoyang2019deepist,xi2019path,compacteta,rade2018wedge,zheng2018learning,wang2018will,wang2014travel,yang2013using,workshop} or aggregation-based% ~\citep{pace,hu2017enabling,dai2016path,yuan2011tdrive} approaches. Function-fitting approaches differ primarily in how they structure the function they are fitting, and aggregation-based approaches primarily differ in how they select or construct records for aggregation. The details of these approaches are orthogonal to the novelty of the \gls{aha} framework. The \gls{aha} framework can be used in conjunction with existing function-fitting as well as aggregation-based approaches, where existing function-fitting approaches can be used as the function-fitting component in \gls{aha} and existing aggregation-based methods can be used for record selection and construction in \gls{aha}. To the best of our knowledge, \gls{aha} is the first approach that combines function-fitting and aggregation-based approaches in a single cohesive framework. We expect that only minor modifications to the output and objective of a function-fitting approach is necessary, and that no changes to the record selection strategies of aggregation-based approaches are necessary. Approaches to travel time and speed estimation can further be categorised as segment-based~\citep{htte,yang2013using,zheng2013time,barnes2020bustr,hu2019stochastic,rade2018wedge,hu2017enabling,wang2014travel,lee2020predicting,workshop}, route-based~\citep{fu2020estimation,pace,wu2019deepeta,shen2019tcl,lan2019travel,taoyang2019deepist,xi2019path,compacteta,zheng2018learning,dai2016path,wang2018will}, origin-destination based ~\citep{stad,hu2020stochastic,yuan2011tdrive}, or station-based~\citep{wei2020spatial,lu2020st,ge2020global,zhang2020novel,zhang2020graph,yin2020attention,zhang2020network,guo2019attention,cui2019traffic,yu2017spatio}, meaning that they target estimation for segments, routes, origin-destination pairs, and traffic measuring stations, respectively. Traffic measuring stations typically represent loop detectors in traffic forecasting applications. In this paper, we have adopted a segment-based and a route-based perspective, but the \gls{aha} framework only requires that the type of data instance, be it a segment, route, origin-destination pair, or a measuring station, is represented as a feature vector. The framework is thus equally compatible with origin-destination-based and station-based approaches, given that appropriate methods of feature vector construction and record selection are available. \COMMENT{ \paragraph{Early Version of \gls{aha}} An early version of the \gls{aha} framework has been presented in the Master's thesis of \citet{estimation-updates}. The present version is generalizes the earlier version s.t.\ the earlier version is a concrete realization of \gls{aha} similar to \gls{agha}. Like \gls{agha}, the early version assumes that travel speed distributions are Gaussian, but only model uncertainty about the distribution means, not the variances, and provides only point estimates of travel time or travel speed. In addition, the early version considers only a post-training computation of the posterior similar to the UniTE-GEN algorithm used in our empirical study. As a result, the early version does not inherently regularize the function-fitting component to account for data imbalances. In addition, the present version of \gls{aha} is not restricted to discrete time like the earlier version, but also supports continuous time. Finally, this work performs presents a more extensive evaluation with an in-depth analysis of the data scalability, data efficiency, and the trade-off between data quantity and data quality when selecting historical records. } \COMMENT{ \begin{itemize} \item Trip travel time can be further segmented into segment-based~\citep{htte,yang2013using,zheng2013time,barnes2020bustr,hu2019stochastic,rade2018wedge,hu2017enabling,wang2014travel}, route-based~\citep{fu2020estimation,pace,wu2019deepeta,shen2019tcl,lan2019travel,taoyang2019deepist,xi2019path,compacteta,zheng2018learning,dai2016path,wang2018will}, and origin-destination based schemes~\citep{stad,hu2020stochastic,yuan2011tdrive}. Although we use \gls{hdse} in a segment-based fashion in our experiments, the \gls{hdse} inputs of (a) a vector representation of the route, and (b) a set of historical driving speed records is scheme-independent. Thus, the \gls{hdse} framework and the Gaussian \gls{hdse} instance presented in \cref{sec:hdse-gaussian} is equally applicable to all these schemes. \item Function-fitting methods~\citep{htte,stad,zheng2013time,hu2019stochastic,hu2020stochastic,taoyang2019deepist,xi2019path,compacteta,rade2018wedge,zheng2018learning,wang2018will,wang2014travel} \item Aggregation-based methods \citep{pace,hu2017enabling,dai2016path,yuan2011tdrive} \item Extrapolation-methods \citep{hu2019stochastic,hu2020stochastic,rade2018wedge,zheng2013time,yang2013using} use function-fitting approaches to learn parameters unique to each individual road segments. \item Aggregation-based methods struggle to find trajectories that cover all possible routes at all times of day and thus resolve to combining the travel time/speed distributions of (sub)paths of historical trajectories, either exactly~\citep{pace}, approximately~\citep{nielsen2020estimating}, or a combination thereof~\citep{pedersen2020anytime}. We expect that such approaches are also relevant to hybrid approaches that fit within the \gls{aha} framework. Such approaches are also relevant Aggregation-based approaches complement \gls{hdse} by providing a means of constructing the set of driving speed records used to compute the posterior parameters. \item Recurrent models~\citep{nielsen2020estimating,xi2019path} \end{itemize} } \section{Time of Week Vector Representation}\label{app:representation-of-time} As explained in \cref{sec:representation-of-time}, we learn a $16$-dimensional time of week vector representation, where $8$ dimensions are used to represent the time of day and $8$ dimensions are used to represent the time of the week. Formally, we learn a time-of-week vector \begin{equation} \boldsymbol\tau_{\mathit{tow}} = \boldsymbol\tau_{\mathit{tod}} \oplus \boldsymbol\tau_{\mathit{dow}} \end{equation} for time $\tau$, where $\boldsymbol\tau_{\mathit{tod}} \in \mathbb{R}^8$ represents the time of day, $\boldsymbol\tau_{\mathit{dow}} \in \mathbb{R}^8$ represents the day of the week, and $\oplus$ denotes vector concatenation. \subsection{Time of Day Representation} To represent time of day in our experiments, we divide the time of day into $96$ $15$-minute intervals $\mathcal{I} = \{I_1, \dots, I_{96} \}$ s.t.\ interval $I_1 = \interval{0:00}{0:15}$, $I_2 = \interval{0:15}{0:30}$, and so forth. Then, we one-hot encode the time of day $\tau_{\mathit{tod}}$ into a $96$-dimensional vector $\boldsymbol\tau'_{tod} = \begin{bmatrix} \mathbbm{1}[\tau_{\mathit{tod}} \in I_1] & \dots & \mathbbm{1}[\tau_{\mathit{tod}} \in I_{96}] \end{bmatrix}$ where $\mathbbm{1}$ is the indicator function. Finally, we multiply the one-hot encoding $\boldsymbol\tau'_{\mathit{tod}}$ by a trainable matrix $\mathbf{W}_{\mathit{tod}} \in \mathbb{R}^{96 \times 8}$ to achieve the time of day representation $\boldsymbol\tau_{\mathit{tod}} = \boldsymbol\tau'_{\mathit{tod}}\mathbf{W}_{\mathit{tod}}$ with dimensionality $8$. \subsection{Day of Week Representation} We represent the day of week in a manner similar to the time of day, but with $7$ dimensions in the one-hot encoding $\boldsymbol\tau'_{\mathit{dow}}$---one for each day of the week---and multiply $\boldsymbol\tau'_{\mathit{dow}}$ by a trainable weight matrix $\mathbf{W}_{\mathit{dow}} \in \mathbb{R}^{7 \times 8}$ to get an $8$-dimensional vector representation $\boldsymbol\tau_{\mathit{dow}}$ of the day of the week. \section{Method} Given an input route $R = (e_1, \dots, e_n)$ and a set of relevant historical trajectories $\mathcal{T}$ that traversed the route $\mathcal{T}$, we wish to output the travel time distribution $\Pr(T_R)$ of $R$. \subsection{Route Convolutional Later} Given a route $R (e_1, \dots, e_n)$ and feature matrices $X^V$, $X^E$, $X^B$, a route convolution performs transforms the route into a matrix $H_R = [h_1, \dots, h_n]$ where transformation of each edge $e_i \in R$ as follows \begin{equation}\label{eq:route-convolution} h_i = \sigma(W \phi^E(e_{i-1}) \oplus \phi^B(b_{i-1}) \phi(e_i) \oplus \phi^B(b_{i+1}) \oplus \phi(e_{i+1})) \end{equation} where $b_{i-1} = (e_{i-1}, e_{i})$ and $b_{i+1} = (e_{i+1}, e_{i})$, $\oplus$ denotes vector concatenation, $\phi^E(e) = X^E_e$ encodes edges, and, for a between-edge $b=((u, v), (v, w))$, $\phi^B(b) = X^B_{b} \oplus X^V_{v}$ encodes between-edges and their connecting intersection. $\sigma$ is an activation function, $W$ a matrix of trainable weights, and $b$ is a bias term. For each road segment $e_i \in R$, \cref{eq:route-convolution} captures the dependencies between the preceding and succeeding road segments. \begin{algorithm} \begin{algorithmic} \State{\Input{Route $R = (e_1, \dots, e_n)$, node feature matrix $X^V$, edge feature matrix $X^E$, between-edge feature matrix $X^B$, and the primal and dual graph representations of the road network $G^P$ and $G^D$, respectively.}} \State $(\theta^V, \theta^E, \theta^B) \gets \textsc{GCN}(X^V, X^E, X^B, G^P, G^D)$ \State $\phi \gets ()$ \For{$1 \leq i \leq n = (v_i, u_i) \in R$} \State \Let{$e_{i-1} = (v_{i-1}, u_{i-1})$} and \Let{$e_{i+1} = (v_{i+1}, u_{i+1})$} \State Draw $\phi_{v_i}$ from $\Pr(\chi_{v_i} \mid \theta^V_{u_i})$ \State Draw $\phi_{u_i}$ from $\Pr(\chi_{u_i} \mid \theta^V_{u_i})$ \State Draw $\phi_{e_i}$ from $\Pr(\chi_{u_i} \mid \theta^V_{e_i})$ \EndFor \end{algorithmic} \begin{algorithmic} \Function{DrawNodeRepresentations}{Graph Elements ()} \Comment{Draws a feature representation for input} return \EndFunction \end{algorithmic} \end{algorithm}o \subsection{Stochastic Graph Convolutional Encoder} \begin{itemize} \item Input: Node, edge, and between-edge feature matrices $X^V$, $X^E$, and $X^B$ \item Output: Transformed node, edge, and between-edge feature matrices $Z^V$, $Z^E$, and $Z^B$ that are drawn from the probability distributions $\Pr(Z^V \mid \theta^V)$, $\Pr(Z^E \mid \theta^E)$, $\Pr(Z^B \mid \theta^B)$, respectively. \item Function: Stochastically transforms the node, edge, and between-edge features matrices by using a graph convolutional mechanism to enrich them with the characteristics of the local neighborhoods of nodes, edges, and between-edges, respectively. The encoder thus exploits spatial correlations among neighbors in the road network. \end{itemize} \subsection{Route Convolutional Layer} \begin{itemize} \item Input: A route $R=(e_1, \dots, e_n)$ and the feature matrices $Z^V$, $Z^E$, and $Z^B$ from the Stochastic Graph Convolutional Encoder. \item Output: A sequence of stochastic variables $(T_1, \dots, T_n)$ that represent the segment-wise travel-time. \item Function: Estimates the travel time distribution for each road segment by taking into account its dependencies with the $k$ preceding and succeeding road segments in the route. Captures the spatial correlations within a route. \item Additional Info: Each stochastic variable $T_i$ is latent, but has observation variables $O_i$ associated with them, thus allowing the insertion of evidence from concrete traversals of the route. This enables the estimate to incorporate information from existing data. \end{itemize} \subsection{Aggregation} \begin{itemize} \item Input: The sequence of stochastic variables $(T_1, \dots, T_n)$ outputted by the Route Convolutional Layer. \item Output: A travel time distribution $T_R$ for the route $R$. \item Function: \end{itemize} \todo{You can use folded or truncated normal distributions to avoid modeling negative speeds} \todo{Useful website on addition and multiplication with Gaussians: http://www.kaspercpa.com/statisticalreview.htm} \begin{figure} \centering \input{illustrations/conceptual_framework.tikz.tex} \caption{Conceptual Framework\label{fig:conceptualFramework}} \todo{I believe the current to next arrow can be switched to produce an equivalent framework. This way we could reflect the dependencies using a HMM-like net.} \end{figure} We first construct a conceptual framework of our approach, illustrated in \cref{fig:conceptualFramework}. We observe that the duration $D^T$ of a trip $T=(e_1, \dots, e_k)$ depends on the traversal speed $S_i$ across each segment $e_i$ in trip $T$. Next, we observe that the traversal speed $S_i$ across segment $e_i \in R$ depends on the driving speed $S_{i-1}$ on the preceding segment $e_{i-1}$ in the trip. For instance, if $e_{i-1}$ is highly congested (thus reducing the traversal speed), a vehicle takes longer to accelerate on segment $e_i$ than if $e_{i-1}$ had no congestion. Similarly, the traversal speed $S_i$ across segment $e_i$ also depends on the traversal speed $S_{i+1}$ across the succeeding segment $e_{i+1}$: if the succeeding segment is congested, a vehicle slows down (voluntarily or otherwise). Beside the traversal speed across preceding and succeeding segments, the trips $T_{1, i}$ and $T_{i, |P|}$, to and from segment $e_i$, respectively, also affects the traversal speed $S_i$ across $e_i$. For instance, if a vehicle is required to make a sharp turn to transition from the preceding segment $e_{i-1}$ to a segment $e_i$, then the vehicle decelerates in anticipation of the turn which affects the speed $S_i$ on segment $e_i$ where time must be spent to accelerate again. Vice versa, if the vehicle is required to make a sharp turn onto the succeeding segment then the vehicle decelerates on segment $e_i$ in anticipation of the turn thus affecting the speed $S_i$ on the segment. In general, the dependency on predecessor and successor segments is not limited to immediate successors. For instance, roundabouts consists of very short segments of a few meters where the immediate predecessors or successors in the trip fail to explain the driving speed completely. \todo{May be relevant to cite some related work on aggregation models here.} \subsection{SequentialRFN} Sequential relational aggregation \subsubsection{Prospect 1: Position-wise change} Given a trip of nodes Compute a relational aggregate of each node in the trip as usual Combine the aggregated representation with the sequential representation - E.g., by concatenating the representations and joining them through a neural network Advantages: Could work for generic graph convolutional networks Disadvantages: - How to take time into account sequentially? \subsubsection{Prospect 2: Position-sensitive fusion function} Given a trip of nodes Compute a relational aggregate, but during fusion use the sequential representation to inform the aggregate - Could be quite expensive since a relational aggregate would be computed for each node in the trip - Is there a way to subvert this? Maybe use that additive fusion is essentially a sum? - Multiply the sequential representation by the number of neighbors and otherwise add the remaining transformations? Advantages: Likely a more powerful model Disadvantage: - Could be computationally expensive. - In addition, the relational fusion networks are relatively unknown and not yet an established method. - How to take time into account sequentially? \subsubsection{Time-dependent model} \paragraph{Estimation of Segment Arrival Times} The arrival time at the $i$th segment of a trajectory $T=(e_1, \dots, e_2)$ is the arrival time at the previous segment plus the time it takes to traverse the previous segment. Formally, the arrival time of the $i$th segment is \begin{equation}\label{eq:arrival-time} T^{\tau}_i = T^{\tau}_{i-1} + \frac{l_{i-1}}{s_{i-1}} \end{equation} where $i > 1$ and $T^{\tau}_1$ is the starting time of the trip in number of time units (e.g., seconds) after midnight. Here, $l_{i-1}$ is the length of segment $e_{i-1}$ and $s_{i-1}$ is the traversal speed across segment $e_{i-1}$. In \cref{eq:arrival-time}, $s_i$ is an unknown quantity and we therefore estimate it using a neural network, but ensure that $s_i > 0$. Note, however, that the estimate of $s_i$ is used only for estimating the arrival time, and not to estimate the travel time of trip $T$. \paragraph{Arrival Time Encoding} Travel times across a segment exhibit strong temporal correlations s.t.\ the travel time is similar within, e.g., a $15$~minute period. We therefore desire an arrival time encoding that assign similar encoding to points in time that are temporally close. To this end, we use a two-dimensional representation that encodes a point of time as a position on a ''clock'': \begin{equation}\label{eq:time-encoding} f_{\tau}(\tau) = \begin{bmatrix}\cos(\tau) & \sin(\tau)\end{bmatrix} \end{equation} where $\tau$ is measured in time units after midnight. The time encoding function $f_{\tau}$ has the property that, e.g., $0:05$ AM and $23:55$ are closer in, e.g., a Euclidean distance sense, than $0:05$ and $0:30$. Thus, $f_{\tau}$ can capture the temporal correlations as desired. However, in practice, there may also be correlations across different times of the day that are relatively distant such as morning and afternoon peaks. In addition, we sum the time, edge, and positional encodings. We therefore seek a time encoding that does not restrict the edge and positional encodings to any particular number of dimensions. To this end, we transform the encoding as follows: \begin{equation} \phi_{\tau}(\tau) = f_{\tau}(\tau)\mathbf{W} \end{equation} where $\mathbf{W} \in \mathbb{R}^{2 \times d}$ is a learnable projection matrix of parameters that projects the time encoding into a $d$-dimensional vector space. Travel time across a route are temporally correlated s.t.\ Previous work has successfully leveraged the strong temporal correlations in driving speeds and travel times A key property A key feature of time The intuition is that the arrival time at the $i$th segment Compute relational aggregate When computing relational aggregate for each node v in a trip t: for each neighbor n of v: perform relational Compute regular relational fusion aggregates For each entry \subsection{Loss Function} \todo{Need to introduce that the travel time of a trip is a stochastic variable.} We observe that, by the laws of physics, the travel time of a trip $T=(r_1, \dots, r_n)$ is equal to the sum of the travel time across each of its constituent edges $e_i \in T$. Thus, the likelihood of a trip $T$ is the joint likelihood of the travel time across its constituent segments $\mathcal{L}(T) = \mathcal{L}(T_{0, 1}, T_{1, 2}, \dots, T_{n-1, n})$. We therefore aim to optimize this likelihood by minimizing the negative log-likelihood of the trips \begin{equation} L_1(T) = \sum_{(e_i, s_i) \in T} -\ln\Bigg( L\bigg( \frac{l_i}{s_i} \bigg) \Bigg) \end{equation} where \todo{Remove assumptions about the Gaussian nature of the speed variables.} We compute a bipartite loss function \begin{equation} \sum_{T \in \mathcal{T}} \alpha L_1(T) + (1-\alpha)L_2(T) \end{equation} where $\alpha, 0 \leq \alpha \leq 1$ regulates the importance between the two sub-losses $L_1$ and $L_2$. The first sub-loss is used to encourage our method to learn accurate trip travel time estimates. It is defined as the negative log-likelihood of $L_1(T) = \sum_{}$ In other words, $y_T = \sum_{e_i \in T} y_{e_i}$ where $y_{e_i} = \frac{l_{e_i}}{s_{e_i}}$. Motivated by this observation, we optimize our proposed method at the edge level. Given a trip $T=(r_1, \dots, r_n)$, we minimize the \emph{negative log-likelihood} of the speed recorded for all road segments in the trip: \begin{equation}\label{eq:edge-loss} L_2(T) = \sum_{(e_i, s_i) \in T, s_i \neq \varnothing} -\ln(\mathcal{L}(s_i)) \end{equation} \todo{should probably normalize s.t.\ the number of known speeds does not affect the magnitude of the loss} The loss in \cref{eq:edge-loss} encourages our method to be robust to the length of the trip since optimization is performed at the edge level. However, as discussed previously, \gls{gps} sampling frequency can cause some road segments not to be associated with any speed and these road segments are filtered out during summation in \cref{eq:edge-loss}. Road segments with missing recorded speeds tend to be relatively short segments and therefore the optimizing the loss in \cref{eq:edge-loss} encourages good predictions on longer segments which are unlikely to have missing speed records in trips. To alleviate this problem we introduce another loss term to \cref{eq:edge-loss}: We divide the loss function in two parts. The first part optimizes the full trip duration prediction. In order to train the network we We now present the loss function for our proposed methodWe design the loss functionw We compute a joint loss. We predict both the By predicting A physical constraint in A key property of our method is the reliance on the physical constraint that \section{Related Work} \section{Experiments} We now proceed to evaluate our proposed method by comparing it against a number of baselines on the task of trip travel time estimation in the Danish road network. \subsection{Data Set} We extract the most important roads in the Danish road network from \gls{osm}~\citep{osm} and convert them into a graph representation as described in \cref{sec:roadNetworkModelling}. This yields a graph with ?? nodes and ?? edges. Then, we mapmatch a set of ?? vehicle trajectories to the road network and divide them into training, validation, and test sets. \subsection{Algorithms} \todo{Our method with and without observation updates} \todo{Sum of independent estimations similar to relational fusion network paper} \subsection{Experimental Setup} We measure the average likelihood of the duration of each trip for our proposed method and all probabilistic baselines. This allows us to evaluate how well our proposed method estimates the trip duration density function compared to the baselines. \todo{Maybe it would be useful to have some idea of optimality? Maybe Oracle models?} Not all the baselines are capable of estimating the trip duration density function. We therefore evaluate all models using \gls{mae}. \todo{Write definition for MAE} In the case of the non-probabilistic models, we measure the \gls{mae} between the model prediction and the ground truth trip duration. For the probabilistic models, we measure the \gls{mae} between the most likely trip durations against the ground truth trip duration. \subsection{Travel Time Estimation} \subsection{Density Estimation} \todo{Perhaps it is possible to estimate the density function of a given route s.t. one can see the probability distribution over trip durations by doing Monte Carlo simulation of each route.} \section{Preliminaries} \subsection{Modeling Road Networks}\label{sec:road-network-modelling} \todo{Rewrite, currently taken directly from RFN paper} We model a road network as an attributed directed graph $G = (V, E, A^V, A^E, A^B)$ where $V$ is the set of nodes and $E$ is the set of edges. Each node $v \in V$ represents an intersection, and each edge $(u, v) \in E$ represents a road segment that enables traversal from $u$ to $v$. Next, $A^V$ and $A^E$ maps intersections and road segments, respectively, to their attributes. In addition, $A^B$ maps a pair of road segments $(u, v), (v, w) \in E$ to their between-segment attributes such as the angle between $(u, v)$ and $(v, w)$ based on their spatial representation. An example of a graph representation of the three-way intersection to the right in figure \cref{fig:compelling-example} is shown in \cref{fig:graph-rep-primal}. Two intersections $u$ and $v$ in $V$ are adjacent if there exists a road segment $(u, v) \in E$ or $(v, u) \in E $. Similarly, two road segments $(u_1, v_1)$ and $(u_2, v_2)$ in $E$ are adjacent if $v_1 = u_2$ or $v_2 = u_1$. The function $N\colon V \cup E \xrightarrow{} 2^V \cup 2^E$ returns the neighborhood, i.e., the set of all adjacent intersections or road segments, of a road network element $g \in V \cup E$. The dual graph representation of $G$ is then $G^D=(E, B)$ where $B = \big\{ \big((u, v), (v, w) \big) \mid (u, v), (v, w) \in E \big\}$ is the set of \emph{between-edges}. Thus, $E$ and $B$ serve as the node and edge sets of the dual graph, respectively. For disambiguation, we refer to $G$ as the primal graph representation. \subsection{Routes and Trips} A route $R = (e_1, \dots, e_n)$ is in a road network with primal graph $G^P=(V, E)$ and dual graph $G^P=(E, B)$ if $\forall e \in R, e \in E$ and $(e_i, e_{i+1}) \in B$ for $1 \leq i \leq n-1$. In addition, let $X^V$, $X^E$, and $X^B$ be the feature matrices that describe nodes, edges, and between-edges, respectively. A \emph{trip} $T=(R, \tau)$ consists of a route $R = (e_1, \dots, e_n)$ in $G$ and a start time $\tau$. For convenience, we say that a segment $e$ is in a trip $T=(R, \tau)$ if $e \in R$ and denote it using the notation $e \in T$. \subsection{Feature Matrices} The feature matrices $\mathbf{X}^V$, $\mathbf{X}^E$, and $\mathbf{X}^B$ encode the attributes of nodes ($A^V$), edges ($A^E$), and between-edges ($A^B$), respectively, as feature vectors s.t.\ each row row corresponds to a node, edge, or between-edge. We use the notation $\mathbf{X}^V_v$, $\mathbf{X}^E_e$, and $\mathbf{X}^B_b$ to access the feature vectors of node $v$, edge $e$, and between-edge $b$, respectively. In some contexts, we use the abbreviated notations $\mathbf{X}^V_i = \mathbf{X}^V_{v_i}$, $\mathbf{X}^E_i = \mathbf{X}^E_{e_i}$, and $\mathbf{X}^B_i = \mathbf{X}^B_{b_i}$. \subsection{Route Convolutions} A route convolution is a function $C$ takes as input a route $R = (e_1, \dots, e_n)$ and maps each edge $e \in R$ to a vector representation $c(v)$, s.t.\ \begin{equation} C(R) = \big(c(e_1), \dots, c(e_n)\big) \end{equation} In this framework, different choices of the convolution function $c$ lead to different route convolutions. For instance, previous work \todo{citation needed} include one-dimensional convolutional layers in their architecture. A one-dimensional convolutional layer with kernel size $k=1$ uses \begin{equation} c(e_i) = \sigma\big(\mathbf{W} (\mathbf{X}^E_{i-1} \oplus \mathbf{X}^E_{i} \oplus \mathbf{X}^E_{i+1}) + \mathbf{b}\big), \end{equation} where $\sigma$ is an activation function, $\mathbf{W}$ is a learnable weight matrix, $\oplus$ denotes vector concatenation, and $\mathbf{b}$ is a bias term. The term route convolution implies that information about the current edge $e_i$ is combined with that of its predecessors and successors in the route \todo{perhaps this is arguable}. However, that is not necessarily the case. For instance, if $c(e_i) = \sigma( \mathbf{W}\phi^E(e_i) + \mathbf{b})$. \COMMENT{ \subsection{Modeling Trips} A \emph{trip} in a road network $G=(V, E)$ is a pair $T=(R, \tau)$ that consists of a route $R = (e_1, \dots, e_n)$ in $G$ and a timestamp $\tau$ denoting the start of the trip. For convenience, we say that a segment $e$ is in a trip $T=(R, \tau)$ if $e \in R$ and denote it using the notation $e \in T$. The travel time of a trip $T$ is a random variable $\mathit{TT}_T$ that has the probability density function $\Pr(TT_T) = \prod_{e \in T} \Pr(TT_e)$ where $TT_e$ is random variable representing the travel time across segment $e$ in trip $T$. A historical trip $T$ For the purposes of estimating A \emph{trip} is a sequence of \emph{trip records} $T=(r_1, \dots, r_n)$. Each trip record $r_i$ in a trip $T$ is triple of the form $(e_i, \tau^a_i, \tau^d_i)$ where $e_i \in E$ is a road segment in a road network $G=(V, E)$. $\tau^a_i$ and $\tau^d_i$ are timestamps indicating the time and date the vehicle arrived to and departed from road segment $e_i$, respectively. If the arrival and departure times are unknown, e.g., due to \gls{gps} sampling frequency, then $\tau^a_i = \varnothing$ and $\tau^d_i = \varnothing$. The \emph{travel time} of a trip $T=(r_1, \dots, r_n)$ is the difference between the arrival time at the first road segment in the trip and the departure time from the last road segment in the trip, i.e., $y_T = \tau^d_n - \tau^a_1$ where $\tau^d_n \neq \varnothing$ and $\tau^a_1 != \varnothing$. \subsection{Relational Fusion Networks} } \subsubsection{Bayesian Parameter Estimation of a Gaussian Mixture} It is well-established in the literature that driving speeds do not follow any simple distribution, such as a Gaussian. We therefore extend the single-component approach described in \cref{sec:parameter-estimation-gaussian} to the Gaussian mixture case. In this case, the model $f$ is required to output prior parameters $w_0$, $\mu_0$, $\lambda_0$, $\alpha_0$, and $\beta_0$ for each of the $K$ components. In brief, we introduce an update rule for the component weight of a Gaussian component $C_k$ in the mixture. In addition, we modify the update rules in \cref{eq:parameter-update} s.t.\ the effect of a driving speed sample $d$ on the update of the prior parameters of a component $C_k$ is proportional to the probability $\Pr(C_k \mid d)$ that $d$ is generated by that component. This ensures that the component means and variances do not all converge towards the sample mean and sample variance when there are is more than one component. Given evidence $E = \{d_1, \dots, d_n\}$, we compute the posterior parameters of each Gaussian component $C_k$ in the mixture as follows. \begin{equation}\label{eq:mixture-component-update} \begin{split} w_n &= \frac{\kappa_0 w_0 + nw_d} {\kappa_0 + n} \\ \mu_n &= \frac{\kappa_0 \mu_0 + p\bar{d}} {\kappa_0 + p} \\ \kappa_n &= \kappa_0 + p \\ \alpha_n &= \alpha_0 + \frac{p}{2} \\ \beta_n &= \beta_0 + \frac{1}{2} p S^2 + \frac{1}{2}\frac{\kappa_0 p {(\bar{d} - \mu_0)}^2} {\kappa_0 + p} \end{split} \end{equation} where \begin{equation*} \begin{split} w_d &= \frac{1}{n} \sum_{i=1}^n\frac{\Pr(C_k \mid d_i)}{\sum_{j=1}^K\Pr(C_j \mid d_i)} \\ p &= \sum_{i=1}^n \frac{{\Pr(C_k \mid d_i)}} {\sum_{j=1}^K \Pr(C_j \mid d_i)}, \\ \bar{d} &= \frac{\sum_{i=1}^n \Pr(C_k \mid d_i) x_i}{\sum_{j=1}^n \Pr(C_k \mid d_j)}, \text{and} \\ S^2 &= \frac{\sum_{i=1}^n{\Pr(C_k \mid d_i)(d_i - \bar{d})}^2}{\sum_{j=1}^n \Pr(C_k \mid d_j)}. \end{split} \end{equation*} Here, $w_d$ represents the component weight of component $C_k$ when estimated from the evidence $E$, and $p$ is a pseudo count representing the expected number of driving to have originated from component $C_k$. Similarly, $\bar{d}_k$ and $S^2_k$ is the expected sample mean and expected biased sample variance of the samples originating from component $C_k$. The primary differences between the Gaussian parameter update rules in \cref{eq:parameter-update} and the Gaussian mixture update rules in \cref{eq:mixture-component-update} is that \begin{enumerate} \item an update rule for the component weights in the mixture is introduced, and \item that the influence of the given samples in the update of a component $C_k$ is proportional to the number of expected samples originating from $C_k$. \end{enumerate} In other words, the samples given as evidence do not increase certainty about the mean $\mu$ and the variance $\sigma^2$ of component $C_k$ substantially if the samples are unlikely to have originated from component $C_k$.

15de198262db1d3681c779235d0e3645c51ae5e9

\section{Introduction} A finite permutation group $G$ acting on a finite set $\Omega$ is \emph{synchronising} if the automaton whose transitions are generators of $G$ together with an arbitrary non-permutation is synchronising. This means that there is a word in the automaton's alphabet (called a \emph{reset} word) such that, after reading this word, the automaton is in a known fixed state, regardless of its starting state. Every synchronising permutation group is primitive \cite[Theorem 3.2]{survey}, and moreover, it has been shown that they are of \emph{affine}, \emph{almost simple}, or \emph{diagonal} type \cite[Proposition 3.3]{survey}. There are many examples known of synchronising groups of affine or almost simple type, but the existence of synchronising groups of diagonal type has been open ever since the birth of this subject (see \cite{Bray}). Bray et. al. \cite{Bray} showed that the structure of a putative synchronising group of diagonal type is constrained, in that its socle contains just two factors. In addition, they showed that if $G$ is primitive of diagonal type, then $G$ is synchronising if and only if it is \emph{separating}. So for primitive permutation groups of diagonal type, we have the following hierarchy: \[ \text{spreading } \implies \text{ separating } \iff \text{ synchronising}. \] See Section \ref{sect:background} for relevant background theory about the synchronisation hierarchy of primitive permutation groups. In this paper, we investigate the group $\PSL(2,q)\times \PSL(2,q)$ in its diagonal action on $\PSL(2,q)$ and find the first known examples of synchronising groups of diagonal type. \begin{theorem}\label{main1} The group $\PSL(2,q)\times \PSL(2,q)$ in its diagonal action on $\PSL(2,q)$ is separating, and hence synchronising, for $q=13$ and $q=17$. \end{theorem} We were led to this family of groups because they form a natural class of diagonal type groups whose socle has only two factors. Neumann \cite[Example 3.3]{Neumann}, see also \cite[\S6.4]{survey}, observed that if a nonabelian simple group $T$ admits an \emph{exact factorisation} into two nontrivial proper subgroups $A$ and $B$ (that is, $T=AB$ with $A\cap B=\{1\}$), then $T\times T$ in its diagonal action on $T$ is non-synchronising (and therefore non-separating and non-spreading). It\^{o} \cite{ito} has shown that $\PSL(2,q)$ has an exact factorisation unless $q\equiv 1\pmod 4$ and $q \notin \{5,29\}$. As a result, the first few values of $q$ for which $\PSL(2,q) \times \PSL(2,q)$ might be synchronising are $q=9$, $q=13$, $q=17$. The group $\PSL(2,9)$ does not have an exact factorisation into subgroups, but we use a slightly more general concept to show that $\PSL(2,9) \times \PSL(2,9)$ is non-synchronising, thereby resolving its position in the synchronisation hierarchy (Proposition \ref{A6}). Then for $q=13$ and $q=17$, we use a combination of theory and computation to show that $\PSL(2,q) \times \PSL(2,q)$ acting in its diagonal action on $\PSL(2,q)$ is synchronising. Finally, we show that none of the groups in this family are spreading. \begin{theorem}\label{main2} The group $\PSL(2,q)\times \PSL(2,q)$ acting in its diagonal action on $\PSL(2,q)$, for $q$ a prime power, is non-spreading. \end{theorem} This leaves the following open problem to completely resolve the synchronisation hierarchy for groups of this type. \begin{problem} For which prime powers $q$ such that $q \equiv 1 \pmod{4}$, $q \geq 25$ and $q \ne 29$ is the group $\PSL(2,q) \times \PSL(2,q)$ acting in its diagonal action on $\PSL(2,q)$ synchronising? \end{problem} The smallest $q$ for which this is unresolved is $q=25$. Any argument for the general case will not be completely straightforward because the group is synchronising for $q=13$ and $q=17$, but not synchronising for $q=29$. To demonstrate that a specific group $G$ is separating or synchronising involves (at least in principle) determining the clique number, coclique number and/or chromatic number for every single $G$-invariant graph. These parameters are notoriously hard to compute for even a single graph, and there can be a large number of $G$-invariant graphs that must be considered. For this reason, we strive to eliminate as many graphs as possible by theoretical means when proving Theorem \ref{main1}. The technique that we introduce results in a significant reduction in the number of graphs that must be considered by computer. \section{Background}\label{sect:background} \subsection{Synchronising, separating, spreading} A transformation semigroup on a set $\Omega$ is called \emph{synchronising} if it contains a constant map (that is, a map $f$ and $\beta\in \Omega$ such that $\alpha^f=\beta$ for all $\alpha\in\Omega$). A finite permutation group $G$ acting on a set $\Omega$ is \emph{synchronising} if for any non-bijective transformation $f$, the transformation semigroup $\langle G,f\rangle$ is synchronising. Alternatively, each of the following conditions is equivalent to $G$ being non-synchronising: \begin{enumerate} \item There is a nontrivial $G$-invariant graph $\Gamma$ having its clique number $\omega(\Gamma)$ equal to its chromatic number of $\Gamma$ (see \cite[Corollary 4.5]{survey}). \item There is a partitition $\mathcal{P}$ of $\Omega$ with transversal $B$ such that $B$ is a transversal of $\mathcal{P}^g$ for all $g\in G$ (see \cite[Theorem 3.8]{survey}). \end{enumerate} We need some notation for both sets and multisets that will be used throughout the paper. Suppose that $A$ is a set or multiset whose elements belong to $\Omega$. Then the \emph{characteristic vector} $\chi_A$ of $A$ is the vector indexed by $\Omega$ where $(\chi_A)_\alpha$ is the multiplicity of $\alpha$ in $A$. The cardinality of $A$ is denoted either $|A|$ or $|\chi_A|$. A transitive permutation group $G$ acting on a set $\Omega$ is \emph{non-spreading} if there exists a non-trivial multiset $A$, a non-trivial set $B$, and a positive integer $\lambda$, such that \begin{enumerate} \item $|A|$ divides $|\Omega|$, \item $|\chi_A \circ \chi_{B^g}|=\lambda$ for all $g\in G$, where $\circ$ is the Schur product of two vectors. \end{enumerate} If in addition, $\lambda = 1$, then $A$ must be a set such that $|A||B|=|\Omega|$, in which case $G$ is said to be \emph{non-separating} \cite[\S5.5]{survey}. We can also use invariant graphs to characterise non-separating group actions. A transitive permutation group $G$ acting on a set $\Omega$ is non-separating if and only if there is a nontrivial $G$-invariant graph $\Gamma$ having $\alpha(\Gamma)\cdot \omega(\Gamma) = |V\Gamma|$ (see \cite[Theorem 4.5]{survey}), where $\alpha(\Gamma)$ and $\omega(\Gamma)$ are the size of the largest coclique and clique of $\Gamma$, respectively. Each of the properties synchronising, separating, and spreading are propagated to overgroups. That is, if $G\le H\le \Sym(\Omega)$ and $G$ is synchronising (resp. separating, resp. spreading) then $H$ is also synchronising (resp. separating, resp. spreading). \subsection{Association schemes} Let $\Omega$ be a set, and let $A_0, A_1, \ldots, A_d$ be symmetric $\{0,1\}$-matrices indexed by $\Omega$. Then $\mathcal{A}=(\Omega, \{A_0, A_1, \ldots, A_d\})$ is an \emph{association scheme} if the following conditions hold: \begin{enumerate} \item $A_0$ is the identity matrix, \item $\sum_{i=0}^d A_i$ is the matrix with every entry equal to $1$, \item There exist constants $p_{ij}^k$ depending only on $i$, $j$, and $k$, such that \[ A_i A_j = \sum_{k=0}^d p_{ij}^k A_k. \] \end{enumerate} The matrices $A_0,A_1,\ldots A_d$ are the \emph{adjacency matrices} of $\mathcal{A}$. Indeed each $A_i$ is the adjacency matrix of an undirected graph. We shall refer to these as the \emph{graphs in the scheme~$\mathcal{A}$}. A graph is the \emph{union of graphs in the scheme~$\mathcal{A}$} if its adjacency matrix is the sum of adjacency matrices of $\mathcal{A}$. It is well known that $\mathbb{C}^\Omega$ decomposes into $d+1$ simultaneous eigenspaces for the adjacency matrices of $\mathcal{A}$. Moreover there are projection matrices $E_0, E_1, \ldots, E_d$ onto each of these eigenspaces, such that \[ E_i = \sum_{j=0}^d Q_{ji} A_j, \] where $Q$ is called the \emph{matrix of dual eigenvalues}. If $C$ is a subset of $\Omega$, then its \emph{inner distribution} is the vector $a = (a_0,a_1,\ldots,a_d)$ defined by \[ a_i= \frac{1}{|C|}\chi_C A_i \chi_C^\top. \] If $Q$ is the matrix of dual eigenvalues of $\mathcal{A}$, then \[ (a Q)_j = \frac{|\Omega|}{|C|} \chi_C E_j \chi_C^\top \] for all $j\ge 0$. This vector is sometimes known as the \emph{MacWilliams transform} of $C$. The \emph{dual degree set} of $C$ is the set of nonzero indices $j$ for which the $j$-th coordinate of its MacWilliams transform is nonzero. Two subsets of $\Omega$ are \emph{design-orthogonal} if their dual degree sets are disjoint. We refer the reader to the classic text by Bannai and Ito \cite{BannaiIto1984} for more information on this subject. \begin{theorem}[\cite{delsarte}, Theorem 3.9 (and discussion thereafter); see also \cite{gm}] \label{thm:CliquesInSchemes} Let $\mathcal{A}$ be an association scheme on $\Omega$ and let $\Gamma$ be a graph whose adjacency matrix is the sum of some of the adjacency matrices of $\mathcal{A}$. If $C$ is a clique and $S$ is a coclique in $\Gamma$, then $|C|\cdot|S|\le |\Omega|$. Equality holds if and only if $C$ and $S$ are design-orthogonal. \end{theorem} We will need the following simple observation, \begin{corollary}[cf., {\cite[3.3 Corollary]{roos}}]\label{intersect1} Let $\mathcal{A}$ be an association scheme on $\Omega$ and let $\Gamma$ be a union of some of the graphs in the scheme. If $C$ is a clique and $S$ is a coclique in $\Gamma$, such that $|C|\cdot|S|= |\Omega|$, then $|C\cap S|=1$. \end{corollary} \subsection{Diagonal actions}\label{sec:diagonal} Let $T$ be a nonabelian simple group. Then $G=T\times T$ acts on $\Omega=T$ in the \emph{diagonal action} as follows: \[ t^{(x,y)}:=x^{-1}ty. \] This action is always primitive. Note that the direct factors $1\times T$ and $T\times 1$ are regular normal subgroups of $G$. Moreover, $G_{1_T}=\{(t,t)\mid t\in T\}$ induces the inner automorphism group $\Inn(T)$ on $\Omega$, and so its orbits on $\Omega$ are the conjugacy classes of $T$. An \emph{orbital} for a group $G$ acting on a set $\Omega$ is an orbit of $G$ on $\Omega\times \Omega$. A graph with vertex set $\Omega$ is $G$-invariant if and only if its edge-set is a union of orbitals. Every such graph for $G=T\times T$ acting in its diagonal action on $T$ is a Cayley graph for $T$ with connection set being a union of conjugacy classes of $T$. Now if $T$ has an exact factorisation $T=AB$, consider $T$ acting on the set $\Sigma$ of right cosets of $A$. Then $B$ is a regular subgroup in this action and so is a transversal for the partition $\mathcal{P}$ of $\Omega=T$ into right cosets of $A$. Following \cite[Example 3.3]{Neumann}, for all $g=(x,y)\in G$, the image of $\mathcal{P}$ under $g$ is the partition of $T$ into right cosets of $A^{x}$. Since $B$ is also a regular subgroup for the action of $T$ on the set of right cosets of $A^{x}$ it follows that $B$ is also a transversal for the partition $\mathcal{P}^g$ and so $G$ is non-synchronising. Similarly, if $A\leqslant T$ and $B$ is a sharply-transitive set for the action of $T$ on the set of right cosets of $A$ then $B$ is a transversal for the partition of $T$ into the set of right cosets of $A^x$ in $T$, for any $x\in T$. (Recall that $B\subseteq T$ is a \emph{sharply-transitive set} for the action of $T$ on a set $\Sigma$ if for all $\alpha,\beta\in \Sigma$ there is a unique $b\in B$ such that $\alpha^b=\beta$.) Hence, $G$ is non-synchronising in this case as well, see \cite[\S 6.4]{survey}. We show that if $T=\PSL(2,9)$ then the diagonal action of $T\times T$ on $T$ is non-synchronising even though $\PSL(2,9)$ does not have an exact factorisation. \begin{proposition}\label{A6} Let $T=\PSL(2,9)$ and let $G$ be $T\times T$ acting in diagonal action on $T$. Then $G$ is non-synchronising. \end{proposition} \begin{proof} First recall that $T\cong A_6$. Let $A$ be $A_5\leqslant T$ and let $B$ be a sharply-transitive set of permutations in $A_6$, in the action of $A_6$ on six points. (Note: $B$ is well-known to exist. For example, take \small{$\{ (), (1\,2)(3\,4\,5\,6), (1\,3)(2\,4\,6\,5), (1\,4)(2\,5\,3\,6), (1\,5)(2\,6\,4\,3), (1\,6)(2\,3\,5\,4) \}$}). Then as discussed in the previous paragraph, $G$ is non-synchronising. \end{proof} \section{Analysis} Let $T=\PSL(2,13)$ and let $G$ be $T\times T$ acting in diagonal action upon $T$. Now $T$ has eight conjugacy classes of nontrivial elements, which we label \cc{2}, \cc{3}, \cc{6}, \cc{7A}, \cc{7B}, \cc{7C}, \cc{13A}, and \cc{13B}, according to the orders of the elements. For $I \subseteq \{\cc{2}, \cc{3}, \cc{6}, \cc{7A}, \cc{7B}, \cc{7C}, \cc{13A}, \cc{13B}\}$, we shall denote by $\Gamma_I$ the Cayley graph on $T$ whose connection set is the set of elements in the union of conjugacy classes determined by $I$. We suppress parentheses in the subscripts, so for example, $\Gamma_{\cc{3},\cc{6}}$ has as connection set the elements of order $3$ or $6$. If $G$ were non-separating, then this would be witnessed by the existence of some $G$-invariant graph $\Gamma$ such that $\alpha(\Gamma)\cdot \omega(\Gamma) = 1092$. We aim to show that no such graph exists, which ostensibly requires examining the 256 graphs of the form $\Gamma_I$. Without loss of generality we may assume that $I$ is non-empty, and as $\alpha(\Gamma) \cdot \omega(\Gamma)$ is invariant under graph complementation, we only need to check one of each pair of complementary graphs, thereby leaving $127$ distinct graphs to check. Denote the union of conjugacy classes with elements of order $7$ and $13$ by \cc{7} and \cc{13}, respectively. The following proposition drastically reduces the number of graphs which must be considered down to $15$. \begin{proposition}\label{prop:FewerGraphs} The group $\PSL(2,13)\times \PSL(2,13)$ acting in its diagonal action on $\PSL(2,13)$ is separating if and only if $\alpha(\Gamma_I) \cdot \omega(\Gamma_I) \neq 1092$ for all $I \subseteq \{\cc{2}, \cc{3}, \cc{6}, \cc{7}, \cc{13}\}$. \end{proposition} To prove Proposition \ref{prop:FewerGraphs}, we shall consider two association schemes. Let $A_I$ be the adjacency matrix of $\Gamma_I$ with respect to $\Omega$. Then up to a reordering of relations, the orbitals of $G$ produce the following association scheme \[ \mathcal{A}=(\Omega, \{A_\cc{1}, A_\cc{6}, A_\cc{2}, A_\cc{3}, A_{\cc{7A}}, A_{\cc{7B}}, A_{\cc{7C}}, A_{\cc{13A}}, A_{\cc{13B}} \}), \] where $A_\cc{1}$ is the identity matrix. We shall fix this ordering of the relations. Note the unusual ordering of the $A_i$, which corresponds to the matrix of dual eigenvalues used in the following calculations. It is well known that $\mathcal{A}$ is equal to an association scheme\footnote{In general, the orbitals of a transitive permutation group do not form a (symmetric) association scheme. They do, however, form a homogeneous coherent configuration.} called the \emph{group scheme} for $T$, and the algebra generated by the adjacency matrices is the centre of the group algebra $\mathbb{C}T$. The matrix of dual eigenvalues $Q^\mathcal{A}$ is readily computable from the character table of $T$: \[ Q^\mathcal{A}:=\small{ \begin{bmatrix} 1& 49& 49& 144& 144& 144& 169& 196& 196\\ 1& -7& -7& 0& 0& 0& 13& -14& 14 \\ 1& -7& -7& 0& 0& 0& 13& 28& -28 \\ 1& 7& 7& 0& 0& 0& 13& -14& -14 \\ 1& 0& 0& -24\cos{\frac{2\pi}{7}}& -24\cos{\frac{4\pi}{7}}& -24\cos{\frac{6\pi}{7}}& -13& 0& 0\\ 1& 0& 0& -24\cos{\frac{6\pi}{7}}& -24\cos{\frac{2\pi}{7}}& -24\cos{\frac{4\pi}{7}}& -13& 0& 0 \\ 1& 0& 0& -24\cos{\frac{4\pi}{7}}& -24\cos{\frac{6\pi}{7}}& -24\cos{\frac{2\pi}{7}}& -13& 0& 0 \\ 1& \frac{7}{2}(1-\sqrt{13})& \frac{7}{2}(1+\sqrt{13})& -12& -12& -12& 0& 14& 14 \\ 1& \frac{7}{2}(1+\sqrt{13})& \frac{7}{2}(1-\sqrt{13})& -12& -12& -12& 0& 14& 14 \end{bmatrix}.} \] Furthermore, we may fuse the relations \cc{7A}, \cc{7B}, and \cc{7C}, and also \cc{13A} and \cc{13B} to form a second association scheme \[ \mathcal{B}=(\Omega, \{A_\cc{1}, A_\cc{6}, A_\cc{2}, A_\cc{3}, A_\cc{7}, A_{\cc{13}} \}), \] with matrix of dual eigenvalues \[ Q^\mathcal{B}:= \begin{bmatrix} 1& 98& 432& 169& 196& 196\\ 1& -14& 0& 13& -14& 14 \\ 1& -14& 0& 13& 28& -28 \\ 1& 14& 0& 13& -14& -14 \\ 1& 0& 12& -13& 0& 0 \\ 1& 7& -36& 0& 14& 14 \end{bmatrix}. \] For both $Q^\mathcal{A}$ and $Q^\mathcal{B}$, we index the rows by the corresponding relation (e.g., \cc{7A} indexes the 5th row of $Q^\mathcal{A}$) and the columns by the corresponding projection matrix. Now, we have a decomposition of $\mathbb{C}^{\Omega}$ into simultaneous eigenspaces of $\mathcal{A}$ with projection matrices $E_0, E_1, \ldots, E_8$, given by \[ E_i = \sum_{j \in \{\cc{1}, \cc{6}, \cc{2}, \cc{3}, \cc{7A}, \cc{7B}, \cc{7C}, \cc{13A}, \cc{13B}\} } Q^{\mathcal{A}}_{ji} A_j. \] Let $I \subseteq \{\cc{6}, \cc{2}, \cc{3}, \cc{7A}, \cc{7B}, \cc{7C}, \cc{13A}, \cc{13B}\}$, and let $C$ and $S$ be a maximum clique and coclique of $\Gamma_I$. Now by Theorem \ref{thm:CliquesInSchemes}, $|C| \cdot |S| = 1092$ if and only if their characteristic vectors $\chi_C$ and $\chi_S$ are design-orthogonal. Let $X$ and $Y$ be the dual degree sets of $\chi_C$ and $\chi_S$ respectively, then \[ \chi_C = \sum_{i \in X} \chi_C E_i = \chi_C \sum_{i \in X} E_i \quad \text{ and } \quad \chi_S = \sum_{i \in Y} \chi_S E_i = \chi_S\sum_{i \in Y} E_i. \] Note that there are no rational vectors in the image of $E_i$ for $i \in \{1,2,3,4,5\}$, since the corresponding columns of $Q^\mathcal{A}$ contain irrational entries and the associated eigenspace is an irreducible $G$-submodule over $\mathbb{C}$. It is clear from studying $Q^\mathcal{A}$ that \begin{center} \begin{tabular}{ lll } $F_0 := E_0, \quad\quad$ & $F_1 := E_1 + E_2, \quad\quad$ & $F_2 := E_3 + E_4 + E_5,$\\ $F_3 := E_6,$ & $F_4 := E_7,$ & $F_5 := E_8,$ \end{tabular} \end{center} are the unique smallest linear combinations of the projection matrices with rational entries. Since $\chi_C$ and $\chi_S$ have rational entries, it is true that \[ \chi_C = \sum_{i \in \overline{X}} \chi_C F_i \quad \text{ and } \quad \chi_S = \sum_{i \in \overline{Y}} \chi_S F_i, \] where $\overline{X}$ and $\overline{Y}$ are the dual degree sets of $\chi_C$ and $\chi_S$ with respect to the new $F_i$ projection matrices. Clearly $\chi_C$ and $\chi_S$ are therefore design-orthogonal with respect to the real $E_i$ projection matrices if and only if they are design-orthogonal with respect to the rational $F_i$ projection matrices. Thus we have reduced the number of non-trivial projection matrices that we need to study from $8$ to $5$. We observe that the $F_i$ projection matrices are precisely the projection matrices onto the simultaneous eigenspaces of $\mathcal{B}$, given by \[ F_i = \sum_{j \in \{1, 6, 2, 3, 7, 13\} } Q^{\mathcal{B}}_{ji} A_j. \] As a result, design-orthogonality of subsets of $\Omega$ with respect to $\mathcal{A}$ is equivalent to design-orthogonality with respect to $\mathcal{B}$. \begin{proof}[Proof of Proposition \ref{prop:FewerGraphs}] Let $C$ be a clique and $S$ be a coclique of $\Gamma_I$ for $I \subseteq \{\cc{6}, \cc{2}, \cc{3}, \cc{7A}, \cc{7B}, \cc{7C}, \cc{13A}, \cc{13B}\}$ such that $|C|\cdot |S|=1092$. Then by Theorem \ref{thm:CliquesInSchemes}, $S$ and $C$ are design-orthogonal with respect to the association scheme $\mathcal{A}$. However, we have shown that $\mathcal{A}$ and $\mathcal{B}$ have the same projection matrices over the rationals, and so $C$ and $S$ must also be design-orthogonal with respect to $\mathcal{B}$. Applying Theorem \ref{thm:CliquesInSchemes} again, we discover that $C$ and $S$ must be a maximum clique and coclique, respectively, for some $\Gamma_{I'}$ for $I' \subseteq \{\cc{6}, \cc{2}, \cc{3}, \cc{7}, \cc{13}\}$. \end{proof} Moreover, we have demonstrated that we can apply the MacWilliams transform with respect to $\mathcal{B}$ (rather than $\mathcal{A}$) to determine design-orthogonality, and hence non-separation of a $G$-invariant graph $\Gamma_I$. We do so now to eliminate many of the remaining $15$ (complementary pairs of) graphs of Proposition \ref{prop:FewerGraphs}. \begin{lemma}\label{sixgraphs} Let $\Gamma$ be a union of graphs of the association scheme $\mathcal{B}$ such that \[ \alpha(\Gamma) \cdot \omega(\Gamma) = 1092. \] Then the inner distribution vectors, $a$ and $b$, of a maximum clique and maximum coclique, respectively, are one of the following pairs: \begin{enumerate}[1.] \item $(1,0,0,0,0,12)$ and $(1,14,7,14,48,0)$. \item $(1,0,0,0,13,0)$ and $(1,26,13,26,0,12)$. \item $(1,0,0,26,0,12)$ and $(1,7-t,t,0,20,0)$, $0\leqslant t\leqslant 7$. \item $(1,0,0,14, 27,0)$ and $(1,13-t,t,0,0,12)$, $0\leqslant t\leqslant 13$. \item $(1,0,13,0,0,12)$ and $(1,14-t,0,t,27,0)$, $0\leqslant t\leqslant 14$. \item $(1,0,7,0,20,0)$ and $(1,26-t,0,t,0,12)$, $0\leqslant t\leqslant 26$. \end{enumerate} \end{lemma} \begin{proof} Let $Q = Q^\mathcal{B}$ henceforth. Since we are only interested in the graphs of $\mathcal{B}$ up to complementation, we may suppose without loss of generality that $a$ has at least three zero entries. Let $a:=(1,a_1,a_2,a_3,a_4,a_5)$ and let $b:=(1,b_1,b_2,b_3,b_4,b_5)$. Note that the sum of the entries of $a$ and $b$, respectively, yields the size of the associated clique or coclique. Since $a$ and $b$ represent a clique and coclique for $\Gamma$, we have \begin{equation} \label{eq:CliqueCoclique} a\circ b =(1,0,\ldots,0). \end{equation} By Theorem \ref{thm:CliquesInSchemes}, \begin{equation} (aQ)\circ (bQ)=(1092,0,\ldots,0). \label{eq:designortho} \end{equation} Moreover, \noindent\begin{minipage}{.5\linewidth} \footnotesize \begin{align} aQ =& \nonumber \\ \big(&\quad a_5+a_4+a_3+a_2+a_1+1, \label{eq:aQ0}\\ &\quad 7(a_5+2a_3-2a_2-2a_1+14), \label{eq:aQ1}\\ &\quad 12(-3a_5+a_4+36), \label{eq:aQ2}\\ &\quad 13(-a_4+a_3+a_2+a_1+13), \label{eq:aQ3}\\ &\quad 14(a_5-a_3+2a_2-a_1+14), \label{eq:aQ4}\\ &\quad 14(a_5-a_3-2a_2+a_1+14) \label{eq:aQ5} \quad \big), \end{align} \end{minipage}% \begin{minipage}{.5\linewidth} \footnotesize \begin{align} bQ =& \nonumber\\ \big(& \quad b_5+b_4+b_3+b_2+b_1+1, \label{eq:bQ0}\\ &\quad 7(b_5+2b_3-2b_2-2b_1+14), \label{eq:bQ1}\\ &\quad 12(-3b_5+b_4+36), \label{eq:bQ2}\\ &\quad 13(-b_4+b_3+b_2+b_1+13), \label{eq:bQ3}\\ &\quad 14(b_5-b_3+2b_2-b_1+14), \label{eq:bQ4}\\ &\quad 14(b_5-b_3-2b_2+b_1+14) \label{eq:bQ5} \quad \big). \end{align} \end{minipage} \vspace{0.5cm} Each entry of $aQ$ and $bQ$ is non-negative (because the $F_j$ are positive semidefinite), and the entries of $a$ and $b$ are non-negative rational numbers. This information yields a series of equations and inequalities. We will undertake a step-by-step analysis. \begin{description} \item[Case $a_1 = a_2 = a_3 = a_4 = 0$ and $a_5>0$] Then (\ref{eq:aQ1}), (\ref{eq:aQ3}), (\ref{eq:aQ4}), and (\ref{eq:aQ5}) are all greater than zero. For $a$ to be non-trivial (\ref{eq:aQ2}) must then be zero, hence $a = (1, 0, 0, 0, 0, 12)$. By (\ref{eq:designortho}) we have (\ref{eq:bQ1}), (\ref{eq:bQ3}), (\ref{eq:bQ4}), and (\ref{eq:bQ5}) must all be zero, and (\ref{eq:bQ0}) equals $84$, and by (\ref{eq:CliqueCoclique}) we have $b_5=0$. So $b = (1, 14, 7, 14, 48, 0)$. \item[Case $a_1 = a_2 = a_3 = a_5 = 0$ and $a_4>0$] Then (\ref{eq:aQ1}), (\ref{eq:aQ2}), (\ref{eq:aQ4}), and (\ref{eq:aQ5}) are all greater than zero. For $a$ to be non-trivial (\ref{eq:aQ3}) must then be zero, hence $a = (1, 0, 0, 0, 13, 0)$. By (\ref{eq:designortho}), (\ref{eq:bQ0}) must equal $78$, and (\ref{eq:bQ1}), (\ref{eq:bQ2}), (\ref{eq:bQ4}), and (\ref{eq:bQ5}) must all be zero. Thus $b=(1,26,13,26,0,12)$. \item[Case $a_1 = a_2 = a_4 = a_5 = 0$ and $a_3 > 0$] Then (\ref{eq:aQ1}), (\ref{eq:aQ2}), (\ref{eq:aQ3}) are all greater than zero. Under the current assumption, for $a$ to be non-trivial, (\ref{eq:aQ4}) and (\ref{eq:aQ5}) must thus both equal to zero. Hence $a = (1, 0, 0, 14, 0, 0)$. However this means the size of the $a$ is $15$, which does not divide 1092, a contradiction. This case does not occur. \item[Case $a_1 = a_3 = a_4 = a_5 = 0$ and $a_2> 0$] Then (\ref{eq:aQ2}), (\ref{eq:aQ3}), (\ref{eq:aQ4}) are greater than zero, hence (\ref{eq:aQ1}) and (\ref{eq:aQ5}) must simultaneously be zero. So $a = (1, 0, 7, 0, 0, 0)$. However the size of the $a$, $8$, does not divide $1092$. This case does not occur. \item[Case $a_2 = a_3 = a_4 = a_5 = 0$ and $a_1> 0$] Then either $a_1 = 7$ or $a_1=14$ such that (\ref{eq:aQ1}) or (\ref{eq:aQ4}) is zero. In either case the size ($8$ or $15$) does not divide $1092$. This case does not occur. \item[Case $a_1 = a_2 = a_3 =0$ and $a_4,a_5> 0$] Now, \eqref{eq:aQ1}, \eqref{eq:aQ4}, and \eqref{eq:aQ5} are all non-zero, hence \eqref{eq:bQ2}, \eqref{eq:aQ4}, \eqref{eq:aQ5} are all zero according to \eqref{eq:designortho}. We have $b_4=b_5=0$ by (\ref{eq:CliqueCoclique}), meaning $b = (1, 14, 7, 14, 0, 0)$. However the size of $b$, $36$, does not divide $1092$. This case does not occur. \item[Case $a_1=a_2=a_4=0$ and $a_3,a_5> 0$] Then $b_3=b_5=0$ and we have $a_3=26$, $a_5=12$, $b_4=20$ as well. Moreover, $a=(1,0,0,26,0,12)$ and $b=(1,7-b_2,b_2,0,20,0)$, $0\le b_2\le 7$. \item[Case $a_1=a_2=a_5=0$ and $a_3,a_4> 0$] Then $b_3=b_4=0$ and we have $b_5=12$, $a_3=14$ as well. Moreover, $a=(1,0,0,14, 27,0)$ and $b=(1,13-b_2,b_2,0,0,12)$, $0\le b_2\le 13$. \item[Case $a_1=a_3=a_4=0$ and $a_2,a_5> 0$] If $a_2,a_5\ne 0$, then $b_2=b_5=0$ and we have $a_2=13$, $a_5=12$, $b_4=27$. Moreover, $a=(1,0,13,0,0,12)$ and $b=(1,14-b_3,0,b_3,27,0)$, $0\le b_3\le 14$. \item[Case $a_1=a_3=a_5=0$ and $a_2,a_4> 0$] If $a_2,a_4\ne 0$, then $b_2=b_4=0$ and we have $a_2=7$, $a_4=20$, and $b_5=12$. Moreover, $a=(1,0,7,0,20,0)$ and $b=(1,26-b_3,0,b_3,0,12)$, $0\le b_3\le 26$. \item[Case $a_1=a_4=a_5=0$ and $a_2,a_3> 0$] So $b_2=b_3=0$. Straight away, we have \eqref{eq:aQ2}, \eqref{eq:aQ3}, \eqref{eq:bQ5}$>0$ and so $-3b_5+b_4+36=0$, $-b_4+b_1+13=0$, and $-a_3-2a_2+14=0$. Hence \eqref{eq:aQ4} $>0$ and $b_5-b_1+14=0$. However, then \eqref{eq:bQ1} $<0$, which is a contradiction. This case does not occur. \item[Case $a_2=a_3=a_4=0$ and $a_1,a_5> 0$] Here we have $b_1=b_5=0$, $a_5=12$ and two cases: \begin{itemize} \item $a=(1,13,0,0,0,12)$, $b=(1,0,0,14,27,0)$. \item $a=(1,26,0,0,0,12)$, $b=(1,0,7,0,20,0)$. \end{itemize} These cases already appear above. \item[Case $a_2=a_3=a_5=0$ and $a_1,a_4> 0$] Here, we have $b_1=b_4=0$, $b_5=12$ and two cases: \begin{itemize} \item $a=(1,7,0,0,20,0)$ and $b=(1,0,0,26,0,12)$. \item $a=(1,14,0,0,27,0)$ and $b=(1,0,13,0,0,0)$. \end{itemize} These cases already appear above. \item[Case $a_2=a_4=a_5=0$ and $a_1,a_3> 0$] If $a_1,a_3\ne 0$, then \[ b=(1,0,91/5,0,156/5,112/5), \] which yields a size that is not an integer, so this case does not occur. \item[Case $a_3=a_4=a_5=0$ and $a_1,a_2> 0$] If $a_1,a_2\ne 0$, then \[ b=(1,0,0,91/2,117/2,63/2), \] which yields a size that is not an integer. This case does not occur.\qedhere \end{description} \end{proof} Each of the inner distribution structures given by Lemma \ref{sixgraphs} defines a complementary pair of graphs of $\mathcal{B}$. For example, the fourth case where the inner distributions are $(1,0,0,14, 27,0)$ and $(1,13-t,t,0,0,12)$ imply that the graph we are taking a maximum clique and coclique from is $\Gamma_{\cc{3,7}}$ or $\Gamma_{\cc{6,2,13}}$ (upon inspection of the positions of zeroes of these vectors). We now have just six graphs to analyse, and by Lemma \ref{sixgraphs}, we know the only possibilities for $\alpha(\Gamma)$ and $\omega(\Gamma)$ when $\alpha(\Gamma) \cdot \omega(\Gamma) = 1092$. We record this information in Table \ref{tab:cliquebounds}. Each of the six graphs has either a clique or a coclique of the size determined by Lemma~\ref{sixgraphs}, these are indicated by an asterisk in the table. \begin{table}[!ht] \begin{center} \begin{tabular}{clcc} \toprule Case&Graph $\Gamma$ & $\omega(\Gamma)$ & $\alpha(\Gamma)$ \\ \midrule 1&$\Gamma_{\cc{13}}$ & 13* & 84\\ 2&$\Gamma_{\cc{7}}$ & 14 & 78* \\ 3&$\Gamma_{\cc{3,13}}$ & 39* & 28\\ 4&$\Gamma_{\cc{3,7}}$ & 42 & 26* \\ 5&$\Gamma_{\cc{6,13}}$ & 26* & 42\\ 6&$\Gamma_{\cc{6,7}}$ & 28 & 39*\\ \bottomrule \end{tabular} \end{center} \caption{Putative non-separating graphs to be considered for $q=13$.}\label{tab:cliquebounds} \end{table} We write $\alpha_I$ and $\omega_{I}$ for the coclique and clique numbers of $\Gamma_I$. We shall say that $\Gamma_I$ is \emph{separating} if $\alpha_I \cdot \omega_I \neq 1092$ (so $G$ is separating if and only if all the graphs $\Gamma_I$ are separating). Recall that if a graph $\Delta$ is a spanning subgraph of a graph $\Gamma$, then $\alpha(\Delta)\ge \alpha(\Gamma)$ and $\omega(\Delta)\le \omega(\Gamma)$. \begin{lemma}\label{twographsdone}\leavevmode \begin{enumerate}[(i)] \item If $\alpha_{\cc{6},\cc{13}}<42$, then $\omega_{\cc{3},\cc{7}}<42$ and the graph $\Gamma_{\cc{3},\cc{7}}$ is separating. \item If $\alpha_{\cc{3},\cc{13}}<28$, then $\omega_{\cc{6},\cc{7}}<28$ and the graph $\Gamma_{\cc{6},\cc{7}}$ is separating. \end{enumerate} \end{lemma} \begin{proof} Since $\Gamma_{\cc{3},\cc{7}}$ is an spanning subgraph of $\Gamma_{\cc{2},\cc{3},\cc{7}}$, we have $\omega_{\cc{3},\cc{7}}\le \omega_{\cc{2},\cc{3},\cc{7}}=\alpha_{\cc{6},\cc{13}}$. Likewise, $\omega_{\cc{6},\cc{7}}\le \omega_{\cc{2},\cc{6},\cc{7}}=\alpha_{\cc{3},\cc{13}}$. \end{proof} \begin{lemma}\label{gamma7} $\Gamma_\cc{7}$ is separating. \end{lemma} \begin{proof} Consider the natural action of $T$ on a set $\Sigma$ of size 14. Then a point stabiliser $T_\omega$ forms a coclique of $\Gamma_\cc{7}$. To see why, notice that if $x,y\in T_\omega$, then $xy^{-1}\in T_\omega$, and in particular, $xy^{-1}$ does not have order 7 (because $|T_\omega|=6\times 13$). Moreover, by Table \ref{tab:cliquebounds}, $T_\omega$ yields a coclique of maximum size. Suppose that $C$ is a clique of $\Gamma_{\cc{7}}$. If $g,h\in C$ such that there exists $\omega\in\Sigma$ for which $\omega^g=\omega^h$, then $gh^{-1}\in T_\omega$. However, this contradicts the elements of $T_\omega$ having order coprime to 7. Thus $C$ is a sharply-transitive set of permutations in $T$ (on $\Sigma$). Such a set does not exist (cf., \cite{mathoverflow}). \end{proof} So we have reduced the problem to three graphs: $\Gamma_{\cc{13}}$, $\Gamma_{\cc{6},\cc{13}}$, $\Gamma_{\cc{3},\cc{13}}$. \begin{lemma} \label{TwoOfTheTrickyOnes} $\alpha(\Gamma_{\cc{6},\cc{13}})=25$ and $\alpha(\Gamma_{\cc{3},\cc{13}})=22$. Therefore, $\Gamma_{\cc{6},\cc{13}}$ and $\Gamma_{\cc{3},\cc{13}}$ are separating. \end{lemma} \begin{proof} We used \textsc{Grape} \cite{grape}, and its basic clique-finding algorithm on a personal laptop. It took less than 10 minutes each in each case to determine that $\alpha(\Gamma_{\cc{6},\cc{13}})=25$ and $\alpha(\Gamma_{\cc{3},\cc{13}})=22$. \end{proof} \begin{lemma} \label{gamma13} $\Gamma_{\cc{13}}$ is separating. \end{lemma} \begin{proof} From Table \ref{tab:cliquebounds}, a maximum clique in $\Gamma_{13}$ has size at most 13. Let $C$ be a cyclic subgroup of $T$ of order 13. Then the elements of $C$ form a maximum clique of $\Gamma_{\cc{13}}$. To see why, let $g,h\in C$. Then $gh^{-1}\in C$ and hence $gh^{-1}$ has order 13. Note that $\Aut(\Gamma_{\cc{13}})$ acts transitively on the vertices of $\Gamma_{\cc{13}}$ as it contains $G$. Let $M$ be the matrix whose rows are the characteristic vectors of each $C^g$, where $g\in \Aut(\Gamma_{\cc{13}})$. Then the characteristic vector of a coclique is a $\{0,1\}$--vector $v$ such that $Mv\preccurlyeq (1,1,\ldots, 1)$. This gives us an integer linear program, where our objective function is the sum of the values of $v$. Using the mixed integer linear programming (MILP) software \textsf{Gurobi} \cite{gurobi}, we find that $\alpha(\Gamma_{\cc{13}})$ is 78. However, $13\times 78<1092$ and hence $\Gamma_{\cc{13}}$ is separating. \end{proof} \begin{proof}[Proof of Theorem \ref{main1}] For $q=13$, the proof follows from Proposition \ref{prop:FewerGraphs}, and Lemmas \ref{sixgraphs}, \ref{twographsdone}, \ref{gamma7}, \ref{TwoOfTheTrickyOnes}, and \ref{gamma13}. For $q=17$ the analysis proceeds in the same manner. An analogous result to Proposition \ref{prop:FewerGraphs} reduces the number of $G$-invariant graphs which need to be considered from $255$ to $31$. With the aid of the computer algebra package, \textsf{Mathematica} \cite{mathematica}, a result analogous to Lemma \ref{sixgraphs} was obtained, leaving only $23$ graphs for consideration. These graphs were ruled out computationally by using \textsc{Grape} \cite{grape} in some instances, and in other instances, we can formulate either an integer linear program or a constraint satisfaction program from the fact (see Corollary \ref{intersect1}) that a maximum clique and maximum coclique would have to meet in precisely one element. We used \textsf{Gurobi} \cite{gurobi} for these constraint satisfaction problems, and verified some cases with \textsf{Minion} \cite{minion}. See Table \ref{tbl:graphs17} in Appendix \ref{append:Cases17} for details on these graphs, including which bounds were obtained, and which methods were used. Graphs $\Gamma_I$ are labelled according to the conjugacy classes of elements of order $2, 3, 4, 8, 9, 17$ in the same manner as for $q=13$. For the reader interested in verifying these computational results, supplementary materials have been made available at \cite{SupplementaryMaterials}. \end{proof} \section{Non-spreading} We already know from our knowledge of exact factorisations of $\PSL(2,q)$ that $G:=\PSL(2,q)\times \PSL(2,q)$ in its diagonal action is non-separating, and hence non-spreading, when $q$ is even or when $q\equiv 3\pmod{4}$. So let $q$ be an odd prime power such that $q\equiv 1\pmod{4}$, and let $T=\PSL(2,q)$. Let $T_1$ be a point stabiliser in the natural action of $T$ on $q+1$ points. Next, let \[ T_1^2:=\{t^2 : t\in T\}. \] This setup will provide us with the ingredients we need to exhibit a multiset $A$ and a set $B$ with the desired properties of a witness to $G$ being non-spreading. \begin{lemma}\label{nicepropPSL}\samepage Let $T=\PSL(2,q)$, with $q\equiv 1\pmod{4}$, and let $T_1$ and $T_2$ be two different point stabilisers in the natural $2$-transitive action of $T$ (on $q+1$ points). Then \[ |T_1^2\cap T_2t|=\tfrac{1}{2}|T_1\cap T_2t| \] for all $t\in T$. \end{lemma} \begin{proof} Since $T$ is $2$-transitive, we may suppose that $T_1$ and $T_2$ are the stabilisers of the points $1$ and $2$ in the action of $T$ on $\{1,\ldots, q+1\}$. First, we note that $T_1$ has structure $[q] \sd C_{(q-1)/2}$ and $T_1^2$ is actually a subgroup with structure $[q]\sd C_{(q-1)/4}$. Let $t\in T$. Now $T_1\cap T_2t$ is the set of elements of $T$ that fix $1$ and map $2$ to $2^t$. The stabiliser $T_1\cap T_2$ is the top group $C_{(q-1)/2}$ acting semiregularly on the remaining points, with two orbits $\mathcal{O}$ and $\mathcal{O}'$, each of size $(q-1)/2$. Therefore, $|T_1\cap T_2t|$ is equal to $0$ or $(q-1)/2$, depending on whether $2^t=1$ or not. Now consider $T_1^2\cap T_2t$. Here, the top group of $T_1^2\cap T_2$ is $C_{(q-1)/4}$ and it splits each orbit $\mathcal{O}$ and $\mathcal{O}'$ in half. Therefore, $|T_1^2\cap T_2t|$ is equal to $0$ or $(q-1)/4$, depending on whether $2^t=1$ or not. \end{proof} \begin{proof}[Proof of Theorem \ref{main2}] Let $A$ be the multiset defined in the following way: place a multiplicity of 2 on each element of $T_1^2$; place a multiplicity of 1 on each element of $T\backslash T_1$. We will show that \begin{enumerate}[(i)] \item $|A|=|T|$; \item $|\chi_A\circ \chi_{T_1^g}|=|T_1|$ for all $g\in G$. \end{enumerate} First, $|A|=2|T_1^2|+|T|-|T_1|=|T|$, so (i) is satisfied. Next, suppose $g\in G$. Then \begin{align*} |\chi_A\circ \chi_{T_1^g} |&=|T_1^2\cap T_1^g|-|(T_1\backslash T_1^2)\cap T_1^g|+|T_1^g|\\ &=|T_1^2\cap T_1^g|-|T_1\cap T_1^g|+|T_1^2\cap T_1^g|+|T_1|\\ &=2|T_1^2\cap T_1^g|-|T_1\cap T_1^g|+|T_1|. \end{align*} Now by Lemma \ref{nicepropPSL}, $|T_1^2\cap T_1^g|=\tfrac{1}{2}|T_1\cap T_1^g|$, because $T_1^g$ is a right coset (in $T$) of a point-stabiliser in $T$ (in the natural $2$-transitive action of $T$). Therefore, $|\chi_A\circ \chi_{T_1^g} |=|T_1|$ and so (ii) is satisfied. Therefore, the pair $(A,T_1)$ witness that $G$ is non-spreading. \end{proof} \subsection*{Acknowledgements} This work forms part of an Australian Research Council Discovery Project DP200101951. \bibliographystyle{abbrv}

e74385dd110b83d35997aefb5d0e20b3027c60ef

\section*{Appendix} \input{sections/A1_Data_Collection} \input{sections/A2_Implementation_Detail} \input{sections/A3_FID} \input{sections/A4_More_Results} \input{sections/A5_Evaluation_on_FAR_PR} \input{sections/A7_User_Study_Detail} \input{sections/A8_Case_Study_Detail} \input{sections/A9_Failed_Attempt} \end{document} \section{Introduction} \label{sec:introduction} \begin{figure}[t] \centering \includegraphics[width= 0.5\textwidth]{figure/teaser-12-11ss.jpg} \caption{Our model takes in a program graph (also called bubble diagram) and a design space in voxel graph representation, and outputs a variety of volumetric designs. Professional architects can convert the output into detailed building design efficiently.} \label{fig:teaser} \end{figure} Volumetric design (also called massing design or schematic design) is the first step when an architect designs a building on a given land site. Based on the local building codes applied to the site, the building can only be designed within a valid design space, which is usually not a regular cuboid. For instance, the daylight restrictions prevent the building from casting too much shadow over its neighboring building by drawing a slant line as upper bound. Within the valid design space, a volumetric design not only depicts the volumetric 3D shape of the building, but also produces 2D program layouts for each story. An example is illustrated in Figure \ref{fig:vol_desgin}. The architect then uses the finalized volumetric design to gradually develop all the details for construction, including façade design, interior design, structure systems, etc. While volumetric design is the foundation of the design and construction process, making a good volumetric design usually requires a significant amount of time and effort. An efficient pipeline to generate volumetric design will bring a great impact on the architecture and construction industry. \begin{figure}[t] \centering \includegraphics[width=0.5\textwidth]{figure/vol_design-06-05ss.jpg} \caption{Left: an example of valid design space. Right: an example of volumetric design within the valid design space} \label{fig:vol_desgin} \end{figure} Generating realistic 2D room layouts has been a popular topic for many years. Existing methods include optimization-based \cite{merrell2010computer, bymw_goodLayout_sigg13} and learning-based \cite{wu2019data, nauata2020house, hu2020graph2plan, di2020end} approaches. Recently, researchers start looking at how to integrate program graphs into layout generation tasks using graph neural networks (GNNs) \cite{nauata2020house, hu2020graph2plan, di2020end}. Program graph, also called bubble diagram, is a graph that illustrates the relations between programs or rooms and is a common representation used by professional architects to explore design ideas. Similar to House-GAN \cite{nauata2020house}, this paper also focuses on the graph-conditioned layout generation task. The task requires the output layouts to be compatible to the condition input \textit{program graphs}. However, there is no literature on extending the task to 3D. Our goal is to produce multiple layouts, which stack up and form a volumetric design for a multi-story building. Though it might seem straight-forward to transfer previous 2D approaches to 3D, there are several challenges and limitations when applying previous approaches: \begin{itemize} \item Compared to the 2D counterparts, 3D program graphs are not only larger in size, but also more complex with additional inter-story relations. The output design space also increases by the number of stories. \item The raw rasterized output used in previous works cannot produce clean corners and edges due to the fine discretization of pixels. For instance, boundaries are usually jagged, rooms can be poorly aligned and overlapping each other, there might be small dents or bulges in some rooms, etc. \item Volumetric images (usually defined as 3D regular grids with uniformly discretized voxels) have the closest structural similarity to rectangular buildings than other 3D representations, such as point clouds or meshes. However, it is not computational and memory efficient to use this dense representation for polygonal rooms. Moreover, there are voxels within the regular grid but not in the irregular valid design space that take unneeded memory and computation. \end{itemize} To overcome these challenges and limitations, we propose \textit{voxel graph}, a novel 3D representation that can encode irregular voxel grids with non-uniform space partitioning. To bridge between the input \textit{program graph} and the output \textit{voxel graph}, we design a pointer-based cross-modal modules in our generative adversarial graph network. The pointer module can be used not only for message passing, but also as a decoder to output probability over a dynamic set of valid programs. We also work with professional architects to create a synthetic dataset that contains 120,000 volumetric designs based on realistic building requirements. We evaluate our model qualitatively and quantitatively, and it outperforms existing method by a large margin in all the three metrics: quality, diversity, and connectivity accuracy. In summary, our main contributions are: 1) a new 3D representation, \textit{voxel graph}; 2) a graph-conditioned generative adversarial network (GAN) using GNN and pointer-based cross-modal module; 3) an automated pipeline to generate valid volumetric designs through simple interaction; and 4) a synthetic dataset that contains 120,000 volumetric design and their corresponding program graphs. We will share the code, model, and dataset. \section{Related Work} \label{sec:related work} \subsection{Voxel Representations} Regular grid representation using voxels, such as occupancy grids, has been studied since the 3D extension of 2D convolution. To achieve 3D shape synthesis, researchers build encoder-decoder models, such as deep belief network \cite{wu20153d}, variational auto-encoder (VAE) \cite{kingma2013auto}, generative adversarial network (GAN) \cite{wu2016learning, smith2017improved}, and energy-based model \cite{xie2018learning}. However, due to the dense representation for sparse occupancy, voxel representation is notorious for its cubic computational cost and poor scalability to higher resolutions and larger sizes. Existing methods to mitigate the problem include sparse convolution \cite{graham2017submanifold, graham20183d, choy20194d} and octree representation \cite{riegler2017octnet, wang2017cnn, wang2018adaptive}. Our proposed voxel graph combines voxel-based and graph-based representations by encoding voxels into graph nodes. Similar idea was proposed in Point-Voxel CNN \cite{liu2019point}. To enhance the local modeling capability, it has a high-resolution point-based branch as well as a low-resolution voxel-based branch for point cloud encoding. Another feature of our voxel graph is the ability to support non-uniform space partition. Polyfit \cite{nan2017polyfit, fang2020connect} reconstructs 3D models by selecting space partition planes extracted from point clouds. BSP-Net \cite{chen2020bsp} learns to generate compact meshes using binary space partitioning. NeuralSim and NeuralSizer \cite{pmlr-v119-chang20a} also use graphs to represent structure grids (i.e., columns and beams) of buildings instead of dense voxels. \begin{figure*}[h] \centering \includegraphics[width=.9\textwidth]{figure/representation-03-04ss.jpg} \caption{Left: the hierarchical program graph. Right: the irregular grid with non-uniform voxel size and the equivalent voxel graph. } \label{fig:graph_representation} \end{figure*} \subsection{Graph-conditioned Layout Generation} To the best of our knowledge, there is no prior works on learning-based 3D layout generation. Alternatively, we review several work on graph-conditioned 2D layout generation. Graph2Plan \cite{hu2020graph2plan} generates bounding boxes for each room, and refines box locations with a cascaded refinement network. The input graphs are retrieved based on user constraints and outline similarity. The user can get various layouts by feeding different graphs, but the model itself cannot produce variation. House-GAN \cite{nauata2020house} proposes a graph-conditioned GAN, where the generator and discriminator are built upon relational architecture - ConvMPN \cite{zhang2020conv}. Xinhan Di \etal \cite{di2020end} uses a similar adversarial approach on interior design with doors, windows, and furniture. LayoutGMN \cite{patil2020layoutgmn} learns to predict structural similarity between two layouts with an attention-based graph matching network. Wamiq Para \etal \cite{para2020generative} explores the idea of generative modeling using constraint generation for layouts. \section{Representation and Data Collection} The goal of this paper is to generate 3D volumetric designs given a program graph and a valid design space. The program graph illustrates the intra-story and inter-story relations between programs. Besides program graph and valid design space, there are other design conditions that are considered by architects in industry practice. Floor area ratio (FAR, derived by dividing the total area of the building by the total area of the parcel), should not exceed a regulation limit. In addition, target program ratio (TPR) defines the approximate ratio between programs. For example, office : corridor : restroom : elevator : stairs $= 50:20:15:5:10$. Both TPR and FAR are encoded into the program graph as described in Section \ref{sec: hpg} and are used as the model input. Another input is a valid design space, which may be irregular due to building codes. The design space can be further partitioned freely based on architect's decisions or statistical heuristics. In practice, before starting the design process, architects usually partition the space by considering construction standards, structure systems, and conventional modules. Inspired by this partitioning process, we invent the representation, \textit{voxel graph}, as described in Section \ref{sec: vg}. \subsection{Data Collection} Since there is no publicly available dataset for volumetric designs from real buildings, we create a synthetic dataset with 120,000 volumetric designs for commercial buildings using parametric models. The site of each design is bounded within $40\times40\times50 m^3$, where different site conditions are randomly generated. The heuristics behind the parametric models are based on the rules and knowledge provided by professional architects. Although these parametric models are able to explore possible volumetric designs, they are not capable of fitting the constraints. Therefore, we generate the designs first and then compute the voxel graph, program graph, FAR, and TPR for each design. Please refer to the supplementary for more details and visualization of the synthetic dataset. The dataset can also be used to explore other learning-based design tools or relevant tasks in computer vision and graphics. \subsection{Hierarchical Program Graph} \label{sec: hpg} Given a building datum, we first construct 2D program graphs for each story. Each program node feature includes the program type and the story level. Here, we consider 6 program types: lobby/corridor, restroom, stairs, elevator, office, and mechanical room. A program edge shows the two programs are connected by a door or opening. To construct the 3D program graph, we stack all 2D program graphs and chain the stairs and elevators, since they are the only paths for moving vertically. In practice, the 3D program graph also represents the circulation of the building. Recall that there are two other design condition inputs: FAR and TPR. The FAR limit is stored as a graph-level feature. As for TPR, we add one hierarchy on top of the 3D program graph. We create one master program node for each program type and connect them to all program nodes of the same type. The edges allow the master node to allocate different area sizes on each program node through message passing. Please refer to left of Figure \ref{fig:graph_representation}. \begin{figure*}[h] \centering \includegraphics[width=\textwidth]{figure/network_arch_d-08ss.jpg} \caption{An overview of Building-GAN. Top: the Program GNN, Voxel GNN, and Cross-Modal Pointer Module for the generator. Bottom: the discriminator with the building and story level decoders.} \label{fig:network_structure} \end{figure*} \subsection{Voxel Graph} \label{sec: vg} To overcome the challenges and limitations listed in Section \ref{sec:introduction}, we invent a 3D representation called \textit{voxel graph}. Each node represents a voxel and the voxel information (coordinate and dimension) is stored as node features. Different from volumetric images with voxel grids, voxel graph does not assume regular grids and consumes memory only for occupied voxels. Moreover, it allows non-uniform space partitioning, which avoids over-discretization when using the uniform voxel size. Theoretically, voxel nodes can encode arbitrary 3D primitives, but in this paper, only cuboids with varying sizes are used to build up the approximated valid design space. When parsing the data, the space partition is defined by the projection of all 2D layouts. In real-world practice, walls tend to align across different stories for structural stability or construction considerations, which leads to a reduced amount of voxels in the space partition. Next, we turn the voxels into graph nodes and store the voxel information (location and dimension) as node features and program type as node labels. Node mask is also stored in case of nodes that are left unused and does not have any program type. Lastly, a voxel edge connects two voxel nodes if they share a face. The final voxel graph should look like an irregular cubic lattice as illustrated in the right of Figure \ref{fig:graph_representation}. \section{Method} \label{sec: method} We formulate the framework as a graph-conditioned GAN. The generator is composed by two GNNs for the program graph and voxel graph, connected by a cross-modal pointer module. The discriminator is composed by a GNN with two decoders to evaluate design from both building and story level. An overview of our model is illustrated in Figure \ref{fig:network_structure}. \subsection{Generator} \label{sec: gen} \subsubsection{Program GNN} Our generator starts with a program graph neural network to encode the input program graph. Denote random program noise as $z^{p}$, FAR limit as $F$, program node feature $i$ as $x_{i}$, neighbor of node $i$ as $Ne(i)$, node cluster of $i$'s program type as $Cl(i)$, target program ratio of $i$'s program type as $r_{Cl(i)}$, multi-layer perceptron as $MLP$, mean pooling as $Mean$, and concatenation operator as $[\cdot,\cdot]$. We first map the node feature to the embedding space (\ref{eq:1}), then compute message passing $T$ times. In each message passing step, we compute the message from neighboring nodes (\ref{eq:2}) and mean pool all nodes with the same program type as the master node embedding (\ref{eq:3}). Lastly, we update the node embeddings with residual learning to avoid gradient vanishing (\ref{eq:4}). After $T=5$ steps of message passing, the final embedding of program node $i$ is denoted as $x^{T}_{i}$. \begin{equation} x_{i}^{0} = MLP_{enc}^{p}([x_{i}, z^{p}_{i}, F]) \label{eq:1} \end{equation} \begin{equation} m_{i}^{t} = \frac{1}{|Ne(i)|}\sum_{j\in Ne(i)} MLP_{message}^{p}([x_{i}^{t}, x_{j}^{t}]) \label{eq:2} \end{equation} \begin{equation} c_{i}^{t} = Mean_{j\in Cl(i)} (\{x_{j}^{t}\}) \label{eq:3} \end{equation} \begin{equation} x_{i}^{t+1} = x_{i}^{t} + MLP_{update}^{p}([x_{i}^{t}, m_{i}^{t}, r_{Cl(i)}c_{i}^{t}, F]) \label{eq:4} \end{equation} \subsubsection{Voxel GNN} The input voxel features $v_{k}$ and voxel noise $z^{v}_{k}$ are first encoded by the voxel GNN encoder. To better encode the story index, we choose positional encoding (PE) as proposed in \cite{vaswani2017attention} and add it to the processed embedding (\ref{eq:5}). Instead of appending the absolute coordinates in voxel features, we use the relative displacements $p_{k}-p_{l}$ in message computation (\ref{eq:6}). Voxel node embeddings are updated with residual learning (\ref{eq:7}). \begin{equation} v_{k}^{0} = MLP_{enc}^{v}([v_{k}, z^{v}_{k}]) + PE(story_{k}) \label{eq:5} \end{equation} \begin{equation} n_{k}^{t} = \sum_{l\in Ne(k)} MLP_{message}^{v}([v_{k}^{t}, v_{l}^{t}, p_{k}-p{l}]) \label{eq:6} \end{equation} \begin{equation} v_{k}^{t} = v_{k}^{t} + MLP_{update}^{v}(v_{k}^{t}, n_{k}^{t}) \label{eq:7} \end{equation} \subsubsection{Pointer-based Cross-Modal Module} After processing the program graph with the program GNN, the final embedding of program nodes can be viewed as the virtual "blueprint" of a design. Therefore, it is necessary to "look" at this blueprint to generate the output. To bridge between the program graph and the voxel graph, we introduce a pointer-based cross-modal module. Inspired by the application \cite{see2017get,nash2020polygen} of the Pointer Network \cite{vinyals2015pointer} in natural language processing and mesh generation tasks, we construct a pointer module to achieve message passing between the voxel nodes and all the program nodes on the same story. We cannot use a fixed length output to model program type distribution since 1) different stories can have different numbers of program nodes to choose from, for example, one floor has five rooms and another one has seven rooms; and 2) if there are two program nodes with the same program type, we want to differentiate between the two nodes, such as two restrooms in the same floor. The pointer module returns three terms: $mask_{k}$, $att_{k}$, and $v_{k}^{t+1}$ (\ref{eq:8}). $mask_{k}$ is used as a soft prediction whether the voxel node $k$ is used or not (\ref{eq:9}). If it is not used, it is left unused and has no program type. Otherwise, $att_{k}$ is the attention distribution over the set of program nodes on the same floor (\ref{eq:10}, \ref{eq:11}). An updated embedding $v_{k}^{t+1}$ is computed by the weighted sum of the program embeddings $x^{T}_{i}$ multiplied by the soft prediction $mask_{k}$ with residual learning (\ref{eq:12}). \begin{equation} mask_{k}, att_{k}, v_{k}^{t+1} = Pointer(v^{t}_{k}, \{x^{T}_{i}\}) \label{eq:8} \end{equation} \begin{equation} mask_{k} = \sigma(MLP(v^{t}_{k})) \label{eq:9} \end{equation} \begin{equation} e_{k,i} = \theta ^{T} \tanh(W_{x}x^{T}_{i} + W_{v}v^{t}_{k}) \label{eq:10} \end{equation} \begin{equation} att_{k} = \textrm{gumbel softmax}(e_{k}) \label{eq:11} \end{equation} \begin{equation} v^{t+1}_{k} = v^{t}_{k}+mask_{k} \sum_{i}att_{k, i}x^{T}_{i} \label{eq:12} \end{equation} We experiment different ways to integrate the pointer module. It can be placed after every several message passing steps in voxel GNN. Our baseline model uses 12 steps of message passing and call the pointer module once every 2 steps. Please refer to the supplementary for the complete model and algorithm. Conceptually, these pointer modules should gradually improve the design. Note that the output $att_{k}$ indicates which program node is associated to the program type of the voxel node, instead of merely the program type prediction. \subsection{Discriminator} \label{sec: dis} Our discriminator is trained to distinguish if a given design is generated by the generator or sampled from the dataset. Therefore, we take a similar architecture as voxel GNN, but without using the pointer modules. The program type predictions are concatenated to the encoded voxel node features. After $T=12$ message passing steps, two separate decoders are used. A graph-level max-pooling decoder evaluates the design as a whole while a story-level max-pooling decoder evaluates the per-story layouts individually. \begin{equation} o^{global} = MLP^{dec}_{global}(\sum_{k} v^{T}_{k}) \label{eq:13} \end{equation} \begin{equation} o^{story} = Mean_{story \; s}(MLP^{dec}_{story}(\sum_{k \in s} v^{T}_{k})) \label{eq:14} \end{equation} \subsection{Loss} We use the WGAN-GP \cite{gulrajani2017improved} loss with gradient penalty set to 10. The two decoder outputs from the discriminator are equally weighted. The gradient penalty is computed by linearly interpolating the cross-modal attention between real data and generated output, while fixing the voxel graph connectivity. \subsection{Evaluation Metric} We evaluate the generated design in terms of quality, diversity, and connectivity accuracy. The quality and diversity of the output design is evaluated with the Fr\'echet Inception Distance (FID) score \cite{heusel2017gans}. FID score has demonstrated high correlation to human judgement and has been widely used in many 2D and 3D studies. Our reference model is based on a larger version of 3D Descriptor Net \cite{xie2018learning}. We replace all convolution layers with 6 residual blocks due to the higher complexity of our data. Then we flatten the embedding to a 128-dimension tensor using convolution operation and pass it to a dense layer for loss computation. The FID score is measured over 10,000 samples. We also run a user study with architects to measure the quality in Section \ref{sec:user_study}. The connectivity accuracy (Con.) is measured by the number of the program (room) connections observed from both the generated design and in the program graph, divided by the amount of all edges in the program graph. Note that only when two rooms are connected in the program graph but disconnected in the voxel graph, it is considered as inaccurate, since there is no shared wall to put a door. It is accurate when two rooms are connected in voxel graph but disconnected in program graph, because designers can decide not to put a door on the shared walls. For more details about model implementation, hyper parameters, training environment, and user study, please refer to the supplementary. \section{Experiments} \begin{figure*}[h] \centering \includegraphics[width=0.7\textwidth]{figure/comparison-06ss.jpg} \caption{For each program graph, volumetric designs are generated by our model and by House-GAN \cite{nauata2020house}.} \label{fig:viz_result} \end{figure*} \subsection{Baseline Comparison and Visualization} We compare our model to a slightly modified version of House-GAN \cite{nauata2020house}. A major difference between our model and House-GAN is that House-GAN directly generates layout masks of size $40 \times 40$ on the nodes of the program graph. Since House-GAN does not use voxel graph representation, it assumes that the valid design space is a regular grid. House-GAN discriminator places the generated masks back to the program nodes as features and determines if the program graph is valid. To extend the House-GAN to 3D, we append the story index of each program node to its feature. Figure \ref{fig:viz_result} shows example designs from ground truth, our model, and House-GAN. Our model shows capability of generating designs that have realistic 3D shapes and clean facade surfaces. We also observe that the number of jagged boundaries are reduced due to the usage of voxel graph representation. In addition, the layouts of individual stories are reasonable. For instance, the functional programs such as elevators, stairs, and restrooms are arranged as clusters and connected by corridors. Last but not least, the functional programs are nicely aligned in the vertical direction. In contrast, though House-GAN seems to generate reasonable layouts story-by-story, they don't align well when stacked vertically. Quantitative results are presented in Table \ref{table:baseline_comparison}. Our model outperforms House-GAN in both FID and connectivity accuracy. \begin{table}[h] \begin{center} \begin{tabular}{|l|c|c|c|} \hline Method & Parameter & FID & Con.\\ \hline\hline House-GAN & - & 17.6003 & 0.403\\ \hline Ours & - & \cellcolor{lightgray} 0.0845 & \cellcolor{lightgray}0.569\\ \hline \hline \multirow{5}{*}{ \begin{tabular}{@{}l@{}} Voxel Layer \\ (Pointer Frequency \\ = every 2 steps)\end{tabular} } & 4 & 1.0463 & 0.432\\ & 6 & 0.4139 & 0.501\\ & 8 & 0.2365 & 0.497\\ & 10 & 0.0997 & 0.534\\ & 12 & \cellcolor{lightgray} 0.0845 & \cellcolor{lightgray} 0.569\\ \hline \multirow{4}{*}{ \begin{tabular}{@{}l@{}}Pointer Frequency \\ (Voxel Layer = 12)\end{tabular} } & first + last & 3.3818 & 0.578\\ & every 6 steps& 0.2179 & 0.547\\ & every 3 steps & 0.1473 & 0.541\\ & every 2 steps & \cellcolor{lightgray} 0.0845 & \cellcolor{lightgray} 0.569\\ \hline \end{tabular} \end{center} \caption{Quantitative evaluation using FID score and connectivity accuracy. We compare our baseline model to House-GAN. We also experiment baseline models with different numbers of voxel layer and pointer frequencies.} \label{table:baseline_comparison} \end{table} \begin{figure*}[h] \centering \includegraphics[width=\textwidth]{figure/variation_viz-08ss.jpg} \caption{Design variations generated by fixing the program graph while changing the voxel graph and noise.} \label{fig:variation} \end{figure*} In Table \ref{table:baseline_comparison}, we also compare models with different hyper-parameter set-ups. First, we fix the frequency of applying the pointer module to 2 and experiment different numbers of message passing layers in voxel GNN. The result shows that using 12 voxel layers yields best performance in both FID and connectivity accuracy. This is not surprising: using more voxel layers allows computing longer-range relations between voxel nodes, which is especially critical for achieving vertical alignment in taller buildings. Next, we fix the number of voxel layer and evaluate the impact of different pointer frequency. The model (first + last) which uses the pointer module only before and after message passing fails to converge. We also run the pointer module every 2, 3, 6 message passing steps in voxel GNN. Using the pointer module every 2 message passing steps yields the best performance in both FID and connectivity accuracy. Reviewing the program graph multiple times during the message passing process might ensure that the information from the program graph is always considered and provide shorter paths for gradient backpropagation. \subsection{Variation Study} In Figure \ref{fig:variation}, we visualize examples generated by fixing the program graph while changing the space partition in the voxel graph and noise. The model is able to generate different designs with different patterns, orientations, etc. based on the given noise and design space partition. \subsection{Ablation Study} We run ablation studies on discriminator, positional encoding, and relative position. The results are summarized in Table \ref{table:ablation}. We found that it is necessary to use both story discriminator and building discriminator as using only one leads to inferior performance. Though having message-passing in the voxel GNN, story discriminator has difficulty evaluating inter-story relations and the overall 3D geometry. The better connectivity accuracy with * actually results from noisy low-quality outputs where the layouts are fragmented. Building discriminator proves to play a more crucial role in learning the task, but adding story discriminator significantly improves the output design quality. We also show that using positional encoding (PE) to encode story indices performs better than directly using it (i.e. $1,2\dots$). Lastly, training without relative position (RP) ends up generating low quality designs. It shows that relative position is an indispensable component for our model, since it can capture the direct spatial relationships between connected voxels. \begin{table}[h] \begin{center} \begin{tabular}{|l|c|c|} \hline Ablation Study & FID & Con.\\ \hline\hline Ours & 0.0845 & 0.569\\ \hline\hline Story discriminator only & 6.8061 & *0.777 \\ Building discriminator only & 1.0464 & 0.459\\ \hline\hline No PE & 0.1512 & 0.507\\ No PE + No RP & 0.8333 & 0.489\\ \hline \end{tabular} \end{center} \caption{Ablation study results on discriminator, positional encoding (PE), and relative position (RP). * The higher accuracy here is caused by fragmented low-quality outputs. } \label{table:ablation} \end{table} \subsection{Intermediate Results} \begin{figure*}[t!] \centering \includegraphics[width=\textwidth]{figure/intermediate_results-03-04ss.jpg} \caption{Visualization of designs generated by all pointer modules for every 2 message passing steps in Voxel GNN during inference.} \label{fig:intermediate} \end{figure*} In voxel GNN, we run 12 message passing layers and use the pointer module every 2 layers. Since every mask and attention computed by the pointer modules represents a design solution, we are curious to see the "design process" of our model by visualizing the intermediate designs during inference. As shown in Figure \ref{fig:intermediate}, before the voxel message passing, the first attention initializes a seemly random design, trying to allocate only the lobby/corridor type. It makes sense since the decision is only based on the program graph and individual voxel nodes. Interestingly, starting from the second attention, the model chooses to start over and gradually grow the voxels. This behavior aligns with the message passing process since the information from a far distance will flow in with more passing steps. The model also tries to refine the design by overwriting some of the decisions made in previous steps. For example, in the first row of Figure \ref{fig:intermediate}, at layer 6, the isolated restroom (in magenta color) is eliminated at layer 8. \subsection{User Study} \label{sec:user_study} To further examine the quality, we conduct user study with 20 professional architects. Each architect is given 48 design pairs that cover all the combinations of the ground truth, House-GAN, and our model. Given a pair of designs, the better design gets 1 point while the worse one gets -1 point. If it's a tie, both get 0 scores. The average score of a method should range between 1 and -1. The results are shown in Figure \ref{fig:user_study} and it should be read row-by-row. Our model and ground truth defeats House-GAN with scores 0.85 and 0.92 respectively. The ground truth score is only 0.37 when compared to our model, which means in many cases architects cannot clearly tell the difference between the ground truth and ours. \begin{figure}[t] \centering \includegraphics[width= 0.2\textwidth]{figure/bg_usertest-11-11.jpg} \caption{The pairwise quality scores between ground truth (G.T.), our model, and House-GAN(H.G.).} \label{fig:user_study} \end{figure} \subsection{Case Study} To understand if our pipeline can be useful to the professional building design process, we invite an architect to create the volumetric design using our pipeline and then complete a detailed building design. As shown in the Figure~\ref{fig:teaser}, the results are realistic and aesthetically appealing. The user do feel the pipeline largely increased the efficiency of the design process. For the creation process and detailed feedback, please refer to the supplementary. \subsection{Failure Case} \begin{figure}[t] \centering \includegraphics[width= 0.4\textwidth]{figure/failure_case-07ss.jpg} \caption{Visualation of the three common flaws in generated volumetric designs. Left: the input program graph. Right: a floor plan in the generated volumetric design given this program graph} \label{fig:failure} \end{figure} We observe three types of common flaws in the volumetric designs generated from our model: 1) missing nodes; 2) missing edges; and 3) disconnected rooms, as visualized in Figure \ref{fig:failure}. One potential cause of these flaws is that our discriminator only considers the program type on each voxel node instead of the attention between voxel nodes and program graph. Therefore, the discriminator lacks information regarding the specific program nodes which the voxel nodes point to. Some of our failed attempts to resolve these flaws are introduced in the supplementary material and we leave the solution for future work. \section{Conclusion} In this paper, we try to provide a novel pipeline, Building-GAN, to improve the efficiency on a realistic professional task, volumetric design in the architectural and construction industry. We invent a 3D representation, voxel graph, to represent building designs, and design a generator with a cross-modal pointer module to connect the program graph and voxel graph. Our extensive evaluations, including user testing and user study, show that architects can create numerous valid and valuable designs by interacting with Building-GAN. Future works include enforcing the constraints, such as connectivity, TPR, and FAR, as well as extending the voxel graph for non-cuboid geometries. We will release our code, model, and dataset, and invite the research community to work together on design-related problems in the industries. \section{Dataset} \begin{figure}[h] \centering \includegraphics[width=0.43\textwidth]{figure/syn1-19.jpg} \caption{Volumetric design samples in the synthetic dataset} \label{fig:syn_1} \end{figure} \begin{figure}[h] \centering \includegraphics[width=0.42\textwidth]{figure/syn1-20.jpg} \caption{Continued} \label{fig:syn_2} \end{figure} The synthetic dataset contains 120,000 volumetric designs that are represented in voxel graph with JSON format. They can also be converted into either conventional voxels, meshes, or solid geometries. For each design, the site conditions such as the shape and maximum height are randomly generated from the distribution shown in Figure \ref{fig:floor_dist} and \ref{fig:site_dist}. The site conditions in Building-GAN are also represented in voxel graphs as part of the input conditions. Other conditions, such as the targeted TPR and FAR, and the input program graph (bubble diagram) are all restored in JSON format. \begin{figure}[h] \centering \includegraphics[width=0.5\textwidth]{figure/floor_dist_test.png} \caption{Number of stories distribution} \label{fig:floor_dist} \end{figure} \begin{figure}[h] \centering \includegraphics[width=0.5\textwidth]{figure/site_dist_test.png} \caption{Site area distribution} \label{fig:site_dist} \end{figure} Figure \ref{fig:syn_1} and \ref{fig:syn_2} present several examples in the synthetic dataset. We apply patterns and rules provided by professional architects to generate this dataset. For example, we can observe that some of the buildings have symmetric elevators facing each other (e.g., top left in Figure \ref{fig:syn_1}), while some corridors form left and right wings for the same set of functional spaces (e.g., bottom right in Figure \ref{fig:syn_1}. Also, the first story usually has a different layout than other stories due to the entrance and lobby requirements for commercial buildings. Moreover, the first story often has a higher ceiling compared to other floors. Besides the interior functionalities, building façades, i.e., the exterior shapes of buildings, is also an important aesthetic factor to consider in the architectural design process. To increase the façade's variety in this dataset, we generate story partition patterns that divide stories into groups. For instance, the top two stories in the top left in Figure \ref{fig:syn_1} have a different shape than the stories below them and therefore create a shape feature in 3D. In practice, such a shape feature can have functional purposes such as a terrace or roof garden. Please refer to Section \ref{sec:A8} Case Study for a architect-generated building design example. \section{Implementation Detail} We implement our model in PyTorch. The latent dimension is 128 and the noise dimension (for both program and voxel graphs) is 32. The message passing steps in program and voxel GNN are 4 and 12 respectively. We summarize MLP specifications in Table \ref{table:MLP_spec} and the pseudo-code of our generator is given in Algorithm \ref{alg:generator}. \begin{table*}[h] \begin{center} \begin{tabular}{|l|c|c|} \hline \multicolumn{3}{|c|}{Generator}\\ \hline \multirow{3}{*}{Program GNN}& Encoder MLP & Dense(input=5+32, output=128)\\ & Message MLP & Dense(input=256, output=128) \\ & Update MLP & Dense(input=384, output=128,act=LeakyReLU) \\ \hline \multirow{5}{*}{Voxel GNN}& Encoder MLP & Dense(input=4+32, output=128)\\ & Message MLP & Dense(input=256+3, output=128) \\ & Update MLP & Dense(input=256, output=128,act=LeakyReLU) \\ & Mask MLP & Dense(input=128, middle=64, output=2)\\ & Attention $W_{x}$, $W_{v}$&Dense(input=128, output=128)\\ \hline \hline \multicolumn{3}{|c|}{Discriminator}\\ \hline \multirow{6}{*}{Voxel GNN}& Feature Encoder MLP & Dense(input=4, output=128)\\ & Label Encoder MLP & Dense(input=6, output=128) \\ & Message MLP & Dense(input=512+3, output=256) \\ & Update MLP & Dense(input=512, output=256,act=LeakyReLU) \\ & Building Decoder MLP & Dense(input=256, middle=128, output=1,act=Sigmoid)\\ & Story Decoder MLP & Dense(input=256, middle=128, output=1,act=Sigmoid)\\ \hline \end{tabular} \end{center} \caption{Model specifications.} \label{table:MLP_spec} \end{table*} \begin{figure*}[h!] \centering \includegraphics[width=\textwidth]{figure/3DDescriptor.png} \caption{The network architecture of the 3D Descriptor Net used for FID computation} \label{fig:3ddescriptor} \end{figure*} \begin{figure*}[h!] \centering \includegraphics[width=\textwidth]{figure/3Ddescriptorresult.png} \caption{Synthesized examples by 3D Descriptor Net} \label{fig:3ddescriptorresult} \end{figure*} The 120,000 data is split to 96,000 data as training set and 24,000 data as test set. We train the models using two NVIDIA GV100 GPUs and an Intel i7 CPU with 16 cores. The total training epoch is 50 using Adam Optimizer ($b_{1}=0.5, \;b_{2}=0.999$) and batch size 8. The learning rates of the generator and the discriminator are both 0.0001. The generator updates its parameters every 5 discriminator update steps. \begin{algorithm}[h] \caption{Generator} \label{alg:generator} \begin{algorithmic} \STATE {\bfseries Input:} program graph $x$, voxel graph $v$, noise $z^{p}, z^{v}$ \\\hrulefill \STATE 1: Program GNN \STATE $x^{0} \gets$ Program Graph Encoder($x, z^{p}$) \FOR{$t=0$ {\bfseries to} $T-1$} \STATE $x^{t+1} \gets$ Program Graph Update($x^{t}$) \ENDFOR \\\hrulefill \STATE 2: Voxel GNN \STATE $v^{0} \gets$ Voxel Graph Encoder($v, z^{v}$) + PE($v$) \STATE \_ , \_ , $v^{0} \gets$ Pointer($v^{0}, x^{T}$) \FOR{$t=0$ {\bfseries to} $T-1$} \STATE $v^{t+1} \gets$ Voxel Graph Update($v^{t}$) \IF{$t<T-1$} \STATE \_ , \_ , $v^{t+1} \gets$ Pointer($v^{t}, x^{T}$) \dots (Optional) \ENDIF \ENDFOR \STATE $mask^{T}, att^{T}, v^{T} \gets$ Pointer($v^{T}, x^{T}$) \end{algorithmic} \end{algorithm} \section{3D Descriptor Network for FID score} The backbone of the inference network used for FID computation is 3D Descriptor Net. We replace all convolution layers with 6 residual blocks due to the higher complexity of our data and insert a dense layer to convert the 128 dimension embedding to a scalar energy value. The model architecture is illustrated in Figure \ref{fig:3ddescriptor}. We train the model for 400 epochs when the average mean square error decreases to 0.0121. Some synthesized volumetric designs generated unconditionally by the trained 3D Descriptor Network are visualized in Figure \ref{fig:3ddescriptorresult}. \begin{figure*}[h] \centering \includegraphics[width=0.85\textwidth]{figure/more-12s.jpg} \caption{More results generated from Building-GAN. Left: Input Program Graph. Right: Output Volumetric Design.} \label{fig:more1} \end{figure*} \section{Additional Results} \subsection{Program Graph to Volumetric Design} Figure \ref{fig:more1} to \ref{fig:more3} shows additional results generated by Building-GAN given the program graphs (bubble diagrams). \begin{figure*}[h] \centering \includegraphics[width=0.95\textwidth]{figure/more-13s.jpg} \caption{Continued.} \label{fig:more2} \end{figure*} \begin{figure*}[h] \centering \includegraphics[width=0.95\textwidth]{figure/more-14s.jpg} \caption{Continued.} \label{fig:more3} \end{figure*} \subsection{Floor Plans from Volumetric Design} Figure \ref{fig:layou1} to \ref{fig:layou3} provide examples of floor plan layouts sliced from volumetric designs generated by Building-GAN. \begin{figure*}[h] \centering \includegraphics[width=0.95\textwidth]{figure/layout-1s.jpg} \caption{Floor plan layouts from volumetric design.} \label{fig:layou1} \end{figure*} \begin{figure*}[h] \centering \includegraphics[width=0.95\textwidth]{figure/layout-2s.jpg} \caption{Continued.} \label{fig:layou2} \end{figure*} \begin{figure*}[h] \centering \includegraphics[width=0.95\textwidth]{figure/layout-3s.jpg} \caption{Continued.} \label{fig:layou3} \end{figure*} \section{FAR and TPR Study} We extend our comparative study and ablation study conducted in Section 5 on two more metrics: Floor Area Ratio (FAR) distance and Target Program Ratio (TPR) accuracy. The results are shown in Table~\ref{table:baseline_comparison-far} and Table~\ref{table:ablation-far} respectively. We use the difference between the actual FAR and the target FAR over the target FAR to calculate the FAR distance for each design. The TPR accuracy of each design is calculated as 1 minus the sum of the absolute difference between the actual program ratio and the target program ratio of each room type. All results are averaged over 10,000 samples. Table \ref{table:baseline_comparison-far} shows that TPR improves slightly when we increase the number of message passing layers, while FAR slightly decreases. We can also see that both FAR and TPR improve while raising the frequency for the pointer module. Table \ref{table:ablation-far} shows that building discriminator helps significantly on both FAR and TPR since these two values are building-level properties. \begin{table}[h] \begin{center} \begin{tabular}{|l|c|c|c|} \hline Method & Parameter & FAR & TPR\\ \hline\hline House-GAN & - & 0.853 & 0.528\\ \hline Ours & - & \cellcolor{lightgray} 1.210 & \cellcolor{lightgray} 0.772\\ \hline \hline \multirow{5}{*}{ \begin{tabular}{@{}l@{}} Voxel Layer \\ (Pointer Frequency \\ = every 2 steps)\end{tabular} } & 4 & 1.075 & 0.744\\ & 6 & 1.159 & 0.759\\ & 8 & 1.117 & 0.749\\ & 10 & 1.144 & 0.754\\ & 12 & \cellcolor{lightgray} 1.210 & \cellcolor{lightgray} 0.772\\ \hline \multirow{4}{*}{ \begin{tabular}{@{}l@{}}Pointer Frequency \\ (Voxel Layer = 12)\end{tabular} } & first + last & 2.026 & 0.512\\ & every 6 steps& 1.290 & 0.743\\ & every 3 steps & 1.393 & 0.734\\ & every 2 steps & \cellcolor{lightgray} 1.210 & \cellcolor{lightgray} 0.772\\ \hline \end{tabular} \end{center} \caption{Quantitative evaluation using FAR distance and TPR accuracy. We compare our baseline model to House-GAN. We also experiment baseline models with different numbers of voxel layer and pointer frequencies.} \label{table:baseline_comparison-far} \end{table} \begin{table}[h] \begin{center} \begin{tabular}{|l|c|c|} \hline Ablation Study & FAR & TPR\\ \hline\hline Ours & 1.210 & 0.772\\ \hline\hline Story discriminator only & 2.026 & 0.207 \\ Building discriminator only & 1.247 & 0.717\\ \hline\hline No PE & 1.260 & 0.731\\ No PE + No RP & 1.065 & 0.673\\ \hline \end{tabular} \end{center} \caption{Ablation study results of FAR distance and TPR accuracy on discriminator, positional encoding (PE), and relative position (RP).} \label{table:ablation-far} \end{table} In our best model, TPR is around 77\% accurate, and FAR is around 1.2, which means about 15-20\% error on each floor. Both numbers are acceptable given there are no explicit loss terms applied to these two conditions. However, we found that these two values are relatively trivial to learn compared to connectivity. First, FAR is highly dependent on the number of stories, i.e., higher buildings have higher FAR. Since the number of stories is given by the program graph, we can expect the FAR of the generated designs to have similar distribution if the model learns the overall shape distribution. Secondly, although TPR values vary in each program graph, they are following some prior design rules. For example, there should not be any building with elevators covering more than 30\% of the floor area. Therefore, as long as the model learns the general type distribution on voxel graphs, TPR will not be significantly off. Based on the analysis above, we choose not to use TPR and FAR as our evaluation metrics. \section{Extrapolation Results} \section{User Study Details} Figure~\ref{fig:user_study_1} shows a screenshot of our user study interface. A subject is presented with annotations of the room types on the top, followed by a pair of generated volumetric designs for each question. In addition, a reference page is provided to show a set of ground-truth volumetric designs, as shown in Figure~\ref{fig:user_study_2}. Each subject is given 48 questions and asked to choose one of the three possible answers (“A is better”, “B is better”, “Similar”), where the entire session takes around 20 minutes to be completed. The generated sample pairs come from pairs of randomly selected models among HouseGAN, Building-GAN, and ground-truth. We enforce that each possible pair of models is selected exactly 16 times during the entire session. \begin{figure}[h] \centering \includegraphics[width=0.4\textwidth]{figure/user_study_1.png} \caption{A screenshot of our user study interface: the annotations appear on the top, followed by a pair of generated samples for each question.} \label{fig:user_study_1} \end{figure} \begin{figure}[h] \centering \includegraphics[width=0.4\textwidth]{figure/user_study_2.png} \caption{A screenshot of the reference page of our user study.} \label{fig:user_study_2} \end{figure} \section{Case Study} \label{sec:A8} \subsection{Workflow and Setup} We invite an architect to go through the design process using Building-GAN and see if he can speed up the current workflow and create good volumetric designs. The typical workflow using conventional tools is as follow: \begin{enumerate} \small \item Collect the project information, including site boundary, height restriction, FAR, and the desired program requirements (TPR) from the client. \item Create the bubble diagram and drawing the partitions or grids on the site. \item Make 2D drawings, 3D models, or physical models using foam or foam board to create volumetric designs. \item Arrange functional space, also called “core” in the architectural industry, including elevators, stairs, restrooms, mechanical rooms. \item Repeat from step 2 to 4 until the design is satisfying. \item Draw and model the interior details as well as the façade design. \item Create drawings, renderings, and presentation slides. \item Present to the client and repeating from any of the previous steps if the client disagrees with the current proposal. \end{enumerate} Based on the evaluation from a professional architecture firm, this process usually takes about two to three weeks with three architectural designers. Building-GAN is designed for speeding up the loop from steps 2 to 5, where the user can quickly specify the requirements and explore the design options interactively by modifying the bubble diagram and voxel partitions. In this case study, we ask the architect to create the volumetric design using Building-GAN and complete step 6 and 7. We measure the time he spends on each step and collect feedback from him. \subsection{Results} \begin{table}[h] \begin{center} \begin{tabular}{|l|c|c|} \hline Tasks & Time (hrs)\\ \hline Draw program graph and site partitions & \cellcolor{lightgray} 0.2\\ Explore the design options & \cellcolor{lightgray} 0.25\\ Modify the generated volumetric design & 1\\ Design the interior and facade & 8\\ Choose materials and set up for rendering & 2.5\\ Render images from different angles & 3\\ \hline \end{tabular} \end{center} \caption{Time spent by the architect in each task of the case study.} \label{table:time_task} \end{table} As shown in Table \ref{table:time_task}, the architect uses only 10 minutes to set up the problem and 15 minutes to find the desired volumetric design. The total labor hours are reduced to less than two days. Although more adjustment and documentation work might be needed for a formal presentation to the client, the architect does feel a significant speed up on exploring valid design options while still being able to drive the design idea by himself. This is an important feedback for us--Building-GAN can help complete professional level tasks while making the architect feel their creativity is not limited nor replaced by providing real-time interaction with the system. In addition, we found that the architect can easily modify the flawed results generated from Building-GAN. As long as the generated design is conceptually or roughly correct, meaningful, or inspiring, the architect can fix those flaws and move on to the next stage. Examples can be found in Figure \ref{fig:modification}. The rendered images of the completed building design are shown in Figure \ref{fig:case1} and \ref{fig:case2}. \begin{figure}[h] \centering \includegraphics[width=0.35\textwidth]{figure/modification-21s.jpg} \caption{Comparison between the Building-GAN output and the modified volumetric design by the architect. There are five major types of modifications: 1) extending the stairs and elevators to the roof; 2) adjusting the boundaries of rooms for alignments; 3) filling the gaps; 4) removing redundant rooms; 5) expanding or shrinking the required space.} \label{fig:modification} \end{figure} \section{Failed Attempts} The failure cases shown in the paper is very likely because our discriminator only observes the program types on the voxel nodes to evaluate whether the entire design is realistic or not. Since the program graph is not taken as input, the discriminator cannot critic on missing nodes and missing edges on the program graph, nor the fragmented layouts (disconnected rooms). To resolve this issue, we experiment with aggregating the voxel node embeddings back to the program nodes. More specifically, the attention output from the generator indicates the program node selected by each voxel node, and we aggregate the voxel node embeddings that point to the same program node together as illustrated in Figure~\ref{fig:failureback}. The aggregated embedding is expected to provide information that tells if there's no voxel node pointing to the program node and if the voxel nodes form connected rooms. Though the attempt fails to generate promising results, we think it is helpful to share the lessons learned. \begin{figure}[h] \centering \includegraphics[width=.45\textwidth]{figure/FailureBack.png} \caption{We tried aggregating the voxel embeddings back to program graph, but the GAN training is not stable.} \label{fig:failureback} \end{figure} We experiment with two approaches that apply this concept. First, a program discriminator can be designed by aggregating the voxel node embeddings to the program nodes. Then the program discriminator classifies each program node. Additional message passing can be added before the aggregation based on "same program type edges (shown in \ref{fig:failureback})" or after the aggregation to identify missing edges. However, the GAN training becomes unstable after adding the program discriminator. It is possibly due to the dynamic aggregation from the generated designs. This operation is also non-differentiable with respect to the attention output computed by the generator since it is used as aggregation indices. The other approach is to take the voxel embeddings from the generator and apply an explicit link prediction loss using contrastive learning. The positive and negative pairs are sampled from program edges and missing edges respectively. Cosine similarity and InfoNCE loss are used to encourage connected program nodes to have more similar aggregated embeddings. The link prediction loss is added to the GAN loss with a specific weight. Our experiments also show unstable GAN training after we add the link prediction loss. From the failed attempts, we learn that it is hard to provide stable gradients by aggregating the voxel embeddings back to program nodes. We leave this challenge for future work. \begin{figure*}[h] \centering \includegraphics[width=0.7\textwidth]{figure/case-16.jpg} \caption{Results of the design process by the architect using Building-GAN.} \label{fig:case1} \end{figure*} \begin{figure*}[h] \centering \includegraphics[width=0.7\textwidth]{figure/case-17.jpg} \caption{Continued.} \label{fig:case2} \end{figure*}

318475be544e44a64ff9a395466125dbecee5ebe

\section{Introduction} Multi-view Stereo (MVS) aims to reconstruct the underlying 3D scene or estimate the dense depth map using multiple neighboring views. It plays a key role in a variety of 3D vision tasks. With high-quality cameras becoming more and more accessible, there are growing interests in developing reliable and efficient stereo algorithms in various applications, such as 3D reconstruction, augmented reality, and autonomous driving. As a fundamental problem in computer vision, MVS has been extensively studied~\cite{Furukawa:2015:MST}. Recent research shows that deep neural networks, especially convolutional neural networks (CNNs), lead to more accurate and robust systems than traditional solutions. Several approaches~\cite{kusupati2019normal,yu2020fast} report exceptional accuracy on challenging benchmarks like ScanNet~\cite{scannet17cvpr} and SUN3D~\cite{xiao2013sun3d}. State-of-the-art CNN-based multi-view approaches typically fall into three categories: 1) Variants of a standard 2D UNet architecture with feature correlation~\cite{mayer2016large,liang2018learning}. However, these approaches work best for rectified stereo pairs, and extending them to multi-view is nontrivial. 2) Constructing a differential 3D cost volume~\cite{huang18deepmvs,im2019dpsnet,yao2018mvsnet,nie2019multi,yao2019recurrent,gu2020cascade}. These algorithms significantly improve the accuracy of MVS, but at the cost of heavy computational burdens. Furthermore, the predicted depth map by 3D convolution usually contains salient artifacts which have to be rectified by a 2D refinement network~\cite{im2019dpsnet}. 3) Maintain a global scene representation and fuse multi-view information through ray-casting features from 2D images~\cite{murez2020atlas}. This paradigm cannot handle large-scale scenes because of the vast memory consumption on maintaining a global representation. \begin{figure} \includegraphics[width=0.85\linewidth]{Figures/teaser.png} \vspace{-0.1in} \caption{Inference frame per second (FPS) vs.\ depth error (AbsRel) on ScanNet~\cite{scannet17cvpr}. Our model achieve significant reduction in inference time, while maintaining state-of-the-art accuracy. } \label{fig:teaser} \vspace{-0.1in} \end{figure} Aside from multi-view depth estimation, we have also witnessed the tremendous growth of single-view depth prediction networks~\cite{lee2019big,yin2019enforcing,xu2017multi,qi2018geonet,Yang_2021_CVPR}. As shown in Table \ref{table:scannet}, Bts~\cite{lee2019big} has achieved impressive result on ScanNet~\cite{scannet17cvpr}. Single-view depth prediction roots in learning feature representations to capture image semantics, which is orthogonal to correspondence-computation in multi-view techniques. A natural question is how to combine single-view depth cues and multi-view depth cues. We introduce MVS2D\xspace that combines the strength of single-view and multi-view depth estimations. The core contribution is an attention mechanism that aggregates features along epipolar lines of each query pixel on the reference images. This module captures rich signals from the reference images. Most importantly, it can be easily integrated into standard CNN architectures defined on the input image, introducing relatively low computational cost. Our attention mechanism possesses two appealing characteristics: 1) Our network only contains 2D convolutions. 2) Besides relying on the expressive power of 2D CNNs, the network seamlessly integrates single-view feature representations and multi-view feature representations. Consequently, MVS2D\xspace is the most efficient approach compared to state-of-the-art algorithms (See Figure~\ref{fig:teaser}). It is $48\times$ faster than NAS~\cite{kusupati2019normal}, $39\times$ faster than DPSNet~\cite{im2019dpsnet}, $10\times$ faster than MVSNet~\cite{yao2018mvsnet}, $4.7\times$ faster than FastMVSNet~\cite{yu2020fast}, and almost $2\times$ speed-up over the most recent fastest approach PatchmatchNet~\cite{wang2021patchmatchnet}. In the mean-time, MVS2D\xspace achieves state-of-the-art accuracy. Intuitively, the benefit of MVS2D\xspace comes from the early fusion of the intermediate feature representations. The outcome is that the intermediate feature representations contain rich 3D signals. Furthermore, MVS2D\xspace offers ample space where we can design locations of the attention modules to address different inputs. One example is when the input camera poses are inaccurate, and corresponding pixels deviate from the epipolar lines on the input reference images. We demonstrate a simple solution, which installs multi-scale attention modules on an encoder-decoder network. In this configuration, corresponding pixels in down-sampled reference images lie closer to the epipolar lines, and MVS2D\xspace detect and rectify correspondences automatically. We conduct extensive experiments on challenging benchmarks ScanNet~\cite{scannet17cvpr}, SUN3D~\cite{xiao2013sun3d}, RGBD~\cite{sturm12iros} and Scenes11~\cite{sturm12iros}. MVS2D\xspace achieves the state-of-the-art performance on nearly all the metrics. Qualitatively, compared to recent approaches~\cite{im2019dpsnet,yao2018mvsnet,kusupati2019normal,yu2020fast}, MVS2D\xspace helps generate higher quality 3D reconstruction outputs. \section{Related Works} \label{Section:Related:Works} \begin{figure*}[ht!] \centering \includegraphics[width=0.8\linewidth]{Figures/network.png} \vspace{-0.1in} \caption{Network architecture of MVS2D\xspace. We employ a 2D UNet structure $\mathcal{F}$ to make the depth prediction on $I_0$, while injecting multi-view cues extracted using $\mathcal{G}$ through the Epipolar Attention Module. Dashed arrows only exist in \textbf{Ours-robust} model (Section \ref{sec3.4}). We highlight that the proposed epipolar attention module can be easily integrated into most 2D CNNs. } \label{fig:network} \vspace{-0.1in} \end{figure*} \myparagraph{Recent advances of multi-view stereo.} Multi-view stereo algorithms can be categorized into depth map-based approaches, where the output is a per-view depth map, or point-based approaches, where the output is a sparse reconstruction of the underlying scene (cf.~\cite{Furukawa:2015:MST}). Many traditional multi-view stereo algorithms follow a match-then-reconstruct paradigm~\cite{furukawa2009accurate} that leverages the sparse-nature of feature correspondences. Such a paradigm typically fails to reconstruct textureless regions where correspondences are not well-defined. Along this line, Zbontar~\etal~\cite{zbontar2015computing} provided one of the first attempts to bring the power of feature learning into multi-view stereo. They proposed a supervised feature learning approach to find the correspondences. Recently, researchers have found that depth map-based approaches~\cite{kar2017learning,huang18deepmvs,yao2018mvsnet,im2019dpsnet,yao2019recurrent} are more favorable than those that follow the match-and-reconstruct paradigm. A key advantage of these approaches is that they can utilize the efficiency of regular tensor operations. \cite{yao2018mvsnet,im2019dpsnet} proposed an end-to-end plane-sweeping stereo approach that constructs learnable 3D cost volume. While MVSNet~\cite{yao2018mvsnet} focuses on the reconstruction of 3D scene, DPSNet~\cite{im2019dpsnet} focuses on evaluating the per-view depth-map accuracy. Researchers have also explored other 3D representations to regularized the prediction, such as point clouds~\cite{chen2019point}, surface normals~ \cite{kusupati2019normal}, or meshes~\cite{wang2020mesh}. There are also several benchmark datasets for this task~\cite{yao2020blendedmvs,xiao2013sun3d,sturm12iros,aanaes2016large,scannet17cvpr,ummenhofer17demon}. \myparagraph{Cost volume for multi-view stereo.} A recent line of works on multi-view stereo utilizes the notion of \textsl{cost volume}, which contains feature matching costs for a pair of images~\cite{hosni2012fast}. This feature representation has been successfully implemented in various pixel-wise matching tasks like optical flow~\cite{sun2018pwc}. Authors of MVSNet~\cite{yao2018mvsnet} and DPSNet~\cite{im2019dpsnet} proposed to first construct a differentiable cost volume and then use the power of 3D CNNs to regularize the cost volume before predicting per-pixel depth or disparity. Most recent state-of-the-art approaches follow such a paradigm~\cite{yao2019recurrent,nie2019multi, gu2020cascade, chen2019point, yu2020fast,long2020multiview}. However, the size of the cost volume ($C\times K\times H\times W$) is linearly related to the number of depth hypotheses $K$. These approaches are typically slow in both training and inference. For example, DPSNet~\cite{im2019dpsnet} takes several days to train on ScanNet; NAS~\cite{kusupati2019normal} takes even longer because of its extra training of a depth-normal consistency module. Recently, Murez~\etal~\cite{murez2020atlas} proposed to construct a volumetric scene representation from a calibrated image sequence for scene reconstruction. However, their approach is very memory demanding due to the high memory requirement of global volumetric representations. \myparagraph{Efficient multi-view stereo.} Several recent works aim at reducing the cost of constructing cost volumes. Duggal~\etal~\cite{duggal2019deeppruner} prune the disparity search range during cost volume construction. Xu~\etal~\cite{xu2020aanet} integrate adaptive sampling and deformable convolution into correlation-based methods~\cite{mayer2016large,liang2018learning} to achieve efficient aggregation. Several other works~\cite{yao2019recurrent,tankovich2020hitnet,gu2020cascade} employ iterative refinement procedures. The above approaches either only work for pairwise rectified stereo matching tasks or have to construct a 3D cost-volume. Alternatively, Poms~\etal~\cite{poms2018learning} learn how to merge patch features for 3D reconstruction efficiently. Badki~\etal~\cite{badki2020bi3d} convert depth estimation as a classification task, but the resulting accuracy is not state-of-the-art. Recently, Yu~\etal~\cite{yu2020fast} proposed constructing a sparse cost-volume through regular sub-sampling and then applying Gauss-Newton iterations to refine the dense depth map. Wang~\etal~\cite{wang2021patchmatchnet} proposed a highly-efficient Patchmatch-inspired approach for MVS tasks. In contrast, we take an orthogonal approach based on the attention-driven 2D convolutions. \myparagraph{Attention in 3D vision} Attention mechanism has shown prominent results on both natural language processing (NLP) tasks~\cite{vaswani2017attention} and vision tasks~\cite{wang2018non}. Recently, self-local attention~\cite{ramachandran2019stand, shaw2018self} has shown promising results compared with the convolution-based counterparts. Several recent works that build an attention mechanism in MVS~\cite{luo2020attention,zhang2021long,long2020multiview}, but still rely on 3D CNNs and cannot avoid constructing a heavy-weighted cost volume. A promising direction is to utilize a geometry-aware 2D attention mechanism. Recent works have shown that this paradigm works well for active sensing~\cite{cheng2018geometry,tung2019learning} and neural rendering~\cite{tobin2019geometry}. Motivated by these works, we propose an epipolar attention module in this paper. The key contribution is a network design that aggregates single-view depth cues and multi-view depth cues to output accurate MVS outputs. \section{Approach} We provide an overview of the network architecture of MVS2D\xspace in Fig.~\ref{fig:network}. We operate in a multi-view stereo setting (Sec.~\ref{sec3.1}), and employ a 2D UNet structure in our network design (Sec.~\ref{sec3.2}). Our core contribution is the epipolar attention module (Sec.~\ref{sec3.3}--\ref{sec3.4}), which is highly accurate and efficient (Sec.~\ref{sec3.5}) for depth estimation (Sec.~\ref{sec3.6}). \subsection{Problem Setup} \label{sec3.1} We aim to estimate the per-pixel depth for a source image $I_0 \in \mathcal{R}^{h\times w \times 3}$, given $n$ reference images $\{I_i\}_{i=1}^{n}$ of the same size captured at nearby views. We assume the source image and the reference images share the same intrinsic camera matrix $\set{K} \in \mathcal{R}^{3\times3}$, which is given. We also assume we have a good approximation of the relative camera pose between the source image and each reference image $T_i = (R_i|\bs{t}_i)$ ,where $R_i\in\text{SO}(3)$ and $\bs{t}_i\in \mathcal{R}^3$. $T_i$ usually comes from the output of a multi-view structure-from-motion algorithm. Our goal is to recover the dense pixel-wise depth map associated with $I_0$. We denote the homogeneous coordinate of a pixel $p_0$ in the source image $I_0$ as $\overline{\bs{p}}_0 = \begin{pmatrix} \overline{p}_{0,1},\overline{p}_{0,2},1 \end{pmatrix}^T$. Given the depth $d_0\in \mathcal{R}$ of $p_0$, the unprojected 3D point of $p_0$ is $$ \bs{p}_0(d_0) = d_0\cdot (\set{K}^{-1}\bs{p}_0). $$ Similarly, we use $\bs{p}_i(d_0)$ and $\overline{\bs{p}}_i(d_0)$ to denote respectively the 3D coordinates and homogeneous coordinates of $\bs{p}_0(d_0)$ in the $i$-th image's coordinate system. They satisfy \begin{align} \bs{p}_i(d_0) &= R_i\bs{p}_0(d_0) + \bs{t}_i, \nonumber \\ \overline{\bs{p}}_i(d_0) &= \mathcal{K}\bs{p}_i(d_0). \label{eq1} \end{align} \subsection{Network Design Overview} \label{sec3.2} In this paper, we innovate developing a multi-view stereo approach that only requires 2D convolutions. Specifically, similar to most single-view depth prediction networks, our approach progressively computes multi-scale activation maps of the source image and outputs a single depth map. The difference is that certain intermediate activation maps combine both the output of a 2D convolution operator applied to the previous activation map and the output of an attention module that aggregates multi-view depth cues. This attention module, which is the main contribution of this paper, matches each pixel of the source image and corresponding pixels on epipolar lines on the reference images. The matching procedure utilizes learned feature activations on both the source image and the reference images. The output is encoded using learned depth codes compatible with the activation maps of the source image. Formally speaking, our goal is to learn a feed-forward network $\set{F}$ with $L$ layers. With $\set{F}_j\in \mathcal{R}^{h_j\times w_j \times m_j}$ we denote the output of the $j$-th layer, where $m_j$ is its feature dimension, $h_j$ and $w_j$ are its height and width. Note that the first layer $\set{F}_1 \in \mathcal{R}^{h_1\times w_1\times 3}$ denotes the input, while the last layer $\set{F}_{L}\in \mathcal{R}^{h_L\times w_L}$ denotes the output layer containing depth prediction. Between two consecutive layers are a general convolution operator $C_j: \mathcal{R}^{h_j\times w_j\times m_j}\rightarrow \mathcal{R}^{h_j\times w_j\times m_{j+1}}$ (it can incorporate standard operators such as down-sampling, up-sampling, and max-pooling) and an optional attention module $\set{A}_j:\mathcal{R}^{h_j\times w_j\times m_j}\rightarrow \mathcal{R}^{h_j\times w_j\times m_j}$: $$ \set{F}_{j+1} = \set{C}_j\circ \set{A}_j \circ \set{F}_j. $$ As we will see immediately, the attention operator $\set{A}_j$ utilizes features extracted from the reference images. Without these attention operations, $\set{F}$ becomes a standard encoder-decoder network for single-view depth prediction. Another characteristic of this network design is that the convolution operator $\set{C}_j$ implicitly aggregates multi-view depth cues extracted at adjacent pixels. This approach promotes consistent correspondences among adjacent pixels that share the same epipolar line or have adjacent epipolar lines. \subsection{Epipolar Attention Module} \label{sec3.3} We proceed to define $\set{A}_j(p_0)$, which is the action of $\set{A}_j$ on each pixel $p_0$. It consists of two parts: \begin{equation} \set{A}_j(p_0) = \set{A}_{j}^{\text{ep}}(p_0, \{I_i\}_{i=1}^{n}) + \set{A}_j^{0}(\set{F}_j(p_0)) . \label{Eq:Attention0} \end{equation} As we will define next, $\set{A}_{j}^{\text{ep}}(p_0, \{I_i\}_{i=1}^{n})$ uses trainable depth codes to encode the matching result between $p_0$ and the reference images. $\set{A}_j^{0}:\mathcal{R}^{m_j}\rightarrow \mathcal{R}^{m_j}$ is composed of an identity map and a trainable linear map that transforms the feature associated with $p_0$ in $\set{F}_j$. The formulation of $\set{A}_{j}^{\text{ep}}(p_0, \{I_i\}_{i=1}^{n})$ uses the \textsl{epipolar context} of $p_0$. It consists of samples on the epipolar lines of $p_0$ on the reference images. These samples are obtained from sampling the depth values $d_0$ of $p_0$ and then applying (\ref{eq1}). With $p_{i}^{k}$ we denote the $k$-th sample on the $i$-th reference image. To match $p_0$ and $p_i^{k}$, we introduce a feature extraction network $\set{G}$ that has identical architecture (except the attention modules) as $\set{F}_{j_{\max}}$ where $j_{\max}$ is the maximum depth of any attention module of $\set{F}$. With $\set{G}_j(I_0,p_0)\in \mathcal{R}^{m_j}$ and $\set{G}_j(I_i, p_i^k)\in \mathcal{R}^{m_j}$ we denote the extracted features of $p_0$ and $p_i^k$, respectively. Following the practice of scaled-dot product attention~\cite{vaswani2017attention}, we introduce two additional trainable linear maps $\bs{f}_{0}^j: \mathcal{R}^{m_j}\rightarrow \mathcal{R}^{m_j}$ and $\bs{f}_{\text{ref}}^j: \mathcal{R}^{m_j}\rightarrow \mathcal{R}^{m_j}$ to transform the extracted features. With this setup, we define the matching score between $p_0$ and $p_i^{k}$ as \begin{equation} w_{ik}^j = \big(\bs{f}_{0}^j(\set{G}_j(I_0,p_0))\big)^T\big(\bs{f}_{\text{ref}}^j(\set{G}_j(I_i, p_i^k))\big). \label{Eq:Matching:Weight} \end{equation} It remains to 1) model samples that are occluded in the reference images, and 2) bridge the weights $w_{ik}^j$ defined in (\ref{Eq:Matching:Weight}) and the input to the convolution operator $\set{C}_j$. To this end, we first introduce trainable mask codes $\bs{c}_{jk}\in \mathcal{R}^{m_j}$ that correspond to the $k$-th depth sample. We then introduce $\bs{v}_{\text{in}}^j\in \mathcal{R}^{m_j}$ and $\bs{v}_{\text{out}}^j\in \mathcal{R}^{m_j}$, which are trainable codes for inside and outside samples, respectively. Define \begin{equation} \bs{v}_{ik}^{j} = \left\{ \begin{array}{cc} \bs{v}_{\text{in}}^j & 0\le \overline{p}_{i,1}^k< w, 0\le \overline{p}_{i,2}^k< h, p_{i,3}^k \ge 0,\\ \bs{v}_{\text{out}}^j & \textup{otherwise} \end{array} \right.\ \label{Eq:Mask:Codes} \end{equation} where $\overline{p}_{i}^k = (\overline{p}_{i,1}^k,\overline{p}_{i,2}^k,1)^T$, $p_{i}^k = (p_{i,1}^k,p_{i,2}^k,p_{i,3}^k)^T$. To enhance the expressive power of $\set{G}_j$, we further include a trainable linear map $\set{A}_j^{1}$ that depends only on feature of $p_0$ and not on the matching results. Combing with (\ref{Eq:Matching:Weight}) and (\ref{Eq:Mask:Codes}), we define \begin{equation} \set{A}_{j}^{\text{ep}}(p_0, \{I_i\}_{i=1}^{n}) =\set{A}_j^{1}(\set{G}_j(p_0))+\sum\limits_{i=1}^{n}\sum\limits_{k=1}^{K} \mathcal{N}(\frac{w_{ik}^j}{\sqrt{m_j}}) (\bs{v}_{ik}^j\odot \bs{c}_k)) \label{Eq:Attention:2} \end{equation} where $\mathcal{N}$ is the softmax normalizing function over $\frac{w_{ik}^j}{\sqrt{m_j}}, 1\leq k \leq K$. Substituting (\ref{Eq:Attention:2}) into (\ref{Eq:Attention0}), the final attention module is given by \begin{align} \set{A}_j(p_0) &= \set{A}_j^{0}(\set{F}_j(p_0)) + \set{A}_j^{1}(\set{G}_j(p_0))\nonumber\\ &+\sum\limits_{i=1}^{n}\sum\limits_{k=1}^{K} \mathcal{N}(\frac{w_{ik}^j}{\sqrt{m_j}})\nonumber (\bs{v}_{ik}^j\odot \bs{c}_k)). \end{align} Note that the attention modules at different layers have different weights. Eq.~\ref{Eq:Matching:Weight} can be viewed as a similarity score between source pixel and correspondence candidates. In Fig.~\ref{fig:visualization_attention}, we visualize the learned attention scores for query pixels. The true corresponding pixels on reference images have larger learned weights along epipolar lines. \begin{figure}[h] \begin{center} \includegraphics[width=0.4\textwidth]{Figures/corres.png} \end{center} \vspace{-0.2in} \caption{Visualization of attention scores. Left: source view with query pixels. Right: reference view with candidate pixels, where opacity is learned attention scores.} \label{fig:visualization_attention} \end{figure} \subsection{Attention Design for Robust Multi-View Stereo} \label{sec3.4} Since the attention module assumes that the corresponding pixels lie on the epipolar lines, the accuracy of MVS2D\xspace depends on the relative poses' accuracy between the reference images and the source image. When input poses are accurate, our experiments suggest a single attention module at the second layer of $\set{F}$ is sufficient. This leads to a highly efficient multi-view stereo network. When the input poses are inexact, we address this issue by installing attention modules at different resolutions of the input images, i.e., at different layers of $\set{F}$. This approach ensures that the corresponding pixels lie sufficiently close to the epipolar lines at those resolutions at coarse resolutions. Since the convolution operations $\set{C}_j$ at different layers of $\set{F}$ aggregate multi-view features extracted at different pixels, we find this simple network design implicitly rectifies inexact epipolar lines. Figure~\ref{fig:network} illustrates the attention modules under these two cases. \begin{table}[!t] \centering \resizebox{\linewidth}{!}{% \begin{tabular}{rccccc} \toprule \multicolumn{1}{c}{Method} & FPS (3)$\uparrow$ & FPS(7)$\uparrow$ & FPS(11) $\uparrow$& Param (M) $\downarrow$ & AbsRel$\downarrow$ \\ \hline Bts~\cite{lee2019big} & 17.0 & - & - & 46.8 & 0.117\\ \hline MVSNet~\cite{yao2018mvsnet} & 4.1 & 2.4 & 1.6 & 1.1 & 0.094\\ DPSNet~\cite{im2019dpsnet} & 1.1 & 0.7 &0.5 & 4.2& 0.094\\ FastMVS~\cite{yu2020fast}& 9.0 & 6.0 & 4.3 & 0.4 & 0.089\\ PatchmatchNet~\cite{wang2021patchmatchnet} & 21.8 & 11.6 & 8.5 &\textbf{0.2} & 0.133\\ NAS~\cite{kusupati2019normal} & 0.9 & 0.6 & 0.4 &18.0 & 0.086\\ \hline Ours-mono & \textbf{94.7} & - & - & 12.3 & 0.145\\ Ours-robust & 17.5 & 10.1 & 7.1& 24.4 & \textbf{0.059}\\ Ours & 42.9 & \textbf{29.1} & \textbf{21.8} & 13.0 & \textbf{0.059}\\ \bottomrule \end{tabular} } \caption{Quantitative comparison on computational efficiency. FPS ($V$) only applies to multi-view methods~\cite{yao2018mvsnet,yu2020fast,im2019dpsnet,kusupati2019normal} and means we use $V$ images to make the prediction. Note that numbers under the AbsRel metric are identical to those in Table~\ref{table:scannet} for ease of comparison. We use a single Nvidia V100 GPU for measuring FPS. Please refer to section~\ref{sec4.4} for additional discussions.} \label{table:speed} \vspace{-0.1in} \end{table} \subsection{Computational Complexity}\label{sec3.5} For the sake of simplicity in notation, we assume the feature channel dimension $C$ is the same in both input and output. We denote the feature height and width as $H$ and $W$ respectively and denote the kernel size of convolution layers as $k$. Suppose there are $K$ depth samples, the complexity of 3D convolution is $\mathcal{O}(C^2HWKk^3)$. For our approach, the computational complexity for executing one layer of $\set{C}\circ \set{A}$ is in total $\mathcal{O}(CHW(Ck^2+K))$. Since $K$ is usually less than $Ck^2$, our module leads to a $Kk$ times reduction in computation. The actual runtime can be found in Table~\ref{table:speed}. \subsection{Training Details}\label{sec3.6} Our implementation is based on Pytorch. For ScanNet and DeMoN, we simply optimize the $L_1$ loss between predicted and ground truth depth. For DTU, we introduce a simple modification, as was done in ~\cite{kendall2017uncertainties}, to simultaneously train a confidence prediction. We use Adam~\cite{DBLP:journals/corr/KingmaB14} optimizer with $\epsilon =10^{-8}$, $\beta = (0.9,0.999)$. We use a starting learning rate $2e^{-4}$ for ScanNet, $8e^{-4}$ for DeMoN and $2e^{-4}$ for DTU. Please refer to supp. material for more training details. \section{Experimental Results} \label{Section:Results} \subsection{Datasets}\label{sec4.1} \myparagraph{ScanNet~\cite{scannet17cvpr}} The ScanNet dataset contains 807 unique scenes with image sequences captured from different camera trajectories. We sample 86324 triple images (one source image and two reference images) for training and 666 triple images for testing. Our setup ensures the scene corresponding to test images is not included in the training set. \myparagraph{DeMoN~\cite{ummenhofer17demon}} We further validate our method on DeMoN, which is a dataset introduced by~\cite{ummenhofer17demon} for multi-view depth estimation. The training set consists of three data sources, SUN3D~\cite{xiao2013sun3d}, RGBD~\cite{sturm12iros}, and Scenes11~\cite{ummenhofer17demon}. SUN3D and RGBD contain real indoor scenes, while Scenes11 is synthetic. In total, there are 79577 training pairs for SUN3D, 16786 for RGBD, and 71820 for Scenes11. \myparagraph{DTU~\cite{aanaes2016large}} While our approach is designed for multi-view depth estimation, we additionally validate our method on the DTU dataset, which has been considered as one of the main test-bed for multi-view reconstruction algorithms. \subsection{Evaluation Metrics}\label{sec4.2} \myparagraph{Efficiency.} We benchmark our methods against baseline methods on the frame per second~(FPS) during inference. We additionally compare the FPS when increasing the number of reference views. \myparagraph{Depth Accuracy.} We use the conventional metrics of depth estimation~\cite{lee2019big} (See Table~\ref{table:metric}). Note that in contrast to monocular depth estimation evaluation, we do not factor out the depth scale before evaluation. The ability to correctly predict scale will render our method more applicable. \myparagraph{Scene Reconstruction Quality.} We further apply MVS2D\xspace for scene reconstruction. We follow PatchmatchNet~\cite{wang2021patchmatchnet} to fuse the per-view depth map into a consistent 3D model. Please refer to supp. material for quantitative and more qualitative comparisons. \myparagraph{Robustness under Noisy Input Pose.} We perturb the input relative poses $T_j$ during training and report the model performance on ScanNet test set in Table \ref{table:perturb}. Please refer to the supp. material for details of the pose perturbing procedures. \begin{table}[!h] \centering \resizebox{\linewidth}{!}{% \begin{tabular}{cc|cc} \toprule AbsRel & $\frac{1}{N}\sum_{i}\frac{|d_i - d_i^{*}|}{d_i^{*}}$ & RMSE & $\sqrt{\frac{1}{N}\sum_{i}(d_i - d_i^{*})^2}$ \\ \xrowht{10pt} SqRel & $\frac{1}{N}\sum_{i}\frac{(d_i - d_i^{*})^2}{ d_i^{*}}$ & RMSELog & $\sqrt{\frac{1}{N}\sum_{i}(\log{d_i} - \log{d_i^{*}})^2}$\\ \xrowht{10pt} AbsDiff & $\sqrt{\frac{1}{N}\sum_{i}|d_i - d_i^{*}|}$ & Log10 & $\frac{1}{N}\sum_{i}|\log_{10}d_i - \log_{10}d_i^{*}|$\\ \xrowht{10pt} $\delta < 1.25^{k}$ & $\frac{1}{N}\sum_{i}(\max(\frac{d_i}{d_i^{*}},\frac{d_i^*}{d_i} ) < 1.25^{k})$ & thre@x & $\frac{1}{N}\sum_{i}I(|d_i-d_i^{*}|<x)$\\ \bottomrule \end{tabular} } \caption{Quantitative metrics for depth estimation. $d_i$ is the predicted depth; $d_i^*$ is the ground truth depth; $N$ corresponds to all pixels with the ground-truth label. $I$ is the indicator function.} \label{table:metric} \end{table} \begin{table*}[!t] \centering \resizebox{0.95\textwidth}{!}{% \begin{tabular}{r|ccccc|ccc} \toprule \multicolumn{1}{c|}{Method} & AbsRel $\downarrow$ & SqRel $\downarrow$& log10 $\downarrow$& RMSE $\downarrow$ & RMSELog $\downarrow$& $\delta<1.25$ $\uparrow$& $\delta<1.25^2$ $\uparrow$& $\delta<1.25^3$ $\uparrow$ \\ \hline Bts~\cite{lee2019big} & 0.117 &0.052& 0.049 & 0.270 & 0.151 & 0.862 & 0.966 & 0.992\\ Bts$^*$~\cite{lee2019big} & 0.088 &0.035& 0.038 & 0.228 & 0.128 & 0.916 & 0.980 & 0.994\\ \hline MVSNet~\cite{yao2018mvsnet} & 0.094 & 0.042 & 0.040 & 0.251 & 0.135 & 0.897 & 0.975 &0.993\\ FastMVS~\cite{yu2020fast} & 0.089 & 0.038 & 0.038 & 0.231 & 0.128 & 0.912 & 0.978 & 0.993 \\ DPSNet~\cite{im2019dpsnet} & 0.094 & 0.041 & 0.043 & 0.258 & 0.141 & 0.883 & 0.970 & 0.992\\ NAS~\cite{kusupati2019normal} & 0.086 & 0.032 & 0.038 & 0.224 & 0.122 & 0.917 & 0.984 & 0.996\\ PatchmatchNet~\cite{wang2021patchmatchnet} & 0.133 & 0.075 & 0.055 & 0.320 & 0.175 & 0.834 & 0.955 & 0.987 \\ \hline Ours-mono & 0.145 & 0.065 & 0.061 & 0.300 & 0.173 & 0.807 & 0.957 & 0.990\\ Ours-mono$^*$ & 0.103 & 0.037 & 0.044 & 0.237 & 0.135 & 0.892 & 0.984 & 0.996\\ Ours-robust & \textbf{0.059} & \textbf{0.016} & \textbf{0.026} & \textbf{0.159} & \textbf{0.083} & \textbf{0.965} & \textbf{0.996} & \textbf{0.999} \\ Ours& \textbf{0.059} & 0.017 & \textbf{0.026} & 0.162 & 0.084 & 0.963 & 0.995 & \textbf{0.999} \\ \bottomrule \end{tabular} } \vspace{-0.1in} \caption{Depth evaluation results on ScanNet~\cite{scannet17cvpr}. We compare against both multi-view depth estimation methods~\cite{yu2020fast, yao2018mvsnet, im2019dpsnet,kusupati2019normal,wang2021patchmatchnet} and a state-of-the-art single-view method~\cite{lee2019big}. Our approach achieve significant improvements over top-performing method NAS~\cite{kusupati2019normal} on AbsRel. The improvements are consistent across all metrics.} \label{table:scannet} \vspace{-0.1in} \end{table*} \begin{table*} \centering \resizebox{0.95\textwidth}{!}{% \begin{tabular}{cr|ccccc|ccc} \toprule & \multicolumn{1}{c|}{Method} & AbsRel $\downarrow$& AbsDiff $\downarrow$& SqRel $\downarrow$& RMSE $\downarrow$& RMSELog $\downarrow$& $\delta<1.25$ $\uparrow$& $\delta<1.25^2$ $\uparrow$& $\delta<1.25^3$ $\uparrow$\\ \hline \parbox[t]{2mm}{\multirow{6}{*}{\rotatebox[origin=c]{90}{SUN3D (Real)}}} & COLMAP~\cite{sfm16cvpr} & 0.623 & 1.327 & 3.236 & 2.316 & 0.661 & 0.327 & 0.554 & 0.718\\ & DeMoN~\cite{ummenhofer17demon} & 0.214 & 2.148 & 1.120 & 2.421 & 0.206 & 0.733 & 0.922 & 0.963\\ & DeepMVS~\cite{huang18deepmvs} & 0.282 & 0.604 & 0.435 & 0.944 & 0.363 & 0.562 & 0.739 & 0.895\\ & DPSNet-U~\cite{im2019dpsnet} & 0.147 & 0.336 & 0.117 & 0.449 & 0.196 & 0.781 & 0.926 & 0.973\\ & NAS~\cite{kusupati2019normal} & 0.127 & 0.288 & 0.085 & 0.378 & 0.170 & 0.830 & 0.944 & 0.978\\ & Ours-robust & \underline{0.100} & \underline{0.231} & \underline{0.057} & \underline{0.313} & \underline{0.140} & \textbf{0.895} & \underline{0.966} & \underline{0.991} \\ & Ours &\textbf{0.099}& \textbf{0.224} & \textbf{0.055} & \textbf{0.304} & \textbf{0.137} & \underline{0.893} & \textbf{0.970} & \textbf{0.993} \\ \hline \parbox[t]{2mm}{\multirow{6}{*}{\rotatebox[origin=c]{90}{RGBD (Real)}}} & COLMAP~\cite{sfm16cvpr} & 0.539 & 0.940 & 1.761 & 1.505 & 0.715 & 0.275 & 0.500 & 0.724\\ & DeMoN~\cite{ummenhofer17demon} & 0.157 & 1.353 & 0.524 & 1.780 & 0.202 & 0.801 & 0.906 & \textbf{0.962}\\ & DeepMVS~\cite{huang18deepmvs} & 0.294 & 0.621 & 0.430 & 0.869 & 0.351 & 0.549 & 0.805 & 0.922\\ & DPSNet-U~\cite{im2019dpsnet} & 0.151 & 0.531 & 0.251 & 0.695 & 0.242 & 0.804 & 0.895 & 0.927\\ & NAS~\cite{kusupati2019normal} & 0.131 & 0.474 & 0.213 & 0.619 & 0.209 & 0.857 & 0.929 & 0.945\\ & Ours-robust & \textbf{0.078} & \textbf{0.311} & \textbf{0.156} & \underline{0.443} & \textbf{0.146} & \textbf{0.926} & \textbf{0.945} & \underline{0.954} \\ & Ours & \underline{0.082} & \underline{0.325} & \underline{0.165} & \textbf{0.440} & \underline{0.147} & \underline{0.921} & \underline{0.939} & 0.948 \\ \hline \parbox[t]{2mm}{\multirow{6}{*}{\rotatebox[origin=c]{90}{Scenes11 (Syn)}}} & COLMAP~\cite{sfm16cvpr} & 0.625 & 2.241 & 3.715 & 3.658 & 0.868 & 0.390 & 0.567 & 0.672\\ & DeMoN~\cite{ummenhofer17demon} & 0.556 & 1.988 & 3.402 & 2.603 & 0.391 & 0.496 & 0.726 & 0.826\\ & DeepMVS~\cite{huang18deepmvs} & 0.210 & 0.597 & 0.373 & 0.891 & 0.270 & 0.688 & 0.894 & 0.969\\ & DPSNet~\cite{im2019dpsnet} & 0.050 & 0.152 & 0.111 & 0.466 & 0.116 & 0.961 & 0.982 & 0.988\\ & NAS~\cite{kusupati2019normal} & \textbf{0.038} & \textbf{0.113} & \underline{0.067} & \textbf{0.371} & \textbf{0.095} & \underline{0.975} & \underline{0.990} & \textbf{0.995}\\ & Ours-robust & \underline{0.041} & \underline{0.141} & \textbf{0.066} & \underline{0.410} & \underline{0.099} & \textbf{0.979} & \textbf{0.991} & \underline{0.994} \\ & Ours & 0.046 & 0.155 & 0.080 & 0.439 & 0.107 & 0.976 & 0.989 & 0.993 \\ \bottomrule \end{tabular} } \vspace{-0.1in} \caption{Depth evaluation results on SUN3D, RGBD, and Scenes11 datasets(synthetic). The numbers for COLMAP, DeMoN, DeepMVS, DPSNet, and NAS are obtained from \cite{kusupati2019normal}. We achieve significant improvements on SUN3D and RGBD. We show the best number in bold and the second best with underline.} \label{table:sun3d} \vspace{-0.1in} \end{table*} \subsection{Baseline Approaches}\label{sec4.3} \myparagraph{MVSNet~\cite{yao2018mvsnet}} is an end-to-end plane sweeping stereo approach based on 3D-cost volume. \myparagraph{DPSNet~\cite{im2019dpsnet}} shares similar spirit of MVSNet~\cite{yao2018mvsnet} but focus on accurate depth map prediction. \myparagraph{NAS~\cite{kusupati2019normal}} is a recent work that jointly predicts consistent depth and normal, using extra normal supervision. \myparagraph{FastMVSNet~\cite{yu2020fast}} is a recent variant to MVSNet which accelerate the computation by computing sparse cost volume. \myparagraph{Bts~\cite{lee2019big}} is a state-of-the-art single view depth prediction network. It incorporates planar priors into network design. Additionally, we use an asterisk sign `$*$' to denote an oracle version \textbf{Bts$^*$}, where we use the ground truth depth map to factor out the global scale. \myparagraph{PatchmatchNet~\cite{wang2021patchmatchnet}} is one of the most recent state-of-the-art efficient MVS algorithm. \myparagraph{Ours-mono} is our method without the epipolar attention module, thus equivalent to single-view depth estimation. Similar to Bts$^*$, we also report the results factoring out the global scale for \textbf{Ours-mono$^*$}. \myparagraph{Ours-robust} is our method with multi-scale epipolar attention module applied on $\mathcal{F}$. \myparagraph{Ours} is our method with epipolar attention module applied only in $\mathcal{F}$'s second layer. \begin{figure*}[!t] \newcommand{\TT}[1]{\raisebox{-0.5\height}{#1}} \setlength{\tabcolsep}{1pt} \footnotesize \def0.28\textwidth{0.16\textwidth} \begin{tabular}{cccccc} Bts~\cite{lee2019big} & FastMVSNet~\cite{yu2020fast} & NAS~\cite{kusupati2019normal} & PatchmatchNet~\cite{wang2021patchmatchnet} & MVS2D\xspace (Ours) & G.T. \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/fig_depth/bts/scene0606_02-0112_0111_0113.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/fig_depth/fastmvsnet/scene0606_02-0112_0111_0113.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/fig_depth/nas/scene0606_02-0112_0111_0113.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/fig_depth/patchmatchnet/scene0606_02-0112_0111_0113.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/fig_depth/ours/scene0606_02-0112_0111_0113.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/fig_depth/gt/scene0606_02-0112_0111_0113.png}} \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/fig_depth/bts/scene0435_02-0001_0000_0002.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/fig_depth/fastmvsnet/scene0435_02-0001_0000_0002.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/fig_depth/nas/scene0435_02-0001_0000_0002.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/fig_depth/patchmatchnet/scene0435_02-0001_0000_0002.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/fig_depth/ours/scene0435_02-0001_0000_0002.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/fig_depth/gt/scene0435_02-0001_0000_0002.png}} \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/fig_depth/bts/scene0599_01-0095_0094_0096.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/fig_depth/fastmvsnet/scene0599_01-0095_0094_0096.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/fig_depth/nas/scene0599_01-0095_0094_0096.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/fig_depth/patchmatchnet/scene0599_01-0095_0094_0096.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/fig_depth/ours/scene0599_01-0095_0094_0096.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/fig_depth/gt/scene0599_01-0095_0094_0096.png}} \\ \end{tabular} \caption{Qualitative results on depth prediction. Each row corresponds to one test example. The region without ground truth depth labels is colored white in GT. Our prediction outperforms both the single-view depth estimation method~\cite{lee2019big} and other multi-view methods. } \label{fig:qualitative} \vspace{-0.1in} \end{figure*} \begin{figure*} \begin{center} \def0.28\textwidth{0.3\textwidth} \begin{tabular}{ccc} \subfigimg[width=0.28\textwidth]{PatchmatchNet}{Figures/dtu_3d/patchmatchnet_48_zoom.png}&\subfigimg[width=0.28\textwidth]{CasMVSNet}{Figures/dtu_3d/cascade_48_zoom.png}&\subfigimg[width=0.28\textwidth]{MVS2D (ours)}{Figures/dtu_3d/mvs2d_48_zoom.png}\\ \subfigimg[width=0.28\textwidth]{PatchmatchNet}{Figures/dtu_3d/patchmatchnet_75_zoom.png}&\subfigimg[width=0.28\textwidth]{CasMVSNet}{Figures/dtu_3d/cascade_75_zoom.png}&\subfigimg[width=0.28\textwidth]{MVS2D (ours)}{Figures/dtu_3d/mvs2d_75_zoom.png} \end{tabular} \vspace{-0.1in} \caption{Qualitative 3D reconstruction results on DTU dataset. MVS2D\xspace produces more complete reconstruction in texture-less region.} \vspace{-0.2in} \label{fig:dtu_3d} \end{center} \end{figure*} \subsection{Result Analysis}\label{sec4.4} \myparagraph{Comparison on Efficiency.} We compare against both single-view methods~\cite{lee2019big} and multi-view methods~\cite{yao2018mvsnet,yu2020fast,im2019dpsnet,kusupati2019normal}. The inference speed of our method is comparable to single-view methods~\cite{lee2019big} and significantly outperforms other multi-view methods~\cite{yao2018mvsnet,yu2020fast,im2019dpsnet,kusupati2019normal}. Evaluations are done on ScanNet~\cite{scannet17cvpr}. Our method is $48\times$ faster than NAS, $39\times$ faster than DPSNet, $10\times$ faster than MVSNet, and $4\times$ faster than the FastMVSNet. Please refer to the supp. for more details. \myparagraph{Comparison on Depth Estimation.} MVS2D\xspace achieves considerable improvements in the depth prediction accuracy (see Table~\ref{table:scannet}). On ScanNet, our approach outperforms MVSNet by large margins, reducing AbsRel error from \textbf{0.094} to \textbf{0.059}. The improvements are consistent across most other metrics. Remarkably, our approach also outperforms NAS, which uses more parameters and runs 48 times slower. We visualize some depth predictions in Figure~\ref{fig:qualitative}. Our approach yields significant improvements over single-view baselines. Adding multi-view cues improves the AbsRel of ours-mono from \textbf{0.145} to \textbf{0.059} on ScanNet. Since single-view has scale-ambiguity, we further investigate whether our methods will still be favorable when factoring out the scale. The results show that when eliminating the scale, ours-mono$^*$ has AbsRel 0.103, which is still a significant improvement. This means our approach does not simply infer the global scaling factor from multi-view cues. Compared with the single-view model, our model only incurs a $5.8\% $ increase in parameters. Such efficiency will enable multi-view methods to embrace a much larger 2D convolutional network which is not possible before. On other datasets, MVS2D\xspace also performs favorably (see Table~\ref{table:sun3d}). We achieve the AbsRel error of 0.078 on the RGBD dataset, while the next best NAS only achieves 0.131. Although MVS2D\xspace excels at adapting to the scene prior, it is encouraging that it also performs well on the Scenes11 dataset, a synthetic scene with randomly placed objects. We ranked second on the Scenes11 dataset on AbsRel. Please refer to the supp. material for experiments on our generalization ability to novel datasets. \myparagraph{Evaluations on DTU.} We evaluate on DTU dataset following the practice of \cite{wang2021patchmatchnet}. We use 4 reference views and 96 depth samples uniformly placed in the inverse depth space ([$\frac{1}{935.}$,$\frac{1}{425.}$]). The quantitative results can be found in Table \ref{tab:evaluation_dtu}. MVS2D\xspace is the best on overall score and the second-best completeness score. Such performance is encouraging since our method is quite simple: it is just a single-stage procedure without using any multi-stage refinement as commonly used in recent MVS algorithms (\cite{yao2019recurrent,wang2021patchmatchnet, gu2020cascade}). We show some qualitative results of 3D reconstruction on DTU objects in Figure \ref{fig:dtu_3d}. Qualitatively, our reconstruction is typically more complete on flat surface areas. The behavior is reasonable because our approach utilizes strong single-view priors. We also compare the inference speed with the recent SOTA PatchmatchNet\cite{wang2021patchmatchnet}. Our approach yield around 2x speed up as shown in Table \ref{table:dtu-fps}. Lastly, as our method was mainly designed for multi-view depth estimation, we additionally examine the depth evaluation metrics. Since DTU does not have ground truth depth for the test set, we report the depth evaluation results on the validation set. As expected, MVS2D\xspace is better than PatchmatchNet in terms of depth metrics, and the performance gap there is wider than in 3D Reconstruction. The results can be found in Table \ref{table:dtu-depth}. \begin{table}[t] \setlength{\abovecaptionskip}{0.1cm} \centering \footnotesize \begin{widetable}{\columnwidth}{cccc} \toprule Methods & Acc.(mm) & Comp.(mm) & Overall(mm)\\ \hline Camp~\cite{campbell2008using} & 0.835 & 0.554 & 0.695 \\ Furu~\cite{furukawa2009accurate} & 0.613 & 0.941 & 0.777 \\ Tola~\cite{tola2012efficient} & 0.342 & 1.190 & 0.766 \\ Gipuma~\cite{galliani2015massively} & \textbf{0.283} & 0.873 & 0.578 \\ SurfaceNet~\cite{ji2017surfacenet} & 0.450 & 1.040 & 0.745 \\ MVSNet~\cite{yao2018mvsnet} & 0.396 & 0.527 & 0.462\\ R-MVSNet~\cite{yao2019recurrent} & 0.383 & 0.452 & 0.417\\ CIDER~\cite{xu2020learning} & 0.417 & 0.437 & 0.427\\ P-MVSNet~\cite{luo2019p} & 0.406 & 0.434 & 0.420\\ Point-MVSNet~\cite{chen2019point} & 0.342 & 0.411 & 0.376\\ Fast-MVSNet~\cite{yu2020fast} & 0.336 & 0.403 & 0.370\\ CasMVSNet~\cite{gu2020cascade} & 0.325 & 0.385 & 0.355\\ UCS-Net~\cite{cheng2020deep} & 0.338 & 0.349 & \underline{0.344} \\ CVP-MVSNet~\cite{yang2020cost} & \underline{0.296} & 0.406 & 0.351 \\ PatchMatchNet~\cite{wang2021patchmatchnet} & 0.427 & \textbf{0.277} & 0.352\\ MVS2D\xspace(Ours) & 0.394 & \underline{0.290} & \textbf{0.342}\\ \bottomrule \end{widetable} \caption{Quantitative results on the evaluation set of DTU~\cite{aanaes2016large}. We bold the best number and underline the second best number.} \label{tab:evaluation_dtu} \end{table} \begin{table} \centering \resizebox{0.95\linewidth}{!}{% \begin{tabular}{rccc} \toprule \multicolumn{1}{c}{Metric} & FPS$640\times 480$ $\uparrow$ &FPS$1280\times 640$ $\uparrow$ & FPS$1536\times 1152$ $\uparrow$ \\ \hline \multicolumn{1}{c}{PatchmatchNet } &16.5 & 6.30 & 4.57 \\ \multicolumn{1}{c}{MVS2D\xspace(Ours) } &\textbf{36.4} & \textbf{10.9} & \textbf{7.3} \\ \bottomrule \end{tabular} } \vspace{-0.1in} \caption{Speed benchmark on DTU dataset. We show FPS (frame per second) on three input resolutions. We use one source image and 4 reference images. } \vspace{-0.1in} \label{table:dtu-fps} \end{table} \begin{table} \centering \resizebox{0.95\linewidth}{!}{% \begin{tabular}{rcccc} \toprule \multicolumn{1}{c}{Metric} & RMSE(mm)$\downarrow$ & thre@0.2$\uparrow$ & thre@0.5$\uparrow$ & thre@1.0$\uparrow$ \\ \hline \multicolumn{1}{c}{PatchmatchNet } & 32.348 &0.169 & 0.387 & 0.610\\ \multicolumn{1}{c}{MVS2D\xspace(Ours) } & \textbf{14.769} & \textbf{0.238} & \textbf{0.504} & \textbf{0.718} \\ \bottomrule \end{tabular} } \vspace{-0.1in} \caption{Depth evaluation on DTU validation set. We show the root mean square error and the percentage of errors fall below 0.2/0.5/1.0mm thresholds.} \vspace{-0.2in} \label{table:dtu-depth} \end{table} \myparagraph{Comparison on Robustness under Noisy Pose.} As shown in Table \ref{table:scannet}, Ours-robust (multi-scale cues) and Ours (single-scale cues) perform similarly when the input poses are accurate. However, as shown in Table \ref{table:perturb}, multi-scale aggregation is preferred when the input poses are noisy. It suggests that when having inaccurate training data, it is necessary to incorporate multi-scale cues, though at a cost of increased computations (as shown in Table~\ref{table:speed}). \begin{table}[!h] \centering \resizebox{0.95\linewidth}{!}{% \begin{tabular}{rcccc} \toprule \multicolumn{1}{c}{Metric} & MVSNet& DPSNet & Ours & Ours-robust \\ \hline \multicolumn{1}{c}{AbsRel $\downarrow$} & 0.094 &0.094 & 0.059 & \textbf{0.059}\\ \multicolumn{1}{c}{AbsRel (p) $\downarrow$} & 0.113 &0.126 & 0.073 & \textbf{0.070} \\ \multicolumn{1}{c}{$\Delta$ $\downarrow$} & 0.019 &0.032 & 0.014 & \textbf{0.011} \\ \hline \multicolumn{1}{c}{$\delta<1.25$ $\uparrow$} & 0.897 & 0.871 & 0.983 & \textbf{0.965} \\ \multicolumn{1}{c}{$\delta<1.25$ (p) $\uparrow$} & 0.851 & 0.807 & 0.947 & \textbf{0.952}\\ \multicolumn{1}{c}{$\Delta$ $\downarrow$} & 0.046 & 0.064 & 0.016 & \textbf{0.013} \\ \bottomrule \end{tabular} } \caption{Different methods' performance under noisy input poses on ScanNet~\cite{scannet17cvpr}. We notice that most methods suffer from significant performance drops. Our method with multi-scale epipolar aggregation shows notable robustness.} \label{table:perturb} \end{table} \myparagraph{Ablation Study on Depth Encoding.} The ablation study of our depth code design can be found in Table \ref{table:embed}. We tested four code types. `Uniform' serves as a sanity check, where we use the same code vector for all depth hypotheses. In other words, the network does not extract useful information from the reference images. `Linear' improves on uniform encoding by scaling a base code vector with the corresponding depth value. `Cosine' codes are identical to the one used in \cite{vaswani2017attention}. `Learned` codes are optimized end-to-end. We can see that learning the codes end-to-end leads to noticeable performance gains. One explanation is that these learned codes can adapt to the single-view feature representations of the source image. \begin{table}[!h] \centering \resizebox{0.95\linewidth}{!}{% \begin{tabular}{rcccc} \toprule \multicolumn{1}{c}{Metric} & Uniform & Linear & Cosine & Learned \\ \hline \multicolumn{1}{c}{AbsRel $\downarrow$} & 0.139 & 0.128& 0.064 & \textbf{0.059} \\ \multicolumn{1}{c}{$\delta < 1.25$ $\uparrow$} & 0.815 & 0.840 & 0.961 & \textbf{0.964} \\ \multicolumn{1}{c}{RMSE $\downarrow$} & 0.293 &0.283 & 0.166& \textbf{0.156 } \\ \bottomrule \end{tabular} } \caption{Ablation study on different depth encodings. We can see that jointly training depth encodings gives the best performance. } \label{table:embed} \vspace{-0.1in} \end{table} \section{Conclusions and Limitations} \noindent\textbf{Conclusions.} We proposed a simple yet effective method for multi-view stereo. The core of our method is to integrate single-view and multi-view cues during the prediction jointly. Such a design not only improves the performance but also has the appealing factor of being efficient. Furthermore, we have demonstrated the trade-off between input pose accuracy and network complexity. When the input pose is exact, we can leverage minimum additional computation to inject more multi-view information through the epipolar attention. \noindent\textbf{Limitations.} One limitation of our approach is that the network is trained in a way that adapted to data distribution well, which might makes it less generalizable to out-of-distribution testing data. In the future, we propose to address this issue by developing robust training losses. Another limitation is that the proposed attention mechanism does not explicitly model the consistency between different pixels on the same epipolar line. We plan to address this issue by developing novel attention mechanisms to explicitly enforce those constraints. \section{Details of Experiments} \subsection{Model Details} We use ResNet-18 (with the fully connected layer and pooling layer removed) as the building block of our networks $\set{G}$ and $\set{F}$. $\set{F}$ additionally includes two up-sampling and convolution layer that output a feature map $F_0 \in \mathcal{R}^{\frac{h}{4}\times \frac{w}{4}\times c}$, where $h$ and $w$ are the input image's height and width respectively. $F_0$ is further feed into a small network with 3 convolutions and 3 de-convolutions to recover a depth probability volume $F_1\in \set{R}^{\frac{h}{4}\times \frac{w}{4}\times k'}$. $F_1$ is finally decoded into a depth map $d_{\frac{1}{4}}\in \mathcal{R}^{\frac{h}{4}\times \frac{w}{4}}$ using the soft-argmax operation as in MVSNet~\cite{yao2018mvsnet}. To supervise the network at full resolution instead of $\frac{1}{4}$ resolution, we further upsample $d_{\frac{1}{4}}$ to the full resolution via nearest interpolation, and add a predicted residual to it to get final prediction $d \in \mathcal{R}^{h\times w}$. Please refer to our released code for more details. \begin{table*}[!ht] \centering \resizebox{0.9\textwidth}{!}{% \begin{tabular}{r|ccccc|ccc} \toprule \multicolumn{1}{c|}{Method} & AbsRel $\downarrow$ & AbsDiff $\downarrow$& SqRel $\downarrow$& RMSE $\downarrow$ & RMSELog $\downarrow$& $\delta<1.25$ $\uparrow$& $\delta<1.25^2$ $\uparrow$& $\delta<1.25^3$ $\uparrow$ \\ \hline COLMAP & 0.384 & 0.843 & 1.26 & 1.480 & 0.500 & 0.482 & 0.663 & 0.840\\ DeMoN &0.311& 1.330& 19.970& 2.607& 0.247& 0.641 &0.902 &0.967\\ DeepMVS &0.231 &0.663 &0.615 & 1.149 &0.302 & 0.674 & 0.887 &0.941\\ DPSNet &0.081 &0.201 &0.097 &0.442 &0.160 &0.885& 0.945& 0.973\\ NAS &\textbf{0.068} &\textbf{0.168}& \textbf{0.056}& 0.375& 0.142 &\textbf{0.905} &0.964 &0.988\\ \hline Ours-robust & 0.100 & 0.231 & 0.057 & \textbf{0.313} & \textbf{0.140} & 0.895 & \textbf{0.966} & \textbf{0.991}\\ Ours & 0.108 & 0.271 & 0.130 & 0.513 & 0.184 & 0.860 & 0.939 & 0.973 \\ \bottomrule \end{tabular} } \caption{Depth evaluation results on the MVS dataset (trained on RGBD, SUN3D, and Scenes11). Please see Sec.~\ref{supp:sec2.4} for discussion.} \label{table:generalization1} \end{table*} \begin{table*}[!ht] \centering \resizebox{0.9\textwidth}{!}{% \begin{tabular}{r|ccccc|ccc} \toprule \multicolumn{1}{c|}{Method} & AbsRel $\downarrow$ & SqRel $\downarrow$& log10 $\downarrow$& RMSE $\downarrow$ & RMSELog $\downarrow$& $\delta<1.25$ $\uparrow$& $\delta<1.25^2$ $\uparrow$& $\delta<1.25^3$ $\uparrow$ \\ \hline MVSNet & 0.154 & 0.125 & 0.067 & 0.478 & 0.212 & 0.779 & 0.927 & 0.973 \\ NAS & 0.134 & 0.094 & 0.064 & 0.434 & 0.190 & 0.789 & 0.932 & 0.979 \\ \hline Ours & \textbf{0.113} & \textbf{0.062} & \textbf{0.049} & \textbf{0.332} & \textbf{0.149} & \textbf{0.871} & \textbf{0.968} & \textbf{0.993} \\ Ours-robust & 0.115 & 0.066 & 0.052 & 0.354 & 0.158 & 0.862 & 0.960 & 0.990\\ \bottomrule \end{tabular} } \caption{Depth evaluation results on the SUN3D dataset (trained on ScanNet). Please see Sec.~\ref{supp:sec2.4} for discussion. } \label{table:generalization2} \end{table*} \subsection{Training Details} We use 4 NVIDIA DGX-V100 GPUs with 32GB memory each to conduct following experiments. \noindent\textbf{ScanNet.} We train for 30 epochs with a batch size 16, and reduce the learning rate by 10 at epoch 25 and 28. The input image size is $640\times 480$. During training, we sample depth hypothesis uniformly in inverse depth space \begin{equation} d_i = 1/((1 - \frac{i}{k})\frac{1}{d_{\text{min}}} + \frac{i}{k} \frac{1}{d_{\text{max}}}), \end{equation} where $k$ is the number of depth hypothesis. We set $k=32$, $d_{\text{min}} = 0.3$, and $d_{\text{max}} = 10.1$. We trained several baselines on ScanNet dataset. Specifically, we trained MVSNet~\cite{yao2018mvsnet} for 500k iterations. We trained FastMVSNet~\cite{yu2020fast} for a total 30 epochs, with the last 10 epochs optimizing the GaussNewton layer. We trained NAS~\cite{kusupati2019normal} for 20 epochs for initialization, and 10 epoch including the normal consistency module. We trained PatchmatchNet~\cite{wang2021patchmatchnet} for 30 epochs. We trained DPSNet~\cite{im2019dpsnet} for 20 epochs. We trained Bts~\cite{lee2019big} for 20 epochs. The training for MVSNet/DPSNet costs around 3 days. The training for NAS costs around 5 days. The training time of FastMVSNet/PatchmatchNet costs around 1 day. \noindent\textbf{SUN3D, RGBD, and Scenes11.} We follow similar setup as the training of ScanNet. During training, we sample depth hypothesis uniformly in inverse depth space: \begin{equation} d_i = 1/((1 - \frac{i}{k})\frac{1}{d_{\text{min}}} + \frac{i}{k} \frac{1}{d_{\text{max}}}), \end{equation} where $d_{\text{min}} = 0.5$, $d_{\text{max}} = 32.0$. \noindent\textbf{DTU.} One modification we made on training DTU dataset is that we add a confidence prediction. The added confidence will be used to filter out unconfident predictions during the final 3D reconstruction. We adopt following loss~\cite{kendall2017uncertainties}: \begin{equation} \mathcal{L} = \frac{|\hat{d} - d_{\text{gt}}|}{\hat{\sigma} } + \log(\hat{\sigma}), \end{equation} where $\hat{\sigma}$ is the predicted confidence map. We implement this confidence prediction by adding a separate head in the final output. Such confidence prediction is trained along with depth prediction without the need for explicit supervision. During the test time, we set a threshold $\tau$ for $\hat{\sigma}$ and prune all predictions $\hat{d_i}$ whose corresponding $\hat{\sigma_i}$ are larger than $\tau$. We train for 80 epochs with a batch size 8, and reduce the learning rate by ten at epoch 40 and 70. The input image size is $1536\times 1152$. The training takes around 5 days. We follow the practice of the robust training scheme \cite{wang2021patchmatchnet} and use randomly chosen 4 reference images for each source image during training. We use $N=96$ depth hypothesis uniformly sampled in the inverse depth space ($d_i = 1/((1 - \frac{i}{k})\frac{1}{d_{\text{min}}} + \frac{i}{k} \frac{1}{d_{\text{max}}})$. We use $d_{\text{min}} = 425.0$, $d_{\text{max}} = 935.0$. During testing, we use top 4 reference images(ranked by a heuristic criterior introduced in ~\cite{yao2018mvsnet}), which is the same as previous approaches~\cite{yao2018mvsnet,chen2019point,wang2021patchmatchnet}. We fuse the depth maps into final 3D model using the fusion code provided in ~\cite{wang2021patchmatchnet}. The qualitative results on DTU test set can be found in Figure \ref{fig:visualization:dtu}. \subsection{Speed Comparison} We use one V100 GPU to conduct the speed benchmark for all methods. to We feed each method an input image of size 640x480. For methods that do not produce full-resolution depth maps (i.e. MVSNet~\cite{yao2018mvsnet}), we further up-sampled them to the full resolution using nearest-neighbor interpolation. The FPS numbers are averaged over 500 random inputs for each method. \subsection{Pose Corruption} We use the following procedure to generate perturbations for input poses. Assume the ground truth relative pose is $T=[R\, |\, t]$ between the source image and a certain reference image. Firstly, we sample $N$ points $\{p_k\}_{k=1}^{N}$ on corresponding ground-truth 3D point cloud of the source image. We use $N=10$ in our experiments. Then, we project those $N$ points into reference image using ground truth relative pose $T$ and camera intrinsics $\mathcal{K}$ to get $\{\overline{p}_k\}_{k=1}^{N}$. We then perturb $\{\overline{p}_k\}_{k=1}^{N}$ by adding noise from a uniform distribution whose maximum value is 10 pixels. We solve a $\text{PnP}$ problem~\cite{andrew2001multiple} using $\{p_k\}_{k=1}^{N}$ and the perturbed pixels $\{\overline{p}_k\}_{k=1}^{N}$ to get the corrupted $\overline{T}=[\overline{R}\, |\, \overline{t}]$. We accept the perturbed $\overline{T}$ if the average pixel offset over the source image is less than 10 pixels. Otherwise, we set $\overline{T} = T$. We pre-compute all perturbations for all image pairs. Figure~\ref{fig:perturb_hist} shows the statistics of pose perturbations. Specifically, we plot the histogram of $\Delta R = \arccos(\frac{\Tr(\overline{R}R^{-1})-1}{2})$, and $\Delta t=\norm{\overline{t} - t}_2$. \begin{figure}[!h] \includegraphics[trim=30 0 30 20,clip,width=0.49\linewidth]{Figures/pose_dR_hist.png} \includegraphics[trim=30 0 30 20,clip,width=0.49\linewidth]{Figures/pose_dt_hist.png} \caption{Pose corruption statistics. Left: histogram of rotation perturbation. Right: histogram of translation perturbation.} \label{fig:perturb_hist} \vspace{-0.1in} \end{figure} \section{Additional Studies} \subsection{Generalization Ability} \label{supp:sec2.4} We did two experiments to evaluate the generalization ability of all methods to unseen datasets. Following the experimental setups of DeMoN and NAS, we use the model trained on SUN3D/RGBD/Scenes11 and test on the MVS dataset, which is an outdoor dataset and the data distributions are very different from the training set. The results can be found in Table~\ref{table:generalization1}. Our methods perform better than COLMAP, DeMoN and DeepMVS, although they still fall behind NAS under some metrics such as $AbsRel$. Such a result is reasonable since our approach is better adapted to the training distribution, which will lead to performance drop on heavily out-of-distribution test data. To further examine each method's performance on unseen datasets whose input data statistics are similar to those in the training sets, we further test models trained using ScanNet on SUN3D test sets (see Table~\ref{table:generalization2}). The input data of SUN3D ScanNet are all indoor scenes. We can see that our methods still perform favorably among other methods. \subsection{Ablation Study on Mask Encoding} \label{supp:sec2.5} To study the benefits of mask encoding, we further experiment with \textbf{Ours-nomask} which removes the mask encoding in Eq 5. The results on ScanNet can be found in Table \ref{table:ablation_mask}. Remove mask encoding (\textbf{Ours-nomask}) leads to worse results than \textbf{Ours}. Such behavior is reasonable since mask encoding provides an easy way for the network to distinguish the valid and invalid interpolation, thus facilitate the training. \begin{table}[!ht] \centering \resizebox{0.46\textwidth}{!}{% \begin{tabular}{r|cccc} \toprule \multicolumn{1}{c|}{Method} & AbsRel $\downarrow$ & $\delta<1.25$ $\uparrow$& $thre@0.2$ $\uparrow$ & $thre@0.5$ $\uparrow$\\ \hline Ours-nomask & 0.0605 & 0.9635 & 0.8543 & 0.9719\\ Ours & \textbf{0.0597} & \textbf{0.9640} & \textbf{0.8585} & \textbf{0.9735}\\ \bottomrule \end{tabular} } \caption{Ablation study on mask encoding. $thre@0.2$/$thre@0.5$ measures the percentage of pixel that has absolute depth error less than 0.2m/0.5m respectively. Removing mask encoding hurts the performance, especially for \textit{AbsRel}} \label{table:ablation_mask} \end{table} \subsection{More Qualitative Results} We show additional visualizations of depth predictions in Figure~\ref{fig:qualitative1}~\&~\ref{fig:qualitative2}. Our method produces higher quality depth estimations compared to other MVS methods and performs better than one of the state-of-the-art single-view depth estimation method Bts. Additionally, we show more qualitative comparisons on 3D reconstruction in Figure \ref{fig:supp_3drecon}. Our method generates comparable or even better visual results with other methods that require expensive 3D convolutions. We also show the reconstruction result on DTU test set in Figure \ref{fig:visualization:dtu}. \newpage \begin{figure*} \newcommand{\TT}[1]{\raisebox{-0.5\height}{#1}} \setlength{\tabcolsep}{1pt} \footnotesize \def0.28\textwidth{0.14\textwidth} \begin{tabular}{ccccccc} Bts & MVSNet & FastMVSNet & DPSNet & NAS & MVS2D (Ours) & G.T. \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0011_00-0080_0079_0081/bts.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0011_00-0080_0079_0081/mvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0011_00-0080_0079_0081/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0011_00-0080_0079_0081/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0011_00-0080_0079_0081/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0011_00-0080_0079_0081/ours.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0011_00-0080_0079_0081/gt.png}} \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0025_01-0045_0044_0046/bts.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0025_01-0045_0044_0046/mvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0025_01-0045_0044_0046/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0025_01-0045_0044_0046/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0025_01-0045_0044_0046/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0025_01-0045_0044_0046/ours.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0025_01-0045_0044_0046/gt.png}} \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0030_01-0075_0074_0076/bts.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0030_01-0075_0074_0076/mvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0030_01-0075_0074_0076/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0030_01-0075_0074_0076/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0030_01-0075_0074_0076/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0030_01-0075_0074_0076/ours.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0030_01-0075_0074_0076/gt.png}} \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0046_01-0011_0010_0012/bts.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0046_01-0011_0010_0012/mvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0046_01-0011_0010_0012/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0046_01-0011_0010_0012/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0046_01-0011_0010_0012/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0046_01-0011_0010_0012/ours.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0046_01-0011_0010_0012/gt.png}} \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0050_00-0001_0000_0002/bts.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0050_00-0001_0000_0002/mvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0050_00-0001_0000_0002/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0050_00-0001_0000_0002/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0050_00-0001_0000_0002/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0050_00-0001_0000_0002/ours.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0050_00-0001_0000_0002/gt.png}} \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0064_01-0033_0032_0034/bts.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0064_01-0033_0032_0034/mvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0064_01-0033_0032_0034/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0064_01-0033_0032_0034/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0064_01-0033_0032_0034/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0064_01-0033_0032_0034/ours.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0064_01-0033_0032_0034/gt.png}} \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0187_01-0090_0089_0091/bts.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0187_01-0090_0089_0091/mvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0187_01-0090_0089_0091/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0187_01-0090_0089_0091/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0187_01-0090_0089_0091/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0187_01-0090_0089_0091/ours.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0187_01-0090_0089_0091/gt.png}} \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0088_00-0007_0006_0008/bts.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0088_00-0007_0006_0008/mvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0088_00-0007_0006_0008/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0088_00-0007_0006_0008/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0088_00-0007_0006_0008/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0088_00-0007_0006_0008/ours.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0088_00-0007_0006_0008/gt.png}} \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0100_02-0039_0038_0040/bts.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0100_02-0039_0038_0040/mvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0100_02-0039_0038_0040/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0100_02-0039_0038_0040/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0100_02-0039_0038_0040/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0100_02-0039_0038_0040/ours.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0100_02-0039_0038_0040/gt.png}} \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0050_01-0128_0127_0129/bts.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0050_01-0128_0127_0129/mvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0050_01-0128_0127_0129/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0050_01-0128_0127_0129/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0050_01-0128_0127_0129/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0050_01-0128_0127_0129/ours.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0050_01-0128_0127_0129/gt.png}} \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0064_00-0014_0013_0015/bts.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0064_00-0014_0013_0015/mvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0064_00-0014_0013_0015/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0064_00-0014_0013_0015/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0064_00-0014_0013_0015/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0064_00-0014_0013_0015/ours.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0064_00-0014_0013_0015/gt.png}} \\ \end{tabular} \caption{[1/2] Qualitative results on depth prediction. Each row corresponds to one test example. The region without ground truth depth labels is colored white in the last column. Our prediction outperforms both the single-view depth estimation method and other multi-view methods. } \label{fig:qualitative1} \end{figure*} \begin{figure*} \newcommand{\TT}[1]{\raisebox{-0.5\height}{#1}} \setlength{\tabcolsep}{1pt} \footnotesize \def0.28\textwidth{0.14\textwidth} \begin{tabular}{ccccccc} Bts & MVSNet & FastMVSNet & DPSNet & NAS & MVS2D (Ours) & G.T. \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0146_01-0018_0017_0019/bts.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0146_01-0018_0017_0019/mvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0146_01-0018_0017_0019/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0146_01-0018_0017_0019/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0146_01-0018_0017_0019/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0146_01-0018_0017_0019/ours.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0146_01-0018_0017_0019/gt.png}} \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0164_00-0075_0074_0076/bts.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0164_00-0075_0074_0076/mvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0164_00-0075_0074_0076/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0164_00-0075_0074_0076/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0164_00-0075_0074_0076/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0164_00-0075_0074_0076/ours.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0164_00-0075_0074_0076/gt.png}} \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0187_00-0004_0003_0005/bts.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0187_00-0004_0003_0005/mvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0187_00-0004_0003_0005/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0187_00-0004_0003_0005/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0187_00-0004_0003_0005/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0187_00-0004_0003_0005/ours.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0187_00-0004_0003_0005/gt.png}} \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0193_01-0003_0002_0004/bts.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0193_01-0003_0002_0004/mvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0193_01-0003_0002_0004/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0193_01-0003_0002_0004/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0193_01-0003_0002_0004/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0193_01-0003_0002_0004/ours.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0193_01-0003_0002_0004/gt.png}} \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0187_00-0084_0083_0085/bts.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0187_00-0084_0083_0085/mvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0187_00-0084_0083_0085/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0187_00-0084_0083_0085/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0187_00-0084_0083_0085/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0187_00-0084_0083_0085/ours.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0187_00-0084_0083_0085/gt.png}} \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0203_00-0019_0018_0020/bts.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0203_00-0019_0018_0020/mvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0203_00-0019_0018_0020/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0203_00-0019_0018_0020/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0203_00-0019_0018_0020/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0203_00-0019_0018_0020/ours.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0203_00-0019_0018_0020/gt.png}} \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0207_00-0010_0009_0011/bts.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0207_00-0010_0009_0011/mvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0207_00-0010_0009_0011/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0207_00-0010_0009_0011/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0207_00-0010_0009_0011/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0207_00-0010_0009_0011/ours.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0207_00-0010_0009_0011/gt.png}} \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0208_00-0003_0002_0004/bts.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0208_00-0003_0002_0004/mvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0208_00-0003_0002_0004/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0208_00-0003_0002_0004/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0208_00-0003_0002_0004/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0208_00-0003_0002_0004/ours.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0208_00-0003_0002_0004/gt.png}} \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0222_00-0012_0011_0013/bts.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0222_00-0012_0011_0013/mvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0222_00-0012_0011_0013/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0222_00-0012_0011_0013/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0222_00-0012_0011_0013/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0222_00-0012_0011_0013/ours.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0222_00-0012_0011_0013/gt.png}} \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0077_01-0046_0045_0044/bts.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0077_01-0046_0045_0044/mvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0077_01-0046_0045_0044/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0077_01-0046_0045_0044/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0077_01-0046_0045_0044/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0077_01-0046_0045_0044/ours.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0077_01-0046_0045_0044/gt.png}} \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0095_01-0025_0024_0026/bts.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0095_01-0025_0024_0026/mvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0095_01-0025_0024_0026/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0095_01-0025_0024_0026/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0095_01-0025_0024_0026/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0095_01-0025_0024_0026/ours.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure3_jet/scene0095_01-0025_0024_0026/gt.png}} \\ \end{tabular} \caption{[2/2] Qualitative results on depth prediction. Each row corresponds to one test example. The region without ground truth depth labels is colored white in the last column. Our prediction outperforms both the single-view depth estimation method and other multi-view methods. } \label{fig:qualitative2} \end{figure*} \begin{figure*} \newcommand{\TT}[1]{\raisebox{-0.5\height}{#1}} \setlength{\tabcolsep}{1pt} \footnotesize \def0.28\textwidth{0.2\textwidth} \begin{tabular}{ccccc} MVSNet &FastMVSNet & DPSNet & NAS & MVS2D (Ours) \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0316_00/mvsnet.png}} &\TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0316_00/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{.igures/Figure4_new/scene0316_00/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0316_00/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0316_00/ours.png}} \\ \vspace{0.2in} \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0277_00/mvsnet.png}} &\TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0277_00/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0277_00/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0277_00/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0277_00/ours.png}} \\ \vspace{0.2in} \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0149_00/mvsnet.png}} &\TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0149_00/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0149_00/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0149_00/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0149_00/ours.png}} \\ \vspace{0.2in} \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0146_00/mvsnet.png}} &\TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0146_00/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0146_00/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0146_00/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0146_00/ours.png}} \\ \vspace{0.2in} \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0300_00/mvsnet.png}} &\TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0300_00/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0300_00/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0300_00/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0300_00/ours.png}} \\ \vspace{0.2in} \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0084_00/mvsnet.png}} &\TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0084_00/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0084_00/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0084_00/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0084_00/ours.png}} \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0488_00/mvsnet.png}} &\TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0488_00/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0488_00/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0488_00/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0488_00/ours.png}} \\ \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0338_00/mvsnet.png}} &\TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0338_00/fastmvsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0338_00/dpsnet.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0338_00/nas.png}} & \TT{\includegraphics[width=0.28\textwidth]{Figures/Figure4_new/scene0338_00/ours.png}} \\ \end{tabular} \caption{Qualitative scene reconstruction results on ScanNet. Our method yields smoother outputs than other baselines. We zoom in parts of a scene (\textcolor{red}{red} box) and show at the corner (\textcolor{blue}{blue} box) to highlight the differences. Best viewed in PDF.} \label{fig:supp_3drecon} \end{figure*} \begingroup \setlength{\tabcolsep}{1.0pt} \begin{figure*}[!ht] \centering \hspace{-1em} \def0.28\textwidth{0.28\textwidth} \def0.25\textheight{0.25\textheight} \def{} \begin{tabular}{ccc} {\includegraphics[trim={100 0 0 0},clip, height=0.25\textheight, width=0.28\textwidth, keepaspectratio]{Figures/dtu_3d/mvs2d_1.png}} & {\includegraphics[trim={200 100 200 0},clip,height=0.25\textheight, width=0.28\textwidth, keepaspectratio]{Figures/dtu_3d/mvs2d_4.png}} & {\includegraphics[trim={400 100 400 200},clip,height=0.25\textheight, width=0.28\textwidth, keepaspectratio]{Figures/dtu_3d/mvs2d_13.png}} \\ {\includegraphics[trim={400 100 550 200},clip,height=0.25\textheight, width=0.28\textwidth, keepaspectratio]{Figures/dtu_3d/mvs2d_10.png}} & {\includegraphics[trim={600 200 300 50},clip,height=0.25\textheight, width=0.28\textwidth, keepaspectratio]{Figures/dtu_3d/mvs2d_11.png}} & {\includegraphics[trim={300 100 200 0},clip,height=0.25\textheight, width=0.28\textwidth, keepaspectratio]{Figures/dtu_3d/mvs2d_114.png}} \\ {\includegraphics[trim={100 00 300 00},clip,height=0.25\textheight, width=0.28\textwidth, keepaspectratio]{Figures/dtu_3d/mvs2d_9.png}} & {\includegraphics[trim={100 0 100 50},clip,height=0.25\textheight, width=0.28\textwidth, keepaspectratio]{Figures/dtu_3d/mvs2d_15.png}} & {\includegraphics[trim={00 0 00 0},clip,height=0.25\textheight, width=0.28\textwidth, keepaspectratio]{Figures/dtu_3d/mvs2d_23.png}} \\ {\includegraphics[trim={400 0 400 0},clip,height=0.25\textheight, width=0.28\textwidth, keepaspectratio]{Figures/dtu_3d/mvs2d_24.png}} & {\includegraphics[trim={650 200 450 0},clip,height=0.25\textheight, width=0.28\textwidth,keepaspectratio]{Figures/dtu_3d/mvs2d_118.png}} & {\includegraphics[trim={600 100 500 50},clip,height=0.25\textheight, width=0.28\textwidth,keepaspectratio]{Figures/dtu_3d/mvs2d_33.png}} \\ {\includegraphics[trim={100 0 100 100},clip,height=0.25\textheight, width=0.28\textwidth, keepaspectratio]{Figures/dtu_3d/mvs2d_34.png}} & {\includegraphics[trim={500 0 100 50},clip,height=0.25\textheight, width=0.28\textwidth, keepaspectratio]{Figures/dtu_3d/mvs2d_110.png}} & {\includegraphics[trim={100 0 100 50},clip,height=0.25\textheight, width=0.28\textwidth, keepaspectratio]{Figures/dtu_3d/mvs2d_32.png}}\\ \end{tabular} \vspace{-5pt} \caption{Qualitative 3D reconstruction results on the DTU test set. We use 4 reference views during the evaluation. The results are obtained from fusing multi-view depth prediction using the tool provided by PatchMatchNet~\cite{wang2021patchmatchnet}.} \label{fig:visualization:dtu}\vspace{-1mm} \end{figure*} \endgroup \section{Introduction} \label{sec:intro} Please follow the steps outlined below when submitting your manuscript to the IEEE Computer Society Press. This style guide now has several important modifications (for example, you are no longer warned against the use of sticky tape to attach your artwork to the paper), so all authors should read this new version. \subsection{Language} All manuscripts must be in English. \subsection{Dual submission} Please refer to the author guidelines on the CVPR\ 2022\ web page for a discussion of the policy on dual submissions. \subsection{Paper length} Papers, excluding the references section, must be no longer than eight pages in length. The references section will not be included in the page count, and there is no limit on the length of the references section. For example, a paper of eight pages with two pages of references would have a total length of 10 pages. {\bf There will be no extra page charges for CVPR\ 2022.} Overlength papers will simply not be reviewed. This includes papers where the margins and formatting are deemed to have been significantly altered from those laid down by this style guide. Note that this \LaTeX\ guide already sets figure captions and references in a smaller font. The reason such papers will not be reviewed is that there is no provision for supervised revisions of manuscripts. The reviewing process cannot determine the suitability of the paper for presentation in eight pages if it is reviewed in eleven. \subsection{The ruler} The \LaTeX\ style defines a printed ruler which should be present in the version submitted for review. The ruler is provided in order that reviewers may comment on particular lines in the paper without circumlocution. If you are preparing a document using a non-\LaTeX\ document preparation system, please arrange for an equivalent ruler to appear on the final output pages. The presence or absence of the ruler should not change the appearance of any other content on the page. The camera-ready copy should not contain a ruler. (\LaTeX\ users may use options of cvpr.sty to switch between different versions.) Reviewers: note that the ruler measurements do not align well with lines in the paper --- this turns out to be very difficult to do well when the paper contains many figures and equations, and, when done, looks ugly. Just use fractional references (\eg, this line is $087.5$), although in most cases one would expect that the approximate location will be adequate. \subsection{Paper ID} Make sure that the Paper ID from the submission system is visible in the version submitted for review (replacing the ``*****'' you see in this document). If you are using the \LaTeX\ template, \textbf{make sure to update paper ID in the appropriate place in the tex file}. \subsection{Mathematics} Please number all of your sections and displayed equations as in these examples: \begin{equation} E = m\cdot c^2 \label{eq:important} \end{equation} and \begin{equation} v = a\cdot t. \label{eq:also-important} \end{equation} It is important for readers to be able to refer to any particular equation. Just because you did not refer to it in the text does not mean some future reader might not need to refer to it. It is cumbersome to have to use circumlocutions like ``the equation second from the top of page 3 column 1''. (Note that the ruler will not be present in the final copy, so is not an alternative to equation numbers). All authors will benefit from reading Mermin's description of how to write mathematics: \url{http://www.pamitc.org/documents/mermin.pdf}. \subsection{Blind review} Many authors misunderstand the concept of anonymizing for blind review. Blind review does not mean that one must remove citations to one's own work---in fact it is often impossible to review a paper unless the previous citations are known and available. Blind review means that you do not use the words ``my'' or ``our'' when citing previous work. That is all. (But see below for tech reports.) Saying ``this builds on the work of Lucy Smith [1]'' does not say that you are Lucy Smith; it says that you are building on her work. If you are Smith and Jones, do not say ``as we show in [7]'', say ``as Smith and Jones show in [7]'' and at the end of the paper, include reference 7 as you would any other cited work. An example of a bad paper just asking to be rejected: \begin{quote} \begin{center} An analysis of the frobnicatable foo filter. \end{center} In this paper we present a performance analysis of our previous paper [1], and show it to be inferior to all previously known methods. Why the previous paper was accepted without this analysis is beyond me. [1] Removed for blind review \end{quote} An example of an acceptable paper: \begin{quote} \begin{center} An analysis of the frobnicatable foo filter. \end{center} In this paper we present a performance analysis of the paper of Smith \etal [1], and show it to be inferior to all previously known methods. Why the previous paper was accepted without this analysis is beyond me. [1] Smith, L and Jones, C. ``The frobnicatable foo filter, a fundamental contribution to human knowledge''. Nature 381(12), 1-213. \end{quote} If you are making a submission to another conference at the same time, which covers similar or overlapping material, you may need to refer to that submission in order to explain the differences, just as you would if you had previously published related work. In such cases, include the anonymized parallel submission~\cite{Authors14} as supplemental material and cite it as \begin{quote} [1] Authors. ``The frobnicatable foo filter'', F\&G 2014 Submission ID 324, Supplied as supplemental material {\tt fg324.pdf}. \end{quote} Finally, you may feel you need to tell the reader that more details can be found elsewhere, and refer them to a technical report. For conference submissions, the paper must stand on its own, and not {\em require} the reviewer to go to a tech report for further details. Thus, you may say in the body of the paper ``further details may be found in~\cite{Authors14b}''. Then submit the tech report as supplemental material. Again, you may not assume the reviewers will read this material. Sometimes your paper is about a problem which you tested using a tool that is widely known to be restricted to a single institution. For example, let's say it's 1969, you have solved a key problem on the Apollo lander, and you believe that the CVPR70 audience would like to hear about your solution. The work is a development of your celebrated 1968 paper entitled ``Zero-g frobnication: How being the only people in the world with access to the Apollo lander source code makes us a wow at parties'', by Zeus \etal. You can handle this paper like any other. Do not write ``We show how to improve our previous work [Anonymous, 1968]. This time we tested the algorithm on a lunar lander [name of lander removed for blind review]''. That would be silly, and would immediately identify the authors. Instead write the following: \begin{quotation} \noindent We describe a system for zero-g frobnication. This system is new because it handles the following cases: A, B. Previous systems [Zeus et al. 1968] did not handle case B properly. Ours handles it by including a foo term in the bar integral. ... The proposed system was integrated with the Apollo lunar lander, and went all the way to the moon, don't you know. It displayed the following behaviours, which show how well we solved cases A and B: ... \end{quotation} As you can see, the above text follows standard scientific convention, reads better than the first version, and does not explicitly name you as the authors. A reviewer might think it likely that the new paper was written by Zeus \etal, but cannot make any decision based on that guess. He or she would have to be sure that no other authors could have been contracted to solve problem B. \medskip \noindent FAQ\medskip\\ {\bf Q:} Are acknowledgements OK?\\ {\bf A:} No. Leave them for the final copy.\medskip\\ {\bf Q:} How do I cite my results reported in open challenges? {\bf A:} To conform with the double-blind review policy, you can report results of other challenge participants together with your results in your paper. For your results, however, you should not identify yourself and should not mention your participation in the challenge. Instead present your results referring to the method proposed in your paper and draw conclusions based on the experimental comparison to other results.\medskip\\ \begin{figure}[t] \centering {\rule{0pt}{2in} \rule{0.9\linewidth}{0pt}} \caption{Example of caption. It is set in Roman so that mathematics (always set in Roman: $B \sin A = A \sin B$) may be included without an ugly clash.} \label{fig:onecol} \end{figure} \subsection{Miscellaneous} \noindent Compare the following:\\ \begin{tabular}{ll} \verb'$conf_a$' & $conf_a$ \\ \verb'$\mathit{conf}_a$' & $\mathit{conf}_a$ \end{tabular}\\ See The \TeX book, p165. The space after \eg, meaning ``for example'', should not be a sentence-ending space. So \eg is correct, {\em e.g.} is not. The provided \verb'\eg' macro takes care of this. When citing a multi-author paper, you may save space by using ``et alia'', shortened to ``\etal'' (not ``{\em et.\ al.}'' as ``{\em et}'' is a complete word). If you use the \verb'\etal' macro provided, then you need not worry about double periods when used at the end of a sentence as in Alpher \etal. However, use it only when there are three or more authors. Thus, the following is correct: ``Frobnication has been trendy lately. It was introduced by Alpher~\cite{Alpher02}, and subsequently developed by Alpher and Fotheringham-Smythe~\cite{Alpher03}, and Alpher \etal~\cite{Alpher04}.'' This is incorrect: ``... subsequently developed by Alpher \etal~\cite{Alpher03} ...'' because reference~\cite{Alpher03} has just two authors. \begin{figure*} \centering \begin{subfigure}{0.68\linewidth} {\rule{0pt}{2in} \rule{.9\linewidth}{0pt}} \caption{An example of a subfigure.} \label{fig:short-a} \end{subfigure} \hfill \begin{subfigure}{0.28\linewidth} {\rule{0pt}{2in} \rule{.9\linewidth}{0pt}} \caption{Another example of a subfigure.} \label{fig:short-b} \end{subfigure} \caption{Example of a short caption, which should be centered.} \label{fig:short} \end{figure*} \section{Formatting your paper} \label{sec:formatting} All text must be in a two-column format. The total allowable size of the text area is $6\frac78$ inches (17.46 cm) wide by $8\frac78$ inches (22.54 cm) high. Columns are to be $3\frac14$ inches (8.25 cm) wide, with a $\frac{5}{16}$ inch (0.8 cm) space between them. The main title (on the first page) should begin 1 inch (2.54 cm) from the top edge of the page. The second and following pages should begin 1 inch (2.54 cm) from the top edge. On all pages, the bottom margin should be $1\frac{1}{8}$ inches (2.86 cm) from the bottom edge of the page for $8.5 \times 11$-inch paper; for A4 paper, approximately $1\frac{5}{8}$ inches (4.13 cm) from the bottom edge of the page. \subsection{Margins and page numbering} All printed material, including text, illustrations, and charts, must be kept within a print area $6\frac{7}{8}$ inches (17.46 cm) wide by $8\frac{7}{8}$ inches (22.54 cm) high. Page numbers should be in the footer, centered and $\frac{3}{4}$ inches from the bottom of the page. The review version should have page numbers, yet the final version submitted as camera ready should not show any page numbers. The \LaTeX\ template takes care of this when used properly. \subsection{Type style and fonts} Wherever Times is specified, Times Roman may also be used. If neither is available on your word processor, please use the font closest in appearance to Times to which you have access. MAIN TITLE. Center the title $1\frac{3}{8}$ inches (3.49 cm) from the top edge of the first page. The title should be in Times 14-point, boldface type. Capitalize the first letter of nouns, pronouns, verbs, adjectives, and adverbs; do not capitalize articles, coordinate conjunctions, or prepositions (unless the title begins with such a word). Leave two blank lines after the title. AUTHOR NAME(s) and AFFILIATION(s) are to be centered beneath the title and printed in Times 12-point, non-boldface type. This information is to be followed by two blank lines. The ABSTRACT and MAIN TEXT are to be in a two-column format. MAIN TEXT. Type main text in 10-point Times, single-spaced. Do NOT use double-spacing. All paragraphs should be indented 1 pica (approx.~$\frac{1}{6}$ inch or 0.422 cm). Make sure your text is fully justified---that is, flush left and flush right. Please do not place any additional blank lines between paragraphs. Figure and table captions should be 9-point Roman type as in \cref{fig:onecol,fig:short}. Short captions should be centred. \noindent Callouts should be 9-point Helvetica, non-boldface type. Initially capitalize only the first word of section titles and first-, second-, and third-order headings. FIRST-ORDER HEADINGS. (For example, {\large \bf 1. Introduction}) should be Times 12-point boldface, initially capitalized, flush left, with one blank line before, and one blank line after. SECOND-ORDER HEADINGS. (For example, { \bf 1.1. Database elements}) should be Times 11-point boldface, initially capitalized, flush left, with one blank line before, and one after. If you require a third-order heading (we discourage it), use 10-point Times, boldface, initially capitalized, flush left, preceded by one blank line, followed by a period and your text on the same line. \subsection{Footnotes} Please use footnotes\footnote{This is what a footnote looks like. It often distracts the reader from the main flow of the argument.} sparingly. Indeed, try to avoid footnotes altogether and include necessary peripheral observations in the text (within parentheses, if you prefer, as in this sentence). If you wish to use a footnote, place it at the bottom of the column on the page on which it is referenced. Use Times 8-point type, single-spaced. \subsection{Cross-references} For the benefit of author(s) and readers, please use the {\small\begin{verbatim} \cref{...} \end{verbatim}} command for cross-referencing to figures, tables, equations, or sections. This will automatically insert the appropriate label alongside the cross-reference as in this example: \begin{quotation} To see how our method outperforms previous work, please see \cref{fig:onecol} and \cref{tab:example}. It is also possible to refer to multiple targets as once, \eg~to \cref{fig:onecol,fig:short-a}. You may also return to \cref{sec:formatting} or look at \cref{eq:also-important}. \end{quotation} If you do not wish to abbreviate the label, for example at the beginning of the sentence, you can use the {\small\begin{verbatim} \Cref{...} \end{verbatim}} command. Here is an example: \begin{quotation} \Cref{fig:onecol} is also quite important. \end{quotation} \subsection{References} List and number all bibliographical references in 9-point Times, single-spaced, at the end of your paper. When referenced in the text, enclose the citation number in square brackets, for example~\cite{Authors14}. Where appropriate, include page numbers and the name(s) of editors of referenced books. When you cite multiple papers at once, please make sure that you cite them in numerical order like this \cite{Alpher02,Alpher03,Alpher05,Authors14b,Authors14}. If you use the template as advised, this will be taken care of automatically. \begin{table} \centering \begin{tabular}{@{}lc@{}} \toprule Method & Frobnability \\ \midrule Theirs & Frumpy \\ Yours & Frobbly \\ Ours & Makes one's heart Frob\\ \bottomrule \end{tabular} \caption{Results. Ours is better.} \label{tab:example} \end{table} \subsection{Illustrations, graphs, and photographs} All graphics should be centered. In \LaTeX, avoid using the \texttt{center} environment for this purpose, as this adds potentially unwanted whitespace. Instead use {\small\begin{verbatim} \centering \end{verbatim}} at the beginning of your figure. Please ensure that any point you wish to make is resolvable in a printed copy of the paper. Resize fonts in figures to match the font in the body text, and choose line widths that render effectively in print. Readers (and reviewers), even of an electronic copy, may choose to print your paper in order to read it. You cannot insist that they do otherwise, and therefore must not assume that they can zoom in to see tiny details on a graphic. When placing figures in \LaTeX, it's almost always best to use \verb+\includegraphics+, and to specify the figure width as a multiple of the line width as in the example below {\small\begin{verbatim} \usepackage{graphicx} ... \includegraphics[width=0.8\linewidth] {myfile.pdf} \end{verbatim} } \subsection{Color} Please refer to the author guidelines on the CVPR\ 2022\ web page for a discussion of the use of color in your document. If you use color in your plots, please keep in mind that a significant subset of reviewers and readers may have a color vision deficiency; red-green blindness is the most frequent kind. Hence avoid relying only on color as the discriminative feature in plots (such as red \vs green lines), but add a second discriminative feature to ease disambiguation. \section{Final copy} You must include your signed IEEE copyright release form when you submit your finished paper. We MUST have this form before your paper can be published in the proceedings. Please direct any questions to the production editor in charge of these proceedings at the IEEE Computer Society Press: \url{https://www.computer.org/about/contact}. {\small \bibliographystyle{ieee_fullname} \section{Introduction} After receiving paper reviews, authors may optionally submit a rebuttal to address the reviewers' comments, which will be limited to a {\bf one page} PDF file. Please follow the steps and style guidelines outlined below for submitting your author response. The author rebuttal is optional and, following similar guidelines to previous CVPR conferences, is meant to provide you with an opportunity to rebut factual errors or to supply additional information requested by the reviewers. It is NOT intended to add new contributions (theorems, algorithms, experiments) that were absent in the original submission and NOT specifically requested by the reviewers. You may optionally add a figure, graph, or proof to your rebuttal to better illustrate your answer to the reviewers' comments. Per a passed 2018 PAMI-TC motion, reviewers should refrain from requesting significant additional experiments for the rebuttal or penalize for lack of additional experiments. Authors should refrain from including new experimental results in the rebuttal, especially when not specifically requested to do so by the reviewers. Authors may include figures with illustrations or comparison tables of results reported in the submission/supplemental material or in other papers. Just like the original submission, the rebuttal must maintain anonymity and cannot include external links that reveal the author identity or circumvent the length restriction. The rebuttal must comply with this template (the use of sections is not required, though it is recommended to structure the rebuttal for ease of reading). \subsection{Response length} Author responses must be no longer than 1 page in length including any references and figures. Overlength responses will simply not be reviewed. This includes responses where the margins and formatting are deemed to have been significantly altered from those laid down by this style guide. Note that this \LaTeX\ guide already sets figure captions and references in a smaller font. \section{Formatting your Response} {\bf Make sure to update the paper title and paper ID in the appropriate place in the tex file.} All text must be in a two-column format. The total allowable size of the text area is $6\frac78$ inches (17.46 cm) wide by $8\frac78$ inches (22.54 cm) high. Columns are to be $3\frac14$ inches (8.25 cm) wide, with a $\frac{5}{16}$ inch (0.8 cm) space between them. The top margin should begin 1 inch (2.54 cm) from the top edge of the page. The bottom margin should be $1\frac{1}{8}$ inches (2.86 cm) from the bottom edge of the page for $8.5 \times 11$-inch paper; for A4 paper, approximately $1\frac{5}{8}$ inches (4.13 cm) from the bottom edge of the page. Please number any displayed equations. It is important for readers to be able to refer to any particular equation. Wherever Times is specified, Times Roman may also be used. Main text should be in 10-point Times, single-spaced. Section headings should be in 10 or 12 point Times. All paragraphs should be indented 1 pica (approx.~$\frac{1}{6}$ inch or 0.422 cm). Figure and table captions should be 9-point Roman type as in \cref{fig:onecol}. List and number all bibliographical references in 9-point Times, single-spaced, at the end of your response. When referenced in the text, enclose the citation number in square brackets, for example~\cite{Alpher05}. Where appropriate, include the name(s) of editors of referenced books. \begin{figure}[t] \centering {\rule{0pt}{0.5in} \rule{0.9\linewidth}{0pt}} \caption{Example of caption. It is set in Roman so that mathematics (always set in Roman: $B \sin A = A \sin B$) may be included without an ugly clash.} \label{fig:onecol} \end{figure} To avoid ambiguities, it is best if the numbering for equations, figures, tables, and references in the author response does not overlap with that in the main paper (the reviewer may wonder if you talk about \cref{fig:onecol} in the author response or in the paper). See \LaTeX\ template for a workaround. \subsection{Illustrations, graphs, and photographs} All graphics should be centered. Please ensure that any point you wish to make is resolvable in a printed copy of the response. Resize fonts in figures to match the font in the body text, and choose line widths which render effectively in print. Readers (and reviewers), even of an electronic copy, may choose to print your response in order to read it. You cannot insist that they do otherwise, and therefore must not assume that they can zoom in to see tiny details on a graphic. When placing figures in \LaTeX, it is almost always best to use \verb+\includegraphics+, and to specify the figure width as a multiple of the line width as in the example below {\small\begin{verbatim} \usepackage{graphicx} ... \includegraphics[width=0.8\linewidth] {myfile.pdf} \end{verbatim} } {\small \bibliographystyle{ieee_fullname}

73995c8047ed5f0a4d9ddccf7d1078c2b58e4e3d

\section{Introduction} Computational electromagnetics (CEM) focuses on numerical algorithms of solving Maxwell's equations via computer techniques\cite{jin2011theory}. Electromagnetic computational algorithms have been widely applied in scientific simulation\cite{knott2004radar,poljak2007advanced}, engineering design\cite{jin2009finite} and information processing\cite{nikolova1999microwave}. Typical electromagnetic computational algorithms include finite difference method (FDM)\cite{jin2011theory}, finite element method (FEM)\cite{jin2015finite} and Method of Moments (MoM)\cite{harrington1993field,Chew2007}. The common approaches of these methods are to solve Maxwell's equations by discretizing and converting them into linear matrix equations. Such linear matrix equations usually have a large number of unknowns which consume lots of computing resources to solve. Many application scenarios require real-time and accurate computational electromagnetic simulations, such as biomedical imaging\cite{nikolova1999microwave} and target detection. Computational efficiency has been a long-standing challenge in developing electromagnetic computational methods. \par One method of acceleration is to reduce the computational complexity based on the physical laws, such as conjugate gradient-fast Fourier transform\cite{sarkar1986application}, adaptive integral method\cite{bleszynski1996aim}, multilevel fast multipole algorithm\cite{rokhlin1985rapid}, etc. Another method is to divide the whole computation into offline and online parts, such as reduced basis method\cite{dang2017quasi}, characteristic basis function\cite{prakash2003characteristic} and model-order reduction method\cite{schilders2008model}. The online computation can be accelerated by pre-computing partial models in the offline process. Similarly, machine learning (ML) techniques are applied to accelerate the online computation via offline training process. Artificial neural network, self-organizing map, radial basis function network are incorporated into FEM\cite{chedid1996automatic,lowther1998automatic,zou2015hybrid} and FDTD\cite{wu1996combination} in order to improve the computational efficiency. ML techniques are also applied into the microwave circuit aided design\cite{wang1997knowledge}, antenna optimizations\cite{haupt2007antenna}, microwave imaging\cite{lee2004neural}, nondestructive testing\cite{massa2005classification}, direction of arrival estimation\cite{el2000neural}, etc. \par With the development of high-performance computing platform, deep learning (DL) is ushering in leapfrog development \cite{lecun2015deep} and has been widely applied in image\cite{simonyan2014very}, video\cite{ciaparrone2020deep} and speech\cite{hinton2012deep} processing. DL promotes the development of fast computational simulations in the field of physics modeling. Pixel-based\cite{li2016fall} and object-based\cite{kipf2018neural} deep learning methods are applied into intuitive physics by learning the relationships between objects and predicting the future physical states\cite{girdhar2020forward}. DL techniques are applied to accelerate the computational fluid, such as smoke simulation\cite{tompson2017accelerating}, steady fluid prediction\cite{Guo2016}, super resolution fluid flow\cite{xie2018tempogan}, etc. In order to improve the generalization ability, DL techniques aim to learn the exact physical laws instead of approximating inner law from the data set\cite{greydanus2019hamiltonian}, such as Hamiltonian neural network motivated by Hamiltonian mechanics\cite{greydanus2019hamiltonian}, Symplectic recurrent neural network with Symplectic integration\cite{chen2019symplectic} and Lagrangian neural network for simulating Lagrangians\cite{cranmer2020lagrangian}. DL are also applied to accelerate the electromagnetic computational simulations by replacing the parts of high computational complexity\cite{yao2018machine,sun2020machine}. In electromagnetic inverse scattering problems, DL-based methods can improve the resolution of inversions and speed up the computing time\cite{wei2018deep,wei2019dominant}. Applications of DL in the antenna engineering have been reported, such as phase synthesis of reflectarrays\cite{8454981}, coding programmable metasurfaces\cite{8988246} and metasurface imager\cite{li2019machine}. \par Recently, various works have been reported to investigate the relationship between DL and partial differential equations (PDEs). DL can help reduce the curse of dimensionalities in the numerical algorithms of PDEs\cite{han2018solving}, such as Schr\"odinger equation\cite{hermann2020deep}, Navier-stokes equation\cite{raissi2018hidden}, Burger's equation\cite{sirignano2018dgm}, Poisson's equation\cite{9062556}, Heat equation\cite{sharma2018weakly}, Advection-diffusion equation\cite{pang2019fpinns}, etc. To improve the performance and generalization ability, DL are combined with traditional numerical algorithms of PDEs, such as variational method\cite{weinan2018deep}, multi-grid method\cite{greenfeld2019learning}, hierarchical matrices\cite{fan2019multiscale}, weak solutions\cite{zang2020weak} and solution ansatz\cite{beidokhti2009solving}. PDEs can be embedded into the loss of neural network and then solved via the automatic differentiation of DL, such as physics-informed neural network\cite{raissi2019physics}, DeepXDE\cite{lu2021deepxde}. Deep neural networks can be interpreted as the PDEs and PDEs' numerical algorithms\cite{lu2018beyond}. Such interpretations can further guide the design of deep neural networks. In \cite{long2018pde} similarities between the differential operators and convolutions are investigated. The parabolic or hyperbolic equations are embedded into the deep neural networks \cite{ruthotto2019deep}. Deep neural networks are designed based on the multi-grid method\cite{he2019mgnet} and runge-kutta method\cite{zhu2019convolutional}. \par In this paper, we propose the physics-informed supervised residual learning (PISRL) as a general framework for 2D electromagnetic forward modeling. PISRL is proposed based on the similarities between ResNet and stationary or non-stationary iterative methods. Then, stationary and non-stationary iterative PISRL neural networks (SIPISRLNN and NSIPISRLNN) are proposed by embedding stationary and non-stationary iterative methods into ResNets. This work is also an example that deep neural network is designed and interpreted from the perspective of traditional numerical algorithms. SIPISRLNN and NSIPISRLNN mimic the iterative process of non-stationary and stationary iterative methods with the same calculation of residuals and addition of modifications\cite{axelsson1996iterative}. The only difference is that convolutional neural networks (CNNs) map between residuals and modifications of the solutions of linear matrix equations in SIPISRLNN and NSIPISRLNN. It is noted that PISRL can be easily expanded to solve different electromagnetic problems because it aims to solve a system of linear matrix equations instead of a specific electromagnetic problem. The generalities of SIPISRLNN and NSIPISRLNN are validated by solving volume integral equations of non-lossy and lossy scatterers. Trained by the data set generated by MoM, both SIPISRLNN and NSIPISRLNN can achieve good computing precisions. In numerical results of non-lossy scatterers, the MSEs of SIPISRLNN and NSIPISRLNN finally converge below $3.152 \times 10^{-4}$ and $4.8925 \times 10^{-7}$; in results of lossy scatterers, the MSEs converge below$1.2775 \times 10^{-4}$ and $1.567 \times 10^{-7}$. Furthermore, the generalization abilities of SIPISRLNN and NSIPISRLNN are verified on the data sets of contrast shapes and incident frequencies that are unseen in the training procedure. \par This paper is organized as follows. Section \uppercase\expandafter{\romannumeral2} investigates the mathematical connections between ResNet and iterative methods, then proposes and analyzes SIPISRLNN and NSIPISRLNN. In Section \uppercase\expandafter{\romannumeral3}, we introduce the volume integral equations and the solving methods including MoM, SIPISRLNN and NSIPISRLNN. In Section \uppercase\expandafter{\romannumeral4}, the generalities of SIPISRLNN and NSIPISRLNN are verified by solving VIEs of non-lossy and lossy scatterers, and the generalization abilities are validated on the data sets of contrast shapes and incident frequencies that are unseen in the training process. Observations and further discussions are summarized in Section \uppercase\expandafter{\romannumeral5}. \begin{figure} \centering \subfigure[] {\includegraphics[width=0.25\linewidth]{resnetblockori.png} \label{resnetblockori}} \subfigure[] {\includegraphics[width=0.25\linewidth]{resnetblockid.png} \label{resnetblockid}} \subfigure[] {\includegraphics[width=0.25\linewidth]{residualblock.png} \label{residualblock}} \caption{Schematics of residual blocks. (a): the general residual block; (b): the residual block with identity mapping; (c): the proposed residual learning block} \end{figure} \section{Physics-informed Supervised Residual Learning} \subsection{Stationary and Nonstationary Iteration Method for Solving Linear Equation System} In electromagnetic forward modeling problems, Maxwell's equations with the boundary conditions will be converted into a linear equation system: \begin{equation} \mathbb{A} \mathrm{x}=\mathrm{b} \label{eq1} \end{equation} If \eq{eq1} is difficult to solve, we can replace $\mathbb{A}$ with the approximated one $\mathbb{A}^{a}$ to make \eq{eq1} easy to solve: \begin{equation} \mathbb{A}^{a} \mathrm{x}=\mathrm{b} \label{eq2} \end{equation} Let $\mathrm{x}^{a}$ and $\mathrm{x}^{*}$ denote the approximated and true solution respectively, then the difference between $\mathrm{x}^{a}$ and $\mathrm{x}^{*}$ satisfies the following relationship: \begin{equation} \mathbb{A} (\mathrm{x}^* - \mathrm{x}^a) = \mathrm{b} - \mathbb{A}\mathrm{x}^a \label{eq3} \end{equation} With $\Delta \mathrm{x}$ denoting the error $\mathrm{x}^* - \mathrm{x}^a$, \eq{eq3} can be written as: \begin{equation} \mathbb{A}\Delta \mathrm{x} =\mathrm{b} - \mathbb{A}\mathrm{x}^a \label{eq4} \end{equation} If the error $\Delta \mathrm{x}$ meets the stop criterion, $x^a$ can be accepted as the solution of \eq{eq1}. Otherwise, it needs to be modified by adding $\Delta \mathrm{x}$: \begin{equation} \mathrm{x}^{a}_{modified} = \mathrm{x}^a + \Delta \mathrm{x} \label{eq5} \end{equation} The $\Delta x$ can be obtained by solving \eq{eq4}. If \eq{eq4} is hard to solve, then the approximated $\Delta \mathrm{x}^{a}$ can be obtained by solving the aproximated one: \begin{equation} \mathbb{A}^{a}\Delta \mathrm{x}^{a} =\mathrm{b} - \mathbb{A}\mathrm{x}^a \label{eq6} \end{equation} The modified approximated solution can be written as: \begin{equation} \mathrm{x}^{a}_{modified} = \mathrm{x}^a + {\mathbb{A}^{a}}^{-1}(\mathrm{b} - \mathbb{A}\mathrm{x}^a) \label{eq7} \end{equation} If $\mathrm{x}^{a}_{modified}$ still cannot meet the stop criterion, the above process can be repeated to obtain the satisfied solutions. Assuming the above process is repeated $K$ times, a sequence of approximation solutions can be obtained: \begin{equation} \mathrm{x}^{a}_{1}, \mathrm{x}^{a}_{2}, \mathrm{x}^{a}_{3}, ... , \mathrm{x}^{a}_{K} \label{eq8} \end{equation} The update equation of $\mathrm{x}^{a}_{k}$ can be written as: \begin{equation} \mathrm{x}^{a}_{k+1} = \mathrm{x}^{a}_{k} + {\mathbb{A}^{a}}^{-1}(\mathrm{b} - \mathbb{A}\mathrm{x}^{a}_{k}) \label{eq9} \end{equation} If we use an operator $\mathcal{L}$ to replace ${\mathbb{A}^{a}}^{-1}(\mathrm{b} - \mathbb{A}\mathrm{x}^{a}_{k})$ in \eq{eq9}, then the equation can be written as: \begin{equation} \mathrm{x}^{a}_{k+1} = \mathrm{x}^{a}_{k} + \mathcal{L} (\mathrm{b} - \mathbb{A}\mathrm{x}^{a}_{k}, {\mathbb{A}^{a}}^{-1}) \label{eq10} \end{equation} By recursively adding the former $K-1$-th iteration, we can get \begin{equation} \mathrm{x}^{a}_{K} = \mathrm{x}^{a}_{1} + \sum_{i=1}^{K-1} \mathcal{L} (\mathrm{b} - \mathbb{A}\mathrm{x}^{a}_{i}, {\mathbb{A}^{a}}^{-1}) \label{eq11} \end{equation} \eq{eq10} and \eq{eq11} can be regarded as the stationary iteration scheme. Note that if the array $\mathbb{A}^{a}$ is different at each iteration step, \eq{eq10} and \eq{eq11} can be transformed as the non-stationary iteration scheme: \begin{equation} \mathrm{x}^{a}_{k+1} = \mathrm{x}^{a}_{k} + \mathcal{L} (\mathrm{b} - \mathbb{A}\mathrm{x}^{a}_{k}, {\mathbb{A}^{a}_{i}}^{-1}) \label{eq12} \end{equation} \begin{equation} \mathrm{x}^{a}_{K} = \mathrm{x}^{a}_{1} + \sum_{i=1}^{K-1} \mathcal{L} (\mathrm{b} - \mathbb{A}\mathrm{x}^{a}_{i}, {\mathbb{A}^{a}_{i}}^{-1}) \label{eq13} \end{equation} \par Two commonly used iteration schemes are Richardson iteration scheme and Jacobi iteration scheme. They can be written in the format of \eq{eq10}. \subsubsection{Richardson Iteration Scheme} The Richardson iteration scheme can be written as: \begin{equation} \mathrm{x}^{a}_{k+1} = \mathrm{x}^{a}_{k} + \omega (\mathrm{b} - \mathbb{A} \mathrm{x}^{a}_{k}) \label{eq14} \end{equation} where $\omega$ is above zero. \par \subsubsection{Jacobi Iteration Scheme} In the Jacobi iteration scheme, the array $\mathbb{A}$ is split into diagonal part $\mathbb{D}$, upper triangle part $\mathbb{L}$ and lower triangle part $\mathbb{U}$: \begin{equation} \mathbb{A} =\mathbb{D}-\mathbb{L}-\mathbb{U} \label{eq15} \end{equation} Then the Jacobi iteration scheme can be written as: \begin{equation} \mathrm{x}^{a}_{k+1} = \mathbb{D}^{-1}(\mathbb{L}+\mathbb{U})\mathrm{x}^{a}_{k} + \mathbb{D}^{-1}\mathrm{b} \label{eq16} \end{equation} It can be written as: \begin{equation} \begin{aligned} \mathrm{x}^{a}_{k+1} & = \mathbb{D}^{-1}(\mathbb{L}+\mathbb{U})\mathrm{x}^{a}_{k} + \mathbb{D}^{-1}\mathrm{b} \\ & = \mathbb{D}^{-1}(\mathbb{L}+\mathbb{U}+\mathbb{D}-\mathbb{D})\mathrm{x}^{a}_{k} + \mathbb{D}^{-1}\mathrm{b} \\ & = \mathrm{x}^{a}_{k} + \mathbb{D}^{-1}(\mathrm{b} - \mathbb{A} \mathrm{x}^{a}_{k}) \\ \label{eq17} \end{aligned} \end{equation} \par \subsection{Physics-informed Supervised Residual Learning Blocks} Residual neural networks (ResNets) have proven effective network architectures and have wide applications in image processing\cite{he2016deep}. ResNets are stacked structures composed of a family of modular blocks and these modular blocks are named as residual blocks\cite{he2016deep}. ResNets can be easily extended to have extremely deep structures with modular blocks. Works on the interpretation of ResNet have been reported. Single block of ResNet can be formulated as the forward Euler discretization of the ordinary differential equations (ODEs)\cite{weinan2017proposal, haber2017stable}. The stability and reversibility of the ResNet are also studied from the perspective of ODEs\cite{chang2018reversible}. ResNet is also interpreted as a transport equation in the control problem\cite{li2017deep} and dynamical systems\cite{chang2017multi}. \fig{resnetblockori} shows the general structure of the residual block which can be expressed as\cite{he2016deep}: \begin{align} \mathrm{y}_{k} = h(\mathrm{x}_{k})+\mathcal{F}(\mathrm{x}_{k},\mathcal{W}_{k}) \label{eq18} \\ \mathrm{x}_{k+1} = \mathcal{N}(\mathrm{y}_{k}) \label{eq19} \end{align} where $\mathrm{x}_{k}$ and $\mathrm{x}_{k+1}$ are input and output of the $k$-th residual block; $h$ denotes the linear projection; $\mathcal{N}$ denotes the nonlinear activation; $\mathcal{F}$ denotes the transform function of the residual and $\mathcal{W}_{k}$ is the parameter set of $\mathcal{F}$. In \eq{eq18}, if we regard $h(\mathrm{x}_k)$ and $\mathrm{y}_k$ as the predicted and desired output, then $\mathcal{F}(\mathrm{x}_{k},\mathcal{W}_{k}) = h(\mathrm{x}_k) - \mathrm{y}_k$ will be the residual of $h(\mathrm{x}_k)$ and $\mathrm{y}_k$. \par In \cite{he2016identity}, by taking $h$ and $\mathcal{N}$ as the identity mappings, residual neural network can achieve better performance with improved generalization ability. Then \eq{eq18} and \eq{eq19} can be re-written as: \begin{equation} \mathrm{x}_{k+1} = \mathrm{x}_{k}+\mathcal{F}(\mathrm{x}_k,\mathcal{W}_k) \label{eq20} \end{equation} \fig{resnetblockid} shows the structure of a residual block with identity mappings. \par Note that the residual block with identity mappings \eq{eq20} is similar to the non-stationary iteration scheme \eq{eq12}. Motivated by such observation, we propose the physics-informed supervised residual learning (PISRL) blocks by embedding the non-stationary iteration scheme into the residual block. The PISRL block can be expressed mathematically as: \begin{equation} \mathrm{x}_{k+1} = \mathrm{x}_{k}+\mathcal{F}(\mathrm{b} - \mathbb{A} \mathrm{x}_{k},\mathcal{W}_k) \label{eq21} \end{equation} \fig{residualblock} illustrates the structure of the proposed PISRL block. With the input $\mathrm{x}_{k}$, the residual is calculated first by: \begin{equation} \mathbb{R}_k = \mathrm{b} - \mathbb{A} \mathrm{x}_{k} \end{equation} where the $\mathbb{A}$ and $\mathrm{b}$ come from the linear equation system \eq{eq1}. Then the calculated residual $\mathbb{R}_k $ goes through the convolution neural network (CNN) and the structure of CNN is determined by the complexity of the linear equation system. The CNN takes the current residual $\mathbb{R}_k$ as input and then outputs the corresponding modification $\Delta \mathrm{x}_{k}$ that is added to $\mathrm{x}_{k}$. \par If the CNN of PISRL block is the same in every step, then the PISRL block can be regarded as embedding the stationary iteration scheme \eq{eq10}: \begin{equation} \mathrm{x}_{k+1} = \mathrm{x}_{k}+\mathcal{F}(\mathrm{b} - \mathbb{A} \mathrm{x}_{k},\mathcal{W}) \label{eq22} \end{equation} \par \begin{figure*} \centering \includegraphics[width=0.8\linewidth]{stationaryNN.png} \caption{Schematics of stationary iterative physics-informed supervised residual learning neural network. The number of channels is denoted in the pictures.} \label{stationaryNN} \end{figure*} \begin{figure*} \centering \includegraphics[width=0.8\linewidth]{nonstationaryNN.png} \caption{Schematics of non-stationary iterative physics-informed supervised residual learning neural network} \label{nonstationaryNN} \end{figure*} \par \subsection{Stationary Iterative Physics-informed Supervised Residual Learning Neural Network} Based on the PISRL blocks embedding the stationary iteration scheme, we propose one general architecture of physics-informed supervised residual learning neural network (PISRLNN): stationary iterative physics-informed supervised residual learning neural network (SIPISRLNN). \par \fig{stationaryNN} shows the schematics of SIPISRLNN. CNNs of PISRL blocks in each iteration have the same structure and share the same parameter set. In each iteration, the mapping between the input and corresponding output is totally different and the shared CNN needs to make reliable predictions. Therefore, the CNN should have powerful learning ability. In this paper, U-net \cite{ronneberger2015u} is adopted as the CNN structure in the PISRL blocks, as illustrated in \fig{stationaryNN}. \par U-net is an effective network with good learning ability and has achieved good performance in image segmentation \cite{ronneberger2015u}. U-net uses $3 \times 3$ convolutional kernels, $2 \times 2$ max pooling, bilinear upsampling, Tanh nonlinear activation and batch normalization throughout the architecture. U-net has a symmetric path consisting of a contracting path and an expansive path \cite{ronneberger2015u}. In the contracting path, U-net extracts the feature maps by $2 \times 2$ max pooling the input and then upsamples the concatenation of high-level and low-level feature maps by bilinear upsampling in the expansive path. In the last two layers, batch normalization is not included because the last two layers aims to reduce the channels of feature maps. \par Note that the input and output of U-net have the same size due to the fully convolutional neural network structure. The input of the U-net is the concatenation of the residual $\mathrm{b} - \mathbb{A} \mathrm{x}_{k}$ and the previous output $\mathrm{x}_{k}$. $\mathrm{x}_{k}$ functions as the hidden state of recurrent neural networks (RNNs). $\mathrm{x}_{k}$ can offer more information and the process of iterations. The channels of input and output are denoted as $2c$ and $c$ depends on the linear equation systems, for example, $c$ will be $1$ in the real-valued linear equations and $c$ will be $2$ in the complex-valued linear equations. \par \subsection{Non-stationary Iterative Physics-informed Supervised Residual Learning Neural Network} With the PISRL blocks embedding the non-stationary iteration scheme, another general architecture of physics-informed supervised residual learning neural network is proposed: non-stationary iterative physics-informed supervised residual learning neural network (NSIPISRLNN). \par \fig{nonstationaryNN} shows the schematics of non-stationary PISRLNN and it can be regarded as the unrolled structure of the SIPISRLNN. The NSIPISRLNN have 7 blocks with the same structure. The CNNs in each PISRL block have the same structure but different parameter sets. In each iteration, independent CNNs can make independent mappings of different pairs of residuals and modifications. Thus, the CNNs need no complicated structures. In this paper, the CNN consists of five stacked modules of $3 \times 3$ convolutional kernels, batch normalization and Tanh nonlinear activation, as shown in \fig{nonstationaryNN}. The output has the same size of the input due to the fully convolutional operations. The channels $c$ of input and output depend on the linear equations and it will be $1$ in real-valued ones and $2$ in complex-valued ones. \subsection{Analysis of Physics-informed Supervised Residual Learning Neural Network} In the SIPISRLNN, the formulation of one block can be written as: \begin{equation} \mathrm{x}_{k+1}^{a} = \mathrm{x}_{k}^{a}+\mathcal{F}(\mathrm{b} - \mathbb{A}\mathrm{x}_{k}^{a},\mathcal{W}) \label{eq23} \end{equation} By summing all output of the former $L-1$ iterations, \eq{eq23} can be written as: \begin{equation} \mathrm{x}_{L}^{a} = \mathrm{x}_1^a + \sum_{i=1}^{L-1} \mathcal{F}(\mathrm{b} - \mathbb{A}\mathrm{x}_{i}^{a},\mathcal{W}) \label{eq24} \end{equation} If the loss function $ \mathcal{E} $ is defined as the mean squared error (MSE), then $ \mathcal{E} $ can be written as: \begin{equation} \begin{aligned} \mathcal{E} & = \frac{1}{N}|| \mathrm{x}_{L}^{a} - \mathrm{x}^{*}||_{F}^2 \\ & = \frac{1}{N} || \mathrm{x}_1^a + \sum_{i=1}^{L-1} \mathcal{F}(\mathrm{b} - \mathbb{A}\mathrm{x}_{i}^{a},\mathcal{W}) - \mathrm{x}^{*}||_{F}^2 \end{aligned} \label{eq25} \end{equation} where the $N$ is the number of element in the $\mathrm{x}^{*}$. Then, based on the chain rule, the back propagation of $ \mathcal{E} $ with respect to the $l$-th input can be written as\cite{he2016identity}: \begin{equation} \frac{\partial \mathcal{E}}{\partial \mathrm{x}_l} = \frac{\partial \mathcal{E}}{\partial \mathrm{x}_L}\frac{\partial \mathrm{x}_L}{\mathrm{x}_l} = \frac{\partial \mathcal{E}}{\partial \mathrm{x}_L}(1+ \frac{\partial}{\mathrm{x}_l}\sum_{i=l}^{L-1}\mathcal{F}(\mathrm{b} - \mathbb{A}\mathrm{x}_{i}^{a},\mathcal{W})) \label{eq26} \end{equation} \eq{eq26} shows that the gradient of $ \mathcal{E} $ with respect to $\mathrm{x}_l$ is unlikely to be zero because $1+ \frac{\partial}{\mathrm{x}_l}\sum_{i=l}^{L-1}\mathcal{F}(\mathrm{b} - \mathbb{A}\mathrm{x}_{i}^{a},\mathcal{W})$ cannot be always zero \cite{he2016identity}. This can prevent the problem of gradient vanishing. Similarly, we can obtain the same conclusions of NSIPISRLNN, of which the block can be written mathematically: \begin{equation} \mathrm{x}_{L}^{a} = \mathrm{x}_1^a + \sum_{i=1}^{L-1} \mathcal{F}(\mathrm{b} - \mathbb{A}\mathrm{x}_{i}^{a},\mathcal{W}_i) \label{eq27} \end{equation} \section{Volume Integral Equation} \begin{figure} \centering \includegraphics[width=0.5\linewidth]{viemodel.png} \caption{The model setup for volume integral equation. Green triangles and yellow squares denote the receivers and transmitters.} \label{viemodel} \end{figure} Volume integral equation (VIE) describes the scattered field of the dielectric scatter in free space $D$, as shown in \fig{viemodel}. The electrical properties of dielectric scatterers only change along the lateral axes. Assuming the permeability of scatterers as vacuum permittivity $\mu_0$, the permittivity can be written as: \begin{equation} \varepsilon(\mathbf{r}) = \varepsilon_0 \varepsilon_r(\mathbf{r}) - j \frac{\sigma(\mathbf{r})}{\omega} \label{eq28} \end{equation} where $\varepsilon_0$, $\varepsilon_r(\mathbf{r})$, $\sigma(\mathbf{r})$, $\omega$, $r$ are vacuum permittivity, relative permittivity, conductivity, angular frequency and the position vector in free space $D$ respectively. The total electric field $E^{tot}(\mathbf{r})$ will be excited when the dielectric scatterers are illuminated by the incident field $E^{inc}(\mathbf{r})$. Taking transverse magnetic (TM) mode into account, the relationship between $E^{tot}(\mathbf{r})$ and $E^{inc}(\mathbf{r})$ satisfies the electric field integral equation (EFIE): \begin{equation} \begin{aligned} \mathbf{E}^{tot}(\mathbf{r})=&E^{inc}(\mathbf{r})\\ &+k_b^2\int_{D} G_D \left(\mathbf{r}, \mathbf{r}^{\prime}\right) \chi\left(\mathbf{r}^{\prime}\right) E^{tot}\left(\mathbf{r}^{\prime}\right) d\mathbf{r}^{\prime} \quad \mathbf{r} \in D \label{eq29} \end{aligned} \end{equation} where $k_b$ is the wavenumber: $k_b^2 = \omega \mu_0 \epsilon_0$, $G_D$ is Green's function in 2D free space: \begin{equation} G_D\left(\mathbf{r}, \mathbf{r}^{\prime}\right) = \frac{1}{4j}H_0^{(2)}(k_b|\mathbf{r} - \mathbf{r}^{\prime}|) \label{eq30} \end{equation} and $\chi\left(\mathbf{r}\right)$ is the contrast: \begin{equation} \chi\left(\mathbf{r}\right) = \frac{\varepsilon(\mathbf{r}) - \varepsilon_0}{\varepsilon_0} = \varepsilon_r(\mathbf{r}) -1 - j\frac{\sigma(\mathbf{r})}{\omega \varepsilon_0} \label{eq31} \end{equation} The incident field will induce the current on the scatterers and the induced current can be described as: \begin{equation} J\left(\mathbf{r}\right) = \chi\left(\mathbf{r}\right) E^{tot}\left(\mathbf{r}\right) \label{eq32} \end{equation} Then the scattered field is radiated from the induced current, which can be calculated by: \begin{equation} E^{sca}(\mathbf{r}^{\prime \prime})=\int_{D} G_S \left(\mathbf{r}^{\prime \prime}, \mathbf{r}^{\prime}\right) \chi\left(\mathbf{r}^{\prime}\right) E^{tot}\left(\mathbf{r}^{\prime}\right) d\mathbf{r}^{\prime} \quad \mathbf{r^{\prime}} \in S \label{eq33} \end{equation} where $ \mathbf{r}^{\prime \prime} $ is the distance between the scatterers and the receivers. \subsection{Method of Moments} To determine the scattered field, \eq{eq33} needs to be solved numerically. Method of moments (MoM) is applied to solve \eq{eq33} in this work. The domain $D$ is discretized with $M \times N$ subdomains. By applying the pulse basis function and delta function, the EFIE \eq{eq33} can be discretized as: \begin{equation} \begin{aligned} &E^{tot}(\mathbf{r}_{p})=E^{inc}(\mathbf{r}_{p}) \\ &+\frac{k_b^2}{4j} \sum_{q=1}^{M \times N} \chi\left(\mathbf{r}_{q}\right) E^{tot}\left(\mathbf{r}_{q}\right) \int_{D_q} H_0^{(2)}(k_b|\mathbf{r}_{p} - \mathbf{r}_{q}^{\prime}|) d\mathbf{r}_{q}^{\prime} \label{eq34} \end{aligned} \end{equation} where $p$ and $q$ are the index of the subdomain. By simplifying the notations, \eq{eq39} can be written as: \begin{equation} E^{tot}_{p}=E^{inc}_{p} + (-\frac{j k_b^2}{4}) \sum_{q=1}^{M \times N} \chi_{q} E^{tot}_{q} \int_{D_q} H_0^{(2)}(k_b|\mathbf{r}_{p} - \mathbf{r}_{q}^{\prime}|) d\mathbf{r}_{q}^{\prime} \label{eq35} \end{equation} For each subdomain, the \eq{eq35} is satisfied and then $M \times N$ equations can be obtained. These equations can be written in the format of matrix equations: \begin{equation} \mathbf{E}^{tot} = \mathbf{E}^{inc} + \mathbf{G}_D \cdot \boldsymbol{\chi} \cdot \mathbf{E}^{tot} \label{eq36} \end{equation} Then \eq{eq36} can be converted as a linear equation system: \begin{equation} (\mathbf{I} - \mathbf{G}_D \cdot \boldsymbol{\chi}) \cdot \mathbf{E}^{tot} = \mathbf{E}^{inc} \label{eq37} \end{equation} where $\mathbf{G}_D$ is a $MN \times MN$ array and can be determined by: \begin{equation} \begin{aligned} \mathbf{G}_{D,pq} &= -\frac{j k_b^2}{4} \int_{D_q} H_0^{(2)}(k_b|\mathbf{r}_{p} - \mathbf{r}_{q}^{\prime}|) d\mathbf{r}_{q}^{\prime} \\ &=\left\{ \begin{array}{lr} -\frac{j}{2} [\pi k_b a_{p}H_1^{(2)}(k_b a_{p}) - 2j], \ p=q \\ -\frac{j\pi k_b a_{p}}{2} J_1(k_b a_{p}) H_0^{(2)}(k_b|\mathbf{r}_{p}-\mathbf{r}_{q}^{\prime}|), \ else \\ \end{array} \right. \end{aligned} \label{eq38} \end{equation} where $a_{p}$ is the radius of the equivalent circle with the same area $S_{p}$ as the square subdomain, which can be calculated as: \begin{equation} a_p = \frac{\sqrt{S_p}}{\pi} \label{eq39} \end{equation} \subsection{SIPISRLNN} The SIPISRLNN is applied to solve \eq{eq37} that is the discretized form of the EFIE (\eq{eq33}). By solving \eq{eq37} based on the stationary iteration method, according to \eq{eq10}, the iteration equation can be written as: \begin{equation} \mathbf{E}^{tot}_{k+1} = \mathbf{E}^{tot}_{k} + \mathcal{L}(\mathbf{E}^{inc} - (\mathbf{I} - \mathbf{G}_D \cdot \boldsymbol{\chi})\cdot \mathbf{E}_k^{tot}, \ {(\mathbf{I} - \mathbf{G}_D \cdot \boldsymbol{\chi})^{a}}^{-1}) \label{eq40} \end{equation} where $(\mathbf{I} - \mathbf{G}_D \cdot \mathcal{\chi})^{a}$ is the approximation array of $(\mathbf{I} - \mathbf{G}_D \cdot \mathcal{\chi})$. Then similar to \eq{eq22}, the iteration equation of the $k$-th stationary PISRL block can be written as: \begin{equation} \mathbf{E}^{tot}_{k+1} = \mathbf{E}^{tot}_{k} + \mathcal{F}(\mathbf{E}^{inc} - (\mathbf{I} - \mathbf{G}_D \cdot \boldsymbol{\chi})\cdot \mathbf{E}_k^{tot}, \mathbf{E}_k^{tot}, \mathcal{W}) \label{eq41} \end{equation} where $\mathcal{F}$ denotes U-Net and $\mathcal{W}$ is the parameters of U-Net. Note that the $\mathcal{W}$ is the same and shared in all iterations due to the stationary iteration scheme. The $k$-th residual in the $k$-th PISRL block can be defined as: \begin{equation} \mathbb{R}_k = \mathbf{E}^{inc} - (\mathbf{I} - \mathbf{G}_D \cdot \boldsymbol{\chi})\cdot \mathbf{E}_k^{tot} \label{eq42} \end{equation} The Green's function $\mathbf{G}_D$ is assumed the same in this paper due to the uniform background. $\boldsymbol{\chi}$ and $\mathbf{E}^{inc}$ are different in different numerical samples. $\mathbf{G}_D$, $\boldsymbol{\chi}$ and $\mathbf{E}^{inc}$ are included in the SIPISRLNN to calculate the $\mathbb{R}_k$ in \eq{eq42}. The objective function of the SIPISRLNN is the mean squared error (MSE), which can be written as: \begin{equation} \mathcal{E} = \frac{1}{N}|| \mathbf{E}_{tot}^{\prime} - \mathbf{E}_{tot}^{*}||_{F}^2 \label{eq43} \end{equation} where $ \mathbf{E}_{tot}^{\prime}$ is the total field calculated by SIPISRLNN, $ \mathbf{E}_{tot}^{*}$ is the total field calculated by MoM, $|| \cdot ||_F$ is the F norm of a matrix and $N$ is the number of elements in a matrix. \subsection{NSIPISRLNN} The NSIPISRLNN is also applied to solve the discretized EFIE \eq{eq37}. Similar to \eq{eq12}, the iteration equation of the non-stationary iteration method can be written as: \begin{equation} \mathbf{E}^{tot}_{k+1} = \mathbf{E}^{tot}_{k} + \mathcal{L}(\mathbf{E}^{inc} - (\mathbf{I} - \mathbf{G}_D \cdot \boldsymbol{\chi})\cdot \mathbf{E}_k^{tot}, \ {(\mathbf{I} - \mathbf{G}_D \cdot \boldsymbol{\chi})^{a}_k}^{-1}) \label{eq44} \end{equation} where $(\mathbf{I} - \mathbf{G}_D \cdot \mathcal{\chi})^{a}_k$ is the approximation array of $(\mathbf{I} - \mathbf{G}_D \cdot \mathcal{\chi})$ but is different and independent in each iteration. Based on \eq{eq21}, the iteration equation of the $k$-th non-stationary PISRL block can be written as: \begin{equation} \mathbf{E}^{tot}_{k+1} = \mathbf{E}^{tot}_{k} + \mathcal{F}_k(\mathbf{E}^{inc} - (\mathbf{I} - \mathbf{G}_D \cdot \boldsymbol{\chi})\cdot \mathbf{E}_k^{tot}, \mathcal{W}_k) \label{eq45} \end{equation} where $\mathcal{F}_k$ is the $k$-th CNN as shown in \fig{nonstationaryNN} and $\mathcal{W}_k$ is the coresponding parameter set. Also, the $k$-th residual can be calculated by: \begin{equation} \mathbb{R}_k = \mathbf{E}^{inc} - (\mathbf{I} - \mathbf{G}_D \cdot \boldsymbol{\chi})\cdot \mathbf{E}_k^{tot} \label{eq46} \end{equation} $\mathbf{G}_D$, $\boldsymbol{\chi}$ and $\mathbf{E}^{inc}$ are needed in the NSIPISRLNN to calcluate the $\mathbb{R}_k$ in \eq{eq42}. The objective function of NSIPISRLNN is MSE (\eq{eq43}) that is the same as SIPISRLNN. \begin{figure} \centering \includegraphics[width=0.9\linewidth]{chap06-nonlossypinc.png} \caption{Incident field of non-lossy scatterers: the first row is the real part and the second is the imaginary part; from left to right: the incident angle is $[0^{\circ}, 45^{\circ}, 90^{\circ}, 135^{\circ}, 180^{\circ}, 225^{\circ}, 270^{\circ}, 315^{\circ}]$.} \label{nonlossypinc} \end{figure} \begin{figure} \centering \includegraphics[width=0.8\linewidth]{chap06-nonlossychi.png} \caption{Examples of non-lossy scatterers: the first row is the real part and the second is the imaginary part.} \label{nonlossychi} \end{figure} \section{Numerical Results and Analysis} In this section, we verify the validity of SIPISRLNN and NSIPISRLNN by numerically solving 2D VIEs. Furthermore, The universality of PISRL is validated by solving 2D VIEs of non-lossy and lossy scatterers because \eq{eq37} are different when the $\boldsymbol{\chi}$ is non-lossy and lossy. \par The model setup is illustrated in \fig{viemodel} and the size of domain $D$ is assumed as $0.15m \times 0.15m$. The domain $D$ is discretized into $32\times 32$ grids. There are 1 transmitter and 32 receivers on a circle with 1.67m centered at the center of the domain $D$. \subsection{Non-lossy scatterers} The incident field's frequency is fixed at 3GHz and angle is randomly selected from $0^{\circ}$, $45^{\circ}$, $90^{\circ}$, $135^{\circ}$, $180^{\circ}$, $225^{\circ}$, $270^{\circ}$, $315^{\circ}$, as shown in \fig{nonlossypinc}. The contrast of non-lossy scatterers in the domain $D$ and can be derived from \eq{eq31}: \begin{equation} \chi\left(\mathbf{r}\right) = \frac{\varepsilon(\mathbf{r}) - \varepsilon_0}{\varepsilon_0} = \varepsilon_r(\mathbf{r}) -1 \label{eq47} \end{equation} The domain $D$ is assumed to have three cylinders which have random positions and radii. The real parts of the three cylinders' contrast vary from 0 to 1, 1 to 2 and 0 to 2 with their imaginary parts are 0, as shown in \fig{nonlossychi}. \par Both SIPISRLNN and NSIPISRLNN are implemented in Pytorch and computed on one Nvidia V100 GPU. The parameters of SIPISRLNN and NSIPISRLNN are optimized by Adam\cite{Kingma2014}. The learning rate is initialized as 0.002 and decayed by $\times 0.8$ every 20 epochs. We use MoM to generate 40000 data samples for SIPISRLNN and NSIPISRLNN of which 32000 are for training and 8000 for testing. \subsubsection{SIPISRLNN} \begin{figure} \centering \includegraphics[width=0.75\linewidth]{sipisrlnolossy-loss.png} \caption{Non-lossy scatterers: MSE convergence curve of SIPISRLNN. } \label{sipisrlnolossy-loss} \end{figure} \begin{figure} \centering \subfigure[] {\includegraphics[width=0.85\linewidth]{sipisrlnolossy1.png}} \subfigure[] {\includegraphics[width=0.85\linewidth]{sipisrlnolossy2.png}} \subfigure[] {\includegraphics[width=0.85\linewidth]{sipisrlnolossy3.png}} \caption{Non-lossy scatterers: comparisons of total field computed by SIPISRLNN and MoM. From left to right: contrast $\chi$, input of SIPISRLNN (iteration initial value, $E^{inc}$), total field computed by MoM $E^{tot}_{MoM}$, total field computed by SIPISRLNN $E^{tot}_{SIPISRLNN}$, and their absolute error distribution. The first row is the real parts and the second is the imaginary parts.} \label{sipisrlnolossy} \end{figure} \begin{figure} \centering \subfigure[] {\includegraphics[width=0.95\linewidth]{sipisrlnolossy1all.png}} \subfigure[] {\includegraphics[width=0.95\linewidth]{sipisrlnolossy2all.png}} \subfigure[] {\includegraphics[width=0.95\linewidth]{sipisrlnolossy3all.png}} \caption{Non-lossy scatterers: updated total fields in each iteration computed by SIPISRLNN. From left to right: contrast $\chi$, total field computed by MoM $E^{tot}_{MoM}$, input of SIPISRLNN (initial value of iteration, $E^{inc}$), total fields computed by SIPISRLNN $E^{tot}_{SIPISRLNN}$ in the first iteration, second iteration and third iteration. The first row is the real parts and the second is the imaginary parts.} \label{sipisrlnolossyall} \end{figure} SIPISRLNN is trained to solve the VIEs of non-lossy scatterers and SIPISRLNN is assumed to have three iterations. \fig{sipisrlnolossy-loss} shows the convergence curve of MSE and MSE finally achieves below $3.152 \times 10^{-4}$. The training MSE and testing MSE agree well with each other and little overfitting exists. U-Net of SIPISRLNN is the same and shared in all iterations. That means U-Net needs to learn the various and complicated mappings between residuals and modifications in all iterations. Thus, the optimization of U-Net is not stable enough and it can be observed that the convergence curve of MSE fluctuates several times during the training process. \par \fig{sipisrlnolossy} illustrates the comparisons of total field computed by SIPISRLNN ($E^{tot}_{SIPISRLNN}$) and MoM ($E^{tot}_{MoM}$) and they are randomly selected from the testing data set. $E^{tot}_{SIPISRLNN}$ and $E^{tot}_{MoM}$ are in a good agreement and only some points in the absolute error distribution are not good enough. For better interpretation of SIPISRLNN, we print and show the updated total fields in each iteration of SIPISRLNN, as shown in \fig{sipisrlnolossyall}. With the incident field $E_{inc}$ as input, the difference between the updated $E^{tot}_{SIPISRLNN}$ and $E^{tot}_{MoM}$ becomes smaller and the corresponding MSE also decreases. \begin{figure} \centering \subfigure[ ] {\includegraphics[width=0.45\linewidth]{sipisrlnolossy-train-realerror.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{sipisrlnolossy-train-imagerror.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{sipisrlnolossy-test-realerror.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{sipisrlnolossy-test-imagerror.png}} \caption{Non-lossy scatterers: histograms of mean absolute error of $E^{tot}_{SIPISRLNN}$. (a) and (b) are real and imaginary parts of $E^{tot}_{SIPISRLNN}$ in the training data set; (c) and (d) are real and imaginary parts of $E^{tot}_{SIPISRLNN}$ in the testing data set.} \label{sipisrlnolossy-hist} \end{figure} \fig{sipisrlnolossy-hist} shows the histogram of mean absolute error (MAE) of $E^{tot}_{SIPISRLNN}$ in the training and testing data sets. The MAE's mean and standard deviation (std) of the training and testing data sets agree well with each other. This observation is consistent with the MSE convergence curve, as shown in \fig{sipisrlnolossy-loss}. \subsubsection{NSIPISRLNN} \begin{figure} \centering \includegraphics[width=0.75\linewidth]{nsipisrlnolossy-loss.png} \caption{Non-lossy scatterers: MSE convergence curve of NSIPISRLNN. } \label{nsipisrlnolossy-loss} \end{figure} \begin{figure} \centering \subfigure[] {\includegraphics[width=0.85\linewidth]{nsipisrlnolossy1.png}} \subfigure[] {\includegraphics[width=0.85\linewidth]{nsipisrlnolossy2.png}} \subfigure[] {\includegraphics[width=0.85\linewidth]{nsipisrlnolossy3.png}} \caption{Non-lossy scatterers: comparisons of total field computed by NSIPISRLNN and MoM. From left to right: contrast $\chi$, input of NSIPISRLNN (iteration initial value, $E^{inc}$), total field computed by MoM $E^{tot}_{MoM}$, total field computed by NSIPISRLNN $E^{tot}_{NSIPISRLNN}$, and their absolute error distribution. The first row is the real parts and the second is the imaginary parts.} \label{nsipisrlnolossy} \end{figure} \begin{figure} \centering \subfigure[] {\includegraphics[width=1\linewidth]{nsipisrlnolossy1all.png}} \subfigure[] {\includegraphics[width=1\linewidth]{nsipisrlnolossy2all.png}} \subfigure[] {\includegraphics[width=1\linewidth]{nsipisrlnolossy3all.png}} \caption{Non-lossy scatterers: updated total fields in each iteration computed by NSIPISRLNN. From left to right: contrast $\chi$, total field computed by MoM $E^{tot}_{MoM}$, input of NSIPISRLNN (initial value of iteration, $E^{inc}$), total fields computed by NSIPISRLNN $E^{tot}_{NSIPISRLNN}$ in the first, third, fifth, sixth and seventh iteration. The first row is the real parts and the second is the imaginary parts.} \label{nsipisrlnolossyall} \end{figure} \begin{figure} \centering \subfigure[ ] {\includegraphics[width=0.45\linewidth]{nsipisrlnolossy-train-realerror.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{nsipisrlnolossy-train-imagerror.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{nsipisrlnolossy-test-realerror.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{nsipisrlnolossy-test-imagerror.png}} \caption{Non-lossy scatterers: histograms of mean absolute error of $E^{tot}_{NSIPISRLNN}$. (a) and (b) are real and imaginary parts of $E^{tot}_{NSIPISRLNN}$ in the training data set; (c) and (d) are real and imaginary parts of $E^{tot}_{NSIPISRLNN}$ in the testing data set.} \label{nsipisrlnolossy-hist} \end{figure} NSIPISRLNN consists of seven PISRL blocks and can be regarded as the unrolled structure of SIPISRLNN. The convergence curve of MSE during the training process is shown in \fig{nsipisrlnolossy-loss}. The training MSE agrees well with testing MSE. Both of them decrease steadily and converge below $4.8925 \times 10^{-7}$. Compared to the converged MSE $3.152 \times 10^{-4}$ of SIPISRLNN, NSIPISRLNN have better computational precision. NSIPISRLNN has more iterations (7 PISRL blocks) but simpler structure of CNN. It still indicates that it is more reasonable to use an independent CNN to map residuals and modifications in each iteration, and better precision can be achieved. \par \fig{nsipisrlnolossy} shows three examples of total fields solved by NSIRPISRLNN, which are randomly chosen from the testing data set. Total fields computed by MoM and NSIPISRLNN agrees well with each other and the corresponding absolute error is at a very small level. \fig{nsipisrlnolossyall} illustrates the updated total field in the iterative process with incident field as input. The updated total field in the first iteration is the coarse-grained approximation of true total field, then it improves and its MSE decreases with the increase of iterations. \fig{nsipisrlnolossy-hist} offers a whole overview of NSIPISRLNN's performance by showing the MAE histogram of training and testing data sets. The mean and std are in a good agreement and it verifies the MSE convergence curve, as shown in the \fig{nsipisrlnolossy-loss}. Small values of std reveal that the range of MAE is small and concentrated. The computational precision of NSIPISRLNN is better and more stable. \subsubsection{Generalization Ability on Contrast Shape} The generalization ability of SIPISRLNN and NSIPISRLNN on unseen contrast shapes is considered. There are 8 types of unseen contrast shapes, including polygon, six-pointed star, five-pointed star, heart shape and the letters T, H, U, EE, as shown in \fig{gennolossychi}. We generate 40 samples for each type of unseen contrast shapes and the range of contrast is in $[0,2]$. The generalization validation data set has 320 data samples ($40 \times 8$). \fig{nsipisrlnolossy-gencpl-hist} shows the MAE histograms of total field computed by SIPISRLNN and NSIPISRLNN on the contrast generalization validation data set. It can be observed that the MAEs of contrast generalization validation data set are at the same level of error as ones of training and testing data sets. The generalization abilities of SIPISRLNN and NSIPISRLNN on unseen contrast shapes are verified. \fig{nolossy-rnn-shape} and \fig{nolossy-shape} show the randomly chosen results of SIPISRLNN and NSIPISRLNN on the data set of unseen contrast shapes. \begin{figure} \centering \includegraphics[width=1\linewidth]{gennolossychi.png} \caption{Examples of the generalization validation data set for unseen contrast shapes. } \label{gennolossychi} \end{figure} \begin{figure} \centering \subfigure[ ] {\includegraphics[width=0.45\linewidth]{sipisrlnolossy-gencpl-realerror.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{sipisrlnolossy-gencpl-imagerror.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{nsipisrlnolossy-validate-realerror.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{nsipisrlnolossy-validate-imagerror.png}} \caption{Non-lossy scatterers: MAE histograms of SIPISRLNN's and NSIPISRLNN's results on generalization validation data set for unseen contrast shapes. (a) and (b) are histograms of real parts and imaginary parts of total fields computed by SIPISRLNN. (c) and (d) are histograms of real parts and imaginary parts of total fields computed by NSIPISRLNN.} \label{nsipisrlnolossy-gencpl-hist} \end{figure} \begin{figure} \centering \subfigure[ ] {\includegraphics[width=0.45\linewidth]{1rnolossygenresult.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{3rnolossygenresult.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{6rnolossygenresult.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{5rnolossygenresult.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{7rnolossygenresult.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{10rnolossygenresult.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{25rnolossygenresult.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{17rnolossygenresult.png}} \caption{Non-lossy scatterers: results of SIPISRLNN on the unseen contrast shape generalization validation data set.} \label{nolossy-rnn-shape} \end{figure} \begin{figure} \centering \subfigure[ ] {\includegraphics[width=0.45\linewidth]{1result.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{2result.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{6result.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{5result.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{7result.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{28result.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{21result.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{18result.png}} \caption{Non-lossy scatterers: results of NSIPISRLNN on the unseen contrast shape generalization validation data set.} \label{nolossy-shape} \end{figure} \subsubsection{Generalization Ability on Incident Frequency} The generalization ability of SIPISRL and NSIPISRL on the frequencies of incident field is also considered. Here, we consider 12 frequencies of incident field: 2GHz, 2.3GHz, 2.6GHz, 2.9GHz, 3.1GHz, 3.2GHz, 3.4GHz, 3.7GHz, 4GHz, 4.5GHz, 5GHz. For each incident frequency, we generate 320 samples as the data set and the shapes and values of contrast are the same as the contrast generalization validation data set, as shown in \fig{gennolossychi}. \fig{nolossy_gen_loss} shows the SIPISRLNN's and NSIPISRLNN's MSEs of different incident frequencies. Both SIPISRLNN and NISIPISRLNN behaved well from 2GHz to 3.2GHz with stable MSEs. When the incident frequency is higher than 3.2GHz, obvious increases in MSE is observed. It indicates that SIPISRLNN and NSIPISRLNN has good generalization ability on lower frequencies compared to the training incident frequency. \fig{nolossy-rnn-freq} and \fig{nolossy-freq} give detailed results of SIPISRLNN and NSIPISRLNN with respect to different incident frequencies. \begin{figure} \centering \subfigure[SIPISRLNN MSE] {\includegraphics[width=1\linewidth]{rnnnolossy_gen_loss.png}} \subfigure[NSIPISRLNN MSE] {\includegraphics[width=1\linewidth]{nolossy_gen_loss.png}} \caption{Non-lossy scatterers: MSE of SIPISRLNN and NSIPISRLNN at different frequency.} \label{nolossy_gen_loss} \end{figure} \begin{figure} \centering \subfigure[ 2.0GHz] {\includegraphics[width=0.45\linewidth]{1rnnnolossyresult2-0.png}} \subfigure[ 2.3GHz] {\includegraphics[width=0.45\linewidth]{7rnnnolossyresult2-3.png}} \subfigure[ 2.6GHz] {\includegraphics[width=0.45\linewidth]{6rnnnolossyresult2-6.png}} \subfigure[ 2.8GHz] {\includegraphics[width=0.45\linewidth]{11rnnnolossyresult2-8.png}} \subfigure[ 2.9GHz] {\includegraphics[width=0.45\linewidth]{11rnnnolossyresult2-9.png}} \subfigure[ 3.1GHz] {\includegraphics[width=0.45\linewidth]{33rnnnolossyresult3-1.png}} \subfigure[ 3.2GHz] {\includegraphics[width=0.45\linewidth]{5rnnnolossyresult3-2.png}} \subfigure[ 3.4GHz] {\includegraphics[width=0.45\linewidth]{1rnnnolossyresult3-4.png}} \subfigure[ 3.7GHz] {\includegraphics[width=0.45\linewidth]{9rnnnolossyresult3-7.png}} \subfigure[ 4.0GHz] {\includegraphics[width=0.45\linewidth]{5rnnnolossyresult4-0.png}} \subfigure[ 4.5GHz] {\includegraphics[width=0.45\linewidth]{2rnnnolossyresult4-5.png}} \subfigure[ 5.0GHz] {\includegraphics[width=0.45\linewidth]{9rnnnolossyresult5-0.png}} \caption{Non-lossy scatterers: results of SIPISRLNN on the frequency generalization validation data set.} \label{nolossy-rnn-freq} \end{figure} \begin{figure} \centering \subfigure[ 2.0GHz] {\includegraphics[width=0.45\linewidth]{15nolossy2_0.png}} \subfigure[2.3GHz ] {\includegraphics[width=0.45\linewidth]{7nolossy2_3.png}} \subfigure[2.6GHz ] {\includegraphics[width=0.45\linewidth]{5nolossy2_6.png}} \subfigure[ 2.8GHz] {\includegraphics[width=0.45\linewidth]{8nolossy2_8.png}} \subfigure[ 2.9GHz] {\includegraphics[width=0.45\linewidth]{1nolossy2_9.png}} \subfigure[ 3.1GHz] {\includegraphics[width=0.45\linewidth]{8nolossy3_1.png}} \subfigure[3.2GHz ] {\includegraphics[width=0.45\linewidth]{2nolossy3_2.png}} \subfigure[3.4GHz ] {\includegraphics[width=0.45\linewidth]{12nolossy3_4.png}} \subfigure[3.7GHz ] {\includegraphics[width=0.45\linewidth]{6nolossy3_7.png}} \subfigure[ 4.0GHz] {\includegraphics[width=0.45\linewidth]{7nolossy4_0.png}} \subfigure[4.5GHz ] {\includegraphics[width=0.45\linewidth]{10nolossy4_5.png}} \subfigure[5.0GHz ] {\includegraphics[width=0.45\linewidth]{10nolossy5_0.png}} \caption{Non-lossy scatterers: results of NSIPISRLNN on the frequency generalization validation data set.} \label{nolossy-freq} \end{figure} \par \subsection{Lossy scatterers} In this section, lossy scatterers in domain $D$ are taken into account and the contrast can be calculated using \eq{eq31}. There are four cylinders in domain $D$ and they have random positions and radii. The ranges of real parts and imaginary parts of contrast can refer to \tab{table01}. \fig{lossychi} shows six examples of contrast distributions. The frequency of incident field is also fixed at 3GHz and the directions of incident field are randomly selected in $[0^{\circ}, 90^{\circ}, 180^{\circ}, 270^{\circ}]$, as shown in \fig{lossypinc}. \par SIPISRLNN and NSIPISRLNN are implemented in Pytorch and the computing platform is one Nvidia V100 GPU. Adam\cite{Kingma2014} is the optimization algorithm of SIPIRSLNN and NSIPISRLNN. The learning rate is initialized as 0.002 and multiplied by 0.8 every 20 epochs. MoM is used to generate 40000 data samples for SIPISRLNN and NSIPISRLNN of which 80\% are for training and 20\% for testing. Note that \eq{eq37} can be regarded as a different type of equations compared to non-lossy ones due to the imaginary parts of contrast. \begin{table} \renewcommand{\arraystretch}{1.3} \caption{Setup of Lossy Scatterers' Contrast } \label{table01} \centering \begin{threeparttable} \begin{tabular}{ccc} \toprule Cylinder & Real Part & Imaginary Part \\ \midrule Cylinder 1 & $[0,1]$ & $[-1,0]$ \\ Cylinder 2 & $[0,1]$ & $[-2,-1]$ \\ Cylinder 3 & $[1,2]$ & $[-1,0]$ \\ Cylinder 4 & $[1,2]$ & $[-2,-1]$ \\ \bottomrule \end{tabular} \end{threeparttable} \end{table} \begin{figure} \centering \includegraphics[width=0.8\linewidth]{chap06-lossychi.png} \caption{Examples of lossy scatterers: the first row is the real part and the second is the imaginary part.} \label{lossychi} \end{figure} \begin{figure} \centering \includegraphics[width=0.9\linewidth]{chap06-lossypinc.png} \caption{Incident field of lossy scatterers: the first row is the real part and the second is the imaginary part; from left to right: the incident angle is $[0^{\circ}, 90^{\circ}, 180^{\circ}, 270^{\circ}]$.} \label{lossypinc} \end{figure} \subsubsection{SIPISRLNN} SIPISRLNN has three iterations, that is, a stationary iterative PSIRL block is iterated three times. \fig{sipisrllossy-loss} shows the convergence curve of SIPISRLNN's MSE and MSE finally converges below $1.2775 \times 10^{-4}$. The training and testing MSEs are in a good agreement. The beginning of the training process is not very stable, which can be observed from the curve of MSE. \par \fig{sipisrllossy} shows three randomly chosen total field $E^{tot}_{SIPISRLNN}$ computed by SIPISRLNN and their comparisons with total fields $E^{tot}_{MoM}$ solved by MoM. With different contrast distributions, total fields solved by SIPISRLNN agree well with the ones solved by MoM. \fig{sipisrllossyall} further illustrates the updated total field in each iteration with the incident field as the initial value of iterations. As shown in \fig{sipisrllossyall}, the difference between the updated $E^{tot}_{SIPISRLNN}$ and $E^{tot}_{MoM}$ decreases with the increase of iterations. The histogram of MAE of $E^{tot}_{SIPISRLNN}$ is illustrated in \fig{sipisrllossy-hist}. The MAE's mean and std of training and testing data set agree well with each other. This observation is also consistent with the convergence curve of MSE. \begin{figure} \centering \includegraphics[width=0.75\linewidth]{sipisrllossy-loss.png} \caption{Lossy scatterers: MSE convergence curve of SIPISRLNN. } \label{sipisrllossy-loss} \end{figure} \begin{figure} \centering \subfigure[] {\includegraphics[width=0.85\linewidth]{sipisrllossy1.png}} \subfigure[] {\includegraphics[width=0.85\linewidth]{sipisrllossy2.png}} \subfigure[] {\includegraphics[width=0.85\linewidth]{sipisrllossy3.png}} \caption{Lossy scatterers: comparisons of total field computed by SIPISRLNN and MoM. From left to right: contrast $\chi$, input of SIPISRLN (iteration initial value, $E^{inc}$), total field computed by MoM $E^{tot}_{MoM}$, total field computed by SIPISRLNN $E^{tot}_{SIPISRLNN}$, and their absolute error distribution. The first row is the real parts and the second is the imaginary parts.} \label{sipisrllossy} \end{figure} \begin{figure} \centering \subfigure[] {\includegraphics[width=0.95\linewidth]{sipisrllossy1all.png}} \subfigure[] {\includegraphics[width=0.95\linewidth]{sipisrllossy2all.png}} \subfigure[] {\includegraphics[width=0.95\linewidth]{sipisrllossy3all.png}} \caption{Lossy scatterers: updated total fields in each iteration computed by SIPISRLNN. From left to right: contrast $\chi$, total field computed by MoM $E^{tot}_{MoM}$, input of SIPISRLNN (initial value of iteration, $E^{inc}$), total fields computed by SIPISRLNN $E^{tot}_{SIPISRLNN}$ in the first iteration, second iteration and third iteration. The first row is the real parts and the second is the imaginary parts.} \label{sipisrllossyall} \end{figure} \begin{figure} \centering \subfigure[ ] {\includegraphics[width=0.45\linewidth]{sipisrllossy-train-realerror.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{sipisrllossy-train-imagerror.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{sipisrllossy-test-realerror.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{sipisrllossy-test-imagerror.png}} \caption{Lossy scatterers: histograms of mean absolute error of $E^{tot}_{SIPISRLNN}$. (a) and (b) are real and imaginary parts of $E^{tot}_{SIPISRLNN}$ in the training data set; (c) and (d) are real and imaginary parts of $E^{tot}_{SIPISRLNN}$ in the testing data set.} \label{sipisrllossy-hist} \end{figure} \subsubsection{NSIPISRLNN} NSIPISRLNN have seven iterations and each iteration have an independent PISRL blocks. The convergence of MSE is shown in \fig{nsipisrllossy-loss} and the final MSE is below $1.567 \times 10^{-7}$. The training MSE is a little smaller than the testing MSE but they are still in a good agreement. The final MSE of SIPISRLNN is $1.2775 \times 10^{-4}$ and NSIPISRLNN has better performance. In the numerical results of both lossy and non-lossy scatterers, NSIPISRLNN has better computing precisions which indicates the non-stationary scheme is more suitable for PISRL. \par \fig{nsipisrllossy} shows three randomly chosen results of NSIPISRLNN. The discrepancy between total fields computed by NSIPISRLNN and MoM is small, and the absolute error is also at a small level. Updated total fields during the iterative process are shown in \fig{nsipisrllossyall}. The updated total fields are improved gradually with the increase of iterations. To evaluate the SIPISRLNN from a global perspective, MAEs of both training and testing data sets are charted as histograms, as shown in \fig{nsipisrllossy-hist}. The testing MAE's mean and std are a little higher than training ones which is still acceptable. This observation is consistent with the convergence curve of MSE. Both training and testing MAE's stds are very small and it indicates the stable precisions of NSIPISRLNN. \begin{figure} \centering \includegraphics[width=0.75\linewidth]{nsipisrllossy-loss.png} \caption{Lossy scatterers: MSE convergence curve of NSIPISRLNN. } \label{nsipisrllossy-loss} \end{figure} \begin{figure} \centering \subfigure[] {\includegraphics[width=0.85\linewidth]{nsipisrllossy1.png}} \subfigure[] {\includegraphics[width=0.85\linewidth]{nsipisrllossy2.png}} \subfigure[] {\includegraphics[width=0.85\linewidth]{nsipisrllossy3.png}} \caption{Lossy scatterers: comparisons of total field computed by NSIPISRLNN and MoM. From left to right: contrast $\chi$, input of NSIPISRLNN (iteration initial value, $E^{inc}$), total field computed by MoM $E^{tot}_{MoM}$, total field computed by NSIPISRLNN $E^{tot}_{NSIPISRLNN}$, and their absolute error distribution. The first row is the real parts and the second is the imaginary parts.} \label{nsipisrllossy} \end{figure} \begin{figure} \centering \subfigure[] {\includegraphics[width=1\linewidth]{nsipisrllossy1all.png}} \subfigure[] {\includegraphics[width=1\linewidth]{nsipisrllossy2all.png}} \subfigure[] {\includegraphics[width=1\linewidth]{nsipisrllossy3all.png}} \caption{Lossy scatterers: updated total fields in each iteration computed by NSIPISRLNN. From left to right: contrast $\chi$, total field computed by MoM $E^{tot}_{MoM}$, input of NSIPISRLNN (initial value of iteration, $E^{inc}$), total fields computed by NSIPISRLNN $E^{tot}_{NSIPISRLNN}$ in the first, third, fifth, sixth and seventh iteration. The first row is the real parts and the second is the imaginary parts.} \label{nsipisrllossyall} \end{figure} \begin{figure} \centering \subfigure[ ] {\includegraphics[width=0.45\linewidth]{nsipisrllossy-train-realerror.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{nsipisrllossy-train-imagerror.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{nsipisrllossy-test-realerror.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{nsipisrllossy-test-imagerror.png}} \caption{Lossy scatterers: histograms of mean absolute error of $E^{tot}_{NSIPISRLNN}$. (a) and (b) are real and imaginary parts of $E^{tot}_{NSIPISRLNN}$ in the training data set; (c) and (d) are real and imaginary parts of $E^{tot}_{NSIPISRLNN}$ in the testing data set.} \label{nsipisrllossy-hist} \end{figure} \subsubsection{Generalization Ability on Contrast Shape} In the numerical results of lossy scatterers, we verify the generalization ability of SIPISRLNN and NSIPISRLNN on contrast shape that are unseen in the training process. Eight types of unseen contrast shapes are taken into account, including polygon, six-pointed star, five-pointed star, heart shape and the letters T, H, U, EE, as shown in \fig{genlossychi}. The real parts and imaginary parts are independently selected in [0,2] and [-2,0] respectively. For each type of shape, 40 data samples are generated and there are 320 samples in the generalization validation data set. The MAE histograms of SIPISRLNN's and NSIPISRLNN's results on the contrast generalization validation data set are plotted in \fig{nsipisrllossy-gencpl-hist}. It can be observed that the means and stds of generalization data set are higher than the training and testing ones. The generalization ability of SIPISRLNN on the unseen contrast shapes is not good and NSIPISRLNN still maintains good generalization ability on the unseen contrast shapes. \fig{lossy-rnn-shape} and \fig{lossy-shape} illustrate the randomly chosen results of SIPISRLNN and NSIPISRLNN on the data set of unseen contrast shapes. \begin{figure} \centering \includegraphics[width=1\linewidth]{genlossychi.png} \caption{Examples of generalization validation data set for unseen contrast shapes. } \label{genlossychi} \end{figure} \begin{figure} \centering \subfigure[ ] {\includegraphics[width=0.45\linewidth]{sipisrllossy-gencpl-realerror.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{sipisrllossy-gencpl-imagerror.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{nsipisrllossy-gen-realerror.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{nsipisrllossy-gen-imagerror.png}} \caption{Lossy scatterers: MAE histogram of SIPISRLNN's and NSIPISRLNN's results on generalization validation data set for unseen contrast shapes. (a) and (b) are histograms of real parts and imaginary parts of total fields computed by SIPISRLNN. (c) and (d) are histograms of real parts and imaginary parts of total fields computed by NSIPISRLNN.} \label{nsipisrllossy-gencpl-hist} \end{figure} \begin{figure} \centering \subfigure[ ] {\includegraphics[width=0.45\linewidth]{19rlossygenresult.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{28rlossygenresult.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{10rlossygenresult.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{36rlossygenresult.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{21rlossygenresult.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{23rlossygenresult.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{39rlossygenresult.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{1rlossygenresult.png}} \caption{Lossy scatterers: results of SIPISRLNN on the unseen contrast shape generalization validation data set.} \label{lossy-rnn-shape} \end{figure} \begin{figure} \centering \centering \subfigure[ ] {\includegraphics[width=0.45\linewidth]{26lossygenresult.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{2lossygenresult.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{13lossygenresult.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{21lossygenresult.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{39lossygenresult.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{3lossygenresult.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{12lossygenresult.png}} \subfigure[ ] {\includegraphics[width=0.45\linewidth]{16lossygenresult.png}} \caption{Lossy scatterers: results of NSIPISRLNN on the unseen contrast shape generalization validation data set.} \label{lossy-shape} \end{figure} \subsubsection{Generalization Ability on Incident Frequency} The generalization abilities of SIPISRLNN and NSIPISRLNN are also considered in the numerical results of lossy scatterers. In this section, we consider 12 frequencies of incident field: 2GHz, 2.3GHz, 2.6GHz, 2.9GHz, 3.1GHz, 3.2GHz, 3.4GHz, 3.7GHz, 4GHz, 4.5GHz, 5GHz. 320 samples are generated by MoM for each incident frequency. The values and shapes of contrast are kept the same as the contrast generalization validation data set for every incident frequency, as shown in \fig{genlossychi}. \fig{lossy_gen_loss} shows the SIPISRLNN's and NSIPISRLNN's MSE of different incident frequencies. SIPISRLNN maintains good levels of error from 2.8GHz to 3.4GHz but behaves not well in the other incident frequencies. The generalization ability of SIPISRLNN in the lossy numerical results is limited compared to the SIPISRLNN's in the non-lossy numerical results. NSIPISRLNN have good computing precisions from 2.0GHz to 3.4GHz and the generalization ability of NSIPISRLNN is verified. Detailed results of SIPISRLNN and NSIPISRLNN are shown in \fig{lossy-rnn-freq} and \fig{lossy-freq} respectively. \begin{figure} \centering \subfigure[SIPISRLNN MSE] {\includegraphics[width=1\linewidth]{rnnlossy_gen_loss.png}} \subfigure[NSIPISRLNN MSE] {\includegraphics[width=1\linewidth]{lossy_gen_loss.png}} \caption{Lossy scatterers: MSE of SIPISRLNN and NSIPISRLNN at different frequency.} \label{lossy_gen_loss} \end{figure} \begin{figure} \centering \subfigure[ 2.0GHz] {\includegraphics[width=0.45\linewidth]{8rnnlossyresult2-0.png}} \subfigure[ 2.3GHz] {\includegraphics[width=0.45\linewidth]{5rnnlossyresult2-3.png}} \subfigure[ 2.6GHz] {\includegraphics[width=0.45\linewidth]{8rnnlossyresult2-6.png}} \subfigure[ 2.8GHz] {\includegraphics[width=0.45\linewidth]{2rnnlossyresult2-8.png}} \subfigure[ 2.9GHz] {\includegraphics[width=0.45\linewidth]{11rnnlossyresult2-9.png}} \subfigure[ 3.1GHz] {\includegraphics[width=0.45\linewidth]{2rnnlossyresult3-1.png}} \subfigure[ 3.2GHz] {\includegraphics[width=0.45\linewidth]{1rnnlossyresult3-2.png}} \subfigure[ 3.4GHz] {\includegraphics[width=0.45\linewidth]{7rnnlossyresult3-4.png}} \subfigure[ 3.7GHz] {\includegraphics[width=0.45\linewidth]{7rnnlossyresult3-7.png}} \subfigure[ 4.0GHz] {\includegraphics[width=0.45\linewidth]{3rnnlossyresult4-0.png}} \subfigure[ 4.5GHz] {\includegraphics[width=0.45\linewidth]{2rnnlossyresult4-5.png}} \subfigure[ 5.0GHz] {\includegraphics[width=0.45\linewidth]{9rnnlossyresult5-0.png}} \caption{Lossy scatterers: results of SIPISRLNN on the frequency generalization validation data set.} \label{lossy-rnn-freq} \end{figure} \begin{figure} \centering \subfigure[ 2.0GHz] {\includegraphics[width=0.45\linewidth]{9lossy2_0.png}} \subfigure[2.3GHz ] {\includegraphics[width=0.45\linewidth]{1lossy2_3.png}} \subfigure[2.6GHz ] {\includegraphics[width=0.45\linewidth]{2lossy2_6.png}} \subfigure[ 2.8GHz] {\includegraphics[width=0.45\linewidth]{10lossy2_8.png}} \subfigure[ 2.9GHz] {\includegraphics[width=0.45\linewidth]{10lossy2_9.png}} \subfigure[ 3.1GHz] {\includegraphics[width=0.45\linewidth]{22lossy3_1.png}} \subfigure[3.2GHz ] {\includegraphics[width=0.45\linewidth]{2lossy3_2.png}} \subfigure[3.4GHz ] {\includegraphics[width=0.45\linewidth]{2lossy3_4.png}} \subfigure[3.7GHz ] {\includegraphics[width=0.45\linewidth]{5lossy3_7.png}} \subfigure[ 4.0GHz] {\includegraphics[width=0.45\linewidth]{10lossy4_0.png}} \subfigure[4.5GHz ] {\includegraphics[width=0.45\linewidth]{10lossy4_5.png}} \subfigure[5.0GHz ] {\includegraphics[width=0.45\linewidth]{6lossy5_0.png}} \caption{Lossy scatterers: results of NSIPISRLNN on the frequency generalization validation data set.} \label{lossy-freq} \end{figure} \par NSIPISRLNN shows good generalization ability on both contrast shapes and incident frequencies that unseen during the training process in numerical cases of non-lossy and lossy scatterers. SIPISRLNN only shows good generalization ability in numerical cases of non-lossy scatterers. Furthermore, numerical results also verify that NSIPISRLNN have better computing precisions than SIPIRLNN. \section{Conclusion} In this paper, we propose the physics-informed residual learning as a general framework of 2D electromagnetic forward modeling. PISRL is designed by embedding the iterative methods for solving linear matrix equations. PISRL aims to solve a system of linear matrix equations instead of a specific electromagnetic problem. Therefore, PISRL can be easily expanded to solve different electromagnetic problems which are finally converted into a system of linear matrix equations. In this paper, the generality of PISRL is verified by solving VIEs of non-lossy and lossy scatterers. SIPISRLNN and NSIPISRLNN are proposed by embedding stationary and non-stationary iterative methods. In solving VIEs of non-lossy scatterers, the MSEs of SIPIRLNN and NSIPISRLNN can achieve below $3.152 \times 10^{-4}$ and $4.8925 \times 10^{-7}$ and in solving VIEs of lossy scatterers, the MSEs can be below $1.2775 \times 10^{-4}$ and $1.567 \times 10^{-7}$. The generalization abilities of SIPISRLNN and NSIPISRLNN are also validated on the data sets of unseen contrast shapes and incident frequencies. Numerical results indicate that NSIPISRLNN has better computing precisions and better generalization abilities due to the independent CNN at each iteration. This work is also an example to design and interpret deep neural networks from the perspective of traditional numerical algorithms. This indicates a great possibility to build fast electromagnetic solvers with strong generalization abilities by combining traditional electromagnetic computational methods with deep learning techniques. \bibliographystyle{IEEEtran}

83ed1c9dd666d15af644b4bc8a719969c73f2150

\section{Introduction} Intelligent reflecting surface (IRS) can adaptively tune the phase shifts of its reflected signals by a large number of low-cost reflecting elements integrated on it to achieve high reflect beamforming gain \cite{Wu2021, Renzo2019}. Therefore, leveraging IRS's intelligent reflection has been regarded as a promising way to improve the spectrum and energy efficiency of future wireless communication networks \cite{Wu2019, Pan2020TWC}. On the other hand, wireless power transfer has been introduced into the rapidly developing Internet-of-Things (IoT) networks to resolve the network devices' energy limitation problem. Such networks are also called wireless power communication networks (WPCNs) \cite{Wu2018}, which provide an efficient way to realize sustainable IoT networks. Recently, both the transmit power minimization study in \cite{WuJSAC2020} and the weighted sum-rate maximization study in \cite{Pan2020} have unveiled that IRSs can achieve high wireless power transfer and information transmission efficiency due to the high reflective beamforming gain. Therefore, integrating IRSs with WPCN is a concrete step toward the realization of low-cost sustainable IoT networks. In \cite{Zheng2020}, the throughput maximization of a two-user IRS-assisted WPCN has been investigated, where IRS reflect beamforming, power allocation and time allocation have been jointly optimized. In \cite{Lyu2020}, joint optimization of IRS reflect beamforming and network resource allocation for throughput maximization have been investigated in a multiuser IRS-assisted WPCN. Meanwhile, non-orthogonal multiple access (NOMA) has been proposed to enable the access of massive numbers of devices in IoT networks \cite{Ding2014}. A recent study in \cite{Zeng2020} shows that IRS-assisted NOMA can achieve a higher uplink sum rate than that of the IRS-assisted orthogonal multiple access (OMA). However, the receiver complexity required by using successive interference cancellation (SIC) in NOMA is significantly higher than that of OMA. Furthermore, the error propagation incurred by subtracting the signals of previous users during the process of SIC will constitute a performance-limiting factor when the number of users is large. Furthermore, NOMA may not always outperform OMA (e.g. time division multiple access (TDMA)) in the IRS-assisted communication systems when the IRS can be dynamically switched \cite{Zhu2020}, and OMA may even have higher energy efficiency than that of the NOMA in WPCN \cite{Wu2018}. Therefore, how to achieve a trade-off between complexity and performance in multiuser IRS-assisted WPCNs is still a challenge. \begin{figure}[!t] \centering \includegraphics[width=0.5\textwidth]{fig1.eps} \caption{An IRS-assisted wireless powered hybrid NOMA and TDMA network.} \label{FigSystem} \end{figure} In this letter, we consider an IRS-assisted WPCN, as shown in Fig. \ref{FigSystem}, where multiple users harvest energy and transmit information from/to a base station (BS) with the assistance of an IRS. Different from existing works on IRS-assisted WPCNs, a novel {\it hybrid} NOMA and TDMA scheme is proposed to balance complexity and performance. In particular, for the uplink information transmission, the users are grouped into several clusters such that different clusters of users transmit information at different times using TDMA, while the users in the same cluster transmit information simultaneously using NOMA. The proposed scheme can be regarded as a generalization of NOMA and TDMA, which not only decreases the complexity of implementing SIC but also offers additional degrees of freedom to improve performance. We aim to maximize the throughput of the users' information transmission by optimizing the time allocation among the BS's energy transfer and different user clusters' information transmission, as well as the IRS reflect beamforming during the BS and the users' transmission periods. Although the considered optimization problem is non-convex and challenging, we propose an efficient algorithm to solve it suboptimally by applying the block coordinate ascent (BCA), semidefinite relaxation (SDR), and sequential rank-one constraint relaxation (SROCR) techniques. Simulation results have verified the performance superiority of the proposed algorithm, as compared to some benchmark algorithms, and have shown the impact of user grouping and clustering setup on the throughput performance. \section{System Model and Problem Formulation} As shown in Fig. \ref{FigSystem}, we consider an IRS-assisted WPCN, which includes a single-antenna BS, $M$ single-antenna users, and an IRS with $N$ reflecting elements. The direct links between the BS and the users are blocked by obstacles, so the IRS is deployed to assist the power transfer and information transmission of the network. Note that the proposed algorithm can be applied directly to the case with direct links or with a multi-antenna BS. To provide a performance upper bound benchmark, we assume that the channel state information (CSI) of all links is perfectly estimated and known by the BS \cite{Wu2021}, and the trade-off between channel estimation and performance optimization will be left for future work. In the downlink power transfer (DPT) phase, the BS broadcasts energy signal to the users for a duration $\tau_0$, and the IRS reflects the energy signal with a reflect beamforming matrix $\Phi_0 \triangleq \text{diag}(\mathbf{w}_0)$, where $\mathbf{w}_0 \triangleq [w_{1,0}, \ldots, w_{N,0}]^T$, and the operator $\text{diag}(\mathbf{a})$ forms a diagonal matrix with the elements of a vector $\mathbf{a}$. For $n=1,\ldots,N$, $w_{n,0}=\alpha_{n,0}e^{j\theta_{n,0}}$, and $\alpha_{n,0}\in [0,1]$ and $\theta_{n,0}\in [0,2\pi)$ are the amplitude and phase shift coefficient of the IRS's $n$th reflecting element during the DPT phase, respectively\footnote{We assume that the phase shifts are continuous, and the obtained performance can be used as a benchmark for the performance of the systems with discrete phase shifts. The impact of discrete phase shifts are evaluated in Section IV.}. To achieve the maximum reflecting power gain of the IRS, we assume the reflection amplitudes of all elements are one, i.e., $\alpha_{n,0} = |w_{n,0}| = 1$, $\forall n$ \cite{Wu2019, Cui2019}. In the uplink information transmission (UIT) phase, the users transmit information to the BS with a hybrid NOMA and TDMA scheme, where all $M$ users are grouped into $K$ clusters with $M_k$ users in the $k$th cluster such that $\sum_{k=1}^K M_k = M$. Different clusters of users transmit information at different times in the order of their cluster index, while users in the same cluster transmit simultaneously. During the transmission of the $k$th cluster of users, whose duration is $\tau_k$, the IRS reflects the user signal with a reflect beamforming matrix $\Phi_k \triangleq \text{diag}(\mathbf{w}_k)$, where $\mathbf{w}_k \triangleq [w_{1,k}, \ldots, w_{N,k}]^T$, and similar to $\mathbf{w}_0$, $|w_{n,k}| = 1$, $\forall n,k$, is satisfied. Note that the hybrid NOMA and TDMA scheme becomes TDMA when $K=M$ and $M_k=1, \forall k$, and becomes NOMA when $K=1$ and $M_1=M$ \cite{Wu2018}. We consider a block quasi-static flat-fading channel model by assuming that all channels are constant over each power transfer and information transmission block and may vary over different blocks. The total duration of one block is denoted by $T$. Let $u_{k,m}$ denote the $m$th user in the $k$th cluster, and the channel coefficients from the BS to the IRS, from the IRS to $u_{k,m}$, from $u_{k,m}$ to the IRS, from the IRS to the BS are denoted by $\mathbf{g}_{\text{BS}} \in \mathbb{C}^{N \times 1}$, $\mathbf{g}_{{k,m}} \in \mathbb{C}^{N \times 1}$, $\mathbf{h}_{{k,m}} \in \mathbb{C}^{N \times 1}$, $\mathbf{h}_{\text{BS}} \in \mathbb{C}^{N \times 1}$, respectively, where $\mathbb{C}^{x \times y}$ denotes the set of $x \times y$ complex-valued matrices. Let $P_0$ denote the transmit power of the BS, and then the harvested energy of $u_{k,m}$ can be expressed by \begin{equation} \label{EquHarEnerg} E_{k,m}= \eta \tau_0 P_0 |\mathbf{g}^{H}_{{k,m}} \Phi_0 \mathbf{g}_{\text{BS}}|^2, \end{equation} where $\eta$ denotes the energy harvesting efficiency of the users\footnote{In \eqref{EquHarEnerg}, we adopt a linear energy harvesting model that is commonly used in existing works \cite{WuJSAC2020, Pan2020, Zheng2020} by considering the harvested energy is in the linear regime of the energy harvester. How to extend our work to the scenario of non-linear energy harvester is left for future work.}. With the harvested energy, the transmit power of $u_{k,m}$ is $P_{k,m} = E_{k,m}/\tau_k$. Thus, the total throughput of the $k$th user cluster in bits/Hertz (bits/Hz) can be written as \cite{Wu2018} \begin{equation} \label{Equ_throu_cluster} R_k = \tau_k \log_2 \Bigg( 1 + \frac{\sum\limits_{m=1}^{M_k} P_{k,m}|\mathbf{h}^{H}_{\text{BS}} \Phi_k \mathbf{h}_{k,m}|^2 }{\sigma^2} \Bigg), \end{equation} where $\sigma^2$ denotes the noise power at the BS. Accordingly, the throughput of all users' information transmission in bits/Hz is $R = \sum_{k=1}^K R_k $. It is observed from \eqref{EquHarEnerg}--\eqref{Equ_throu_cluster} that the durations of power transfer and information transmission and the IRS reflect beamforming have a great impact on the value of $R$. In this letter, we aim to maximize the throughput $R$ by jointly optimizing the time allocation among the BS's energy transfer and different user clusters' information transmission, i.e., $\tau_0$ and $\{\tau_k\}$, as well as the IRS reflect beamforming at different time, i.e., $\mathbf{w}_0$ and $\{\mathbf{w}_k\}$. The considered problem can be formulated as \begin{subequations} \label{EquOriProb} \begin{align} \text{(P1)}: \max \limits_{\tau_0,\{\tau_k\},\mathbf{w}_0,\{ \mathbf{w}_k\}} & \sum_{k=1}^K R_k \label{EquObjFun} \\ \textrm{s.t.}\; \quad \quad & |w_{n,0}| =1, \; \forall n, \label{EquIRSCon0} \\ &|w_{n,k}| =1, \; \forall n, k, \label{EquIRSConk} \\ &\tau_0 + \sum\limits_{k=1}^K \tau_k \leq T, \label{EquTotalDurCon} \\ & \tau_0 \geq 0, \; \tau_k \geq 0, \; \forall k. \label{EquDurPos} \end{align} \end{subequations} In problem (P1), the optimization variables $\tau_0$, $\{\tau_k\}$, $\mathbf{w}_{0}$, and $\{\mathbf{w}_{\text{k}}\}$ are intricately coupled in the objective function, and the objective function is not jointly concave with respect to them. Furthermore, the left-hand-sides (LHSs) of the equality constraints \eqref{EquIRSCon0} and \eqref{EquIRSConk} are non-linear functions, so (P1) is a highly non-convex optimization problem whose optimal solution is difficult to obtain in polynomial time. Nevertheless, we propose an efficient algorithm to obtain its suboptimal solution in the next section. \section{Proposed Algorithm} It is observed that in (P1), \eqref{EquIRSCon0} is constraints on $\mathbf{w}_{0}$, \eqref{EquIRSConk} is constraints on $\{\mathbf{w}_{\text{k}}\}$, and \eqref{EquTotalDurCon} and \eqref{EquDurPos} are constraints on $\tau_0$ and $\{\tau_k\}$. This motivates us to resolve the optimization variable coupling in the objective function of (P1) by using the BCA technique, which divides the optimization variables of (P1) into three blocks, i.e., $\mathbf{w}_{0}$, $\{\mathbf{w}_{k}\}$, and $( \tau_0, \{\tau_k\})$, and alternately optimizes one block with the others fixed until achieving convergence. Specifically, the following three sub-problems need to be solved: sub-problem 1 optimizes $\mathbf{w}_{0}$ under given $\{\mathbf{w}_{k}\}$, $\tau_0$ and $\{\tau_k\}$, sub-problem 2 optimizes $\{\mathbf{w}_{k}\}$ under given $\mathbf{w}_{0}$, $\tau_0$ and $\{\tau_k\}$, and sub-problem 3 optimizes $\tau_0$ and $\{\tau_k\}$ under given $\mathbf{w}_{0}$ and $\{\mathbf{w}_{k}\}$. The procedures of solving all sub-problems and the overall proposed algorithm are presented below. \subsection{Optimizing $\mathbf{w}_{0}$ with Given $\{\mathbf{w}_{\text{k}}\}$, $\tau_0$ and $\{\tau_k\}$} By letting $\lambda_k=\frac{\eta\tau_0P_0}{\tau_k\sigma^2}$, $b_{k,m} = |\mathbf{h}^{H}_{\text{BS}} \Phi_k \mathbf{h}_{k,m}|^2$, $\mathbf{\hat{g}}_{k,m} = \mathbf{g}_{k,m}\odot\mathbf{g}_{\text{BS}}$, where $\odot$ denotes the Hadamard product operator, sub-problem 1 can be formulated as \begin{subequations} \begin{align} \text{(P2)}: \max \limits_{\mathbf{w}_0}& \sum_{k=1}^K\tau_k\log_2 \bigg( 1 + \lambda_k \sum_{m=1}^{M_{k}}|\mathbf{w}_{0}^{H} \mathbf{\hat{g}}_{k,m}|^2b_{k,m}\bigg) \label{EquOriProbObj} \\ \textrm{s.t.}\;\;&|w_{n,0}| = 1, \; \forall n. \label{EquConwInq2} \end{align} \end{subequations} Problem (P2) is difficult to solve since its objective function is non-concave with respect to $\mathbf{w}_0$, and the LHSs of \eqref{EquConwInq2} is non-linear. We apply the SDR technique to overcome such a difficulty. First, we write the summation term in the logarithmic function of \eqref{EquOriProbObj} into the following form: \begin{equation} \sum_{m=1}^{M_{k}}|\mathbf{w}_{0}^{H} \mathbf{\hat{g}}_{k,m}|^2b_{k,m}=\mathbf{w}_{0}^{H} \mathbf{G}_{k}\mathbf{w}_{0} = \text{tr}(\mathbf{G}_{k}\mathbf{w}_{0}\mathbf{w}_{0}^{H} ), \end{equation} where $\mathbf{G}_{k}=\sum_{m=1}^{M_{k}}b_{k,m}\mathbf{\hat{g}}_{k,m}\mathbf{\hat{g}}^{H}_{k,m}$, and $\text{tr}(\cdot)$ denotes the trace operator. Besides, we write constraint \eqref{EquConwInq2} into \begin{equation} \mathbf{w}_{0}^{H} \mathbf{B}_n \mathbf{w}_{0} = \text{tr}(\mathbf{B}_{n}\mathbf{w}_{0}\mathbf{w}_{0}^{H} ) = 1, \; \forall n, \end{equation} where the matrix $\mathbf{B}_n$ is constructed by letting its $(i,j)$th element, denoted by $[\mathbf{B}_n]_{i,j}$, satisfy \begin{equation} [\mathbf{B}_n]_{i,j} = \begin{cases} 1 & i=j=n \\ 0 & \text{otherwise}. \end{cases} \end{equation} Then, we let $\mathbf{W}_{0} \triangleq \mathbf{w}_{0} \mathbf{w}_{0}^{H}$ and re-express problem (P2) as \begin{subequations} \begin{align} \text{(P3)}: \max \limits_{\mathbf{W}_0 \succeq 0 } &\sum_{k=1}^K\tau_k\log_2 \bigg( 1 + \lambda_k \text{tr}(\mathbf{G}_{k}\mathbf{W}_{0})\bigg) \\ \textrm{s.t.}\;\;&\text{tr}(\mathbf{B}_{n}\mathbf{W}_{0})=1, \; \forall n \label{EquConModuOne} \\ &\text{rank}(\mathbf{W}_{0}) =1. \label{Rank2_one} \end{align} \end{subequations} Although problem (P3) is NP-hard due to the rank-one constraint \eqref{Rank2_one}, we propose an SROCR-based algorithm to solve it. Unlike the existing approach for the (P3)-type problems \cite{Wu2019, Cui2019}, which first solves its relaxed problem that ignores the rank-one constraint and then constructs a feasible solution based on the solution to the relaxed problem by using e.g., Gaussian randomization method, the SROCR-based algorithm relaxes the rank-one constraint gradually over iterations and can find a locally optimal solution to (P3) eventually \cite{Cao2017}. Suppose that $\mathbf{W}_{0}^{(i)}$ is the obtained solution in iteration $i$. In iteration $i+1$, the proposed algorithm constructs and solves the following relaxed problem: \begin{subequations} \begin{align} \text{(P4)}: & \max \limits_{\mathbf{W}_0 \succeq 0 } \sum_{k=1}^K\tau_k\log_2 \bigg( 1 + \lambda_k \text{tr}(\mathbf{G}_{k}\mathbf{W}_{0})\bigg) \\ \textrm{s.t.}\;& \eqref{EquConModuOne}, \\ & \mathbf{u}_{\max}(\mathbf{W}_{0}^{(i)})^H \mathbf{W}_{0} \mathbf{u}_{\max}(\mathbf{W}_{0}^{(i)}) \geq v^{(i)}\text{tr}(\mathbf{W}_{0}), \label{SROCR} \end{align} \end{subequations} where $\mathbf{u}_{\max}(\mathbf{W}_{0}^{(i)})$ denotes the eigenvector corresponding to the largest eigenvalue of $\mathbf{W}_{0}^{(i)}$, and $v^{(i)}$ denotes an introduced relaxation parameter. Problem (P4) is a semidefinite programming (SDP) problem and can be efficiently solved by the interior-point method. The relaxation from \eqref{Rank2_one} to \eqref{SROCR} becomes tighter and tighter, as $v^{(i)}$ increases from $0$ to $1$ over iterations. When $v^{(i)}=1$, \eqref{SROCR} is equivalent to \eqref{Rank2_one}, and thus the solution obtained by solving (P4) is a solution to (P3). The detail of the proposed algorithm is presented in Algorithm 1, where $\lambda_{\max}(\mathbf{W}_{0})$ denotes the largest eigenvalue of $\mathbf{W}_{0}$, $g_0(\mathbf{W}_{0})$ denotes the objective value of (P4) with solution $\mathbf{W}_{0}$, and $\epsilon_1$ and $\epsilon_2$ are thresholds indicating the convergence accuracy. After obtaining the solution to (P3), denoted by $\mathbf{W}_{0}^*$, the solution to sub-problem 1 can be obtained by a rank-one decomposition on $\mathbf{W}_{0}^*$. \begin{algorithm}[!t] \caption{SROCR-Based Algorithm for Problem (P3).} \begin{algorithmic}[1] \STATE \textbf{Initialization:} Solve problem (P4) when $v^{(0)}=0$ and denote the solution by $\mathbf{W}_{0}^{(0)}$. Define an initial step size $\delta^{(0)} \in \big(0, 1 - \frac{ \lambda_{\max}(\mathbf{W}_{0}^{(0)}) }{\text{tr}(\mathbf{W}_{0}^{(0)})} \big]$ and set $i=0$. \REPEAT \STATE Solve problem (P4) under given $\{v^{(i)},\mathbf{W}_{0}^{(i)}\}$. \IF{Problem (P4) is solvable} \STATE Denote the solution by $\mathbf{W}_{0}^{(i+1)}$ and set $\delta^{(i+1)}=\delta^{(i)}$. \ELSE \STATE Set $\mathbf{W}_{0}^{(i+1)}=\mathbf{W}_{0}^{(i)}$ and $\delta^{(i+1)}=\delta^{(i)}/2$. \ENDIF \STATE Set $v^{(i+1)} = \min \big(1, \frac{ \lambda_{\max}(\mathbf{W}_{0}^{(i+1)}) }{\text{tr}(\mathbf{W}_{0}^{(i+1)})} +\delta^{(i+1)}\big)$. \STATE Set $i=i+1$. \UNTIL $v^{(i-1)} \geq \epsilon_1$ \& $|g_0(\mathbf{W}_{0}^{(i)}) - g_0(\mathbf{W}_{0}^{(i-1)})| \leq \epsilon_2$. \end{algorithmic} \end{algorithm} \subsection{Optimizing $\{\mathbf{w}_{\text{k}}\}$ with Given $\mathbf{w}_{0}$, $\tau_0$ and $\{\tau_k\}$ } By letting $c_{k,m} = |\mathbf{w}_{0}^{H} \mathbf{\hat{g}}_{k,m}|^2$ and $\mathbf{\hat{h}}_{k,m} = \mathbf{h}_{\text{BS}}\odot\mathbf{h}_{k,m}$, sub-problem 2 can be formulated as \begin{subequations} \begin{align} \text{(P5)}: \max \limits_{\{ \mathbf{w}_k \} }& \sum_{k=1}^K\tau_k\log_2 \bigg( 1 + \lambda_k \sum_{m=1}^{M_{k}}|\mathbf{w}^{H}_{k}\mathbf{\hat{h}}_{k,m}|^2c_{k,m}\bigg) \\ \textrm{s.t.}\;\;&|w_{n,k}| = 1, \; \forall n,k. \end{align} \end{subequations} It is worth noting that the objective function and constraints of problem (P5) can be decoupled with respect to different user clusters, which means that it can be solved by optimizing $\mathbf{w}_1, \ldots, \mathbf{w}_K$ independently. In particular, let $\mathbf{H}_{k}=\sum_{m=1}^{M_{k}}c_{k,m}\mathbf{\hat{h}}_{k,m}\mathbf{\hat{h}}^{H}_{k,m}$ and $\mathbf{W}_{k} \triangleq \mathbf{w}_{k}\mathbf{w}^{H}_{k}$ and the problem optimizing $\mathbf{w}_k$ can be reduced to \begin{subequations} \begin{align} \text{(P6)}: \max \limits_{ \mathbf{W}_k \succeq 0 } \; \;&\text{tr}(\mathbf{H}_{k}\mathbf{W}_{k}) \\ \textrm{s.t.}\;\;&\text{tr}(\mathbf{B}_{n}\mathbf{W}_{k})=1, \; \forall n, \\ &\text{rank}(\mathbf{W}_{k}) =1 \label{EquRankWk}. \end{align} \end{subequations} Problem (P6) can be solved similarly by an SROCR-based algorithm like Algorithm 1, and the detail is omitted here for brevity. \subsection{Optimizing $\tau_0$ and $\{\tau_k\}$ with Given $\mathbf{w}_{0}$ and $\{\mathbf{w}_{\text{k}}\}$} Sub-problem 3 can be formulated as \begin{subequations} \begin{align} \text{(P7)} : \max \limits_{\tau_0,\{\tau_k\}} & \sum\limits_{k=1}^K\tau_k\log_2 \Bigg( 1 + \frac{ \eta\tau_0 P_0\sum\limits_{m=1}^{M_{k}} b_{k,m} c_{k,m}}{ \tau_k \sigma^2} \Bigg) \\ \text{s.t.}\;\;& \eqref{EquTotalDurCon}, \eqref{EquDurPos}. \end{align} \end{subequations} Problem (P7) is a convex optimization problem, so by analyzing its Karush-Kuhn-Tucker (KKT) conditions, its optimal solution can be obtained as \begin{equation} \label{EquTauSol} \tau_0^*= \frac{T}{1+\sum\nolimits_{k=1}^K\frac{\gamma_k}{ x_k^*}}, \; \; \tau_k^*=\frac{\gamma_k}{ x_k^*}\tau_0^*, \; \forall k, \end{equation} where $\gamma_k= \frac{ \eta P_0 }{ \sigma^2 } \sum\nolimits_{m=1}^{M_{k}} b_{k,m} c_{k,m} $ and $x_k^*$ is the solution of the following equation \begin{equation} \label{EquEqux} \log_2(1+x_k)-\frac{x_k\log_2(e)}{1+x_k} -\sum\limits_{k=1}^K\frac{\gamma_k \log_2(e)}{1+x_k} = 0. \end{equation} Note that $x_k^*$ can be obtained by a numerical method such as the bisection method. The procedure of obtaining \eqref{EquTauSol} is similar to that in \cite{Wu2018} and thus omitted here for simplicity. \begin{algorithm}[!t] \caption{Proposed IRS Reflect Beamforming and Time Allocation Algorithm.} \begin{algorithmic}[1] \STATE \textbf{Initialization:} Set initial value for $\mathbf{w}_{0}$, $\{ \mathbf{w}_k \}$, $\tau_0$, and $\{ \tau_{k} \}$. Set $l=0$ and $R^{(0)} = f ( \mathbf{w}_{0}, \{ \mathbf{w}_{k} \}, \tau_0, \{ \tau_{k} \} )$. \REPEAT \STATE Set $l=l+1$. \STATE Under given $\{\mathbf{w}_k\}$, $\tau_0$ and $\{ \tau_{k} \}$, update $\mathbf{w}_{0}$ by solving problem (P3). \STATE Under given $\mathbf{w}_0$, $\tau_0$ and $\{ \tau_{k} \}$, update $\{\mathbf{w}_k\}$ by solving problem (P6). \STATE Under given $\mathbf{w}_{0}$ and $\{ \mathbf{w}_k \}$, update $\tau_0$ and $\{ \tau_{k} \}$ by solving problem (P7). \STATE Set $R^{(l)} = f ( \mathbf{w}_{0}, \{ \mathbf{w}_{k} \}, \tau_0, \{ \tau_{k} \} )$. \UNTIL {$\frac{R^{(l)} - R^{(l-1)}}{R^{(l)}} < \epsilon$.} \end{algorithmic} \end{algorithm} \subsection{Overall Algorithm and Complexity Analysis} The overall algorithm for problem (P1) is present in Algorithm 2, where $f( \mathbf{w}_{0}, \{ \mathbf{w}_{k} \}, \tau_0, \{ \tau_{k} \} )$ denotes the objective value of (P1) with solution $\mathbf{w}_{0}$, $\{ \mathbf{w}_{k} \}$, $\tau_0$, and $\{ \tau_{k} \}$ and $\epsilon$ is a small positive threshold indicating the convergence accuracy. The algorithm solves sub-problems 1--3 in steps 4--6, respectively. Since the obtained solutions to sub-problems 1 and 2 are locally optimal and that to sub-problem 3 is globally optimal, the objective value of (P1) over iterations is non-decreasing. Furthermore, it must be upper-bounded by a finite value, so Algorithm 2 will converge. The main complexities are in executing steps 4 and 5, for which the complexities are $\mathcal{O}(N^{4.5})$ and $\mathcal{O}(KN^{4.5})$, respectively. Thus, the complexity of Algorithm 2 is $\mathcal{O}( (K+1) N^{4.5} )$. Note that Algorithm 2 can be executed at the BS, and the obtained result can be sent to the IRS and users over a reliable control link \cite{Wu2021}. \section{Simulation Results} This section provides numerical results to validate the effectiveness of the proposed algorithm. The simulation parameters are set as follows. There are $M=12$ users, and the total block length is set as $T=0.1$ s. The channel coefficients are set as $\mathbf{g}_{\text{BS}} =\sqrt{\zeta_0 ( d_0 / d_{\text{BI}} )^{ \alpha_{\text{BI}} } }\mathbf{\tilde{g}}_{\text{BS}}$, $\mathbf{g}_{k,m} =\sqrt{\zeta_0 ( d_0 / d_{k,m} )^{ \alpha_{k,m}}}\mathbf{\tilde{g}}_{k,m}$, $\mathbf{h}_{\text{BS}} =\sqrt{\zeta_0 ( d_0 / d_{\text{BI}} )^{ \alpha_{\text{BI}}}}\mathbf{\tilde{h}}_{\text{BS}}$, and $\mathbf{h}_{k,m} =\sqrt{\zeta_0 ( d_0 / d_{k,m} )^{ \alpha_{k,m}}}\mathbf{\tilde{h}}_{k,m}$, where $\zeta_0=-30$ dB denotes the path loss at the reference distance $d_0= 1$ m, $d_{\text{BI}}$ and $d_{k,m}$ denote the distance between the BS and the IRS and that between $u_{k,m}$ and the IRS, respectively, $\alpha_{\text{BI}}=2.2$ and $\alpha_{k,m}=2.5$ are the path loss exponents, and $\mathbf{\tilde{g}}_{\text{BS}}$, $\mathbf{\tilde{g}}_{k,m}$, $\mathbf{\tilde{h}}_{\text{BS}}$, $\mathbf{\tilde{h}}_{k,m}$ are small-scale fading components modeled by the Rician fading model with Rician factors setting as $1$ \cite{Cui2019}. The other parameters are set as $d_{\text{BI}}=1$ m, $\sigma^2= -110$ dBm, $\epsilon_1=0.95$, $\epsilon_2=10^{-3}$, and $\epsilon = 10^{-3}$. We denote the proposed algorithm by `Optimized IRS w/ time allocation (TA)', and compare it to the following benchmark algorithms. 1) `Optimized IRS w/o TA': it fixes the durations as $\tau_0 = \tau_k = T/(K+1)$, $\forall k$, and optimizes IRS reflect beamforming by using steps 4-5 of Algorithm 2. 2) `Random IRS w/ TA': it uses random IRS reflect beamforming and optimizes $\tau_0$ and $\{\tau_k\}$ by using step 6 of Algorithm 2. 3) `Random IRS w/o TA': it uses random IRS reflect beamforming and sets $\tau_0 = \tau_k = T/(K+1)$, $\forall k$. 4) `Optimized IRS w/ TA (same IRS)': it uses the same IRS reflect beamforming over the whole block, i.e., $\mathbf{w}_0=\mathbf{w}_k$, $\forall k$, and optimizes IRS beamforming and time allocation by an iterative algorithm similar to Algorithm 2. 5) `Upper Bound': it constructs relaxed problems of (P3) and (P6) by dropping the rank-one constraints \eqref{Rank2_one} and \eqref{EquRankWk} from them, respectively, and solves the relaxed problems of (P3) and (P6), as well as (P7) alternatively to obtain a throughput upper bound of the proposed algorithm. 6) `Discrete phase': it is similar to the `Optimized IRS w/ TA' algorithm except that the phase shift of each IRS reflecting element is discretized into $2^b$ levels over the interval $[0, 2\pi)$ uniformly, where $b$ denotes the control bit number for each reflecting element. The following results are obtained by averaging the performances of 1000 random channel realizations. \begin{figure*}[!t] \centering \subfloat[Throughputs versus $N$.]{\includegraphics[width=0.23\textwidth]{fig2a.eps} \label{FigThr_N}} \hfil \subfloat[Throughputs versus $d_{k,m}$.]{\includegraphics[width=0.23\textwidth]{fig2b.eps} \label{FigThr_D}} \hfil \subfloat[Impact of user grouping.]{\includegraphics[width=0.23\textwidth]{fig2c.eps} \label{FigThr_Cluster}} \hfil \subfloat[Impact of number of user groups.]{\includegraphics[width=0.23\textwidth]{fig2d.eps} \label{FigThr_Setup}} \caption{Simulation results.} \label{fig_sim} \end{figure*} First, we randomly group all users into $K=3$ clusters with $M_k = 4$ users in each cluster, and show the throughputs of different algorithms versus the IRS's reflecting element number $N$ and the distance between the IRS and the users $d_{k,m}$ when $P_0=40$ \text{dBm} in Figs. 2(a) and 2(b), respectively. $d_{k,m}=5$ m in Fig. 2(a), and $N=64$ in Fig. 2(b). It is observed that the throughput of the proposed algorithm is very close to its upper bound and is significantly higher than that of the other benchmark algorithms. The proposed algorithm outperforms the `Optimized IRS w/ TA (same IRS)' algorithm, because it has more degrees of freedom for optimization than the latter. Besides, the proposed algorithm with continuous IRS phase shifts always outperforms its `Discrete phase' counterpart with $b=1, 2, 3$ control bits, and the performance gap between them decreases as $b$ grows. This is because signal misalignment caused by low-resolution phase shift degrades system performance. Furthermore, the throughput gain of the proposed algorithm over the 'Optimized IRS w/o TA' algorithm and that of the `Random IRS w/ TA' algorithm over the `Random IRS w/o TA' algorithm demonstrate the effectiveness of time allocation in improving throughput. Besides, the throughput gain of the proposed algorithm over the `Random IRS w/ TA' algorithm demonstrates the effectiveness of IRS reflecting beamforming in improving throughput. Thus, the proposed algorithm can fully reap the gains brought by IRS reflect beamforming and time allocation. Fig. 2(c) shows the throughput of the proposed algorithm under three different user grouping schemes versus $N$ under the same condition with Fig. 2(a) except that $d_{k,m}$ is random in $[5,15]$ m. To describe the user grouping schemes, we rename and order the users as $u_1, \ldots, u_N$ such that $||\mathbf{h}_1|| \geq \ldots \geq ||\mathbf{h}_N||$, where $\mathbf{h}_n$ denotes the channel coefficient from user $n$ to the IRS, $n=1,\ldots,N$. The first user grouping scheme is called large-channel-strength-difference (LCSD) scheme, which lets the user-IRS channel strength differences among the users in the same cluster as large as possible, and the user clusters obtained by this scheme can be expressed as $\{ u_1, u_4, u_7, u_{10} \}$, $\{ u_2, u_5, u_8, u_{11} \}$, and $\{ u_3, u_6, u_9, u_{12} \}$. The second scheme is called small-channel-strength-difference (SCSD) scheme, which lets the user-IRS channel strength differences among the users in the same cluster as small as possible, thus its obtained user clusters can be expressed as $\{ u_1, u_2, u_3, u_{4} \}$, $\{ u_5, u_6, u_7, u_{8} \}$, and $\{ u_9, u_{10}, u_{11}, u_{12} \}$. The third scheme is called random user grouping scheme, which groups the users randomly. It is observed from Fig. 2(c) that the LCSD scheme achieves the highest throughput while the SCSD scheme achieves the lowest. This is because, as compared to the other two schemes, the LCSD scheme achieves the smallest overall channel strength difference among different user clusters and thus can achieve the highest performance gain via IRS reflect beamforming and time allocation optimization. Next, we show the throughput of the proposed algorithm under three different user clustering setups versus $N$ when $P_0=40$ dBm and $d_{k,m}=5$ m in Fig. 2(d), where all users are randomly grouped into $K=12/3/1$ clusters with $M_k = 1/4/12$ user(s) in each cluster. It is observed that the more clusters that the users are grouped, the higher throughput can be achieved. This is because, with more user clusters, there are more degrees of freedom for both IRS reflect beamforming and time allocation optimization, which results in higher throughput. \section{Conclusion} This letter studied an IRS-assisted wireless powered hybrid NOMA and TDMA network, where IRS reflect beamforming and time allocation are jointly designed to maximize the throughput of the network. Based on the BCA, SDR, and SROCR techniques, an efficient algorithm was proposed to solve the considered challenging problem. Simulation results have shown that the proposed algorithm achieves significantly higher throughput than other benchmark algorithms. It is also found that grouping users into more clusters brings more degrees of freedom for the joint IRS beamforming and time allocation and thus can achieve better throughput performance.

62532159d4a7afe455e7e81e19a5aa4b8ee0fd83

\section{Introduction} Let $q = p^k$ be a power of a prime number $p$, let $\GF{q}$ be a finite field with $q$ elements, and let $\GF{q}[x]$ be the ring of polynomials over $\GF{q}$. We call $f(x) \in \GF{q}[x]$ a \emph{permutation polynomial} (PP) over $\GF{q}$ if its associated polynomial mapping $f:c\mapsto f(c)$ from $\GF{q}$ to itself is a bijection. It is well known that every permutation of $\GF{q}$ can be expressed as a permutation polynomial over $\GF{q}$, of degree at most $q-1$. Permutation polynomials over finite fields have been a hot topic of study for many years, partially due to their applications in coding theory \cite{TIT:DingH13,FFA:Laigle07}, cryptography\cite{ CRYPTO:LidlM83, CACM:RivestSA78, EL:SchwenkH98}, combinatorial designs \cite{JCT:DingY06}, and other areas of mathematics and engineering. More background material and information about properties, constructions, and applications of permutation polynomials may be found in \cite[Chapter 7]{FFields:LidlN97} and \cite[Chapter 8]{HandbookFFields:MullenP13}. For a detailed survey of open questions and recent results we refer the reader to \cite{FFA:Hou15} and \cite{IndexSurvey}. Recently, Akbary, Ghioca and Wang have derived the following useful criterion to study permutation functions on finite sets. It first appears in \cite{AGW} and is further developed in \cite{DCC:LiQW17, FFA:YuanD11, FFA:YuanD14, DCC:ZhengYP16}, among others. \begin{lemma}[The AGW Criterion]\label{lemma:AGWCriterion} Let $A$, $S$ and $\bar{S}$ be finite sets with $\order{S} = \order{\bar{S}}$, and let $f:A\rightarrow A$, $\bar{f}:S\rightarrow \bar{S}$, $\lambda:A\rightarrow S$, and $\bar{\lambda}:A\rightarrow \bar{S}$ be maps such that $\bar{\lambda} \circ f=\bar{f} \circ\lambda$ (see the following commutative diagram). \[ \xymatrix{ A \ar[r]^{f}\ar[d]^{\lambda} & A \ar[d]^{\bar{\lambda}}\\ S \ar[r]^{\bar{f}} & \bar{S} } \] If both $\lambda$ and $\bar{\lambda}$ are surjective, then the following statements are equivalent: \begin{itemize} \item $f$ is a bijection from $A$ to $A$ (a permutation over $A$); \item $\bar{f}$ is a bijection from $S$ to $\bar{S}$ and $f$ is injective on $\lambda^{-1}(s)$ for each $s\in S$. \end{itemize} \end{lemma} The importance of the AGW criterion depends on that it can be used not only to explain some previous constructions of PPs, but also to construct numerous new classes. For example, Akbary, Ghoica and Wang \cite{AGW} applied their approach into different cases (e.g., multiplicative group case, elliptic curve case, additive group case) and obtained many interesting results. In the additive group case, for any polynomial $g \in \mathbb{F}_{q^n}[x]$, any additive polynomials $\varphi, \psi,\bar{\psi}\in\mathbb{F}_{q^n}[x]$ satisfying $\varphi \circ \psi = \bar{\psi} \circ \varphi$ and $\#\psi(\mathbb{F}_{q^n})=\#\bar{\psi}(\mathbb{F}_{q^n})$, and any polynomial $h \in \mathbb{F}_{q^n}[x]$ such that $h(\psi(\mathbb{F}_{q^n}))\subseteq \mathbb{F}_q^*$, the permutation polynomials of the form $f(x) := h(\psi(x)) \varphi (x) + g(\psi(x))$ over $\mathbb{F}_{q^n}$ were characterized. One of two necessary and sufficient conditions requires $\ker(\varphi) \cap \ker(\psi) = \{0\}$, equivalently, $\varphi$ induces a bijection betwenen $\ker(\psi)$ and $\ker(\bar{\psi})$. Later on, Yuan and Ding \cite{FFA:YuanD11} extended their study to PPs with the form $f(x)=g(B(x))+\sum_{i=1}^{r}\left(L_i(x)+\delta_i\right)h_i(B(x))$ over $\mathbb{F}_{q^n}$, where $B(x), L_1(x), \ldots, L_r(x) \in \mathbb{F}_q[x]$ are $q$-polynomials, $g(x) \in \mathbb{F}_{q^n}[x]$, $h_1(x), \ldots, h_r(x) \in \mathbb{F}_q[x]$, and $\delta_1, \ldots, \delta_r \in \mathbb{F}_{q^n}$ such that $B(\delta_i) \in \mathbb{F}_q$ and $h_i(B(\mathbb{F}_{q^n}))\subseteq \mathbb{F}_q$. In this case, the condition $\ker(B) \cap \ker(\sum_{i=1}^r L_i h_i(y)) = \{0\}$ for each $y \in B(\mathbb{F}_{q^n})$ reduced to $\gcd(\sum_{i=1}^r l_i(x) h_i(y), b(x)) =1$ for any $y \in \mathbb{F}_q$, where $l_i(x)$ and $b(x)$ are conventional $q$-associate of $L_i(x)$ and $B(x)$. Several interesting classes of PPs of the form $L(x) + g(x^q -x + \delta) \in \mathbb{F}_{q^n}[x]$, where $L(x) $ is a linearized polynomial and $g(x)^q = g(x)$ were also given in \cite{FFA:YuanD14}. Further generic applications of AGW criterion over $\mathbb{F}_{q^2}$ can be found in Zheng, Yuan and Pei \cite{DCC:ZhengYP16} and Li, Qu and Wang \cite{DCC:LiQW17}. In this paper we focus on the subclass of permutation polynomials of $\mathbb{F}_{q^n}$ with the form \begin{equation}\label{eq:main}P(x)=f(L(x))+k(L(x))\cdot M(x),\end{equation} where $f\in \mathbb{F}_{q^n}[x]$, and $L, M\in \mathbb{F}_{q^n}[x]$ are $q$-linearized polynomials with coefficients in $\mathbb{F}_q$. We further relate the problem of constructing PPs of $\mathbb{F}_{q^n}$ to the problem of factorizing $x^n-1$ in $\mathbb{F}_q[x]$. Using explicit factors of $x^n-1$ in $\mathbb{F}_q[x]$, we can construct PPs of $\mathbb{F}_{q^n}$ by taking their linearized $q$-associates for $L(x)$. In the special case that $L(x)=\mathrm{Tr}_{q^n/q}(x):=x^{q^{n-1}}+\cdots+x$ is the trace polynomial, we provide more explicit results concerning the construction of permutations and their inverses, and also the construction of complete permutation polynomials (i.e., complete mappings). Most notably, Theorem~\ref{thm:const} provides a general method to produce permutations of $\mathbb{F}_{q^n}$ from permutations of $\mathbb{F}_q$, by simply solving a system of equations of the form $\mathrm{Tr}_{q^n/q}(x_i)=y_i$. In fact, the same method can be applied to construct complete mappings of $\mathbb{F}_{q^n}$ from complete mappings of $\mathbb{F}_q$. We further show that the results on the trace case are extended to polynomials $L(x)=\sum_{i=0}^{n-1}a^ix^{q^{n-1-i}}$ with $a\in \mathbb{F}_q$ and $a^n=1$. It is worthy of mention that past works have considered permutation polynomials like in Eq.~\eqref{eq:main} with $L(x)$ being the trace map~\cite{LWZ, TZJ, WY, ZTT}. In some cases, the polynomial $M$ can even be replaced by a more general one, but the permutation criteria become less explicit. Nevertheless, the polynomial $f$ is rather too restricted: in particular, $f$ has at most $3$ nonzero coefficients. The paper is organized as follows. In Section 2 we provide background material and present some initial results, including a criterion for when a polynomial given by Eq.~\eqref{eq:main} permutes $\mathbb{F}_{q^n}$. In Section 3 we specialize our study to permutations arising from the trace function and discuss the construction of permutation and complete permutation polynomials, and their inverses. Finally, in Section 4, we provide concluding remarks and propose directions for future research. \section{Some preliminary results} Throughout this paper, $\mathbb{F}_q$ denotes the finite field with $q$ elements and $\overline{\mathbb{F}}_q$ denotes its algebraic closure. A $q$-linearized polynomial is a polynomial of the form $\sum_{i=0}^ma_ix^{q^i}$, where $a_i\in \overline{\mathbb{F}}_q$. From the well-known identity $(a+b)^q=a^q+b^q$, we observe that $q$-linearized polynomials with coefficients in $\mathbb{F}_{q^t}$ induce $\mathbb{F}_q$-linear maps over every finite extension of $\mathbb{F}_{q^t}$. \subsection{Background material} \begin{definition} For $f\in \mathbb{F}_q[x]$ with $f(x)=\sum_{i=0}^ma_ix^i$, the linearized $q$-associate of $f$ is the polynomial $L_f(x)=\sum_{i=0}^ma_ix^{q^i}$. \end{definition} The following lemma provides some properties of the $q$-associates through basic operations. Its proof follows by direct verification so we omit details. \begin{lemma}\label{lem:q-associate} For $f, g\in \mathbb{F}_q[x]$, we have that $L_{f+g}(x)=L_f(x)+L_g(x)$, $L_f(L_g(x))=L_{fg}(x)$ and $\gcd(L_f(x), L_g(x))=L_{\gcd(f, g)}(x)$. \end{lemma} We obtain the following corollary. \begin{corollary}\label{cor:roots} For a positive integer $n$ and a nonzero polynomial $f\in \mathbb{F}_q[x]$, the equation $L_f(y)=0$ has $q^{\deg(F)}$ solutions over $\mathbb{F}_{q^n}$, where $F(x)=\gcd(f(x), x^n-1)$. Moreover, such solutions comprise an $\mathbb{F}_q$-vector space and, for a monic divisor $g\in \mathbb{F}_q[x]$ of $x^n-1$, the image set $L_g(\mathbb{F}_{q^n})$ equals the set of roots of $L_G(x)$, where $G(x)=\frac{x^n-1}{g(x)}$. \end{corollary} \begin{proof} We observe that, for $h(x)=x^n-1$, the roots of $L_{h}(x)=x^{q^n}-x$ are simple and comprise the field $\mathbb{F}_{q^n}$. From Lemma~\ref{lem:q-associate}, we have that $\gcd(L_f(x), x^{q^n}-x)=L_F(x)$. In particular, $L_F$ splits completely over $\mathbb{F}_{q^n}$ into distinct linear factors. Therefore, $L_f(y)=0$ has $\deg(L_F)=q^{\deg(F)}$ solutions over $\mathbb{F}_{q^n}$. As $L_F$ is $q$-linearized, the evaluation map $c\mapsto L_f(c)$ is an $\mathbb{F}_q$-linear map over $\mathbb{F}_{q^n}$. Therefore, the set of solutions of $L_f(y)=0$ comprise an $\mathbb{F}_q$-vector space. For the last statement, observe that $L_G(L_g(y))=y^{q^n}-y=0$ for every $y\in \mathbb{F}_{q^n}$, hence the image set $L_g(\mathbb{F}_{q^n})$ is contained in the set of the roots of $L_G(x)$. From the Rank-Nullity Theorem and the fact that $g(x)$ divides $x^n-1$, $L_g(\mathbb{F}_{q^n})$ has $q^{n-\deg(g)}=\deg(L_G)$ elements, from where the result follows. \end{proof} We obtain the following result. \begin{proposition}\label{prop:values} Let $f, g\in \mathbb{F}_q[x]$, and let $V$ be the set of distinct roots of $L_f(x)=0$ over $\mathbb{F}_{q^n}$. Then the map $c\mapsto L_g(c)$ is a one to one correspondence of $V$ with itself if and only if $\gcd(f(x), g(x), x^n-1) =1$. \end{proposition} \begin{proof} First, we show that $L_g(V)\subseteq V$. In fact, for $v\in V$, we have that $L_f(L_g(v))=L_{fg}(v)=L_g(L_f(v))=L_g(0)=0$ and so $L_g(v)\in V$. Therefore, from Corollary~\ref{cor:roots} and the fact that $y\mapsto L_g(y)$ is $\mathbb{F}_q$-linear, the following are equivalent: \begin{enumerate}[(i)] \item the map $c\mapsto L_g(c)$ is a one to one correspondence of $V$ with itself; \item the only root of $L_g(x)$ lying in $V$ is $x=0$; \item the only common root of $L_g(x)$ and $L_f(x)$ lying in $\mathbb{F}_{q^n}$ is $x=0$; \item $\gcd(f(x), g(x), x^n-1)=1$. \end{enumerate} \end{proof} \subsection{A permutation criterion} We observe that a polynomial like in Eq.~\eqref{eq:main} can be written as $f(L_g(x))+k(L_g(x))\cdot L_h(x)$, where $g, h\in \mathbb{F}_q[x]$ are polynomials such that $g(x)$ divides $x^n-1$. The following theorem provides a general permutation criterion on when such a polynomial permutes $\mathbb{F}_{q^n}$ and it is the starting point of our results. \begin{theorem}\label{agw-linearized} Let $P\in \mathbb{F}_{q^n}[x]$ be such that $P(x)=f(L_g(x))+k(L_g(x))\cdot L_h(x)$, where $g, h\in \mathbb{F}_q[x]$, $g(x)$ is a monic divisor of $x^n-1$ and $k(L_g(\mathbb{F}_{q^n}))\subseteq \mathbb{F}_q^*$. Then $P$ is a PP of $\mathbb{F}_{q^n}$ if and only if the following hold: \begin{enumerate}[(i)] \item $\gcd(g, h)=1$; \item the polynomial $Q(x)=L_g(f(x))+k(x)\cdot L_h(x)$ permutes the set $L_g(\mathbb{F}_{q^n})$. \end{enumerate} \end{theorem} \begin{proof} We employ a special case of the AGW criterion~\cite{AGW}. In the notation of Theorem~1.5 in~\cite{AGW}, take $\Psi=\overline{\Psi}=L_g$ and $\varphi=L_h$. Then $P$ permutes $\mathbb{F}_{q^n}$ if and only if $Q(x)=L_g(f(x))+k(x)\cdot L_h(x)$ permutes the set $L_g(\mathbb{F}_{q^n})$ and $0\in \mathbb{F}_{q^n}$ is the only common root of $L_g$ and $L_h$, defined over $\mathbb{F}_{q^n}$. Since $g$ divides $x^n-1$, Proposition~\ref{prop:values} entails that the latter holds if and only if $\gcd(g, h)=1$. \end{proof} We observe that the previous theorem is not effective if $g(x)= 1$. So we may require that $g(x)$ is not trivial, i.e., we are looking for polynomials with $g$ of positive degree (preferably high). The following result provides some PP's arising from a generic divisor of $x^n-1$. \begin{proposition}\label{prop:generic-g} Let $P\in \mathbb{F}_{q^n}[x]$ be such that $P(x)=f(L_g(x))+L_h(x)$, where $g, h\in \mathbb{F}_q[x]$ and $g(x)$ is a divisor of $x^n-1$. Suppose that $L_g(f(L_g(y))) = 0$ for every $y\in \mathbb{F}_{q^n}$, i.e., the image of the set $L_g(\mathbb{F}_{q^n})$ by $f(x)$ is contained in the kernel of $L_g$. Then $P$ is a PP of $\mathbb{F}_{q^n}$ if and only if $\gcd(x^n-1, h(x))=1$. The former holds if, for instance, $f$ is of the form $L_{G}(f_0(x))$ with $f_0\in \mathbb{F}_{q^n}[x]$ and $G(x)=\frac{x^n-1}{g(x)}$. \end{proposition} \begin{proof} We employ Theorem~\ref{agw-linearized} with $k(x)=1$. From hypothesis, $Q(z)=L_g(f(z))+L_h(z)=L_h(z)$ for every $z\in L_g(\mathbb{F}_{q^n})$. Hence Theorem~\ref{agw-linearized} implies that $P$ is a PP of $\mathbb{F}_{q^n}$ if and only if $\gcd(g, h)=1$ and $L_h(x)$ permutes the set $L_g(\mathbb{F}_{q^n})$. From Corollary~\ref{cor:roots}, $L_g(\mathbb{F}_{q^n})$ equals the set of distinct roots of $L_{g_0}(x)$, where $g_0(x)=\frac{x^n-1}{g(x)}$ and then, by Proposition~\ref{prop:values}, the former holds if and only if $\gcd(g_0, h)=1$. In conclusion, $P$ is a PP of $\mathbb{F}_{q^n}$ if and only if $\gcd(g, h)=\gcd(g_0, h)=1$. Since $g(x)\cdot g_0(x)=x^n-1$, the latter is equivalent to $\gcd(x^n-1, h(x))=1$. \end{proof} The following proposition provides a family of PP's that arise from Theorem~\ref{agw-linearized} with a special restriction on the polynomials $g$ and $h$. \begin{proposition}\label{prop:PP's} Let $g, h\in \mathbb{F}_q[x]$ be relatively prime polynomials such that $g(x)$ divides $x^n-1$ and $g(x)\cdot h(x)$ is divisible by $x^n-1$. For every polynomial $f\in \mathbb{F}_{q^n}[x]$, $P_{f, g, h}(x):=f(L_g(x))+L_h(x)$ is a PP of $\mathbb{F}_{q^n}$ if and only if $L_g(f(x))$ permutes the set $L_g(\mathbb{F}_{q^n})$. In particular, if $f(L_g(\mathbb{F}_{q^n}))\subseteq L_g(\mathbb{F}_{q^n})$, the latter holds if and only if $f$ permutes the set $L_g(\mathbb{F}_{q^n})$. \end{proposition} \begin{proof} The first part follows directly from Theorem~\ref{agw-linearized} since $L_{gh}(y)=L_{x^n-1}(y)=y^{q^n}-y=0$ for every $y\in \mathbb{F}_{q^n}$. For the second part, it suffices to prove that $L_g$ permutes the set $L_g(\mathbb{F}_{q^n})$. From Corollary~\ref{cor:roots}, $L_g(\mathbb{F}_{q^n})$ is just the set of the roots of the polynomial $L_{g_0}(x)$, where $g_0(x)=\frac{x^n-1}{g(x)}$. Since $x^n-1$ divides $g(x)\cdot h(x)$, it follows that $g_0$ divides $h$ and, since $\gcd(g, h)=1$, we conclude that $\gcd(g, g_0)=1$. The latter combined with Proposition~\ref{prop:values} implies that $L_g$ permutes the set $L_g(\mathbb{F}_{q^n})$. \end{proof} \section{The case $g(x)=\frac{x^n-1}{x-1}$} Here we focus on PP's arising from the factor $g(x)=\frac{x^n-1}{x-1}$. In other words, we are considering permutation polynomials of the form $$f(\mathrm{Tr}_{q^n/q}(x))+k(\mathrm{Tr}_{q^n/q}(x))\cdot L_h(x),$$ where $k, h\in \mathbb{F}_{q}[x]$, $f\in \mathbb{F}_{q^n}[x]$ and $k(\mathbb{F}_{q})\subseteq \mathbb{F}_q^*$. The following definition is frequently used. \begin{definition} For a polynomial $f(x)=\sum_{i=0}^da_ix^i\in \mathbb{F}_{q^n}[x]$, we set $$T_n[f](x)=\sum_{i=0}^d\mathrm{Tr}_{q^n/q}(a_i)\cdot x^i\in \mathbb{F}_q[x].$$ \end{definition} We obtain the following result. \begin{theorem}\label{thm:trace} The polynomial $P(x)=f(\mathrm{Tr}_{q^n/q}(x))+k(\mathrm{Tr}_{q^n/q}(x))\cdot L_h(x)\in \mathbb{F}_{q^n}[x]$ with $h\in \mathbb{F}_q[x]$ and $k(\mathbb{F}_{q})\subseteq \mathbb{F}_q^*$ is a PP over $\mathbb{F}_{q^n}$ if and only if the following conditions hold: \begin{enumerate} \item $\gcd\left(h(x), \frac{x^n-1}{x-1}\right)=1$; \item $Q(x):=T_n[f](x)+k(x)\cdot h(1)\cdot x\in \mathbb{F}_q[x]$ is a PP over $\mathbb{F}_q$. \end{enumerate} In affirmative case, if $R$ is the inverse PP of $Q$ over $\mathbb{F}_q$, then the inverse PP of $P$ over $\mathbb{F}_{q^n}$ equals $$P_0(x)=F(\mathrm{Tr}_{q^n/q}(x))+k(R( \mathrm{Tr}_{q^n/q}(x)))^{q-2}\cdot L_H(x),$$ where $H\in \mathbb{F}_q[x]$ and $F\in \mathbb{F}_{q^n}[x]$ are given as follows: \begin{enumerate}[(i)] \item if $p\mid n$, then $H\in \mathbb{F}_q[x]$ is the unique polynomial of degree at most $n-1$ such that $h(x)\cdot H(x)\equiv 1\pmod {x^n-1}$ and $F$ is any polynomial satisfying $F(x)\equiv -k(R(x))^{q-2}\cdot L_H(f(R(x)))\pmod {x^{q}-x}$; \item if $p\nmid n$, then $H\in \mathbb{F}_q[x]$ is the unique polynomial of degree at most $n-2$ such that $h(x)\cdot H(x)\equiv 1\pmod {\frac{x^n-1}{x-1}}$ and $F$ is any polynomial satisfying $F(x)\equiv M(R(x))\pmod {x^q-x}$, where $$M(x)=-k(x)^{q-2}\cdot L_H(f(x))+x\cdot \frac{1-h(1)\cdot H(1)}{n}.$$ \end{enumerate} We remark that the polynomial $H$ in case (i) always exists since the condition $\gcd\left(h(x), \frac{x^n-1}{x-1}\right)=1$ is equivalent to $\gcd(h(x), x^n-1)=1$ if $p \mid n$. \end{theorem} \begin{proof} Observe that, for $a\in \mathbb{F}_q$, $\mathrm{Tr}_{q^n/q}(f(a))=T_n[f](a)$. In particular, $\mathrm{Tr}_{q^n/q}(f(y))+k(y)\cdot L_h(y)=Q(y)$ for every $y\in \mathbb{F}_q$. Since $\mathrm{Tr}_{q^n/q}(\mathbb{F}_{q^n})=\mathbb{F}_q$, the permutation criterion on $P$ follows by Theorem~\ref{agw-linearized}. It remains to prove the statement regarding the inverse of $P$. For $a\in \mathbb{F}_{q^n}$, set $e_{a}=\mathrm{Tr}_{q^n/q}(a)\in \mathbb{F}_q$, hence $P(a)=f(e_a)+k(e_a)\cdot L_h(a)$ and so $\mathrm{Tr}_{q^n/q}(P(a))=Q(e_a)$ for every $a\in \mathbb{F}_{q^n}$. We obtain that \begin{align*}P_0(P(a))& =F(\mathrm{Tr}_{q^n/q}(P(a))+k(R(\mathrm{Tr}_{q^n/q}(P(a))))^{q-2}\cdot L_H(P(a))\\ {} &= F(Q(e_a))+k(R(Q(e_a)))^{q-2}\cdot L_H(f(e_a)+k(e_a)\cdot L_h(a))\\ {}& = F(Q(e_a))+k(e_a)^{q-2}\cdot (L_H(f(e_a))+k(e_a)\cdot L_{hH}(a))\\ {}& = F(Q(e_a))+k(e_a)^{q-2}\cdot L_H(f(e_a))+ L_{hH}(a),\end{align*} where in the last equality we used the fact that $k(\mathbb{F}_q)\subseteq \mathbb{F}_q^*$, i.e., $k(e_a)^{q-1}=1$. We split the proof into the cases $p\mid n$ and $p\nmid n$. \begin{enumerate}[(i)] \item If $p\mid n$ and $F(x)\equiv -k(R(x))^{q-2}\cdot L_H(f(R(x)))\pmod{x^q-x}$, we obtain that \begin{align*}P_0(P(a))&=-k(R(Q(e_a)))^{q-2}\cdot L_H(f(R(Q(e_{a}))))+k(e_a)^{q-2}\cdot L_H(f(e_{a}))+L_{hH}(a)\\{} & =-k(e_a)^{q-2}\cdot L_H(f(e_a))+k(e_a)^{q-2}\cdot L_H(f(e_{a}))+L_{hH}(a)=a ,\end{align*} where in the last equality we used the fact that $h(x)\cdot H(x)\equiv 1\pmod {x^n-1}$. \item If $p\nmid n$, from $\mathrm{Tr}_{q^n/q}(1)=n\ne 0\in \mathbb{F}_q$, it follows that $a=e_a/n+\delta$ for some $\delta\in \mathbb{F}_{q^n}$ with $\mathrm{Tr}_{q^n/q}(\delta)=0$. Since $h(x)\cdot H(x)\equiv 1\pmod {\frac{x^n-1}{x-1}}$, we have that $$L_{hH}(a)=L_{hH}(\delta+e_a/n)=\delta+L_{hH}(e_a/n)=\delta+e_a\cdot H(1)\cdot h(1)/n.$$ Therefore, $$P_0(P(a))=F(Q(e_a))+k(e_a)^{q-2}\cdot L_H(f(e_a))+\delta+e_a\cdot H(1)\cdot h(1)/n.$$ Since $F(x)\equiv M(R(x))\pmod {x^q-x}$ and $M(x)=-k(x)^{q-2}\cdot L_H(f(x))+x\cdot \frac{1-h(1)\cdot H(1)}{n}$, we have that $$P_0(P(a))=M(e_a)+k(e_a)^{q-2}\cdot L_H(f(e_a))+\delta+e_a\cdot \frac{H(1)\cdot h(1)}{n}=\delta+\frac{e_a}{n}=a.$$ \end{enumerate} \end{proof} If $f, h\in \mathbb{F}_q[x]$, then $L_h(f(x))\equiv h(1)\cdot f(x)\pmod {x^q-x}$. In particular, Theorem~\ref{thm:trace} readily implies the following corollary. \begin{corollary}\label{cor:1} The polynomial $P(x)=f(\mathrm{Tr}_{q^n/q}(x))+k(\mathrm{Tr}_{q^n/q}(x))\cdot L_h(x)$ with $f, h\in \mathbb{F}_q[x]$ and $k(\mathbb{F}_q)\subseteq \mathbb{F}_q^*$ is a $PP$ over $\mathbb{F}_{q^n}$ if and only if $\gcd\left(h(x), \frac{x^n-1}{x-1}\right)=1$ and $Q(x)=n\cdot f(x)+h(1)k(x)\cdot x$ induces a permutation of $\mathbb{F}_q$. More specifically, $P(x)=f(\mathrm{Tr}_{q^n/q}(x))+k(\mathrm{Tr}_{q^n/q}(x))\cdot L_h(x)$ is a PP of $\mathbb{F}_{q^n}$ if and only if one of the following holds: \begin{enumerate}[(i)] \item $p \mid n$, $\gcd(h(x), x^n-1)=1$ and $k(x)\cdot x\in \mathbb{F}_q[x]$ is a PP over $\mathbb{F}_q$; \item $p\nmid n$, $\gcd\left(h(x), x^n-1\right)=x-1$ and $f\in \mathbb{F}_q[x]$ is a PP over $\mathbb{F}_q$; \item $p\nmid n$, $\gcd\left(h(x), x^n-1\right)=1$ and $nf(x)+h(1)\cdot k(x)\cdot x$ is a PP over $\mathbb{F}_q$. \end{enumerate} \end{corollary} Recall that $b\in \mathbb{F}_q[x]$ is a complete permutation polynomial (CPP) or complete mapping over $\mathbb{F}_q$ if both $b(x)$ and $b(x)+x$ are permutations of $\mathbb{F}_q$. We obtain the following corollary. \begin{corollary}\label{cor:cpp} Let $P(x)=f(\mathrm{Tr}_{q^n/q}(x))+L_h(x)\in \mathbb{F}_{q^n}[x]$ be a polynomial such that $h\in \mathbb{F}_q[x]$. Then $P(x)$ is a CPP over $\mathbb{F}_{q^n}$ if and only if the following hold: \begin{enumerate}[(i)] \item $\gcd\left(h(x)\cdot (h(x)+1), \frac{x^n-1}{x-1}\right)=1$; \item $Q(x):=T_n[f](x)+h(1)\cdot x\in \mathbb{F}_q[x]$ is a CPP over $\mathbb{F}_q$. \end{enumerate} \end{corollary} \begin{proof} We observe that if $h_0(x)=h(x)+1$, then $P(x)+x=f(\mathrm{Tr}_{q^n/q}(x))+L_{h_0}(x)$ and $T_n[f](x)+h_0(1)\cdot x=Q(x)+x$. The result follows from Theorem~\ref{thm:trace} by taking $k(x)=1$. \end{proof} \subsection{From PP's of $\mathbb{F}_q$ to PP's of $\mathbb{F}_{q^n}$} From Theorem~\ref{thm:trace}, we obtain the following method for producing permutations of the extension field $\mathbb{F}_{q^n}$ from permutations of the base field $\mathbb{F}_q$. \begin{theorem}\label{thm:const} Let $b\in \mathbb{F}_q[x]$ be a permutation polynomial, let $n$ be a positive integer and let $h, k\in \mathbb{F}_q[x]$ be polynomials such that $\gcd\left(h(x), \frac{x^n-1}{x-1}\right) =1$ and $k(\mathbb{F}_q)\subseteq \mathbb{F}_q^*$. Then for every polynomial $f\in \mathbb{F}_{q^n}[x]$ such that $T_n[f](x)\equiv b(x)-h(1)\cdot k(x)\cdot x\pmod {x^q-x}$, the polynomial $$\mathcal P_{b, f, h, k, n}(x):=f(\mathrm{Tr}_{q^n/q}(x))+k(\mathrm{Tr}_{q^n/q}(x))\cdot L_h(x),$$ permutes $\mathbb{F}_{q^n}$. \end{theorem} In simpler terms, Theorem~\ref{thm:const} provides the following procedure for constructing permutations of $\mathbb{F}_{q^n}$. \begin{algorithm} \begin{itemize} \item Step 1: pick a permutation polynomial $b \in \mathbb{F}_q[x]$. \\ \item Step 2: pick polynomials $h, k\in \mathbb{F}_{q}[x]$ such that $\gcd\left(h(x), \frac{x^n-1}{x-1}\right) =1$ and $k(\mathbb{F}_q)\subseteq \mathbb{F}_q^*$. \item Step 3: find $f\in \mathbb{F}_{q^n}[x]$ such that $$T_n[f](x)\equiv b(x) - k(x)\cdot h(1)\cdot x\pmod {x^q-x}.$$ \item Step 4: construct the permutation polynomial $$\mathcal P_{b, f, h, k, n} (x) = f(\mathrm{Tr}_{q^n/q}(x))+k(\mathrm{Tr}_{q^n/q}(x))\cdot L_h(x).$$ \end{itemize} \end{algorithm} We observe that the functional equation $$T_n[f](x)\equiv b(x)-k(x)\cdot h(1)\cdot x\pmod {x^q-x},$$ with the restriction $\deg(f)<q$ gives rise to a linear system of $q$ equations involving trace functions. More specifically, if we write $f(x)=\sum_{i=0}^{q-1}y_ix^i, b(x)=\sum_{i=0}^{q-1}b_ix^i$ and $k(x)=\sum_{i=0}^{q-1}k_ix^i$, the functional equation is equivalent to the following system of equations: \begin{align*} \mathrm{Tr}_{q^n/q}(y_0) &= b_0,\\ \mathrm{Tr}_{q^n/q}(y_1) &= b_1 - k_0\cdot h(1)-k_{q-1}\cdot h(1), \\ \mathrm{Tr}_{q^n/q}(y_i) &= b_i-k_{i-1}\cdot h(1), \; 2\le i\le n. \end{align*} In particular, for fixed $n, b, h$ and $k$, there are $q^{(n-1)q}$ distinct choices for $f$ under the condition $\deg(f)<q$. Moreover, such condition implies that the corresponding polynomials $\mathcal P_{b, f, h,k, n}$ are all incongruent modulo $x^{q^n}-x$. Hence we obtain $q^{(n-1)q}$ distinct PP's of $\mathbb{F}_{q^n}$. \begin{remark}\label{rem:main} The following are easily verified. \begin{enumerate}[(i)] \item We may also iterate the procedure above, hence obtaining a sequence $\{\mathcal P_i\}_{i\ge 1}$ of polynomials such that $\mathcal P_i$ permutes $\mathbb{F}_{q^{n^i}}$ for every $i\ge 1$. \item In the context of Theorem~\ref{thm:const}, if we also know the inverse of the permutation $b(x)$, we may employ Theorem~\ref{thm:trace} and obtain the inverse PP of any permutation $\mathcal P_{b, f, h, k, n}$. \item From Corollary~\ref{cor:cpp}, the method above can be extended to the construction of complete permutation polynomials of $\mathbb{F}_{q^n}$ with $k(x)=1$: we start with $b \in \mathbb{F}_q[x]$, a CPP of $\mathbb{F}_q$, and pick $h\in \mathbb{F}_q[x]$ such that $\gcd\left(h(x)\cdot (h(x)+1), \frac{x^n-1}{x-1}\right)=1$. Similarly, from such construction, we obtain $q^{(n-1)q}$ distinct CPP's of $\mathbb{F}_{q^n}$. \end{enumerate} \end{remark} The following result exemplifies the applicability of Theorem~\ref{thm:const} to the explicit construction of permutations and complete permutation polynomials of $\mathbb{F}_{q^n}$ with $k(x)=1$. Its proof follows directly by Theorem~\ref{thm:const} and the fact that $\mathrm{Tr}_{q^n/q}(\theta^q-\theta)=0$ for every $\theta\in \mathbb{F}_{q^n}$. \begin{corollary} Let $n$ be a positive integer, not divisible by the characteristic $p$ of $\mathbb{F}_q$. Let $h\in \mathbb{F}_q[x]$ be a polynomial such that $\gcd\left(h(x), \frac{x^n-1}{x-1}\right) =1$ and let $b(x) = \sum_{i=0}^{m} b_i x^i$ be any permutation polynomial of $\mathbb{F}_q$. For every $\theta_0, \ldots, \theta_{q-1}\in \mathbb{F}_{q^n}$, the polynomial \begin{equation*} P(x)=L_h(x)- \frac{h(1)}{n} \cdot \mathrm{Tr}_{q^n/q}(x)+ \displaystyle{\sum_{i=0}^{m }\frac{b_i}{n}\cdot \mathrm{Tr}_{q^n/q}(x)^i + \sum_{i=0}^{q-1}(\theta_i^q-\theta_i )\cdot \mathrm{Tr}_{q^n/q}(x)^i }, \end{equation*} permutes $\mathbb{F}_{q^n}$. In addition, if $\gcd\left(h(x)+1, \frac{x^n-1}{x-1}\right)=1$ and $b(x)$ is a CPP of $\mathbb{F}_q$, then $P(x)$ is also a CPP of $\mathbb{F}_{q^n}$. \end{corollary} For a permutation monomial $b(x) = ax^m$ of $\mathbb{F}_q$, Theorem~\ref{thm:const} immediately implies the following result. \begin{corollary} Let $h\in \mathbb{F}_q[x]$ be a polynomial such that $\gcd\left(h(x), \frac{x^n-1}{x-1}\right) =1$ and let $m$ be a positive integer such that $m<q-1$ and $\gcd(m, q-1)=1$. The following hold: \begin{enumerate}[(i)] \item if $m=1$, then for every $\alpha, \theta_0, \ldots, \theta_{q-1}\in \mathbb{F}_{q^n}$ such that $\mathrm{Tr}_{q^n/q}(\alpha)\ne -h(1)$, the polynomial $$P(x)=L_h(x)+\alpha\cdot \mathrm{Tr}_{q^n/q}(x)+\sum_{i=0}^{q-1}(\theta_i^q-\theta_i)\cdot \mathrm{Tr}_{q^n/q}(x)^i,$$ permutes $\mathbb{F}_{q^n}$. In particular, if $n$ is not divisible by the characteristic $p$ of $\mathbb{F}_q$, one may take $\alpha$ as any element in $\mathbb{F}_q\setminus\{-\frac{h(1)}{n}\}$. \item if $m>1$, then for every $\alpha, \beta, \theta_0, \ldots, \theta_{q-1}\in \mathbb{F}_{q^n}$ such that $\mathrm{Tr}_{q^n/q}(\alpha)=-h(1)$ and $\mathrm{Tr}_{q^n/q}(\beta)\ne 0$, the polynomial $$P(x)=L_h(x)+\alpha\cdot \mathrm{Tr}_{q^n/q}(x)+\beta\cdot \mathrm{Tr}_{q^n/q}(x)^m+\sum_{i=0}^{q-1}(\theta_i^q-\theta_i)\cdot \mathrm{Tr}_{q^n/q}(x)^i,$$ permutes $\mathbb{F}_{q^n}$. In particular, if $n$ is not divisible by the characteristic $p$ of $\mathbb{F}_q$, one may take $\alpha=\frac{-h(1)}{n}$ and $\beta$ as any nonzero element of $\mathbb{F}_q$. \end{enumerate} \end{corollary} \subsection{Extension to generic divisors of degree $n-1$.} The previous result can be easily extended to any other divisor of $x^n-1$ over $\mathbb{F}_q$ with degree $n-1$. If $a\in \mathbb{F}_{q}$ is a root of the polynomial $x^n-1$ (this holds for some $a\ne 1$ if $\gcd(q-1, n)>1$), and $g_a(x)=\frac{x^n-1}{x-a} = x^{n-1} + ax^{n-2} + \cdots + a^{n-1}$, then we can obtain an analogue of Theorem~\ref{thm:trace}. In this case, $L_{g_a}(x) = x^{q^{n-1}} + ax^{q^{n-2}} + \cdots + a^{n-1} x$ and $L_{g_a}(\mathbb{F}_{q^n}) = \delta \cdot \mathbb{F}_q$, where $\delta \in \mathbb{F}_{q^n}$ any element satisfying $\delta^q = a \delta$. In this analogy, we replace conditions (1) and (2) there by $\gcd\left(\frac{x^n-1}{x-a}, h(x)\right)=1$ and $Q_a(x)=T_n^{(a)}[f](x)+k(x)\cdot h(a)\cdot x$ permutes the set $\delta \cdot \mathbb{F}_{q}$, where $\delta\in \mathbb{F}_{q^n}$ is any element with $\delta^q=a\delta$ and for $f(x)=\sum_{i=0}^ma_ix^i\in \mathbb{F}_{q^n}$, $$T_n^{(a)}[f](x)=a^{-1}\sum_{i=0}^m\left(\sum_{j=0}^{n-1}a^{(i-1)j}\cdot a_i^{q^j}\right)x^i=a^{-1}\sum_{i=0}^m\left(\sum_{j=0}^{n-1}\delta^{(i-1)(q^j-1)}\cdot a_i^{q^j}\right)x^i.$$ We observe that $Q_a(x)$ permutes $\delta \cdot \mathbb{F}_{q}$ if and only if $\delta^{-1}Q_a(\delta x)$ permutes $\mathbb{F}_q$. By a direct computation, we obtain the following polynomial: $$ \bar{Q}_a(x) = \delta^{-1}Q_a(\delta x) = \frac{1}{a} \sum_{i=0}^m \mathrm{Tr}(\delta^{i-1} a_i) x^i + k(\delta x) h(a) x. $$ In terms of AGW criterion, we have the following commutative diagram. \[ \xymatrixcolsep{12pc} \xymatrix{ \mathbb{F}_{q^n} \ar[r]^{P(x) = f(L_{g_a}(x) )+k(L_{g_a}(x))\cdot L_h(x) } \ar[d]_{L_{g_a}(x) } & \mathbb{F}_{q^n} \ar[d]^{L_{g_a}(x) }\\ \delta\cdot \mathbb{F}_q \ar[r]^{Q_a(x)=T_n^{(a)}[f](x)+k(x)\cdot h(a)\cdot x} \ar[d]_{\delta^{-1} x } & \delta\cdot \mathbb{F}_q \ar[d]^{ \delta^{-1}x}\ \\ \mathbb{F}_q \ar[r]^{\bar{Q}_a(x)=\frac{1}{a} \sum_{i=0}^m \mathrm{Tr}(\delta^{i-1} a_i) x^i + k(\delta x) h(a) x} & \mathbb{F}_q } \] We summarize our result as follows: \begin{theorem}\label{thm:tracevariant} Let $a \in \mathbb{F}_q$ such that $a^n=1$ and $\delta \in \mathbb{F}_{q^n}$ such that $\delta^q = a \delta$. The polynomial $P(x)=f(L_{g_a}(x) )+k(L_{g_a}(x))\cdot L_h(x)\in \mathbb{F}_{q^n}[x]$ with $L_{g_a}(x) = x^{q^{n-1}} + ax^{q^{n-2}} + \cdots + a^{n-1} x$, $f(x) = \sum_{i=0}^m a_i x^i$, $h(x)\in \mathbb{F}_q[x]$ and $k(\delta \mathbb{F}_{q})\subseteq \mathbb{F}_q^*$ is a PP over $\mathbb{F}_{q^n}$ if and only if the following conditions hold: \begin{enumerate} \item $\gcd\left(h(x), \frac{x^n-1}{x-a}\right)=1$; \item $\bar{Q}_a(x) = \frac{1}{a} \sum_{i=0}^m \mathrm{Tr}(\delta^{j-1} a_j) x^i + k(\delta x) h(a) x$ is a PP over $\mathbb{F}_q$. \end{enumerate} In affirmative case, if $R$ is the inverse PP of $\bar{Q}_a$ over $\mathbb{F}_q$, then the inverse PP of $P$ over $\mathbb{F}_{q^n}$ equals $$P_0(x)=F(\delta^{-1} L_{g_a}(x))+k(\delta R(\delta^{-1} L_{g_a}(x)))^{q-2}\cdot L_H(x),$$ where $H\in \mathbb{F}_q[x]$ and $F\in \mathbb{F}_{q^n}[x]$ are given as follows: \begin{enumerate}[(i)] \item if $p\mid n$, then $H\in \mathbb{F}_q[x]$ is the unique polynomial of degree at most $n-1$ such that $h(x)\cdot H(x)\equiv 1\pmod {x^n-1}$ and $F$ is any polynomial satisfying $F(x)\equiv -k(\delta R(x))^{q-2}\cdot L_H(f(\delta R(x)))\pmod {x^{q}-x}$; \item if $p\nmid n$, then $H\in \mathbb{F}_q[x]$ is the unique polynomial of degree at most $n-2$ such that $h(x)\cdot H(x)\equiv 1\pmod {\frac{x^n-1}{x-a}}$ and $F$ is any polynomial satisfying $F(x)\equiv M(\delta R(x))\pmod {x^q-x}$, where $$M(x)=-k(x)^{q-2}\cdot L_H(f(x))+ax\cdot \frac{1-h(a)\cdot H(a)}{n}.$$ \end{enumerate} We remark that the polynomial $H$ in case (i) always exists since the condition $\gcd\left(h(x), \frac{x^n-1}{x-a}\right)=1$ is equivalent to $\gcd(h(x), x^n-1)=1$ if $p\mid n$. \begin{proof} In this case, $L_{g_a}(\delta) = \delta^{q^{n-1}} + a \delta^{q^{n-2}} + \cdots + a^{n-1} \delta = n a^{n-1} \delta=na^{-1}\delta$. For each $b\in \mathbb{F}_{q^n}$, we set $e_b = L_{g_a}(b) \in \delta\cdot \mathbb{F}_q$. In particular, $b= a\cdot \frac{e_{b}}{n} + \alpha$ for some $\alpha$ such that $L_g(\alpha) =0$. The rest of proof follows as the proof of Theorem~\ref{thm:trace}. \end{proof} \end{theorem} Similarly, we can recursively construct permutation polynomials of $\mathbb{F}_{q^n}$ from a permutation polynomial $b(x)=\sum_{i=0}^{q-1}b_ix^i$ of $\mathbb{F}_q$. Let $k(x)=\sum_{i=0}^{q-1}k_ix^i\in \mathbb{F}_{q^n}[x]$ such that $k(\delta x) \in \mathbb{F}_q[x]$ and $k(\delta \mathbb{F}_q) \subseteq \mathbb{F}_q^*$. From the functional equation $\bar{Q}_a(x) \equiv b(x) \pmod {x^q-x}$, we obtain the following system of equations: \begin{align*} \mathrm{Tr}_{q^n/q}(\delta^{-1} y_0) &=a b_0,\\ \mathrm{Tr}_{q^n/q}(y_1) &= a(b_1 - k_0\cdot h(a)-k_{q-1}\delta^{q-1} \cdot h(a)), \\ \mathrm{Tr}_{q^n/q}(\delta^{i-1} y_i) &= a(b_i-k_{i-1} \delta^{i-1} \cdot h(a)), \; 2\le i\le q-1. \end{align*} In particular, for fixed $n, b, h$ and $k$, there are $q^{(n-1)q}$ distinct choices for $f$ under the condition $\deg(f)<q$. Moreover, such condition implies that the corresponding polynomials are all incongruent modulo $x^{q^n}-x$. Hence we obtain $q^{(n-1)q}$ distinct PP's of $\mathbb{F}_{q^n}$. \section{Conclusion and future research} In this paper we have explored the permutational property of polynomials of the form $f(L_g(x))+k(L_g(x))\cdot L_h(x)\in \mathbb{F}_{q^n}[x]$ over $\mathbb{F}_{q^n}$, where $L_g, L_h$ are the $q$-associates of polynomials $g, h\in \mathbb{F}_q[x]$ with $g(x)$ a divisor of $x^n-1$ and $k\in \mathbb{F}_q[x]$ satisfies $k( \delta \mathbb{F}_q)\subseteq \mathbb{F}_q^*$. With the help of the AGW criterion, we have provided some general results on PP's of this form, relating to arithmetic properties of the polynomials $g$ and $h$, like the condition $\gcd(g, h)=1$. By specializing to the case when the divisor $g(x)$ is of degree $n-1$, we obtained results on the construction of PP's, CPP's and their inverses. It would be interesting to explore the construction of PP's arising from other factors of $x^n-1$. The natural next step is to consider divisors of $x^n-1$ of degree $n-2$ over $\mathbb{F}_q$, where we have to discuss the permutational property of polynomials over $\mathbb{F}_q$-vector spaces of dimension $2$. \newcommand{\etalchar}[1]{$^{#1}$}

c51980c767bd6e60a481c3c68b9234e393d273e1

\section{Introduction: dynamical and phenomenological study at different regime of QCD phase diagram} \section{Introduction} The phase transition and phase structure of the strongly interacting matter at finite temperature and density are central topics in high-energy nuclear physics. Lattice QCD simulations show that the phase transition at vanishing baryon chemical potential ($\mu_B\simeq0$) is a smooth crossover \cite{Aoki:2006we,Ding:2015ona,Bazavov:2019lgz,Ratti:2018ksb}. Effective field theory models predict a first-order phase transition boundary at large chemical potential, together with a critical endpoint~\cite{Fukushima:2010bq,Fukushima:2013rx,Fischer:2018sdj}. However, lattice QCD suffers from the notorious sign problem at large chemical potential region, and theoretical predictions for the critical point's location remain model-dependent~\cite{Stephanov:2004wx,Stephanov:2007fk}. On the experimental side, relativistic heavy-ion collisions at Relativistic Heavy-Ion Collider (RHIC) and the Large Hadron Collider (LHC) aim to create the quark-gluon plasma (QGP) and explore its phase transition at zero and finite baryon chemical potential. At top RHIC and the LHC energies, the created QGP behaves like a nearly perfect liquid with almost vanishing chemical potential at mid-rapidity~\cite{Gyulassy:2004vg,Gyulassy:2004zy,Kolb:2003dz,Muller:2012zq}. To probe the QCD phase diagram at finite baryon chemical potential and search the QCD critical point, RHIC has carried out Beam Energy Scan (BES) program for Au+Au collisions with collision energies ranges from $3 \sim 200$ A GeV~\cite{Aggarwal:2010cw,Luo:2017faz,Bzdak:2019pkr}. Future experimental programs, such as Facility for Antiproton and Ion Research (FAIR) in Darmstadt \cite{Friman:2011zz}, Nuclotron-based Ion Collider fAcility (NICA) in Dubna \cite{NICA-web}, HIAF in Huizhou \cite{Ruan:2018fpo} and J-PARC-HI in Tokai \cite{J_PARC-HI-web}, will further explore the phase diagram at higher baryon density to search for the QCD critical point, the first-order phase transition boundary, and study the strongly interacting matter at high baryon density. On the theoretical and phenomenological side, significant progress has been achieved to extract the QGP transport properties at top RHIC and LHC energies~\cite{Song:2010mg,Song:2012ua,Bernhard:2016tnd, Bernhard:2019bmu,Nijs:2020roc,Everett:2020yty,Everett:2020xug}. In particular, integrated hybrid models have been developed to quantitatively describe the complex multi-stage evolution of the QCD matter created in relativistic heavy-ion collisions \cite{Song:2010mg,Song:2010aq,Hirano:2012kj,Shen:2014vra,Niemi:2015qia,Ryu:2017qzn,Pang:2018zzo,Shen:2017bsr}. To extend the theoretical frameworks to heavy-ion collisions at $\sqrt{s_\mathrm{NN}} \sim \mathcal{O}(10)$\,GeV, new theory ingredients in multiple aspects need to be included. In particular, the QCD phase transition must be properly encoded in the dynamical model to describe the experimental measurements quantitatively. Phenomenological studies with upcoming precise measurements can extract robust QGP transport coefficients in a baryon-rich environment. This short review will summarize recent progress in improving the dynamical model descriptions of relativistic nuclear collisions as they go through crossover, critical point, and first-order phase transition regions in the QCD phase diagram. \section{Quantitative characterization of the QCD crossover region} In this section, we focus on dynamical descriptions of relativistic heavy-ion collisions with collision energies $\sqrt{s_\mathrm{NN}} \ge 20$\,GeV. These collisions probe the hot QCD matter with net baryon chemical potential $\mu_B \le 250$\,MeV \cite{Andronic:2017pug, Adamczyk:2017iwn}, where the QGP and hadronic phases are connected with a smooth crossover. \underline{\textbf{Hydrodynamics with dynamical initialization}} \underline{\textbf{schemes:}} Hydrodynamics and hybrid models are important tools to describe the QGP fireball evolution and study soft observables for relativistic heavy-ion collisions at RHIC and LHC energies \cite{Heinz:2013th,Gale:2013da,Jeon:2015dfa,Song:2017wtw}.For collision energies $\sqrt{s_\mathrm{NN}} = 20 - 60$\,GeV, the duration for two colliding nuclei with radii $R$ to pass through each other can be estimated by $\tau_\mathrm{overlap} \sim 2R / \sqrt{\left(\frac{\sqrt{s_\mathrm{NN}}}{2m_N}\right)^2 - 1}$, which gradually increases with the decrease of collision energy~\cite{Karpenko:2015xea, Shen:2017bsr, Shen:2017fnn}. Meanwhile, the longitudinal boost invariance violates significantly~\cite{Noronha:2018atu, Shen:2020jwv}. Therefore, it is important to study the role of pre-equilibrium dynamics during the overlapping period within the framework of (3+1)D hydrodynamics. In recent years, dynamical initialization schemes have been developed to model this extended interaction region in heavy-ion collisions \cite{Okai:2017ofp, Shen:2017ruz, Shen:2017bsr,Du:2018mpf, Akamatsu:2018olk, Kanakubo:2019ogh}. Commonly, they interweave the initial collision stage with hydrodynamics on a local basis while the two nuclei pass through each other. The initial state energy-momentum and conserved charge density currents are treated as sources to feed the hydrodynamic fields, \begin{eqnarray} \partial_\mu T^{\mu\nu} &=& J^\nu_\mathrm{source}(\tau, \vec{x}) \\ \partial_\mu J^\mu_i &=& \rho_{i,\mathrm{source}}(\tau, \vec{x}). \end{eqnarray} Here the conserved quantum charges for light flavor quarks are baryon, strangeness, and electric charges, $i = B, S, Q$. Dynamical initialization schemes require initial state models to provide the (3+1)D space-time and momentum information of the energy-momentum and charge distributions. There has been a stream of collective effort to develop 3D initial state models. The complex 3D collision dynamics can be approximated by parametric energy and charge depositions \cite{Hirano:2005xf, Bozek:2010vz, Bozek:2015bna, Bozek:2017qir, Shen:2020jwv, Sakai:2020pjw}. More dynamical models involve simulating energy loss during individual nucleon-nucleon collisions. Such initial state models have been built based on classical string deceleration \cite{Shen:2017bsr, Bialas:2016epd}. And there are 3D initial conditions based on hadronic and partonic transport simulations \cite{Pang:2012he, Karpenko:2015xea,Du:2018mpf,Xu:2016hmp,Fu:2020oxj}. These models provide non-trivial correlations between the longitudinal energy distribution and flow velocity. Furthermore, recent theory developments to understand early-stage baryon stopping from the Color Glass Condensate-based approaches in the fragmentation region \cite{Li:2018ini, McLerran:2018avb}. The initial energy density and baryon charge distributions were also studied from a holographic approach at intermediate couplings \cite{Attems:2018gou}. Measurements of the rapidity-dependent particle production and flow correlations can provide valuable constraints on initial state longitudinal fluctuations and baryon stopping. \underline{\textbf{Equation of State (EoS):}} The equation of state for nuclear matter for $\mu_B \le 250$\,MeV has been constructed based on the Taylor series technique with high-order susceptibility coefficients computed from lattice QCD calculations~\cite{Ratti:2018ksb, Bazavov:2018mes, Monnai:2019hkn, Noronha-Hostler:2019ayj, Parotto:2018pwx}. See a recent review for more details on this topic~\cite{Monnai:2021kgu}. These equations of state at finite densities are essential to enable dynamics of Au+Au collisions at the RHIC BES energies. Furthermore, one needs to build in constraints on the strangeness neutrality and electric charge $n_Q \simeq 0.4 n_B$ \cite{Monnai:2019hkn, Noronha-Hostler:2019ayj} for heavy-ion collisions with gold and lead nuclei. The strangeness neutrality condition was crucial to reproduce the multi-strangeness baryon yields and could significantly change the fireball phase trajectories in the QCD phase diagram \cite{Monnai:2019hkn}. \underline{\textbf{Transport coefficients:}} Transport coefficients are important inputs in hydrodynamic simulations. At the top RHIC and LHC energies, the Bayesian inference method has been adopted to quantitatively constrain the shear and bulk viscosity and their temperature dependence with the soft hadron measurements~\cite{Bernhard:2016tnd, Bernhard:2019bmu,Nijs:2020roc,Everett:2020yty,Everett:2020xug}. Extending this approach to analyze flow measurements at the RHIC Beam Energy Scan program would further extract the $\mu_B$-dependence of the specific QGP viscosity with reliable uncertainty. Studies have shown that the collision energy and rapidity dependence of the anisotropic flow coefficients has strong constraining power of QGP's $\eta/s(T, \mu_B)$ \cite{Karpenko:2015xea, Denicol:2015nhu, Shen:2020jwv}. To take advantage of those measurements in the Bayesian analysis, full-scale (3+1)D hydrodynamic simulations are essential. The dynamics of conserved charge currents allow us to study new transport properties of the QGP, namely the charge diffusion processes inside the fluid. Driven by the local gradients of $\mu_i/T$, the net baryon, strangeness, and electric charges can flow with different velocities than the energy density. Recent works showed that the net baryon diffusion has important effects on the longitudinal dynamics of the net baryon current \cite{Denicol:2018wdp, Li:2018fow, Du:2019obx, Wu:2021ypv}. A theoretical framework that incorporates diffusion effects on multiple conserved charge currents has been developed \cite{Fotakis:2019nbq}. Such a framework can help us access the full diffusion coefficient matrix, which takes the cross charge correlation into account \cite{Greif:2017byw, Rose:2020sjv}. \underline{\textbf{Hadronic transport:}} Hydrodynamics are generally connected with a hadronic transport model for a better description of the non-equilibrium late hadronic evolution~\cite{Song:2010aq,Hirano:2012kj,Shen:2014vra,Ryu:2017qzn}. As the QGP fireball evolves into the phase-transition region, individual fluid cells are converted to hadrons according to the Cooper-Frye prescription \cite{Cooper:1974mv}. These hadrons can further scatter with each other and decay to stable states described by hadronic transport model models, such as UrQMD \cite{Bass:1998ca, Bleicher:1999xi}, JAM \cite{Nara:1999dz}, and SMASH \cite{Weil:2016zrk}. As the collision energy reduces from 200 GeV to 20 GeV, the role of the late-stage hadronic transport becomes more and more crucial \cite{Monnai:2019hkn}. Because the hydrodynamic phase in heavy-ion collisions at $\sqrt{s} \sim 20-60$\,GeV is shorter than those at TeV collision energies at the LHC, both radial and anisotropic flow further develop during the hadronic evolution due to the remaining spatial inhomogeneity~\cite{Shen:2011zc, Denicol:2018wdp}. \section{Dynamic models near the critical point} This section will briefly review critical phenomena in equilibrium and then highlight the dynamical descriptions near the QCD critical point. \subsection{Equilibrium critical phenomena } One of the landmark features for an equilibrium system near the critical point is the long-range correlation, which results in many unique properties, such as critical opalescence, universal scaling, singular behavior of the equation of state, large fluctuations of thermodynamics variables, and so on. This subsection will briefly review the critical phenomena for an equilibrium hot QCD system. \underline{\textbf{ Universal Scaling of the hot QCD system}}: Due to the infinite long-range correlation and divergence of the correlation length, the system at the critical point has no characteristic scale, which behaviors self-similarly under the scale transformation. As a result, different systems in the same universality class, determined by the system's dimension and the number of order parameter components, share universal critical properties. From the symmetry analysis, it is generally believed that QCD critical point and 3D-Ising model belong to the same static universality class~\cite{Pisarski:1983ms,Wilczek:1992sf,Rajagopal:1992qz}. For finite volume systems, such as the QGP fireball created in heavy-ion collisions, the universal scaling argument should take finite size scaling into account~\cite{Palhares:2009tf,Fraga:2011hi}. \underline{\textbf{EoS with a critical point}}: Recent lattice QCD calculations have narrowed the location of the critical point, which may exist in the region of $T<140,\mu_B>300$ MeV \cite{Ding:2020rtq}. However, it is still challenging for lattice simulations to precisely predict the position of the critical point due to the sign problem at finite $\mu_B$. On the other hand, because the hot QCD system belongs to the same universality class of the 3D-Ising model~\cite{Rajagopal:1992qz,Berges:1998rc,Halasz:1998qr,Karsch:2001nf}, non-universally map the EoS between these two systems~\cite{Nonaka:2004pg,Bluhm:2006av,Parotto:2018pwx,Pradeep:2019ccv} provide a proper parametrization for EoS with a QCD critical point. Recently, a more comprehensive EoS near the critical point has been constructed, which combines the singular part from the Ising model and the background part obtained from lattice EoS at $\mu_B=0$~\cite{Parotto:2018pwx,Mroczek:2020rpm,Stafford:2021wik}. It is crucial to study such EoS effects on final observables, which has not been done yet. \underline{\textbf{ Multiplicity fluctuations}}: The strong fluctuations of the order parameter field near the critical point could lead to large fluctuations of final hadrons produced in heavy-ion collisions~\cite{Stephanov:1998dy,Stephanov:1999zu}, especially for the event-by-event net-proton fluctuations~\cite{Hatta:2003wn,Kitazawa:2012at,Kitazawa:2011wh}. However, the correlation length is largely limited due to the finite size and finite time effects of the QGP fireball created in heavy-ion collisions~\cite{Berdnikov:1999ph,Nonaka:2004pg}, which make it hard to probe the critical point from the second-order cumulant of the event-by-event multiplicity fluctuations. Ref.~\cite{Stephanov:2008qz} proposed that the higher-order cumulants are more sensitive to correlation length, some of which change signs near the QCD phase boundary~\cite{Stephanov:2011pb,Athanasiou:2010kw,Asakawa:2009aj}. Recently, the higher-order cumulants of the net-proton have been systematically measured in the RHIC BES program. The kurtosis of net-proton in the most central collisions presents non-monotonic behavior, which qualitatively agrees with the theoretical expectation and indicating the potential discovery of QCD critical point. However, the sign of skewness data fails to describe the experiment measurement~\cite{Jiang:2015hri} and requires to include the dynamical effect, which will be discussed in the following subsection. \subsection{Dynamical fluctuations near the QCD critical point} For systems evolving near the critical point, the dynamical universality class is classified according to the symmetry of order parameter, dimensionality, conservation laws of the conserved densities, and the Poisson bracket among them~\cite{Hohenberg:1977ym}. Ref.~\cite{Son:2004iv} suggested that the evolving hot QCD system is in the class of model H, which describes a dynamical system with the conserved order parameter, conserved momentum density, and the Poisson bracket between them. Owing to the complexity of numerical implementations of model H, model A and B have been served as simplified dynamical models near the QCD critical point, which will be briefly reviewed as follow. We will also review the recent progress on other dynamical models near the critical point, such as Non-equilibrium chiral fluid dynamics and the Hydro+ formalism. \underline{\textbf{ Model A}} only evolves the non-conserved order parameter field $\sigma$, {which serves as a simplified version of the dynamic model near the QCD critical point}. It was found that the critical slowing down effects not only limit the growth of the correlation length, as predicted in early papers~\cite{Nonaka:2004pg,Berdnikov:1999ph}, but also substantially modifies the temporal evolution of cumulants and even reverse the signs of skewness and kurtosis compared to the equilibrium values~\cite{Mukherjee:2015swa,Mukherjee:2016kyu,Jiang:2017mji,Wu:2018twy}. \underline{\textbf{ Model B}} only evolves the conserved quantity, which can be considered as another simplified model for the evolving dynamical systems near the QCD critical point. It is motivated by the fact that the dynamics of the non-conserved order parameter field becomes insignificant by mixing with the conserved baryon number density, which makes the dynamics of conserved quantity dominate at a long time scale \cite{Son:2004iv}. Besides the normal critical slowing down effects \cite{Nahrgang:2018afz,Sakaida:2017rtj,Wu:2019qfz}, one of the intriguing results for the conserved critical dynamics is the competition between the growth of correlation length and the diffusion effect, which leads to a non-monotonic behavior for the multiplicity fluctuations of the conserved charges as the rapidity window increased \cite{Sakaida:2017rtj}. For more realistic implementation, model B should be extended with the spatial-temporal evolution of the fireball, which is an ongoing efforts~\cite{Kitazawa:2020kvc}. \underline{\textbf{Non-equilibrium chiral fluid dynamics (N$\chi$FD)}} is a dynamical model that couples the chiral condensate with the evolving fluid~\cite{Nahrgang:2011mg,Nahrgang:2011mv,Nahrgang:2011vn,Herold:2014zoa, Herold:2016uvv,Herold:2018ptm,Paech:2003fe}. The chiral condensate $\sigma$, treated as the critical mode, evolves according to the Langevin equation with the effective potential obtained from the effective theory ({\it e.g.,} (P)QM model \cite{Sasaki:2011sd}, linear sigma model \cite{GellMann:1960np,Scavenius:2000qd}). It also provides source terms for the hydrodynamic equations for the heat bath evolution of the fluid. In~\cite{Herold:2014zoa,Herold:2016uvv,Herold:2018ptm}, N$\chi$FD has been numerical implemented to study various dynamical effects near the QCD phase transition. It was found that variance and kurtosis for the net-proton multiplicities are enhanced near the critical point compared with the values with the crossover phase transition. \underline{\textbf{Fluctuating Hydrodynamics}} extends traditional dissipative hydrodynamics with stochastic fluctuations according to the fluctuation-dissipation theorem~\cite{Kapusta:2011gt}, where fluctuations are treated as white noise. Fluctuating Hydrodynamics was first proposed in a non-relativistic form~\cite{Landau:1959,Lifshitz:1980} and extended to the relativistic case in~\cite{Kapusta:2011gt}. In principle, by identifying the slow mode associated with critical fluctuations and properly implementing the EoS and transport coefficients, one could study the dynamics of critical fluctuations and their influence on experimental observables using the fluctuating hydrodynamics framework. For (0+1)D boost invariant expanding background, analytical calculations showed that the magnitude of the correlation function enhances due to the increased thermal conductivity near the QCD critical point~\cite{Kapusta:2012zb}. In practice, especially for numerical implementations, one should carefully deal with the {multiplicative} noise~\cite{Murase:2013tma,Kovtun:2014hpa,Arnold:1999va} and the dependence of the grid sizes, especially for the case near the phase transition. The renormalization of the equation of state, the transport coefficients~\cite{Kovtun:2011np,Chafin:2012eq}, and the proper treatment for the numerical simulations are still challenging and under development~\cite{Murase:2016rhl,Hirano:2018diu,Nahrgang:2017oqp,Bluhm:2018plm,Singh:2018dpk}. \underline{\textbf{ {Hydrodynamics-kinetics}}} evolves stochastic fluctuations with the deterministic kinetic equation for two-point correlation function, while treating the background evolution hydrodynamically~\cite{Akamatsu:2016llw,An:2019osr,Akamatsu:2017rdu,Akamatsu:2018vjr,Martinez:2018wia}. It introduces renormalized transport coefficients and the equation of state, which naturally absorbs the lattice spacing dependence in the numerical implementation of fluctuating hydrodynamics mentioned above. The hydrodynamics-kinetic approach could reproduce the long-time tails of the hydrodynamics fluctuations~\cite{Kovtun:2012rj,Chafin:2012eq,Martinez:2017jjf,Martinez:2018wia} and estimate the relevant scale of critical fluctuations near the critical point~\cite{Akamatsu:2018vjr}. Besides the two-point function, the kinetic equation for three-point and four-point functions has also been developed recently~\cite{Pratt:2019fbj, An:2020vri}. \underline{\textbf{Hydro+}} encodes the critical fluctuations in the kinetic equation of the slow modes and couples it with the hydrodynamic evolution~\cite{Stephanov:2017ghc}. Here, the slowest mode $\phi_{\bm{Q}}(t,\bm{x})$ is identified as the Wigner transformed two-point function of entropy per baryon, which is driven to out-of-equilibrium near the critical point. With the generalized EoS, shear viscosity and bulk viscosity are modified by the slow mode $\phi_{\bm{Q}}(t,\bm{x})$, dynamics of critical fluctuations naturedly couples with the evolving fluid background. Hydro+ has been numerically implemented in a simplified system for a rapidity-independent fireball undergoing radial flow~\cite{Rajagopal:2019xwg}, and ideal Gubser flow \cite{Du:2020bxp}. The dynamics of the slow modes $\phi_{Q}$ have negligible effects on the fluid evolution. The out-of-equilibrium corrections to generalized entropy is within 0.1\% in the numerical simulations. Besides the dynamics of the two-point function for the slowest modes, the extension of the hydro+ by involving other slow modes, called hydro++, are also under development~\cite{An:2019csj, An:2020jjk}. \underline{\textbf{Dynamical universal scaling}} comes from the competition between the relaxation rate of the slow mode near the critical point and the expanding rate of the evolving system. Due to the critical slowing down effects, the critical modes will be driven out of equilibrium, which lead to correlated regions with characteristic length scales, called Kibble-Zurek scales, defined by the correlation length at which these two rates equal to each other. It was realized that, within the framework of Kibble-Zurek mechanism (KZM), one could construct some universal variables near the critical point that are independent of some non-universal factors, such as the evolving trajectory of the system approaching the critical point~\cite{Francuz:2015zva,Nikoghosyan:2013fqa}. Recently, the universal behavior of non-conversed order parameter field and conserved quantities near the QCD critical point has been studied within the framework of model A and model B~\cite{Mukherjee:2016kyu,Wu:2018twy,Wu:2019qfz,Akamatsu:2018vjr}. It was found that the oscillating behavior for the higher cumulants of net protons can be drastically suppressed, which converge into approximate universal curves with these constructed Kibble-Zurek functions~\cite{Wu:2018twy}. \section{Softening and clustering across the first-order phase boundary} For the system evolving across the first-order phase boundary, the bulk matter tends to separate into the two coexisting phases with the development of instabilities. According to~\cite{Gibbs}, such processes in the coexistence region are classified into two categories: nucleation and spinodal decomposition. \underline{\textbf{Nucleation}} belongs to nonlinear instability, which requires the formation of a sufficient large nucleus of the stable phase to overcome the barrier of free energy between stable and metastable minimums. The classical nucleation theory has been developed many years ago~\cite{Becker}, together with the following extensions~\cite{Lothe,Cahn, Langer:1967ax, Langer:1969bc, Zeng:1991}. The essential point is the nucleation rate that describes the decay probability per time and per volume of a metastable state, $\Gamma \sim e^{-\Delta\mathcal{F}_c}$, is mainly determined by the cost of the free energy $\Delta\mathcal{F}_c=\Delta \mathcal{F}(R_c)$ (where $R_c$ is the critical size of the bubble formed in a system). In heavy-ion physics, the nucleation rate of hadronic bubbles formed in the QGP is estimated through extending the classical nucleation theory~\cite{Csernai:1992tj,Csernai:1992as,Csernai:1992bs,Zabrodin:1998dk,Mishustin:1998eq,Alamoudi:1999ti, Shukla:2000dx,Shukla:2001xv,Bessa:2008nw}, in which the supercooling (or reheating) and the probability of bubble formation are influenced by many factors, such as the barrier height between two minimums of free energy, viscosity, and the expansion scenario. The effects of the supercooling or reheating have recently been observed in dynamical model simulations with the first-order phase transition, such as model A based on Langevin dynamics~\cite{Jiang:2017mji} or N$\chi$FD~\cite{Nahrgang:2011mv,Nahrgang:2011vn,Paech:2003fe,Herold:2014zoa}. In the rapidly expanding fireball, such systems could further evolve into the unstable region where the spinodal instabilities become dominant, which will be addressed follow. \underline{\textbf{Spinodal decomposition}} belongs to linear instability with small fluctuation of the order parameter field growing instantly and uniformly throughout the volume when the system evolves into the negative curvature region of the free energy. The basic theory of spinodal decomposition is the Cahn-Hilliard equation~\cite{Cahn}, which has been widely studied in metallurgy. In heavy-ion physics, the spinodal decomposition has been investigated within the framework dissipative hydrodynamics using an approximate equation of state for the coexistence region~\cite{Randrup:2010ax,Randrup:2009gp,Steinheimer:2012gc,Steinheimer:2013xxa,Steinheimer:2013gla}. In this model, the nuclear spinodal separation between confined and deconfined phases was observed, and its experimental consequences, such as the increasing density moments arising from the spinodal amplification of spatial irregularities, have also been studied. Although some progress has been made during the past years (Please also refer to Refs~\cite{Skokov:2009yu, Pratt:2017lce}), the dynamics near the first-order phase transition is still poorly understood due to the complexities of the evolving system. For instance, systems at the unstable region are easily driven out of equilibrium, which requires an additional study on the validity of hydrodynamics near the first-order phase transition region. More attention and effort from the theoretical and phenomenological side are still needed to study and predict the related experimental observable. \section{Experimental observables at RHIC BES} This section will briefly review the experimental observable measured at the RHIC BES program phase I, primarily focusing on soft observables to probe the bulk properties of the QGP and observables to probe the critical point and the first-order phase transition. \subsection{Soft observables to probe the bulk properties of QGP} \underline{\textbf{Particle yields, $p_T$ spectra, and flow }} in Au+Au collisions at the RHIC BES program are important soft observables to probe the QGP properties and the QCD phase structure. More specifically, the $p_T$-integrated yield at mid-rapidity provides detailed information on hadronic chemistry, from which the chemical freeze-out temperature and chemical potentials can be extracted using the statistical hadronization model~\cite{Andronic:2017pug,Andronic:2008gu, Adamczyk:2017iwn}. The mean $p_T$ of identified particles are directly proportional to the amount of radial flow of the evolving system, which helps to elucidate the size of pressure gradients in the initial profiles, the speed of sound, and bulk viscosity of the evolving systems~\cite{Ryu:2015vwa,Song:2009rh}. Various flow observables are sensitive to the transport properties and initial state fluctuation of the QGP fireball~\cite{Gale:2012rq,Qiu:2012uy,Heinz:2013bua,Teaney:2013dta, Zhu:2016puf,Pang:2015zrq,Zhao:2017yhj,Zhao:2017rgg,Moravcova:2020wnf,Giacalone:2020byk}. A comparison of the $\sqrt{s}$ dependence of the elliptic flow with hydrodynamics + transport hybrid models showed that the effective shear viscosity increases at low collision energies \cite{Karpenko:2015xea}. The work~\cite{Shen:2020jwv} further demonstrated that the elliptic flow measurements as a function of collision energy and rapidity could set strong constraints on the $T$ and $\mu_B$ dependence of the QGP specific shear viscosity. Recently, the STAR collaboration found that the particle multiplicity scaled triangular flow $v_3$ as a function of collision energy exhibited a minimum at $\sqrt{s_\mathrm{NN}} \sim 20$\,GeV \cite{Adamczyk:2016exq}, which hinted at a softening of the equation of state around the minimum. However, this non-monotonic behavior can be reproduced by the event-by-event hybrid framework without a critical point \cite{Shen:2020gef}. Therefore, it is essential to understand the interplay among the duration of dynamical initialization, the variation of the speed of sound, and the $T$ and $\mu_B$ dependence of the specific shear viscosity. \underline{\textbf{The net proton rapidity distributions}} have been measured in a few experiments during the past years~\cite{E-802:1998xum,Ahle:1999in,Barrette:1999ry,Arsene:2009aa,Anticic:2010mp}. These measurements can elucidate the initial baryon stopping and the dynamics of conserved charge density currents inside the QGP. The latter is controlled by the medium's charge diffusion constants and heat conductivity. The baryon diffusion during the hydrodynamic phase evolves baryon charges from forwarding rapidity back to the central rapidity region~\cite{Denicol:2018wdp, Li:2018fow, Du:2019obx, Wu:2021ypv}. These dynamics are driven by the inward-pointing spatial gradients of $\mu_B/T$, which act against local pressure gradients. Because the net proton rapidity distribution is sensitive to both the initial state stopping and baryon diffusion \cite{Denicol:2018wdp}, independent experimental observables are needed to disentangle these two effects. Recently, charge balance functions were proposed to independently constrain the charge diffusion constants \cite{Pratt:2019pnd, Pratt:2021xvg}. A larger diffusion constant in the QGP medium leads to wider azimuthal distribution for the $K^+K^-$ and $p\bar{p}$ correlations. \subsection{Observables to probe the critical point and the first-order phase transition} \underline{\textbf{Multiplicity fluctuations of net protons}} serves as a proxy for the net-baryon fluctuations, which is regarded as a promising experimental observable for searching the QCD critical point~\cite{Hatta:2003wn,Kitazawa:2012at,Kitazawa:2011wh}. The skewness and kurtosis of net proton distribution $S\sigma$ and $\kappa\sigma^2$, with the momentum coverage $|y|<0.5$ and $0.4<p_T<2$ GeV, have been systematically measured in Au+Au collisions from 7.7 to 200 A GeV~\cite{Adam:2020unf,Abdallah:2021fzj}, which shows potential features that may hint the critical point and first-order phase transition. With collision energy decreases, the net proton $\kappa\sigma^2$ first decreases and then increases with a minimum at $\sqrt{s_\mathrm{NN}} \sim 20$\,GeV. However, the error bars in the RHIC BES I measurements below 19.6 GeV remain too large to confirm any non-monotonic $\sqrt{s}$ dependence. \underline{\textbf{Light nuclei productions}} are argued to sensitive to the relative density fluctuation at the freeze-out surface, which is proposed as a sensitive observable to probe the QCD critical point and the first-order phase transition~\cite{Sun:2017xrx,Sun:2018jhg}. Recently, the STAR Collaboration has systematically measured the yields of light nuclei at the RHIC BES program and found that the coalescence parameters of (anti-)deuteron and the yield ratio of light nuclei, $N_t\cdot N_p/N_d^2$ show a non-monotonic energy dependence with a dip and a peak around $\sqrt{s_\mathrm{NN}}=20$\,GeV in central Au+Au collisions, respectively~\cite{Adamczyk:2016gfs,Adam:2019wnb,Zhang:2019wun}. Dynamical models without critical point or first-order phase transition fail to reproduce the non-monotonic behavior of light nuclei ratio ~\cite{Liu:2019nii,Sun:2020uoj,Deng:2020zxo,Zhao:2020irc}. It was recently found that the attractive and repulsive interaction between nucleons near the critical point was found to play an important role in the clustering of the light nuclei \cite{Shuryak:2018lgd,Shuryak:2019ikv,Shuryak:2020yrs,DeMartini:2020hka,DeMartini:2020anq}. Ref.~\cite{Sun:2020pjz} also study the light nuclei production near first-order phase transition within the framework of transport model that includes a first-order chiral phase transition and found the enhanced yield ratio near the transition region. Last but not least, a recent work \cite{Oliinychenko:2020znl} pointed out that the weak decay corrections to the proton spectrum need careful attention as it could potentially contaminated the collision energy dependence of the measured $N_t\cdot N_p/N_d^2$ ratio. \section{Outlook} Over the past two decades, the collective advancements in developing quantitative theoretical frameworks for relativistic nuclear collisions and high-quality flow measurements at top RHIC and the LHC energies have been driving our field to a precision era. With the increasing complexity in theoretical models, statistics and data science techniques such as Bayesian Inference have become a standard tool, which systematically constrains the QGP transport properties and initial-state fluctuations from various available soft observables~\cite{Everett:2020yty, Everett:2020xug,Moreland:2018gsh,Bernhard:2019bmu}. The RHIC Beam Energy Scan program has excited a wave of theoretical developments on dynamical modeling with full (3+1)D simulations that go beyond Bjorken's boost-invariant assumption. With dynamical initialization schemes which interweave the initial 3D collision dynamics with hydrodynamic simulations, we are starting to quantify initial baryon stopping and study the collectivity of the QGP in a baryon-rich environment. Confronting flow measurements from the upcoming RHIC BES program phase II and future FAIR/NICA experiments, we can elucidate how the QGP transport properties change as a function of net baryon density. We also expect that the Bayesian analysis with the BES phase II measurements at RHIC will quantitatively characterize the QGP viscosity and charge diffusion constant. In the meantime, experiment measurements on the multiplicity fluctuations of net proton and light nuclei productions at the RHIC BES program, together with the model studies, hint for the possible existence of the critical point and the first-order phase transition boundary. Quantitative description of these measurements requires more realistic and sophisticated model development on the theoretical side. Especially, the phenomenological model with the critical slowing down and first-order phase transition effects, together with a proper equation of state for the hot QCD system, needs to be developed consistently with the expanding collective background for such short-lived and tiny fireballs created at heavy-ion collisions. In parallel, high statistic measurements from the coming RHIC BES phase II could provide more confidence on the potential critical point signals. Furthermore, the larger rapidity acceptance and collisions at lower energies through fixed target setups in the RHIC BES II allowsfor detecting a broader range of the QCD phase diagram than that in the phase I program. Together with the description from sophisticated dynamical models, our community will provide deeper insights into the QCD phase structure. \\ \textbf{Acknowledgments} S.W. and H.S. is supported by the NSFC under grant No. 12075007 and No. 11675004. S.W. is also supported by the China Postdoctoral Science Foundation under Grant No. 2020M680184 and NSFC under grant No. 11947236. C.S. is supported by the U.S. Department of Energy under grant number DE-SC0013460 and the National Science Foundation under grant number PHY-2012922. This work is also supported in part by the U.S. Department of Energy, Office of Science, Office of Nuclear Physics, within the framework of the Beam Energy Scan Theory (BEST) Topical Collaboration. \bibliographystyle{apsrev}

8ba68a54bc2d90400c61ef8fa316ae5d77ca1259

\section{A Brief History and Introduction} The Riemann zeta function $\zeta(s)$ is one of the most important special functions of Mathematics. While the critical strip $0 < \Re\left(s\right) < 1$ is undoubtedly the most important region in the complex plane on account of the unsolved problem regarding location of non-trivial zeros of $\zeta(s)$, namely, the Riemann Hypothesis, the right-half plane $\Re\left(s\right) > 1$ also has its own share of interesting unsolved problems to contribute to. It is quite well known that many number theoretic properties of odd zeta values are nowadays still unsolved mysteries, such as the rationality, transcendence and existence of closed-forms. Only in 1978 did Ap\'ery \cite{Apery} famously proved that $\zeta(3)$ is irrational. This was later reproved in a variety of ways by several authors, in particular Beukers \cite{Beukers} who devised a simple approach involving certain integrals over $[0, 1]^3$. In the early 2000s, an important work of Rivoal \cite{Rivol}, and Ball and Rivoal \cite{Brivol} determined that infinitely many values of $\zeta$ at odd integers are irrational, and the work of Zudilin \cite{Zudilin} proved that at least one among $\zeta(5), \zeta(7), \zeta(9)$ and $\zeta(11)$ is irrational. A very recent result due to Rivoal and Zudilin \cite{RW1} states that at least two of the numbers $\zeta(5), \zeta(7), \ldots , \zeta(69)$ are irrational. Moreover, for any pair of positive integers $a$ and $b$, Haynes and Zudilin \cite[Theorem 1]{Haynes} have shown that either there are infinitely many $m \in \mathbb{N}$ for which $\zeta(am + b)$ is irrational, or the sequence $\{q_m\}_{m=1}^{\infty}$ of common denominators of the rational elements of the set $\{\zeta(a + b),\ldots, \zeta(am + b)\}$ grows super-exponentially, that is, $q_m^{1/m} \to \infty$ as $m \to \infty$. Despite these advances, to this day no value of $\zeta(2n + 1)$ with $n\geqslant 2$ is known to be irrational. A folklore conjecture states that the numbers $\pi$, $\zeta(3), \zeta(5), \zeta(7), \ldots$ are algebraically independent over the rationals. This conjecture is predicted by the Grothendieck’s period conjecture for mixed Tate motives. But both conjectures are far out of reach and we do not even know the transcendence of a single odd zeta value. One should mention that Brown \cite{Brown} has in the past few years outlined a simple geometric approach to understand the structures involved in Beukers’s proof of irrationality of $\zeta(3) $ and how this may generalize to other odd zeta values. Ramanujan made many beautiful and elegant discoveries in his short life of 32 years. One of the most remarkable formulas suggested by Ramanujan that has attracted the attention of several mathematicians over the years is the following intriguing identity involving the odd values of the Riemann zeta function \cite[1.2]{Berndt}: \begin{thm}[Ramanujan's Formula for $ \zeta(2m+1) $]\label{MainThm} If $ \alpha $ and $ \beta $ are positive numbers such that $ \alpha\beta = \pi^2 $ and if $ m $ is a positive integer, then we have \begin{align}\alpha^{-m}\,\left\{\dfrac{1}{2}\,\zeta(2m + 1) + \sum_{n = 1}^{\infty}\dfrac{n^{-2m - 1}}{e^{2\alpha n} - 1}\right\} &-(- \beta)^{-m}\,\left\{\dfrac{1}{2}\,\zeta(2m + 1) + \sum_{n = 1}^{\infty}\dfrac{n^{-2m - 1}}{e^{2\beta n} - 1}\right\}\nonumber\\ &=2^{2m}\sum_{k = 0}^{m + 1}\dfrac{\left(-1\right)^{k-1}B_{2k}\,B_{2m - 2k+2}}{\left(2k\right)!\left(2m -2k+2\right)!}\,\alpha^{m - k + 1}\beta^k. \label{maineq}\end{align} where $ B_m $ denotes the $ m $-th Bernoulli number with the exponential generating function $$\sum_{m\geqslant0}B_{m}\frac{z^m}{m!}=\dfrac{z}{\exp\left(z\right)-1}.$$ \end{thm} Theorem \ref{MainThm} appears as Entry 21 in Chapter 14 of Ramanujan's second notebook \cite[173]{Notebooks 2}. The first published proof of Theorem \ref{MainThm} is due to S.L. Marulkar \cite{SL} although he was not aware that this formula can be found in Ramanujan’s Notebooks. Grosswald too rediscovered this formula and studied it more generally in \cite{Gross1}, \cite{Gross2}. Berndt \cite[Theorem 2.2]{BB1} derived a general formula from which both Euler’s identity for even zeta values (1.1) and Ramanujan’s identity for odd zeta values follow as special cases, thus showing that Euler’s and Ramanujan’s formulas are natural companions of each other. Ramanujan's identity has a number of applications. For example, a contemporary interpretation of the above identity, as given for instance in \cite{Encode}, is that it encodes fundamental transformation properties of Eisenstein series on the full modular group and their Eichler integrals. This observation is extended in \cite[Section 5]{BBAS} to weight $2k + 1$ Eisenstein series of level $2$ through secant Dirichlet series. Moreover, Ramanujan’s identity also has applications in theoretical computer science \cite{CS} in the analysis of special data structures and algorithms. More specifically, it is used there to achieve certain distribution results on random variables related to data structures called \emph{tries}. \vspace{0.03in} \\ B. Berndt and A. Straub in their article \cite[3.1]{Berndt} show that identity (\ref{maineq}) is equivalent to \begin{align}\alpha^{-m}\sum_{n = 1}^{\infty}\dfrac{\coth\left(\alpha n\right)}{n^{2m+1}} &- (-\beta)^{-m}\sum_{n = 1}^{\infty}\dfrac{\coth\left(\beta n\right)}{n^{2m+1}}\nonumber \\&= - 2^{2m+1}\sum_{k = 0}^{m + 1}\dfrac{\left(-1\right)^{k}B_{2k}\,B_{2m + 2 - 2k}}{\left(2k\right)!\left(2m + 2 - 2k\right)!}\,\alpha^{m + 1 - k}\beta^k.\end{align} Letting $ \alpha = \beta = \pi $ and replacing $ m $ with $ 2m + 1 $, we get the following \cite[3.2]{Berndt}: \begin{prop} \label{special_case} The following identity holds \begin{equation}\label{eqspecial}\sum_{n = 1}^{\infty}\dfrac{\coth\left(\pi n\right)}{n^{4m + 3}} = 2^{4m+2}\pi^{4m + 3}\sum_{k = 0}^{2m + 2}\dfrac{\left(-1\right)^{k+1}B_{2k}\,B_{4m + 4 -2k}}{\left(2k\right)!\left(4m+4-2k\right)!}\end{equation} where as before $ m \in \mathbb{N} $ and $ B_m $ denotes the $ m $-th Bernoulli number. \end{prop} This variation was apparently first established by M. Lerch \cite{Lerch}. Later proofs were given by G.N. Watson \cite{Watson}, H.F. Sandham \cite{Sadham}, J.R. Smart \cite{Smart} and F.P. Sayer \cite{Sayer}. The main purpose of this article is to present an elementary proof of Theorem \ref{MainThm} and Proposition \ref{special_case}. Our proof solely relies on a Mittag-Leffler type expansion for hyperbolic cotangent function and Euler's identity for even zeta value \section{Preliminaries} \begin{prop}[Euler's Formula for $ \zeta(2m) $] \label{prop21} For $m \geqslant 1$, we have \begin{equation}\label{zeta2m}\zeta(2m) = \left(2\pi\right)^{2m}\dfrac{\left(-1\right)^{m - 1}B_{2m}}{2\left(2m\right)!}.\end{equation} \end{prop} Euler's identity for even zeta values (\ref{zeta2m}) not only provides an elegant formula for evaluating $ \zeta(2m) $ for any integer $ m \geqslant 1 $, but it also gives us information about the arithmetical nature of even zeta values. For a fascinating account of the history and proof of this formula, we refer the reader to \cite{Apostol}. \begin{prop}[Mittag-Leffler expansion for hyperbolic cotangent function] \label{cor23} Let $ x \in \mathbb{C} \setminus{\{0\}} $, then the following identity holds \begin{equation}\label{ML2} \coth\left(\pi x\right)=\dfrac{1}{\pi x}+\dfrac{2x}{\pi}\sum_{k=1}^{\infty}\dfrac{1}{x^2+k^2}. \end{equation} \end{prop} Note that equation (\ref{ML2}) appears as identity 1.421.4 in \cite{GR}. \section{Proof of Proposition \ref{special_case}} Replacing $ x $ with $ n $, dividing both sides of identity $ (\ref{ML2}) $ by $ n^{4m+2} $ and summing over $ n $ produces \[\dfrac{1}{2}\sum_{n = 1}^{\infty}\dfrac{\pi\coth(\pi n)}{n^{4m + 3}} - \dfrac{1}{2}\sum_{n = 1}^{\infty}\dfrac{1}{n^{4m + 4}} = \sum_{n = 1}^{\infty}\sum_{k = 1}^{\infty}\dfrac{1}{n^{4m + 2}\,(k^2 + n^2)}.\] Since the above double sum is absolutely convergent, we may switch the order of summation to get \[\dfrac{1}{2}\left(\sum_{n = 1}^{\infty}\dfrac{\pi\coth(\pi n)}{n^{4m + 3}} - \sum_{n = 1}^{\infty}\dfrac{1}{n^{4m + 4}}\right) = \sum_{k = 1}^{\infty}\sum_{n = 1}^{\infty}\dfrac{1}{n^{4m + 2}\,(k^2 + n^2)}.\] Next, we simply notice that \[ \dfrac{1}{n^{4m + 2}\left(k^2 + n^2\right)} = \dfrac{1}{n^{4m}k^2}\left(\dfrac{1}{n^2} - \dfrac{1}{k^2 + n^2}\right).\] Thus, we have \[ \sum_{n = 1}^{\infty}\sum_{k = 1}^{\infty}\dfrac{1}{n^{4m + 2}\left(k^2 + n^2\right)}= \zeta(2)\,\zeta(4m + 2) - \sum_{k = 1}^{\infty}\sum_{n = 1}^{\infty}\dfrac{1}{n^{4m}k^2\left(k^2 + n^2\right)}.\] Similarly, we find that \[\dfrac{1}{n^{4m}k^2\,(k^2 + n^2)} = \dfrac{n^2}{n^{4m}k^4}\left(\dfrac{1}{n^2} - \dfrac{1}{k^2 + n^2}\right).\] Thus, we have \[\sum_{k = 1}^{\infty}\sum_{n = 1}^{\infty}\dfrac{1}{n^{4m}k^2\left(k^2 + n^2\right)}= \zeta(4)\,\zeta(4m) - \sum_{k = 1}^{\infty}\sum_{n = 1}^{\infty}\dfrac{1}{n^{4m - 2}k^4\left(k^2 + n^2\right)} .\] This recursive pattern that produces even zeta values continues as follows: \[\sum_{k = 1}^{\infty}\sum_{n = 1}^{\infty}\dfrac{1}{n^{4m - 2}k^4\left(k^2 + n^2\right)} = \zeta(6)\,\zeta(4m-2) - \sum_{k = 1}^{\infty}\sum_{n = 1}^{\infty}\dfrac{1}{n^{4m - 4}k^6\left(k^2 + n^2\right)}\] \[\sum_{k = 1}^{\infty}\sum_{n = 1}^{\infty}\dfrac{1}{n^{4m - 4}k^6\left(k^2 + n^2\right)} = \zeta(8)\,\zeta(4m-4) - \sum_{k = 1}^{\infty}\sum_{n = 1}^{\infty}\dfrac{1}{n^{4m - 6}k^8\left(k^2 + n^2\right)}\] and so on. At the end, we get \[\sum_{k = 1}^{\infty}\sum_{n = 1}^{\infty}\dfrac{1}{n^{2}k^{4m}\left(k^2 + n^2\right)} = \zeta(4m + 2)\,\zeta(2) - \sum_{k = 1}^{\infty}\sum_{n = 1}^{\infty}\dfrac{1}{k^{4m + 2}\left(k^2 + n^2\right)}.\] Therefore, we obtain the final result \begin{align} \sum_{k = 1}^{\infty}\sum_{n = 1}^{\infty}\dfrac{1}{n^{4m + 2}\left(k^2 + n^2\right)}&= \zeta(2)\,\zeta(4m + 2) - \zeta(4)\,\zeta(4m) + \cdots + \zeta(4m + 2)\,\zeta(2) \nonumber \\ &- \sum_{k = 1}^{\infty}\sum_{n = 1}^{\infty}\dfrac{1}{k^{4m + 2}\left(k^2 + n^2\right)}\label{convoeq} \end{align} which can be easily proved by induction on $m$. Next, we observe that, by symmetry, \[\sum_{k = 1}^{\infty}\sum_{n = 1}^{\infty}\dfrac{1}{n^{4m + 2}\left(k^2 + n^2\right)} = \sum_{k = 1}^{\infty}\sum_{n = 1}^{\infty}\dfrac{1}{k^{4m + 2}\left(k^2 + n^2\right)}.\] Therefore we have \begin{align*}2\sum_{k = 1}^{\infty}\sum_{n = 1}^{\infty}\dfrac{1}{n^{4m + 2}\left(k^2 + n^2\right)} &= \left(-1\right)^m \zeta(2m+2)\,\zeta(2m+2) \\&+ 2\sum_{k=0}^{m-1} \left(-1\right)^k \zeta(4m+2-2k)\,\zeta(2k+2).\end{align*} Putting all things together produces \begin{align*}\sum_{n = 1}^{\infty}\dfrac{\pi\coth(\pi n)}{n^{4m + 3}}&= \zeta(4m + 4) + (-1)^m \zeta(2m+2)\,\zeta(2m+2) \\&+ 2\sum_{k=1}^m(-1)^{k-1}\,\zeta(4m-2k+4)\,\zeta(2k)\end{align*} where $k$ has been replaced with $k-1$ in the summand. Thus, it suffices to prove that \begin{align} \zeta(4m + 4) + \left(-1\right)^m &\zeta(2m+2)\,\zeta(2m+2) + 2\sum_{k=1}^m\left(-1\right)^{k-1}\zeta(4m-2k+4)\,\zeta(2k) \nonumber\\ = \,& \,2^{4m+2}\pi^{4m+4}\sum_{k = 0}^{2m+2}\dfrac{\left(-1\right)^{k-1}B_{2k}\,B_{4m+4-2k}}{\left(2k\right)!\left(4m+4-2k\right)!}. \end{align} After replacing $m$ with $m - 1 $ in the above equation, it suffices to prove that \begin{align} \zeta(4m) + \left(-1\right)^{m-1}&\zeta(2m)\,\zeta(2m) + 2\sum_{k=1}^{m-1}\left(-1\right)^{k-1}\zeta(4m-2k)\,\zeta(2k)\nonumber \\=\,&\,2^{4m-2}\pi^{4m}\sum_{k = 0}^{2m}\dfrac{\left(-1\right)^{k-1}B_{2k}\,B_{4m-2k}}{\left(2k\right)!\left(4m-2k\right)!}\label{Berno1} \end{align} which is indeed just a matter of substitution; using Euler's identity for $\zeta(2m)$, we have \begin{align} &\zeta(4m) + \left(-1\right)^{m-1}\zeta(2m)\,\zeta(2m) + 2\sum_{k=1}^{m-1}(-1)^{k-1}\,\zeta(4m-2k)\,\zeta(2k) \nonumber\\&= 2^{4m-2}\pi^{4m}\left\{-\dfrac{B_{4m}}{\left(4m\right)!} + \dfrac{\left(-1\right)^{m-1}B_{2m}\,B_{2m}}{\left(2m\right)!\left(2m\right)!} +2\sum_{k=1}^{m-1}\dfrac{\left(-1\right)^{k-1}B_{2k}\,B_{4m-2k}}{\left(2k\right)!\left(4m-2k\right)!}\right\}\label{Berno2} \end{align} The summand in equation (\ref{Berno1}) is invariant under the substitution $ k \mapsto 2m - k $, so if we split it according to $k \in \{0, 2m\},\, k \in \{1,2,\ldots,m + 1, \ldots, 2m - 1\}$ and $k = m$, we get \begin{equation}\sum_{k = 0}^{2m}\dfrac{\left(-1\right)^{k-1}B_{2k}\,B_{4m-2k}}{\left(2k\right)!\left(4m-2k\right)!}= -\dfrac{B_{4m}}{\left(4m\right)!} + \dfrac{\left(-1\right)^{m-1}B_{2m}\,B_{2m}}{\left(2m\right)!\left(2m\right)!} +2\sum_{k=1}^{m-1}\dfrac{\left(-1\right)^{k-1}B_{2k}\,B_{4m-2k}}{\left(2k\right)!\left(4m-2k\right)!}.\label{rfreq}\end{equation} Combining equations (\ref{Berno1}), (\ref{Berno2}) and (\ref{rfreq}) gives us the desired result. \hspace{\stretch{1}} $\blacksquare$ \section{Proof of Theorem \ref{MainThm}} Substituting $x\mapsto\left(\alpha k/\pi\right)$, with $\alpha\in\mathbb{R}_{>0}$, in identity (\ref{ML2}) yields \begin{equation}\label{ML1} \dfrac{1}{e^{2\alpha k}-1}=-\dfrac{1}{2} + \dfrac{1}{2\alpha k} + \sum_{m=1}^{\infty}\dfrac{\alpha k}{\pi^{2}m^{2}+\alpha^{2}k^{2}}. \end{equation} Dividing both sides of the above identity by $\alpha^{n}k^{2n+1}$ and summing over $k$ produces \[\alpha^{-n}\left\{ \sum_{k=1}^{\infty}\dfrac{k^{-2n-1}}{e^{2\alpha k}-1}+\dfrac{1}{2}\,\zeta(2n+1)\right\} =\dfrac{\zeta(2n+2)}{2\alpha^{n+1}}+\sum_{k=1}^{\infty}\sum_{m=1}^{\infty}\dfrac{\alpha^{n+1}}{\left(\alpha^{2}k^{2}\right)^{n}\left(\pi^{2}m^{2}+\alpha^{2}k^{2}\right)}.\] Next, we notice that, for an arbitrary positive integer $n$ and two arbitrary non-vanishing sequences $\{x_{k}\}$ and $\{z_{m}\}$, we have \begin{equation}\label{MM1} \sum_{k=1}^{\infty}\sum_{m=1}^{\infty}\frac{1}{x_{k}^{n}\left(x_{k}+z_{m}\right)}=\sum_{k=1}^{\infty}\frac{1}{x_{k}^{n}}\sum_{m=1}^{\infty}\frac{1}{z_{m}}-\sum_{k=1}^{\infty}\sum_{m=1}^{\infty}\frac{1}{x_{k}^{n-1}z_{m}\left(x_{k}+z_{m}\right)}\end{equation} assuming that the sums involved are convergent. Using the same recursive technique as in the previous section, we find that \[\sum_{k=1}^{\infty}\sum_{m=1}^{\infty}\frac{1}{x_{k}^{n}\left(x_{k}+z_{m}\right)}=\sum_{k=1}^{\infty}\frac{1}{x_{k}^{n}}\sum_{m=1}^{\infty}\frac{1}{z_{m}}-\sum_{k=1}^{\infty}\frac{1}{x_{k}^{n-1}}\sum_{m=1}^{\infty}\frac{1}{z_{m}^{2}}\] \begin{equation}\label{MM2}+\sum_{k=1}^{\infty}\frac{1}{x_{k}^{n-2}}\sum_{m=1}^{\infty}\frac{1}{z_{m}^{3}}+\cdots+(-1)^{n-1}\sum_{k=1}^{\infty}\frac{1}{x_{k}}\sum_{m=1}^{\infty}\frac{1}{z_{m}^{n}}+(-1)^{n}\sum_{k=1}^{\infty}\sum_{m=1}^{\infty}\frac{1}{z_{m}^{n}\left(x_{k}+z_{m}\right)}. \end{equation} For notational convenience, set \[ \sum_{k=1}^{\infty}\frac{1}{x_{k}^{n}}=\zeta_{x}(n),\quad\sum_{k=1}^{\infty}\frac{1}{z_{m}^{n}}=\zeta_{z}(n). \] These are also known as \emph{quasisymmetric zeta function} (the correct analog of a Riemann zeta function in the ring of quasisymmetric functions). Thus, equation (\ref{MM2}) can now be rewritten as \begin{equation}\label{MM3} \sum_{k=1}^{\infty}\sum_{m=1}^{\infty}\frac{1}{x_{k}^{n}\left(x_{k}+z_{m}\right)}=\sum_{p=0}^{n-1}(-1)^{p}\zeta_{x}(n-p)\,\zeta_{z}(p+1)+(-1)^{n}\sum_{k=1}^{\infty}\sum_{m=1}^{\infty}\frac{1}{z_{m}^{n}\left(x_{k}+z_{m}\right)} \end{equation} Substituting $x_{k}\mapsto\alpha^{2}k^{2}$ and $z_{m}\mapsto\pi^{2}m^{2}$ in identity (\ref{MM3}) produces \begin{align} \sum_{k=1}^{\infty}\sum_{m=1}^{\infty}\dfrac{1}{\left(\alpha^2k^2\right)^n\left(\alpha^2k^2 + \pi^2m^2\right)} &=\sum_{p = 0}^{n-1}\left(-1\right)^{p}\dfrac{\zeta(2n-2p)}{\alpha^{2n-2p}}\,\dfrac{\zeta(2p + 2)}{\pi^{2p + 2}}\nonumber \\&+ \left(-1\right)^n\sum_{k=1}^{\infty}\sum_{m=1}^{\infty}\dfrac{1}{\left(\pi^2m^2\right)^n\left(\alpha^2k^2 + \pi^2m^2\right)}.\label{MM4} \end{align} Euler's formula for $\zeta(2m)$ allows us to transform the zetas into Bernoullis as follows: \[\sum_{p = 0}^{n-1}\left(-1\right)^{p}\dfrac{\zeta(2n-2p)}{\alpha^{2n-2p}}\dfrac{\zeta(2p + 2)}{\pi^{2p + 2}} = \left(-1\right)^n2^{2n}\sum_{p = 0}^{n-1}\left(\dfrac{\pi}{\alpha}\right)^{2n-2p}\dfrac{\left(-1\right)^{p-1}B_{2n - 2p}\,B_{2p + 2}}{\left(2n - 2p\right)!\left(2p + 2\right)!}.\] Moreover, using $ \alpha\beta = \pi^2 $ and replacing $ p $ with $p - 1$ in the summand produces \[\sum_{p = 0}^{n-1}\left(-1\right)^{p}\,\dfrac{\zeta(2n-2p)}{\alpha^{2n-2p}}\dfrac{\zeta(2p + 2)}{\pi^{2p + 2}} = \left(-1\right)^n2^{2n}\sum_{p = 1}^{n}\left(\dfrac{\beta}{\alpha}\right)^{n-p+1}\dfrac{\left(-1\right)^{p}B_{2n - 2p+2}\,B_{2p}}{\left(2n - 2p+2\right)!\left(2p\right)!}.\] Substituting this expression in equation (\ref{MM4}) we get \begin{align} \sum_{k=1}^{\infty}\sum_{m=1}^{\infty}\dfrac{1}{\left(\alpha^2k^2\right)^n\left(\alpha^2k^2 + \pi^2m^2\right)} & = \left(-1\right)^n2^{2n}\sum_{p = 1}^{n}\left(\dfrac{\beta}{\alpha}\right)^{n-p+1}\dfrac{\left(-1\right)^{p}B_{2n - 2p+2}\,B_{2p}}{\left(2n - 2p+2\right)!\left(2p\right)!}\nonumber \\&+ \left(-1\right)^n \sum_{k=1}^{\infty}\sum_{m=1}^{\infty}\dfrac{1}{\left(\pi^2m^2\right)^n\left(\alpha^2k^2 + \pi^2m^2\right)}\label{MM5} \end{align} Putting all things together produces \[\alpha^{-n}\left\{\sum_{k = 1}^{\infty}\dfrac{k^{-2n-1}}{e^{2\alpha k}-1} + \dfrac{1}{2}\,\zeta(2n+1)\right\}=\left(-1\right)^n\alpha^{n+1}\sum_{k=1}^{\infty}\sum_{m=1}^{\infty}\dfrac{1}{(\pi^2m^2)^n(\alpha^2k^2 + \pi^2m^2)}\] \[+\, \dfrac{\zeta(2n+2)}{2\alpha^{n+1}} + \alpha^{n+1}\left\{= \left(-1\right)^n2^{2n}\sum_{p = 1}^{n}\left(\dfrac{\beta}{\alpha}\right)^{n-p+1}\dfrac{\left(-1\right)^{p}B_{2n - 2p+2}\,B_{2p}}{\left(2n - 2p+2\right)!\left(2p\right)!}\right\}.\] Next, we notice that the remaining zeta term \[\dfrac{\zeta(2n+2)}{2\alpha^{n+1}} = \dfrac{\left(-1\right)^n}{2\alpha^{n+1}}\dfrac{ \left(2\pi\right)^{2n+2}B_{2n+2}}{2\left(2n+2\right)!} = \dfrac{\left(-1\right)^n\beta^{n+1}2^{2n}B_{2n+2}}{\left(2n+2\right)!}\] is the missing $ p = 0 $-th term in the summand. Thus, we have \begin{align}\alpha^{-n}\left\{ \sum_{k=1}^{\infty}\dfrac{k^{-2n-1}}{e^{2\alpha k}-1}+\dfrac{1}{2}\,\zeta(2n+1)\right\} &- \alpha^{n+1}\left\{ \dfrac{\left(-1\right)^{n}}{\pi^{2n}}\sum_{k=1}^{\infty}\sum_{m=1}^{\infty}\dfrac{1}{m^{2n}\left(\alpha^{2}k^{2}+m^{2}\pi^{2}\right)}\right\}\nonumber \\&=\left(-1\right)^{n}2^{2n}\sum_{p=0}^{n}\dfrac{\left(-1\right)^{p}B_{2n-2p+2}\,B_{2p}}{\left(2n-2p+2\right)!\left(2p\right)!}\,\alpha^{p}\beta^{n-p+1}.\label{MM6} \end{align} Notice that, using $\alpha\beta=\pi^{2}$ we obtain \[ \sum_{k=1}^{\infty}\sum_{m=1}^{\infty}\dfrac{1}{m^{2n}\left(\alpha^{2}k^{2}+m^{2}\pi^{2}\right)}=\dfrac{1}{\alpha}\sum_{k=1}^{\infty}\sum_{m=1}^{\infty}\dfrac{1}{m^{2n}\left(\alpha k^{2}+\beta m^{2}\right)}=\dfrac{1}{\alpha}\sum_{k=1}^{\infty}\sum_{m=1}^{\infty}\dfrac{1}{k^{2n}\left(\alpha^{2}m^{2}+\beta k^{2}\right)}, \] where the last equality is obtained by replacing $k$ by $m$ and $m$ by $k$ so that we get \begin{equation}\label{MM7} \sum_{k=1}^{\infty}\sum_{m=1}^{\infty}\dfrac{1}{m^{2n}\left(\alpha^{2}k^{2}+m^{2}\pi^{2}\right)}=\dfrac{\beta}{\alpha}\sum_{k=1}^{\infty}\sum_{m=1}^{\infty}\dfrac{1}{k^{2n}\left(\beta^{2}k^{2}+m^{2}\pi^{2}\right)} \end{equation} Dividing both sides of identity (\ref{ML2}) by $k^{2n+1}$ and summing over $k$ produces \begin{equation}\label{MM8} \beta\sum_{k=1}^{\infty}\sum_{m=1}^{\infty}\dfrac{1}{k^{2n}\left(\beta^{2}k^{2}+m^{2}\pi^{2}\right)}=\sum_{k=1}^{\infty}\dfrac{k^{-2n-1}}{e^{2\beta k}-1}+\dfrac{1}{2}\,\zeta(2n+1)-\dfrac{1}{2\beta}\,\zeta(2n+2) \end{equation} where we have replaced $\alpha$ with $\beta$. Thus, we have \[ \sum_{k=1}^{\infty}\sum_{m=1}^{\infty}\dfrac{1}{m^{2n}\left(\alpha^{2}k^{2}+m^{2}\pi^{2}\right)}=\dfrac{1}{\alpha}\left\{ \sum_{k=1}^{\infty}\dfrac{k^{-2n-1}}{e^{2\beta k}-1}+\dfrac{1}{2}\,\zeta(2n+1)-\dfrac{1}{2\beta}\,\zeta(2n+2)\right\}. \] Finally, substituting this expression in equation (\ref{MM6}) yields \begin{align*} & \alpha^{-n}\left\{ \sum_{k=1}^{\infty}\dfrac{k^{-2n-1}}{e^{2\alpha k}-1}+\dfrac{1}{2}\,\zeta(2n+1)\right\} -\left(-1\right)^{n}2^{2n}\sum_{p=0}^{n}\dfrac{\left(-1\right)^{p}B_{2n-2p+2}\,B_{2p}}{\left(2n-2p+2\right)!\left(2p\right)!}\,\alpha^{p}\,\beta^{n-p+1}\\ & =\dfrac{\left(-1\right)^{n}\alpha^{n}}{\pi^{2n}}\left\{ \sum_{k=1}^{\infty}\dfrac{k^{-2n-1}}{e^{2\beta k}-1}+\dfrac{1}{2}\,\zeta(2n+1)\right\} - \dfrac{\left(-1\right)^{n}\alpha^{n+1}}{\pi^{2n}}\left\{\dfrac{1}{2\alpha\beta}\,\zeta(2n+2)\right\ \end{align*} \[=(-\beta)^{-n}\left\{ \sum_{k=1}^{\infty}\dfrac{k^{-2n-1}}{e^{2\beta k}-1}+\dfrac{1}{2}\,\zeta(2n+1)\right\} -\dfrac{\left(-1\right)^{n}\alpha^{n+1}}{\pi^{2n+2}}\dfrac{\left(-1\right)^{n}2^{2n+2}\pi^{2n+2}B_{2n+2}}{4\left(2n+2\right)!}\] Next, we notice that the remaining term: \vspace{0.05in} \[ -\dfrac{\left(-1\right)^{n}\alpha^{n+1}}{\pi^{2n+2}}\dfrac{\left(-1\right)^{n}2^{2n+2}\pi^{2n+2}B_{2n+2}}{4\left(2n+2\right)!}=\dfrac{\left(-1\right)^{2n+1}2^{2n}B_{2n+2}}{\left(2n+2\right)!}\,\alpha^{n+1} \] is the $p=n+1$-st term in the summand. Therefore, we deduce that \begin{align} \alpha^{-n}\left\{ \sum_{k=1}^{\infty}\dfrac{k^{-2n-1}}{e^{2\alpha k}-1}+\dfrac{1}{2}\,\zeta(2n+1)\right\} &-\left(-\beta\right)^{-n}\left\{ \sum_{k=1}^{\infty}\dfrac{k^{-2n-1}}{e^{2\beta k}-1}+\dfrac{1}{2}\,\zeta(2n+1)\right\} \\ & =\left(-1\right)^{n}2^{2n}\sum_{p=0}^{n+1}\dfrac{\left(-1\right)^{p}B_{2n-2p+2}\,B_{2p}}{\left(2n-2p+2\right)!\left(2p\right)!}\,\alpha^{p}\beta^{n-p+1}.\nonumber \end{align} Replacing $p$ with $n-p+1$ in the above identity gives us the desired result. \hspace{\stretch{1}} $\blacksquare$ \section{Acknowledgements} The author would like to thank Christophe Vignat for his guidance and support throughout the completion of this work. The author is very much thankful to the reviewer for his/her valuable comments towards the improvement of the manuscript. The author would also like to thank Bruce Berndt and Atul Dixit for endorsing him on arXiv.

3582b5e208044058f35f7d320a68469b0fdff554

\section{Introduction} Lasers with stable, narrow frequency spectra have numerous applications in ultra-high precision measurements including state of the art clocks \cite{JILAlatticeClock, UshijimaLatticeClock, McGrewYbLatticeClock, SYRTEclocks, PTBclocks, INRIMclock, NPLclock} and gravitational wave detection \cite{LIGO, LISA}. These frequency references can be passive, in which case independently generated laser light is compared to an accurate reference. The state of the art performance of such systems is limited today by the thermal motion at the atomic level in the cavity mirrors \cite{Numata2004, Cole2013, MateiPRL2017, Robinson2019}. This limitation can be partially circumvented in an active reference, in which the system could consist of an atomic ensemble within an optical cavity \cite{KazakovEnsembles, KazakovColdAtoms, HollandLaser, ChenCaLaser, JingbiaoChen2005, ruggedLaser}. Here the idea is to pump the atoms to an excited state with a long lifetime relative to the decay of the cavity field, thus operating in the bad cavity regime \cite{ChenOE2019}. In this regime the cavity will enhance the emission rate by the atoms without pulling the frequency strongly to the cavity resonance, as the phase information is primarily stored in the atoms. This makes active frequency references operating in the bad cavity regime a promising technology for generating highly stable and reproducible frequencies, and has sparked interest in the physics of bad-cavity lasing.\\ Superradiance can be observed when an atomic ensemble is coupled to a preferred electromagnetic mode. This process was originally investigated theoretically in the 1950s by R. H. Dicke \cite{Dicke}, and subsequently observed in the optical regime in 1973 \cite{HFsuperradiance}. Coupling an atomic ensemble to a broad cavity resonance, superradiant emission can be realized deep in the bad cavity regime. Superradiant pulses are characterized by a hyperbolic secant shape in the time-domain. The dynamics of superradiant lasing pulses have been studied in the context of narrow optical transitions on the 375 Hz wide $^1$S$_0 - ^3$P$_1$ line in Calcium \cite{CalciumDelayStatistics} and on the mHz wide $^1$S$_0 - ^3$P$_0$ transition in Strontium \cite{NorciaClockPulses}, where the effects of cavity pulling have also been characterized \cite{NorciaClockFreq}. In systems where the atomic transition and cavity linewidths are comparable, a superradiant crossover regime can be realized, characterized by both atomic coherence and a cavity field population, which acts back on the atoms. As a result these pulses deviate from the hyperbolic secant shape. The dynamics in this regime have been studied in experiments using the $^1$S$_0 - ^3$P$_1$ line in Strontium confined in an optical lattice at $\upmu$K temperatures \cite{NorciaCrossover} and in our system at mK temperatures \cite{qmetLasing}. Though lower temperatures imply less inhomogeneous broadening from Doppler shifts, these low temperatures are technically demanding to achieve. Thus a superradiant laser using hotter atoms could show great technological promise if the influence of Doppler shifts can be suppressed in the lasing frequency. Such a scheme has recently been proposed for operation in the continuous regime \cite{ruggedLaser} and reduces the technical challenges compared to continuous superradiance based on ultracold atoms \cite{KazakovColdAtoms}.\\ In this paper we investigate the spectral properties and phase dynamics of an ensemble of unconfined $^{88}$Sr atoms emitting superradiant pulses of light via the mode of an optical cavity. In this system the temperature of the atoms results in a Doppler-broadened lasing transition linewidth that is larger than the cavity linewidth. Despite this, operating the system around a specific cavity detuning enables us to exploit the velocity-dependent behavior of moving atoms in order to suppress cavity pulling in the pulsed lasing regime. We present the experimental setup in Section II and the theoretical model of the system in Section III. This model is applied in the following sections to extract additional information about the physics at the atomic level, and simulations are shown along with the experimental results. We first describe the phase behavior of the lasing process in Section IV, then in Section V we present the spectral properties of the lasing pulses and relevant dynamics. In Section VI we analyze cavity pulling in the system and under which conditions this can be minimized, and in Section VII we evaluate the sensitivity of the system to cavity noise. In Section VIII we discuss improvements which are relevant for metrological applications. \section{Experimental system} The experimental system consists of a 5 mK ensemble of $^{88}$Sr atoms cooled and trapped in a 3D Magneto-Optical Trap (MOT) from a Zeeman-slowed atomic beam, using the $^1$S$_0 \leftrightarrow ^1$P$_1$ transition at 461 nm. The atomic cloud partially overlaps with an optical cavity and can be coherently pumped to $^3$P$_1$ using an off-axis pump laser, see Fig. \ref{fig:setup}. The cavity has a linewidth of $\kappa = 2\pi\times 620$~kHz and can be tuned to resonance with the $^1$S$_0 \leftrightarrow ^3$P$_1$ intercombination line ($\gamma = 2\pi\times 7.5$~kHz). For comparison the Doppler FWHM is $\Gamma_D = 2\pi \times 2.3$~MHz. Due to a cavity waist radius of $w_0 = 450$~$\upmu$m the number of atoms within the cavity, $N_{cav}$, is of order $10^7$, while the total number of atoms in the cloud, N, is of order $10^8$, as estimated from fluorescence measurements and lasing pulse intensity.\\ \begin{figure}[t] \includegraphics[width=\columnwidth]{Figures/setupBeat.pdf} \caption{(a) Schematic of the experimental system showing a thermal ensemble of strontium atoms partially overlapping with the cavity mode. Coherent pumping prepares the atoms in $^3$P$_1$ and allows subsequent emission of a lasing pulse into the cavity mode. Light couples out of both ends of the cavity with the cavity linewidth $\kappa$. At one end the cavity output is overlapped with a reference laser beam, and the beat signal is detected. (b) Level scheme for $^{88}$Sr showing the cooling and lasing transitions. Wavelength ($\lambda$) and natural transition linewidths ($\gamma$) are indicated.\label{fig:setup}} \end{figure} Once the atomic cloud is established, the slowing and cooling beams are turned off, and the atoms are pumped to $^3$P$_1$ with an approximate $\pi$-pulse of 190 ns from an angle of 45$^\circ$ with respect to the cavity axis. Due to the large dimension of the cloud and the finite temperature, the population transferred to $^3$P$_1$ is limited to an estimated 85 \% within the cavity mode. Following a delay of a few microseconds, a lasing pulse is emitted by the atoms into the cavity. To characterize the spectrum of the laser pulses we beat the light leaking from one side of the cavity with a laser beam which is generated independently from the setup and stabilized to a transportable cavity on loan from SYRTE \cite{SYRTEcavity}. The reference laser frequency is detuned by 171 MHz from the atomic resonance, and the detected beat signal is then mixed at 120 MHz. This yields a signal with a frequency near 51 MHz which is band pass-filtered and recorded with a fast oscilloscope. Such a beat signal is shown in Fig. \ref{fig:specEvolHeatmap}a, for the case where the cavity is detuned by 1 MHz from the atomic resonance. This promotes oscillatory energy exchange between the atoms and cavity mode which results in damped oscillations in the emitted power. In Fig. \ref{fig:specEvolHeatmap}(b) a Fourier transformation is used to extract the spectrum for increasing window sizes, with the start fixed at $t = 0$. This method showcases how the spectrum changes throughout the lasing process; we see bifurcations each time additional intensity oscillations enter the analysis window. We also see that the spectral peak is shifted further towards the atomic resonance frequency as the window is expanded.\\ \begin{figure}[t!] \includegraphics[width=\columnwidth]{Figures/specEvolHeatmap41.pdf} \caption{(a) Beat signal of a lasing pulse at a cavity detuning of 1 MHz. (b) Lasing intensity spectrum obtained by squaring the Fourier transformed beat signal for increasing window sizes, starting at $t = 0$. $\Delta_{le} = 0$ corresponds to the atomic transition frequency. Minima in emitted power are marked by green dashed lines.\label{fig:specEvolHeatmap}} \end{figure} \section{Modeling the laser spectrum} To model the laser pulses we numerically integrate a set of equations for the atomic and cavity states, derived in first order mean-field theory from a Tavis-Cummings Hamiltonian. The details of the model and lasing dynamics are described in \cite{qmetLasing}. To characterize the emitted spectrum in the model we extend the Hamiltonian with terms representing an array of imaginary filter cavities, following the method used by K. Debnath \emph{et al} \cite{Debnath}. \begin{eqnarray} H &=& \hbar \omega_c a^\dagger a + \sum_{j=1}^N \hbar \omega_{e} \sigma_{ee}^j + \sum_{k=1}^{N_f} \hbar \omega_{f}^{k} f_k^\dagger f_k \\\nonumber &+& \sum_{j=1}^N \hbar g_c^j \left( \sigma_{ge}^j + \sigma_{eg}^j \right) \left( a + a^\dagger \right)\\\nonumber &+& \sum_{j=1}^N \hbar \frac{\chi_p^j}{2} \left( \sigma_{ge}^j + \sigma_{eg}^j \right) \left( e^{ i \mathbf{k}_p \cdot \mathbf{r}_j - i \omega_p t } + e^{ -i \mathbf{k}_p \cdot \mathbf{r}_j + i \omega_p t } \right)\\\nonumber &+& \sum_{k=1}^{N_f} \hbar g_f \left( a + a^\dagger \right) \left( f_k + f_k^\dagger \right). \end{eqnarray} Here the lowering operators $a$, $f_k$ and $\sigma_{ge}^j$ represent the cavity field, the $k$'th filter cavity, and the $j$'th atom, respectively. The angular frequencies are given by $\omega_{c}$ for the cavity, $\omega_{e}$ for the atom transition, $\omega_{p}$ for the pump pulse laser and $\omega_{f}^k$ for the $k$'th filter cavity. $\mathbf{k_p}$ is the pump pulse wavevector and $\mathbf{r_j}$ is the position vector of the $j$'th atom at a given time. $g_c^j$ is the position-dependent atom-cavity coupling, $\chi^j_p$ is the Rabi frequency of the semiclassical pump pulse laser, and $g_f$ is the coupling between the cavity and filter cavities.\\ In this model we account for the motion and cavity coupling of each atom separately, due to the mK temperature and large spatial extent of the atomic cloud relative to the cavity waist. Since we describe expectation values our model neglects quantum noise in the system, which has little influence on the final spectrum due to the low Purcell decay rate [$\gamma_P = 1/($37 ms$)$] relative to coherent interactions. The filter cavities enable us to simulate the spectrum of the light leaking from the main cavity. Each of these have a very narrow linewidth $\kappa_F^j$ and is very weakly coupled (with coupling constant $g_f$) so that we can neglect any back-action on the lasing cavity. This appropriately models the experimental situation where laser photons do not re-enter the lasing cavity once they have left the output coupler towards the beat signal detector. The time evolution of the system operator expectation values then obey four distinct types of equations: \begin{eqnarray} \label{eq:OBE} \dot{\left\langle \sigma_{ge}^j \right\rangle} &=& -\left( i\Delta_{ep} + \frac{\gamma}{2} \right) \left\langle \sigma_{ge}^j \right\rangle \\\nonumber &&+ i \left( g_c^j \langle a \rangle + \frac { \chi_p^j}{2} e ^ { -i \mathbf{k}_p \cdot \mathbf{r}_j } \right) \left( \left\langle \sigma_{ee}^j \right\rangle - \left\langle \sigma_{gg}^j \right\rangle \right)\\\nonumber \dot{\left\langle \sigma_{ee}^j \right\rangle} &=& - \gamma \left\langle \sigma_{ee}^j \right\rangle + i \left( g_c^j \left\langle a^\dagger \right\rangle + \frac{\chi_p^j}{2} e^{ i \mathbf{k}_p \cdot \mathbf{r}_j} \right) \left\langle \sigma_{ge}^j \right\rangle\\\nonumber &&- i \left( g_c^j \langle a \rangle + \frac{\chi_p^j}{2} e^{ -i \mathbf {k}_p \cdot \mathbf{r}_j } \right) \left\langle \sigma_{eg}^j \right\rangle.\\\nonumber \dot {\langle a \rangle} &=& - \left( i \Delta_{cp} + \frac { \kappa } { 2 } \right) \langle a \rangle - \sum_{j=1}^N i g_c^j \left\langle \sigma_{ge}^j \right\rangle\\\nonumber \dot {\langle f_k \rangle} &=& - \left( i \Delta^k_{fp} + \frac { \kappa^k_f } { 2 } \right) \langle f_k \rangle - i g_f \left\langle a \right\rangle. \end{eqnarray} Here the detunings $\Delta_{ij} = \omega_i - \omega_j$ have been introduced and a rotating frame has been chosen at $\omega_p$. The intensity spectrum from system initiation up to a time $t$ is then given by the relative populations of the filter cavities at time $t$ as a function of their resonance frequencies, $\abs{\expval{f_k(\omega_f^k)}}^2$, with a normalization depending on $g_f$. The number of filter cavities, $N_f$ and the range of frequencies they span determines the spectral resolution of the simulation.\\ \section{Phase behavior} The phase coherence of the cavity field and its evolution is directly related to the spectral properties of the laser pulses and informs us about the underlying physical processes. The mean phase evolution can be traced in simulations from $\arg(\expval{a}$) in the semiclassical approximation. In Fig. \ref{fig:detc0phaseEvol} we show the simulated cavity output power for three random instances of lasing pulses when the cavity frequency is tuned to atom-resonance [Fig. \ref{fig:detc0phaseEvol}(a)] and the corresponding phase evolution of the cavity field [Fig. \ref{fig:detc0phaseEvol}(b)] in the rotating frame of the stationary atomic transition. The pump pulse initially imprints different phases on the Bloch vectors of all the atoms, depending on their positions and velocities. This prepares the ensemble with a non-zero macroscopic coherence, however it is tiny and randomized due to the 45$^{\circ}$ angle between pump pulse and cavity, and the random initial positions of the atoms at the wavelength level. Furthermore, in the semiclassical model a mean cavity field amplitude builds up during the pumping pulse corresponding to about 0.1 photons. Both the nonzero coherence and cavity field amplitude can separately lead to buildup of a lasing pulse. The atomic coherence can be removed after pumping experimentally by cycling the atoms on the $^1S_0$ - $^1P_1$ transition. If this is simulated, the remaining cavity field will imprint its random phase onto the atomic coherence as the power starts to build up. In the simulations we can also explicitly remove the field built up during pumping. In this case the randomized atomic coherence will trigger the dynamics. The resulting cavity field phase will depend on the macroscopically favored configuration originating from each atoms' state and coupling to the cavity mode. Keeping both the field built up during pumping and the initial random atomic coherence, the simulation shows both playing a role in the buildup of the laser pulse: after pumping, the phase is not stable, but the atoms build up coherence via the cavity field, which settles on a phase within 1 $\upmu$s after the pump pulse, as the lasing pulse starts to build up. At this point the output power has reached on the order of 0.1 nW, and the random delay between pumping and the lasing pulse is closely related to the time it takes for the phase to settle. With the choice of reference frame in Fig. \ref{fig:detc0phaseEvol}(b) a change in the phase of the cavity field over time corresponds to the cavity field being off resonance with the slowest atoms. This means the gain in the system is lower, slowing the pulse build-up and causing a longer delay of the pulse. A third process that can contribute to building up the lasing pulse is spontaneous emission into the cavity mode, which is not included in the model. With approximately $10^7$ atoms within the cavity waist, on the order of 60 photons are emitted into the cavity within 1 $\upmu$s given the Purcell rate of the system. This is the dominant process in triggering the lasing pulse in experiments, and is a source of phase noise that plays a role in the early stages of phase-synchronization.\\ \begin{figure}[t!] \includegraphics[width=\columnwidth]{Figures/detc0phaseEvol.pdf} \caption{Time evolution of the cavity output power (a) and cavity field phase (b) for three simulated laser pulses after a pump pulse (on during the red shaded interval). The three instances have the same ensemble parameters but show random variation. (a) The laser pulses have random delays and are followed by small intensity oscillations when the cavity is on atom-resonance. The blue intervals mark every even number oscillation of the pulse with shortest delay. (b) The phase is random after the pump pulse, then evolves and stabilizes as the laser power builds up. For every intensity oscillation the phase flips sign. (c) Calculating $\sin{(\phi)} \cdot \sqrt{P_{out}}$ illustrates how a demodulated beat signal looks when measured in experiment. (d) Moving mean of three experimentally measured demodulated beat signals (shaded: raw data). We see qualitative agreement with the simulated behavior. \label{fig:detc0phaseEvol}} \end{figure} Though the ensemble preparation, scattering and spontaneous emission can initiate lasing and have different influences on the phase, the subsequent dynamics and spectral properties are highly deterministic. The Bloch vectors of the atoms in the ensemble all oscillate with different Rabi frequencies, depending on the atoms' velocities and positions with respect to the cavity waist and standing wave. Initially most atoms emit light, but at peak output power, most atoms start to reabsorb. By the end of the first intensity oscillation [e.g. at $t = 3$ us for the blue traces in Fig. \ref{fig:detc0phaseEvol}(a) to (c)], the atoms are now out of phase due to their varying velocities and positions. The atoms that were absorbing at the end of the first lasing pulse begin to emit, and those that were emitting start to absorb, causing a $\pi$ phase shift of the second power oscillation. When the cavity is on atom-resonance, this pattern repeats during the subsequent intensity oscillations, as seen in Fig. \ref{fig:detc0phaseEvol}(b). We compare the simulations to measurements of the beat signal with an in-cavity reference laser on resonance with the cavity, but detuned 1 FSR from the atomic resonance, in Fig. \ref{fig:detc0phaseEvol}(c) (simulations) and \ref{fig:detc0phaseEvol}(d) (experiments). The simulation and experimental samples shown in this section are picked randomly from larger samples. We find that the measured beat signals also show the phase flipping behavior during the intensity oscillations, as the beat signals change sign between each oscillation.\\ In the off-resonant case, where the cavity is detuned from the atomic resonance and pronounced after-pulsing is observed, the phase evolution is different (see Fig. \ref{fig:detc1MphaseEvol}). To compare the phase evolution with our experiments, the phase shown here has been transformed to a rotating frame at $\omega_c$. We find that the phase shift is close to $\pi$ between the first and second intensity oscillation, where the output power goes to zero. This explains the splitting of the spectrum in the experimental measurement of Fig. \ref{fig:specEvolHeatmap}(b) during the second oscillation, and the dip forming at the peak frequency of the initial pulse (these dynamics occur between the two leftmost green dashed lines). In subsequent oscillations the power does not go to zero and the phase shifts occur more gradually. The chirp in frequency towards the atomic resonance is also evident in the rate of change of the phase, which increases as the spectrum is pulled closer to atomic resonance. \begin{figure}[t!] \includegraphics[width=\columnwidth]{Figures/detc1MphaseEvolFrameshift.pdf} \caption{Time evolution of the cavity output power (a) and cavity field phase (b) for two simulated laser pulses after a pump pulse (on during the red shaded interval) at a cavity detuning of 1 MHz. (a) Oscillations in the output power are pronounced when the cavity is detuned. The blue areas mark every even number oscillation of the pulse with shortest delay. (b) The phase evolution during the laser pulses in a rotating reference frame at $\omega_c$. The sloped behavior is related to pulling of the lasing frequency away from the cavity resonance and towards the atomic transition frequency. (c) Calculated and (d) moving mean of experimental demodulated beat signals (shaded: raw data). Purple intervals mark every second oscillation in the experimentally measured output power. \label{fig:detc1MphaseEvol}} \end{figure} \section{Lasing spectra and underlying dynamics} To compare simulations to our beat experiments in Fig. \ref{fig:specEvolHeatmap}, the simulated lasing dynamics of a pulse at a cavity detuning of 1 MHz is shown in Fig. \ref{fig:specEvolHeatmapSim}(a), and the corresponding evolution of the lasing spectrum in Fig. \ref{fig:specEvolHeatmapSim}(b). Here we recover the general structure found in Fig. \ref{fig:specEvolHeatmap}, including bifurcations and frequency chirp towards atom resonance. The pattern is found to closely resemble the velocity-dependent absorption/emission behavior previously reported in this regime \cite{qmetLasing} and is a fingerprint of the Doppler shifts of those atoms which interact strongly with the cavity mode. The final spectrum is shown in Fig. \ref{fig:specEvolHeatmapSim}(c), where the lasing spectrum is compared to relevant linewidths in our system: the natural linewidth $\gamma$ of a stationary atom, the Doppler linewidth of the ensemble, and the cavity linewidth shifted by $\Delta_{ce} = 2\pi \times 1$~MHz from the atomic resonance. Despite the large Doppler broadening and the cavity detuning of 1 MHz, the lasing spectrum has a peak component only 300 kHz away from the atomic resonance. This corresponds to the Doppler shift of atoms moving at 0.2 m/s along the cavity axis, which were previously shown to constantly emit during the intensity oscillations in this regime \cite{qmetLasing}. The width of the spectrum is Fourier limited, and thus highly dependent on the duration of the main laser pulse. As the atom number increases the pulse evolves faster, resulting in a shorter main laser pulse. Reducing the temperature of the atoms also leads to faster pulse evolution as it increases the atom number in the lower velocity classes. In addition the lasing peak near $\Delta_{le}$ = $2\pi \times 300$~kHz will become more prominent relative to the structures at higher frequencies. A less efficient pumping leads to a smaller pulse, similarly to the effect of having fewer atoms. Fluctuations in temperature, atom number and pumping efficiency will thus lead to variations in spectra in experiments, even if the cavity could be kept perfectly stable.\\ \begin{figure}[t!] \includegraphics[width=\columnwidth]{Figures/specEvolHeatmapSimDetc1MHzNewcolors.pdf} \caption{(a) Dynamics of the output power from one side of the cavity in a simulation, for a cavity detuning of 1 MHz. (b) Normalized lasing intensity spectrum as a function of window length in the simulation (starting at t = 0). As the intensity oscillations become included in the window, the peaks in the spectrum split up into multiple branches, and the dominant frequency is pulled towards the resonance frequency of the stationary atoms. Minima in output power are marked by green dashed lines. (c) The final spectrum at t = 8 $\upmu$s (black) compared to relevant lineshapes in the system (dashed blue: stationary atom, dash-dotted green: ensemble Doppler profile, red: cavity).\label{fig:specEvolHeatmapSim}} \end{figure} \section{Cavity pulling} To characterize the sensitivity of our experimental system to cavity noise we measure the beat signals for a range of cavity detunings. We vary the cavity detuning in steps of 100 kHz in a randomly generated order to prevent systematic bias, and gather 100 beat signals for each cavity detuning. We find the frequency intensity spectra of individual laser pulses as shown in Fig. \ref{fig:specEvolHeatmap}.\\ \begin{figure}[t!] \includegraphics[width=\columnwidth]{Figures/SpecNormVerticalSimExpIntens.pdf} \caption{Intensity spectra from many lasing pulses for varying cavity detuning: (a) simulations, (b) experiments. Each column in the images constitute the spectrum of a single simulated/measured lasing pulse. The spectra are normalized such that the peak of each column equals one. The dotted gray lines indicate the center of mass of the lasing spectrum, which has a minimal slope of 0.25 at a cavity detuning of $\pm$1.2 MHz in panel a. The dashed green lines refer to the simulation in Fig. \ref{fig:velDatVertical} which illustrates the dynamics in this regime where location of the spectral peak is independent of cavity detuning. \label{fig:cavPull}} \end{figure} \begin{figure}[t!] \includegraphics[width=\columnwidth]{Figures/VelDatVertical300kHz.pdf} \caption{Simulated time evolution of the rate of emission (yellow) and absorption (blue) of cavity photons, depending on the velocity of atoms along the cavity axis. The dynamics are shown for a cavity detuning of 300 kHz. The dotted green line shows where the Doppler shift equals the cavity detuning on the right axis, and the red line shows the spectral peak frequency in comparison, with a window starting at t = 0.\label{fig:velDatVertical}} \end{figure} We can compare the experimentally measured spectra to simulated spectra, in order to understand the dependency on cavity detuning, see Fig. \ref{fig:cavPull}. We use a cavity detuning ($\Delta_{ce}/2\pi$) resolution of 25 kHz in the simulations [Fig. \ref{fig:cavPull}(a)]. In the experiments [Fig. \ref{fig:cavPull}(b)] 100 data samples are taken in steps of 100 kHz in nominal cavity detuning and distributed along the axis. Here the cavity detuning also fluctuates on the order of 100 kHz from its nominal value from shot to shot. We can characterize the immunity of the system to cavity noise by a local cavity pulling coefficient in terms of how much the lasing frequency changes when the cavity detuning is changed: $c_{pull} = d\Delta_{le}/d\Delta_{ce}$. The lasing spectrum can be partially characterized by its peak and center (defined by the center of mass) frequencies, which we will use to characterize cavity pulling. In an ideal active frequency reference the lasing frequency is completely immune to shifts in cavity resonance, which would yield $c_{pull}$ = 0, while in the good cavity regime where $\gamma \gg \kappa$, $c_{pull}$ approaches 1 as the cavity alone determines the lasing frequency. Simulations interestingly show that the peak frequency of the spectrum is immune to shifts in cavity resonance frequency for a range of detunings around $\pm$300 kHz, thus $c_{pull}^{peak}$ = 0. This local immunity to cavity fluctuations occurs because of the velocity-dependent interactions between the atoms and cavity in the pulsed lasing regime, as illustrated in the simulations in Fig. \ref{fig:velDatVertical}. The frequency of the initial lasing pulse starts out with a shift of about 2/3 of the cavity detuning, but during the subsequent cycles of reabsorption and reemission the spectral peak is pulled to 150 kHz off the atom resonance. In this interval the center frequency of the spectrum is still highly sensitive to the cavity detuning; $c_{pull}^{center}$ = 1. The minimal cavity pulling obtained when considering the center of the spectrum is $c_{pull}^{center}$ = 0.25 at $\Delta_{ce}/2\pi$ = $\pm$ 1.2 MHz, and in this range $c_{pull}^{peak}$ = 0.5. The dynamics in this regime are very similar to the behavior shown in Fig. \ref{fig:specEvolHeatmapSim}. Beyond detunings of 1.5 MHz cavity pulling begins to increase significantly as the cavity is so far away from resonance with the slowest atoms that their interaction becomes too weak for these atoms to filter the light. At these values the pulses start to become smaller, with fewer after-pulses than at 1 MHz, and the lasing process instead takes place mainly in a subset of fast-moving atoms. \section{Influence of cavity noise} The non-linear dependency of cavity pulling on its frequency detuning suggests that the frequency stability of an active laser should also be dependent on the chosen cavity detuning during operation. This could allow for supression of classical noise from cavity fluctuations beyond the expected resonant cavity pulling factor. To investigate this quantitatively for the laser pulses shown in Fig. \ref{fig:cavPull} we extract the peak frequencies of the lasing spectra and plot their standard deviation in Fig. \ref{fig:freqStab}. We discard outliers based on a 5$\sigma$ criterium, which filters away data samples where e.g. the lasing pulse did not build up, and also some spectra for negative $\Delta_{ce}$ where the spectral peak occured in the bifurcating structure which is depicted in Fig. \ref{fig:specEvolHeatmapSim}(b) and (c) near $\Delta_{le}/2\pi$ = 1 MHz. Since the laser pulse often fails to build up for cavity detunings outside $\pm$1.5 MHz, the remaining analyzed samples in this regime will tend to be sampled from the lower range of the fluctuations, which are on the order of 100 kHz. To compare this to simulations we first assume that experimental fluctuations in the cavity detuning are the dominant contribution to the standard deviation of the experimental peak frequencies. This is supported by the fact that the frequency fluctuations are smallest near $\pm$300 kHz and largest near resonance or at large detunings, in agreement with the cavity pulling behavior. Therefore we interpolate the simulated peak frequency values as a function of cavity detuning, increasing the resolution by a factor of 50, and run a Monte Carlo-sampling of the interpolated peak frequencies obtained for a Gaussian distribution of cavity detunings with width $\sigma_{jitter}$. We pick 1000 points for each cavity detuning. The standard deviations of the obtained peak frequency distributions are shown in Fig. \ref{fig:freqStab} for $\sigma_{jitter}$ = 30, 100 and 300 kHz. We see that the experimental fluctuations are in reasonable agreement with the simulated ones for $\sigma_{jitter}$ = 100 kHz. For $\sigma_{jitter}$ = 30 kHz the lower values of the peak frequency deviations vary erratically. This is because the random variations in frequency from pulse to pulse start to become significant compared to the effect of cavity pulling.\\ \begin{figure}[t!] \includegraphics[width=\columnwidth]{Figures/sensitivity.pdf} \caption{The sensitivity of the lasing peak frequencies found in experiments and simulations for different cavity detunings. Purple points show experimental frequency stability with errorbars indicating 1$\sigma$ uncertainty. The curves show simulations with added $\sigma_{jitter}$ of 30 kHz (dark grey) 100 kHz (orange) and 300 kHz (light grey). Variations in atom number and pumping efficiency causes additional variations in the peak frequencies found in experiments when compared to simulations. The fluctuations are minimal for a cavity detuning near $\pm$300 kHz. \label{fig:freqStab}} \end{figure} \section{Future improvements} Though the spectral peak frequency can be made highly insensitive to the cavity detuning, the spectral center of mass and other components still vary significantly with cavity detuning. By cooling the atoms to $\upmu$K temperatures (for example by additional cooling on the $^1$S$_0 - ^3$P$_1$ transition), the influence of Doppler shifts is reduced, and in this regime our simulations indicate that $c_{pull}^{peak}$ and $c_{pull}^{center}$ can simultaneously be reduced to 0.15, or separately reach 0 for different parameters. Here the initial inversion of the ensemble (which is determined by the pump pulse duration) also influences the lasing spectrum. This has previously been observed on the $^{87}$Sr clock transition in \cite{NorciaClockFreq}. This allows one to optimize both the cavity detuning and pump pulse duration to operate in a regime where the lasing frequency is approximately independent of the cavity frequency, and thus noise in the cavity.\\ For metrological applications it is desireable to extend the emission time beyond a few microseconds. To reach the quasi-continuous or ultimately true continuous-wave lasing regime, where the gain is replenished while emission is taking place, incoherent repumping schemes could be employed, such as demonstrated in \cite{NorciaCrossover}. In these regimes the nature of the velocity-dependent dynamics and spectral properties would change significantly, as energy is replenished to emitting atoms with different velocities, or new atoms enter a populated cavity field. The velocity-dependent Rabi oscillations in the pulsed regime may on a longer timescale be replaced by synchronization among atoms within a range of velocities, similar to the dynamics simulated in \cite{DebnathFreqClasses}. The spectral properties due to these dynamics could have a very different dependency on the cavity detuning than the pulsed behavior and not necessarily lead to zero-crossings of the cavity pulling coefficient. \section{Conclusion} We have investigated the spectral and phase properties of pulsed lasing on the $^1$S$_0 - ^3$P$_1$ line in a mK Strontium ensemble both experimentally and in simulations, finding general agreement. We have described the processes involved in the random phase built up during the lasing pulses and investigated the phase-flipping behavior. In the resonant-cavity regime, the phase shifts abruptly by $\pi$ between each intensity oscillation, while the phase shifts are more gradual in the detuned-cavity regime. The spectral properties are complementary to the phase behavior, and we find that Doppler shifts play a significant role in the spectral properties. However even for a detuned cavity, slow atoms can pull the lasing frequency closer towards the stationary atomic resonance and suppress the influence of cavity drift on the spectrum. The extent of cavity pulling generally depends on a wide range of experimental parameters that can be optimized, including the cavity detuning and pump pulse duration. We notably find that the the peak spectral component is to first order insensitive to variations in cavity detuning for detunings around $\pm$300 kHz, due to the velocity-dependent absorption and emission dynamics in this regime. The presented simulation method, applied here to the pulsed regime of superradiance, can also be used to gain insight into turn-on and transient behavior of continuous superradiant lasers. \begin{acknowledgments} The authors would like to thank Rodolphe Le Targat and J\'er\^ome Lodewyck at SYRTE (Observatoire de Paris, CNRS, PSL, Sorbonne Universit\'e, LNE) allowing us to borrow a reference cavity and helpful advice in setting it up. They would also like to thank Yuan Zhang for useful theoretical discussions, and Sofus L. Kristensen, Valentin P. Cambier, Eliot A. Bohr and Julian C. R. Tait for useful discussions about the article. This project has received funding from the European Union’s (EU) Horizon 2020 research and innovation programme under grant agreement No 820404 (iqClock project), the USOQS projet (17FUN03) under the EMPIR initiative, and the Q-Clocks project under the European Comission's QuantERA initiative. We acknowledge support from VILLUM FONDEN via research grant 17558. \end{acknowledgments} \newpage \section{Introduction} Lasers with stable, narrow frequency spectra have numerous applications in ultra-high precision measurements including state of the art clocks \cite{JILAlatticeClock, UshijimaLatticeClock, McGrewYbLatticeClock, SYRTEclocks, PTBclocks, INRIMclock, NPLclock} and gravitational wave detection \cite{LIGO, LISA}. These frequency references can be passive, in which case independently generated laser light is compared to an accurate reference. The state of the art performance of such systems is limited today by the thermal motion at the atomic level in the cavity mirrors \cite{Numata2004, Cole2013, MateiPRL2017, Robinson2019}. This limitation can be partially circumvented in an active reference, in which the system could consist of an atomic ensemble within an optical cavity \cite{KazakovEnsembles, KazakovColdAtoms, HollandLaser, ChenCaLaser, JingbiaoChen2005, ruggedLaser}. Here the idea is to pump the atoms to an excited state with a long lifetime relative to the decay of the cavity field, thus operating in the bad cavity regime \cite{ChenOE2019}. In this regime the cavity will enhance the emission rate by the atoms without pulling the frequency strongly to the cavity resonance, as the phase information is primarily stored in the atoms. This makes active frequency references operating in the bad cavity regime a promising technology for generating highly stable and reproducible frequencies, and has sparked interest in the physics of bad-cavity lasing.\\ Superradiance can be observed when an atomic ensemble is coupled to a preferred electromagnetic mode. This process was originally investigated theoretically in the 1950s by R. H. Dicke \cite{Dicke}, and subsequently observed in the optical regime in 1973 \cite{HFsuperradiance}. Coupling an atomic ensemble to a broad cavity resonance, superradiant emission can be realized deep in the bad cavity regime. Superradiant pulses are characterized by a hyperbolic secant shape in the time-domain. The dynamics of superradiant lasing pulses have been studied in the context of narrow optical transitions on the 375 Hz wide $^1$S$_0 - ^3$P$_1$ line in Calcium \cite{CalciumDelayStatistics} and on the mHz wide $^1$S$_0 - ^3$P$_0$ transition in Strontium \cite{NorciaClockPulses}, where the effects of cavity pulling have also been characterized \cite{NorciaClockFreq}. In systems where the atomic transition and cavity linewidths are comparable, a superradiant crossover regime can be realized, characterized by both atomic coherence and a cavity field population, which acts back on the atoms. As a result these pulses deviate from the hyperbolic secant shape. The dynamics in this regime have been studied in experiments using the $^1$S$_0 - ^3$P$_1$ line in Strontium confined in an optical lattice at $\upmu$K temperatures \cite{NorciaCrossover} and in our system at mK temperatures \cite{qmetLasing}. Though lower temperatures imply less inhomogeneous broadening from Doppler shifts, these low temperatures are technically demanding to achieve. Thus a superradiant laser using hotter atoms could show great technological promise if the influence of Doppler shifts can be suppressed in the lasing frequency. Such a scheme has recently been proposed for operation in the continuous regime \cite{ruggedLaser} and reduces the technical challenges compared to continuous superradiance based on ultracold atoms \cite{KazakovColdAtoms}.\\ In this paper we investigate the spectral properties and phase dynamics of an ensemble of unconfined $^{88}$Sr atoms emitting superradiant pulses of light via the mode of an optical cavity. In this system the temperature of the atoms results in a Doppler-broadened lasing transition linewidth that is larger than the cavity linewidth. Despite this, operating the system around a specific cavity detuning enables us to exploit the velocity-dependent behavior of moving atoms in order to suppress cavity pulling in the pulsed lasing regime. We present the experimental setup in Section II and the theoretical model of the system in Section III. This model is applied in the following sections to extract additional information about the physics at the atomic level, and simulations are shown along with the experimental results. We first describe the phase behavior of the lasing process in Section IV, then in Section V we present the spectral properties of the lasing pulses and relevant dynamics. In Section VI we analyze cavity pulling in the system and under which conditions this can be minimized, and in Section VII we evaluate the sensitivity of the system to cavity noise. In Section VIII we discuss improvements which are relevant for metrological applications. \section{Experimental system} The experimental system consists of a 5 mK ensemble of $^{88}$Sr atoms cooled and trapped in a 3D Magneto-Optical Trap (MOT) from a Zeeman-slowed atomic beam, using the $^1$S$_0 \leftrightarrow ^1$P$_1$ transition at 461 nm. The atomic cloud partially overlaps with an optical cavity and can be coherently pumped to $^3$P$_1$ using an off-axis pump laser, see Fig. \ref{fig:setup}. The cavity has a linewidth of $\kappa = 2\pi\times 620$~kHz and can be tuned to resonance with the $^1$S$_0 \leftrightarrow ^3$P$_1$ intercombination line ($\gamma = 2\pi\times 7.5$~kHz). For comparison the Doppler FWHM is $\Gamma_D = 2\pi \times 2.3$~MHz. Due to a cavity waist radius of $w_0 = 450$~$\upmu$m the number of atoms within the cavity, $N_{cav}$, is of order $10^7$, while the total number of atoms in the cloud, N, is of order $10^8$, as estimated from fluorescence measurements and lasing pulse intensity.\\ \begin{figure}[t] \includegraphics[width=\columnwidth]{Figures/setupBeat.pdf} \caption{(a) Schematic of the experimental system showing a thermal ensemble of strontium atoms partially overlapping with the cavity mode. Coherent pumping prepares the atoms in $^3$P$_1$ and allows subsequent emission of a lasing pulse into the cavity mode. Light couples out of both ends of the cavity with the cavity linewidth $\kappa$. At one end the cavity output is overlapped with a reference laser beam, and the beat signal is detected. (b) Level scheme for $^{88}$Sr showing the cooling and lasing transitions. Wavelength ($\lambda$) and natural transition linewidths ($\gamma$) are indicated.\label{fig:setup}} \end{figure} Once the atomic cloud is established, the slowing and cooling beams are turned off, and the atoms are pumped to $^3$P$_1$ with an approximate $\pi$-pulse of 190 ns from an angle of 45$^\circ$ with respect to the cavity axis. Due to the large dimension of the cloud and the finite temperature, the population transferred to $^3$P$_1$ is limited to an estimated 85 \% within the cavity mode. Following a delay of a few microseconds, a lasing pulse is emitted by the atoms into the cavity. To characterize the spectrum of the laser pulses we beat the light leaking from one side of the cavity with a laser beam which is generated independently from the setup and stabilized to a transportable cavity on loan from SYRTE \cite{SYRTEcavity}. The reference laser frequency is detuned by 171 MHz from the atomic resonance, and the detected beat signal is then mixed at 120 MHz. This yields a signal with a frequency near 51 MHz which is band pass-filtered and recorded with a fast oscilloscope. Such a beat signal is shown in Fig. \ref{fig:specEvolHeatmap}a, for the case where the cavity is detuned by 1 MHz from the atomic resonance. This promotes oscillatory energy exchange between the atoms and cavity mode which results in damped oscillations in the emitted power. In Fig. \ref{fig:specEvolHeatmap}(b) a Fourier transformation is used to extract the spectrum for increasing window sizes, with the start fixed at $t = 0$. This method showcases how the spectrum changes throughout the lasing process; we see bifurcations each time additional intensity oscillations enter the analysis window. We also see that the spectral peak is shifted further towards the atomic resonance frequency as the window is expanded.\\ \begin{figure}[t!] \includegraphics[width=\columnwidth]{Figures/specEvolHeatmap41.pdf} \caption{(a) Beat signal of a lasing pulse at a cavity detuning of 1 MHz. (b) Lasing intensity spectrum obtained by squaring the Fourier transformed beat signal for increasing window sizes, starting at $t = 0$. $\Delta_{le} = 0$ corresponds to the atomic transition frequency. Minima in emitted power are marked by green dashed lines.\label{fig:specEvolHeatmap}} \end{figure} \section{Modeling the laser spectrum} To model the laser pulses we numerically integrate a set of equations for the atomic and cavity states, derived in first order mean-field theory from a Tavis-Cummings Hamiltonian. The details of the model and lasing dynamics are described in \cite{qmetLasing}. To characterize the emitted spectrum in the model we extend the Hamiltonian with terms representing an array of imaginary filter cavities, following the method used by K. Debnath \emph{et al} \cite{Debnath}. \begin{eqnarray} H &=& \hbar \omega_c a^\dagger a + \sum_{j=1}^N \hbar \omega_{e} \sigma_{ee}^j + \sum_{k=1}^{N_f} \hbar \omega_{f}^{k} f_k^\dagger f_k \\\nonumber &+& \sum_{j=1}^N \hbar g_c^j \left( \sigma_{ge}^j + \sigma_{eg}^j \right) \left( a + a^\dagger \right)\\\nonumber &+& \sum_{j=1}^N \hbar \frac{\chi_p^j}{2} \left( \sigma_{ge}^j + \sigma_{eg}^j \right) \left( e^{ i \mathbf{k}_p \cdot \mathbf{r}_j - i \omega_p t } + e^{ -i \mathbf{k}_p \cdot \mathbf{r}_j + i \omega_p t } \right)\\\nonumber &+& \sum_{k=1}^{N_f} \hbar g_f \left( a + a^\dagger \right) \left( f_k + f_k^\dagger \right). \end{eqnarray} Here the lowering operators $a$, $f_k$ and $\sigma_{ge}^j$ represent the cavity field, the $k$'th filter cavity, and the $j$'th atom, respectively. The angular frequencies are given by $\omega_{c}$ for the cavity, $\omega_{e}$ for the atom transition, $\omega_{p}$ for the pump pulse laser and $\omega_{f}^k$ for the $k$'th filter cavity. $\mathbf{k_p}$ is the pump pulse wavevector and $\mathbf{r_j}$ is the position vector of the $j$'th atom at a given time. $g_c^j$ is the position-dependent atom-cavity coupling, $\chi^j_p$ is the Rabi frequency of the semiclassical pump pulse laser, and $g_f$ is the coupling between the cavity and filter cavities.\\ In this model we account for the motion and cavity coupling of each atom separately, due to the mK temperature and large spatial extent of the atomic cloud relative to the cavity waist. Since we describe expectation values our model neglects quantum noise in the system, which has little influence on the final spectrum due to the low Purcell decay rate [$\gamma_P = 1/($37 ms$)$] relative to coherent interactions. The filter cavities enable us to simulate the spectrum of the light leaking from the main cavity. Each of these have a very narrow linewidth $\kappa_F^j$ and is very weakly coupled (with coupling constant $g_f$) so that we can neglect any back-action on the lasing cavity. This appropriately models the experimental situation where laser photons do not re-enter the lasing cavity once they have left the output coupler towards the beat signal detector. The time evolution of the system operator expectation values then obey four distinct types of equations: \begin{eqnarray} \label{eq:OBE} \dot{\left\langle \sigma_{ge}^j \right\rangle} &=& -\left( i\Delta_{ep} + \frac{\gamma}{2} \right) \left\langle \sigma_{ge}^j \right\rangle \\\nonumber &&+ i \left( g_c^j \langle a \rangle + \frac { \chi_p^j}{2} e ^ { -i \mathbf{k}_p \cdot \mathbf{r}_j } \right) \left( \left\langle \sigma_{ee}^j \right\rangle - \left\langle \sigma_{gg}^j \right\rangle \right)\\\nonumber \dot{\left\langle \sigma_{ee}^j \right\rangle} &=& - \gamma \left\langle \sigma_{ee}^j \right\rangle + i \left( g_c^j \left\langle a^\dagger \right\rangle + \frac{\chi_p^j}{2} e^{ i \mathbf{k}_p \cdot \mathbf{r}_j} \right) \left\langle \sigma_{ge}^j \right\rangle\\\nonumber &&- i \left( g_c^j \langle a \rangle + \frac{\chi_p^j}{2} e^{ -i \mathbf {k}_p \cdot \mathbf{r}_j } \right) \left\langle \sigma_{eg}^j \right\rangle.\\\nonumber \dot {\langle a \rangle} &=& - \left( i \Delta_{cp} + \frac { \kappa } { 2 } \right) \langle a \rangle - \sum_{j=1}^N i g_c^j \left\langle \sigma_{ge}^j \right\rangle\\\nonumber \dot {\langle f_k \rangle} &=& - \left( i \Delta^k_{fp} + \frac { \kappa^k_f } { 2 } \right) \langle f_k \rangle - i g_f \left\langle a \right\rangle. \end{eqnarray} Here the detunings $\Delta_{ij} = \omega_i - \omega_j$ have been introduced and a rotating frame has been chosen at $\omega_p$. The intensity spectrum from system initiation up to a time $t$ is then given by the relative populations of the filter cavities at time $t$ as a function of their resonance frequencies, $\abs{\expval{f_k(\omega_f^k)}}^2$, with a normalization depending on $g_f$. The number of filter cavities, $N_f$ and the range of frequencies they span determines the spectral resolution of the simulation.\\ \section{Phase behavior} The phase coherence of the cavity field and its evolution is directly related to the spectral properties of the laser pulses and informs us about the underlying physical processes. The mean phase evolution can be traced in simulations from $\arg(\expval{a}$) in the semiclassical approximation. In Fig. \ref{fig:detc0phaseEvol} we show the simulated cavity output power for three random instances of lasing pulses when the cavity frequency is tuned to atom-resonance [Fig. \ref{fig:detc0phaseEvol}(a)] and the corresponding phase evolution of the cavity field [Fig. \ref{fig:detc0phaseEvol}(b)] in the rotating frame of the stationary atomic transition. The pump pulse initially imprints different phases on the Bloch vectors of all the atoms, depending on their positions and velocities. This prepares the ensemble with a non-zero macroscopic coherence, however it is tiny and randomized due to the 45$^{\circ}$ angle between pump pulse and cavity, and the random initial positions of the atoms at the wavelength level. Furthermore, in the semiclassical model a mean cavity field amplitude builds up during the pumping pulse corresponding to about 0.1 photons. Both the nonzero coherence and cavity field amplitude can separately lead to buildup of a lasing pulse. The atomic coherence can be removed after pumping experimentally by cycling the atoms on the $^1S_0$ - $^1P_1$ transition. If this is simulated, the remaining cavity field will imprint its random phase onto the atomic coherence as the power starts to build up. In the simulations we can also explicitly remove the field built up during pumping. In this case the randomized atomic coherence will trigger the dynamics. The resulting cavity field phase will depend on the macroscopically favored configuration originating from each atoms' state and coupling to the cavity mode. Keeping both the field built up during pumping and the initial random atomic coherence, the simulation shows both playing a role in the buildup of the laser pulse: after pumping, the phase is not stable, but the atoms build up coherence via the cavity field, which settles on a phase within 1 $\upmu$s after the pump pulse, as the lasing pulse starts to build up. At this point the output power has reached on the order of 0.1 nW, and the random delay between pumping and the lasing pulse is closely related to the time it takes for the phase to settle. With the choice of reference frame in Fig. \ref{fig:detc0phaseEvol}(b) a change in the phase of the cavity field over time corresponds to the cavity field being off resonance with the slowest atoms. This means the gain in the system is lower, slowing the pulse build-up and causing a longer delay of the pulse. A third process that can contribute to building up the lasing pulse is spontaneous emission into the cavity mode, which is not included in the model. With approximately $10^7$ atoms within the cavity waist, on the order of 60 photons are emitted into the cavity within 1 $\upmu$s given the Purcell rate of the system. This is the dominant process in triggering the lasing pulse in experiments, and is a source of phase noise that plays a role in the early stages of phase-synchronization.\\ \begin{figure}[t!] \includegraphics[width=\columnwidth]{Figures/detc0phaseEvol.pdf} \caption{Time evolution of the cavity output power (a) and cavity field phase (b) for three simulated laser pulses after a pump pulse (on during the red shaded interval). The three instances have the same ensemble parameters but show random variation. (a) The laser pulses have random delays and are followed by small intensity oscillations when the cavity is on atom-resonance. The blue intervals mark every even number oscillation of the pulse with shortest delay. (b) The phase is random after the pump pulse, then evolves and stabilizes as the laser power builds up. For every intensity oscillation the phase flips sign. (c) Calculating $\sin{(\phi)} \cdot \sqrt{P_{out}}$ illustrates how a demodulated beat signal looks when measured in experiment. (d) Moving mean of three experimentally measured demodulated beat signals (shaded: raw data). We see qualitative agreement with the simulated behavior. \label{fig:detc0phaseEvol}} \end{figure} Though the ensemble preparation, scattering and spontaneous emission can initiate lasing and have different influences on the phase, the subsequent dynamics and spectral properties are highly deterministic. The Bloch vectors of the atoms in the ensemble all oscillate with different Rabi frequencies, depending on the atoms' velocities and positions with respect to the cavity waist and standing wave. Initially most atoms emit light, but at peak output power, most atoms start to reabsorb. By the end of the first intensity oscillation [e.g. at $t = 3$ us for the blue traces in Fig. \ref{fig:detc0phaseEvol}(a) to (c)], the atoms are now out of phase due to their varying velocities and positions. The atoms that were absorbing at the end of the first lasing pulse begin to emit, and those that were emitting start to absorb, causing a $\pi$ phase shift of the second power oscillation. When the cavity is on atom-resonance, this pattern repeats during the subsequent intensity oscillations, as seen in Fig. \ref{fig:detc0phaseEvol}(b). We compare the simulations to measurements of the beat signal with an in-cavity reference laser on resonance with the cavity, but detuned 1 FSR from the atomic resonance, in Fig. \ref{fig:detc0phaseEvol}(c) (simulations) and \ref{fig:detc0phaseEvol}(d) (experiments). The simulation and experimental samples shown in this section are picked randomly from larger samples. We find that the measured beat signals also show the phase flipping behavior during the intensity oscillations, as the beat signals change sign between each oscillation.\\ In the off-resonant case, where the cavity is detuned from the atomic resonance and pronounced after-pulsing is observed, the phase evolution is different (see Fig. \ref{fig:detc1MphaseEvol}). To compare the phase evolution with our experiments, the phase shown here has been transformed to a rotating frame at $\omega_c$. We find that the phase shift is close to $\pi$ between the first and second intensity oscillation, where the output power goes to zero. This explains the splitting of the spectrum in the experimental measurement of Fig. \ref{fig:specEvolHeatmap}(b) during the second oscillation, and the dip forming at the peak frequency of the initial pulse (these dynamics occur between the two leftmost green dashed lines). In subsequent oscillations the power does not go to zero and the phase shifts occur more gradually. The chirp in frequency towards the atomic resonance is also evident in the rate of change of the phase, which increases as the spectrum is pulled closer to atomic resonance. \begin{figure}[t!] \includegraphics[width=\columnwidth]{Figures/detc1MphaseEvolFrameshift.pdf} \caption{Time evolution of the cavity output power (a) and cavity field phase (b) for two simulated laser pulses after a pump pulse (on during the red shaded interval) at a cavity detuning of 1 MHz. (a) Oscillations in the output power are pronounced when the cavity is detuned. The blue areas mark every even number oscillation of the pulse with shortest delay. (b) The phase evolution during the laser pulses in a rotating reference frame at $\omega_c$. The sloped behavior is related to pulling of the lasing frequency away from the cavity resonance and towards the atomic transition frequency. (c) Calculated and (d) moving mean of experimental demodulated beat signals (shaded: raw data). Purple intervals mark every second oscillation in the experimentally measured output power. \label{fig:detc1MphaseEvol}} \end{figure} \section{Lasing spectra and underlying dynamics} To compare simulations to our beat experiments in Fig. \ref{fig:specEvolHeatmap}, the simulated lasing dynamics of a pulse at a cavity detuning of 1 MHz is shown in Fig. \ref{fig:specEvolHeatmapSim}(a), and the corresponding evolution of the lasing spectrum in Fig. \ref{fig:specEvolHeatmapSim}(b). Here we recover the general structure found in Fig. \ref{fig:specEvolHeatmap}, including bifurcations and frequency chirp towards atom resonance. The pattern is found to closely resemble the velocity-dependent absorption/emission behavior previously reported in this regime \cite{qmetLasing} and is a fingerprint of the Doppler shifts of those atoms which interact strongly with the cavity mode. The final spectrum is shown in Fig. \ref{fig:specEvolHeatmapSim}(c), where the lasing spectrum is compared to relevant linewidths in our system: the natural linewidth $\gamma$ of a stationary atom, the Doppler linewidth of the ensemble, and the cavity linewidth shifted by $\Delta_{ce} = 2\pi \times 1$~MHz from the atomic resonance. Despite the large Doppler broadening and the cavity detuning of 1 MHz, the lasing spectrum has a peak component only 300 kHz away from the atomic resonance. This corresponds to the Doppler shift of atoms moving at 0.2 m/s along the cavity axis, which were previously shown to constantly emit during the intensity oscillations in this regime \cite{qmetLasing}. The width of the spectrum is Fourier limited, and thus highly dependent on the duration of the main laser pulse. As the atom number increases the pulse evolves faster, resulting in a shorter main laser pulse. Reducing the temperature of the atoms also leads to faster pulse evolution as it increases the atom number in the lower velocity classes. In addition the lasing peak near $\Delta_{le}$ = $2\pi \times 300$~kHz will become more prominent relative to the structures at higher frequencies. A less efficient pumping leads to a smaller pulse, similarly to the effect of having fewer atoms. Fluctuations in temperature, atom number and pumping efficiency will thus lead to variations in spectra in experiments, even if the cavity could be kept perfectly stable.\\ \begin{figure}[t!] \includegraphics[width=\columnwidth]{Figures/specEvolHeatmapSimDetc1MHzNewcolors.pdf} \caption{(a) Dynamics of the output power from one side of the cavity in a simulation, for a cavity detuning of 1 MHz. (b) Normalized lasing intensity spectrum as a function of window length in the simulation (starting at t = 0). As the intensity oscillations become included in the window, the peaks in the spectrum split up into multiple branches, and the dominant frequency is pulled towards the resonance frequency of the stationary atoms. Minima in output power are marked by green dashed lines. (c) The final spectrum at t = 8 $\upmu$s (black) compared to relevant lineshapes in the system (dashed blue: stationary atom, dash-dotted green: ensemble Doppler profile, red: cavity).\label{fig:specEvolHeatmapSim}} \end{figure} \section{Cavity pulling} To characterize the sensitivity of our experimental system to cavity noise we measure the beat signals for a range of cavity detunings. We vary the cavity detuning in steps of 100 kHz in a randomly generated order to prevent systematic bias, and gather 100 beat signals for each cavity detuning. We find the frequency intensity spectra of individual laser pulses as shown in Fig. \ref{fig:specEvolHeatmap}.\\ \begin{figure}[t!] \includegraphics[width=\columnwidth]{Figures/SpecNormVerticalSimExpIntens.pdf} \caption{Intensity spectra from many lasing pulses for varying cavity detuning: (a) simulations, (b) experiments. Each column in the images constitute the spectrum of a single simulated/measured lasing pulse. The spectra are normalized such that the peak of each column equals one. The dotted gray lines indicate the center of mass of the lasing spectrum, which has a minimal slope of 0.25 at a cavity detuning of $\pm$1.2 MHz in panel a. The dashed green lines refer to the simulation in Fig. \ref{fig:velDatVertical} which illustrates the dynamics in this regime where location of the spectral peak is independent of cavity detuning. \label{fig:cavPull}} \end{figure} \begin{figure}[t!] \includegraphics[width=\columnwidth]{Figures/VelDatVertical300kHz.pdf} \caption{Simulated time evolution of the rate of emission (yellow) and absorption (blue) of cavity photons, depending on the velocity of atoms along the cavity axis. The dynamics are shown for a cavity detuning of 300 kHz. The dotted green line shows where the Doppler shift equals the cavity detuning on the right axis, and the red line shows the spectral peak frequency in comparison, with a window starting at t = 0.\label{fig:velDatVertical}} \end{figure} We can compare the experimentally measured spectra to simulated spectra, in order to understand the dependency on cavity detuning, see Fig. \ref{fig:cavPull}. We use a cavity detuning ($\Delta_{ce}/2\pi$) resolution of 25 kHz in the simulations [Fig. \ref{fig:cavPull}(a)]. In the experiments [Fig. \ref{fig:cavPull}(b)] 100 data samples are taken in steps of 100 kHz in nominal cavity detuning and distributed along the axis. Here the cavity detuning also fluctuates on the order of 100 kHz from its nominal value from shot to shot. We can characterize the immunity of the system to cavity noise by a local cavity pulling coefficient in terms of how much the lasing frequency changes when the cavity detuning is changed: $c_{pull} = d\Delta_{le}/d\Delta_{ce}$. The lasing spectrum can be partially characterized by its peak and center (defined by the center of mass) frequencies, which we will use to characterize cavity pulling. In an ideal active frequency reference the lasing frequency is completely immune to shifts in cavity resonance, which would yield $c_{pull}$ = 0, while in the good cavity regime where $\gamma \gg \kappa$, $c_{pull}$ approaches 1 as the cavity alone determines the lasing frequency. Simulations interestingly show that the peak frequency of the spectrum is immune to shifts in cavity resonance frequency for a range of detunings around $\pm$300 kHz, thus $c_{pull}^{peak}$ = 0. This local immunity to cavity fluctuations occurs because of the velocity-dependent interactions between the atoms and cavity in the pulsed lasing regime, as illustrated in the simulations in Fig. \ref{fig:velDatVertical}. The frequency of the initial lasing pulse starts out with a shift of about 2/3 of the cavity detuning, but during the subsequent cycles of reabsorption and reemission the spectral peak is pulled to 150 kHz off the atom resonance. In this interval the center frequency of the spectrum is still highly sensitive to the cavity detuning; $c_{pull}^{center}$ = 1. The minimal cavity pulling obtained when considering the center of the spectrum is $c_{pull}^{center}$ = 0.25 at $\Delta_{ce}/2\pi$ = $\pm$ 1.2 MHz, and in this range $c_{pull}^{peak}$ = 0.5. The dynamics in this regime are very similar to the behavior shown in Fig. \ref{fig:specEvolHeatmapSim}. Beyond detunings of 1.5 MHz cavity pulling begins to increase significantly as the cavity is so far away from resonance with the slowest atoms that their interaction becomes too weak for these atoms to filter the light. At these values the pulses start to become smaller, with fewer after-pulses than at 1 MHz, and the lasing process instead takes place mainly in a subset of fast-moving atoms. \section{Influence of cavity noise} The non-linear dependency of cavity pulling on its frequency detuning suggests that the frequency stability of an active laser should also be dependent on the chosen cavity detuning during operation. This could allow for supression of classical noise from cavity fluctuations beyond the expected resonant cavity pulling factor. To investigate this quantitatively for the laser pulses shown in Fig. \ref{fig:cavPull} we extract the peak frequencies of the lasing spectra and plot their standard deviation in Fig. \ref{fig:freqStab}. We discard outliers based on a 5$\sigma$ criterium, which filters away data samples where e.g. the lasing pulse did not build up, and also some spectra for negative $\Delta_{ce}$ where the spectral peak occured in the bifurcating structure which is depicted in Fig. \ref{fig:specEvolHeatmapSim}(b) and (c) near $\Delta_{le}/2\pi$ = 1 MHz. Since the laser pulse often fails to build up for cavity detunings outside $\pm$1.5 MHz, the remaining analyzed samples in this regime will tend to be sampled from the lower range of the fluctuations, which are on the order of 100 kHz. To compare this to simulations we first assume that experimental fluctuations in the cavity detuning are the dominant contribution to the standard deviation of the experimental peak frequencies. This is supported by the fact that the frequency fluctuations are smallest near $\pm$300 kHz and largest near resonance or at large detunings, in agreement with the cavity pulling behavior. Therefore we interpolate the simulated peak frequency values as a function of cavity detuning, increasing the resolution by a factor of 50, and run a Monte Carlo-sampling of the interpolated peak frequencies obtained for a Gaussian distribution of cavity detunings with width $\sigma_{jitter}$. We pick 1000 points for each cavity detuning. The standard deviations of the obtained peak frequency distributions are shown in Fig. \ref{fig:freqStab} for $\sigma_{jitter}$ = 30, 100 and 300 kHz. We see that the experimental fluctuations are in reasonable agreement with the simulated ones for $\sigma_{jitter}$ = 100 kHz. For $\sigma_{jitter}$ = 30 kHz the lower values of the peak frequency deviations vary erratically. This is because the random variations in frequency from pulse to pulse start to become significant compared to the effect of cavity pulling.\\ \begin{figure}[t!] \includegraphics[width=\columnwidth]{Figures/sensitivity.pdf} \caption{The sensitivity of the lasing peak frequencies found in experiments and simulations for different cavity detunings. Purple points show experimental frequency stability with errorbars indicating 1$\sigma$ uncertainty. The curves show simulations with added $\sigma_{jitter}$ of 30 kHz (dark grey) 100 kHz (orange) and 300 kHz (light grey). Variations in atom number and pumping efficiency causes additional variations in the peak frequencies found in experiments when compared to simulations. The fluctuations are minimal for a cavity detuning near $\pm$300 kHz. \label{fig:freqStab}} \end{figure} \section{Future improvements} Though the spectral peak frequency can be made highly insensitive to the cavity detuning, the spectral center of mass and other components still vary significantly with cavity detuning. By cooling the atoms to $\upmu$K temperatures (for example by additional cooling on the $^1$S$_0 - ^3$P$_1$ transition), the influence of Doppler shifts is reduced, and in this regime our simulations indicate that $c_{pull}^{peak}$ and $c_{pull}^{center}$ can simultaneously be reduced to 0.15, or separately reach 0 for different parameters. Here the initial inversion of the ensemble (which is determined by the pump pulse duration) also influences the lasing spectrum. This has previously been observed on the $^{87}$Sr clock transition in \cite{NorciaClockFreq}. This allows one to optimize both the cavity detuning and pump pulse duration to operate in a regime where the lasing frequency is approximately independent of the cavity frequency, and thus noise in the cavity.\\ For metrological applications it is desireable to extend the emission time beyond a few microseconds. To reach the quasi-continuous or ultimately true continuous-wave lasing regime, where the gain is replenished while emission is taking place, incoherent repumping schemes could be employed, such as demonstrated in \cite{NorciaCrossover}. In these regimes the nature of the velocity-dependent dynamics and spectral properties would change significantly, as energy is replenished to emitting atoms with different velocities, or new atoms enter a populated cavity field. The velocity-dependent Rabi oscillations in the pulsed regime may on a longer timescale be replaced by synchronization among atoms within a range of velocities, similar to the dynamics simulated in \cite{DebnathFreqClasses}. The spectral properties due to these dynamics could have a very different dependency on the cavity detuning than the pulsed behavior and not necessarily lead to zero-crossings of the cavity pulling coefficient. \section{Conclusion} We have investigated the spectral and phase properties of pulsed lasing on the $^1$S$_0 - ^3$P$_1$ line in a mK Strontium ensemble both experimentally and in simulations, finding general agreement. We have described the processes involved in the random phase built up during the lasing pulses and investigated the phase-flipping behavior. In the resonant-cavity regime, the phase shifts abruptly by $\pi$ between each intensity oscillation, while the phase shifts are more gradual in the detuned-cavity regime. The spectral properties are complementary to the phase behavior, and we find that Doppler shifts play a significant role in the spectral properties. However even for a detuned cavity, slow atoms can pull the lasing frequency closer towards the stationary atomic resonance and suppress the influence of cavity drift on the spectrum. The extent of cavity pulling generally depends on a wide range of experimental parameters that can be optimized, including the cavity detuning and pump pulse duration. We notably find that the the peak spectral component is to first order insensitive to variations in cavity detuning for detunings around $\pm$300 kHz, due to the velocity-dependent absorption and emission dynamics in this regime. The presented simulation method, applied here to the pulsed regime of superradiance, can also be used to gain insight into turn-on and transient behavior of continuous superradiant lasers. \begin{acknowledgments} The authors would like to thank Rodolphe Le Targat and J\'er\^ome Lodewyck at SYRTE (Observatoire de Paris, CNRS, PSL, Sorbonne Universit\'e, LNE) allowing us to borrow a reference cavity and helpful advice in setting it up. They would also like to thank Yuan Zhang for useful theoretical discussions, and Sofus L. Kristensen, Valentin P. Cambier, Eliot A. Bohr and Julian C. R. Tait for useful discussions about the article. This project has received funding from the European Union’s (EU) Horizon 2020 research and innovation programme under grant agreement No 820404 (iqClock project), the USOQS projet (17FUN03) under the EMPIR initiative, and the Q-Clocks project under the European Comission's QuantERA initiative. We acknowledge support from VILLUM FONDEN via research grant 17558. \end{acknowledgments} \newpage

c13f3e03abfc8007eb6b15fd289b42308ede8855

\section{A Generalized Lipschitz Condition}\label{section6} In this section we explore a different class of morphisms between QLR, generalizing the usual Lipschitz condition. Notably, we show that in this setting the QLR satisfying reflexivity and transitivity can be lifted to all simple types. \subsection{From Lipschitz to Locally Lipschitz functions} As observed in previous sections, the Lipschitz condition has been widely investigated in program semantics, but is considered problematic when dealing with fully higher-order languages. Does the picture change when we step from models like $\mathsf{Met}_{Q}$ to categories of QLR? \begin{remark} For simplicity, from now on we will suppose that QLR are always reflexive and symmetric. \end{remark} To answer this question we must first find a suitable extension of the Lipschitz condition to this setting. The first step is to introduce a notion of \emph{finiteness}: since a quantale is a complete lattice, we must avoid that \emph{any} function $f:X\to Y$ between QLR admits the trivial Lipschitz constant $\top$. \begin{definition} Let $Q$ be a commutative and integral quantale. A \emph{finiteness filter of $Q$} is a downward set $\ff Q\subseteq Q$ such that $a,b\in \ff Q$ implies $a+b\in \ff Q$. A \emph{finitary QRL} is a tuple $(X,Q,\ff Q,a)$ such that $(X,Q,a)$ is a QLR, $\ff Q$ is a finitary filter of $Q$ and $Im(a)\subseteq \ff Q$. \end{definition} The positive reals $\BB R_{\geq 0}$ form a finiteness filter of $\BB R^{\infty}_{\geq 0}$. Moreover, if $\ff Q$ and $\ff R$ are finiteness filters of $Q$ and $R$, then $\ff Q\times \ff R$ is a finiteness filter of $Q\times R$, and for all set $X$, $(\ff Q)^{X}$ is a finiteness filter of $Q^{X}$. Now, a basic observation is that if a function $f:X\to Y$ between metric spaces is $L$-Lipschitz, then there is a monoid homomorphism $\varphi: \BB R^{+}_{\geq 0}\to \BB R^{+}_{\geq 0}$ given by $\varphi(x)=L\cdot x$, such that $d(f(x),f(y))\leq \varphi(d(x,y))$. This suggests the following: \begin{definition}[generalized Lipschitz maps] Let $(X,Q,\ff Q, a)$, $(Y,R,\ff R,b)$ be finitary QLR. A function $f:X\to Y$ is a \emph{generalized Lipschitz map} from $X$ to $Y$ if there exists a monoid homomorphism $\varphi:Q\to R$ satisfying: \begin{align*} \forall \alpha\in \ff Q \ & \varphi(\alpha)\in \ff R \tag{finiteness} \\ a(x,y)\leq \alpha \ & \Rightarrow \ b(f(x),f(y))\leq \varphi(\alpha) \tag{Lipschitz} \end{align*} \end{definition} Observe that any Lipschitz function $f: \BB R\to \BB R$ in the usual sense is a generalized Lipschitz map between the finitary and reflexive QLR given by the Euclidean metric. Finitary QLR and generalized Lipschitz maps form a category $\mathbf L$ with cartesian structure defined as in $\mathbf{Q}$. Moreover, given finitary QLR $(X,Q,\ff Q,a)$ and $(Y,R,\ff R,b)$ there is a finitary QLR $(\mathbf L(X,Y), R^{X},(\ff R)^{X}, b^{X})$ where $b^{X}(f,g)(x)=b(f(x),g(x))$ (note that also symmetry and reflexivity are preserved). Yet, with this definition $\mathbf L$ is still \emph{not} cartesian closed. For instance, consider the function $f(x)(y): \BB R\to \BB R^{\BB R}$ given by $f(x)(y)=x \cdot y$. As a function of two variables, $f$ is Lipschitz in both $x$ and $y$, with Lipschitz constants $|y|$ and $|x|$; one can use this fact to show that $f\in \mathbf{L} (\BB R, \BB R^{\BB R})$. Now, if $\mathbf{L}$ were cartesian closed, using the canonical isomorphism $\mathbf{L}(\BB R, \BB R^{\BB R})\simeq \mathbf{L}(\BB R\times \BB R, \BB R)$, we could deduce that also the function $\mathsf{ev}(f)(\langle x,y\rangle)=f(x)(y)$ is Lipschitz. However, there is no way to deduce, from the two piecewise Lipschitz constants $|x|$ and $|y|$ for $x$ and $y$, a uniform Lipschitz constant for \emph{both} variables. In fact, all we can say is that, for any choice of points $x,y\in \BB R$, we can deduce a Lipschitz constant $L_{x,y}=|x|\cdot |y|$ for $\mathsf{ev}(f)$, although there is no way to define one in a uniform way. This observation suggests to replace the Lipschitz condition with the \emph{local} Lipschitz condition. Recall that a function $f:X\to Y$ between two metric spaces is \emph{locally Lipschitz continuous} when for all $x\in X$ there exists a constant $L_{x}$ such that the inequality $d(f(y),f(z))\leq L_{x}\cdot d(y,z)$ holds in some open neighborhood of $x$. \begin{remark} From now on we will suppose that quantales $(Q,\geq)$ are \emph{continuous} as lattices, and we indicate by $\alpha \ll \beta$ the usual \emph{way below} relation. It is clear that the Lawvere quantale and all quantales obtained from it by applying products are continuous lattices. \end{remark} \begin{definition} Let $(X,Q,\ff Q,a)$ and $(Y,R,\ff R,b)$ be finitary QLR. A function $f: X\to Y$ is said \emph{generalized locally Lipschitz} (in short LL), if there exists a function $\varphi:X\times Q\to R$ (called a \emph{family of LL-constants for $f$}) such that $\varphi(x,\_)$ is additive in its second variable, and the following hold for all $x\in X$: \begin{align*} &\forall \alpha \in \ff Q \ \varphi(x,\alpha)\in \ff R & \text{\emph{(finiteness)}} \\ & \exists \delta_{x} \gg 0\forall y,z\in X \ a(x,y),a(x,z)\leq \delta_{x}\Rightarrow \\ & \ \ a(y,z)\leq \alpha \ \Rightarrow \ b(f(y),f(z))\leq \varphi(x,\alpha) & \text{\emph{(local\ Lipschitz)}} \end{align*} \end{definition} Any locally Lipschitz function $f: \BB R\to \BB R$ yields a LL-map between the finitary QLR given by the Euclidean metric. The finitary QLR with LL maps form a category $\mathbf{LL}$: the identity function $\mathrm{id}_{X}$ has the LL constants $\lambda x\alpha.\alpha$. Moreover, the composition of LL functions $f:X\to Y$ and $g: Y\to Z$ is LL: if $\varphi$ is a family of LL constants for $f$ and $\psi$ is a family of LL constants for $g$, then the map $(x,\alpha)\mapsto \psi(f(x), \varphi(x,\alpha))$ is a family of LL constants for $g\circ f$ (observe that identity and composition of LL constants work precisely as in $\mathbf{Q}$). One can also consider a slightly different category $\mathbf{LL}^{*}$ defined as follows. First, for a QLR $(X,Q,a)$, let $\simeq_{a}$ be the equivalence relation over $X$ defined by $x\simeq_{a}x'$ if $a(x,x')=0$. We indicate by $X/a$ the quotient of $X$ by $\simeq_{a}$. By definition, the QLR $(X/a,Q,a)$ is separated. The objects of $\mathbf{LL}^{*}$ are the same as those of $\mathbf{LL}$, while the arrows between $(X,Q,\ff Q,a)$ and $(Y,R,\ff R,b)$ are pairs $(f,\varphi)$, where $f:X\to Y$ is LL and stable under $\simeq_{a}$-classes (i.e.~$a(x,y)=0$ implies $b(f(x),f(y))=0$), and $\varphi$ is a family of LL-constants for $f$ and is also stable under $\simeq_{a}$-classes (i.e.~$a(x,y)=0$ implies $\varphi(x,\alpha)=\varphi(y,\alpha)$). There is a forgetful functor $U: \mathbf{LL}^{*} \to \mathbf{LL}$ given by $U{(X,Q,\ff Q,a)}= (X/a, Q,\ff Q, Ua)$, where $Ua([x],[y])=a(x,y)$, and $U{(f,\varphi)}=\widetilde f$, where $\widetilde f([x]_{a})=[f(x)]_{b}$. Given finitary QLR $(X,Q,\ff Q,a)$ and $(Y,R,\ff R,b)$ we can define the two finitary QLR \longv{\\ } $( \mathbf{LL}(X,Y), R^{X}, (\ff R)^{X},b^{X})$ and $(\mathbf{LL}^{*}(X,Y), R^{X},(\ff R)^{X}, b^{X}\circ \pi_{1})$. Observe that if $X$ and $Y$ satisfy transitivity, so do $\mathbf{LL}(X,Y)$ and $\mathbf{LL}^{*}(X,Y)$, and if $Y$ is a standard metric space, $\mathbf{LL}(X,Y)$ is a standard metric space, while $\mathbf{LL}^{*}(X,Y)$ is a pseudo-metric space. Moreover, if the QLR $X,Y,Z$ satisfy transitivity, we can define an isomorphism \\ $\begin{tikzcd}\mathbf{LL}^{*}(Z\times X, Y)\ar[bend left=5]{r}[above]{\lambda} & \mathbf{LL}^{*}(Z, \mathbf{LL}^{*}(X,Y)) \ar[bend left=5]{l}[below]{\mathsf{ev}} \end{tikzcd}$ as follows: \begin{itemize} \item the map $\lambda(f, \varphi)=( \langle \lambda(f), \lambda_{0}(\varphi)\rangle, \lambda_{1}(\varphi))$ is defined by \begin{align*} \lambda(f)(z)(x) & =f(\langle z,x\rangle) \\ \lambda_{0}(\varphi)(z)(\langle x,\alpha\rangle) & =\varphi(\langle z,x\rangle, \langle 0,\alpha\rangle)\\ \lambda_{1}(\varphi)(\langle z,\zeta\rangle)(x)& = \varphi(\langle z,x\rangle, \langle \zeta, 0\rangle) \end{align*} \item the map $\mathsf{ev}(\langle g,\psi\rangle, \chi)=\langle \mathsf{ev}(g), \mathsf{ev}(\psi, \chi)\rangle $ is defined by \begin{align*} \mathsf{ev}(f)(\langle z,x\rangle) &= f(z)(x) \\ \mathsf{ev}(\psi, \chi) (\langle z,x\rangle, \langle \zeta, \alpha\rangle) &= \chi(\langle z,\zeta\rangle)(x)+ \psi(z)(\langle x,\alpha\rangle) \end{align*} \end{itemize} Reflexivity and transitivity are essential for the isomorphism above to hold: for all $(f,\varphi)\in \mathbf{LL}^{*}(Z\times X, Y)$, to show that \begin{align*} b(f( z,x), f( z,x))\leq \lambda_{1}(\varphi)(\langle z, 0\rangle)(x)=0 \end{align*} one makes essential use of the fact that $b(f( z,x), f( z,x))=0$ holds in $Y$. Conversely, given $(\langle g,\psi\rangle, \chi)\in \mathbf{LL}^{*}(Z,\mathbf{LL}^{*}(X,Y))$, to show that \begin{align*} b(f(z,x), f( z',x'))&\leq \mathsf{ev}(\psi,\chi)(\langle x,z\rangle, \langle c(z,z'),a(x,x')\rangle) \\ &= \chi(\langle z, c(z,z')\rangle)(x)+\psi (z)(\langle x,a(x,x')\rangle) \end{align*} one makes essential use of the transitivity of $Y$ to deduce the above from $b(f(z,x), f(z',x))\leq \chi(z, c(z,z'))(x)$ and $b(f(z',x), f(z',x'))\leq \psi(z)(x,a(x,x'))$. All this leads to the following result: \begin{proposition}\label{prop:uu} The full sub-category $\mathbf{LL}_{\mathsf{Met}}\hookrightarrow\mathbf{LL}$ of standard metric spaces is cartesian closed. The full sub-category $\mathbf{LL}_{\mathsf{pMet}}\hookrightarrow \mathbf{LL}^{*}$ of pseudo-metric spaces is cartesian closed. Moreover, the restriction of $U$ as a functor from $\mathbf{LL}_{\mathsf{Met}}$ to $\mathbf{LL}_{\mathsf{pMet}}$ is a cartesian closed functor. \end{proposition} \longv{\begin{proof} We first check the cartesian closure of $\mathbf{LL}_{\mathsf{pMet}}$. \begin{itemize} \item[($\Rightarrow$)] the map $\lambda(f, \varphi)=( \langle \lambda(f), \lambda_{0}(\varphi)\rangle, \lambda_{1}(\varphi))$ is defined by \begin{align*} \lambda(f)(z)(x) & =f(\langle z,x\rangle) \\ \lambda_{0}(\varphi)(z)(\langle x,\alpha\rangle) & =\varphi(\langle z,x\rangle, \langle 0,\alpha\rangle)\\ \lambda_{1}(\varphi)(\langle z,\zeta\rangle)(x)& = \varphi(\langle z,x\rangle, \langle \zeta, 0\rangle) \end{align*} For all $z\in Z$, then map $\lambda_{0}(\varphi)(z)(\_,\_)$ is additive in its second variable; moreover, for all $z\in Z$ and $x\in X$ there is $\langle \zeta_{z},\alpha_{x}\rangle \gg 0$ (which implies $\zeta_{z}\gg 0$ and $\alpha_{x}\gg 0$) such that, whenever $c(z,z'),c(z,z'')\leq \zeta_{z}$ and $a(x,x'),a(x,x'')\leq \alpha_{x}$, $\lambda_{0}(\varphi)(z)(\langle x,a(x',x'')\rangle)\geq b(\lambda (f)(z)(x'), \lambda (f)(z)(x''))= b(f(\langle z,x'\rangle), f(\langle z,x''\rangle))$. This proves that $\langle\lambda (f), \lambda_{0}(\varphi)\rangle(z)\in \mathbf{LL}_{\mathsf{Met}}(X,Y)$. Finally, any $z$ is contained in an open ball such that, whenever $z',z''$ belong to it, \\ $\lambda_{1}(\varphi)(\langle z, c(z',z'')\rangle)(x)\geq b(\lambda (f)(z')(x), \lambda (f)(z'')(x))= b(f(\langle z',x\rangle), f(\langle z'',x\rangle))$, so we can conclude that $\lambda (f, \varphi)\in \mathbf{LL}_{\mathsf{pMet}}(Z, \mathbf{LL}_{\mathsf{pMet}}(X,Y))$. \item[($\Leftarrow$)] the map $\mathsf{ev}(\langle g,\psi\rangle, \chi)=\langle \mathsf{ev}(g), \mathsf{ev}(\psi, \chi)\rangle $ is defined by \begin{align*} \mathsf{ev}(f)(\langle z,x\rangle) &= f(z)(x) \\ \mathsf{ev}(\psi, \chi) (\langle z,x\rangle, \langle \zeta, \alpha\rangle) &= \chi(\langle z,\zeta\rangle)(x)+ \psi(z)(\langle x,\alpha\rangle) \end{align*} The map $\mathsf{ev}(\psi, \chi)$ is additive in its second variable. In fact we have \begin{align*} \mathsf{ev}(\psi,\chi)(\langle z,x\rangle, \langle 0, 0\rangle) & = \chi(\langle z,0\rangle)(x)+ \psi(z)(\langle x,0\rangle)\\ & =0+0=0 \end{align*} and \begin{align*} & \mathsf{ev}(\psi,\chi)(\langle z,x\rangle, \langle \zeta+\zeta', \alpha+\alpha'\rangle) \\ & =\chi(\langle z,\zeta+\zeta'\rangle)(x)+ \psi(z)(\langle x,\alpha+\alpha'\rangle) \\ & =\chi(\langle z,\zeta\rangle)(x)+\chi(\langle z,\zeta')(x) + \psi(z)(\langle x,\alpha\rangle)+\psi(z)(\langle x,\alpha'\rangle) \\ & = \chi(\langle z,\zeta\rangle)(x)+ \psi(z)(\langle x,\alpha\rangle)+\chi(\langle z,\zeta')(x) +\psi(z)(\langle x,\alpha'\rangle) \\ & = \mathsf{ev}(\psi,\chi)(\langle z,x\rangle, \langle \zeta,\alpha\rangle)+ \mathsf{ev}(\psi,\chi)(\langle z,x\rangle, \langle \zeta',\alpha'\rangle) \end{align*} Moreover, for all $z\in Z$ and $x\in X$ there exists $\zeta_{z}\gg 0, \alpha_{x}\gg0$ (which implies $\langle \zeta_{z},\alpha_{x}\rangle \gg 0$) such that whenever $c(z,z'),c(z,z'')\leq \zeta_{z}$ and $a(x,x'),a(x,x'')\leq \alpha_{x}$ \begin{align*} &\mathsf{ev}(\psi,\chi)(\langle z,x\rangle, \langle \zeta, \alpha \rangle)\\ & \geq b(f(z')(x'), f(z')(x''))+ b(f(z')(x''),f(z'')(x'')) \\ & \geq b(f(z')(x'), f(z'')(x'')\\ & = b(\mathsf{ev}(f)(\langle z',x'\rangle), \mathsf{ev} (f)(\langle z'',x''\rangle)) \end{align*} We can thus conclude that $\mathsf{ev}(\langle g,\psi\rangle, \chi)\in \mathbf{LL}_{\mathsf{pMet}}(Z\times X, Y)$. \end{itemize} It remains to show that $\lambda$ and $\mathsf{ev}$ inverse each-other: \begin{itemize} \item on one side we have $$\mathsf{ev}( \langle \lambda (f), \lambda_{0}(\varphi)\rangle, \lambda_{1}(\varphi)\rangle) = \langle \mathsf{ev}(\lambda (f)), \mathsf{ev}(\lambda_{0}(\varphi), \lambda_{1}(\varphi))\rangle =\langle f, \varphi\rangle $$ since $\mathsf{ev}(\lambda_{0}(\varphi), \lambda_{1}(\varphi))(\langle z,x\rangle, \langle \zeta, \alpha\rangle)= \varphi(\langle z,x\rangle, \langle \zeta, 0\rangle)+ \varphi(\langle z,x\rangle, \langle 0, \alpha\rangle) = \varphi(\langle z,x\rangle, \langle \zeta, \alpha\rangle)$ by the additivity of $\varphi$. \item on the other side we have $$\lambda (\mathsf{ev}(\langle g,\psi\rangle, \chi))= \lambda ( \mathsf{ev}(g), \mathsf{ev}(\psi, \chi))= ( \langle \lambda(\mathsf{ev}(g)), \lambda_{0}(\mathsf{ev}(\psi, \chi))\rangle, \lambda_{1}(\mathsf{ev}(\psi, \chi))\rangle= (\langle g,\psi\rangle, \chi)$$ since $\lambda_{0}(\mathsf{ev}(\psi, \chi))(z)(\langle x,\alpha\rangle)= \psi(z)(\langle x, \alpha\rangle) + \chi(\langle z,0\rangle)(x)=\psi(z)(\langle x, \alpha\rangle)$ and \\ $\lambda_{1}(\mathsf{ev}(\psi, \chi))(\langle z, \zeta\rangle)(x)= \psi(z)(\langle x, 0\rangle) + \chi(\langle z,\zeta\rangle)(x)=\chi(\langle z,\zeta\rangle)(x)$. \end{itemize} The cartesian closure of $\mathbf{LL}_{\mathsf{Met}}$ is proved as follows: if $f\in \mathbf{LL}_{\mathsf{Met}}(Z\times X,Y)$, then $f$ admits a family of LL-constants $\varphi$. Then for all $z\in Z$, $\lambda_{0}(\varphi)(z)$ is a family of LL-constants for $\lambda(f)(z)$, which implies that $\lambda (f)(z)\in \mathbf{LL}_{\mathsf{Met}}(X,Y)$; moreover, $\lambda_{1}(\varphi)$ is a family of LL-constants for the application $z\mapsto \lambda(f)(z)$, so we can conclude that $\lambda(f)\in \mathbf{LL}_{\mathsf{Met}}(Z,Y^{X})$. If now $f\in \mathbf{LL}_{\mathsf{Met}}(Z, Y^{X})$, then for all $z\in Z$, the set of families of LL-constants for $f(z)$ is non-empty; by the axiom of choice, there exists then a function $\psi$ yielding, for all $z\in Z$, a family of LL-constants for $f(z)$. Moreover $f$ itself admits a family of LL-constants $\chi$. Then the map $\mathsf{ev}(\psi,\chi)$ is a family of LL-constants for $\mathsf{ev}(f)$, so we deduce $\mathsf{ev}(f)\in \mathbf{LL}_{\mathsf{Met}}(X,Y)$. It remains to prove that $U$ is a cartesian closed functor. This descends from the following facts: \begin{itemize} \item ${X\times Y}/a\times b\simeq( X/a)\times( Y/b)$: in fact $\langle x,y\rangle \simeq_{a\times b}\langle x',y'\rangle $ iff $x\simeq_{a}x'$ and $y\simeq_{b}y'$. \item ${\mathbf{LL}_{\mathsf{pMet}}(X,Y)}/{b^{X}} \simeq ({Y}/{b})^{({X}/{a})}$: first, observe that $(f,\varphi)\simeq_{b^{X}}(g,\psi)$ iff for all $x\in X$, $f(x)\simeq_{b} g(x)$ iff for all $x,y\in X$, $a(x,y)=0$ implies $f(x)\simeq_{b}g(y)$ (since $f,g$ are stable under $\simeq_{a}$-classes). Now, for all $\simeq_{a}$-stable functions $f,g$, let $f\sim g$ iff for all $x,y\in X$, $a(x,y)=0$ implies $f(x)\simeq_{b}g(y)$. Then the claim follows from the observation that the equivalence classes of $\sim$ are in bijection with the functions from $\simeq_{a}$-classes to $\simeq_{b}$-classes. \end{itemize} Finally, since for all pseudo-metric space $(X,Q,a)$ we have that $Ua([x],[y])=a(x,y)$, from $b(f(y),f(z))\leq \varphi(x, a(y,z))$ we deduce $Ub(Uf([y]), Uf([z]))\leq \tilde\varphi([x], Ua([y],[z]))$. We conclude then that $\tilde\varphi$ is a family of LL-constants for $Uf$. \end{proof} } The category $\mathbf{LL}^{*}$ is in some sense more constructive than $\mathbf{LL}$ since to show that cartesian closure of the latter one needs the axiom of choice (see Appendix). \begin{example} In $\mathbf{LL}$ the space of locally Lipschitz functions $\mathbf{LL}(\BB R, \BB R)$ is endowed with the \emph{pointwise} metric $d_{\mathsf{Point}}(f,g): \BB R\to \BB R_{\geq 0}$, where $d_{\mathsf{Point}}(f,g)(x)=d_{\mathsf{Euc}}(f(x),g(x))$. \end{example} \subsection{Locally Lipschitz Models} For any cartesian closed category, $\BB C$, we let a \emph{LL-model of $\BB C$} be a cartesian closed functor $F: \BB C\to \mathbf{LL}_{\mathsf{pMet}}$, Observe that a LL-model $F: \BB C\to \mathbf{LL}_{\mathsf{pMet}}$ induces a cartesian closed functor $U\circ F:\BB C\to \mathbf{LL}_{\mathsf{Met}}$. Concretely, a LL-model consists in the following data: \begin{itemize} \item for any object $X$ of $\BB C$, a finitary pseudo-metric space $(\model X, \nudel X, \fini X , a_{X})$; \item for any morphism $f\in \BB C(X,Y)$, a LL-map $\model f: \model X\to \model Y$ stable on the $a_{X}$-classes, and a family of LL-constants $\nudel f: \model X\times \nudel X\to \nudel Y$ for $\model f$, \end{itemize} where the application $f\mapsto \nudel f$, which plays the role of the derivative in this setting, satisfies a bunch of properties that we discuss in some more detail below We now define a concrete model of the simply typed $\lambda$-calculus over a set of locally Lipschitz functions. For all $n>0$, let us fix a set $\C L_{n}$ of {locally Lipschitz} functions $f: \BB R^{n}\to \BB R$ (in the usual sense), and for each $f\in \C L_{n}$, let us fix a function $\mathsf{Lip}(f): \BB R^{n}\to [0,+\infty)$ associating each $\vec x\in \BB R^{n}$ with a local Lipschitz constant $\mathsf{Lip}(f)(\vec x)$ so that when $\vec y, \vec z$ are in some open neighborhood of $\vec x$, $$ |f(\vec y)- f(\vec z)| \leq \mathsf{Lip}(f)(\vec x)\cdot d^{n}_{\mathsf{Euc}}(\vec y,\vec z) $$ where $d^{n}_{\mathsf{Euc}}(\vec y, \vec z)=\sqrt{\sum_{i}{(y_{i}-z_{i})^{2}}}$. For any simple type $\sigma$, a finitary pseudo-metric space $(\model\sigma, \nudel \sigma, \fini{\sigma}, a_{\sigma})$ is defined by first letting $ \model\mathsf{Real} = \BB R$, $\fini{\mathsf{Real}} =\BB R^{\infty}_{\geq 0}$, $ \nudel\mathsf{Real} =[0,\infty]_{+} $, $ a_{\mathsf{Real}} =d_{\mathsf{Euc}}$ and then lifting the definition to all other types exploiting the cartesian closed structure of $\mathbf{LL}^{*}$. For any simple type $\sigma$, $U(\model \sigma, \nudel \sigma, \fini \sigma, a_{\sigma})$ is then a standard metric space (observe in particular that one has $Ua_{\sigma\to \tau}(f,g)(x)=Ua_{\sigma}(f(x),g(x))$). Moreover, given a context $\Gamma=\{x_{1}:\sigma_{1},\dots, x_{n}:\sigma_{n}\}$ and a term $t$ of type $\Gamma \vdash t:\sigma$ (that we take as representative of a class of terms of type $(\prod_{i=1}^{n}\sigma_{i})\to \sigma$), the functions $\model t: \prod_{i=1}^{n}\model{\sigma_{i}} \to \model \sigma$ and $\nudel t: \prod_{i=1}^{n}\model{\sigma_{i}}\times \prod_{i=1}^{n}\nudel{\sigma_{i}}\to \nudel \sigma$ are defined by a straightforward induction on $t$. We illustrate below only the definition of $\nudel t$: \begin{align*} \nudel{\TT r} (\vec x, \vec \alpha)& = 0 \\ \nudel{\TT f}(\vec x, \vec \alpha) & = \mathsf{Lip}(f)(\vec x)\cdot(\sum \vec \alpha) \\ \nudel{x_{i}}(\vec x, \vec \alpha) & = \alpha_{i}\\ \nudel{\langle t,u\rangle}(\vec x, \vec \alpha)& =\langle \nudel t(\vec x, \vec \alpha),\nudel u(\vec x, \vec \alpha)\rangle\\ \nudel{t\pi_{i}}(\vec x, \vec \alpha) & = \pi_{i}(\nudel t(\vec x,\vec \alpha)) \\ \nudel{\lambda y.t}(\vec x, \vec \alpha) & = \lambda y.{ \nudel t}(\vec x*y, \vec \alpha*0 ) \\ \nudel{tu}(\vec x, \vec \alpha) & = {\nudel t}(\vec x, \vec \alpha)(\model u(\vec x)) \\ & \quad + \model{t}_{1}(\vec x, \vec \alpha)(\model u(\vec x), \nudel u(\vec x, \vec \alpha)) \end{align*} where recall that for $t$ of type $\tau\to \sigma$, $\model t$ is a pair $ \langle \model t_{0}, \model t_{1}\rangle$ with $\model t_{0}(\vec x,\vec \alpha)\in \model \sigma^{\model \tau}$ and $\model t_{1}(\vec x, \vec \alpha) \in \nudel \sigma^{\model \tau \times \nudel \tau}$. \begin{theorem}[Soundness]\label{thm:llstlc} For all simply typed term $t$ such that $\Gamma \vdash t:\tau$, $(\model t, \nudel t)\in \mathbf{LL}_{\mathsf{pMet}}(\model \Gamma, \model \sigma)$. Moreover, if $t\longrightarrow_{\beta} u$, then $\model t=\model u$ and $\nudel t=\nudel u$. \end{theorem} Observe that since the QLR $(\model \sigma, \nudel \sigma, a_{\sigma})$ are metric spaces, the Fundamental Lemma reduces in this case to the remark that $a_{\sigma}(\model t, \model t)=0$ holds for all term $t$ of type $\sigma$. Instead, one can prove a ``local'' version of the contextuality lemma: \begin{corollary}[local contextuality of distances] For all terms $\vdash t,u:\sigma$ there exists $\delta_{t}\in \nudel{\sigma}$, with $\delta_{t}\gg 0$, such that for all contexts $\TT C[\ ]: \sigma \vdash \tau$ $$ a_{\tau}(\model{\TT C[t]}, \model{\TT C[u]}) \leq \nudel{\TT C}( \model t, a_{\sigma}(\model t, \model u)) $$ holds whenever $a_{\sigma}(\model t,\model u)\leq \delta_{t}$. \end{corollary} \subsection{Lipschitz Derivatives and Cartesian Differential Categories} Due to their different function spaces, the derivatives constructed in $\mathbf{LL}_{\mathsf{pMet}}$ (i.e.~the maps $\nudel t$) behave differently with respect to the derivatives from $\mathbf{Q}$. In particular, the former behave more closely to the derivatives found in \emph{Differential $\lambda$-Categories} \cite{BUCCIARELLI2010213} (in short D$\lambda$C), the categorical models of the differential $\lambda$-calculus \cite{ER} We recall that a D$\lambda$C is a left-additive \cite{Blute2009} category $\BB C$ in which every morphism $f\in \BB C(X, Y)$ is associated with a morphism $\mathsf{D}(f)\in \BB C(X\times X,Y)$ satisfying a few axioms: the axioms (D1)-(D7) of Cartesian Differential Categories \cite{Blute2009}, plus an additional axiom ($\mathsf{D}$-curry) \cite{BUCCIARELLI2010213} relating derivatives and the function space. We list below the properties of the application $f\mapsto \nudel f$ in a QLR model inside $\mathbf{LL}_{\mathsf{pMet}}$. We let $\lambda_{\BB C}, \mathsf{ev}_{\BB C}$ indicate the isomorphism $\BB C(Z\times X,Y)\simeq \BB C(Z,\BB C(X,Y))$, $\mathsf{ev}^{*}_{\BB C}= \mathsf{ev}_{\BB C}(\mathrm{id}_{\BB C(X,Y})$, and similarly $\mathsf{ev}^{*}=\mathsf{ev}(\mathrm{id}_{\mathbf{LL}_{\mathsf{pMet}}(X,Y)})$: \begin{itemize} \item[(1)] $\nudel{\mathrm{id}}=\pi_{1}$, $\nudel{g \circ f}=\nudel g\circ \langle f\circ \pi_{1},\nudel f\rangle$; \item[(2)] $\nudel f(x,0)=0$, $\nudel f(x,\alpha+\beta)=\nudel f(x,\alpha)+\nudel f(x,\beta)$; \item[(3)] $\nudel{\pi_{1}}=\pi_{1}\circ \pi_{1}$, $\nudel{\pi_{2}}=\pi_{2}\circ \pi_{1}$; \item[(4)] $\nudel{\langle f,g\rangle}= \langle \nudel f, \nudel g\rangle$; \item[(5)] $\nudel{\lambda_{\BB C}(f)}= \lambda_{X} ( \nudel f \circ \langle \pi_{1}\times \mathrm{id}_{X}, \pi_{2}\times 0 \rangle)$\\ (where for $g:Z\times X\to Y$, $\lambda_{X}(g)=\lambda x.g(\langle\_,x\rangle)$) \item[(6)] $\nudel{\mathsf{ev}^{*}_{\BB C}\circ\langle h,g\rangle}= \mathsf{ev}^{*}\circ \langle\nudel h , g\circ \pi_{1}\rangle + \nudel{\mathsf{ev}_{\BB C}(h)} \circ \langle \langle \pi_{1},g\circ\pi_{1}, \langle 0, \nudel g\rangle\rangle$, (where $h\in\BB C(Z,\BB C(X,Y))$, $g\in \BB C(Z,X)$). \end{itemize} The properties above literally translate the fact that a QLR model is a cartesian closed functor: \begin{itemize} \item (1) says that $f\mapsto \nudel f$ is functorial; \item (2) says that $\nudel f$ is additive in its second variable; \item (3) and (4) say that the cartesian structure of $\BB C$ commutes with that of $\mathbf{LL}_{\mathsf{pMet}}$; \item (5) and (6) say that the cartesian closed structure of $\BB C$ commutes with that of $\mathbf{LL}_{\mathsf{pMet}}$. \end{itemize} (1)-(2)-(3)-(4) coincide with axioms (D2)-(D3)-(D4)-(D5) of Cartesian Differential Categories (in short, CDC). Actually, this is not very surprising, since these axioms describe the fact that the application $f\mapsto \langle f, \mathsf{D}(f)\rangle$ in a CDC $\BB C$ yields a cartesian functor (known as the \emph{tangent functor}, see \cite{Cockett2011}). Observe that the other axioms of CDCs do not make sense in our setting, because $\mathbf{LL}_{\mathsf{pMet}}$ is not left-additive and there are no ``second derivatives'' in $\mathbf{LL}_{\mathsf{pMet}}$. Finally, property (5) is precisely axiom ($\mathsf{D}$-curry) of D$\lambda$Cs, and property (6) can be deduced in any D$\lambda$C from the other axioms (cf.~\cite{BUCCIARELLI2010213}, Lemma 4.5). \section{Introduction} In the literature on program semantics much attention has been devoted to program equivalence, and, accordingly, to the study of program transformations which do not produce observable changes of behavior. However, in fields involving {numerical} or {probabilistic} forms of computation one often deals with transformations that \textit{do} alter program behavior, replacing a piece of program with one which is only approximately equivalent. For example, numerical methods (e.g.~linear regression, numerical integration) are based on the replacement of computationally expensive operations with more efficient, although less precise, ones. On another scale, statistical learning algorithms compute approximations of a desired function by fitting with a finite sample. The challenge that accompanies the use of such \emph{approximate} program transformations \cite{chaudhuri} is to come up with methods to measure and bound the error they produce. This has motivated much literature on \emph{program metrics} \cite{ARNOLD1980181, VANBREUGEL20011,Gabo2019, Escardo1999, BAIER1994171,DalLago2015, 10.1007/978-3-662-44584-6_4, 10.1007/978-3-662-54434-1_13, 10.1145/3209108.3209149}, that is, on semantics in which types are endowed with a notion of distance. This approach has found widespread applications, for example in differential privacy \cite{Gaboardi2017, 10.1007/978-3-642-29420-4_3, Barthe_2012} and reinforcement learning \cite{Panangaden2011}. A natural framework for the study of program metrics and their abstract properties is provided by so-called \emph{generalized metrics}. Since Lawvere's \cite{Lawvere1973} it has been known that some of the basic axioms of standard metric spaces (notably, the \emph{reflexivity} and \emph{transitivity} axioms $d(x,x)=0$ and $d(x,z)+d(z,y)\geq d(x,y)$) can be seen, at a higher level of abstraction, as describing the structure of a category \emph{enriched} over some quantitative algebra. Typically, when this algebra is the usual semi-ring of positive reals (i.e.~when ``0'' actually means zero, and ``+'' actually means plus), one gets the metric spaces everyone is used to. However, one can consider generalized distance functions $d: X\times X\to Q$, where $Q$ is now a different algebra (typically a \emph{quantale} or a \emph{quantaloid} \cite{Hofmann2014}), and the monoidal structure of $Q$ determines the actual meaning of the metric axioms. Well-investigated examples of this generalized approach are given by \emph{ultra}-metric spaces \cite{VANBREUGEL20011,Escardo1999}, \emph{partial} metric spaces \cite{matthews, Bukatin1997,AGT7849,Stubbe2018} and \emph{probabilistic} metric spaces \cite{Sklar1983, Hofmann:2013aa}. Generalized program metrics have been applied in several areas of computer science, e.g.~to co-algebraic \cite{DBLP:journals/lmcs/BaldanBKK18,Pouzet2020}, and concurrent \cite{10.1007/978-3-662-44584-6_4} systems, and to algebraic effects \cite{Plotk,10.1145/3209108.3209149}. However, the application of program metrics to even basic \emph{higher-order} languages like the simply typed $\lambda$-calculus $\mathsf{ST\lambda C}$ has so far proved unsatisfactory. One can mention both theoretical and practical reasons for this failure. At the abstract level, for instance, there is the well-known fact that standard categories of metric spaces, even generalized, are usually \emph{not} cartesian closed, and thus only account for linear or sub-exponential variants of $\mathsf{ST\lambda C}$ \cite{10.1145/1932681.1863568, Gaboardi_2013, Gaboardi2017}. At a more practical level, there is the observation that even with such restrictions, the distance between two functional programs computed in such models is often not very informative, as it estimates the error of replacing one program by the other one \emph{in the worst case}, and thus independently of the current \emph{context} in which these programs are placed. In this paper we introduce a new class of program metric semantics for $\mathsf{ST\lambda C}$ which overcome the aforementioned difficulties. These semantics arise from the study of a class of quantitative models based on what we call \emph{quantitative logical relations} (in short, QLR). A QLR is just what remains of a generalized metric space when one discards the reflexivity and transitivity axioms; in other words, it is nothing more than a function $a: X\times X\to Q$ relating pairs of points $x,y\in X$ with an element $a(x,y)$ of some quantitative algebra $Q$. At the same time, such functions can be seen as quantitative analogs of standard logical relations. The difference is that while with the latter two programs may or may not be related, with QLR two programs are always related \emph{to a certain degree}. We believe that models for $\mathsf{ST\lambda C}$ should be as elementary as possible. By the way, the category of sets is itself a denotational model of $\mathsf{ST\lambda C}$. For this reason, we do not, at first, impose any restriction (e.g.~continuity, Lipschitz continuity) over the set-theoretic functions between QLR. Importantly, maps of QLR can relate functions measuring distances over \emph{different} quantitative algebras. For this reason, set-theoretic maps are accompanied by a second map, a sort of \emph{derivative}, relating \emph{errors} in input with errors in output. This idea, which extends similar ones from \emph{differential logical relations} \cite{dallago,dallago2} and \emph{diameter spaces} \cite{Geoffroy2020}, mark the main difference between our approach and usual metric semantcs (in which one usually considers a \emph{fixed} quantale), and is a key ingredient to obtain models of the full $\mathsf{ST\lambda C}$. Our first contribution is to show that several variants of QLR form cartesian closed categories and that some standard results about logical relations have a quantitative analog in the realm of QLR. These results show that QLR-models capture quantitative relational reasoning of higher-order programs in a fully compositional way. However, recall that our starting point was program metric semantics, and QLR, by their very definition, are \emph{not} metric spaces. Yet, since generalized metrics are particular cases of QLR, the latter provide an ideal environment to investigate \emph{which} families of generalized metrics (i.e.~which choices of the ``0'' and the ``+'') adapt well to the cartesian closed structure. Our second contribution is a characterization of the class of generalized metric spaces that give rise to cartesian closed categories of QLR. These results demonstrate the existence of a variety of compositional metric semantics of $\mathsf{ST\lambda C}$ which extend the Euclidean metrics over the reals to all simple types. Finally, we show that the derivatives found in QLR-models can be compared with those appearing in other quantitative models of $\mathsf{ST\lambda C}$, like those arising from the \emph{differential $\lambda$-calculus} \cite{ER,Blute2009, BUCCIARELLI2010213}. \medskip \paragraph*{Outline} After motivating the introduction of QLR in Section \ref{section2}, in Section \ref{section3} we recall the definition of some classes of generalized metric spaces; in Section \ref{section4} we introduce two cartesian closed categories $\mathbf{Q}$ and $\mathbf{Q}^{\mathsf{r}}$ of QLR, and we describe the interpretation of $\mathsf{ST\lambda C}$ in them. In Section \ref{section5} we investigate the generalized metrics which form cartesian closed sub-categories of $\mathbf{Q}$ and $\mathbf{Q}^{\mathsf{r}}$. Finally, in Section \ref{section5} we construct a different cartesian closed category $\mathbf{LL}_{\mathsf{Met}}$ of generalized metric spaces based on a ``locally Lipschitz'' condition for QLR morphisms. \section{Higher-Order Metric Semantics}\label{section2} \subsection{Program Metrics and Higher-Order Languages} Program metrics have been widely investigated to capture properties like program \emph{similarity} and \emph{sensitivity}. The fundamental idea is usually to associate types $\sigma,\tau$ with metric spaces, and programs $f:\sigma\to \tau$ with \emph{non-expansive}, or more generally \emph{Lipschitz continuous} functions. This means that for all programs $t,u$ of type $\sigma$, the distance between $f(t)$ and $f(u)$ does not exceed that between $t$ and $u$ by more than a fixed factor $L$ (formally, $d(f(t),f(u))\leq L\cdot d(t,u)$). However, the approach just sketched is not satisfactory for the interpretation of \emph{higher-order} languages, as those based on $\mathsf{ST\lambda C}$. The main problem is that the category $\mathsf{Met}_{Q}$ of metric spaces over a quantale $Q$ and non-expansive maps \cite{Hofmann2014}, which provides the abstract setting for usual program metrics, is not compatible with the usual structure of models of $\mathsf{ST\lambda C}$. More precisely, while the space $ \mathsf{Met}_{Q}(X,Y)$ of non-expansive functions can be endowed with a metric (the $\sup$-metric $d_{\sup}(f,g)=\sup\{d(f(x),g(x))\mid x\in X\}$), this construction does \emph{not} yield a right-adjoint to the categorical product. For this reason $\mathsf{Met}_{Q}$ is not a \emph{cartesian closed} category (although $\mathsf{Met}_{Q}$ still admits some interesting cartesian closed \emph{sub}-categories, see \cite{CLEMENTINO20063113, Clementino:2009aa}). This abstract issue is not the only one has to face, though. After all, category theory is usually invoked in program semantics as a way to enforce \emph{compositionality}, i.e.~the property by which the semantics of a composed program is expressed in terms of the semantics of its components. Yet, even if we accept to restrict ourselves to higher-order languages compatible with the categorical structure of $\mathsf{Met}_{Q}$ (like e.g.~the system $\mathsf{Fuzz}$ \cite{10.1145/1932681.1863568}), the metric $d_{\sup}$ still does not account for the behavior of higher-order programs in a sufficiently {compositional}, and, in the end, informative way. For example, as observed in \cite{dallago}, consider the two Lipschitz functions $f=\lambda x.\sin(x)$ and $g=\lambda x.x$: since $f$ and $g$ get arbitrarily far from each other in the worst case (i.e.~as $x$ approaches $\pm\infty$), one can deduce that $d_{\sup}(f,g)$ is infinite. Hence, the distance $d_{\sup}(f,g)$ provides no significant information in any situation in which $f$ is replaced by $g$ as a {component} of a larger program: for instance, if $\TT C[\ ]$ is a context applying a function on values close to 0, the programs $\TT C[f]$ by $\TT C[g]$ will likely turn out close, yet there is no way to predict this fact on the basis of $d_{\sup}(f,g)$. A related issue occurs with \emph{contextual} notions of distance, as those found e.g.~in probabilistc extensions of the $\lambda$-calculus \cite{DalLago2015}. These metrics extend usual contextual equivalence, by letting the distance $d_{\mathsf{ctx}}(t,u)$ between two objects of type $\sigma$ be the sup of all observable distances $d_{\mathsf{Euc}}(\TT C[f], \TT C[g])$, for any context $\TT C[\ ]:\sigma\Rightarrow \mathsf{Real}$. As shown in \cite{10.1007/978-3-662-54434-1_13}, the non-linearity of $\mathsf{ST\lambda C}$ can be used to define contexts that arbitrarily {amplify} distances, with the consequence that the metric $d_{\mathsf{ctx}}$ trivializes onto plain contextual equivalence. \subsection{From Program Metrics to Quantitative Logical Relations} To overcome these issues, in Section \ref{section4} we introduce \emph{quantitative logical relations}, a quantitative extension of usual logical relations (generalizing previous approaches \cite{dallago,dallago2, Geoffroy2020}) which, on the one hand, applies to higher-order programs without restrictions (e.g.~Lipschitz-continuity), and, on the other hand, enables reasoning about behavioral similarity in a fully compositional way. Semantically, logical relations for a programming language $\C L$ can be introduced starting from a denotational model of $\C L$ (for simplicity, we consider a simple set-theoretic model, associating each type $\sigma$ with a set $\model \sigma$ and each program $t:\sigma\to \tau$ with a function $\model t:\model \sigma \to \model \tau$); one then constructs a more refined model whose objects are binary relations $r: \model \sigma \times \model \sigma\to\{0,1\}$, and whose arrows are those functions from our original model which send related points into related points (in more abstract terms, this construction is an instance of the \emph{glueing} construction, see \cite{Schalk2003}). The so-called \emph{Fundamental Lemma} tells us then that any program $t:\sigma\to \tau$ of $\C L$ yields a morphism in this model, i.e.~preserves relatedness. While in logical relations relatedness is measured over a fixed algebra (the Boolean algebra $\{0,1\}$), in QLR relatedness is measured over a larger class of quantales. Hence, a QLR is of the form $a: \model \sigma \times \model \sigma\to \nudel \sigma$, where $\nudel \sigma$ is some quantale associated with $\sigma$. Typically, when $\sigma$ is a functional type, $\nudel{\sigma}$ will be some quantale of functions mapping differences in input into differences in output. To interpret a program $t:\sigma\to \tau$ we must accompany the function $\model t$ with a second function $\nudel t: \model \sigma \times \nudel \sigma \to \nudel \tau$ mapping differences in $\nudel\sigma$ \emph{around some point of $\model\sigma$} into differences in $\nudel \tau$. The function $\nudel t$ can be seen as sort of \emph{derivative} of $\model t$, and is the key ingredient to reason about $t$ in a compositional way: if $\alpha\in \nudel \sigma$ measures the similarity of two programs $u,v$ and $\TT C[\ ]:\sigma \to \tau$ is a context with derivative $\nudel{\TT C}$, then by composing $\nudel{\TT C}$ with $\model u$ and $\alpha$, we obtain a measure of the similarity between $\TT C[u]$ and $\TT C[v]$. Notably, the Fundamental Lemma of logical relations translates in this setting into a result showing that any program $t$ from $\C L$ translates into a derivative $\nudel t$, yielding a fully compositional semantics for $\C L$. For instance, take $\model \mathsf{Real}=\BB R$ and $\nudel \mathsf{Real}=\BB R_{\geq 0}$; if $f,g: \mathsf{Real}\to \mathsf{Real}$ are the two programs $\lambda x.\sin(x), \lambda x.x$ seen before and $\TT C[\ ]=[\ ]0:(\mathsf{Real}\to \mathsf{Real})\to \mathsf{Real}$ is the context that applies a function to $0$, in our setting we can reason as follows: first, the difference $d(\model f, \model g)$ will be itself a function mapping small differences in input \emph{around} 0 onto small differences in output; secondly, the derivative $\nudel{\TT C}$ will be such that that the value $\nudel{\TT C}(f,\varphi)$ only depends on how much $\varphi$ grows on small neighborhoods of $0$; hence, the difference between $\TT C[f]$ and $\TT C[g]$, computed by applying $\nudel{\TT C}$ to $\model f $ and to $d(\model f, \model g)$, will yield a value close to 0. Similar ideas already appear in \cite{dallago2, Geoffroy2020} and have been shown to provide a compositional account of techniques from \emph{incremental computing} and \emph{approximate programming} (e.g.~{loop perforation} \cite{loopperf} and {numerical integration}). The study of QLR, that we develop here, is intended to capture the basic structure underlying such (non-equivalent) constructions, and to characterize a much larger family of quantitative and metric models to which those from \cite{dallago2, Geoffroy2020} belong. \subsection{...and back to Generalized Metric Spaces} While a QLR $a: \model \sigma \times \model \sigma \to \nudel\sigma$ needs not be a metric, several classes of generalized metric spaces can be seen as QLR satisfying further properties. One can thus ask \emph{which} families of generalized metrics can be lifted to all simple types within a given QLR-model. In Section \ref{section5} we investigate generalized metrics in categories of QLR with unrestricted morphisms (that is, with no continuity or Lipschitz restriction). We show that, under some mild assumptions, lifting metrics to simple types forces distances to be idempotent (i.e.~to satisfy $\alpha=\alpha+\alpha$). This implies that the generalized metrics that can be lifted to all simple types are of two kinds: firstly, the \emph{ultra-metric} and \emph{partial ultra-metric} spaces, that is, those metrics based on an idempotent quantitative algebra; secondly, those generalized metrics whose distance function can be \emph{factorized} through an idempotent metric. By extending a construction from \cite{Geoffroy2020} relating partial metrics with lattice-valued distances, we show that the Euclidean metric, as well as many other standard {metrics} and {partial metrics}, belong to this second class. In Section \ref{section6} we investigate generalized metrics in categories of QLR where morphisms satisfy suitable generalizations of the \emph{Lipschitz} and \emph{locally Lipschitz} continuity conditions. We first show that the first condition does \emph{not} yield a cartesian closed category, for reasons very similar to those found when considering metrics over a fixed quantale. We then show that the second does yields, instead, a model of $\mathsf{ST\lambda C}$ in which types are interpreted by generalized metric spaces. \section{Generalized Metric Spaces}\label{section3} In this paper we consider several variants of metric spaces. It is thus useful to adopt a general and abstract definition of what we take a (generalized) metric space to be. We exploit the abstract formulation of generalized metric spaces as enriched categories dating back to Lawvere's \cite{Lawvere1973}, who first observed that a metric space in the standard sense can be seen as a category enriched in the monoidal poset $([0,+\infty), \geq, 0,+)$ of positve real numbers under reversed ordering and addition. \subsection{Metrics over an Arbitrary Quantale} The standard axioms of metric spaces involve an order relation and a monoidal operation (addition) with a neutral element 0. This structure is characterized by a \emph{monoidal poset}, that is, a tuple $(M,\geq, 0,+)$ where $(M,\geq)$ is a poset and $(M,0,+)$ is a monoid such that $+$ is monotone. In practice, one is usually interested in measuring distances in monoidal posets where $\sup$s and $\inf$s always exist. This leads to consider (commutative and integral) quantales: \begin{definition} A (commutative) quantale is a commutative monoidal poset $(Q,0,+,\geq)$ such that $(Q,\geq)$ is a complete lattice satisfying $ \alpha +\bigwedge S= \bigwedge\{\alpha+\beta\mid \beta\in S\}$, for all $S\subseteq Q$. A quantale $(Q,0,+,\geq)$ is \emph{integral} when $0=\bot$. A commutative quantale $Q$ is a \emph{locale} when $0=\bot$ and $\alpha=\alpha+\alpha$ holds for all $\alpha\in Q$ (or, equivalently, when $\alpha+\beta=\alpha\lor\beta$). \end{definition} \begin{remark} With respect to common presentations of quantales, we adopt here the \emph{reversed} order (so that $\bigvee$s and $\bigwedge$s are inverted), as this is more in accordance with the quantitative intuition. \end{remark} \begin{example}[The Lawvere quantale] The structure $(\BB R^{\infty}_{\geq 0}, 0, +, \leq)$, where $\BB R^{\infty}_{\geq 0}$ is the set of positive reals plus $\infty$, is a commutative and integral quantale, and is usually referred to as the \emph{Lawvere quantale} \cite{Hofmann2014}. If we replace $+$ with $\sup$, the resulting structure $(\BB R^{\infty}_{\geq 0}, 0, \sup, \leq)$ is a locale. \end{example} \begin{example}\label{example:subsets} For any commutative monoid $(M,0,+)$, the structure $(\wp(M), \{0\},+,\subseteq)$, is a commutative quantale, where $A+B=\{x+y\mid x\in A, y\in B\}$. \end{example} \begin{example} All products $\Pi_{i\in I}Q_{i}$ of (commutative and integral) quantales, with the pointwise order, are still commutative and integral quantales. \end{example} In a quantale $Q$ one can define the following two operations: \begin{align*} \alpha \multimapinv \beta = \bigwedge \{\delta\mid \beta + \delta \geq \alpha\}\qquad \alpha \Leftarrow \beta = \bigwedge \{\delta\mid \beta \vee \delta \geq \alpha\} \end{align*} In any quantale $\delta \geq \alpha \multimapinv \beta$ holds iff $\delta + \beta \geq \alpha$, that is, $\multimapinv$ is \emph{right-adjoint} to $+$. A quantale in which $\Leftarrow$ is right-adjoint to $\vee$, i.e.~$\delta \geq \alpha \Leftarrow \beta$ holds iff $\delta \vee \beta \geq \alpha$, is called a \emph{Heyting quantale} \cite{Hofmann2014, CLEMENTINO20063113}. The Lawvere quantale and all other quantales obtained from it by product are Heyting. Moreover, all locales are Heyting. \begin{example} In the Lawvere quantale $x\multimapinv y= \max\{0, x-y\}$ and $x\Leftarrow y$ is $0$ if $x\leq y$ and is $x$ otherwise. \end{example} Over any quantale $Q$ we can define generalized metric spaces as follows: \begin{definition}\label{def:gms1} A \emph{generalized metric space} is a triple $(X,Q,a)$ where $X$ is a set, $Q$ is a commutative quantale, and $a:X\times X\to Q$ satisfies, for all $x,y,z\in X$: \begin{align} 0 & \geq a(x,x) \tag{reflexivity}\\ a(x,y)+a(y,z) & \geq a(x,z) \tag{transitivity} \end{align} A generalized metric space is said: \begin{itemize} \item \emph{symmetric} if $a(x,y)=a(y,x)$; \item \emph{separated} if $a(x,y)=0$ implies $x=y$. \end{itemize} \end{definition} Observe that, when $Q$ is integral, from the reflexivity axiom it follows that $a(x,x)=0$ holds for all $x\in X$. Following usual terminology, we let a \emph{pseudo-metric space} be a symmetric metric space $(X,Q,a)$, and a \emph{standard metric space} be a separated pseudo-metric space. The \emph{Euclidean metric} is the standard metric space $(\BB R, \BB R^{\infty}_{\geq 0}, d_{\mathsf{Euc}})$ where $d_{\mathsf{Euc}}(x,y)=|x-y|$. \begin{example} A standard metric space $(X,Q,a)$ in which $Q$ is a locale is usually called a \emph{ultra-metric space}. The transitivity axiom reads in this case as $a(x,y)\vee a(y,z)\geq a(x,z)$. For instance, the \emph{sequence metric} on the set $X^{\BB N}$ of $X$-sequences $(x_{n})_{n\in \BB N}$ is the ultra-metric space $(X^{\BB N}, \BB R^{\infty}_{\geq 0}, d_{\mathsf{seq}})$ given by $d_{\mathsf{seq}}(x_{n},y_{n})=2^{-c(x_{n},y_{n})}$, where $c(x_{n},y_{n})$ is the length or the largest common prefix of $x_{n}$ and $y_{n}$. \end{example} \begin{example}\label{ex:prob} A standard metric space $(X,\Delta,a)$ in which $\Delta$ is the quantale of \emph{distributions}, i.e.~the left-continuous maps $f: \BB R_{\geq 0}\to[0,1]$ with pointwise ordering and monoidal operation $(f\oplus g)(r)=\bigwedge_{s+t=r}f(s)\cdot g(t)$, is an example of \emph{probabilistic metric space} \cite{Sklar1983, Hofmann:2013aa}. Observe that the transitivity axiom reads in this case as $a(x,y)(r)+a(y,z)(s)\geq a(x,y)(r+s)$ \end{example} \subsection{Partial Metric Spaces} In several approaches to program metrics one encounters distance functions which do not satisfy the reflexivity axiom $0\geq a(x,x)$. A basic example (see \cite{matthews}) is obtained when the sequence metric $d_{\mathsf{seq}}$ is extended to the set $\widehat X=\bigcup_{n}^{\infty}X^{n}\cup X^{\BB N}$ of \emph{finite and infinite} $X$-sequences (this kind of spaces are common, for instance, in domain theory): whenever $x_{n}$ is a sequence of length $k$, we have that $d_{\mathsf{seq}}(x_{n},x_{n})=2^{-k}>0$. The simplest way to define a metric with non-zero self-distances is simply to drop the reflexivity axiom. This yields the \emph{relaxed metrics} from \cite{Bukatin1997}. An even more drastic relaxation of the metric axioms is the one considered in \cite{dallago}, where transitivity is also weakened to\footnote{Actually, \cite{dallago} does not define a distance function $d:X\times X\to Q$ but rather a distance \emph{relation} $\rho \subseteq X\times Q\times X$ obeying a relaxed transitivity of the form $\rho(x,\alpha,y), \rho(y,\beta,y), \rho(y,\gamma,z) \Rightarrow \rho(x,\alpha+\beta+\gamma,y)$. In fact, this is the same thing as a function $d_{\rho}: X\times X\to \wp(Q)$ (where $\wp(Q)$ indicates the quantale of subsets of $Q$ from Example~\ref{example:subsets}) satisfying \eqref{eq:superrelaxed}.\label{foot1}} \begin{equation}\label{eq:superrelaxed} a(x,y)\leq a(x,z)+a(z,z)+a(z,y) \end{equation} We will refer to the latter as \emph{hyper-relaxed metrics}. A different approach consists in considering distance functions that \emph{do} satisfy both metric axioms, but relative to a \emph{different} monoidal structure over $Q$. The \emph{partial metric spaces} \cite{matthews,Bukatin1997}, developed to account for domains of objects akin to the set $\widehat X$, provide an example of this approach, as shown by the elegant presentation from \cite{Stubbe2018, STUBBE201495}, that we recall below. For any commutative integral quantale $Q$, let $\C D(Q)$ be the category whose objects are all elements of $Q$, and where $\C D(Q)(\alpha,\beta)$ is the complete lattice of \emph{diagonals} from $\alpha$ to $\beta$, i.e.~those $\delta\in Q$ satisfying $$\alpha+ (\delta\multimapinv \alpha) =\delta= (\delta\multimapinv \beta)+\beta$$ The identity morphism $\mathrm{id}_{\alpha}$ is just $\alpha$ (moreover, $\alpha$ is the smallest element of $\C D(Q)(\alpha,\alpha)$); the composition of two diagonals $\delta\in \C D(Q)(\beta,\alpha)$ and $\eta\in \C D(Q)(\gamma,\beta)$ is the diagonal $$ \eta+_{\beta} \gamma := \eta +(\gamma \multimapinv \beta)\in \C D(Q)(\gamma,\alpha) $$ The category $\C D(Q)$ is an example of \emph{quantaloid} (see~\cite{STUBBE201495}). \begin{example} In the Lawvere quantale, a diagonal from $x$ to $y$ is any real number $z\geq x,y$, and the composition law reads as $x+_{z}y:= x+y-z$. \end{example} \begin{remark}\label{rem:loca} When $Q$ is a locale, $\C D(Q)(\alpha,\beta)=\{\gamma\mid \alpha\vee \beta \leq \gamma\}$ and the composition law of $\C D(Q)$ coincides with that of $Q$, since $\alpha\vee (\beta\Leftarrow \gamma)= \alpha\vee \beta$ holds for all $\gamma \leq \beta$. Using this fact, the definition of the category of diagonals can be extended to the case in which $Q$ is just a complete lattice (and thus needs not be a locale), by letting $\C D(Q)(\alpha,\beta)=\{\gamma\mid \alpha\vee \beta \leq \gamma\}$, with identities $\mathrm{id}_{\alpha}=\alpha$ and composition given by $\lor$. The category $\C D(Q)$ is then a quantaloid precisely when $Q$ is a locale. \end{remark} Partial metric spaces can be defined as metric spaces with respect to the monoidal structure of diagonals: \begin{definition} A \emph{partial metric space} is a tuple $(X,Q,t,a)$ where $X$ is a set, $Q$ is a (commutative and integral) quantale, $t:X\to Q$ and $a:X\times X\to Q$ are such, for all $x,y,z\in X$, $a(x,y)\in \Delta(Q)(ty,tx)$ and: \begin{align} \mathrm{id}_{tx} & \geq a(x,x) \tag{reflexivity} \\ a(x,y) +_{ty} a(y,z) & \geq a(x,z) \tag{transitivity} \end{align} A partial metric space is said: \begin{itemize} \item \emph{symmetric} if $a(x,y)=a(y,x)$; \item \emph{separated} if $a(x,y)=a(x,x)=a(y,y)$ implies $x=y$. \end{itemize} \end{definition} \begin{remark} When $Q$ is integral, reflexivity forces $tx=a(x,x)$, so the partial metric structure is entirely determined by the triple $(X,Q,a)$. \end{remark} A symmetric and separated partial metric over the Lawvere quantale $a:X\times X\to \BB R^{\infty}_{\geq 0}$ satisfies the axioms below: \begin{description} \item[PMS1] \ $a(x,x)\leq a(x,y),a(y,x)$; \item[PMS2] \ $a(x,y)=a(y,x)$; \item[PMS3] \ if $a(x,x)=a(x,y)=a(y,x)$, then $x=y$; \item[PMS4] \ $a(x,y)\leq a(x,z) + a(z,y) - a(z,z) $. \end{description} Observe that a (symmetric and separated) metric is the same as a (symmetric and separated) partial metric with $a(x,x)=0$. Moreover, any (symmetric and separated) partial metric $a:X\times X\to Q$ gives rise to a (symmetric and separated) metric \begin{align*} a^{*}(x,y)= (a(x,y)\multimapinv a(x,x))+(a(x,y)\multimapinv a(y,y)) \end{align*} The terminology for \emph{pseudo-}, \emph{standard} and \emph{ultra-}metrics extends straightforwardly to from metric to partial metric spaces. For example, the sequence metric $d_{\mathsf{seq}}$ extended to $\widehat X$ yields a partial ultra-metric space. Another standard example of partial metric over the Lawvere quantale is the one defined over the set $\C I$ of \emph{closed intervals} $\{[r,s]\mid r\leq s\}$ by $p([r,s],[r',s'])= \max\{s,s'\}- \min\{r,r'\}$. \begin{remark} The definition of partial ultra-metric spaces can be extended, as we will do in Section \ref{section5}, to the case in which $Q$ is just a complete lattice, using Remark \ref{rem:loca}. However, one must be careful that all properties that rely on the existence of the right-adjoint $\Leftarrow$ need not hold in this case. \end{remark} \section{Quantitative Logical Relations}\label{section4} In this section we introduce two categories $\mathbf{Q}$ and $\mathbf{Q}^{\mathsf{r}}$ of quantitative logical relations. After describing their cartesian closed structure, we describe the interpretation of $\mathsf{ST\lambda C}$ in these categories and we show that some standard results about logical relations scale to QLR in a quantitative sense. \subsection{Two Categories of QLR} A \emph{quantitative logical relation} $(X,Q,a)$ (in short, a \emph{QLR}) is the given of a set $X$, a commutative quantale $Q$ and a function $a: X\times X\to Q$. A \emph{map of quantitative logical relations} $(X,Q,a)$, $(Y,R,b)$ is a pair $(f,\varphi)$, where $f:X\to Y$, $\varphi: X\times Q\to R$ and for all $x,y\in X$, $$ a(x,y)\leq \alpha \ \Rightarrow \ b(f(x),f(y)) \leq \varphi(x,\alpha) $$ QLR and their maps form a category $\mathbf{Q}$ having as identities the pairs $(\mathrm{id}_{X}, \lambda x\alpha.\alpha)$, and composition defined by $(g,\psi)\circ (f,\phi)= (g\circ f, \psi\circ \langle f\circ \pi_{1}, \varphi\rangle)$. The category $\mathbf{Q}$ is cartesian closed: given QLR $(X,Q,a)$ and $(Y,R,b)$, their cartesian product is the QLR $(X\times Y, Q\times R, a\times b)$, with unit $(\{\star\}, \{\star\}, \langle\star,\star\rangle\mapsto \star)$, and their exponential is the QLR $(Y^{X}, R^{X\times Q},d^{\mathbf{Q}}_{a,b})$ where \shortv{ \begin{center} \resizebox{0.48\textwidth}{!}{ $ d^{\mathbf{Q}}_{a,b}(f,g)(x,\alpha) = \sup \{d(f(x),g(y)), d(f(x),f(y)) \mid a(x,y)\leq \alpha\} $ } \end{center}} \longv{ $$ d^{\mathbf{Q}}_{a,b}(f,g)(x,\alpha) = \sup \{d(f(x),g(y)), d(f(x),f(y)) \mid a(x,y)\leq \alpha\} $$ } The isomorphism $\begin{tikzcd}\mathbf{Q}(Z\times X, Y)\ar[bend left=5]{r}{\lambda} & \mathbf{Q}(Z,Y^{X}) \ar[bend left=5]{l}[below]{\mathsf{ev}}\end{tikzcd}$ defining the cartesian closed structure is given by $\lambda(f,\varphi)=(\lambda (f),\lambda(\varphi))$ and $\mathsf{ev}(f,\varphi)=(\mathsf{ev}(f),\mathsf{ev}(\varphi))$, where \begin{align*} \lambda (f)(z)(x) & =f(\langle z,x\rangle)\\ \lambda (\varphi)(\langle z,\gamma\rangle)(\langle x, \alpha\rangle)&= \varphi(\langle\langle z,x\rangle, \langle \gamma,\alpha\rangle\rangle)\\ \mathsf{ev}(f)(\langle z,x\rangle)&=f(z)(x)\\ \mathsf{ev}(\psi)(\langle\langle z,x\rangle, \langle \gamma,\alpha\rangle\rangle)& = \psi(\langle z,\gamma\rangle)(\langle x, \alpha\rangle) \end{align*} Given QLR $(X,Q,a)$ and $(Y,R,b)$, for any function $f:X\to Y$ there exists a \emph{smallest} function $\mathsf{D}(f):X\times Q\to R$ such that $(f,\mathsf{D}(f))\in \mathbf{Q} (X,Y)$, defined by \begin{align} \mathsf{D}(f)(x,\alpha)= \sup \{ b(f(x),f(y))\mid a(x,y)\leq \alpha\} \end{align} We call $\mathsf{D}(f)$ the \emph{derivative} of $f$. Derivatives in $\mathbf{Q}$ satisfy the following properties: \begin{align} \mathsf{D}(\mathrm{id}_{X})(x, \alpha) & = \alpha \tag{D1} \label{eq:prop2}\\ \mathsf{D}( \pi_{i})(\langle x_{1},x_{2}\rangle, \langle \alpha_{1},\alpha_{2}\rangle) & = \alpha_{i} \tag{D2}\label{eq:prop3}\\ \mathsf{D}(\langle f,g\rangle)(x,\alpha) & =\langle \mathsf{D}(f)(x,\alpha), \mathsf{D}(g)(x,\alpha)\rangle \tag{D3}\label{eq:prop4}\\ \mathsf{D}(g\circ f)(x,\alpha) & \leq \mathsf{D}(g)(f(x),\mathsf{D}(f)(x,\alpha)) \tag{D4}\label{eq:prop5} \\ \mathsf{D}(\lambda(f))(x, \alpha) & \leq \lambda (\mathsf{D}(f))(x,\alpha) \tag{D5}\label{eq:prop6} \\ \mathsf{D}(\mathsf{ev}(f))(x,\alpha) & \leq \mathsf{ev}(\mathsf{D}(f))(x,\alpha) \tag{D6}\label{eq:prop7} \end{align} Properties \eqref{eq:prop2}-\eqref{eq:prop4} recall some of the axioms of \emph{Cartesian Differential Categories} \cite{Blute2009}, a well-investigated formalization of abstract derivatives. Property \eqref{eq:prop5} is a \emph{lax} version of the chain rule, and properties \eqref{eq:prop6} and \eqref{eq:prop7} state that $\mathsf{D}$ commutes with the cartesian closed isomorphisms in a lax way. \begin{remark} Derivatives $\partial(f)$ in Cartesian Differential Categories are \emph{additive} in their second variable, i.e.~they satisfy $\partial(f)(x,0)=0$ and $\partial(f)(x,\alpha+\beta)=\partial(f)(x,\alpha)+\partial(x,\beta)$. By contrast, it is not difficult to construct counter-examples to the additivity of $\mathsf{D}(f)$. Let $f,g:\BB R\to \BB R$ be given by $$ f(x)= \begin{cases} x & \text{ if } |x|\leq 1 \\ 2x & \text{ otherwise} \end{cases} \qquad g(x)= \begin{cases} 2x & \text{ if } |x|\leq 1 \\ x & \text{ otherwise} \end{cases} $$ Then $3=\mathsf{D}(f)(0,1+1)> \mathsf{D}(f)(0,1)+\mathsf{D}(f)(0,1)=2$ and $ 3= \mathsf{D}(g)(0,1+1) < \mathsf{D}(g)(0,1)+\mathsf{D}(g)(0,1)=4 $. \end{remark} The distance function on $Y^{X}$ in $\mathbf{Q}$ can be characterized using derivatives as follows: given QLR $(X,Q,a)$ and $(Y,R,b)$ and functions $f,g\in Y^{X}$, let $(2, \{0<\infty\}, d_{\mathsf{disc}})$ be the QLR given by the discrete metric on $2=\{0,1\}$. Let $\mathbf h_{f,g}: 2\times X\to Y$ be the function given by $\mathbf h_{f,g}(0,x)=f(x)$ and $\mathbf h_{f,g}(1,x)=g(x)$. A simple calculation yields then: \begin{lemma}\label{lemma:distanceder} $d^{\mathbf{Q}}_{a,b}(f,g)(x,\alpha)=\mathsf{D}(\mathbf h_{f,g})(\langle\langle 0,x\rangle, \langle \infty, \alpha\rangle\rangle)$. \end{lemma} \longv{\begin{proof} We have that \begin{align*} \mathsf{D} & (\mathbf h_{f,g}) (\langle\langle 0,x\rangle, \langle \infty, \alpha\rangle\rangle) \\ &= \sup\{ b(\mathbf h_{f,g}(\langle 0,x\rangle), \mathbf h_{f,g}(\langle i,y\rangle))\mid d_{\mathsf{disc}}(0,i)\leq \infty, a(x,y)\leq \alpha\}\\ &= \sup\{ b(f(x), f(y)), b(f(x),g(y))\mid a(x,y)\leq \alpha\}\\ &= d^{\mathbf{Q}}_{a,b}(f,g)(x,\alpha) \end{align*} \end{proof} } A consequence of Lemma~\ref{lemma:distanceder} is that the self-distance of $f\in Y^{X}$ coincides with its derivative, that is: \begin{align}\label{eq:law} d^{\mathbf{Q}}_{a,b}(f,f)=\mathsf{D}(f \end{align} Observe that this property implies that the self-distance of $f$ is (constantly) zero precisely when $f$ is a constant function. We now define a category $\mathbf{Q}^{\mathsf{r}}$ of \emph{reflexive} QLR: $\mathbf{Q}^{\mathsf{r}}$ is the full subcategory of $\mathbf{Q}$ made of QLR $(X,Q, a)$ such that $Q$ is Heyting and satisfies the property below: \begin{align}\label{eq:heyting} \text{if }\alpha \leq \beta \ \text{ then } \ \beta \leq \beta \Leftarrow \alpha \tag{$\star\star$} \end{align} and such that $a(x,x)=0$ holds for all $x\in X$. The Lawvere quantale satisfies property \eqref{eq:heyting}, and this property is stable by product. In particular, $\mathbf{Q}^{\mathsf{r}}$ inherits the cartesian product from $\mathbf{Q}$. Instead, the exponential of $(X,Q,a)$ and $(Y,R,b)$ in $\mathbf{Q}^{\mathsf{r}}$ is the QLR $(Y^{X}, R^{X\times Q}, d^{\mathbf{Q}^{\mathsf{r}}}_{a,b})$, where $$ d^{\mathbf{Q}^{\mathsf{r}}}_{a,b}(f,g):= d^{\mathbf{Q}}_{a,b}(f,g) \Leftarrow \mathsf{D}(f) $$ Observe that $d^{\mathbf{Q}^{\mathsf{r}}}_{a,b}(f,f)= \mathsf{D}(f)\Leftarrow\mathsf{D}(f)=0$. The isomorphism $\begin{tikzcd}\mathbf{Q}^{\mathsf{r}}(Z\times X, Y)\ar[bend left=5]{r}{\lambda^{\mathsf r}} & \mathbf{Q}^{\mathsf{r}}(Z,Y^{X}) \ar[bend left=5]{l}[below]{\mathsf{ev}^{\mathsf r}}\end{tikzcd}$ is given by \begin{align*} \lambda^{\mathsf{r}} (f,\varphi)& = (\lambda (f), \lambda(\varphi)\Leftarrow \lambda z. \mathsf{D}( f (\langle z,\_))) \\ \mathsf{ev}^{\mathsf{r}}(f,\varphi)&= (\mathsf{ev}(f), \mathsf{ev}(\varphi) \vee \lambda z. \mathsf{D}(f(z)(\_))) \end{align*} \begin{remark} In the absence of property \eqref{eq:heyting}, reflexive QLR only form a cartesian \emph{lax}-closed category \cite{Seely}. In particular, one has that $\mathsf{ev}^{\mathsf r}(\lambda^{\mathsf r}(f,\varphi))=\varphi$ and $\lambda^{\mathsf r}(\mathsf{ev}^{\mathsf r}(f,\psi))\leq \psi$ (in other words, $\beta$-reduction is preserved while $\eta$-reduction \emph{decreases} the interpretation). \end{remark} \begin{remark}\label{rem:distances} In $\mathbf{Q}$ and $\mathbf{Q}^{\mathsf{r}}$ we can define a ``na\"ive'' lifting of the Euclidean metric to all simple types built over the reals. This yields the two distance functions $d$ and $e$ on $\BB R^{\BB R}$ below: \medskip \noindent \adjustbox{scale=0.9}{ \begin{minipage}{\linewidth} \begin{align*}\label{eq:relaxed} d(f,g)(x,\alpha)& =\sup\{d_{\mathsf{Euc}}(f(x),f(y)), d_{\mathsf{Euc}}(f(x),g(y))\shortv{\\ &\qquad\qquad\qquad\qquad\qquad\qquad }\mid d_{\mathsf{Euc}}(x,y)\leq \alpha\} \\ e(f,g)(x,\alpha)&= \begin{cases} d(f,g)(x,\alpha) & \text{ if } d(f,g)(x,\alpha) > \mathsf{D}(f)(x,\alpha) \\ 0 & \text{ otherwise} \end{cases} \end{align*} \end{minipage} } \end{remark} One can also consider categories $\mathbf{Q}^{\mathsf{s}}, \mathbf{Q}^{\mathsf{rs}}$ of \emph{symmetric} (resp.~reflexive and symmetric) QLR. One has the following: \begin{lemma}\label{prop:symmetric0} Let $(X,Q,a)$, $(Y,R,b)$ be symmetric QLR. If $R$ is a locale, then their exponential QLR in $\mathbf{Q}$ is still symmetric. \end{lemma} \longv{ \begin{proof}[Proof of Lemma~\ref{prop:symmetric0}] If $R$ is a locale, then we have that for all $x,y\in X$, $\alpha\in Q$ with $a(x,y)\leq \alpha$, $b(g(x),f(y)) \leq b(g(x),f(x))\vee b(f(x),f(y))= b(f(x),f(y))\vee b(f(x),g(x))\leq d^{\mathbf{Q}}_{a,b}(f,g)(x,\alpha)$ and $b(g(x),g(y)) \leq b(g(x),f(x))\vee b(f(x),g(y))= b(f(x),g(x))\vee b(f(x),g(y))\leq d^{\mathbf{Q}}_{a,b}(f,g)(x,\alpha)$, since $b$ is symmetric. From this we deduce that $d^{\mathbf{Q}}_{a,b}(g,f)(x,\alpha)= \sup\{ b(g(x),g(y)), b(g(x), f(y))\mid a(x,y)\leq \alpha\}\leq d^{\mathbf{Q}}_{a,b}(f,g)(x,\alpha)$ and conversely. \end{proof} } As a consequence, the categories $\mathbf{Q}_{\land}^{\mathsf s}$ and $\mathbf{Q}^{\mathsf{rs}}_{\land}$ of symmetric (resp.~reflexive and symmetric) QLR $(X,Q,a)$ where $Q$ is a locale, {are} cartesian closed subcategories of $\mathbf{Q}, \mathbf{Q}^{\mathsf{r}}$, respectively. We will meet these two categories in the next section. The locale-valued symmetric QLR are essentially the only ones to inherit the cartesian closed structure of $\mathbf{Q}$ and $\mathbf{Q}^{\mathsf{r}}$, as shown be the lemma below\longv{ (which is proved in the next section)}. \begin{lemma}\label{prop:symmetric} Let $(X,Q,a)$, $(Y,R,b)$ be symmetric QLR, where $Y$ is \emph{injective} (\cite{Espinola:2001aa,CLEMENTINO20063113}\longv{, see also Section \ref{section5}}) and $X$ contains two points $v_{0},v_{1}$ with $a(v_{0},v_{1})\neq 0$. Then, if the exponential of $X$ and $Y$ in $\mathbf{Q}$ is symmetric, then for all $\alpha\in R$ such that $\alpha+\alpha\in Im(b)$, $\alpha=\alpha+\alpha$. \end{lemma} \subsection{QLR Models} We now describe the interpretation of the simply typed $\lambda$-calculus inside $\mathbf{Q}$ and $\mathbf{Q}^{\mathsf{r}}$. Concretely, this means associating each simple type with a QLR and each typed program with a morphism of QLR. We describe this situation abstractly through the notion of \emph{QLR-model}, introduced below. \begin{definition} Let $\BB C$ be a cartesian closed category. A \emph{$\mathbf{Q}$-model} (resp. \emph{$\mathbf{Q}^{\mathsf{r}}$-model}) of $\BB C$ is a cartesian closed functor $F: \BB C\to \mathbf{Q}$ (resp. $F:\BB C\to \mathbf{Q}^{\mathsf{r}}$). \end{definition} Concretely, a $\mathbf{Q}$-model consists in the following data: \begin{itemize} \item for any object $X$ of $\BB C$, a QLR $(\model X, \nudel X, a_{X})$; \item for any morphism $f\in \BB C(X,Y)$, functions $\model f: \model X\to \model Y$ and $\nudel f: \model X\times \nudel X\to \nudel Y$ such that $(\model f,\nudel f)$ is a QLR morphism from $\model X$ to $\model Y$, \end{itemize} where the application $f\mapsto \nudel f$ satisfies suitable equations resembling Eq.~\ref{eq:prop2}-\ref{eq:prop7} (however, with equality in place of $\leq$). Observe that $\nudel f$ is in general only an \emph{approximation} of the derivative $\mathsf{D}{(\model f)}$ (that is, one has $\mathsf{D}{(\model f)}\leq \nudel f$). We now describe a concrete $\mathbf{Q}$-model for a simply typed $\lambda$-calculus $\mathsf{ST\lambda C}(\mathcal F)$ over a type $\mathsf{Real}$ for real numbers. More precisely, simple types are defined by the grammar $$\sigma, \tau:=\mathsf{Real}\mid \sigma\to \tau\mid \sigma\times \tau$$ We fix a family $\C F=(\C F_{n})_{n>0}$ of sets of functions from $\BB R^{n}$ to $\BB R$. We consider the usual Curry-style simply-typed $\lambda$-calculus, with left and right projection $\pi_{1}$ and $\pi_{2}$, and with pair constructor $\langle \_,\_\rangle$, enriched with the following constants: for all $r\in \BB R$, a constant $\TT r:\mathsf{Real}$; for all $n>0$ and $f\in \C F_{n}$, a constant $\TT f: \mathsf{Real} \to \dots \to \mathsf{Real} \to \mathsf{Real}$. The usual relation of $\beta$-reduction is enriched with the following rule, extended to all contexts: for all $n > 0$, $f \in \C F_{n}$, and $ r_{1}, \dots, r_{n} \in \BB R$, $\TT f\TT r_{1} \dots \TT r_{n} \longrightarrow_{\beta}\TT s$, where $s = f(r_{1},...,r_{n})$. By standard arguments \cite{Krivine}, this calculus has the properties of subject reduction, confluence and strong normalization. We let $\Lambda(\C F)$ be the cartesian closed category whose objects are the simple types and where $\Lambda(\C F)(\sigma, \tau)$ is the quotient of the set of closed terms of type $\sigma\to \tau$ under $\beta\eta$-equivalence, and composition of $[\lambda x.t]\in \Lambda(\C F)(\sigma, \tau)$ and $[\lambda x.u]\in \Lambda(\C F)(\tau, \rho)$ is $[\lambda x.u(tx)]$. A $\mathbf{Q}$-model of $\mathsf{ST\lambda C}(\C F)$ is defined by setting $\model \mathsf{Real} = \BB R$, $\nudel \mathsf{Real} = \BB R^{\infty}_{\geq 0}$, $a_{\mathsf{Real}}=d_{\mathsf{Euc}} $ and extending the definition of the QLR $(\model\sigma, \nudel \sigma, a_{\sigma})$ to all simple types $\sigma$ using the cartesian closed structure of $\mathbf{Q}$. Moreover, given a context $\Gamma=\{x_{1}:\sigma_{1},\dots, x_{n}:\sigma_{n}\}$ and a term $t$ of type $\Gamma \vdash t:\sigma$ (that we take as representative of a class of terms of type $(\prod_{i=1}^{n}\sigma_{i})\to \sigma$), the functions $\model t: \prod_{i=1}^{n}\model{\sigma_{i}} \to \model \sigma$ and $\nudel t: \prod_{i=1}^{n}\model{\sigma_{i}}\times \prod_{i=1}^{n}\nudel{\sigma_{i}}\to \nudel \sigma$ are defined by a straightforward induction on $t$. We unroll below the definition of $\nudel t$: \begin{align*} \nudel{\TT r} (\vec x, \vec \alpha)& = 0 \\ \nudel{\TT f}(\vec x, \vec \alpha) & = \mathsf{D}(f)(\vec x, \vec \alpha) \\ \nudel{x_{i}}(\vec x, \vec \alpha) & = \alpha_{i}\\ \nudel{\langle t,u\rangle}(\vec x, \vec \alpha)& =\langle \nudel t(\vec x, \vec \alpha),\nudel u(\vec x, \vec \alpha)\rangle\\ \nudel{t\pi_{i}}(\vec x, \vec \alpha) & = \pi_{i}(\nudel t(\vec x,\vec \alpha)) \\ \nudel{\lambda y.t}(\vec x, \vec \alpha) & = \lambda y\alpha. \nudel t(\vec x*y, \vec \alpha*\alpha ) \\ \nudel{tu}(\vec x, \vec \alpha) & = {\nudel t}(\vec x, \vec \alpha )(\model u(\vec x),\nudel u(\vec x,\vec \alpha)) \end{align*} where $\vec x*y$ indicates the concatenation of $\vec x$ with $y$. \begin{theorem}[Soundness]\label{thm:stlc} For all simply typed terms $t$ such that $\Gamma \vdash t:\sigma$, $(\model t, \nudel t)\in \mathbf{Q}(\model \Gamma, \model \sigma)$. Moreover, if $t\longrightarrow_{\beta}u$, then $\model t=\model u$ and $\nudel t=\nudel u$. \end{theorem} The following fact is an immediate consequence of Theorem \ref{thm:stlc} and Eq.~\eqref{eq:law}, and can be seen as a quantitative analog of the \emph{Fundamental Lemma} of logical relations, stating that any program $t$ is related to itself by $\nudel t$: \begin{corollary}[Fundamental Lemma for QLR]\label{lemma:fundamental} For all terms $t$ such that $\vdash t:\sigma $, $ a_{\sigma}(\model t, \model t)\leq \nudel t $. \end{corollary} Another quite literal consequence of Theorem \ref{thm:stlc} is that program distances are \emph{contextual}: given a distance between programs $t$ and $u$, for any context $\TT C[\_]$ we can obtain a distance between $\TT C[t]$ and $\TT C[u]$: \begin{corollary}[contextuality of distances]\label{cor:context} For all terms $t,u$ such that $\vdash t,u:\sigma$ holds and for all context $\TT C[\ ]: \sigma \vdash \tau$, $$ a_{\tau}(\model{\TT C[t]}, \model{\TT C[u]}) \leq \nudel{\TT C}( \model t, a_{\sigma}(\model t, \model u)) $$ \end{corollary} In a similar way one can define a $\mathbf{Q}^{\mathsf{r}}$-model of $\mathsf{ST\lambda C}(\C F)$ and prove analogs of the results above (where Corollary \ref{lemma:fundamental} now reads as $a_{\sigma}(\model t, \model t)=0$). \begin{remark} Corollaries \ref{lemma:fundamental} and \ref{cor:context} generalize properties established in the setting of differential logical relations (cf.~Lemma 15 in \cite{dallago}). \end{remark} \begin{remark} One can define an alternative interpretation of $\mathsf{ST\lambda C}$ by letting $\nudel t$ be the ``true'' derivative $\mathsf{D}(\model t)$. However, while Corollaries \ref{lemma:fundamental} and \ref{cor:context} still hold, the operation $t\mapsto (\model t, \mathsf{D}(\model t))$ only yields a \emph{colax} functor (since one only has $\mathsf{D}(\model{u}\circ \model t)\leq \mathsf{D}(\model u)(\model t, \mathsf{D}(\model t))$). \end{remark} \section{Metrizability}\label{section5} In this section we investigate generalized metrics in sub-categories of $\mathbf{Q}$ and $\mathbf{Q}^{\mathsf{r}}$. We first show that the relaxed and hyper-relaxed metrics all form cartesian closed subcategories of $\mathbf{Q}$; we then turn to metrics and partial metrics: we show that, under suitable assumptions, the exponential QLR formed from two metric or partial metric spaces $X$ and $Y$ is a metric or a partial metric space precisely when the metric of $Y$ is idempotent (i.e.~distances satisfy $\alpha=\alpha+\alpha$). This result can be used to show that ultra-metrics and partial ultra-metrics form cartesian closed subcategories of $\mathbf{Q}$ and $\mathbf{Q}^{\mathsf{r}}$; at the same time it shows that the na\"ive lifting of the Euclidean metric (as well as of any non-idempotent metric) in either $\mathbf{Q}$ or $\mathbf{Q}^{\mathsf{r}}$ is \emph{not} a generalized metric. Nevertheless, we show that liftings to all simple types can be defined for those metrics and partial metrics (including the Euclidean metric), whose distance function factors as the composition of an idempotent metric and a \emph{valuation} \cite{ONeill, 10.1016/j.tcs.2003.11.016}. \subsection{Relaxed metrics} It is not difficult to check that whenever $(X,Q,a)$ and $(Y,R,b)$ are two relaxed or hyper-relaxed metrics, so is their exponential in $\mathbf{Q}$. For the relaxed metrics, given $f,g,h\in Y^{X}$, using the triangular law of $Y$ we deduce that for all $x,y\in X$ and $\alpha\geq a(x,y)$, \begin{align*} b(f(x), g(y)) & \leq b(f(x), h(x))+b(h(x),g(y)) \\ &\leq d^{\mathbf{Q}}_{a,b}(f,h)(x, \alpha)+d^{\mathbf{Q}}_{a,b}(h,g)(x,\alpha) \end{align*} and thus that $d^{\mathbf{Q}}_{a,b}(f,g)\leq d^{\mathbf{Q}}_{a,b}(f,h)+d^{\mathbf{Q}}_{a,b}(h,g)$. This argument straightforwardly scales to the hyper-relaxed metrics, yielding: \begin{proposition}\label{prop:relaxed} The full subcategories of $\mathbf{Q}$ made of relaxed and hyper-relaxed metrics are cartesian closed. \end{proposition} An immediate consequence is that the distance $d$ from Remark \ref{rem:distances} is a relaxed metric. We will show below that we cannot actually say \emph{more} of $d$: it is not a partial metric. \subsection{Ultra-metrics} For all metric spaces $(X,Q,a)$ and $(Y,R,b)$, whenever $R$ satisfies $\alpha+\beta=\alpha\vee \beta$ (or, equivalently, $\alpha=\alpha+\alpha$ and $0=\bot$), it is not difficult to check that the transitivity axiom lifts to the exponential in $\mathbf{Q}$: in fact, for all $f,g,h\in Y^{X}$ and $x,y\in X$ with $a(x,y)\leq \alpha$ one has \begin{align*} b(f(x),g(y)) & \leq b(f(x),h(x)) \vee b(h(x),g(y)) \\ & \leq d^{\mathbf{Q}}_{a,b}(f,h)(x,\alpha) \vee d^{\mathbf{Q}}_{a,b}(h,g)(x,\alpha) \end{align*} from which we deduce $d^{\mathbf{Q}}_{a,b}(f,g)(x,\alpha)\leq d^{\mathbf{Q}}_{a,b}(f,h)(x,\alpha) \vee d^{\mathbf{Q}}_{a,b}(h,g)(x,\alpha)$. A similar argument can be developed for the distance $d^{\mathbf{Q}^{\mathsf{r}}}_{a,b}$, leading to: \begin{proposition}\label{prop:ultra} The full subcategories of $ \mathbf{Q}^{\mathsf{rs}}_{\land}$ and $\mathbf{Q}_{\land}^{\mathsf s}$ made of ultra-metric spaces and partial ultra-metric spaces are cartesian closed. \end{proposition} \longv{\begin{proof} Let $(X,Q,a), (Y,R,b)$ be objects of $\mathbf{Q}^{\mathsf{rs}}_{\land}$. It suffices to show that the QLR $Y^{X}$ satisfies transitivity. Since $R$ is a locale, $\alpha+ \beta=\alpha\vee \beta$ holds for all $\alpha,\beta\in R$. Let $f,g,h\in Y^{X}$. Then we have that $\mathsf{D}(f)\vee (d^{\mathbf{Q}}_{a,b}(f,h)+d^{\mathbf{Q}}_{a,b}(h,g))=(\mathsf{D}(f)\vee d^{\mathbf{Q}}_{a,b}(f,h))\vee d^{\mathbf{Q}}_{a,b}(h,g)$, so in particular for all $x,y\in X$ and $\alpha\geq a(x,y)$, $(\mathsf{D}(f)\vee (d^{\mathbf{Q}}_{a,b}(f,g)+d^{\mathbf{Q}}_{a,b}(h,g)))(x,\alpha) = ((\mathsf{D}(f)\vee d^{\mathbf{Q}}_{a,b}(f,g))(x,\alpha)) \vee((\mathsf{D}(f)\vee d^{\mathbf{Q}}_{a,b}(h,g))(x,\alpha))\geq d^{\mathbf{Q}}_{a,b}(f(x), h(x))\vee d^{\mathbf{Q}}_{a,b}(h(x),g(y)) \geq d^{\mathbf{Q}}_{a,b}(f(x),g(y))$, from which we deduce that $(d^{\mathbf{Q}}_{a,b}(f,g)+d^{\mathbf{Q}}_{a,b}(h,g))(x,\alpha) \Leftarrow \mathsf{D}(f)(x,\alpha)\geq( d^{\mathbf{Q}}_{a,b}(f,g)(x,\alpha))\Leftarrow (\mathsf{D}(f) (x,\alpha))=d^{\mathbf{Q}^{\mathsf{r}}}_{a,b}(f,g)(x,\alpha)$. A similar argument can be developed for $\mathbf{Q}_{\land}^{\mathsf s}$, using the fact that in a locale $\alpha+_{\gamma}\beta=\alpha\vee \beta$. \end{proof} } When $Q$ is a locale, also the category $\mathsf{Met}_{Q}$ is cartesian closed \cite{Smyth}. These categories have been mostly used to account for \emph{intensional} properties of higher-order programs (e.g.~measuring program approximations or the number of computation steps \cite{Escardo1999}). In the categories $\mathbf{Q}^{\mathsf{rs}}_{\land}$ and $\mathbf{Q}_{\land}^{\mathsf s}$ we can define metrics describing more \emph{extensional} properties (i.e.~measuring distances between program outputs) as the one below. \begin{example}\label{ex:intervals} Let $\C I(\BB R)$ be the complete lattice of \emph{closed intervals} $[x,y]$ (where $x,y\in \BB R$ and $x\leq y$), enriched with $\emptyset$ and $\BB R$. We can define a partial ultra-metric $u: \BB R\times \BB R\to \C I(\BB R)$ by letting $u(x,y)=[\min\{x,y\}, \max\{x,y\}]$. The metric $u$ lifts in $\mathbf{Q}_{\land}^{\mathsf s}$ to a partial ultra-metric $d^{\mathbf{Q}}_{u,u}$ over real-valued functions where, for all $x\in \BB R$ and $I\in \C I(\BB R)$, $d^{\mathbf{Q}}_{u,u}(f,g)(x,I)$ is the smallest interval containing all $f(y)$ and $g(y)$, for $y\in I\vee\{x\}$ (see also \cite{Geoffroy2020}). \end{example} We now establish a sort of converse to Proposition \ref{prop:ultra}: under suitable conditions, if the exponential of two metric spaces $X$ and $Y$ satisfies the transitivity axiom, then the distances over $Y$ are idempotent\shortv{:}\longv{. Let us first recall the notion of \emph{injective} metric space \cite{Espinola:2001aa,CLEMENTINO20063113}, that will be essential in our argument. A map $f:X\to X$ between two metric spaces $(X,Q,a), (Y,Q,b)$ over the \emph{same} quantale is said an \emph{extension} if for all $x,y\in X$, $b(f(x),f(y))=a(x,y)$, and is said \emph{non-expansive} if for all $x,y\in X$, $b(f(x),f(y))\leq a(x,y)$. A metric space $(X,Q,a)$ is \emph{injective} when for all non-expansive map $f: Y\to X$ and extension $e:Y\to Z$ there exists a non-expansive map $h:Z\to X$ such that $f=h\circ e$. Injective metric spaces (also known as \emph{hyperconvex} metric spaces) enjoy several nice properties (see \cite{Espinola:2001aa}). In particular, they form a cartesian closed subcategory of $\mathsf{Met}$ \cite{CLEMENTINO20063113}, which includes the Euclidean metric. Here we will use such spaces to establish a few negative results.} \begin{lemma}\label{lemma:trivialmetric} \begin{itemize} \item[i.]Let $(X,Q,a)$ and $(Y,R,b)$ be two metric spaces, where $X$ has at least two distinct points and $Y$ is \emph{injective}\shortv{ (\cite{Espinola:2001aa,CLEMENTINO20063113})}. If the reflexive QLR $(Y^{X}, R^{X\times Q}, d^{\mathbf{Q}^{\mathsf{r}}}_{a,b})$ is a metric space then for all $\alpha,\beta\in R$ such that $\alpha+\beta\in Im(b)$, $\alpha+\beta=\alpha\vee \beta$. \item[ii.] Let $(X,Q,a)$ and $(Y,R,b)$ be two partial metric spaces, where $X$ has at least two distinct points and $Y$ is injective. If the QLR $(Y^{X}, R^{X\times Q}, d^{\mathbf{Q}}_{a,b})$ is a partial metric space then for all $\alpha,\beta\in R$ such that $\alpha+\beta\in Im(b)$, $\alpha+\beta=\alpha\vee \beta$. \end{itemize} \end{lemma} \longv{ \begin{proof} \begin{itemize} \item[i.] Let $\alpha,\beta\in R$ and $u_{0},u_{2}\in Y$ be such that $b(u_{0},u_{2})=\alpha+\beta$. Let $Y'=Y\cup \{v_{1}\}$ and $b'$ be as $b$ on $Y$ and satisfying $b(u_{0},v_{1})=\alpha$, $b(v_{1},u_{2})=\beta$. The injection $\iota:Y\to Y'$ is an expansion, hence, since $Y$ is injective, there exists a non-expansive function $f: Y'\to Y$ such that $f\circ \iota= \mathrm{id}_{Y}$. This implies in particular that, by letting $u_{1}:= f(v_{1})$, $b(u_{0},u_{1})\leq\alpha$, $b(u_{1},u_{2})\leq \beta$. Let now $x_{0},x_{1}$ be two distinct points in $X$ and let $f,g,h:X\to Y$ be the following functions: $f(x)$ is constantly $u_{0}$ except for $f(x_{1})=u_{1}$; $g(x)$ is constantly $u_{2}$ and $h(x)$ is constantly $u_{1}$. We have then that $\mathsf{D}(f)(x,a(x_{0},x_{1}))\leq \alpha$, $\mathsf{D}(g)=\mathsf{D}(h)=0$. Moreover, for all $x'\in X$ with $a(x_{0},x')\leq a(x_{0},x_{1})$, $b(f(x_{0}),f(x')), b(f(x_{0}),h(x'))\leq b(u_{0},u_{1})\leq \mathsf{D}(f)(x,a(x_{0},x_{1}))=\mathsf{D}(f)(x_{0},a(x_{0},x_{1})) \vee 0 $, that is $d^{\mathbf{Q}}_{a,b}(f,h)(x_{0},a(x_{0},x_{1}))\leq \mathsf{D}(f)(x,a(x_{0},x_{1}))$, and thus \\ $d^{\mathbf{Q}^{\mathsf{r}}}_{a,b}(f,h)(x,a(x_{0},x_{1}))=d^{\mathbf{Q}}_{a,b}(f,h)\Leftarrow \mathsf{D}(f))(x_{0},a(x_{0},x_{1}))\leq 0$ Then, since by hypothesis $e_{a,b}$ is a metric, we deduce that \begin{align*} \alpha+\beta& = b(u_{0},u_{2})=b(f(x_{0}),g(x_{1})) \\ & \leq d^{\mathbf{Q}}_{a,b}(f,g)(x_{0},a(x_{0},x_{1})) \\ & \leq \big (\mathsf{D}(f)\vee d^{\mathbf{Q}^{\mathsf{r}}}_{a,b}(f,g)\big )(x_{0},a(x_{0},x_{1})) \\ & \leq\big (\mathsf{D}(f)\vee(d^{\mathbf{Q}^{\mathsf{r}}}_{a,b}(f,h)+d^{\mathbf{Q}^{\mathsf{r}}}_{a,b}(h,g))\big )(x_{0},a(x_{0},x_{1}))\\ &\leq \alpha \vee (0+\beta)= \alpha\vee \beta \end{align*} \item[ii.] As in the proof of point i.~let $\alpha,\beta\in R$ and $u_{0},u_{1},u_{2}\in Y$ be such that $b(u_{0},u_{1})\leq \alpha$, $b(u_{1},u_{2})\leq \beta$ and $b(u_{0},u_{2})=\alpha+\beta$. We can suppose w.l.o.g. that $b$ is symmetric. Let now $x_{0},x_{1}$ be two distinct points in $X$ and let $f,g,h:X\to Y$ be the following functions: $f(x)$ is constantly $u_{0}$, $h(x)$ is constantly $u_{1}$ except for $h(x_{1})=u_{0}$ and $g(x)$ is constantly $u_{1}$ except for $g(x_{1})=u_{2}$. Then we have that $d^{\mathbf{Q}}_{a,b}(f,g)(x_{0},a(x_{0},x_{1}))=b(u_{0},u_{2})=\alpha+\beta$, $d^{\mathbf{Q}}_{a,b}(f,h)(x_{0},a(x_{0},x_{1}))=d^{\mathbf{Q}}_{a,b}(h,h)(x_{0},a(x_{0},x_{1}))=b(u_{0},u_{1})\leq \alpha$ and $d^{\mathbf{Q}}_{a,b}(h,g)(x_{0},a(x_{0},x_{1}))= b(u_{0},u_{1})\vee b(u_{1},u_{2})\leq \alpha \vee \beta$. Then, since by hypothesis $d^{\mathbf{Q}}_{a,b}$ is a partial metric, we deduce that \noindent \adjustbox{scale=0.95}{ \begin{minipage}{\linewidth} \begin{align*} \alpha+\beta& = b(u_{0},u_{2})=b(f(x_{0}),g(x_{1})) \\ & \leq d^{\mathbf{Q}}_{a,b}(f,g)(x_{0},a(x_{0},x_{1}))\\ & \leq \big ((d^{\mathbf{Q}}_{a,b}(f,h)\multimapinv d^{\mathbf{Q}}_{a,b}(h,h))+d^{\mathbf{Q}}_{a,b}(h,g)\big )(x_{0},a(x_{0},x_{1})) \\ & = \big ((d^{\mathbf{Q}}_{a,b}(h,h)\multimapinv d^{\mathbf{Q}}_{a,b}(h,h))+d^{\mathbf{Q}}_{a,b}(h,g)\big )(x_{0},a(x_{0},x_{1})) \\ & = d^{\mathbf{Q}}_{a,b}(h,g)(x_{0},a(x_{0},x_{1})) \leq \alpha \vee \beta \end{align*} \end{minipage} } \end{itemize} \end{proof} } To give the reader an \shortv{idea of the proof}\longv{illustration} of Lemma \ref{lemma:trivialmetric}, we \shortv{illustrate}\longv{show} in Fig.~\ref{fig:counterexamples} counter-examples to transitivity for the na\"ive extensions of the Euclidean metric (cf.~Remark \ref{rem:distances}). \longv{ Along similar lines we can also prove Lemma \ref{prop:symmetric} from the previous section. \begin{proof}[Proof of Lemma~\ref{prop:symmetric}] Let $\alpha\in R$ and $x_{1},x_{2}\in Y$ be such that $b(x_{1},x_{2})=\alpha+\alpha$. Let $(Z,R,c)$ be a metric space where $Z=X\cup\{u_{0},u_{3}\}$ and $c$ is defined so that $c(u_{0},u_{0})=c(u_{3},u_{3})=0$ and the following hold: \begin{align*} c(u_{0},u_{1}), c(u_{0},u_{2}), c(u_{0},u_{3})& =\alpha \\ c(u_{1},u_{2}), c(u_{2},u_{3}), c(u_{3},u_{1}) & = \alpha+\alpha \end{align*} Since $Y$ is injective, there exists a non-expansive map $f:Y\to X$ such that $f\circ \iota=\mathrm{id}_{X}$, where $\iota$ is the injection $\iota: X\to Z$ (which is obviously an expansion). Hence there exist points $x_{0},x_{3}\in X$ such that $ b(x_{0},x_{1}), b(x_{0},x_{2}), b(x_{0},x_{3}) =\alpha$ and $b(x_{1},x_{2}), b(x_{2},x_{3}), b(x_{3},x_{1}) \leq \alpha+\alpha$. Let $f,g\in Y^{X}$ be defined by \begin{align*} f(w)= \begin{cases} x_{1} & \text{ if } w=v_{0} \\ x_{2} & \text{ otherwise} \end{cases} \qquad g(w)= \begin{cases} x_{3} & \text{ if } w=v_{0} \\ x_{4} & \text{ otherwise} \end{cases} \end{align*} where $v_{0},v_{1}$ are two distinct points of $X$ such that $a(v_{0},v_{1})\neq 0$. If $d^{\mathbf{Q}}_{a,b}(f,g)=d^{\mathbf{Q}}_{a,b}(g,f)$, we deduce that \begin{align*} \alpha & \geq \sup\{ b(f(v_{0}),f(w)), b(f(v_{0}),g(w))\mid a(v_{0},w)\leq a(v_{0},v_{1})\}\\ & = d^{\mathbf{Q}}_{a,b}(f,g)(v_{0}, a(v_{0},v_{1}))\\ & =d^{\mathbf{Q}}_{a,b}(g,f)(v_{0},a(v_{0},v_{1})) \\ & = \sup\{ b(g(v_{0}),g(w)),b(g(v_{0}),f(w))\mid a(v_{0},w)\leq a(v_{0},v_{1})\}\\ & = \alpha+\alpha \end{align*} \end{proof} } \begin{figure*} \fbox{ \begin{subfigure}{0.47\textwidth} \begin{center} \begin{tikzpicture}[domain=-2:2] \draw[dashed, |<->|] (-2,0) -- (2,0); \draw[<->] (0,-0.4) -- (0,3); \node at (0,0)[circle,fill,inner sep=1pt]{}; \node(z) at (0.2,-0.2) {$x$}; \node(-e) at (-2,-0.2) {$x-r$}; \node(e) at (2,-0.2) {$x+r$}; \node(f) at (0,2.6)[circle,fill,inner sep=1pt]{}; \node(f) at (0,1.45)[circle,fill,inner sep=1pt]{}; \node(f) at (0,0.3)[circle,fill,inner sep=1pt]{}; \node(f) at (-0.2,2.4) {\tiny$g(x)$}; \node(f) at (-0.2,1.6) {\tiny$h(x)$}; \node(f) at (-0.2,0.5) {\tiny $f(x)$}; \node(ff) at (1.3,2.3) {{$g$}}; \node(gg) at (1.2,1.1) {{$h$}}; \node(gg) at (1,0.45) {{$f$}}; \draw[color=orange] plot (\x,{2.25+0.8*(-cos(\x r))}) ; \draw[color=red] plot (\x,{0.65+0.8*(cos(\x r))}) ; \draw[color=blue] plot (\x,{0.3}) ; \draw[dotted] (0,0.3) -- (2.4,0.3); \draw[dotted] (0,1.45) -- (2.4,1.45); \draw[dotted] (0,2.6) -- (2.4,2.6); \draw[dashed,|<->| ] (2.4,2.6) -- node[right] {\tiny$d(h,g)$} (2.4,1.45); \draw[dashed,|<->| ] (2.4,1.45) -- node[right] {\tiny$d(f,h)=d(h,h)$} (2.4,0.3); \draw[dashed,|<->| ] (2.8,2.6) -- node[right] {\tiny$d(f,g)$} (2.8,0.3); \end{tikzpicture} \end{center} \caption{\small The distance $d$ from Remark \ref{rem:distances} is not a partial metric. The example above shows that $d(f,g)> d(f,h)+d(h,g)- d(h,h)$ (with all distances computed in $(x,r)$). A similar example can be found in~\cite{Geoffroy2020}. \\ \ \longv{\\ \ } } \end{subfigure} \ \ \ \begin{subfigure}{0.47\textwidth} \begin{center} \begin{tikzpicture}[domain=-2:2] \draw[dashed, |<->|] (-2,0) -- (2,0); \draw[<->] (0,-0.4) -- (0,3); \node at (0,0)[circle,fill,inner sep=1pt]{}; \node(z) at (0.2,-0.2) {$x$}; \node(-e) at (-2,-0.2) {$x-r$}; \node(e) at (2,-0.2) {$x+r$}; \node(f) at (0,2.6)[circle,fill,inner sep=1pt]{}; \node(f) at (0,1.45)[circle,fill,inner sep=1pt]{}; \node(f) at (0,0.3)[circle,fill,inner sep=1pt]{}; \node(f) at (-0.2,2.4) {\tiny$g(x)$}; \node(f) at (-0.2,1.6) {\tiny$h(x)$}; \node(f) at (-0.2,0.5) {\tiny $f(x)$}; \node(ff) at (1.3,2.3) {{$g$}}; \node(gg) at (1.2,1.1) {{$h$}}; \node(gg) at (1,0.45) {{$f$}}; \draw[color=blue] plot (\x,{1.1+0.8*(-cos(\x r))}) ; \draw[color=orange] plot (\x,{1.8+0.8*(cos(\x r))}) ; \draw[color=red] plot (\x,{1.45}) ; \draw[dotted] (0,0.3) -- (2.4,0.3); \draw[dotted] (0,1.45) -- (2.4,1.45); \draw[dotted] (0,2.6) -- (2.4,2.6); \draw[dashed,|<->| ] (2.4,2.6) -- node[right] {\tiny$d(h,g)$} (2.4,1.45); \draw[dashed,|<->| ] (2.4,1.45) -- node[right] {\tiny$d(f,f)=d(f,h)$} (2.4,0.3); \draw[dashed,|<->| ] (2.8,2.6) -- node[right] {\tiny$d(f,g)$} (2.8,0.3); \end{tikzpicture} \end{center} \caption{\small The distance $e$ from Remark \ref{rem:distances} is not a metric. In the example above (with all values computed in $(x,r)$), $e(f,h)=0$, since each $h(y)$ is no farther from $f(x)$ than $f(x+r)$, $e(h,g)=d(h,g)$ and $e(f,g)$ is $d(f,f)+d(f,g)$. Hence transitivity fails since $e(f,g)= d(f,h)+d(h,g) > 0+d(h,g)= e(f,h)+e(h,g)$.} \end{subfigure} } \caption{The distances $d$ and $e$ from Remark \ref{rem:distances} do not satisfy the transitivity axioms of metric and partial metric spaces.} \label{fig:counterexamples} \end{figure*} \subsection{Decomposing Partial Metrics through Valuations} Lemma~\ref{lemma:trivialmetric} suggests that one cannot hope to lift the Euclidean metric to all simple types inside $\mathbf{Q}$ or $\mathbf{Q}^{\mathsf{r}}$. Nevertheless, we will show that the Euclidean metric, as well as many other non-idempotent metrics and partial metrics, can be lifted to all simple types inside the categories $\mathbf{Q}_{\land}^{\mathsf s}$ and $\mathbf{Q}^{\mathsf{rs}}_{\land}$, by exploiting a well-investigated connection between partial metrics and lattice-valued metrics. A basic intuition comes from the observation that the Euclidean distance can be \emph{decomposed} as $$ \begin{tikzcd} \BB R\times \BB R \ar{r}{u} & \C I(\BB R) \ar{r}{\mu} & \BB R^{+}_{\geq 0} \end{tikzcd} $$ where $u$ is the partial ultra-metric from Example \ref{ex:intervals} and $\mu$ is the Lebesgue measure. This observation can be generalized using the theory of \emph{valuations} \cite{ONeill, 10.1007/BFb0053546,10.1016/j.tcs.2003.11.016}. A \emph{join-valuation} \cite{10.1016/j.tcs.2003.11.016} on a join semi-lattice $L$ is a monotone function $\mathcal F: L\to \BB R^{+\infty}_{\geq 0}$ which satisfies the condition \begin{equation}\label{eq:submodular} \C F( a \lor b) \leq \C F(a) + \C F(b) -\C F(a\land b) \end{equation} for all $a,b$ such that $a\land b$ exists in $L$. When $L$ is a $\sigma$-algebra, join-valuations on $L$ are thus sort of relaxed measures on $L$. Any join-valuation $\C F:L\to \BB R^{+\infty}_{\geq 0}$ induces a join semi-lattice $L_{\C F}$ obtained by quotienting $L$ under the equivalence $$ a \simeq_{\C F}b \ \text{iff} \ (a\leq b \text{ or } b\leq a) \text{ and }\C F(a)=\C F(b) $$ One can obtain then a separated and symmetric partial metric $p_{\C F}:L_{\C F}\times L_{\C F}\to \BB R^{+\infty}_{\geq 0}$ by letting $p_{\C F}(a,b)=\C F(a\vee b)$. The transitivity axiom is checked as follows: \begin{align*} \C F(a \lor b) & \leq\C F((a\lor c)\lor (c\lor b)) \\ & \leq\C F(a\lor c)+ \C F(c\lor b)- \C F((a\lor c) \land (c\lor b)) \\ & \leq \C F(a\lor c)+ \C F(c\lor b)- \C F(c\lor c) \end{align*} \begin{remark} The connection between partial metrics and valuations has a converse side \cite{10.1016/j.tcs.2003.11.016}: any (symmetric and separated) partial metric $p:X\times X\to \BB R^{+\infty}_{\geq 0}$ defines an order $\sqsubseteq_{p}$ over $X$ given by $x\sqsubseteq_{p}y $ iff $p(x,y)\leq p(x,x)$. Then, whenever the poset $(X, \sqsubseteq_{p})$ is a join semi-lattice, the self-distance function $ X \stackrel{\Delta}{\to} X\times X \stackrel{p}{\to} \BB R^{+\infty}_{\geq 0}$ is a join-valuation. \end{remark} Extending this observation to arbitrary (commutative and integral) quantales leads to the following: \begin{definition} A \emph{(generalized) valuation space} (noted $\COV{L}{\C F}{Q}$) is the given of a monotone function from a complete lattice $L$ to a quantale $Q$ satisfying \begin{equation} \C F(a\lor b) \leq \C F(a) + (\C F(b) \multimapinv\C F(a\land b)) \end{equation} for all $a,b\in L$ such that $a\land b\neq \bot$. \end{definition} By arguing as above, any valuation space $\COV{L}{\C F}{Q}$ yields a (symmetric and separated) partial metric $\C F: L_{\C F}\times L_{\C F}\to Q$. This leads to the following definition: \begin{definition} A \emph{partial metric valuation space} is a triple $(X, \COV{L}{\C F}{Q}, a)$, where $ \COV{L}{\C F}{Q}$ is a valuation space and $UX=(X, L_{\C F},a)$ is a (symmetric and separated) partial ultra-metric space. A map of partial metric valuation spaces $(X, \COV{L}{\C F}{Q},a)$ and $(Y,\COV{M}{\C G}{R},b)$ is an arrow $(f,\varphi)$ in $\mathbf{Q}_{\land}^{\mathsf s}( UX, UY)$. \end{definition} Observe that any partial metric valuation space $(X, \COV{L}{\C F}{Q}, a)$ yields \emph{both} a partial ultra-metric $a:X\times X\to L_{\C F}$ and a (separated) partial metric $\C F\circ a:X\times X\to Q$. \begin{example} The Euclidean metric can be presented as a partial metric valuation space in two ways: either using the Lebesgue measure as shown before, or by considering the valuation space $\COV{\C I(\BB R)^{-}}{\mathsf{diam}}{\BB R^{+\infty}_{\geq 0}}$ where $\C I(\BB R)^{-}$ is the join-semilattice $\C I(\BB R)-\{\emptyset\}$ and $\mathsf{diam}$ is the diameter function (which is in fact \emph{modular} over intersecting intervals, see \cite{Geoffroy2020}). \end{example} Observe that for any map $(f,\varphi)$ of spaces $(X, \COV{L}{\C F}{Q},a)$ and $(Y, \COV{M}{\C G}{R},b)$, we have that for all $x,y\in X$ and $\alpha\in L$, \begin{align*} \C G(b(f(x),f(y))\leq \C G(\varphi(x,\alpha)) \end{align*} In other words, the composition of derivatives and valuations provides a compositional way to compute distance bounds. We let $\mathsf p\mathbf{V}$ indicate the category of partial metric valuation spaces. Since the functor $U: \mathsf p\mathbf{V}\to \mathbf{Q}_{\land}^{\mathsf s}$ is by definition full and faithful, $\mathsf p\mathbf{V}$ inherits the cartesian closed structure from $\mathbf{Q}_{\land}^{\mathsf s}$. In particular, given partial metric valuation spaces $(X, \COV{L}{\C F}{Q},a)$ and $(Y, \COV{M}{\C G}{R},b)$, their product and exponential are as follows: \begin{center} $ (X\times Y, \begin{tikzcd}L\times M\ar{r}{\C F\times\C G} & R\times Q\end{tikzcd}, a\times b) $\\ $ (Y^{X}, \begin{tikzcd} (R_{\C G})^{X\times L_{\C F}}\ar{r}{ \C G\circ \_ }& Q^{X\times L_{\C F}}\end{tikzcd} , d_{a,b}) $ \end{center} \begin{example}\label{ex:pmet} The exponential of the Euclidean metric in $\mathsf p\mathbf{V}$ is the partial metric $p: (\BB R^{\BB R}\times \BB R^{\BB R}) \to (\BB R^{+}_{\geq 0})^{\BB R\times \C I(\BB R)}$ given by $$ p(f,g)(x, I)=\mathsf{diam}\{ b(f(y), g(z))\mid y,z\in \{x\}\vee I\} $$ We can compare $p$ with the na\"ive lifting $d$ in Fig.~\ref{fig:counterexamples}, by considering the interval $I=[x-r,x+r]$. One has $p(f,h)(x,I)=d(f,h)$ but $p(h,g)(x,I)= d(h,h)+d(h,g)$. Hence transitivity holds for $p$, since $p(f,g)(x,I)= p(f,h)(x,I)+ p(h,g)(x,I)-p(h,h)(x,I)$. \end{example} This construction can be adapted to metric spaces. Let a \emph{dual join-valuation} be a monotone map $L^{\mathsf{op}}\times L \stackrel{\C D}{\to} Q$ (where $L^{\mathsf{op}}$ is the complete lattice with the reversed order) satisfying \begin{align*} \C D(a,a) =0\qquad \qquad \C D(a,b\vee c) \leq \C D(a, b)+ \C D(b\land c, c) \end{align*} One defines the quotient $L_{\C D}$ by $a\simeq_{\C D}b$ iff $a\leq b$ or $b\leq a$ and $\C D(a,a\vee b)=\C D(b,b\vee a)=0$. For any dual join valuation $\C D$, the function $d_{\C D}:L_{\C D}\times L_{\C D}\to Q$ given by $d(a,b)=\C D(a,a\vee b)+\C D(b,b\vee a)$ is a symmetric and separated metric. Moreover, any join-valuation $L\stackrel{\C F}{\to} Q$ yields the dual join valuation $\C F'(a,b)=\C F(b)\multimapinv\C F(a)$. Let a \emph{metric valuation space} be a triple $(X, L^{\mathsf{op}}\times L \stackrel{\C D}{\to} Q, a)$, where $L^{\mathsf{op}}\times L \stackrel{\C D}{\to} Q$ is a dual join valuation and $UX=(X,L_{\C D}, a)$ is a symmetric and separated ultra-metric space. One obtains then a category $\mathbf{V}$ of metric valuation spaces, with $\mathbf{V}(X,Y)=\mathbf{Q}^{\mathsf{rs}}_{\land}(UX,UY)$. \begin{theorem} The categories $\mathsf p\mathbf{V}$ and $\mathbf{V}$ are cartesian closed. \end{theorem} \begin{example} The Euclidean metric lives in $\mathbf{V}$ as it arises from the dual join valuation $\C D: \C I(\BB R)^{\mathsf{op}}\times \C I(\BB R)\to \BB R^{\infty}_{\geq 0}$ given by $\C D(I,J)=\mathsf{diam}(J)\multimapinv \mathsf{diam}(I)$. Its lifting to $\BB R^{\BB R}$ inside $\mathbf{V}$ yields the metric $m(f,g)=2p(f,g)-p(f,f)-p(g,g)$, where $p$ is the partial metric from Example \ref{ex:pmet}. \end{example} \input{Lipschitz} \section{Related Works} Logical relations \cite{Plotkin1973, STATMAN198585} are a standard method to establish program equivalence and other behavioral properties of higher-order programs, also related to the concept of \emph{relational parametricity} \cite{Reynolds1983}. The primary source of inspiration for the QLR are differential logical relations (DLR) \cite{dallago,dallago2}, whose cartesian closed structure is very similar to that of the category $\mathbf{Q}$. While DLR can be seen as special cases of QLR (see footnote \ref{foot1}), the only metric structure studied for the DLR in \cite{dallago} are what we called here hyper-relaxed metrics. A precursor of this approach is \cite{chaudhuri}, which develops a System F-based system for approximate program transformations, but without explicitly mentioning any metric structure. The category $\mathbf{V}$ from Section \ref{section5} is reminiscent of the diameter spaces from \cite{Geoffroy2020}, which form a cartesian \emph{lax}-closed category based on a similar factorization of partial metric spaces. A main difference is that in \cite{Geoffroy2020} the factorization is considered as a \emph{property} of (suitable) partial metric spaces, rather than an additional \emph{structure}, as we do here Several \emph{relational logics} have been developed to formalize logical relations and, more generally, higher-order relational reasoning \cite{Plotkin1993, Ahmed2011, 10.1145/2775051.2676980, 10.1145/3009837.3009877,10.1145/3110265}, including quantitative reasoning \cite{Barthe_2012, Gabo2019b}. An important question, which transcends the scope of this paper, is whether one can describe a QLR semantics for at least some of these logics, or if a different relational logic has to be developed in order to capture quantitative relational reasoning based on QLR. The literature on program metrics in denotational semantics is vast. Since \cite{ARNOLD1980181} metric spaces have been exploited as an alternative framework to standard, domain-theoretic, denotational semantics. Notably, \emph{Banach's fixed point theorem} plays the role of standard order-theoretic fixpoint theorems in this setting (see \cite{VANBREUGEL20011} and \cite{BAIER1994171}). More recently, program metrics have been applied in the field of differential privacy \cite{10.1145/1932681.1863568, 10.1007/978-3-642-29420-4_3, Barthe_2012}, by relying on Lipschitz-continuity as a foundation for the notion of program sensitivity. To this line of research belongs also the literature on System $\mathsf{Fuzz}$ \cite{10.1145/1932681.1863568}, a sub-exponential PCF-style language designed for differential privacy, which admits an elegant semantics based on metric spaces and metric CPOs \cite{10.1145/1932681.1863568, Gaboardi2017}. Ultra-metrics are widely applied in program metrics, mostly to describe intensional aspects (e.g.~traces, computation steps) \cite{VANBREUGEL20011, MAJSTERCEDERBAUM1991217, Escardo1999}, also for the $\lambda$-calculus, due to the fact that when $Q$ is a locale, $\mathsf{Met}_{Q}$ is cartesian closed. Partial metrics were introduced in \cite{matthews} with the goal of modeling partial objects in program semantics, and independently discovered in sheaf theory as \emph{$M$-valued sets} \cite{Hohle:1992aa}. \cite{Bukatin1997} shows that partial metrics and relaxed metrics can be used to characterize the topology of continuous Scott domains with a countable bases. This work was, to our knowledge, the first to acknowledge the correspondence between partial metrics and lattices, which was later developed through the theory of valuations \cite{10.1007/BFb0053546, ONeill, 10.1016/j.tcs.2003.11.016}. \cite{AGT7849} provides a topological characterization of partial metric spaces. Fuzzy and probabilistic partial metric spaces are well-investigated too \cite{Yueli:2015aa, Wu:2017aa, HE201999}. Our description of generalized partial metric spaces was based on the elegant presentation from \cite{Stubbe2018, STUBBE201495} in the language of quantaloid-enriched categories. Together with standard real-valued metrics, Lawvere's generalized metrics \cite{Lawvere1973} have also played a major role in these research lines. More generally, the abstract investigation of metric spaces as quantale and quantaloid-enriched categories is part of the growing field of \emph{monoidal topology} \cite{Hofmann2014}. To this approach we can ascribe the already mentioned description of partial metric spaces from \cite{Stubbe2018, STUBBE201495}, as well as the general characterization of \emph{exponentiable} metric spaces and quantaloid-enriched categories in \cite{CLEMENTINO20063113,Clementino:2009aa}. Quantitative approaches based on generalized metric spaces have been developed for bisimulation metrics \cite{Bonchi2014, Bonchi2018, DBLP:journals/lmcs/BaldanBKK18} and algebraic effects \cite{Plotk, 10.1145/3209108.3209149}. Generalized metrics based on Heyting quantales have been used to investigate properties of graphs and transition systems (see \cite{Pouzet2020} for a recent survey). Finally, research on axiomatizations of abstract notions of differentiation has been a very active domain of research in recent years \cite{Blute2009, Cockett2011, Cockett2018, Blute2019,10.1007/978-3-030-17127-8_3, 10.1007/978-3-030-45231-5_4}, supported by the growth of interest in algorithms based on automatic differentiation. The two notions of derivative discussed in this paper can be compared with two lines of research on abstract differentiation. On the one hand, the derivatives arising from differential logical relations (which essentially coincide with the derivatives from $\mathbf{Q}$) have been compared \cite{dallago2} with those found in some recent literature on discrete differentiation (e.g.~finite difference operators, Boolean derivatives), and approaches based on the so-called \emph{incremental $\lambda$-calculus} \cite{Cai2014, 10.1007/978-3-030-17184-1_19, 10.1007/978-3-030-45231-5_4}. On the other hand, the derivatives from Section \ref{section6} can be compared with the literature on Cartesian Differential Categories, originating in Ehrhard and Regnier's work on \textit{differential linear logic} and the \textit{differential $\lambda$-calculus} \cite{ER}. Very recently, Cartesian Difference Categories \cite{10.1007/978-3-030-45231-5_4} have been proposed as a framework unifying these two lines of research. \section{Conclusion} This paper provides just a first exploration of the program metrics semantics that arise from the study of quantitative logical relations, and leaves a considerable number of open questions. We indicate a few natural prosecutions of this work. While our focus here was only on cartesian closure, it is natural to look for QLR-models with further structure (e.g.~coproducts, recursion, monads etc.). For instance, by extending the picture to \emph{quantaloid}-valued relations \cite{STUBBE201495}, one can define a coproduct of QLR with nice properties. The correspondence between metrics and enriched categories suggests to consider the transitivity axiom as a ``vertical'' composition law for distances. An interesting question is whether one can define higher-dimensional categories of program distances with a nice compositional structure, in analogy with well-investigated higher-dimensional models in \emph{categorical rewriting} \cite{Meseguer1992, Miyoshi1996}. At a more formal level, the same observation also suggests to investigate \emph{relational logics} to formalize the metric reasoning justified by QLR-models, in line with the program logics developed for standard logical relations \cite{Plotkin1993, Ahmed2011} and for quantitative relational reasoning \cite{10.1145/2775051.2676980, 10.1145/3009837.3009877,10.1145/3110265, Barthe_2012, Gabo2019b}. \bibliographystyle{plain}

3f8df07778018d7bb9505f26499724df07514675

\section{Introduction} The Web is continuously becoming a bigger part of everyday life. The first thing many people do when they want some information is turn to their web browser and type into the search bar. People will search the Web for news, articles, weather reports, or even tutorials if they want to learn a new hobby or skill. We rely on the Web browser to allow us to navigate the Web safely and securely. Web browsers will process information about the webpage as it loads and present us with appropriate user interface cues and messages about the page, keeping us informed as we navigate the Web. For example, browsers can detect if a connection is secure, and we have become used to the browser warning us when a website is insecure or could potentially take our information. The HTTPS secure lock icon is there to quickly tell us that the page we are viewing is safe. Browsers also check for more information about the page other than just security. If a tab is playing music or a video, the browser can add an icon to that tab to let us know that the website displayed by that tab is attempting to play audio. The Web is a great resource for information, however it is extremely dynamic. Not only are new webpages being created, but existing ones are changing or being deleted. For these reasons, archiving the Web has become increasingly important. Archiving a webpage produces what is known as a memento \cite{vandesompel2009memento}. Mementos allow you to go back in time on the Web, which is useful for either studying the past Web or retrieving information from a webpage that was changed or deleted. One of the easiest ways you can view mementos is by visiting a web archive, such as the Internet Archive \cite{internet-archive} or Trove \cite{trove}. Mementos can also be found outside of web archives \cite{memento:mediawiki}. For example, \url{w3.org} stores old revisions of their webpages as mementos, which is useful for visiting the page exactly as it was in the past. Figure \ref{fig:w3-example} shows one of these old revisions and circled in red is where you can find the date the page is from, which for this particular memento is April 8th, 2020. Currently, archived webpages are not recognized by web browsers, meaning that the browser does not react any differently if the page is archived or a part of the live Web. Because of this, the user has to look out for visual cues on the page itself to see if the webpage they are viewing is from the past or current Web. Web archives do a great job of showing the datetime the memento was captured, in addition to displaying other information about the archived page (Figure \ref{fig:ia-example}). However, mementos such as the one shown in Figure \ref{fig:w3-example} may have a datetime visible on the page that is hard to locate, if they do have a datetime visible at all. As archived webpages become increasingly common, this can become an issue and confuse users. A user could even be led to believe that they are on a different page than the one they actually navigated to. \begin{figure*}[ht] \centering \includegraphics[width=0.91\textwidth, frame]{images/w3-example.png} \caption{Memento of an old revision of a w3.org page.} \label{fig:w3-example} \end{figure*} \begin{figure*}[ht] \centering \includegraphics[width=0.91\textwidth, frame]{images/ia-example.png} \caption{Memento via the Internet Archive's Wayback Machine \cite{wayback}. Memento Datetime and web archive information are clearly visible to the user.} \label{fig:ia-example} \end{figure*} \section{Basics of Modifying Chromium} \label{section:design} The Memento-aware Browser was created by adding on to the implementation of Chromium, Google’s open source web browser. This section describes important details about working with the Chromium code base, such as some differences between how it works on different operating systems and how to successfully add on to the implementation. The section also details the proposed features of the Memento-aware Browser and the proposed design that was implemented. Much of the initial design of the features as well as their final implementation was shaped around the current implementation and user experience (UX) of the Chromium browser. \subsection{Getting Started with the Source Code} \label{section:working-with-chromium} To begin working with Chromium, the repository must first be properly downloaded along with the necessary tools. The Chromium source code \cite{chromium-source} is readily available for download. Additionally, Google makes their tools that are required to build and run Chromium, called \lstinline{depot_tools} \cite{depot_tools}, available for download. \subsection{Downloading and Building Chromium} Google provides detailed instructions to download, build, and run the Chromium source code for Linux, macOS, and Windows. The Chromium source code also has a GitHub mirror \cite{chromium-github}, however the documentation advises that you do not download and run the code from GitHub. Instead, Google advises that you first download their \lstinline{depot_tools} and run \lstinline{fetch chromium} to download the latest source code for your operating system. Since running \lstinline{fetch chromium} downloads operating system specific code, the same code base will not be able to build and run the Chromium browser on all operating systems, which is why cloning the GitHub mirror is not recommended. Before being able to run \lstinline{fetch chromium} and get the source code, you must first download \lstinline{depot_tools} which provides the necessary tools for running Chromium specific commands. Instructions on how to set up the path for \lstinline{depot_tools} varies between operating systems. Since the Chromium code base varies slightly between operating systems, it is important to note that the Memento-aware Browser was primarily implemented on Linux, specifically Ubuntu 20.04. Additionally, the browser was tested on Windows. Releases for Linux and Windows were generated and are available to download on the Memento-aware Browser GitHub Repository \cite{repo}. The May 2020 Chromium version that was used as a base for the browser did not run on the current version of macOS Catalina, so a macOS version was not created. \subsection{Editing the Chromium Code} The majority of the Chromium implementation that was edited for the Memento-aware Browser was in C++. Editing these C++ files makes changes for all operating systems. Occasionally, it is necessary to make operating system specific changes to the code. Listing \ref{lst:os_spec_header_files} shows an example of including different header files depending on the operating system. As you can see in the listing, if the operating system is Windows then \lstinline{base/base_paths_win.h} will be included. If the operating system is macOS or Android, then the header files specific to those operating systems will be included. In addition to including header files depending on the operating system, it is also possible to include different code blocks depending on the operating system. Listing \ref{lst:os_spec_if_statement} shows an if statement executing only if the current operating system is macOS. \begin{lstlisting}[language=C++, caption=Include different header files depending on the operating system., label=lst:os_spec_header_files, ] #if defined(OS_WIN) #include "base/base_paths_win.h" #elif defined(OS_MACOSX) #include "base/base_paths_mac.h" #elif defined(OS_ANDROID) #include "base/base_paths_android.h" #endif \end{lstlisting} \begin{lstlisting}[language=C++, caption=Creating operating system specific code., label=lst:os_spec_if_statement, ] #if defined(OS_MACOSX) // Operating system specific code here #endif // defined(OS_MACOSX) \end{lstlisting} \subsection{Debugging the Code} When running Chromium in debug mode, the easiest way to print debug logs is with \lstinline{DVLOG(0)}. You can further control when debug statements are printed by passing different parameters to the \lstinline{DVLOG()} function as outlined in the Chromium docs \cite{chromium-logging}. Listing \ref{lst:dvlog} shows an example of \lstinline{DVLOG(0)} being used to print a debug statement, and Listing \ref{lst:dvlog-output} shows the output of that line. \begin{lstlisting}[language=C++, caption=Using DVLOG(0) for debugging., label=lst:dvlog, ] DVLOG(0) << "Datetime: " << memento_datetime; \end{lstlisting} \begin{lstlisting}[language=bash, caption=Output of DVLOG(0)., label=lst:dvlog-output, ] Datetime: Tue, 05 Mar 2019 09:38:34 GMT \end{lstlisting} \subsection{Adding Additional Classes} In order to implement the features of the Memento-aware Browser, it was necessary to not only alter existing C++ classes but also to create entirely new classes. Creating and adding new files to add a new class to the Chromium code is a simple multi-step process. First, you must create the files you need and save them to the appropriate directory. For example, to add the files \lstinline{example.cc} and \lstinline{example.h} to \lstinline{Memento-aware-Browser/src/chrome/common} you must create and save them in the \lstinline{Memento-aware-Browser/src/chrome/common} directory. Next, you must locate the \lstinline{BUILD.gn} file within the \lstinline{Memento-aware-Browser/src/chrome/common} directory. Add the lines \lstinline{"example.cc"} and \lstinline{"example.h"} to the \lstinline{BUILD.gn} as shown in Figure \ref{fig:build-gn}. With the appropriate changes made to the source code, the application can be built with the new classes now included. \begin{figure*}[ht] \centering \includegraphics[width=0.6\textwidth, frame]{images/build-gn.png} \caption{Adding a new class to \lstinline{Memento-aware-Browser/src/chrome/common/BUILD.gn}} \label{fig:build-gn} \end{figure*} \subsection{Adding Additional Icons} The Chromium browser uses the Skia Graphics Library \cite{skia} as its graphics engine. Any icons rendered by the browser are in the Skia format. The contents of the Skia icon file for the HTTPS secure lock icon are shown in Listing \ref{lst:skia-example}. In the listing, three different icon sizes are shown, where the first icon size is the largest and the last icon size is the smallest. The first is 36 x 16, the second is 24 x 34, and so on. The browser picks which size to use depending on what the icon is being rendered for. One way to generate a new icon to use in the browser is to first create an SVG of the icon and then ``Skiafy" the icon using an online tool. The tool that was used to create the memento icon for the Memento-aware Browser was generated from an SVG using a tool called Skiafy by Evan Stade \cite{skiafy}. The output produced by the Skiafy tool is a great starting point to generate any icon to use in the Chromium browser, however it was necessary to make small changes to the output from the tool in order to get the sizing of the icon correct. \begin{lstlisting}[float, language=bash, caption=Skia file for the HTTPS secure lock icon., label=lst:skia-example, ] MOVE_TO, 36, 16, R_H_LINE_TO, -2, R_V_LINE_TO, -4, R_CUBIC_TO, 0, -5.52f, -4.48f, -10, -10, -10, CUBIC_TO, 18.48f, 2, 14, 6.48f, 14, 12, R_V_LINE_TO, 4, R_H_LINE_TO, -2, R_CUBIC_TO, -2.21f, 0, -4, 1.79f, -4, 4, R_V_LINE_TO, 20, R_CUBIC_TO, 0, 2.21f, 1.79f, 4, 4, 4, R_H_LINE_TO, 24, R_CUBIC_TO, 2.21f, 0, 4, -1.79f, 4, -4, V_LINE_TO, 20, R_CUBIC_TO, 0, -2.21f, -1.79f, -4, -4, -4, CLOSE, MOVE_TO, 24, 34, R_CUBIC_TO, -2.21f, 0, -4, -1.79f, -4, -4, R_CUBIC_TO, 0, -2.21f, 1.79f, -4, 4, -4, R_CUBIC_TO, 2.21f, 0, 4, 1.79f, 4, 4, R_CUBIC_TO, 0, 2.21f, -1.79f, 4, -4, 4, CLOSE, [...] CLOSE \end{lstlisting} \section{Memento-Aware Design} \subsection{Basic Memento Detection} \label{section:basic-memento-detection} The first task in implementing memento detection is to first consider the most basic of possibilities. This would be when the entire root webpage is archived and considered to be a memento. You can easily view this case by looking at an archived page within a web archive. Figure \ref{fig:wayback-mitre} is an example of an archived page from the Internet Archive’s Wayback Machine \cite{wayback}. In this example, we want the browser to classify the visible page as a memento and alert the user that the page is archived and not live. We cannot accomplish this by looking at the URL of the page since there is no defined format for a memento URL. If we look at the URL for an archived page from the Wayback Machine as shown in Figure \ref{fig:wayback-mitre}, we see that the URL contains a datetime as well as the URL of the original live page. However, if we look at a URL for an archived page from Perma.cc \cite{permacc} we see that there is no datetime or original URL. Instead, there is just a hash. This means memento detection is not as simple as just parsing the URL. A better method is to parse the HTTP response headers to find if the webpage is archived since archived webpages return the Memento-Datetime HTTP response header. The Memento-Datetime header indicates that the response contains a representation of a memento where the value of the header is the datetime of the original resource \cite{mementoweb}. The Memento-Datetime header can be observed in Figure \ref{fig:curl-wayback-mitre} where a simple HEAD request with Curl was made for \href{https://web.archive.org/web/20100412125057/http://www.mitre.org/}{web.archive.org/web/20100412125057/http://www.mitre.org/}. \begin{figure*}[ht] \centering \includegraphics[width=0.8\textwidth, frame]{images/wayback-mitre.png} \caption{Memento of mitre.org in 2010 via the Wayback Machine \cite{wayback}} \label{fig:wayback-mitre} \end{figure*} \begin{figure*}[ht] \centering \includegraphics[width=0.8\textwidth, frame]{images/curl-wayback-mitre.png} \caption{Curl request for https://web.archive.org/web/20100412125057/http://www.mitre.org/} \label{fig:curl-wayback-mitre} \end{figure*} \subsection{Detecting Embedded Mementos} A memento does not have to be the root webpage. It is possible that a live root webpage could contain one or more embedded archived frames within itself. Most commonly this occurs when the root webpage contains an iframe that loads an archived page. When it comes to these embedded mementos, there are three possibilities: \begin{itemize} \item The first possibility is that it is intended for the iframe memento to make up the whole webpage and the root page should be considered a memento even though the URL of the root page does not return the Memento-Datetime header. This case can be seen in web archives that display the memento within an iframe on the root page, rather than making the memento the root page itself. Examples of such web archives are Trove \cite{trove} and Perma.cc \cite{permacc}. Figure \ref{fig:trove-iframe-memento} illustrates this case by showing a memento displayed in an iframe from Trove. Additionally, the figure shows a diagram of the rendered frames on the page. In the diagram the green portion represents the root webpage and the purple portion represents the embedded iframe. This page structure can be translated to the tree structure as shown in Figure \ref{fig:iframe-memento-tree} where the root webpage is the root node of the tree and the iframe memento is a child node of the root page. \item The next possibility is that a live webpage is simply displaying a memento without intending for the memento to make up the whole page. A simple example of this would be when a live page decides to place an iframe memento within their website so their readers can view the content of that memento. To demonstrate this, an example page was created at \url{https://www.cs.odu.edu/~amabe/oneiframe.html}. This is similar to the first possibility in that there is a single memento associated with the current page being viewed. However, the iframe memento is not intended to make up the entire webpage. Instead, it is intended to be an additional element on the page. This possibility is shown in Figure \ref{fig:oneiframe} where the test webpage is shown on the right and on the left is a diagram of the page structure. In the diagram, the gray portion represents the root webpage with no Memento-Datetime header and the blue portion represents the iframe memento. This can also be translated into a tree structure, shown in Figure \ref{fig:oneiframe-tree}. Note that this tree structure is identical to the structure of the memento from Trove shown in Figure \ref{fig:iframe-memento-tree}. As a user, it is easy to differentiate the Trove iframe example from the test webpage example, but from the tree structure the two examples appear the same. \item The final possibility, shown in Figure \ref{fig:multiframe}, is when a live webpage has multiple mementos on the page. The context of this is similar to the previous possibility, but instead of a single memento embedded on the currently viewed page there are multiple embedded mementos. An example webpage was created to demonstrate this and can be viewed at \url{https://www.cs.odu.edu/~amabe/test.html}. Again, this can be thought of as a tree structure as shown in Figure \ref{fig:multiframe-tree} where the root node (root webpage) has 3 child nodes (iframe mementos). \end{itemize} \begin{figure*}[ht] \centering \includegraphics[width=0.8\textwidth, frame]{images/trove-iframe-memento.png} \caption{Memento displayed within an iframe via Trove} \label{fig:trove-iframe-memento} \end{figure*} \begin{figure*}[ht] \centering \includegraphics[width=0.8\textwidth, frame]{images/iframe-memento-tree.png} \caption{Tree structure of a live root webpage from Trove containing an iframe Memento} \label{fig:iframe-memento-tree} \end{figure*} \begin{figure*}[ht] \centering \includegraphics[width=0.85\textwidth, frame]{images/oneiframe.png} \caption{Memento displayed within an iframe as an additional page element on a live webpage} \label{fig:oneiframe} \end{figure*} \begin{figure*}[ht] \centering \includegraphics[width=0.85\textwidth, frame]{images/oneiframe-tree.png} \caption{Tree structure of a live root webpage containing an iframe Memento} \label{fig:oneiframe-tree} \end{figure*} \begin{figure*}[ht] \centering \includegraphics[width=0.85\textwidth, frame]{images/multiframe.png} \caption{Three mementos displayed within iframes as additional page elements on a live webpage} \label{fig:multiframe} \end{figure*} \begin{figure*}[ht] \centering \includegraphics[width=0.85\textwidth, frame]{images/multiframe-tree.png} \caption{Tree structure of a live root webpage containing three iframe Mementos} \label{fig:multiframe-tree} \end{figure*} Each of these possibilities should be detected by the browser in order to present the user with the correct message about the content they are viewing. Considering the first case, if the user were to view an archived page through Trove or Perma.cc the browser should detect the datetime from the archived content in the iframe and detect that the iframe is formatted to make up the entire page. When these things are detected, the browser should then tell the user that the page they are viewing is an archived page from a past date. In the second possibility, the browser should detect the archived content within the page but also detect that the embedded iframe is an additional element on the live page. The user should be able to tell through the native UI elements in the browser that the page they are viewing contains archived content and they should also be able to read the datetime that the archived content is from. In the third and final case, the browser should behave similarly as to when the second possibility occurs. The only difference is that the browser should list all datetimes from all archived content on the page rather than just the one datetime. \subsection{Embedded Elements Within Mementos} \label{section:embedded-elements-within-mementos} As covered in the previous section, a root webpage can display a number of embedded elements. It is possible that the root webpage is an archived page, or an embedded element on the root page is archived. The previous cases covered were about live webpages displaying archived content. The next possible cases to cover involve an archived page or embedded archived element displaying its own embedded elements that may or may not also be archived. The archived page displaying embedded content could be the root page or an embedded element itself. Either way, an archived page or element can display its own embedded elements and they may or may not also be archived which changes how the browser should react to the archived content. \begin{itemize} \item The first case is that an archived frame (could be the root page or an element on the root page) contains one or more embedded elements that are displaying content from the live web. This is a rare yet undesirable occurrence since the archived webpage is intended to present content from the past, yet there is a portion or portions of the page that can display content from the live web inside the memento. When this occurs, it is not necessarily obvious to the user that the live web is leaking into the archived page. Such an occurrence could be called a ``zombie" \cite{zombies}. A common example of this occurring is when the archived page contains Javascript or an embedded element that pulls content from the live web. Figure \ref{fig:best_one} illustrates this. In the figure, it can be seen that the memento is from 2008 but there is an advertisement from 2012. \item The second case is that an archived page contains one or more embedded elements that are displaying additional archived frames, each with their own datetime. This is similar to the first case in that the memento is displaying embedded content, but the content is not from the live web. Typically when this case occurs it is not an issue as the embedded frames that are displayed within the memento were archived with the memento and do not allow the live web to leak in. \end{itemize} In these cases, whether the archived frame that is displaying embedded content is the root page or an embedded element itself, the browser should parse response headers for any content within the archived frame. If Memento-Datetime headers are found, the user should be alerted that frames with different datetimes exist within the page. If frames within the archived content do not return the Memento-Datetime header, the user should be alerted that the archive content they are viewing contains the live web. \begin{figure*}[ht] \centering \includegraphics[width=0.8\textwidth, frame]{images/best_one.png} \caption{``Zombie" memento from 2008 that is displaying an advertisement from 2012 \cite{zombies}} \label{fig:best_one} \end{figure*} \subsection{Bookmark as Archive} The concept of a Memento-aware browser goes beyond the browser detecting archived content. In addition to making users aware of any archived content they are viewing, users should also be able to easily archive any live content that they want to save. Essentially, when the user saves a bookmark they should have the option to also archive it so that when they go to their bookmarks, they have the option to go to the live web or view the archived version in a public web archive \cite{weigle-wadl18}. To design a feature that allows users to submit a live webpage to be archived, it is best to first think of what a user currently does to save a webpage. The common way that a user saves a webpage for viewing at a later date is by bookmarking the page through their browser. The bookmark functionality saves a link to the page within the browser settings and allows the user to revisit that link in the location they saved it. For example, saving a bookmark to the bookmarks bar places a link to the page in a readily accessible location in the browser UI. The user also has the option to save the bookmark to a folder. The user can create a bookmark folder named “School” and save any relevant webpages to their “School” bookmark folder as shown in Figure \ref{fig:bookmark-folders}. This bookmarking functionality is a great way to save and organize webpages for later viewing. However, bookmarking only saves a link to the page. If the page were to be taken down, the bookmark link would lead to a 404. Additionally, if the content on the webpage were to change, the bookmark link would still lead to the correct page but the content that the user originally saw and wanted to save would no longer be present. In many cases we can assume that the page the user wanted to save contained important information, which is why they wanted to bookmark it. In the case of 404s and changed content, that important information that the user wanted to save is, at best, temporarily lost and the information on the page may be available again at a later date or is now available elsewhere. In the worst case scenario that content is permanently lost. \begin{figure*}[ht] \centering \includegraphics[width=0.8\textwidth, frame]{images/bookmark-folders.png} \caption{Bookmark folders within the Chromium bookmark manager} \label{fig:bookmark-folders} \end{figure*} A way to make these bookmarked pages constant is through the use of public web archives. When the user wants to bookmark a webpage, the browser should also allow them to submit the webpage to a public web archive. This way, the browser bookmarking functionality would work as normal where a link to the webpage would be placed in the proper bookmark folder. However in addition to this, the browser would also submit the webpage to the user selected public web archive. The browser would create an additional bookmark folder for this archive and place the link to the webpage in the archive within this new folder. The end result of this design is that the user has their traditional bookmark system, but if they encounter a 404 or a changed webpage they can access their archived bookmarks and visit the page as it was the day that they originally wanted to save it. \section{Implementation} The design and features described in the previous section were implemented in a way that allowed the features to be added on to the native implementation of the browser with minimal changes to the native code. The majority of the code for the Chromium web browser is C++ so most of the implementation was done in C++. The Chromium browser does also have HTML5, Javascript, CSS3, and Python files that make up things such as settings pages and the bookmark manager for the browser. Some of these files were edited as needed. To build and run the browser, Google provides their own \lstinline{depot_tools} and the Chromium browser uses the Ninja build system, so the instructions listed by The Chromium Projects \cite{install-dependencies} were followed in order to build and run the browser. The primary environment used to develop the Memento-aware Browser has the following specifications: \begin{itemize} \item Ubuntu 20.04.1 LTS 64-bit operating system \item 16 GB memory \item AMD Ryzen 5 3600 3.6 GHz 6-Core Processor \item 1TB solid state drive \item Python 3.8.5 \item HTML5 and CSS3 for webpages standard to the browser (settings, profile, etc.) \end{itemize} Additionally, the Memento-aware Browser was tested on a Windows system with the following specifications: \begin{itemize} \item Windows 10 Home 64-bit operating system \item 16 GB memory \item Intel(R) Core(TM) i7-8550U CPU \item 1TB solid state drive \item Python 3.8.0 \item HTML5 and CSS3 for webpages standard to the browser (settings, profile, etc.) \end{itemize} The implemented features were primarily done in C++ code that works the same between operating systems. The primary difference between operating systems that should be noted is required dependencies. The GitHub repository for the Memento-aware Browser \cite{repo} can be built and run on Linux and Windows as is. With the proper dependencies and up to date Chromium version as a base, the code could potentially also be run on macOS. \subsection{Chromium Page Structure} \label{section:chromium-page-structure} The implementation for the Memento-aware Browser is built off of the Chromium browser page structure and the state of the page security information. The Chromium browser considers a single tab as an entry where the current active tab is known as an Active Entry. The Active Entry contains information about the page and all resources that make up that page. A part of the information about the page contained within the Active Entry information is the Security State. The Security State contains security information about the page such as whether the protocol of the root webpage is secure or insecure, and if the webpage contains any content that is insecure. The Active Entry also consists of a tree of rendered frames that exist on the page. The root node of the tree is the root webpage, and child nodes of the root node are rendered frames on the root page. This tree structure is illustrated in Figure \ref{fig:chromium-page-structure}. Also shown in the figure is a single insecure HTTP node. Since this one child node on the tree is considered insecure, this makes the entire Active Entry considered insecure since it contains a mix of secure and insecure content. \begin{figure*}[ht] \centering \includegraphics[width=0.6\textwidth, frame]{images/chromium-page-structure.png} \caption{Chromium treats rendered frames as nodes in a tree with the root node being the root page. An insecure HTTP node can cause an insecure security state for the page.} \label{fig:chromium-page-structure} \end{figure*} To include archived page information in the Active Entry, the Security State has been extended to contain Memento-Datetime response header information found while parsing response headers. \subsection{Detecting the Memento-Datetime Header} \label{section:detecting-memento-header} To implement the memento information portion of the Security State, the Memento-Datetime response header first needs to be detected. Additionally, as described in the Section \ref{section:design}, there are a variety of ways that archived content needs to be detected and a variety of messages to present to the user based on the presence of archived content. Properly detecting archived content begins with parsing any received response headers when a page navigation is requested to find the Memento-Datetime header. When this header is found, the datetime also needs to be extracted and associated with the resource that returned that particular memento datetime. Within the Chromium browser, response headers from all resources are parsed within the URLLoader class. The URLLoader class constructs a URLResponseHead object to hold response information for that particular resource. A URLResponseHead object holds information about the resource such as the URL itself and the \lstinline{content_length} of the resource. To implement memento detection, the URLResponseHead object has been extended to hold a \lstinline{memento_datetime} string that is an empty string by default. If the Memento-Datetime header is found when the response headers are parsed, then the \lstinline{memento_datetime} member of the URLResponseHead object is set to the value of the Memento-Datetime header (Listing \ref{lst:urlresponsehead-mementodatetime}). \begin{lstlisting}[language=C++, caption=Finding the Memento-Datetime header and setting the \lstinline{memento_datetime} of the URLResponseHead object., label=lst:urlresponsehead-mementodatetime, ] // Check for Memento-Datetime header. if (response->headers->HasHeader("Memento-Datetime") || response->headers->HasHeader("memento_datetime")) { response->memento_info = true; response->memento_datetime = response->headers->GetMementoDatetime(); } \end{lstlisting} The URLResponseHead object can then be accessed by the currently committing NavigationRequest. Listing \ref{lst:navigationrequest-mementodatetime} shows the member function \lstinline{GetMementoDatetime()} of the NavigationRequest class where the URLResponseHead object returns the \lstinline{memento_datetime}. \begin{lstlisting}[language=C++, caption=Accessing the Memento-Datetime value from the URLResponseHead., label=lst:navigationrequest-mementodatetime, ] std::string NavigationRequest::GetMementoDatetime() { if(response()) { return response()->memento_datetime; } return ""; } \end{lstlisting} \subsection{Root Page} With parsing response headers for the Memento-Datetime header implemented within the \lstinline{URLLoader} class, root page memento detection could be further implemented. When the user navigates to a webpage, a \lstinline{NavigationRequest} is committed. The \lstinline{Navigator} class initiates the navigation and utilizes the \lstinline{NavigationController} class to control the currently rendered frame (\lstinline{RenderFrameHostImpl}) and the navigation entry (\lstinline{NavigationEntry}). Since the \lstinline{URLResponseHead} object for the root page contains the \lstinline{memento_datetime} member variable, this datetime string can be accessed when a \lstinline{NavigationRequest} is committed. Using the \lstinline{NavigationRequest::GetMementoDatetime()} function, the currently rendering frame member variable, \lstinline{params->memento_datetime}, is set to the \lstinline{memento_datetime} of the currently committing navigation request (Listing \ref{lst:renderframehostimpl-mementodatetime}). \begin{lstlisting}[float, language=C++, caption=Within \lstinline{render_frame_host_impl.cc} set the \lstinline{memento_datetime} of the current rendered frame to the \lstinline{memento_datetime} of the root page resource., label=lst:renderframehostimpl-mementodatetime, ] void RenderFrameHostImpl::DidCommitPerNavigationMojoInterfaceNavigation(...) { ... params->memento_datetime = committing_navigation_request->GetMementoDatetime(); ... } \end{lstlisting} With the rendering frame's \lstinline{memento_datetime} parameter properly set to the value that was stored in the \lstinline{URLResponseHead}, the \lstinline{memento_datetime} can then be set for the entire \lstinline{NavigationEntry} within the \lstinline{NavigationController} class (Listing \ref{lst:activeentry-mementodatetime}). \begin{lstlisting}[float, language=C++, caption=Set the \lstinline{memento_datetime} of the \lstinline{NavigationRequest} object to the value stored in \lstinline{params.memento_datetime}., label=lst:activeentry-mementodatetime, ] if (params.memento_datetime != "") { active_entry->SetMementoDatetime(params.memento_datetime); active_entry->SetMementoInfo(true); } \end{lstlisting} The \lstinline{NavigationEntry} object that now holds the proper memento information variables for the page acts as the bridge between the back-end and front-end of the browser as the page loads. Figure \ref{fig:rootpage-loading-backend} illustrates the previously described process of detecting the Memento-Datetime header when the URL is loaded and bringing the value up to the where it can be accessed by the front-end. In the figure, when the user loads a webpage and a \lstinline{NavigationRequest} is initiated, it creates a \lstinline{URLResponseHead} object that contains the memento information member variables, along with other necessary variables for the loaded URL. The \lstinline{URLLoader} class loads the URL for the root page and parses the response headers. If the Memento-Datetime header is found, the values of the memento information member variables of the \lstinline{URLResponseHead} are set accordingly. Next, the \lstinline{Navigator} can begin processing the navigation to the webpage. The \lstinline{Navigator} uses a \lstinline{NavigationController} to control the \lstinline{RenderFrameHostImpl} object (the currently rendering frame) and the \lstinline{NavigationEntry} object (the entry for the navigation to that particular webpage). The parameters of the rendering frame are set based off of the \lstinline{URLResponseHead} object for that particular frame and its respective resources. There is a single main \lstinline{NavigationEntry} object being used for a webpage navigation, but the currently rendering frame changes depending on what frame within the page is currently loading. Because of this, the \lstinline{NavigationEntry} variables are updated as the currently rendering frame parameters are updated. Thus, the \lstinline{NavigationEntry} contains information regarding all rendered frames within that navigation once loading is fully complete. \begin{figure*}[ht] \centering \includegraphics[width=0.8\textwidth, frame]{images/rootpage-loading-backend.png} \caption{The sequence of classes involved in loading a webpage. These are the main classes that were edited to implement rootpage memento detection.} \label{fig:rootpage-loading-backend} \end{figure*} The front-end interacts with the \lstinline{NavigationEntry} through the \lstinline{ContentUtils} class. Within \lstinline{content_utils.cc}, various parameters that make up the \lstinline{VisibleSecurityState} are set according to the values within the \lstinline{NavigationEntry} (Listing \ref{lst:rootpage-contentutils}). \begin{lstlisting}[float, language=C++, caption=Set the \lstinline{VisibleSecurityState} according to the values that were set for the current \lstinline{NavigationEntry} within \lstinline{content_utils.cc}., label=lst:rootpage-contentutils, ] // Flag for if the current NavigationEntry is an error page state->is_error_page = entry->GetPageType() == content::PAGE_TYPE_ERROR; // String that hold the Memento-Datetime header value of the NavigationEntry state->memento_datetime = entry->GetMementoDatetime(); // Flag for if the current NavigationEntry is considered a Memento state->memento_info = entry->GetMementoInfo(); \end{lstlisting} With the \lstinline{VisibleSecurityState} for the webpage set, the \lstinline{PageInfo} class can then access the \lstinline{VisibleSecurityState} object and use its parameters to create an \lstinline{IdentityInfo} object for the page. The \lstinline{IdentityInfo} object will then be used to set the values within \lstinline{PageInfoUI} to construct the user interface according to the various security attributes within the \lstinline{VisibleSecurityState} (Listings \ref{lst:rootpage-pageinfo} and \ref{lst:rootpage-pageinfoui}). \begin{lstlisting}[language=C++, caption=Create an \lstinline{IdentityInfo} object and set the memento information members within \lstinline{page_info.cc}., label=lst:rootpage-pageinfo, ] PageInfoUI::IdentityInfo info; info.memento_status = memento_status_; info.memento_datetime = memento_datetime_; \end{lstlisting} \begin{lstlisting}[float, language=C++, caption=Construct the user interface according to the values passed from the \lstinline{IdentityInfo}., label=lst:rootpage-pageinfoui, ] if (memento_status && memento_datetime != "") { security_description->memento_summary = l10n_util::GetStringUTF16(IDS_PAGE_INFO_MEMENTO_SUMMARY); std::string datetime_string = "The page displayed is a memento captured on " + memento_datetime; security_description->memento_info = base::UTF8ToUTF16(datetime_string); } \end{lstlisting} This process of the front-end accessing the \lstinline{NavigationEntry} object to construct the final set of UI elements for the page is illustrated in Figure \ref{fig:rootpage-loading-frontend}. The \lstinline{ContentUtils} class controls the use of the \lstinline{NavigationEntry} object to set the \lstinline{VisibleSecurityState} of the page. \lstinline{ContentUtils} is continuously updating as pieces of the webpage load and the \lstinline{NavigationEntry} updates with new information. The \lstinline{PageInfo} class pulls information from the \lstinline{VisibleSecurityState} and constructs an \lstinline{IdentityInfo} object to be passed to \lstinline{PageInfoUI} where the parameters that make up the front-end of the browser are set accordingly. This is where any messages presented to the user are constructed, such as the "Connection Secure" text or the "Your connection to this site is not secure" message. For the implementation of root page memento detection in the Memento-aware Browser, the message "The page displayed is a memento captured on <datetime>" was added to the list of possible messages where the datetime from the response headers would be added on to the message. \begin{comment} if the user should be presented with the HTTPS secure lock icon or the "Not secure" warning. For the Memento-aware Browser implementation, this is also where it is decided if the user should be presented with a memento icon that alerts them that the webpage they are viewing is archived or contains archived content. \end{comment} \begin{figure*}[ht] \centering \includegraphics[width=0.86\textwidth, frame]{images/rootpage-loading-frontend.png} \caption{The sequence of classes involved in constructing the browser UI for a loaded webpage. These are the main classes that were edited to implement the UI for rootpage memento detection.} \label{fig:rootpage-loading-frontend} \end{figure*} \subsection{Displaying the Memento Icon} As of this point in the implementation, the front-end of the browser is aware of whether or not the root webpage is an archived page and also the datetime the page was archived, if applicable. The browser can also set the string variable containing the appropriate message about the page being archived. Now the user needs to be presented with the memento icon to be alerted that the page they are viewing is an archived page. This icon should also be clickable so that the user can open up an additional menu providing more details about the archived page they are viewing. To do this, an additional location icon was added to the \lstinline{LocationBarView} as shown in Listing \ref{lst:location_icon_memento} where \lstinline{location_icon_view} is the standard icon that can show the HTTPS secure lock icon and \lstinline{location_icon_memento} is the new memento icon. \begin{minipage}{\linewidth} \begin{lstlisting}[language=C++, caption=Add an additional icon to the location bar within \lstinline{location_bar_view.cc}., label=lst:location_icon_memento, ] auto location_icon_view = std::make_unique<LocationIconView>(font_list, this, this); auto location_icon_memento = std::make_unique<LocationIconView>(font_list, this, this); \end{lstlisting} \end{minipage} With this extra icon added, there needed to be a way within the icon class to differentiate which is the original icon and which is the memento icon. A boolean flag was added to the \lstinline{LocationIconView} class implementation to tell if that particular icon is the memento icon. By default, this flag is set to false to signify that the icon is the standard icon (Listing \ref{lst:location_icon_flag}). This flag will be used throughout the \lstinline{LocationIconView} class to set various attributes for the icon. For example, if the root webpage is an archived webpage then the datetime for the root page should be displayed next to the memento icon in \lstinline{YYYY-MM-DD} format. This means that if \lstinline{is_memento_icon_} is set to true, then the \lstinline{LocationIconView::ShouldShowText()} function should always return true since the icon should be displaying text. Additionally, the icon resource being displayed by the \lstinline{LocationIconView} should always be set to the memento icon resource since the standard icon is already displaying other necessary information such as the HTTPS secure lock icon. \begin{lstlisting}[language=C++, caption=Flag for whether or not the \lstinline{LocationIconView} object is the memento icon., label=lst:location_icon_flag, ] // Whether the icon is the original icon // or the new memento icon bool is_memento_icon_ = false; \end{lstlisting} Now the memento icon exists within the implementation but does not appear within the location bar by default. The memento icon should only be displayed if the webpage was determine to be a memento. In order for the memento icon \lstinline{LocationIconView} object to determine if it should appear in the location bar, it needs to be able to access the \lstinline{VisibleSecurityState}. To do this, the \lstinline{LocationIconView} class can use the \lstinline{LocationBarModelImpl} to call an \lstinline{IsMemento()} function (Listing \ref{lst:ismemento}). \begin{lstlisting}[float, language=C++, caption=The memento icon needs to reach the \lstinline{VisibleSecurityState} through the \lstinline{LocationBarModelImpl} to determine if it should be visible in the location bar., label=lst:ismemento, ] bool LocationIconView::ShouldShowMementoInfo() const { return delegate_->GetLocationBarModel()->IsMemento(); } \end{lstlisting} The \lstinline{LocationBarModelImpl} class has access to the security information (\lstinline{SecurityState}) that was set in the previous section and can return the \lstinline{memento_status} (Listing \ref{lst:ismemento-status}). \begin{lstlisting}[float, language=C++, caption=Return the \lstinline{memento_status} stored within the \lstinline{VisibleSecurityState}., label=lst:ismemento-status, ] bool LocationBarModelImpl::IsMemento() const { std::unique_ptr<security_state::VisibleSecurityState> visible_security_state = delegate_->GetVisibleSecurityState(); return visible_security_state->memento_status; } \end{lstlisting} With the \lstinline{VisibleSecurityState} information being used to determine if the memento icon should become visible, the root page memento detection and user interface is complete. As shown in Figure \ref{fig:rootpage-memento-UI}, when the user loads an archived webpage such as a memento from the Portuguese Web Archive, arquivo.pt \cite{portugeuse}, the browser is able to detect the Memento-Datetime HTTP response header and update the status of the page accordingly, making the memento icon and the datetime of the webpage easily visible for the user. \begin{figure*}[ht] \centering \includegraphics[width=0.86\textwidth, frame]{images/rootpage-memento-UI.png} \caption{The memento icon, datetime in \lstinline{YYYY-MM-DD} format, and popup with additional information for rootpage memento detection.} \label{fig:rootpage-memento-UI} \end{figure*} \subsection{Iframe Elements} \label{section:iframe-elements-section} The previous section described detecting when the root webpage returns the Memento-Datetime HTTP response header. As described in the Design section, mementos can appear as embedded elements, not just as the root webpage. There are three possibilities of this occurring to account for in the implementation: \begin{itemize} \item A memento is being displayed as an embedded element and is intended to make up the whole webpage. The best example of this is how the archives Trove and Perma.cc display Mementos. \item A live webpage is displaying a memento within an embedded element as an additional part of the page, but the whole webpage is not intended to be considered a memento. \item A live webpage is displaying multiple mementos within embedded elements as additional parts of the page, but the while webpage is not intended to be considered a memento. \end{itemize} The first case of detecting mementos such as the ones displayed by Trove and Perma.cc has the same end goal as detecting archived root webpages. The Memento-Datetime HTTP response header should be detected by the browser and the user should be presented with the memento icon and the datetime the page was archived. To implement this detection, properties of the root page detection implementation were altered or added on to. For example, as described in the Root Page section, the \lstinline{NavigationEntry} object updates with new information as each frame on the webpage is rendered. However, with the implementation in that section only the root page Memento-Datetime header is considered as a page parameter and passed to the \lstinline{NavigationEntry}. All resources already pass through the \lstinline{URLLoader} class and have any discovered Memento-Datetime header values associated with their corresponding \lstinline{URLResponseHead} object, but these are not immediately passed on to the page parameters since they involve more processing and consideration than the root page datetime. These datetimes associated with rendered frames on the page need to be considered as the entire webpage loads. In the case of a memento via Trove, the datetime from the iframe source should be associated with the entire webpage. Meaning that the \lstinline{VisibleSecurityState} will hold the datetime of the iframe source and the UI will appear the same as if the root page were a memento. The implementation for this is shown in Figure \ref{fig:iframe-memento-sequence}, where the path of the datetime value of frames within the root page is different from the path of the datetime value of the root page itself. The dotted gray line represents where the root page datetime would be passed from the root page \lstinline{RenderFrameHostImpl} object to the \lstinline{NavigationEntry}. For a datetime associated with a frame within the root page, the datetime cannot be immediately passed to the \lstinline{NavigationEntry} because it may or may not be associated with the root page. Instead, the datetime gets passed back up to the \lstinline{Navigator} and then down again to the \lstinline{NavigationController} where it can be determined if it should be taken as the datetime for the entire \lstinline{NavigationEntry}. If the \lstinline{NavigationController} determines that the iframe source is intended to act as the root webpage, the datetime is set as the datetime for the entire page. This is determined by the loading sequence of the nodes on the page. The root page processes last since all resources need to load first, meaning iframes will first be processed as subframes. If the iframe is processed a second time as the root page is processed, then the iframe is considered as a main piece of the root page and not just an element on the page. The \lstinline{switch} used to determine this is shown in Listing \ref{lst:processing-switch}. If the navigation type is \lstinline{NAVIGATION_TYPE_AUTO_SUBFRAME} then the element is an iframe. For iframe mementos that are to be considered as the root page, this \lstinline{switch} case is met the first time the frame is processed. The second time it is processed, the iframe is processed along with the root page. Essentially, if the datetime that was found was the only datetime and the embedded element was processed a second time with the root page, then its datetime will be considered as the datetime for the whole page. \begin{figure*}[ht] \centering \includegraphics[width=0.85\textwidth, frame]{images/iframe-memento-sequence.png} \caption{The loading sequence for embedded page elements that return the Memento-Datetime header.} \label{fig:iframe-memento-sequence} \end{figure*} \begin{lstlisting}[float, language=C++, caption=Switch statement that directs the processing of different navigation types where the processing of the \lstinline{AUTO_SUBFRAME} type is for iframe elements and used in the memento Detction implementation., label=lst:processing-switch, ] switch (details->type) { case NAVIGATION_TYPE_NEW_PAGE: ... case NAVIGATION_TYPE_EXISTING_PAGE: ... case NAVIGATION_TYPE_SAME_PAGE: ... case NAVIGATION_TYPE_NEW_SUBFRAME: ... case NAVIGATION_TYPE_AUTO_SUBFRAME: // iframes processed here case NAVIGATION_TYPE_NAV_IGNORE: // If a pending navigation was in progress, this canceled it. We should // discard it and make sure it is removed from the URL bar. After that, // there is nothing we can do with this navigation, so we just return to // the caller that nothing has happened. if (pending_entry_) DiscardNonCommittedEntries(); return false; case NAVIGATION_TYPE_UNKNOWN: NOTREACHED(); break; } \end{lstlisting} The fact that embedded elements are processed in this \lstinline{switch} case means we can collect any datetimes associated with elements into a list, completing the implementation for the other two possibilities for iframe mementos (Listing \ref{lst:collect-memento-dates}). Doing this allows a list of datetimes that exist on the page to be stored in the \lstinline{NavigationEntry}, meaning that the \lstinline{NavigationEntry} class now has a total of 3 variables regarding the memento information for the whole page. \begin{lstlisting}[float, language=C++, caption=For the \lstinline{AUTO_SUBFRAME} case collect any datetimes returned by embedded elements., label=lst:collect-memento-dates, ] case NAVIGATION_TYPE_AUTO_SUBFRAME: if (!RendererDidNavigateAutoSubframe(rfh, params, navigation_request)) { NavigationEntry* visible_entry = GetVisibleEntry(); if (datetime != "" && root != frame_tree_node && frame_tree_node->depth() < 2) { root->AddMementoDate(datetime); visible_entry->SetMementoDates(root->GetMementoDates()); } visible_entry->SetIterations(iterations); visible_entry->SetIsMixedMementoLiveWeb(mixed_memento_live_web); // We don't send a notification about auto-subframe PageState during // UpdateStateForFrame, since it looks like nothing has changed. Send // it here at commit time instead. NotifyEntryChanged(GetLastCommittedEntry()); return false; } break; \end{lstlisting} \begin{itemize} \item \lstinline{bool memento_info} - Flag for whether or not the root page is an archived page. \item \lstinline{string memento_datetime} - Datetime of the root page if it is a memento, otherwise set to "None". \item \lstinline{vector<string> memento_dates} - List of datetimes for archived elements on the page. \end{itemize} With the list of datetimes on the page, \lstinline{memento_dates}, set within the \lstinline{NavigationEntry}, this list can be passed to the \lstinline{VisibleSecurityState} and later used to construct the proper UI elements. Figure \ref{fig:one-iframe-ui} shows the UI for when a single iframe memento is displayed on a live web page and Figure \ref{fig:multi-iframe-ui} shows when multiple iframe mementos are displayed on a live webpage. \begin{figure*}[ht] \centering \includegraphics[width=0.8\textwidth, frame]{images/one-iframe-ui.png} \caption{When a single iframe memento is displayed on a live webpage the user is alerted that they are viewing a mix of live and archival content and the datetime of the archived resource is listed.} \label{fig:one-iframe-ui} \end{figure*} \begin{figure*}[ht] \centering \includegraphics[width=0.8\textwidth, frame]{images/multi-iframe-ui.png} \caption{When multiple iframe mementos are displayed on a live webpage the user is alerted that they are viewing a mix of live and archival content and the datetimes of all archived resources are listed.} \label{fig:multi-iframe-ui} \end{figure*} \subsection{Live Web within Mementos} When a page is loaded, all elements on the root page are loaded, including embedded elements within the embedded elements. Consider a webpage displaying an iframe, and that iframe displaying another iframe. The first iframe is a child node of the root page, and the second iframe is a child node of the first iframe. When the tree structure of the page is considered, this means the root page is the root node, the first iframe has a depth of 1, and the second iframe has a depth of 2. As shown in Figure \ref{fig:height-3-tree}, the nodes in the tree could also have any number of other child nodes aside from the ones described, meaning each frame could have any number of its own embedded frames. \begin{figure*}[ht] \centering \includegraphics[width=0.52\textwidth, frame]{images/height-3-tree.png} \caption{Root page with an embedded frame that contains another embedded frame.} \label{fig:height-3-tree} \end{figure*} The browser considers all these child nodes when constructing the \lstinline{VisibleSecurityState} for the whole page. As was shown in Figure \ref{fig:chromium-page-structure}, an insecure node at any depth of the tree could raise a security alert. The implementation described in Section \ref{section:iframe-elements-section} allows for datetimes returned by embedded elements at any depth to be collected. However, when a memento element is encountered its child nodes should be considered differently than regular child nodes. This is because children of the memento should return the same datetime as the memento itself. If the child nodes return no datetime, the memento is not correctly displaying content from the datetime it has returned and the live web is leaking into the archived element. This would create a webpage that never truly existed at the datetime returned by the resource, meaning that the user should be alerted that although it appears they are looking at an archived page from a certain date and time, the page never truly existed. Since the live web is being displayed within the archived page, this means a ``zombie" webpage is being displayed and the UI elements of the browser should alert the user that there is a mix of live and archival content on the page. This occurrence was described in Section \ref{section:embedded-elements-within-mementos} as one of the possible cases of embedded elements displaying within a memento. To account for this, the implementation of the \lstinline{Navigator} class was extended. If a currently processing child node of a memento does not have a datetime associated with its \lstinline{URLResponseHead} object, then the information for the root page needs to be updated so that it is known by the front-end that the page is displaying a mix of live and archival content (Listing \ref{lst:mixed-memento-flag}). Now the \lstinline{NavigationEntry} holds another variable for the memento information on the page: \begin{itemize} \item \lstinline{bool memento_info} - Flag for whether or not the root page is an archived page. \item \lstinline{string memento_datetime} - Datetime of the root page if it is a memento, otherwise set to "None". \item \lstinline{vector<string> memento_dates} - List of datetimes for archived elements on the page. \item \lstinline{bool mixed_memento_live_web} - Flag for whether or not the root page is a memento that is displaying the current web or contains a memento that is displaying the current web. \end{itemize} \begin{lstlisting}[language=C++, caption=When a child frame of a memento does not return a datetime set the \lstinline{mixed_memento_live_web} flag for the root., label=lst:mixed-memento-flag, ] // Set the mixed_memento_live_web flag root->SetIsMixedMementoLiveWeb(true); \end{lstlisting} With the \lstinline{mixed_memento_live_web} information for the root page appropriately updated, this can then be passed on to the \lstinline{NavigationEntry} so that the \lstinline{VisibleSecurityState} can have the information. Passing the information on to the \lstinline{VisibleSecurityState} allows the front-end to update as described in Figure \ref{fig:rootpage-loading-frontend}. The user will be presented with the memento icon as well as a message stating ``live + archival content" (Figure \ref{fig:mixed-memento-UI}). \begin{figure*}[ht] \centering \includegraphics[width=0.6\textwidth, frame]{images/mixed-memento-UI.png} \caption{Memento icon and message for when an archived element is displaying the live web.} \label{fig:mixed-memento-UI} \end{figure*} \subsection{Memento Detection Evaluation} Root page memento detection was tested by loading archived pages from a variety of web archives. Web archives may display mementos differently, however the browser should still react the same by updating the UI to display the memento icon and datetime. The root page memento detection was tested with the following web archives: \begin{itemize} \item Archive-It \cite{archive-it} \item Archive.today \cite{archivetoday} \item Australian Web Archive (Trove) \cite{trove} \item BAnQ \cite{banq} \item Bibliotheca Alexandrina Web Archive \cite{bibalex} \item Icelandic Web Archive \cite{icelandic} \item Internet Archive \cite{internet-archive} \item Library and Archives Canada \cite{canada} \item Library of Congress \cite{congress} \item National Records of Scotland \cite{scotland} \item Perma Archive (Perma.cc) \cite{permacc} \item Portugeuse Web Archive \cite{portugeuse} \item Stanford Web Archive \cite{stanford} \item UK National Archives Web Archive \cite{uk-national-webarchives} \item UK Parliament Web Archive \cite{uk-parliament} \item UK Web Archive \cite{uk-webarchive} \end{itemize} Each of these web archives was navigated with the browser so that memento detection could be tested. While navigating the archive, the memento icon would not display until an actual archived page was selected, meaning the memento detection was working as expected by not displaying the icon while selecting a memento from the archive and then bringing up the icon when the memento was actually displayed in the browser tab. The testing of these web archives is shown in Figure \ref{fig:testing-webarchives}. \begin{figure*}[ht] \centering \includegraphics[width=0.8\textwidth]{images/final-memento-detection.png} \caption{Testing memento detection by loading archived pages from a variety of web archives.} \label{fig:testing-webarchives} \end{figure*} \subsection{Bookmark as Archive Implementation} \label{section:bookmark-implementation} The bookmark as archive feature was implemented in a way that fits with the current Chromium bookmarking. Normally when the user wants to bookmark a page, they would click the star icon in the upper right. This action immediately generates a \lstinline{BookmarkNode} for that URL and brings up a menu where the user can edit the bookmark options such as the name and location, or click ``Remove" to get rid of the bookmark immediately after they added it (Figure \ref{fig:bookmark-popup}). To add the archive dropdown, the implementation of the original bookmarking dropdown for selecting the location was copied to get started. The class that implements the location dropdown is \lstinline{RecentlyUsedFoldersComboModel}, and the title comes from the fact that the dropdown orders the bookmark locations based off of when they were most recently used. The version of this class for the archive dropdown is \lstinline{WebArchiveComboModel}. Changes in the class implementation between the two dropdowns occurs in the constructor which prepares the possible options that could be selected in the dropdown. The dropdowns may only display permanent bookmark nodes such as the ``Bookmarks bar" or the ``Other bookmarks" options. With the copy of the \lstinline{RecentlyUsedFoldersComboModel} class created to create the Web archive dropdown, this new class could be called alongside the old class from the \lstinline{BookmarkBubbleView} class so the new dropdown could be added with the original (Listing \ref{lst:new-dropdown}). \begin{figure*}[ht] \centering \includegraphics[width=0.6\textwidth, frame]{images/bookmark-popup.png} \caption{The user can quickly hit the star icon to bookmark a URL. This creates a new bookmark node for that URL and brings up a popup where they can remove the bookmark or change its location.} \label{fig:bookmark-popup} \end{figure*} This creates a second dropdown that is identical to the first. To put different options into the archive dropdown, the new options first need to be added to the implementation as permanent nodes. The options for this new dropdown are public web archives that can easily accept a submission: The Internet Archive, Archive.Today \cite{archivetoday}, and Megalodon.jp \cite{megalodon}. To add these web archives as permanent nodes, they were first added to the \lstinline{BookmarkModel} class header and the \lstinline{BookmarkNode} class header. Additionally, the option of ``No archive" was added so this would also appear in the dropdown (Listings \ref{lst:add-permanent-nodes} and \ref{lst:add-new-nodes}). \begin{lstlisting}[float, language=C++, caption=Adding a second combobox to the add bookmark popup., label=lst:new-dropdown, ] // Add the original dropdown auto parent_folder_model = std::make_unique<RecentlyUsedFoldersComboModel>( model, model->GetMostRecentlyAddedUserNodeForURL(url_)); // Add the archive dropdown auto parent_folder_model2 = std::make_unique<WebArchiveComboModel>( archive_model, archive_model->GetMostRecentlyAddedUserNodeForURL(url_)); \end{lstlisting} \begin{lstlisting}[float, language=C++, caption=Create variables for the new permanent nodes within the bookmark model., label=lst:add-permanent-nodes, ] BookmarkPermanentNode* bookmark_bar_node_ = nullptr; BookmarkPermanentNode* no_archive_node_ = nullptr; // New permanent node BookmarkPermanentNode* archive_today_node_ = nullptr; // New permanent node BookmarkPermanentNode* internet_archive_node_ = nullptr; // New permanent node BookmarkPermanentNode* megalodon_node_ = nullptr; // New permanent node BookmarkPermanentNode* other_node_ = nullptr; BookmarkPermanentNode* mobile_node_ = nullptr; \end{lstlisting} \begin{lstlisting}[float, language=C++, caption=Add new permanent node types for the bookmark as archive dropdown., label=lst:add-new-nodes, ] enum Type { URL, FOLDER, BOOKMARK_BAR, NO_ARCHIVE, // New permanent node type ARCHIVE_TODAY, // New permanent node type INTERNET_ARCHIVE, // New permanent node type MEGALODON, // New permanent node type OTHER_NODE, MOBILE }; \end{lstlisting} In addition to adding the new possible permanent types, they also needed their own unique GUIDs. Variables for this were added to the \lstinline{BookmarkNode} class header so that they could be given values within the class implementation (Listing \ref{lst:add-new-guids}). \begin{lstlisting}[float, language=C++, caption=Add the variables for the archive node GUIDs and then values will be given within the \lstinline{BookmarkNode} class implementation., label=lst:add-new-guids, ] static const char kRootNodeGuid[]; static const char kBookmarkBarNodeGuid[]; static const char kNoArchiveNodeGuid[]; // New GUID static const char kArchiveTodayNodeGuid[]; // New GUID static const char kInternetArchiveNodeGuid[]; // New GUID static const char kMegalodonNodeGuid[]; // New GUID static const char kOtherBookmarksNodeGuid[]; static const char kMobileBookmarksNodeGuid[]; static const char kManagedNodeGuid[]; \end{lstlisting} In addition to adding the new permanent nodes to the \lstinline{BookmarkModel} and the \lstinline{BookmarkNode} classes, references to the nodes also need to be added in various places within the \lstinline{BookmarkCodec} class. The \lstinline{BookmarkCodec} class handles encoding and decoding the bookmark nodes from the Chromium config file that contains the JSON information of the user's bookmarks. Anywhere the \lstinline{bookmarks_bar} permanent node was accessed or altered in the \lstinline{BookmarkCodec} class, the new permanent nodes were added and accessed or altered in the same fashion. After adding the permanent nodes to the implementation, they could be pushed into the list of nodes displayed in the archive dropdown. This was done within the constructor for the archive dropdown in the \lstinline{WebArchiveComboModel} class (Listing \ref{lst:add-archive-options}). \begin{lstlisting}[float, language=C++, caption=Push the new permanent archive nodes onto the stack of options that will be displayed in the archive dropdown., label=lst:add-archive-options, ] items_.push_back(Item(model->no_archive_node(), Item::TYPE_NODE)); items_.push_back(Item(model->archive_today_node(), Item::TYPE_NODE)); items_.push_back(Item(model->megalodon_node(), Item::TYPE_NODE)); items_.push_back(Item(model->internet_archive_node(), Item::TYPE_NODE)); \end{lstlisting} At this point in the implementation of the bookmark as archive feature, the archive dropdown was successfully displayed within the bookmark popup and the user could view the different web archive options, which are ``None", ``Archive.today", ``Internet Archive", and ``Megalodon" (Figure \ref{fig:archive-dropdown}). Clicking ``Done" to save changes in the popup does not do anything with the selected option in the archive dropdown. To implement this, the \lstinline{MaybeChangeParent} function within the \lstinline{WebArchiveComboModel} must be modified. This function is called when the ``Done" button is clicked as it is intended to ``maybe change the parent" of the bookmark depending on if the user changed its folder location. The bookmark folder dropdown class will keep this implementation of possibly changing the parent folder of the bookmark, but the archive dropdown class will need to submit the bookmarked URL to the selected public web archive. To submit to a public web archive, the browser can use ArchiveNow \cite{jcdl18:archivenow}, a tool to push web resources into web archives \cite{archivenow}. ArchiveNow is created with Python3, so the Chromium C++ implementation needs to call Python code in order to use ArchiveNow. This can be done using \lstinline{system} to make a system command line call to the Python script. However, this \lstinline{system} call will need to be different depending on the operating system. The command can be changed according to the operating system using the same technique that is shown in Listing \ref{lst:os_spec_header_files}. To construct the command, we need: \begin{itemize} \item The location of the current working directory \item The title of the archive to submit to \item The URL of the webpage the user wants to submit \end{itemize} \begin{figure*}[ht] \centering \includegraphics[width=0.6\textwidth, frame]{images/bookmark-popup.png} \caption{The archive dropdown appears in the popup that displays when the user clicks the star icon to bookmark a page.} \label{fig:archive-dropdown} \end{figure*} The location of the current working directory can be found with the function shown in Listing \ref{lst:cur-directory}. \begin{lstlisting}[float, language=C++, caption=Function to get the path of the current working directory., label=lst:cur-directory, ] std::string get_current_dir() { char buff[FILENAME_MAX]; //create string buffer to hold path (void)GetCurrentDir( buff, FILENAME_MAX ); std::string current_working_dir(buff); return current_working_dir; } \end{lstlisting} The title of the selected archive is the title of the bookmark node the user selected in the dropdown, so this can be easily accessed in the following way: \begin{lstlisting}[language=C++, caption=Access the title of the bookmark node the user selected from the archive dropdown., label=lst:selected-archive, ] items_[selected_index].node->GetTitle() \end{lstlisting} The URL of the webpage the user wants to archive can be accessed through the bookmark node that was created when the user clicked the star icon: \begin{lstlisting}[language=C++, caption=Access the URL of the page the user just bookmarked and wants to archive., label=lst:archiving-url, ] node->url().spec() \end{lstlisting} On macOS and Linux, the final Python command to call the script that utilizes ArchiveNow will look as follows: \lstinline{python3} \lstinline{path/to/script.py} \lstinline{'Internet Archive'} \lstinline{'https://example.com'} \lstinline{ &}. The ``\&" runs the Python script in the background. The command will look similar on Windows but the method of telling the script to run in the background is different. With the script to utilize ArchiveNow called, the ``Done" button in the bookmark popup is successfully submitting the bookmarked URL to be archived. However, this does not mean the user can access the archived page from the browser at this point. The star icon creates a standard bookmark node for the URL when clicked, and if the user has not selected ``None" in the archive dropdown then when they click the done button, the URL will be submitted to their selected web archive. This implementation so far has not created a node for the archived version of the webpage. The node for the archived webpage is also unable to have the proper URL to the page in the Web archive until it is done archiving, and archiving a page can take anywhere from a few seconds to several minutes. To mitigate this, the original plan was to add an archived bookmark node that links to the original page and the title of the node says that it is still archiving, then update the URL of that node in the background with the URL to the archived page once it is done archiving. Adding the placeholder archived bookmark node was simple and was done as shown in Listing \ref{lst:add-archived-node} within the \lstinline{MaybeChangeParent} function of the \lstinline{WebArchiveComboModel} class. \begin{lstlisting}[language=C++, caption=Add a placeholder bookmark node for the archived page with the intent to update the URL with the URL to the page in the Web archive once it has completed archiving., label=lst:add-archived-node, ] bookmark_model_->AddURL(new_parent, new_parent->children().size(), base::UTF8ToUTF16(std::string("Archiving " + node->url().spec())), node->url().spec()); \end{lstlisting} In Listing \ref{lst:add-archived-node}, \lstinline{node}\lstinline{->}\lstinline{url()}\lstinline{.}\lstinline{spec()} is the string URL of the bookmarked page so, for example, the title of the archived page bookmark node becomes something like ``Archiving https://example.com". And the URL that the archive bookmark node leads to is initially set to be the URL to the original page. When the page has finished archiving, the following parts of the archive bookmark node need to be updated: \begin{itemize} \item The title of the node needs to be updated from ``Archiving <URL>" to something such as ``Archive.today example.com 2020-03-04". \item The URL that the node leads to needs to be updated to the URL to the archived version of the webpage. \end{itemize} However, updating the bookmark node proved to be very difficult. Any interaction with the bookmark nodes runs on the UI thread, meaning that the front-end updates immediately whenever something happens with the bookmarks. This is why when the user bookmarks a page and places it on the bookmarks bar, the browser immediately displays the new bookmark node on the bookmarks bar. Since all bookmark actions run on the UI thread and the archiving runs on the background thread since it can take several minutes, when the archiving is complete the bookmark node cannot be updated without crashing the browser with an invalid sequence error. Since this issue could not be quickly resolved, an alternative was implemented. Instead of creating a placeholder bookmark node that links to the actual webpage since it is expected to be updated, a bookmark node that links to that webpage in the archive at the time the ``Done" button was clicked is created. The Internet Archive's Wayback Machine and Archive.today both support URLs with 14 digit date strings. For example, the URL \href{https://web.archive.org/web/20100412125057/http://www.mitre.org/}{web.archive.org/web/20100412125057/http://www.mitre.org/} links to an archived webpage in the Wayback Machine. If you change the datestring to one for which the Wayback Machine does not have a memento, it will redirect to the closest datetime that they do have. Meaning if you put the datestring for the current date and time, it will take you to the latest memento for that webpage. The same is true for Archive.today, if you try to load the same URL for a memento of mitre.org but change the hostname to archive.is, it will take you to Archive.today's closest memento to that date: \href{https://archive.is/20100412125057/http://www.mitre.org/}{https://archive.is/20100412125057/http://www.mitre.org/}. That URL will actually load a memento of mitre.org in Archive.today from 2012, so the closest memento Archive.today has of mitre.org to 2010 is from 2012 (Figure \ref{fig:mitre-2012}). \begin{figure*}[ht] \centering \includegraphics[width=0.9\textwidth, frame]{images/mitre-2012.png} \caption{If you try to load a memento of mitre.org from Archive.today from the datetime 20100412125057, it will redirect to a memento from 2012 since that is the closest date Archive.today has to the datetime given.} \label{fig:mitre-2012} \end{figure*} Since the Internet Archive's Wayback Machine and Archive.today support the 14 digit datetime format and will redirect to the memento closest to that date, an archive bookmark node can be created with a working URL without waiting for the page to be archived. The URL will be constructed of the hostname for the chosen web archive, the 14 digit datetime string for the time the user submitted bookmark edits, and the URL of the page the user wanted to archive. This constructed URL will take the user to the memento in the archive closest to that date, which initially will be the latest memento for that URL, but once the page is finished archiving again from the script initiated by clicking the ``Done" button, the datetime in the URL will be closest to that memento so that is where it will redirect to. For example, if on March 4th, 2021 at 3:00pm the user clicks the ``Done" button to submit the website example.com to be archived by the Wayback Machine, the 14 digit datetime string will be 20210304030000. The URL for the archive bookmark node will be https://web.archive.org/web/20210304030000/https://example.com. If the webpage finishes archiving two minutes later at 3:02pm, then the true 14 digit datetime string for the archived page will be 20210304030200. Since the datetimes only have a two minute difference, the datetime for when the user clicked the button at exactly 3:00pm will redirect to the memento that was archived at 3:02pm. This workaround cannot be done for Megalodon.jp since it does not redirect to the nearest datetime. All archived page URLs are additionally printed to the file \lstinline{archive_urls.txt} so that they can be accessed once the page finishes archiving even though they cannot currently be added to the browser bookmarks. In addition to creating the original bookmark node and the archive bookmark node, the browser also needs to place each node into a location that makes sense. When the user initially clicks the star icon to bookmark a page, the browser will create the node for that page and place it into the last bookmark folder used. Typically, this is the bookmarks bar. The archive bookmark could be placed into this location alongside the original bookmark, but the user has the option to archive the page again from the edit bookmark popup, thus creating more archive bookmark nodes. The original bookmark and the archive bookmark nodes should, by default, be placed into a location together since they all lead to the same webpage. That way the user can easily choose if they want the live page or one of the archived versions they saved. However since the user can archive a page multiple times and create multiple archive bookmark nodes, all these nodes must be managed and kept together efficiently. To implement this, when the user chooses to archive a webpage when they bookmark it, when the archive bookmark node is created a folder will also be created. The original bookmark node will be placed into this folder alongside the archive bookmark node. The title of the folder will be the URL to the live webpage. This way, the user will see their single bookmark folder node on the bookmarks bar (or wherever they chose to save the original bookmark) and when they click it they will see a dropdown showing all the possible versions of that page they saved (Figure \ref{fig:archived-bookmark-folder}). \begin{figure*}[ht] \centering \includegraphics[width=0.8\textwidth, frame]{images/archived-bookmark-folder.png} \caption{When the user archives a page they have bookmarked, a folder gets created to hold all the bookmark nodes that lead to that webpage, either live or archived.} \label{fig:archived-bookmark-folder} \end{figure*} To accomplish this, the implementation within the \lstinline{MaybeChangeParent} function of the \lstinline{WebArchiveComboModel} class was expanded (Listing \ref{lst:add-bookmark-folder}). \begin{lstlisting}[language=C++, caption=Add a bookmark folder node to store the nodes for a specific webpage., label=lst:add-bookmark-folder, ] const BookmarkNode* new_parent = bookmark_model_->AddFolder( node->parent(), node->parent()->children().size(), base::UTF8ToUTF16(node->url().spec())); bookmark_model_->Move(node, new_parent, new_parent->children().size()); \end{lstlisting} In the listing, a new folder is added and set as the \lstinline{new_parent} variable. Next, the original bookmark node is moved into this \lstinline{new_parent}, meaning that the newly created folder becomes the parent of the original bookmark node. The only issue with this implementation is that a new folder will be created each subsequent time that webpage is archived. To mitigate this, the code can be wrapped in a condition where it only creates that new folder if it has not already been created and set as the parent of the original bookmark node (Listing \ref{lst:conditionally-add-bookmark-folder}). \begin{minipage}{\linewidth} \begin{lstlisting}[language=C++, caption=Wrap the bookmark folder node in a condition so that it does not execute if the page is being archived again., label=lst:conditionally-add-bookmark-folder, ] if (items_[selected_index].node->GetTitle() != base::UTF8ToUTF16(std::string("None"))) { const BookmarkNode* new_parent = bookmark_model_->AddFolder( node->parent(), node->parent()->children().size(), base::UTF8ToUTF16(node->url().spec())); bookmark_model_->Move(node, new_parent, new_parent->children().size()); } \end{lstlisting} \end{minipage} \subsection{Bookmark as Archive Evaluation} Since the workaround for the bookmark as archive URLs does not link to the exact mementos and instead depends on being redirected to the nearest, there is a chance it will link to a memento that is not the one created from the user's submission. Since archiving takes some time, it is also possible that there is a reasonable offset that can be added to the 14 digit date string of when the user clicks the ``Done" button. For example, if a page typically takes 60 seconds to archive, then 60 seconds could be added to the 14 digit date string in order to increase the chance the URL will redirect to the exact correct memento. To evaluate this, 17 timemaps were analyzed to see how often the mementos in the timemap are very close together. The 17 timemaps were of popular news websites \cite{nwala2020365} and were chosen since they are likely to have been archived frequently. Additionally, only the portions of the timemaps from 2021 were considered since earlier portions of the timemaps (such as from around 2000) are likely to not have many mementos. For each timemap, a datetime for each second between January 1, 2021 and the current date was generated and the memento closest to that datetime was found. Then, an offset of 30 seconds was added to the datetime and the memento closest to that new datetime was found. Next, these two mementos were compared. It was found that 99.1\% of the time the memento closest to the random datetime and the memento closest to that random datetime plus 30 seconds were the same. A 60 second offset was tested and the mementos were the same 98.4\% of the time. With a 120 second offset, the accuracy was 96.9\%. In this evaluation, only seconds were considered since the memento datetimes have second granularity. Since webpages always take some time to archive, if we assume this to be around 60 seconds, then adding a 60 second offset to the datetime used to construct the URL should lead to the correct memento about 98\% of the time. The timeline for mementos and offsets is shown in Figure \ref{fig:bookmark-eval}. There are four possible cases, two that are successful (Case 1 and Case 3) and two that result in linking to an incorrect memento (Case 2 and Case 4). In the figure, \textsf{M1} stands for the latest memento before the bookmark is created, \textsf{t1} is the time the bookmark request is made, \textsf{t1+x} is the datetime used for the bookmark where \textsf{x} is the offset value we are testing, and \textsf{MC} is when the bookmark memento is actually created. The difference between Case 1 and Case 2 is how long it takes for the archive to create the new memento. Just looking at Cases 1 and 2, we might think it best to set the offset \textsf{x} to be large. But the cases (Cases 3 and 4) where another memento, \textsf{M2}, is created in the archive (by some other process) after the bookmarking process has started must also be considered. If the offset results in the bookmark node URL being too close to \textsf{M2} (Case 4), then the bookmark node will link to the incorrect memento. The archive will redirect to the closest existing datetime, so we want to set \textsf{t1+x} so that in most cases, it is closer to \textsf{MC} than any other mementos, without knowing exactly how long the archive will take to capture the page or what other mementos exist. The current feature implementation applies no offset, however from this evaluation a 30 second or 60 second offset would be reasonable to add. \begin{figure*}[ht] \centering \includegraphics[width=0.7\textwidth, trim=0 0 375 0, clip, frame]{images/bookmark-eval-3.png} \caption{Four cases of timelines for mementos and offsets for ``bookmark as archive".} \label{fig:bookmark-eval} \end{figure*} \section{Usage} This section first describes the browser from the user's perspective and what memento-aware features they can use. It also describes what partially implemented features can be seen from the UI. Also described is how to run the browser from the source code on GitHub \cite{repo} on both Linux and Windows, as well as instructions for switching between generating a debug build or a release build. The source code for the Memento-aware Browser on GitHub can only be built and run on Linux and Windows. \subsection{System Walkthrough} \label{section:system-walkthrough} The main feature of the Memento-aware Browser is memento detection. If the user visits a webpage that is archived, they should expect to see the memento icon plus the datetime in YYYY-MM-DD format to appear to the left of where the HTTPS secure lock icon appears. Figure \ref{fig:final-memento-detection} shows an archived page with the memento icon visible and the expanded popup. \begin{figure*}[ht] \centering \includegraphics[width=0.7\textwidth, frame]{images/memento-detection-UX.png} \caption{If the root page is considered to be an archived page, the memento icon will display along with the datetime in YYYY-MM-DD format and the user may click the icon to bring up a popup with more information.} \label{fig:final-memento-detection} \end{figure*} Additionally, just like the user can click the HTTPS secure lock icon to bring up the connection popup for more connection security information, they can also click on the memento icon for more information about the archived page. The current browser implementation brings up a popup very similar to the connection popup, except the text is about the webpage being archived. As shown in the left hand side of Figure \ref{fig:memento-popups}, the popup states that the user is currently viewing a memento, and that the page being displayed by the current browser tab is an archived page that was captured at a certain date and time. Additionally, there are two buttons on the popup. The first button says ``About this Memento" and the second is the standard ``Site settings" button that also appears on the connection popup. The ``About this Memento" button leads to a currently blank popup (right hand side of Figure \ref{fig:memento-popups}) that is further discussed in \ref{section:future-enhancements-section}. As for the ``Site settings" button, it is there by default with the popup and was not removed when the connection popup was copied to create the memento info popup since it could be used in the future for settings regarding archived content. \begin{figure*}[ht] \centering \begin{tabular}{cc} \includegraphics[width=70mm]{images/viewing-a-memento.png} & \includegraphics[width=90mm]{images/about-memento-popup.png} \\ (a) memento info popup & (b) About this memento blank popup \\ \end{tabular} \caption{The memento info popup appears when the user clicks the memento icon and the blank popup appears when they click ``About this memento"} \label{fig:memento-popups} \end{figure*} The next part of the memento detection feature that the user can experience is the browser detecting archived content within a live webpage. For this, two example webpages are available where the first contains one memento (\url{https://www.cs.odu.edu/~amabe/oneiframe.html}) and the second contains three (\url{https://www.cs.odu.edu/~amabe/test.html}). The first example webpage is a live webpage with an iframe element that is displaying a memento. This scenario is supposed to simulate a live webpage displaying archived content as a part of the page, and not the whole page. For example, an article on the live web may want to show an embedded memento so that their readers can see the archived content they are referring to. The second example webpage is similar, except there are three embedded mementos instead of one. When the user visits the first example webpage they can expect the Memento-aware Browser to display the memento icon, but instead of displaying a datetime next to the icon, it will show the message ``Mixed archival content". When the user visits the second example webpage, they can expect the browser to display the memento icon and message just as it did for the first example page, but the popup will list three datetimes, one for each archived frame on the page. This is shown in Figure \ref{fig:three-mementos-UI}. \begin{figure*}[ht] \centering \includegraphics[width=0.6\textwidth, frame]{images/three-mementos-UI.png} \caption{The popup will display all three datetimes that exist on the page.} \label{fig:three-mementos-UI} \end{figure*} The final piece of the memento detection feature the user can expect is the detection of ``zombies" \cite{zombies}, or archived pages that are displaying content from the live web. With this feature, when an archived page is visited, the browser displays the memento icon as usual and detects if any of the frames being displayed by that archived page are from the live web. If live content is detected in the archived page, instead of showing the datetime by the memento icon, the browser will display the message ``Memento + live content". Additionally, the popup that appears when the user clicks the memento icon will display the same message as usual when an archived page is displayed (Figure \ref{fig:memento-plus-live}). \begin{figure*}[ht] \centering \includegraphics[width=0.6\textwidth, frame]{images/memento-plus-live.png} \caption{The browser alerts the user that the archived page they are viewing is displaying content from the live web but also allows them to see the datetime the archived page is supposed to be from.} \label{fig:memento-plus-live} \end{figure*} In addition to the memento detection, the user may also use the bookmark as archive feature. When bookmarking a page through the star icon, the edit bookmark popup will appear which allows the user to change the name of the bookmark, change the location, or submit that page to a public web archive. To archive the page from this edit bookmark popup, the user may select a web archive from the archive dropdown and then either click the ``Done" button to submit or click out of the window. To exit the window without applying edits made to the bookmark, the user must click ``Cancel" to cancel all edits or ``Remove" to remove the bookmark entirely. Essentially, all edits made in the edit bookmark popup are immediately applied when the user clicks out of the window or clicks ``Done" which is standard in the Chromium implementation. With the archive dropdown being present on top of the original implementation, this means that the page is immediately submitted to the selected archive when the user clicks out of the popup or clicks the ``Done" button. Figure \ref{fig:archive-options} shows the edit bookmark popup with the newly added archive dropdown and its options. The available archives are the Internet Archive, Archive.today, and Megalodon.jp. \begin{figure*}[ht] \centering \includegraphics[width=0.58\textwidth, frame]{images/archive-options.png} \caption{The bookmark as archive option appears in the edit bookmark popup.} \label{fig:archive-options} \end{figure*} When the edit bookmark popup changes are applied and the user has selected a web archive, they can expect the browser to create a bookmark node folder for that webpage and place the original bookmark for the page into that folder. The browser will also create an archive bookmark node and place it into the folder alongside the original bookmark node. If the user had selected either the Internet Archive or Archive.today, then the archive bookmark node that is created will link to an archived page in the chosen web archive from the datetime that the bookmark edits were applied. If the user has selected Megalodon.jp, then the archive bookmark node will link to the live webpage. In the background, the user can expect the browser to be archiving the webpage. When it is finished archiving, the URL will be added to the file \lstinline{archive_urls.txt}. This way the user can always access the URL to the archived page even though the browser cannot update the archive bookmark node. The reason as to why the URL is set differently depending on the archive and the archive bookmark node cannot be updated is explained in Section \ref{section:bookmark-implementation}. To walk through the example shown in Figure \ref{fig:bookmark-as-archive-ux}, say the user wanted to use the bookmark as archive feature to submit example.com to Archive.today on March 10, 2021 at 4:18pm. The user can expect the browser to immediately create a bookmark node folder for the webpage and place two bookmark nodes in the folder. The first bookmark node in the folder would be the standard bookmark that leads to the live webpage and the second bookmark node would be the one for the archived version of the page. The URL for the archive bookmark node would have been constructed using the method described in Section \ref{section:bookmark-implementation} and would initially redirect to Archive.today's latest memento of example.com, which at that time would be from February 28, 2021. Eventually, when the page is done archiving a few minutes later, the archive bookmark node URL will redirect to the memento of example.com that was archived a few minutes after the user submitted it with the bookmark as archive feature at 4:18pm. Additionally, the user can also expect the URL to the memento to appear in the \lstinline{archive_urls.txt} file. Rather than redirecting, this URL will lead directly to the memento of example.com that was archived a few minutes after the user submitted it with the bookmark as archive feature at 4:18pm. \begin{figure*}[ht] \centering \includegraphics[width=0.99\textwidth, frame]{images/bookmark-as-archive-ux.png} \caption{The archive bookmark node will redirect to the memento in the archive that is closest to the datetime the page was submitted from the edit bookmark popup. Initially, this may be an older memento.} \label{fig:bookmark-as-archive-ux} \end{figure*} \subsection{Running from Source} The Memento-aware Browser can be run from the source code on GitHub on either Linux or Windows. The README file of the GitHub repository walks through running the code on either operating system. There are instructions for running on macOS, however these do not work due to missing dependencies and the issue could be fixed in the future. When running the source code, by default a debug version will be generated. A release version can be built by adding \lstinline{is_debug=false} to the file \lstinline{out/Default/args.gn}. In order to run from source, it is also necessary to download \lstinline{depot_tools} as described in Section \ref{section:working-with-chromium}. It is recommended to have about 100GB of free space before attempting to build the browser from source. Additionally, it is recommended to have at least 8GB of RAM to build and run the browser from source. \subsection{Release Builds} The GitHub repository includes release builds for Linux and Windows \cite{repo}. In each of these releases, the memento detection feature is available without any additional dependencies. As for the bookmark as archive feature, ArchiveNow \cite{archivenow} and its dependencies will need to be installed as instructed on the release page \cite{releases}. \section{Future Enhancements} \label{section:future-enhancements-section} This section outlines additional features and enhancements to the Memento-aware Browser. The features mentioned can be built off of the existing features, specifically the memento detection feature since it stores the memento datetimes for all resources but is currently only considering the datetimes returned by the main URL. \subsection{``About this Memento" Popup} \label{section:about-this-memento-popup} The current memento detection provides information about whether or not the page displayed is an archived page as well as the datetime the page was archived. However, there is more information about the archived page that could be useful to the user. An example of this is a feature provided by the Internet Archive's Wayback Machine. The feature is a drop down labelled ``About this capture" that allows the user to see a list of all the archived resources and their respective datetimes that make up the entire webpage (Figure \ref{fig:about-this-capture}). Since the Memento-aware Browser stores the Memento-Datetime value for all resources as they are loaded as explained in Section \ref{section:detecting-memento-header}, implementing this feature would mostly consist of completing the front-end. Already in the browser is a button that would bring up the popup to display the information. The button is the ``About this memento" option that was detailed in Section \ref{section:system-walkthrough} and appears when the user clicks on the memento icon to get more information about the memento. Currently the button brings up an empty popup that was made from the collected cookies popup. If the feature were to be implemented, this popup would display a list of all the resources and their respective datetimes that make up the archived page. Additionally, it would be ideal if the user could click on a listing for a resource within the popup and that resource would highlight somehow within the browser tab. For example, if the user saw a listing for an image in the popup and they clicked that listing, the browser would highlight that image within the browser tab with a red box. \begin{figure*}[ht] \centering \includegraphics[width=0.99\textwidth, frame]{images/about-this-capture.jpg} \caption{The Internet Archive's Wayback Machine allows the user to see a list of archived resources that make up the archived page.} \label{fig:about-this-capture} \end{figure*} \subsection{List of Known Web Archives} The current implementation of the memento detection just reacts to the presence of the Memento-Datetime HTTP response header and also the level the frame that returned the header exists within the page structure tree. It is possible for a live webpage to return the Memento-Datetime header. A possible way to account for this is for the browser to only acknowledge the header if the frame that is returning it is from a known web archive. This would involve the browser having a hard coded list of known and accepted web archives. Ideally, this list would be a part of the browser settings so that individual users could add or remove archives from the list stored in their local browser configuration. Initially the browser would come with its own list and then the user could customize it from there. A list of some possible web archives for the browser to recognize initially is the list of web archives the memento detection has been tested with. This list can be found within the docs section of the Memento-aware Browser GitHub repository \cite{repo-docs}. \subsection{Temporal Coherence} Another addition to the memento detection would be to consider the temporal coherence of the memento \cite{ht15:as-presented}. If you consider the HTML webpage and all the resources such as CSS and images required for the full and complete page, these pieces together can be defined as a composite memento \cite{ainsworth:composite:tr}. Just like the Chromium page structure defined in Section \ref{section:chromium-page-structure}, composite mementos can be thought of as a tree structure where the root resource (the main HTML webpage) has other resources connected to it. Consider an example of a weather report that was captured on December 9, 2004 at the time 19:09:26 GMT (Figure \ref{fig:weather-report}). The Memento-Datetime value returned by the page would be 2004-12-09 19:09:26 GMT, however if you look closer at the it would appear that the content does not make sense. The description of Current Conditions within the weather report states light drizzle, yet the rader shows no clouds. Looking at the datetimes for the page resources, it can be found that the radar image was captured 9 months later than the webpage itself, meaning that this weather report never actually existed on the live web. When something like this occurs, it would be ideal for the browser to alert the user that the page they are viewing is temporally incoherent and that although it appears they are looking at an archived webpage from a certain date and time, that webpage never actually existed as it is being presented. \begin{figure*}[ht] \centering \includegraphics[width=0.7\textwidth, frame]{images/weather-report.png} \caption{Weather report archived on 2004-12-09 19:09:26 GMT \cite{ht15:as-presented}} \label{fig:weather-report} \end{figure*} As mentioned in Section \ref{section:about-this-memento-popup}, the Memento-aware Browser already stores the Memento-Datetime value for any resources that return it. Since the browser already has the necessary information about any composite mementos that are loaded, then the main part of the implementation for this feature would be classifying the different scenarios and presenting the appropriate UI elements. A suggested method to alert the user of a temporally incoherent memento is to display an infobar \cite{infobar}. Infobars appear in the Chromium browser for a variety of reasons. Some examples are shown in Figure \ref{fig:infobar-examples}. Many infobars displayed by the browser also have a ``Dismiss" button. So the user is alerted by the infobar appearing and once they have read and noted the information they can dismiss the popup. An infobar would be a great way to alert the user that the memento they are viewing is temporally incoherent since the memento icon will already be displaying datetime information. The user could also dismiss the warning since in some scenarios it may not matter to them that the archived page is temporally incoherent so they may want to view the page without any warnings. Figure \ref{fig:infobar-UI} shows an example of how this popup might look. The use of an infobar for temporal coherence information would also allow for different colored popups to be used depending on the situation, or even different options on the bar in addition to ``Dismiss". Another form of alert that the Chromium browser offers is a dialog box. However, a dialog box is not an optimal method of giving the user this information since dialog boxes should only be used if we want to block the user from doing something unless they take action, or the box appears in response to a user action. We do not want to force the user to acknowledge the temporal incoherence warning since it is not information that requires immediate action. Rather, it is additional information being offered to the user so they may keep this in mind when viewing the archived page. \begin{figure*}[ht] \centering \includegraphics[width=0.74\textwidth, frame]{images/infobar-examples.png} \caption{Examples of infobar alerts that appear in the Chromium browser (\url{https://www.chromium.org/user-experience/infobars}).} \label{fig:infobar-examples} \end{figure*} \begin{figure*}[ht] \centering \includegraphics[width=0.9\textwidth, frame]{images/infobar-UI.png} \caption{Mock up of using an infobar to display an alert about temporal incoherence.} \label{fig:infobar-UI} \end{figure*} \section{Conclusion} A successful proof of concept Memento-aware Browser was created by extending Google's open source web browser Chromium \cite{chromium-download}. Throughout the process of adding the memento-aware features to the Chromium code base, it was found that these features work well within the native implementation. Detecting the Memento-Datetime HTTP response header proved to not be too difficult since the browser already needs to parse all response headers; it was mostly a matter of accounting for this header where the headers were already being parsed. Adding the basic memento detection described in Section \ref{section:basic-memento-detection} also worked with the native browser implementation very well. The memento detection features became more complex when different cases were accounted for, like the iframe memento case such as the one that can be see when viewing archived pages from Trove \cite{trove}. However once fully implemented, the user experience of the memento detection works well and makes sense when compared to the user experience when using the standard Chromium browser. The bookmark as archive feature implementation also fit into the native implementation well. However, there were challenges encountered while implementing this feature. It was found that since all bookmark actions run on the UI thread, bookmarks could not be updated on completion of the archiving that was running on the background thread. A workaround was implemented that utilized the fact that the Internet Archive and Archive.today redirect links that contain datetimes they do not have mementos for to the memento closest to that datetime. The Memento-aware Browser source code is on GitHub \cite{repo} and can be built and run on Linux and Windows. Additionally, there are release builds for Linux and Windows attached to the GitHub repository. There are some existing issues with the browser, mainly in that the bookmark as archive feature is not considered complete. The implemented workaround for this feature only works for when the user selects the Internet Archive or Archive.today. Megalodon.jp \cite{megalodon} does not redirect links to the closest datetime so the workaround was not implemented for this archive. The memento detection feature works well but there likely are some corner cases that have not been accounted for. There is much room for improvement and additional features in the browser, however the current version serves as a great proof of concept for the user experience and the ability to work these features into the native browser implementation. \clearpage

23f62d0f8b2222faec3076959cf3143c7fa42e83

\section{Introduction} \IEEEPARstart{T}{he} problem of determining the optimal graph representation of a given dataset has been considered for various tasks in different fields, ranging from signal processing to machine learning, yet remains largely unresolved. Hyperspectral image (HSI) data, with its rich and descriptive spatial and spectral information contained in several hundred bands, encapsulates several layers of dependencies between the high-dimensional pixels and their corresponding labels \cite{guide}, and a plethora of methods have sought to exploit these under different modelling assumptions.\\ In hyperspectral image classification, a classifier is sought to assign a class label to each pixel, in light of arising difficulties including high spectral dimensionality, large spatial variability and limited availability of labels. While the majority of frameworks involve a supervised classifier, due to the time and cost associated with obtaining labelled samples, semi-supervised learning (SSL) has experienced a rapid development by tackling the small sample problem through effective exploitation of both labeled and unlabelled samples \cite{ssl}. \\ Multiple views or features of data can provide rich insights into the underlying data structure, while exploiting diverse information and thus gaining robustness for subsequent tasks. The identification of suitable features, however, represents a problem in itself. Inspired by the field of multi-view clustering \cite{Nie1} and in an effort to incorporate more sophisticated model priors, we adopt the perspective of learning a graph from multiple data features (or views) in order to effectively capture the complexity of hyperspectral data, while further integrating the label information into the graph so as to exploit prior knowledge of label dependencies. We further resort to superpixel segmentation of the data in order to effectively reduce the complexity of the classification task, while simultaneously capturing a first level of spectral-spatial dependencies through the clustering into local homogeneous regions. \\ In this paper, we introduce a novel semi-supervised framework for HSI classification which involves the design of a graph classification function which is smooth with respect to both the intrinsic structure of the data, as described via superpixel features, as well as the label space. The proposed graph learning framework is special due to its edge-efficient analytic solution, known to satisfy graph-optimality constraints, and ability to incorporate multiple features, as well as due to its optimization of both the graph and classification function, resulting in pseudo-labels. In addition, in light of the range of parameters required to tune a multi-feature superpixel graph, we propose a variation of the framework which, instead of incorporating pre-set feature weights, learns them by imposing an additional smoothness functional. We summarize the main contributions as follows: \begin{itemize} \item We extend multi-view graph learning to the domain of superpixels and HSI data, with tunable pseudo-label generation incorporated into an updateable graph. In particular, we employ an initial soft graph on which labels are firstly propagated among nearest neighbours to generate pseudo-label features, which are subsequently utilized to inform an improved graph. \item We propose a pseudo-label-guided framework for HSI feature selection, weighting and subsequent graph construction, enhanced by dynamic pseudo-label-features. To the best of our knowledge, the issue of feature contribution on a multi-feature graph for HSI-data has not yet been tackled, beyond a simple parameter search. \item We extensively validate our proposed approaches on the basis of three benchmark datasets and demonstrate their superiority with respect to comparable state-of the-art approaches. \end{itemize} \noindent The proposed framework facilitates edge-efficient graph and label learning, while flexibly incorporating multiple features which capture spectral-, spatial- and label-dependencies within the given data. The remainder of this paper is organized as follows: Section II discusses related work, Section III explores the preliminaries, and details of the proposed methods are stated in Section IV. Section V presents the experimental results of our methods in comparison with state-of-the-art approaches, and Section VI contains concluding remarks. \begin{figure*}[h] \centering \includegraphics[width=5.5in]{w10} \caption{Main Workflow: The proposed MGL and PMGL methods are mainly composed of spectral dimensionality reduction, superpixel segmentation, superpixel feature extraction and superpixel label regularization, followed by a pseudo-label driven graph learning stage, and culminating in the final classification result via label propagation. } \label{fig:w} \end{figure*} \section{Related Work} Works in SSL can be categorized into generative models, which employ probabilistic generative mixture models, co-and self-training methods, low-density separation methods, which seek a decision boundary through low density regions, and graph-based methods \cite{ssl}. Popular modelling assumptions and concepts have included spectral clustering, data manifolds and local-global consistency \cite{lgc}, according to which nearby points and points residing on the same cluster or manifold are likely to have the same label \cite{ssl}. In the context of HSI data, earlier approaches comprise kernel methods such as the purely spectral-based SVM \cite{svm} and spectral-spatial multiple kernel learning \cite{sp}. Feature extraction methods, such as \cite{IFRF}, \cite{LBP}, \cite{lcmr}, have sought to characterize a lower dimensional subspace which best captures the spectral-spatial information of the data. Further, classical signal processing concepts such as sparse representation, low-rank, and wavelet analysis have been notably incorporated into SSL methods \cite{lowrank}, \cite{dict}, \cite{guide}.\\ More recent methods have gained in complexity by generating multi-stage workflows with different pre- and post-processing levels as well as by combining the strengths of different classifiers in an effort to create more sophisticated models which exploit multiple dependencies of HSI data and counteract the small sample problem through a gradual learning process \cite{guide}. \\ Graph-based SSL methods have become increasingly popular due to the superior modelling capabilities of graphs, particularly as a means to counteract the limited amount of labels available, and have included pixel-based \cite{gssl} as well as superpixel-based graph constructions \cite{sp2}, \cite{rlf}, \cite{sperw}, ranging from end-to-end approaches, that utilize graphs for both data modelling and label propagation, to hybrid approaches. We note that many existing methods, such as EPF \cite{epf}, which makes use of the bilateral filter, exhibit underlying graph-like qualities, while not directly or only in part employing graphs, and as a result have exhibited superior performance. In addition, approaches such as \cite{hong}, have sought to incorporate label information early on in the workflow, known as \textit{pseudo-labelling}, which is progressively refined, in an effort to inform data modelling; nevertheless, the graph construction is not necessarily analytic. While identification of the correct model graph is essential for SSL \cite{graph}, extensive parameter analysis is still inevitable in order to study their influence, as performance is strongly affected by all components of the graph; nevertheless, previous studies have generally not found explorable patterns \cite{graph}, and stability is preferred over excellence in a narrow parameter range.\\ A preceding body of work on multi-view clustering has sought to unify the tasks of data similarity learning (i.e. graph construction) and clustering/label propagation in a joint optimization framework \cite{Nie1}, \cite{multigraph}, \cite{auto}, which, to the best of our knowledge, has not been fully investigated for the joint challenge of HSI data and SSL. Graph learning and label propagation are separate tasks but recent efforts have opted to combine them, thereby updating and incorporating label information into the graph, under the driving assumption that a single static graph is not sufficient to solve the entire SSL task successfully. Nevertheless, and not least of all for complex data such as HSI, the issue of propagating error noise into each task needs to be addressed. There are multiple ways in which the label information can be incorporated into the graph, which include space fusion approaches, i.e. the graph and label space are fused via the addition of a label correlation matrix \cite{DLP} or the removal of differently labelled edges \cite{RMGT}, and implicit approaches, which instead consider the graph as a function of the distances between labels \cite{Nie1}.\\ Deep learning (DL) methods have penetrated graph-based classification in the form of Graph Convolutional Networks (GCNs), which usually require a given pre-constructed graph, such as the spectral-spatial GCN \cite{gcn1}, and have been employed to extract features automatically \cite{zhu}. Under the assumption that the mapping from the feature to the label space is sufficiently smooth, the theoretical relation between label propagation (LP) and GCNs, as instances which propagate labels and features respectively, has been notably shown to satisfy a smoothness inequality \cite{gcnlpa}. Nevertheless, DL-, and non-DL methods alike, ultimately rely on the optimality of the graph while often performing worse with limited labels. Neural networks have further been employed to learn graphs from scratch (see e.g. \cite{gcn2}), which inevitably comes at a high computational cost, however, for complex and rich datasets such as HSI the formulation of sophisticated (graph) model priors is paramount to the successful performance of any classifier. \\ Graphs ultimately capture model constraints in the form of linear dependencies between data points (as per the graph Laplacian \cite{ansyn}) and, as such, present versatile modelling tools which can help simplify more complex modelling assumptions. Most graph-based methods operate under the assumption that the given graph is optimal and/or that the given data naturally resides on a known graph, so any subsequent tasks and operations are subject to noise pertaining to an imperfect model. Given a data set, approaches have opted to either hand-craft the graph, or learn it automatically through the minimization of a chosen optimization function. In the former case, this has included local and/or adaptive graph weighting and connectivity schemes, which take into account variations in i.a. (spectral) data density, and are increasingly refined but also bear a risk of distorting neighborhood information \cite{spectraldens}. While handcrafted approaches usually lead to higher accuracy, automated ones can be more robust to variable datasets, requiring less parameter tuning while nevertheless being more costly due to lack of an analytic solution. In an effort to combine the advantages of both, we therefore seek an analytic solution to a well-defined optimization problem with the possibility of sophisticated parameter learning (i.e. reduction) through regularization. \section{Preliminaries and Prior Work: Graph (Laplacian) Learning} An undirected graph $G=(V,E)$ is characterized by a set of vertices $V$ and a set of edges $E$, and its connectivity is encapsulated in the symmetric adjacency matrix ${\bf W}$, with $W_{i,j}>0$ if there is an edge between vertices $i$ and $j$, and $W_{i,j}=0$ otherwise. The non-normalized graph Laplacian matrix is defined as ${\bf L}={\bf D}-{\bf W}$, where ${\bf D}=diag({\bf W}{\bf 1})$ is the diagonal degree matrix, with ${\bf 1}$ denoting the vector of 1's. The construction of a graph which optimally represents a given dataset ${\mathcal X}=\{{\bf x}_1,...,{\bf x}_N\}$ is a multi-part task, which generally establishes relations of the form $W_{i,j}=A_{i,j} w({\bf x}_i,{\bf x}_j)$ between data points ${\bf x}_i, {\bf x}_j\in \mathbb{R}^d$, where $w:\mathbb{R}^d\times\mathbb{R}^d\rightarrow \mathbb{R}$ is a similarity function and ${\bf A}$ an adjacency matrix, for which $A_{i,j}=1$ iff $j\in\mathcal{N}_i$ for some pre-determined neighborhood $\mathcal{N}_i$ of node $i$, and $A_{i,j}=0$ otherwise. From the smoothness of $w$ to the range of influence $\mathcal{N}$, each component is vital in ensuring effective discriminative data representation. \\ In graph signal processing, Graph Laplacian Learning (GLL) considers the minimization of the graph Laplacian quadratic form with respect to data (aka graph signal) matrix ${\bf X}$: \[Tr({\bf X}^T {\bf L} {\bf X})=\frac{1}{2}\sum_{i,j} W_{i,j}||{\bf x}_i-{\bf x}_j||_2^2=\frac{1}{2}||{\bf W}\circ {\bf Z}||_{1,1},\] with trace operator $Tr(\cdot)$ and $Z_{i,j}=||{\bf x}_i-{\bf x}_j||_2^2$, and can be alternatively expressed as a weighted sparsity $l_1$-norm of ${\bf W}$, whose minimization enforces connectivity (large $W_{i,j}$) between similar features ${\bf x}_i$ and ${\bf x}_j$. Previous works (see e.g. \cite{kalofolias}, \cite{xdong}), have considered variations of the general framework: \begin{equation}\label{eq:opt} \min_{{\bf W}\in\mathcal{W}}||{\bf W}\circ {\bf Z}||_{1,1}+f({\bf W}),\end{equation} consisting of the graph Laplacian quadratic form and a, possibly sparsity-promoting, regularization term $f({\bf W})$, subject to graph constraints, e.g. $\mathcal{W}=\{{\bf W}\in\mathbb{R}_{\geq 0}^{N\times N}: {\bf W}={\bf W}^T, diag({\bf W})=0\}$. We select $f({\bf W})=\alpha||{\bf W}||^2_F,\alpha\in\mathbb{R}$, with Frobenius norm $||{\bf W}||_F=\sqrt{\sum_{i,j} |W_{i,j}|^2}$, which controls sparsity by preventing the occurrence of strong edges, and due to the edge-locality of the functional, facilitates the decomposition of the problem into a sum over graph edges, and hence an analytic solution.\\ Specifically, the optimization problem of Eq. (\ref{eq:opt}) becomes separable for $f({\bf W})=\alpha||{\bf W}||^2_F$, and can be rewritten in row-wise form \cite{Nie1}, for row ${\bf W}_i$ of ${\bf W}$, as \begin{equation} \label{eq:pro} \min_{{\bf W}_i{\bf 1}=1, {\bf W}_i\geq {\bf 0}, W_{i,i}=0} \sum_{j=1}^N Z_{i,j} {W}_{i,j} +\alpha\sum_{j=1}^N W_{i,j}^2 \end{equation} yielding the closed-form solution \[{W}_{i,j}=\left(\eta_i-\frac{Z_{i,j}}{2\alpha}\right)_{+},\] where $(x)_{+}=\max (0,x)$ and scalar $\eta_i$. Since the second term creates a dense edge pattern, one can further enforce kNN connectivity by determining the maximal $\alpha_i$ per row s.t. the optimal ${\bf W}_i$ has exactly $k$ non-zeros (see \cite{Nie1} for details). This leads to \begin{equation} \label{eq:gg} W_{i,j}= \begin{cases} \frac{Z_{i,k+1}-Z_{i,j}}{k Z_{i,k+1}-\sum_{h=1}^k Z_{i,h}}, & j\leq k\\ 0, & j> k. \end{cases} \end{equation} where entries $\{Z_{i,1},...,Z_{i,N}\}$ are assumed to be ordered from small to large wlog, and we have $\eta_i=\frac{1}{k}+\frac{\sum_{h=1}^k Z_{i,h}}{2k\alpha_i}$ and $\alpha_i=(k/2) Z_{i,k+1}-(1/2)\sum_{h=1}^k Z_{i,h}$. Here, $W_{i,i}=0$ is enforced and $Z_{i,i}=0$ appended at the end. Symmetrization is achieved through ${\bf L}={\bf D}-\frac{{\bf W}+{\bf W}^T}{2}$. \\ It has been noted that this approach is computationally efficient due to its analytic solution and in-built sparsity which does not require oblique tuning of $\alpha$ (instead only requiring the straight-forward number of edges $k$) and is further scale-invariant w.r.t. feature vectors ${\bf x}_i$.\\ \\ For different features (views) of type $v$, we consider $Z_{i,j}=\sum_v c_v Z_{i,j}^v$ with feature coefficients $c_v\in\mathbb{R}$. In order to reduce parameters, it has been proposed to constrain the coefficients to be proportional to pre-set feature-dependent functionals and subject to regularization \cite{multigraph}, \cite{auto}. \section{Proposed Method: Superpixel-based Multi-feature Graph Learning} In the proposed method, several instances of data dependencies are exploited in a multi-level workflow: after conducting an initial spectral dimensionality reduction using PCA \cite{pca}, we consider a priori the segmentation of the hyperspectral image into superpixels to define local regions of homogeneous spectral content and to reduce spatial dimensionality. Subsequently, we compute analytic superpixel features, which capture different image properties, and extrapolate given pixel labels to superpixel labels via a simple averaging filter. We then compute an initial graph $G_0$ and form pseudo-label features through a soft label propagation to nearest neighbors on $G_0$. This is subsequently updated and refined, before a final graph classifier is applied (see Fig. \ref{fig:w}). \subsection{Superpixel-based Feature Extraction} \label{sp1} Given the raw HSI data cube ${\bf I}\in\mathbb{R}^{X\times Y\times B}$, we apply dimensionality reduction in the first instance in the spectral domain using PCA to obtain the reduced $\tilde{{\bf I}}\in\mathbb{R}^{X\times Y\times b},\ b<<B$. \\ Subsequently, the first PC component is used to conduct superpixel segmentation via SLIC \cite{slic}, resulting in the 2D superpixel labelling map $\tilde{{\bf S}}\in\mathbb{R}^{X\times Y}$ with $N$ superpixels: \[{\bf S}_k \ \text{s.t.}\ {\bf S}_k=\{{\bf S}_{(i,j)} | {\bf S}_{(i,j)}=k\}, \ \tilde{{\bf S}}=\cup_{k=1}^N{\bf S}_k.\] While it has been established that the goodness and scale of the superpixel segmentation is foundational for the success of subsequent data modelling and classification tasks, it is not the objective of this work to optimize this particular instance of the workflow; as such, we select SLIC \cite{slic} as the base superpixel segmentation algorithm and determine the number of superpixels approximately according to \cite{colsup}, which takes into account both the size and resolution of the HSI image (albeit not the scene complexity).\\ Let ${\bf Y}\in\mathbb{R}^{XY\times c}$ denote the initial class indicator matrix with $Y_{i,j}=1$ if pixel $i$ belongs to class $j$ for $c$ classes. Following superpixel segmentation, we regularize this to ${\bf Y}^S\in\mathbb{R}^{N\times c}$ by averaging over the existing pixel labels per superpixel. Specifically, $Y^S_{i,j}$ records the number of pixels per superpixel ${\bf S}_i$ which belong to class $j$ divided by the total number of pixels in ${\bf S}_i$. \\ As statistical descriptors for the superpixels, we consider the features as proposed in \cite{sp}, comprising the mean feature vector ${\bf s}_k^M$ \[{\bf s}_k^M=\frac{\sum_{i,j}^{N_{k}} \tilde{{\bf I}}_{(i,j)}}{N_k}, \ {\bf S}_{(i,j)}=k,\ k=1,...,N\] which takes a simple average of the $N_k$ pixels per superpixel $k$, and the spatial-mean feature vector ${\bf s}_k^S$ \[{\bf s}_k^S=\sum_{i=1}^J w_{k,a_i} {\bf s}_{a_i}^M, \ w_{k,a_i}=\frac{\exp{(-||{\bf s}^M_{a_i}-{\bf s}^M_k ||_2^2/h)}}{\sum_{i=1}^J \exp{(-||{\bf s}^M_{a_i}-{\bf s}^M_k ||_2^2/h)}}\] which constitutes a weighted sum of the mean feature vectors of adjacent superpixels, given by index set $\mathcal{A}_k=\{a_1, ...,a_J\}$ for the $k$-th superpixel and with pre-set scalar $h\in\mathbb{R}$. Further, we consider the centroidal location of each superpixel as: \[{\bf s}_k^C=\frac{\sum_{j=1}^{N_k} l_{k,j}}{N_k}\] where $l_{k,j}$ denotes the 2D image coordinate of the $k$-th superpixel. We note that an optimized graph can only be as good as the extracted features it is built upon, however, the task of feature optimization, in line with superpixel segmentation, represents a problem in itself and is not the focus of this work. \subsection{Dynamic Graph Learning and Label Propagation} In the following, we wish to learn a superpixel HSI graph and conduct classification through label propagation. \\Consider the joint optimization problem over the graph ${\bf W}$ and the labelling function ${\bf F}$ \begin{equation} \label{eq:op} \min_{{\bf W}, {\bf F}}\sum_{i,j} (Z_{i,j}+\gamma Z_{i,j}^F)W_{i,j}+\alpha W_{i,j}^2,\end{equation}\[ \ \text{s.t. }\ \sum_j W_{i,j}=1,\ W_{i,j}\geq 0, \ {\bf F}_l={\bf Y}^S_l.\] with $Z_{i,j}^F=||{\bf f}_i-{\bf f}_j||_2^2$, where ${\bf f}_i$ denotes the $i$-th row of ${\bf F}$, and ${\bf Y}^S_l\in\mathbb{R}^{l\times c}$ is the labelled submatrix of ${\bf Y}^S$. Via alternating optimization, the optimal graph ${\bf W}$ can be computed according to Eq. (\ref{eq:gg}) with $Z_{i,j}\rightarrow Z_{i,j}+\gamma Z_{i,j}^F$ and by absorbing $\alpha$ into $k$, the number of edges per row, while for ${\bf F}$, this yields the solution \begin{equation} \label{eq:sol} {\bf F}_u=-{\bf L}_{u,u}^{-1} {\bf L}_{u,l}{\bf Y}^S_l \end{equation} for the unlabelled superpixel nodes, which is also known as harmonic label propagation \cite{auto}, \cite{harmonic}. Here, ${\bf L}_{u,u}$ denotes the graph Laplacian submatrix with rows and columns corresponding to unlabelled nodes. The final superpixel labels are assigned via the decision function \begin{equation}\label{eq:final} y_i=\arg \max_j F_{i,j},\ \ \forall j=1,..., c. \end{equation} Accordingly, we require the optimal graph to exhibit smoothness with respect to both the pre-designed superpixel features, as extracted from the HSI data, as well as the labelling function ${\bf F}$. The functional $Tr({\bf F}^T{\bf L}{\bf F})=\frac{1}{2}\sum_{i,j}Z_{i,j}^F W_{i,j}$, given ${\bf L}$, has been employed as a standalone graph classifier (e.g. \cite{lgc}, \cite{harmonic}), while in the present framework, it is further utilized to enrich the graph construction process, rendering it dynamic.\\ \\ Let ${\bf Z}^v$ with $Z^v_{i,j}=||{\bf s}^v_i-{\bf s}_j^v||^2_2$ denote the Euclidean distance matrix between superpixel feature vectors of type (or view) $v\in\{M,S,C\}$, as previously defined in Sect.\ \ref{sp1}, and $c_v$ the feature weight. In a variation of the above, we propose to employ the superpixel multi-feature dynamic graph $G$ of the basic form as in Eq. (\ref{eq:gg}) with weighted pairwise distances $Z_{i,j}=\sum_v c_v Z^v_{i,j}$ and $Z_{i,j}^{\tilde{F}}=||\tilde{{\bf f}}_i-\tilde{{\bf f}}_j||_2^2$, the latter of which we define as \textit{pseudo-label features} \[\tilde{{\bf f}}_i[j]=\sum_k{ P}_{i,k}^0{Y}_{k,j}^S,\ j=1,...,c \] where $\tilde{{\bf f}}_i$ denotes the $i$-th row of $\tilde{{\bf F}}\in\mathbb{R}^{N\times c}$, ${\bf W}_0$ the initial graph based on $Z_{i,j}$ and ${\bf P}^0={\bf D}^{-1}_0{\bf W}_0$ its random-walk normalized version. In particular, $\tilde{{\bf F}}$ constitutes one instance of a random walk, as opposed to the fully converged solution in Eq. (\ref{eq:sol}). The motivation behind this construction is to provide a soft pre-labelling approach by propagating given labels among their nearest neighbors, as determined through an initial superpixel graph ${\bf W}_0$ based solely on HSI superpixel features, thereby merging information from the label and superpixel dependencies. The resulting distance matrix ${\bf Z}^{\tilde{F}}$ is then utilized as an additional component to rebuild the graph, with large values penalizing nodes not in the same class, further rendering the graph construction dynamic and implicit. Final label propagation on this graph is conducted via the converged harmonic solution in Eq. (\ref{eq:sol}) and class assignment via Eq. (\ref{eq:final}). We summarize the approach in Algorithm 1. \begin{algorithm} \caption{Multi-Feature Graph SSL for HSI Data} \begin{algorithmic}[1] \STATE {\bf INPUT:} raw HSI cube ${\bf I}$, label matrix ${\bf Y}$, parameters: k, $\{c_v\}_v$, $\gamma$. \STATE {\bf OUTPUT:} Classification map ${\bf F}$. \STATE Apply PCA on ${\bf I}$ to obtain $\tilde{{\bf I}}$. \STATE Conduct superpixel segmentation to obtain $\tilde{{\bf S}}$ and label regularization to obtain ${\bf Y}^S$. \STATE Extract superpixel features $\{{\bf s}^M,{\bf s}^S,{\bf s}^C\}$. \STATE Compute initial superpixel-feature graph ${\bf W}_0$ with $Z_{i,j}=\sum_v c_v Z_{i,j}^v$, pre-set $c_v$, $v\in\{S,M,C\}$, via Eq. (\ref{eq:gg}) $\&$ symmetrize. \STATE Compute pseudo-label features $\tilde{{\bf F}}={\bf P}^0{\bf Y}^S$ and ${\bf Z}^{\tilde{F}}$. \STATE Update graph with $\tilde{Z}_{i,j}=Z_{i,j}+\gamma Z_{i,j}^{\tilde{F}}$ and symmetrize. \STATE Compute unknown labels with graph classifier ${\bf F}_u=-{\bf L}_{u,u}^{-1} {\bf L}_{u,l}{\bf Y}^S_l$. \STATE Assign final classes via Eq. (\ref{eq:final}). \end{algorithmic} \end{algorithm} \\ \textit{Remark:} The RBF kernel with $W_{i,j}=\exp\left(-\frac{||{\bf x}_i-{\bf x}_j||_2^2}{\sigma^2}\right)$ constitutes a prominent approach to model graph weights \cite{graph} and is incidentally the result of the optimization problem in Eq. (\ref{eq:opt}) with $f({\bf W})=\sigma^2\sum_{i,j}W_{i,j}\log W_{i,j}$ in normalized form. Nevertheless, performance is strongly affected by the tuning of $\sigma$ and graph connectivity, the latter of which is not embedded into the graph solution and both of which are non-trivial. \\ While the proposed approach simplifies the issue of tuning edge connectivity and neighbourhood range, thus increasing robustness, it still comprises a range of parameters, which are categorized into: $k$ (the number of nearest neighbour edges), $\{c_v\}_v$ (superpixel feature weights), and $\gamma$ (pseudo-label feature weight). As the goodness of the graph is dependent on the goodness of its features, we proceed to simultaneously learn feature weights and update the graph with the goal to help guide as well as minimize uninformed parameter-tuning. \subsection{Parameter-optimal Multi-feature Graph Learning} While a reduction of parameters generally occurs at the sacrifice of performance and cannot replace a thorough parameter search, we investigate the possibility of a pseudo-label guided parameter reduction and further propose a variation of the preceding framework, inspired by \cite{multigraph}, in an effort to facilitate training and generalizability to diverse datasets.\\ Consider individual superpixel feature graphs, denoted with ${\bf A}^v$, whose entries are computed from $Z_{i,j}^v$ using Eq. (\ref{eq:gg}) with $k$ edges per row, and initialize the global graph ${\bf W}=\sum_{v=1}^V c_v {\bf A}^v$ with $c_v=\frac{1}{V}$. We assume that the deviation $r^v=||{\bf W}-{\bf A}^v||_F^2$ from ${\bf W}$ is inversely related to the feature importance $c_v$. After an initial pseudo-label computation $\tilde{{\bf F}}$ using Eq. (\ref{eq:sol}), we update the global graph ${\bf W}$ by solving \begin{equation} \label{eq:gl} \min_{{\bf W}_i\geq 0, {\bf W}_i {\bf 1}=1}\left|\left|{\bf W}_i+\frac{(\gamma_1/2) {\bf Z}_{i}^{WF}-\sum_v c_v {\bf A}_i^v}{\sum_v c_v}\right|\right|_2^2\end{equation} In contrast to the previous method, we determine the pseudo-labels $\tilde{{\bf F}}$ here via the converged harmonic solution in Eq. (\ref{eq:sol}) and apply a binary mask ${W}$ which holds the nonzero locations of the current global graph estimate ${\bf W}$, giving the Hadamard product ${\bf Z}^{WF}={W}\circ {\bf Z}^{\tilde{F}}$. As such, when solving Eq. (\ref{eq:gl}) the positions of the non-zero graph weights remain fixed, while their values are perturbed, i.e. re-weighted or eliminated according to pseudo-label information. This prevents the formation of noisy edges when ${\bf Z}^{\tilde{F}}$ is dense, and constitutes an alternative to soft pseudo-labelling (for which we previously employed one random walk). Subsequently, we regularize the weights $c_v$ via an $l_2$-norm term: \begin{equation} \label{eq:coef} \min_{{\bf c}} \sum_v c_v ||{\bf W}-{\bf A}^v||_F^2+\gamma_2 ||{\bf c}||_2^2,\ \text{s.t.}\ c_v\geq 0, \ {\bf c}^T{\bf 1}=1\end{equation} which can be simplified to \begin{equation} \label{eq:coef2}\min_{c_v \geq 0, {\bf c}^T{\bf 1}=1}\left|\left|{\bf c}+\frac{{\bf r}}{2\gamma_2}\right|\right|_2^2.\end{equation} Notably, Eq. (\ref{eq:pro}) can be written in the same form as Eqs. (\ref{eq:gl}) and (\ref{eq:coef2}), which constitute the Euclidean projection onto the probabilistic simplex \cite{simplex}; however, as we cannot apply the same kNN simplification, we solve the latter two iteratively, whereby we employ Newton's method to enforce the unity sum constraint. As such, both the graph edge learning and feature weight learning stage are essentially the same. \\ By recomputing ${\bf Z}^{WF}$ and then ${\bf W}$ with corresponding parameter $\gamma_3$, we obtain a pseudo-label enhanced graph which is used by the graph classifier of Eq. (\ref{eq:sol}) to obtain the final solution.\\ Overall, $\gamma_2$ controls the disparity between feature weights, while replacing $V$ parameters with one, while $\gamma_1$ and $\gamma_3$ regulate the pseudo-label contribution at different stages. Here we employ pseudo-labels in a two-fold way to inform feature contribution as well as to form a separate feature embedded in the graph. While this can be tuned with a single parameter $\gamma_1$, in practice, we observe that performance benefits from weighting the steps separately by introducing $\gamma_3$, as we will demonstrate in Sect. V. We summarize the graph learning stage of the approach in Algorithm 2; here, each computed graph is a posteriori symmetrized via $\frac{{\bf W}+{\bf W}^T}{2}$. The approach is similar to solving the joint optimization problem of Eqs. (\ref{eq:op}) (with respect to ${\bf F}$) and (\ref{eq:coef}) alternately, as each step constitutes an optimal closed-form solution; however, we refrain from further iterations between steps ${\bf 4}$ and ${\bf 6}$ to limit possible noise resulting from pseudo-labelling.\\ We note that large deviations in (feature) scales, and thus in ${\bf r}$, result in binary weights (i.e. single-feature selection), which can be remedied, in part, by tuning $\gamma_2$, as well as by refining the feature selection. For this approach, we introduce two composite superpixel feature measures, which merge the centroidal feature, which is less informative for a standalone feature graph, with either of the spectral-content features. Specifically, we consider the multiplicative $Z^v_{i,j}\circ Z_{i,j}^C$ and additive $Z_{i,j}^v+\lambda Z_{i,j}^C, v\in\{M, S\}$, with $v$ chosen as per dataset and $\lambda \sim \sigma_v/\sigma_C$, where $\sigma_v=\sum_{i,j} Z_{i,j}^v/N^2$ denotes the scale per feature. \begin{algorithm} \caption{Parameter-optimal Multi-Feature Graph SSL} \begin{algorithmic}[1] \STATE Initialize superpixel-feature graphs ${\bf A}^v$ with ${\bf Z}^v$ via Eq. (\ref{eq:gg}) and ${\bf W}^0=\sum_v c_v {\bf A}^v$ with $c_v=\frac{1}{V}$, then symmetrize. \STATE Compute pseudo-label features $\tilde{{\bf F}}$ via Eq. (\ref{eq:sol}) and ${\bf Z}^{WF}=\textit{W}\circ{\bf Z}^{\tilde{F}}$. \STATE Update graph with $\tilde{Z}_{i,j}=(\frac{\gamma_1}{2} Z_{i,j}^{WF}-\sum_v c_v A^v_{i,j})/\sum_v c_v$ with pre-set $\gamma_1$ by solving Eq. (\ref{eq:gl}), symmetrize. \STATE Compute feature weights $c_v$ with pre-set $\gamma_2$ via Eq. (\ref{eq:coef2}). \STATE Update pseudo-labels $\tilde{{\bf F}}$ via Eq. (\ref{eq:sol}). \STATE Update graph with pre-set pseudo-label weight $\gamma_3$ in Eq. (\ref{eq:gl}), then symmetrize. \STATE Compute final labels via graph classifier ${\bf F}_u=-{\bf L}_{u,u}^{-1} {\bf L}_{u,l}{\bf Y}^S_l$ and Eq. (\ref{eq:final}). \end{algorithmic} \end{algorithm}\\ \noindent \textit{Remark}: One could further consider the constraint $r^v=\sum_{i,j} Z^v_{i,j} W_{i,j}$ in Eq. (\ref{eq:coef}) as a non-separable way to estimate the feature weights, however, we observe that discrepancies in scaling render the parameter $\gamma_2$ more difficult to tune. Instead, we opt to separate the graph into the sum of individual feature graphs. While this bears the bias of reduced global interaction between the features (i.e. in Eq. (\ref{eq:gg}) the sum $Z_{i,j}=\sum_{v} c_vZ^v_{i,j}$ drives the assignment of the nearest $k$ edges), we remedy this by incorporating pseudo-labels into the framework as a means to perturb the solution and reinforce inter-and intra-class relations as well as by introducing composite features. \section{Experimental Results} \subsection{Dataset Description} We validate our approach on three benchmark HSI datasets:\\ {\bf Indian Pines:} This data set was gathered by an Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) over an agricultural site in Indiana and consists of $145\times 145$ pixels with a spatial resolution of 20 m per pixel. The AVIRIS sensor has a wavelength range from $0.4$ to $2.5$ $\mu m$, which is divided into 224 bands, of which 200 are retained for experiments. There are 16 classes, the distribution of which is imbalanced, with Alfalfa, Oats and Grass/Pasture-mowed containing relatively few labeled samples.\\ {\bf Salinas:} The second data set was similarly acquired by the AVIRIS sensor over Salinas Valley, California, comprising $512\times 217$ pixels with a notably higher spatial resolution of $3.7$ m per pixel. Further, 204 bands are retained. The scene contains 16 classes, covering i.a. soils and fields.\\ {\bf University of Pavia:} The third data set was collected by the Reflective Optics System Imaging Spectrometer (ROSIS), containing $610\times 340$ pixels with a spatial resolution of $1.3$ m per pixel, the highest of the three datasets. The spectral range from $0.43$ to $0.86$ $\mu m$ is divided into 115 spectral bands, of which 103 are retained. The urban site contains 9 classes and covers the Engineering School at the University of Pavia. \subsection{Experimental Design} In the following, we conduct the experimental evaluation of our method on the benchmarking datasets and demonstrate its superiority compared to state-of-the-art algorithms. For evaluation, experiments are repeated 10 times and performance is assessed on the basis of the average and standard deviation of three quality indeces: the overall accuracy (OA), as the percentage of correctly classified pixels, the average accuracy (AA), as the mean of the percentage of correctly classified pixels per class, and the Kappa Coefficient ($\kappa$), as the percentage of correctly classified pixels corrected by the number of agreements expected purely by chance. Further, we compare performance with the Local Covariance Matrix Representation (LCMR) \cite{lcmr}, the Edge-Preserving Filtering (EPF) \cite{epf}, the Image Fusion and Recursive Filtering (IFRF) \cite{IFRF}, and the SVM \cite{svm} methods, which, as established state-of-the-art methods, were specifically chosen as comparisons due to their inherent spectral-spatial modelling techniques (with exception of the purely spectral SVM). The SVM algorithm is implemented in the LIBSVM library \cite{libsvm}, adopting the Gaussian kernel with fivefold cross validation for the classifier. We further adopt model-parameters as specified in these works. The proposed methods are abbreviated as Multi-Feature Graph Learning (MGL) and Parameter-optimal Multi-Feature Graph Learning (PMGL) respectively. \subsection{Parameter Specification} For both proposed methods, we conduct PCA on the standardized data to explain $99.8\%$ of the data's variance and employ SLIC \cite{slic} for superpixel segmentation with a compactness of $10$, where we fix the number $K$ of superpixels per dataset to $K=1287$ (Indian Pines), $K=2237$ (Salinas) and $K=3080$ (University of Pavia). For all graph constructions, we set $k=10$ as the number of edges per node, and for the spatial mean feature construction, we select $h=15$.\\ For the first method (MGL), we employ pseudo-label feature weight $\gamma=10$ and superpixel feature weights $c_S=1$, $c_M=0.5$, $c_C=10^{-2}$ (Indian Pines), $c_S=0.5$, $c_M=5$, $c_C=10^{-5}$ (University of Pavia), and $c_S=1$, $c_M=0.1$, $c_C=10^{-4}$ (Salinas).\\ For the second method (PMGL), we employ $\gamma_1=0$, $\gamma_2=30$, $\gamma_3=1$ (Indian Pines), $\gamma_1=20$, $\gamma_2=40$, $\gamma_3=0$ (University of Pavia) and $\gamma_1=1$, $\gamma_2=30$, $\gamma_3=1$ (Salinas). Further, for the latter method we construct three individual feature graphs respectively based on the following feature distances: $\{{\bf Z}^M, {\bf Z}^S, {\bf Z}^C\circ {\bf Z}^S\}$ (Indian Pines), $\{{\bf Z}^M, {\bf Z}^S, {\bf Z}^M+\lambda {\bf Z}^C\}$ (University of Pavia), $\{{\bf Z}^S, {\bf Z}^M+\lambda {\bf Z}^C, {\bf Z}^S+\lambda {\bf Z}^C\}$ (Salinas), which were deemed to summarize most effectively the main properties of the different HSI scenes. \subsection{Experimental Results \& Discussion} Performance is evaluated using a very limited amount of labels for training, with rates of $3-20$ samples per class, which are randomly selected, in several stages of experiments. In the first instance, we conduct numerical evaluations and comparisons of classification accuracy of the proposed methods with state-of-the-art approaches, as detailed above, followed by an evaluation of the corresponding visual classification maps. \\ {\bf E1}: We begin by evaluating the OA and Kappa coefficient in comparison using a reduced label rate of 3-20 randomly selected labels per class, whose results are graphically displayed in Fig. \ref{figg1} for the three benchmark data sets. \begin{figure*}[!t] \subfloat[Indian Pines]{ \includegraphics[width=2.2in]{IPOAnewcor.pdf}} \subfloat[Pavia University]{ \includegraphics[width=2.2in]{UPOAnewcor.pdf}} \subfloat[Salinas]{ \includegraphics[width=2.2in]{SOAnewcor.pdf}} \subfloat[Indian Pines]{ \includegraphics[width=2.2in]{IPKappanewcor.pdf}} \subfloat[Pavia University]{ \includegraphics[width=2.2in]{UPKappanewcor.pdf}} \subfloat[Salinas]{ \includegraphics[width=2.2in]{SKappanewcor.pdf}} \caption{\footnotesize{Comparison of the classification accuracy (OA) and Kappa coefficient of different methods with varying number of training samples over 10 trials. The solid lines represent the mean while the shaded area covers the standard deviation from the mean.}} \label{figg1} \end{figure*} We observe that both proposed methods consistently outperform the other methods over the entire range of label rates for all three benchmark data sets. In particular, the performance gain is highest for lower label rates, signifying that the proposed pseudo-label guided graph-based methods perform strongly even when extremely few labels are available as a result of their superior model. LCMR and IFRF form the closest competitors for the Indian Pines and Salinas data sets with a maximum gap of approximately $10\%$ and $8 \%$ respectively to the closest competitor at 3 labelled samples, with LCMR being the dominant competitor for the more complex Pavia University data set with a gap of $10\%$ to PMGL. Further, among the two proposed methods, in the lower label limit of the Pavia dataset, PMGL exhibits up to $4\%$ gain in OA performance, while for the other data sets, this gain is vanishingly small with the two methods performing comparably. This indicates that a more refined parameter selection can be beneficial for structurally complex data sets. \noindent {\bf E2}: To demonstrate the influence of the pseudo-labels and feature coefficients in PMGL on performance, we consider the overall accuracy at a fixed label rate of 7 samples per class with varying feature parameters over 10 trials. In particular, in order to illustrate the interaction between pseudo-label contribution and feature contribution, we fix pseudo-label parameter $\gamma_3$, while varying $\gamma_1$, along with the feature weight-regularization parameter $\gamma_2$, and consider the mean OA in a 3D plot. We report results for the University of Pavia data set, as the most structurally complex of the three, in Fig. \ref{figg2} with the OA plotted against $\gamma_1$ and $\gamma_2$ and $\gamma_3=0$ in (a), and $\gamma_3=1$ in (b), with (c) showing the resulting feature coefficient distribution for (a). We observe that in (a), OA is highest when both $\gamma_1$ and $\gamma_2$ are increased, generating a perturbation toward more evenly distributed coefficients, which, as shown in (c), corresponds to the gradual matching of composite coefficient $c_{M+\lambda C}$ and spectral mean coefficient $c_{M}$, while the contribution of the spatial mean coefficient $c_{S}$ is negligible. When additionally, the pseudo-label feature is incorporated into the graph via $\gamma_3$, the coefficient distribution for the best OA changes, instead overall moving toward one dominant superpixel feature. While we observe interactions between pseudo-label and superpixel features which drive performance, the feature regularization parameter is ultimately dependent on the data set and the selected features at hand. \begin{figure*} \subfloat[]{ \includegraphics[width=2.4in]{UPOAP11cor}} \subfloat[]{ \includegraphics[width=2.4in]{UPOAP22cor}} \subfloat[]{ \includegraphics[width=2.4in]{UPcoefcor}} \caption{University of Pavia: results for 7 training samples per class, 3D plot of pseudo-label weight $\gamma_1$ vs feature regularization weight $\gamma_2$ against OA of PMGL with (a) $\gamma_3=0$, (b) $\gamma_3=1$, and (c) feature coefficient distribution for $\gamma_3=0$.} \label{figg2} \end{figure*}\\ \noindent {\bf E3}: For each data set we use 7 labeled samples per class and run the methods again to calculate the OA, Kappa coefficient, AA, and a full class by class accuracy breakdown over 10 trials. For all three data sets, the proposed methods MGL and PMGL consistently outperform the competitors in OA, Kappa coefficient and AA, and for the majority of per class accuracies. For the Indian Pines data set, PMGL performs only slightly better than MGL in the first three measures, and around $10 \%$ improvement in OA over its closest competitor, LCMR. In the case of the University of Pavia data set, the gain of PMGL over MGL is even larger with 3-4$\%$, followed by LCMR with $10$$\%$ improvement in OA. Lastly, for the Salinas data set, MGL and PMGL perform comparativey well, with a gain of $3$$\%$ over their closest competitor, IFRF. Overall, we observe that while PMGL utilizes more intricate relations and selective feature contributions, MGL is still close in performance, with a gain of the former becoming more evident for increasingly complex data sets, such as Pavia University. \begin{table*}[!t] \caption{OA, Kappa, AA and per class breakdown in $\%$ with 7 training labels per class} \label{table1} \centering \begin{tabular}{|c|c|c||c|c|c|c|} \cline{1-7} \multicolumn{7}{|c@{}|}{{\bf Indian Pines}}\\ \hline {CLASS} & MGL & PMGL & LCMR \cite{lcmr} & EPF \cite{epf}& IFRF \cite{IFRF}& SVM \cite{svm} \\ \hline C1 & $97.95\pm1.03$ & $97.95\pm1.03$ & ${\bf 99.23\pm1.24}$ & $95.13\pm7.69$ & ${\bf 99.23\pm1.24}$ & $72.05\pm13.54$ \\ C2 & $66.17\pm10.17$ & $70.63\pm11.86$ & ${\bf 73.36\pm8.24}$ & $39.32\pm9.52$ & $65.58\pm9.98$ & $33.12\pm4.41$ \\ C3 & $76.11\pm9.73$ & ${\bf 86.12\pm8.26}$ & $62.08\pm8.98$ & $55.65\pm12.92$ & $76.1\pm9.16$ & $44.63\pm6.07$ \\ C4 & ${\bf 96.17\pm4.01}$ & $92.3\pm7.54$ & $93.48\pm6.45$ & $78.13\pm20.28$ & $87.3\pm7.99$ & $48.96\pm12.01$ \\ C5 & ${\bf 88.84\pm9.09}$ & $88.47\pm8.83$ & $88.55\pm8.04$ & $83.8\pm10.94$ & $80.8\pm6.47$ & $70.82\pm10.22$ \\ C6 & ${\bf 94.54\pm10.67}$ & ${\bf 93.43\pm8.9}$ & $86.43\pm8.34$ & $90.06\pm8.05$ & $86.02\pm13.87$ & $68.49\pm10.61$ \\ C7 & ${\bf 100.0\pm0.0}$ & ${\bf 100.0\pm0.0}$ & ${\bf 100.0\pm0.0}$ & $98.1\pm2.46$ & ${\bf 100.0\pm0.0}$ & $83.33\pm15.27$ \\ C8 & ${\bf 100.0\pm0.0}$ & ${\bf 100.0\pm0.0}$ & $97.62\pm4.47$ & $68.3\pm16.46$ & $99.62\pm0.71$ & $56.28\pm9.54$ \\ C9 & ${\bf 100.0\pm0.0}$ & ${\bf 100.0\pm0.0}$ & ${\bf 100.0\pm0.0}$ & ${\bf 100.0\pm0.0}$ & ${\bf 100.0\pm0.0}$ & $89.23\pm12.67$ \\ C10 & ${\bf 88.05\pm7.41}$ & ${\bf 89.97\pm7.04}$ & $76.29\pm11.93$ & $61.06\pm17.8$ & $79.19\pm8.29$ & $45.67\pm10.66$ \\ C11 & ${\bf 87.53\pm6.94}$ & ${\bf 82.07\pm10.0}$ & $61.55\pm9.07$ & $48.35\pm13.16$ & $64.08\pm12.69$ & $38.08\pm7.69$ \\ C12 & ${\bf 86.64\pm12.78}$ & ${\bf 93.16\pm10.15}$ & $79.68\pm7.3$ & $57.59\pm25.11$ & $68.45\pm12.47$ & $40.39\pm11.65$ \\ C13 & ${\bf 99.49\pm0.0}$ & ${\bf 99.49\pm0.0}$ & $99.19\pm0.9$ & ${\bf 99.49\pm0.0}$ & $98.59\pm1.52$ & $90.86\pm4.36$ \\ C14 & ${\bf 97.02\pm8.58}$ & $94.42\pm8.77$ & $95.86\pm5.15$ & $76.92\pm23.07$ & $82.22\pm7.23$ & $60.51\pm16.01$ \\ C15 & ${\bf 94.51\pm9.39}$ & ${\bf 94.64\pm9.29}$ & $92.32\pm9.22$ & $41.42\pm13.57$ & $86.15\pm7.41$ & $31.77\pm7.52$ \\ C16 & $93.72\pm1.74$ & $93.72\pm1.74$ & ${\bf 99.19\pm2.57}$ & $95.23\pm6.79$ & $97.09\pm4.04$ & $85.58\pm10.11$ \\ \hline OA & ${\bf 86.75\pm2.31}$ & ${\bf 86.93\pm2.79}$ & $77.88\pm4.07$ & $60.7\pm6.92$ & $75.77\pm4.65$ & $47.84\pm3.29$ \\ Kappa & ${\bf 84.96\pm2.6}$ & ${\bf 85.19\pm3.12}$ & $75.14\pm4.51$ & $56.16\pm7.59$ & $72.84\pm5.06$ & $42.07\pm3.6$ \\ AA & ${\bf 91.67\pm1.31}$ & ${\bf 92.27\pm1.11}$ & $87.8\pm2.27$ & $74.28\pm4.98$ & $85.65\pm2.77$ & $59.99\pm3.48$ \\ \hline \\ \cline{1-7} \multicolumn{7}{|c@{}|}{{\bf University of Pavia}}\\ \hline {CLASS} & MGL & PMGL & LCMR \cite{lcmr} & EPF \cite{epf}& IFRF \cite{IFRF}& SVM \cite{svm} \\ \hline C1 & $76.23\pm8.91$ & $78.42\pm6.35$ & ${\bf 79.92\pm9.14}$ & $76.0\pm11.24$ & $53.47\pm9.97$ & $62.63\pm8.75$ \\ C2 & ${\bf 92.53\pm7.9}$ & ${\bf 96.12\pm5.36}$ & $81.38\pm11.88$ & $59.17\pm10.17$ & $78.64\pm6.42$ & $55.55\pm8.96$ \\ C3 & ${\bf 87.5\pm10.43}$ & ${\bf 90.19\pm8.26}$ & $83.98\pm10.72$ & $60.88\pm13.2$ & $55.76\pm10.56$ & $55.05\pm9.57$ \\ C4 & $84.03\pm7.83$ & $84.76\pm6.85$ & ${\bf 93.72\pm4.43}$ & $86.49\pm5.0$ & $65.45\pm16.27$ & $88.46\pm5.49$ \\ C5 & $93.27\pm6.35$ & $93.68\pm5.89$ & $97.2\pm7.54$ & $96.76\pm10.23$ & ${\bf 99.13\pm0.62}$ & $96.63\pm8.49$ \\ C6 & ${\bf 94.94\pm6.51}$ & ${\bf 98.23\pm2.63}$ & $84.82\pm8.41$ & $63.46\pm13.17$ & $83.4\pm7.23$ & $57.34\pm9.54$ \\ C7 & ${\bf 96.02\pm1.84}$ & ${\bf 99.46\pm0.26}$ & $85.15\pm12.58$ & $94.32\pm9.97$ & $77.41\pm14.24$ & $86.22\pm8.86$ \\ C8 & ${\bf 82.41\pm19.95}$ & ${\bf 90.71\pm5.06}$ & $71.03\pm9.2$ & $74.57\pm15.13$ & $71.8\pm10.57$ & $64.32\pm10.18$ \\ C9 & $92.96\pm3.37$ & $95.9\pm2.03$ & $94.46\pm2.06$ & ${\bf 99.78\pm0.56}$ & $50.54\pm10.7$ & $99.66\pm0.58$ \\ \hline OA & ${\bf 88.7\pm3.56}$ & ${\bf 92.08\pm2.48}$ & $82.58\pm5.74$ & $68.81\pm4.31$ & $72.63\pm3.09$ & $63.15\pm3.64$ \\ Kappa & ${\bf 85.3\pm4.44}$ & ${\bf 89.61\pm3.11}$ & $77.77\pm6.74$ & $61.27\pm4.75$ & $64.86\pm3.7$ & $54.59\pm3.77$ \\ AA & ${\bf 88.87\pm2.35}$ & ${\bf 91.94\pm1.6}$ & $85.74\pm2.64$ & $79.05\pm4.05$ & $70.62\pm2.78$ & $73.98\pm2.75$ \\ \hline \\ \cline{1-7} \multicolumn{7}{|c@{}|}{{\bf Salinas}}\\ \hline {CLASS} & MGL & PMGL & LCMR \cite{lcmr} & EPF \cite{epf}& IFRF \cite{IFRF}& SVM \cite{svm} \\ \hline C1 & ${\bf 100.0\pm0.0}$ & ${\bf 100.0\pm0.0}$ & $99.89\pm0.21$ & $99.77\pm0.5$ & $98.23\pm5.61$ & $97.85\pm1.26$ \\ C2 & ${\bf 100.0\pm0.0}$ & $ {\bf 100.0\pm0.0}$ & $89.79\pm5.67$ & $99.74\pm0.4$ & $96.36\pm4.47$ & $98.14\pm1.42$ \\ C3 & ${\bf 100.0\pm0.0}$ & ${\bf 100.0\pm0.0}$ & $98.55\pm2.5$ & $86.26\pm19.94$ & $99.97\pm0.08$ & $81.24\pm18.02$ \\ C4 & $94.18\pm5.92$ & $94.17\pm8.91$ & ${\bf 100.0\pm0.0}$ & $99.88\pm0.13$ & $98.61\pm2.63$ & $99.15\pm0.59$ \\ C5 & $93.28\pm4.31$ & $94.72\pm0.0$ & $96.53\pm0.69$ & ${\bf 97.62\pm1.81}$ & $92.43\pm5.1$ & $96.07\pm2.37$ \\ C6 & ${\bf 99.59\pm0.07}$ & ${\bf 99.59\pm0.06}$ & $99.02\pm0.99$ & $99.51\pm0.79$ & $99.53\pm0.65$ & $98.02\pm2.1$ \\ C7 & ${\bf 99.89\pm0.06}$ & ${\bf 100.0\pm0.0}$ & $97.9\pm2.15$ & $99.79\pm0.07$ & $98.82\pm3.33$ & $98.57\pm0.93$ \\ C8 & ${\bf 98.45\pm0.71}$ & ${\bf 97.46\pm1.84}$ & $85.58\pm5.54$ & $64.96\pm18.35$ & $85.93\pm7.01$ & $57.35\pm11.09$ \\ C9 & ${\bf 100.0\pm0.0}$ & ${\bf 100.0\pm0.0}$ & $92.23\pm9.51$ & $99.49\pm0.61$ & $99.94\pm0.15$ & $98.23\pm0.91$ \\ C10 & $89.47\pm7.41$ & $90.37\pm5.68$ & ${\bf 97.44\pm1.37}$ & $89.24\pm9.66$ & $97.16\pm3.84$ & $80.48\pm9.14$ \\ C11 & $95.57\pm5.16$ & $97.67\pm0.85$ & ${\bf 99.91\pm0.08}$ & $97.15\pm2.03$ & $96.03\pm3.32$ & $87.51\pm3.7$ \\ C12 & $97.59\pm0.31$ & $97.59\pm0.31$ & $99.23\pm2.38$ & ${\bf 100.0\pm0.0}$ & $98.3\pm1.45$ & $96.18\pm3.08$ \\ C13 & $97.67\pm0.62$& $97.48\pm0.36$ & $98.26\pm0.85$ & ${\bf 98.92\pm0.38}$ & $95.73\pm5.84$ & $98.22\pm0.6$ \\ C14 & $92.91\pm4.8$ & $94.65\pm2.95$ & $93.13\pm5.84$ & $95.3\pm1.66$ & ${\bf 96.32\pm3.28}$ & $88.86\pm2.69$ \\ C15 & ${\bf 96.32\pm1.28}$ & ${\bf 98.23\pm0.97}$ & $82.71\pm7.54$ & $71.74\pm21.97$ & $93.98\pm3.76$ & $59.05\pm12.78$ \\ C16 & ${\bf 100.0\pm0.0}$ & ${\bf 100.0\pm0.0}$ & $91.23\pm7.49$ & $94.71\pm4.17$ & $98.12\pm3.02$ & $92.33\pm4.68$ \\ \hline OA & ${\bf 97.67\pm0.48}$ & ${\bf 97.93\pm0.45}$ & $91.99\pm1.34$ & $87.14\pm3.22$ & $94.89\pm1.7$ & $81.98\pm1.54$ \\ Kappa & ${\bf 97.41\pm0.54}$ & ${\bf 97.7\pm0.5}$ & $91.1\pm1.49$ & $85.73\pm3.57$ & $94.32\pm1.89$ & $80.01\pm1.69$ \\ AA & ${\bf 97.18\pm0.69}$ & ${\bf 97.62\pm0.65}$ & $95.09\pm1.01$ & $93.38\pm2.24$ & $96.59\pm0.85$ & $89.2\pm1.5$ \\ \hline \end{tabular} \end{table*}\\ \noindent {\bf E4}: At last, we show the full classification maps produced for training with 7 samples per class for all methods in comparison. In Fig. \ref{ind_im}, classification maps for the Indian Pines data set illustrate increased smoothness and local homogeneity for the proposed graph-based MGL and PMGL, exemplifying their superiority. Their closest competitor, LCMR exhibits more noisy regions. Fig. \ref{pu_im} shows the University of Pavia classification maps, which as the most structurally complex scene of the three with some scattered classes, similarly exhibits smooth yet spatially refined classification results for the proposed MGL and PMGL, despite the inherent crudeness of superpixel segmentation and label regularization. At last, for the Salinas data set in Fig. \ref{sal_im}, which due to its locally homogeneous regions and overall spatial regularity represents a simpler scene, the proposed methods still to manage to improve over existing methods, achieving near perfect classification.\\ It becomes evident that the use of superpixels, ensuring local homogeneity, as well as that of graphs, for refined local and global modelling, which incorporates pixel-and superpixel-level as well as spectral-spatial-label dependencies, facilitates a significant performance gain. Notably, EPF and IFRF employ a spectral-spatial filtering approach which can be likened to graph filtering, with the distinction that the latter is more flexible and versatile to model; nevertheless, the use of the SVM classifier in all competitors, as opposed to a graph-filter in the latter case, contributes to a comparatively decreased accuracy owing to the noise of the spectral-based SVM. \begin{figure*}[!t] \centering \subfloat[Colour]{ \includegraphics[width=1.2in]{ip_colour.png}}\hfil \subfloat[GT]{ \includegraphics[width=1.2in]{ind_gt.png}}\hfil \subfloat[MGL]{ \includegraphics[width=1.2in]{indmgl}}\hfil \subfloat[PMGL]{\includegraphics[width=1.2in]{indpmgl.png}}\hfil \subfloat[LCMR]{ \includegraphics[width=1.2in]{LCMR.png}}\hfil \subfloat[EPF]{ \includegraphics[width=1.2in]{EPF.png}}\hfil \subfloat[IFRF]{ \includegraphics[width=1.2in]{IFRF.png}}\hfil \subfloat[SVM]{ \includegraphics[width=1.2in]{SVM.png}}\hfil \vspace{2mm} {\includegraphics[width=4.2in]{ip_legend.png}} \caption{Indian Pines: (a) Colour composite, (b) Ground truth, (c)-(h) classification maps produced using 7 labelled samples per class. \label{ind_im} \end{figure*} \begin{figure*}[!t] \centering \subfloat[Colour]{ \includegraphics[width=0.84in]{pavia_colour.png}}\hfil \subfloat[GT]{ \includegraphics[width=0.84in]{paviau_gt.png}}\hfil \subfloat[MGL]{ \includegraphics[width=0.84in]{pumgl.png}}\hfil \subfloat[PMGL]{ \includegraphics[width=0.84in]{pupmgl.png}}\hfil \subfloat[LCMR]{ \includegraphics[width=0.84in]{LCMR_Pavia.png}}\hfil \subfloat[EPF]{ \includegraphics[width=0.84in]{EPF_Pavia.png}}\hfil \subfloat[IFRF]{ \includegraphics[width=0.84in]{IFRF_Pavia.png}}\hfil \subfloat[SVM]{ \includegraphics[width=0.84in]{SVM_Pavia.png}}\hfil % \vspace{2mm} {\includegraphics[width=4in]{pu_legend.png}} \caption{Pavia University: (a) Colour composite, (b) Ground truth, (c)-(h) classification maps produced using 7 labelled samples per class.} \label{pu_im} \end{figure*} \begin{figure*}[!t] \centering \subfloat[Colour]{ \includegraphics[width=0.84in]{sal_colour.png}}\hfil \subfloat[GT]{ \includegraphics[width=0.84in]{sal_gt.png}}\hfil \subfloat[MGL]{ \includegraphics[width=0.84in]{smgl.png}}\hfi \subfloat[PMGL]{ \includegraphics[width=0.84in]{spmgl.png}}\hfi \subfloat[LCMR]{ \includegraphics[width=0.84in]{LCMR_Sal.png}}\hfi \subfloat[EPF]{ \includegraphics[width=0.84in]{EPF_Sal.png}}\hfi \subfloat[IFRF]{ \includegraphics[width=0.84in]{IFRF_Sal.png}}\hfi \subfloat[SVM]{ \includegraphics[width=0.84in]{SVM_Sal.png}}\hfi % \vspace{2mm} {\includegraphics[width=6in]{s_legend.png}} \caption{Salinas: (a) Colour composite, (b) Ground truth, (c)-(h) classification maps produced using 7 labelled samples per class.} \label{sal_im} \end{figure*} \section{Conclusion} In this work, we have developed a pseudo-label-guided and superpixel-based graph learning framework for semi-supervised classification of HSI data. Specifically, we have presented two methods: while the former constructs a single dynamic graph by fusing different superpixel features along with a pseudo-label feature, the latter constructs a global graph as the sum of individual feature graphs whose contribution is informed through pseudo-label regularization. We have demonstrated on the basis of benchmark data sets the superiority in performance through quantitative and qualitative results in comparison to state-of-the-art methods, particularly in the small labelled sample limit.\\ The incorporation of multiple features as well as label information through pseudo-labels into the graph facilitate refined modelling of the complex dependencies present in HSI data, and ultimately leverage these for an improved classification performance. Furthermore, the multi-stage workflow, which employs superpixels and a flexible, inherently sparse graph design with the option to reduce parameters through regularization, is versatile by allowing for multiple components and exploiting multiple levels of spectral-spatial and label-dependencies. Nevertheless, it remains a challenge to completely eliminate parameters whilst maintaining competitive performance as well as to select the ideal features upon which the goodness of subsequent steps depends; ultimately, the process of feature selection/extraction and feature weighting presented is not exhaustive. In our future work, we wish to explore automated DL approaches for both feature selection and graph construction which can be guided by sophisticated model-based priors. \section*{Acknowledgment} The authors would like to thank Prof. D. Landgrebe from Purdue University and the NASA Jet Propulsion Laboratory for providing the hyperspectral data sets. They would also like to thank Prof D. Coomes from the Department of Plant Sciences, University of Cambridge, for his advice and support. This research was carried out as part of the INTEGRAL project funded by GCRF (EPSRC) grant EP/T003553/1. In addition, CBS acknowledges support from the Philip Leverhulme Prize, the Royal Society Wolfson Fellowship, the EPSRC grants EP/S026045/1, EP/N014588/1, EP/T017961/1, the Wellcome Innovator Award RG98755, the Leverhulme Trust project Unveiling the invisible, the European Union Horizon 2020 research and innovation programme under the Marie Skodowska-Curie grant agreement No. 777826 NoMADS, the Cantab Capital Institute for the Mathematics of Information and the Alan Turing Institute. \bibliographystyle{IEEEtran}

7c42a19a34a8e0e339fd08d117677fb273ed8a48

\section{Introduction} As from November 28, 2021, the SARS-CoV-2 Coronavirus has spread to 213 countries around the world, has over 262 million confirmed cases and has caused over 5.2 million deaths \cite{1} \cite{22}.\\ The SARS-CoV-2 Coronavirus causes a respiratory infection called COVID-19. The COVID-19 virus spreads primarily through droplets of saliva or discharge from the nose when an infected person coughs or sneezes \cite{36}.\\ Given that access to vaccines is limited and treatments are still in development, minimizing the spread through timely testing and isolating infected people is one of the best effective defenses humanity has at its disposal against COVID-19. However, the ability to deploy this strategy depends on each country’s capability to test significant fractions of their populations, including those that are not yet in contact with the health system. Agile, scalable and proactive testing have become the key differentiator for managing this pandemic. \subsection{Diagnostic strategies} The main types of tests used for COVID-19 are: \begin{itemize} \item Rapid serology antibody tests (ICT, LFA) \item Enzyme Linked Immunosorbent Assays (ELISA) \item Real-time polymerase chain reaction (RT-PCR) through swab tests \end{itemize} Rapid serology tests and ELISAs detect the presence of IgG / IgM antibodies produced by the immune system in response to a SARS-CoV-2 infection, while RT-PCR tests are molecular tests that detect the presence of viral RNA of SARS-CoV-2.\\ Triage remains as the fastest method for pre-screening a person, and this method is the one that can be scaled faster and at a low cost using exponential technologies. Triage is the process of sorting people based on their need for medical treatment as compared to their chance of benefiting from such care. Triage is usually performed in emergency rooms, disasters and wars, when limited medical resources must be allocated to maximize the number of survivors. During infectious disease outbreaks, triage is particularly important to separate patients likely to be infected with the pathogen of concern.\\ The main route of transmission of COVID-19 is through respiratory droplets generated when an infected person coughs or sneezes. Any person who is in close contact with someone who has respiratory symptoms (e.g., sneezing, coughing, etc.) is at risk of being exposed to potentially infective respiratory droplets \cite{37}.\\ Recently, the US FDA \cite{18} has written about the importance of expanding rapid COVID-19 serological antibody testing. Current laboratory diagnostic tests for COVID-19 are based on labor-intensive molecular techniques (RT-PCR swab tests) and have generally been reserved for patients whose disease severity, age and/or comorbidities place them in high risk of serious illness, specially in countries with limited testing resources.\\ Recent work has also described that SARS-CoV-2 viral shedding often occurs in early pre-symptomatic stages, which could affect the dynamics and diagnostic accuracy of SARS-CoV-2 RT-PCR swab tests, especially if patients do not undergo molecular diagnostic tests until several days after the onset of symptoms. On the other hand, COVID-19 rapid serology antibody tests detect IgM/IgG antibodies, which are generated 1-3 weeks after the onset of symptoms.\\ Together, the SARS-CoV-2 RT-PCR swab test, combined with a rapid COVID-19 serological antibody test provides a more complete picture of COVID-19 disease progression, in a given individual and across populations. \subsection{Importance of early diagnosis} Many authors have highlighted that the key to safely returning back to normal life \cite{35}, especially while we wait for the COVID-19 vaccination process to be completed, does not lie in a single "perfect" COVID-19 test - but through multiple rapid and affordable tests and non-clinic pre-screening tools.\\ RT-PCR tests have been the gold standard for COVID-19 diagnosis. Nevertheless, their high cost and relatively long processing times make them a less appealing option when deploying population wide testing strategies. On the other hand, antibody and antigen detection tests have proved to be an excellent complement to RT-PCR tests. Their lower costs, shorter processing times and higher availability allows for a more wide spread testing strategy and also shorter testing-isolation times. \subsection{Difficulties with early diagnosis} Currently, existing RT-PCR tests take considerably long to deliver results, are expensive to run frequently, or are often considered uncomfortable, which discourages people from going to get tested.\\ There are already some rapid tests for COVID-19 on the market. The Sofia SARS antigen test performed by the Quidel Corporation is approximately 99\% accurate, takes about 15 minutes to deliver results, and is already in use. A pharmacy in Arkansas, USA, reported that they had dispensed more than 800 in about two weeks. The test costs U\$D 95, which is a considerable price for most people’s budget, especially if they have to take the test periodically.\\ OpenCovidScreen \cite{17}, a coalition of scientists and biotech leaders working to drive innovation in COVID-19 testing and, in turn, accelerate return to work and school, says that a good COVID-19 testing solution needs four items: frequency (at least weekly), more test points, an affordable price (around U\$D 10-20), and an easy method of collection (e.g. saliva, quick test, cough). Currently, there are tests available on the market that address one or two of these considerations, but there is nothing that brings all four together. \subsection{Audio signals for pre-screening and disease monitoring} Audio signals generated by the human body (e.g. sighs, breathing, heart, digestion) have often been used by physicians and clinical researchers in the diagnosis and monitoring of diseases. However, until recently, such signals were generally collected by manual auscultation during scheduled visits to a medical facility. Medical research has begun to use digital technology to collect body sounds (e.g. digital stethoscopes) and perform automated data analysis \cite{1}, (e.g. for wheezing detection in asthma \cite{2} \cite{3}).\\ Researchers have also been testing the use of the human voice to aid in the early diagnosis of a variety of diseases. For example, in Parkinson's disease the patient’s speech is analyzed for variations in smoothness to determine if there exists a lack of coordination in the vocal muscles \cite{4} \cite{5}. Also vocal tone, rhythm, and volume have also been explored to diagnose illnesses such as post-traumatic stress disorder \cite{7}, traumatic brain injury, and some psychiatric conditions \cite{8}.\\ The use of human-generated audio as a biomarker for various diseases offers enormous potential for early diagnosis, as well as for affordable solutions that could be widely applied if integrated into devices such as cell phones. This is even more powerful if such solutions could monitor people throughout their daily lives in a discreet way.\\ Clinical work has focused on the use of voice analysis for specific diseases: for example, in Parkinson's disease, microphone and laryngoscope equipment have been used to detect the softness of speech resulting from lack of coordination on the vocal muscles \cite{6} \cite{12}. Voice characteristics have also been used to diagnose bipolar disorder; and to correlate pitch, rhythm, and volume with signs of invisible conditions such as post-traumatic stress disorder \cite{7}, traumatic brain injury, and depression. Voice frequency has been linked to coronary artery disease (as a result of hardening of the arteries that can affect voice production) \cite{6}. Companies such as Israel-based Beyond Verbal and Mayo Clinic have indicated in press releases that they are testing these approaches.\\ More recently, the microphone in basic devices such as smartphones and wearables has been used for sound analysis. In \cite{12}, the microphone is used to understand the context of the user and this information is added to generate a view of the surrounding environment of the recorded sound.. In \cite{13} the authors analyze the sounds emitted while the user is sleeping, to identify episodes of sleep apnea. Similar work has also used sound to detect asthma and wheezing \cite{2} \cite{3}. \subsection{IATos: proposal and background} Machine learning methods have been designed to recognize and diagnose respiratory diseases from sounds \cite{1} and more specifically cough: Bales et al.\cite{14} used convolutional neural networks (CNN) to detect cough within ambient audio and diagnose three potential diseases (bronchitis, bronchiolitis and whooping cough) based on its unique audio characteristics.\\ With the advent of COVID-19, researchers have begun to explore whether cough sounds could be diagnostic \cite{15}. A COVID-19-related cough detection study is presented in \cite{15} using a cohort of 48 patients tested with COVID-19 versus other pathological coughs, in which a set of models is trained. In \cite{11} the voice recordings of COVID-19 patients are analyzed to automatically classify the health status of the patients in four aspects, namely, the severity of the disease, the quality of sleep, fatigue and anxiety. Quatieri et al. \cite{26} showed that changes in vocal patterns could be a potential biomarker for COVID-19.\\ This research project aims to explore the use of human cough sounds as pre-diagnostic markers for COVID-19. Within the framework of this research study, we worked on the data acquisition of audio samples from people who quarantined in out-of-hospital isolation units for COVID-19 mild cases and close contacts or who were admitted to hospitals.\\ This research project differs from other similar initiatives in two main aspects. In the first place, the acquisition of audio samples was done through WhatsApp, and not through web pages. This allowed for a more streamlined data collection from the study participants and simplified the signing of informed consents. On the other hand, only people who underwent a PCR or antigen swab test participated in the study. Also, their results had to be available in the Buenos Aires City Health Ministry COVID-19 results database. This helped us classify the audio samples of each individual for the first 5k audio dataset, according to the result of their RT-PCR results, considered as the current "gold standard" for COVID-19 diagnosis. The second dataset, which we used to re-train the neural network, has 140k audio files. This second dataset is unbalanced (10\% corresponds to COVID-19 positive individuals and 90\% corresponds to COVID-19 negative individuals) and includes both audio files from people that had undergone an RT-PCR and antigen tests.\\ The expected outcome is to be able to analyze the patient’s cough as an additional marker that can be incorporated into the medical triage strategy. We seek to improve the pre-screening process that is carried out today prior to undergoing diagnosis testing methods such as PCR or rapid antigen tests. Having a digital Triage, which can be performed from anywhere using WhatsApp, allows us to exponentially scale the detection strategy for symptomatic and asymptomatic cases. \section{Methodology} Based on the objective of studying cough as a way to pre screen individuals for COVID-19, the research question that guided our work was: what is the sensitivity and specificity of a machine learning based COVID-19 cough classifier, compared to the standard diagnosis test for COVID-19?\\ In order to answer this question and prior to the development of the neural network, we consider the following steps as our methodology of work. The steps of this process are: \begin{itemize} \item Generate a data acquisition process \item Generate a dataset \begin{itemize} \item Define the selection process for study participants \item Contact people who agree to participate in the study \item Start the audio acquisition flow \end{itemize} \item Establish a preprocessing strategy for the collected audios \end{itemize} \subsection{Data description} Observational or experimental studies require a priori sample size calculation in order to achieve the necessary statistical power to reject the null hypothesis \cite{25}. In the field of pattern recognition though, there is no standardized priori determination method to evaluate the generated models performance based on sample size of the datasets used for training.\\ Review of previous literature from past research uses an empirical approach through learning curves and the use of surrogate models to provide an answer to the problem of sample size \cite{26}. However, these ex-post-facto solutions do not take into account the nature of the classificatory problem to be solved, the complexity of the model used, the use of feature augmentation techniques, or the use of pre-trained models with the transfer learning method.\\ As an approach to determining the necessary sample size, we revised the literature that refers to the use of pattern recognition models applied to the audio signal produced by coughing. We reviewed pre-existing literature and studied the differences of different samples sizes and their related metrics (see Table \ref{sample-table}).\\ \begin{table}[h] \caption{Bibliographic search of works that refer to the use of pattern recognition models applied to the audio signal produced by the act of coughing.} \begin{minipage}{\textwidth} \begin{tabular}{lcrrrrr} \hline\hline Author (Year) & Samples & COVID + & Others\footnote{Another non-COVID-19 positive respiratory disease.}& Control & S/E & AUC\\ \hline Imran (2020) & 1838 & 70 & 226 & 247 & .96/.95 & NR\\ \noalign{\vspace {.5cm}} Brown (2020) & 491 & 141 & 52 & 298 & NR & 0.875\\ \noalign{\vspace {.5cm}} Bales (2020) & 268 & - & 268 & - & .89/.948 & NR\\ \hline\hline \end{tabular} \vspace{-2\baselineskip} \end{minipage} \label{sample-table} \end{table} In light of the studied literature we established that a sample size of 5000 cough audio sounds could enable the correct performance of the generated model.\\ The data acquisition process was approved by the Ethics Committee of the Elizalde Hospital (Comité de Etica - Hospital Elizalde). In August 2020 we started the process of collecting cough audios through the Buenos Aires City Government chatbot with the aim of creating an Open Voice dataset for COVID-19 cough discrimination.\\ We only used samples with two conditions: \begin{itemize} \item The person providing the cough had undergone a PCR test. \item The audio files were recorded within 3 days of that PCR test. \end{itemize} At the end of the data acquisition process, we had collected 6,000 audio files, from 2,821 individuals. This collection of audio files was split into two groups: \begin{itemize} \item Training and validation: 2,412 audio files from individuals that tested negative for COVID-19, and 2,477 audio files for individuals that tested positive These audio files belong to 2,329 individuals. \item Test: 995 audio files that belong to 492 individuals. \end{itemize} Each person included in the study had to generate a variety of representative audios of the act of coughing. We collected on average 3 coughs per subject accompanied by general subject information: age, sex, date of the swab/cough, location, outcome of medical diagnosis, and finally information regarding timing of onset of COVID-19 signs and symptoms (fever, tiredness, sore throat, difficulty breathing, persistent pain or pressure in the chest, diarrhoea and coughing). \begin{center} \begin{figure}[htp] \centering \includegraphics[width=12cm]{DATASET} \caption{Dataset general information.} \label{fig:dataset} \end{figure} \end{center} Data was anonymized before being collected on our secure server and samples were saved without compression in WAV format. Samples that had no coughs, too much noise or were silent were removed. No segmentation was performed on the cough recordings used to train and test the neural network. \subsection{Data acquisition} Individuals who participated in the study were asked to send one audio per day, for 3 consecutive days following the PCR test. The primary goal was to obtain multiple samples per person. Secondarily, we seek to study the evolution of audio patterns to detect if there are changes that show evidence of the evolution of the virus.\\ When someone went to a testing center, health system personnel invited them to participate in this study. If the person agreed, they were guided to start a conversational flow on the Buenos Aires city government WhatsApp chatbot specifically designed to address citizen inquiries related to the coronavirus pandemic (COVID-19). From the beginning of the study until September 30, 2020 people signed a printed informed consent to participate in this study. From September 30, 2020 the informed consent was digitized into the WhatsApp chatbot conversation. \begin{figure}[htp] \centering \includegraphics[width=12cm]{DATAACQUISITION} \caption{Data acquisition timeline.} \label{fig:timeline} \end{figure} The audio files that were collected for this study belong to individuals who were swabbed in the City of Buenos Aires in public and private facilities, out-of-hospital isolation units for patients with confirmed COVID mild cases and testing points of the strategic testing program "DetectAR". The health personnel who performed the swab briefly explained the study to the individuals and invited them to participate. \begin{figure}[htp] \centering \includegraphics[width=12cm]{MAPA} \caption{Data acquisition sites withing the Buenos Aires health system.} \label{fig:sites} \end{figure} Data collection sites: 19 Febrile Emergency Units (FEU), 12 testing points of the strategic testing program "DetectAR", 14 out-of-hospital isolation units for COVID-19 mild cases and close contacts and 1 private medical center as well as government personnel. \subsubsection{Update on data acquisition} A second dataset containing 140.530 audio coughs was collected during the months of April to October 2021, and has 18.271 audios from individuals that tested positive for COVID-19 and 122.259 from individuals that tested negative.\\ This second data acquisition process was carried on from April 2021 until October 2021, and the same method for audio collection was applied. In this case, we incorporated information related to the clinical triage.\\ Each individual answered if they had symptoms of Covid-19 at the time of sending the audio, if they were in close contact with someone with COVID-19, if they had breathing problems or if they were part of a risk group. Additionally, the data corresponding to the type of test that was carried out was stored as a variable for future studies. On this occasion, the samples were not analyzed looking for the percentage of men and women who participated, nor the age groups, assuming that the sample is a representation of the population of the City of Buenos Aires that uses the public testing system. \subsection{Data preprocessing} The preprocessing pipeline consists of sampling, denoising and cough discrimination from noise and silence audio chunks.\\ Although these could be extended to more stages, our current proposal (Fig. X) consists of: \begin{itemize} \item Detects the presence of a cough in an audio, and cuts an audio segment of 1 second maximum duration. \item Normalize the audios in amplitude so that they all have a similar amplitude domain. \item Denoise audio files that have a low signal-to-noise ratio (SNR). \end{itemize} \begin{figure}[htp] \centering \includegraphics[width=12cm]{PREPROSSECING} \caption{Pre-processing pipeline} \label{fig:prepro} \end{figure} As the first stage of preprocessing the audios, they are converted from the .ogg format to the .wav format, and all the audios are sampled equally at 16 kHz.\\ Subsequently, the audio files are cut into 1 second chunks. Then, using YAMNet and a network trained from the collected audio files, audio chunks that contain coughs are separated from those that are noise or silence. Finally, the resulting fragments are filtered to remove any noise that the signals may have.\\ The audios are then preprocessed in order to extract their main characteristics. This was achieved through a spectral analysis and selection of features of the signal. Finally the audio samples were normalized in amplitude.\\ From the information extracted, images were generated and introduced into a deep learning model, in order to be able to classify new cough audio signals.\\ For data processing, we mainly used a transformation that has broad consensus in the scientific community. In the first place, from the audio signal, the Mel \cite{27} spectrogram was obtained, extracting the most relevant characteristics of the sound using human auditory perception as a criterion. The transformation consists of splitting the signal into short frames, applying the discrete Fourier transform (DFT) to generate the spectrogram, then applying the Mel Filterbank -to calculate the energy in each Mel filter- and finally calculating the logarithm of the energy for each filterbank \cite{28}.\\ We used regularization techniques in our data since, given the complexity of the model to be applied, this will allow a better generalization of them. Specifically, to feed the model we used data augmentation, that is, we performed small random transformations to create new images, for example scaling and/or rotating the data. This technique allowed us to generalize the information and also obtain an improvement in the performance over audio signal data. \section{Neural network model} A deep learning-based COVID-19-associated cough detector was created. That is, a three-class classifier based on Convolutional Neural Networks that uses Mel's spectrograms as input information. Due to the subtle differences that exist between images of coughs corresponding to healthy patients, with various respiratory diseases, or with COVID-19, a neural network with more layers was required. Therefore, we used a CNN architecture.\\ \begin{figure}[htp] \centering \includegraphics[width=10cm]{CNN2} \caption{CNN Architecture.} \label{nn} \end{figure} Regarding the architecture, CNNs have a set of convolutional layers, to which pooling layers are then added and the dimensions of the matrix are adjusted. This is repeated successively and in the end, a fully connected layer is added to perform the prediction through a softmax.\\ For this project, three widely accepted scientific metrics will be used: Specificity, Recall/Sensitivity and Accuracy. In the training process of the model, a constant monitoring of the precision metric was carried out in order to generate the least amount of false positives. Due to the way COVID-19 spreads, it will always be a priority to maintain a high criterion of reliability and certainty before recommending a patient to seek assistance at a health facility. In parallel, the AUC measurement was monitored, which represents the area under the ROC curve that allows visualizing the sensitivity versus specificity.\\ We started by using a classification threshold of 0.5. Our proposed architecture, drawn in Fig. \ref{nn}, takes a recording with one or more coughs, performs pre-processing steps and inputs it into a CNN based model to output a pre-screening result. As pre-processing, each input audio file is split into 1 second audio chunks, padded as needed, processed with the MFCC package and subsequently passed through the neural network classifier. The output of these steps becomes the input to a CNN.\\ In order to train the classifier model, the set of spectrograms obtained were divided into train (80\%), validation (10\%) and internal test (10\%). Finally, with the model obtained, inferences were made to a second set of audio files that belong to a test set of 492 individuals. In this way we were able to evaluate the performance of the model on new data that it had never analyzed before. \section{Results} As we explained in the previous section, in order to evaluate the model we use the performance metrics of accuracy, specificity and sensitivity/recall on the test set. The accuracy here refers to the overall accuracy of the model. These performance metrics are based on mean confusion matrices from cross-validation. Tuning of the various hyper-parameters (number of hidden layers, learning rate, activation functions, dropout rate) of deep neural network-based models has also been performed, based on the cross-validation accuracy. Furthermore, the decay of model loss versus the number of epochs has been investigated to rule out the possibility of overfitting. \subsection{Cough discrimination accuracy} When we receive an audio file, the first step is to discriminate whether it has a cough or not. For this step we combine the use of YAMNet and a neural network that we specially trained to distinguish coughs. The analyzed audio file is cut into 1-second chunks and these fragments are evaluated by this detection stage, to discard the fragments that do not have the presence of cough. We achieved an accuracy of 97\% determining which fragments have a cough and which have noise compared to a random selected, manually classified sample of 100 1 second audio chunks. This allowed us to move on to the next stage, which is to determine if the audio fragments tagged as “cough” by the YAMNet algorithm, correspond to that of an individual who is potentially COVID-19 positive or not. \subsection{COVID-19 detection} At the moment and based on the collected data, the overall accuracy of the deep learning based classifier is 86.00\%, Recall is 0.89 and F1-score is 0.87. The mean normalized confusion matrix resulting from this approach is shown in Fig. \ref{matrixnn}. Future work will continue to improve this model as more training data becomes available. Fig. XX shows the mean loss versus epochs of the neural network classifier, for both training and validation data sets. Both the curves start to saturate after around 70 epochs, indicating a reasonable learning time, without overfitting. \begin{figure}[htp] \centering \includegraphics[width=10cm]{matrixnn} \caption{Confusion matrix.} \label{matrixnn} \end{figure} \section{Discussion and future work} We have shown that COVID-19 can be discriminated with 86\% accuracy using only a forced-cough audio recording. We find most remarkable that our model performs better with coughs from positive patients than those from negative ones.\\ This first stage of developing the model focused on training it on a large dataset to learn good features for discriminating COVID-19 forced-coughs. The results when evaluating the testing set (made with coughs from individuals diagnosed with a validated PCR and antigen tests) serve as an indicator that the model would have similar accuracy when deployed. The next stage of this project is turning on the NN classifier together with the triage that is currently being used in the Buenos Aires City WhatsApp chatbot. We will also gather more quality data that can further train, fine-tune, and validate the model in a real world environment.\\ Since there might be cultural and age differences in coughs, future work could also focus on tailoring the model to different age groups and regions of the world using the metadata captured, and possibly including other sounds or input modalities such as vision or natural language symptom descriptions. \subsection{Potential use cases of IATos} Motivated by an urgent need, we developed a research protocol that was intended to evaluate the usefulness of an AI-based preliminary diagnostic tool for COVID-19, ubiquitously scalable through WhatsApp audio files containing a recording of an individual's forced cough.\\ The central idea of such a tool is inspired by previous studies \cite{10} \cite{16} showing that coughing can be used as a pre-screening tool for the diagnosis of a variety of respiratory diseases using AI techniques. The main objective is that the cough analysis is a complement and never a replacement for other diagnostic methods.\\ \begin{figure}[htp] \centering \includegraphics[width=10cm]{triage} \caption{Triage flow with IATOS.} \label{triage} \end{figure} This non-invasive, free to access and real-time pre-screening tool may prove to have a great potential to complement current efforts to contain the disease spread in many different contexts. Possible use cases could include: \begin{itemize} \item Population daily screening tool: As workers go back to work, students go back to school, and commuters use public transport, to name a few, methods are required to screen infected COVID-19 carriers, especially asymptomatics. The only screening method currently available is using thermometers, however this study \cite{36} showed only 45\% of mild-moderate COVID-19 cases have fever (this represents 9\% of COVID-19 positives when asymptomatics are included). Meanwhile our AI tool has shown to discriminate against a forced cough of COVID-19 positives with 86\% accuracy and could act as a complement to the medical triage questions. \item Pre-selection of candidates for test pooling: The test pooling strategy is expected to be employed in many countries, especially in low-incidence areas to rapidly identify a subgroup of individuals likely to be infected, however, “preliminary results show there is no dilution and no decrease on test sensitivity when mini pools of five samples each are used” \cite{37}. Group testing with our tool as shown in Fig. \ref{triage}, could pre-screen school classrooms, factories or even countries on a daily basis signalling probable infected candidate groups for smaller test pooling batches. \end{itemize} This technology could be assembled with low levels of effort and risk within the virtual triage that is carried out as a follow-up protocol for close contacts of positive patients.\\ In the future, subject to the learning curve and validation of the model, additional use cases could include \cite{10}: \begin{itemize} \item Complementing temperature scanners at airports, borders or other key places where the virus circulates. \item Allowing remote pre-monitoring for anyone, anywhere, at any time, regardless of the existing infrastructure/testing facilities. \item Providing a centralized record of tests with location and time stamps. The data collected from the app could serve as input for long-term health care planning and health policy formulation. \end{itemize} Despite the good performance that is preliminarily observed in these tools, they are not intended as a replacement for clinical tests but as a complement. The goal is to investigate the potential for a single functional tool to monitor, track and control the rampant spread of the global pandemic in a timely, cost-effective and most importantly safe manner, by allowing a pre-test for anyone with access to a smartphone and an internet connection. \section{Conclusion} We presented an AI pre-screening tool that uses a neural network classifier that is able to discriminate COVID-19 positives with 86\% accuracy from a forced-cough recording, at essentially no cost and with an exponential scaling capacity.\\ Despite its performance, IATOS is not meant to compete with clinical testing. Instead, it offers a unique functional tool for timely, cost-effective and most importantly safe monitoring, tracing, tracking and thus, controlling the rampant spread of the global pandemic by virtually enabling testing for everyone.\\ The proposed solution could be seen as a group outbreak detection tool for pre-screening whole-populations on a daily basis, taking the pre-screening to each house.. At the same time, this tool will allow us to send more people, potentially asymptomatic, to be tested. In this way we may be able to detect many positive asymptomatic cases that would normally go undetected.\\ As part of our ongoing pilots, data pipelines with FEUs and detection points have been set up to continue to improve the neural network. We plan on leveraging this data to further train and validate our models with the aim of improving pandemic management practices.\\ As part of our efforts to make data public and contribute to open innovation, we are opening our research methods and our dataset to collaborate and inspire others to develop similar tools to fight the COVID-19 pandemic, and hopefully be better prepared for the next pandemic. \section{Ethical aspects} The study will be guided by the rules of good clinical practice, the Declaration of Helsinki and the regulations of the Government of the City of Buenos Aires in force. Authorization was requested from the Research Ethics Committee, the DGDIYDP (Dirección General de Docencia, Investigación y Desarrollo Profesional) and the SSPLSAN (Subsecretaría de Planificación Sanitaria. Ministerio de Salud). Informed consent of those responsible, subjects and / or consent of minors will be requested as appropriate. Due to the context of the study and the specific SOPs of the Ethics Committee of the Elizalde Hospital (CEI-HGNPE, Hospital General de Niños Pedro de Elizalde), the documentation of the process will be protected in images of the forms duly signed by the subjects or their managers, who will keep the originals. The audio data of the patients and their COVID-19 test results will be stored without direct or indirect identifiers. Because it is an observational study and is framed in the context of the COVID-19 pandemic, the CEI-HGNPE requested its expeditious evaluation and approval.\\ \textbf{Data Availability} All the data referred to in the manuscript are available in https://data.buenosaires.gob.ar/.\\ \textbf{Declaration of competing interest} The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.\\ \textbf{Acknowledgement} This work is dedicated to those affected by the COVID-19 pandemic and those who are helping to fight this battle in any way they can.\\

c22c181b42bb1cfacd76299f2fa26d17b7559bbe

\section{Evaluation on Quantum Causal Synthetic Data: Depolarizing Quantum Channel (Model \ref{model:depolar})}\label{eval:moreDepolar} To find the best choice of hyper-parameters in {\sc \textbf{\texttt{QInferGraph}}}~i.e., $\alpha$ and $\beta$, we evaluate the performance of {\sc \textbf{\texttt{QInferGraph}}}~for causal structure discovery in Model \ref{model:depolar} by considering different intervals for the value of $\beta$ and different values of $\alpha$. Then, we will discuss the best choice of $\alpha$ and $\beta$ in these settings. \paragraph{Part I: Latent Graph.} In this part, we evaluate the performance of {\sc \textbf{\texttt{QInferGraph}}}~for the identification of latent graphs in Model \ref{model:depolar} (Part I) as follows. We run {\sc \textbf{\texttt{QLatentSearch}}}~(Algorithm 1) on 50 different values of $\beta$ uniformly spaced in the interval (0.2,0.3), (0.6,0.7), (0.7,0.8), and (0.8,0.9), respectively. {\sc \textbf{\texttt{QInferGraph}}}~(Algorithm 2) calls {\sc \textbf{\texttt{QLatentSearch}}}, and for each $\beta$ the algorithm {\sc \textbf{\texttt{QLatentSearch}}}~is executed for 500 iterations. For different values of $\alpha=0.7, 0.8, 0.9, 1$ the results are summarized in Table \ref{t:5}. In the table, \textcolor{blue}{T} means that {\sc \textbf{\texttt{QInferGraph}}}~(Algorithm 2) identifies the latent graph correctly. But, \textcolor{red}{F} means that the algorithm fails to identify the latent graph. Our proposed algorithm i.e., {\sc \textbf{\texttt{QInferGraph}}}~works well where $\beta$ is in (0.2,0.3), (0.6,0.7), (0.7,0.8), and (0.8,0.9). The results confirm our observations that we made in Model \ref{ex:generalqsc} (Part I). However, in this case {\sc \textbf{\texttt{QInferGraph}}}~has a higher performance quality, in almost all cases. \begin{table}[!htb] \caption{Validation of Latent Graph in Model \ref{model:depolar} (Part I) via {\sc \textbf{\texttt{QInferGraph}}}, where $\alpha=0.7, 0.8,0.9,1$ and $\beta\in (0.2,0.3), (0.6,0.7), (0.7,0.8), (0.8, 0.9)$.}\label{t:5} \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p_2$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.01& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.1& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.2& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.3& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.4& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ $p_1$&0.5& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.6& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.7& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.8& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.9& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.99& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ \cline{2-13} \end{tabular} \end{table} \paragraph{Part II: Direct Graph} In this part we consider direct graphs in Model \ref{model:depolar} (Part II), and we use the same parameters explained in Part I of this section. In all cases for the different values of $\alpha$ considered, the results show that there is no false positive in this case indicating the desired causal inference. The results, Table \ref{t:9}, indicates that in quantum noisy channels with very small probability of errors, {\sc \textbf{\texttt{QInferGraph}}}~very likely fails to draw the right conclusion about the identification of latent confounders. \begin{table}[!htbp] \caption{Validation of Direct Graph in Model \ref{model:depolar} (Part II).} \label{t:9} \begin{subtable}{.5\linewidth} \caption{$\beta=(0.2,0.3)$} \centering \resizebox{\columnwidth}{!}{% \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.7&\textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ &0.8 & \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ $\alpha$&0.9&\textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ &1&\textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ \cline{2-13} \end{tabular}% } \end{subtable}% \begin{subtable}{.5\linewidth} \centering \caption{$\beta=(0.6,0.7)$} \resizebox{\columnwidth}{!}{% \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.7&\textcolor{red}{F}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ &0.8 & \textcolor{blue}{T}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ $\alpha$&0.9&\textcolor{red}{F}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ &1&\textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ \cline{2-13} \end{tabular}% } \end{subtable} \vspace{.5cm} \begin{subtable}{.5\linewidth} \caption{$\beta=(0.7,0.8)$} \centering \resizebox{\columnwidth}{!}{% \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.7&\textcolor{red}{F}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ &0.8 & \textcolor{red}{F}& \textcolor{red}{F} &\textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ $\alpha$&0.9&\textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ &1&\textcolor{red}{F}& \textcolor{red}{F} &\textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ \cline{2-13} \end{tabular}% } \end{subtable}% \begin{subtable}{.5\linewidth} \centering \caption{$\beta=(0.8,0.9)$} \resizebox{\columnwidth}{!}{% \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.7&\textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ &0.8 & \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ $\alpha$&0.9&\textcolor{red}{F}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ &1&\textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ \cline{2-13} \end{tabular}% } \end{subtable} \end{table} \section{Evaluation on Quantum Causal Synthetic Data: Generalized Quantum Symmetric Channel (Model \ref{ex:generalqsc})}\label{eval:moreGQSC} In this section, we apply {\sc \textbf{\texttt{QInferGraph}}}~to a generalized version of binary symmetric channel, as explained in Model \ref{ex:generalqsc}, where a qubit is transmitted over a binary symmetric channel (to illustrate the case of no confounder), or a qubit is transmitted over two separate channels (to illustrate the case of latent confounder). To build the model, we follow the same scheme that we explained in Model \ref{model:classicchannel}. \begin{model}[Generalized Quantum Symmetric Channel: Latent and Direct Graph]\label{ex:generalqsc} \emph{\textbf{Part I: Latent Graph.}} Assume that there are real numbers $\gamma$ and $\lambda$ such that $\gamma^2+\lambda^2=1$, and $X_1=Y_1=Z_1=\gamma |0\rangle + \lambda|1\rangle$, and $X_2=Y_2=Z_2=\gamma |0\rangle - \lambda|1\rangle$. Let $Z$ be in the mixed state $|Z_1\rangle |X_1\rangle |Y_1\rangle$ with probability $q$, and $|Z_2\rangle |X_2\rangle |Y_2\rangle$ with probability $1-q$. We consider a generalization of Quantum Symmetric Channel in the following model, in which the phase of the qubit is reversed with certain probability. The second and third qubits are transmitted over two separate quantum symmetric channels with error probability $p_1$ and $p_2$, respectively, and are labeled $X$ and $Y$, respectively. Thus, the joint density matrix of $X,Y$ and $Z$, $\rho_{ZXY}^{\gamma,\lambda}$, can be written as mixtures of the following pure density matrices: \begin{equation*} \left\{ \begin{array}{lll} |Z_1\rangle |X_1\rangle |Y_1\rangle (|Z_1\rangle |X_1\rangle |Y_1\rangle)^\dagger & & q(1-p_1)(1-p_2)\\ |Z_1\rangle |X_1\rangle |Y_2\rangle (|Z_1\rangle |X_1\rangle |Y_2\rangle)^\dagger & & q(1-p_1)p_2 \\ |Z_1\rangle |X_2\rangle |Y_1\rangle (|Z_1\rangle |X_2\rangle |Y_1\rangle)^\dagger & & qp_1(1-p_2) \\ |Z_1\rangle |X_2\rangle |Y_2\rangle (|Z_1\rangle |X_2\rangle |Y_2\rangle)^\dagger & & qp_1p_2\\ |Z_2\rangle |X_1\rangle |Y_1\rangle (|Z_2\rangle |X_1\rangle |Y_1\rangle)^\dagger & & (1-q)p_1p_2\\ |Z_2\rangle |X_1\rangle |Y_2\rangle (|Z_2\rangle |X_1\rangle |Y_2\rangle)^\dagger & & (1-q)p_1(1-p_2) \\ |Z_2\rangle |X_2\rangle |Y_1\rangle (|Z_2\rangle |X_2\rangle |Y_1\rangle)^\dagger & & (1-q)(1-p_1)p_2 \\ |Z_2\rangle |X_2\rangle |Y_2\rangle (|Z_2\rangle |X_2\rangle |Y_2\rangle)^\dagger & & (1-q)(1-p_1)(1-p_2) \end{array} \right. \end{equation*} Now, we trace out $Z$ to obtain the density matrix for the latent graph $X\leftrightarrow Y$. Then, we apply {\sc \textbf{\texttt{QInferGraph}}}~(Algorithm \ref{alg:qInferGraph}) on $\rho_{XY}$ to verify that $X$ and $Y$ are confounded by $Z$. For this purpose, we use {\sc \textbf{\texttt{QLatentSearch}}}~(Algorithm \ref{alg:qLatentSearch}) on 50 different values of $\beta$, uniformly spaced in the interval $(0.7, 0.8)$. More results for $\beta\in (0.2,0.3)$, $\beta\in (0.6,0.7)$, and $\beta\in (0.8,0.9)$ can be found at the end of this section. We run {\sc \textbf{\texttt{QLatentSearch}}}~for 500 iterations each time. We use the conditional mutual information threshold of 0.05. In other words, of the algorithm outputs for the 50 $\beta$ values used, we pick the smallest entropy $W$ discovered by the algorithm among those that ensure $I(X; Y |W) \le 0.05$. Table \ref{t:LatentQSC} summarizes the results for different $S(W)$ threshold that is determined by $\theta=\alpha\min\{S(X),S(Y)\}$. For different values of $\alpha=0.7,0.8,0.9,1$, the results are the same, and are given in Table \ref{t:LatentQSC}. We let $q=0.4$. In the Table, $\textcolor{blue}{T}$ means that {\sc \textbf{\texttt{QInferGraph}}}~(Algorithm \ref{alg:qInferGraph}) identifies the latent graph correctly. But, $\textcolor{red}{F}$ means that the algorithm fails to identify the latent graph. For very small or very large $p_i$'s the case of latent confounder or direct graph are hard to separate, while the proposed algorithm works well in most other cases. \begin{table} \vspace{-.3in} \caption{Validation of Latent Graph in Model \ref{ex:generalqsc} (Part I) for $\alpha=0.7,0.8,0.9,1$ with the density matrix obtained from $\rho_{ZXY}^{1/\sqrt{2},1/\sqrt{2}}$ via tracing out $Z$. }\label{t:LatentQSC} \centering \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p_2$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.01& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.1& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.2& \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ &0.3& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.4& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ $p_1$&0.5& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.6& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.7& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.8& \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ &0.9& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.99& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ \cline{2-13} \end{tabular} \end{table} \emph{\textbf{Some highlights for results in Part I:}} (1) Note that where the probability of errors i.e., $p_1$ and $p_2$ are very small, the latent confounder $Z$ is hardly distinguishable from $X$ (or $Y$) and {\sc \textbf{\texttt{QInferGraph}}}~fails to discover the latent graph. (2) Note that {\sc \textbf{\texttt{QLatentSearch}}}~tries to find the stationary point(s) of the loss function $L$ in Equation (\ref{eq:qlossfunc}), and there is no guarantee to find the global optimum. However, the performance of {\sc \textbf{\texttt{QInferGraph}}}~in this case is acceptable: true positive rate (recall) = 0.77, false positive rate (fall-out) = 0, false negative rate (miss rate) = 0.23, accuracy = 0.77. (3) The hyperparameter $\alpha$ does not affect significantly on the quality of results in our experimental settings that indicates {\sc \textbf{\texttt{QInferGraph}}}~is not very sensitive to hyperparameters. \emph{\textbf{Part II: Direct Graph.}} We consider a generalization of Quantum Symmetric Channel in the following model, in which the phase of the qubit is reversed with certain probability. We consider a mixed state $|X_1\rangle |Y_1\rangle$ with probability $q$, and $|X_2\rangle |Y_2\rangle$ with probability $1-q$. Further, the second qubit is transmitted over the quantum symmetric channel with the error probability $p$. After the transmission, the two qubits are labeled $X$ and $Y$, respectively. Thus, the joint density matrix of $X$ and $Y$ is the mixture of the following pure density matrices: \begin{equation*} \left\{ \begin{array}{lll} |X_1\rangle |Y_1\rangle (|X_1\rangle |Y_1\rangle)^\dagger & & q(1-p)\\ |X_1\rangle |Y_2\rangle (|X_1\rangle |Y_2\rangle)^\dagger & & qp \\ |X_2\rangle |Y_1\rangle (|X_2\rangle |Y_1\rangle)^\dagger & & (1-q)p \\ |X_2\rangle |Y_2\rangle (|X_2\rangle |Y_2\rangle)^\dagger & & (1-q)(1-p) \end{array} \right. \end{equation*} We already know that $X$ is the cause of $Y$ in this scenario, i.e., $X\to Y$ is the corresponding directed graph. To verify this, we use Algorithm \ref{alg:qLatentSearch} and \ref{alg:qInferGraph} as we explained earlier in this model. The results are summarized in Table \ref{t:DirectQSC} for $q=0.4$. $\textcolor{blue}{T}$ means that {\sc \textbf{\texttt{QInferGraph}}}~(Algorithm \ref{alg:qInferGraph}) identifies the direct graph correctly. But, $\textcolor{red}{F}$ means that the algorithm fails to identify the direct graph. In all cases for the different values of $\alpha$ considered, the results show that there is no false positive in this case indicating the desired causal inference. \begin{table}[!htbp] \caption{Validation of Direct Graph in Model \ref{ex:generalqsc} (Part II) with the density matrix $\rho_{XY}^{1/\sqrt{2},1/\sqrt{2}}$.} \label{t:DirectQSC} \centering \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.7&\textcolor{blue}{T} & \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ &0.8 & \textcolor{blue}{T}& \textcolor{blue}{T} &\textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ $\alpha$&0.9&\textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ &1&\textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ \cline{2-13} \end{tabular} \end{table} \end{model} Figure \ref{fig:ISplane} illustrates the trade-off between the von Neumann entropy of the latent common cause and the conditional quantum mutual information of the observed systems for two examples of Model \ref{ex:generalqsc} in Part I and Part II. The conditional quantum mutual information between 0 and 0.05 in the I-S plot indicates that the suitable conditional quantum mutual information threshold for this setting is $\approx 0.05$. \begin{figure} \centering \includegraphics[width=.5\linewidth]{ISplane.pdf} \caption{Trade-off curve discovered by {\sc \textbf{\texttt{QLatentSearch}}}~for: (i) Model \ref{ex:generalqsc} (Part I) with $q=0.4$ and $p=0.1$ [\textcolor{orange}{right-side curve}], (ii) Model \ref{ex:generalqsc} (Part II) with $q=0.4$, $p_1=0.2$, and $p_2=0.3$ [\textcolor{blue}{left-side curve}].}\label{fig:ISplane} \end{figure} \begin{remark}\label{rem:cVSq} Note that the generalized quantum symmetric channel is a \emph{rotated} version of the classical binary symmetric channel. So, in this case we can convert the joint density matrix to a joint probability distribution using the rotational procedure in Algorithm \ref{alg:rotate}, we discuss this procedure later in Example \ref{ex:counterex}. Then, we can apply the classical causal inference algorithm i.e., Algorithm \ref{alg:InferGraph} to identify latent confounders. However, our results in Table \ref{t:LQSCclassic} and \ref{t:DQSCclassic} indicate that even in this classical scenario our quantum approach outperforms the classical causal inference method. Note that there always exist a trade off for choosing $\alpha=0.7, 0.8, 0.9, 1$ (and hence the $H(Z)$ threshold). As suggested and verified in \citep{KocaogluNEURIPS2020}, $\alpha=0.8$ seems an appropriate value for $\alpha$ in practice. Our results verifies this suggestion as well. \begin{table}[!htb] \caption{Validation of Latent Graph in Model \ref{ex:generalqsc} (Part I) via classical causal inference (Algorithm \ref{alg:InferGraph}).}\label{t:LQSCclassic} \begin{subtable}{.5\linewidth} \caption{$\alpha=0.7$} \centering \resizebox{\columnwidth}{!}{% \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p_2$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.01& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.1& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.2& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.3& \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ &0.4& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ $p_1$&0.5& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.6& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.7& \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ &0.8& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.9& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.99& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ \cline{2-13} \end{tabular}% } \end{subtable}% \begin{subtable}{.5\linewidth} \centering \caption{$\alpha=0.8$} \resizebox{\columnwidth}{!}{% \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p_2$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.01& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.1& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.2& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.3& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.4& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ $p_1$&0.5& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.6& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.7& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.8& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.9& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.99& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ \cline{2-13} \end{tabular}% } \end{subtable} \vspace{.5cm} \begin{subtable}{.5\linewidth} \caption{$\alpha=0.9$} \centering \resizebox{\columnwidth}{!}{% \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p_2$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.01& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.1& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.2& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.3& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.4& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ $p_1$&0.5& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.6& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.7& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.8& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.9& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.99& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F}\\ \cline{2-13} \end{tabular}% } \end{subtable}% \begin{subtable}{.5\linewidth} \centering \caption{$\alpha=1$} \resizebox{\columnwidth}{!}{% \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p_2$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.01& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.1& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.2& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.3& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.4& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ $p_1$&0.5& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.6& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.7& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.8& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.9& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.99& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ \cline{2-13} \end{tabular}% } \end{subtable} \end{table} \begin{table}[!htbp] \caption{Validation of Direct Graph in Model \ref{ex:generalqsc} (Part II) via classical causal inference (Algorithm \ref{alg:InferGraph}).} \label{t:DQSCclassic} \centering \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.7&\textcolor{blue}{T} & \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ &0.8 & \textcolor{blue}{T}& \textcolor{blue}{T} &\textcolor{blue}{T}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ $\alpha$&0.9&\textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ &1&\textcolor{red}{F}& \textcolor{red}{F} &\textcolor{red}{F}& \textcolor{red}{F} &\textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F} &\textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}\\ \cline{2-13} \end{tabular} \end{table} \end{remark} \clearpage To find the best choice of hyper-parameters in {\sc \textbf{\texttt{QInferGraph}}}~i.e., $\alpha$ and $\beta$, we evaluate the performance of {\sc \textbf{\texttt{QInferGraph}}}~for causal structure discovery in Model \ref{ex:generalqsc} by considering different intervals for the value of $\beta$ and different values of $\alpha$. Then, we will discuss the best choice of $\alpha$ and $\beta$ in these settings. \paragraph{Part I: Latent Graph.} In this part, we evaluate the performance of {\sc \textbf{\texttt{QInferGraph}}}~for the identification of latent graphs in Model \ref{ex:generalqsc} (Part I) as follows. We run {\sc \textbf{\texttt{QLatentSearch}}}~(Algorithm 1) on 50 different values of $\beta$ uniformly spaced in the intervals (0.2,0.3), (0.6,0.7), (0.7,0.8), and (0.8,0.9), respectively. {\sc \textbf{\texttt{QInferGraph}}}~(Algorithm 2) calls {\sc \textbf{\texttt{QLatentSearch}}}, and for each $\beta$ the algorithm {\sc \textbf{\texttt{QLatentSearch}}}~is executed for 500 iterations. For different values of $\alpha=0.7, 0.8, 0.9, 1$ the results are summarized in the following tables (Table \ref{t:1}, \ref{t:2}, \ref{t:3}, and \ref{t:4}), respectively. In the tables, \textcolor{blue}{T} means that {\sc \textbf{\texttt{QInferGraph}}}~(Algorithm 2) identifies the latent graph correctly. But, \textcolor{red}{F} means that the algorithm fails to identify the latent graph. For very small or very large $p_i$'s (probability of errors, see Model \ref{ex:generalqsc} for details) the case of latent confounder or direct graph are hard to separate, while the proposed algorithm i.e., {\sc \textbf{\texttt{QInferGraph}}}~works well in most other cases where $\beta$ is in (0.6,0.7), (0.7,0.8), and (0.8,0.9). \begin{table}[!htb] \caption{Validation of Latent Graph in Model \ref{ex:generalqsc} (Part I) via {\sc \textbf{\texttt{QInferGraph}}}, where $\alpha=0.7, 0.8$.}\label{t:1} \begin{subtable}{.5\linewidth} \caption{$\beta=(0.2,0.3)$} \centering \resizebox{\columnwidth}{!}{% \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p_2$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.01& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.1& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.2& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.3& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.4& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ $p_1$&0.5& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.6& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.7& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.8& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.9& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.99& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ \cline{2-13} \end{tabular}% } \end{subtable}% \begin{subtable}{.5\linewidth} \centering \caption{$\beta=(0.6,0.7)$} \resizebox{\columnwidth}{!}{% \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p_2$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.01& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.1& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.2& \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ &0.3& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.4& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ $p_1$&0.5& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.6& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.7& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.8& \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ &0.9& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.99& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ \cline{2-13} \end{tabular}% } \end{subtable} \vspace{.5cm} \begin{subtable}{.5\linewidth} \caption{$\beta=(0.7,0.8)$} \centering \resizebox{\columnwidth}{!}{% \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p_2$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.01& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.1& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.2& \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ &0.3& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.4& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ $p_1$&0.5& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.6& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.7& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.8& \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ &0.9& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.99& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ \cline{2-13} \end{tabular}% } \end{subtable}% \begin{subtable}{.5\linewidth} \centering \caption{$\beta=(0.8,0.9)$} \resizebox{\columnwidth}{!}{% \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p_2$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.01& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.1& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.2& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.3& \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ &0.4& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ $p_1$&0.5& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.6& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.7& \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ &0.8& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.9& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.99& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ \cline{2-13} \end{tabular}% } \end{subtable} \end{table} \begin{table}[!htb] \caption{Validation of Latent Graph in Model \ref{ex:generalqsc} (Part I) via {\sc \textbf{\texttt{QInferGraph}}}, where $\alpha=0.9$.}\label{t:2} \begin{subtable}{.5\linewidth} \caption{$\beta=(0.2,0.3)$} \centering \resizebox{\columnwidth}{!}{% \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p_2$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.01& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.1& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.2& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.3& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.4& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ $p_1$&0.5& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} \\ &0.6& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.7& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.8& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.9& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.99& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ \cline{2-13} \end{tabular}% } \end{subtable}% \begin{subtable}{.5\linewidth} \centering \caption{$\beta=(0.6,0.7)$} \resizebox{\columnwidth}{!}{% \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p_2$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.01& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.1& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.2& \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ &0.3& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.4& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ $p_1$&0.5& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.6& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.7& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.8& \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ &0.9& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.99& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ \cline{2-13} \end{tabular}% } \end{subtable} \vspace{.5cm} \begin{subtable}{.5\linewidth} \caption{$\beta=(0.7,0.8)$} \centering \resizebox{\columnwidth}{!}{% \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p_2$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.01& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.1& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.2& \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ &0.3& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.4& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ $p_1$&0.5& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.6& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.7& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.8& \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ &0.9& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.99& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ \cline{2-13} \end{tabular}% } \end{subtable}% \begin{subtable}{.5\linewidth} \centering \caption{$\beta=(0.8,0.9)$} \resizebox{\columnwidth}{!}{% \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p_2$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.01& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.1& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.2& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.3& \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ &0.4& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ $p_1$&0.5& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.6& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.7& \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ &0.8& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.9& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.99& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ \cline{2-13} \end{tabular}% } \end{subtable} \end{table} \begin{table}[!htb] \caption{Validation of Latent Graph in Model \ref{ex:generalqsc} (Part I) via {\sc \textbf{\texttt{QInferGraph}}}, where $\alpha=1$.}\label{t:3} \begin{subtable}{.5\linewidth} \caption{$\beta=(0.2,0.3)$} \centering \resizebox{\columnwidth}{!}{% \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p_2$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.01& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.1& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.2& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.3& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.4& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ $p_1$&0.5& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.6& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.7& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.8& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.9& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.99& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ \cline{2-13} \end{tabular}% } \end{subtable}% \begin{subtable}{.5\linewidth} \centering \caption{$\beta=(0.6,0.7)$} \resizebox{\columnwidth}{!}{% \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p_2$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.01& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.1& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.2& \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ &0.3& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.4& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ $p_1$&0.5& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.6& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.7& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.8& \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ &0.9& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.99& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ \cline{2-13} \end{tabular}% } \end{subtable} \vspace{.5cm} \begin{subtable}{.5\linewidth} \caption{$\beta=(0.7,0.8)$} \centering \resizebox{\columnwidth}{!}{% \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p_2$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.01& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.1& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.2& \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ &0.3& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.4& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ $p_1$&0.5& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.6& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.7& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.8& \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ &0.9& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.99& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ \cline{2-13} \end{tabular}% } \end{subtable}% \begin{subtable}{.5\linewidth} \centering \caption{$\beta=(0.8,0.9)$} \resizebox{\columnwidth}{!}{% \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p_2$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.01& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.1& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.2& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.3& \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ &0.4& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ $p_1$&0.5& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.6& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.7& \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ &0.8& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.9& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.99& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ \cline{2-13} \end{tabular}% } \end{subtable} \end{table} \clearpage \paragraph{Part II: Direct Graph} In this part we consider direct graphs in Model \ref{ex:generalqsc} (Part II), and we use the same parameters explained in Part I of this section. In all cases for the different values of $\alpha$ considered, the results show that there is no false positive in this case indicating the desired causal inference. \begin{table}[!htbp] \caption{Validation of Direct Graph in Model \ref{ex:generalqsc} (Part II) for $\beta$ uniformly spaced in the intervals (0.2,0.3), (0.6,0.7), (0.7,0.8), and (0.8,0.9).} \label{t:4} \centering \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.7&\textcolor{blue}{T} & \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ &0.8 & \textcolor{blue}{T}& \textcolor{blue}{T} &\textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ $\alpha$&0.9&\textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ &1&\textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ \cline{2-13} \end{tabular} \end{table} \section{Review of Classical Causal Inference Framework in \citep{KocaogluNEURIPS2020}}\label{sec:classic} In this section, we briefly review the proposed approach in \citep{KocaogluNEURIPS2020} for confounder discovery via solving an optimization problem that its aim is to discover the trade-off between the entropy of the latent variable and the conditional mutual information of the observed variables. Consider that the joint distribution $P(X,Y)$ between two observed variables is given. The goal is to find a random variable $Z$ that makes $X$ and $Y$ conditionally independent given $Z$. Possible cases that can represent this situation is shown in Figure \ref{fig:possibilities}. \begin{figure}[ht] \centering \[\begin{tikzpicture}[transform shape] \tikzset{vertex/.style = {shape=circle,draw,minimum size=1.5em}} \tikzset{edge/.style = {->,> = latex'}} \node[vertex,thick] (a) at (2,0) {$Y$}; \node[state,rectangle] (b) at (1,1) {$Z$}; \node[vertex,thick] (c) at (0,0) {$X$}; \node (d) at (1,-1) {$(a)$}; \draw[edge,thick] (b) to (a); \draw[edge,thick] (b) to (c); \node[vertex,thick] (e) at (5,0) {$Y$}; \node[vertex,thick] (f) at (4,1) {$Z$}; \node[vertex,thick] (g) at (3,0) {$X$}; \node (h) at (4,-1) {$(b)$}; \draw[edge,thick] (f) to (e); \draw[edge,thick] (f) to (g); \draw[edge,thick] (g) to (e); \node[vertex,thick] (h) at (8,0) {$X$}; \node[vertex,thick] (i) at (6,0) {$Y$}; \node[vertex,thick] (j) at (8,1) {$Y$}; \node[vertex,thick] (k) at (6,1) {$X$}; \node (l) at (7,-1) {$(c)$}; \node (m) at (7,.5) {$or$}; \draw[edge,thick] (k) to (j); \draw[edge,thick] (i) to (h); \node[vertex,thick] (n) at (9,0) {$X$}; \node[state,rectangle] (o) at (10.5,0) {$M$}; \node[vertex,thick] (p) at (12,0) {$Y$}; \node (q) at (10.5,-1) {$(d)$}; \draw[edge,thick] (n) to (o); \draw[edge,thick] (o) to (p); \end{tikzpicture}\] \caption{(a) Latent Graph, (b) Triangle Graph, (c) Direct Graph, and (d) Mediator Graph.} \label{fig:possibilities} \end{figure} \begin{algorithm}[!ht] \caption{\textbf{LatentSearch} \citep{KocaogluNEURIPS2020}}\label{alg:LatentSearch} \SetAlgoLined \scriptsize\KwIn{Supports of $X,Y$, and $Z$, respectively; Joint probability distribution $p(x,y)$; Number of iterations $N$; $\beta$ in the loss function $L=I(X; Y|Z)+\beta H(Z)$, Initialization of $q_1(z|x,y)$.} \KwOut{Joint distribution $q(x, y, z)$.} \For{$i = 1:N$}{ \tcc{Form the joint distribution:} $q_i(x,y,z)\gets q_i(z|x,y)p(x,y), \forall x,y,z$\; \textbf{Calculate}: $$q_i(z|x)\gets \frac{\sum_{y\in Y} q_i(x,y,z)}{\sum_{y\in Y,z\in Z} q_i(x,y,z)}, q_i(z|y)\gets \frac{\sum_{x\in X} q_i(x,y,z)}{\sum_{x\in X,z\in Z} q_i(x,y,z)},q_i(z)\gets \sum_{x\in X,y\in Y} q_i(x,y,z)$$ \textbf{Update}: $$q_{i+1}(z|x,y)\gets \frac{1}{F(x,y)}\frac{q_i(z|x)q_i(z|y)}{q_i(z)^{1-\beta}}, \textrm{ where } F(x,y)=\sum_{z\in Z}\frac{q_i(z|x)q_i(z|y)}{q_i(z)^{1-\beta}}$$ } \textbf{return} $q(x,y,z):= q_{N+1}(z|x,y)p(x,y)$. \end{algorithm} In the classical causal inference, \cite{KocaogluNEURIPS2020} distinguished between \textit{latent graph} in Figure \ref{fig:possibilities}(a) from others in Figure \ref{fig:possibilities} based on unmeasured confounder having low Shannon entropy under certain assumptions. Formally, the following was assumed: \begin{assumption}\label{mainassump1} Consider any causal model with observed variables $X$ and $Y$. Let $Z$ represents the variable that captures all latent confounders between $X$ and $Y$. Then $H(Z)<\theta$, where $H(Z) = -\sum_{i=1}^n P(x_i)\log(x_i)$. \end{assumption} \begin{assumption}\label{mainassump2} Consider a causal model where $X$ causes $Y$. If $X$ causes $Y$ only through a latent mediator $Z$, i.e., $X\to Z\to Y$, then $H(Z)\ge\theta$. \end{assumption} Note that $I(X;Y|Z)=0$ means that $Z$ makes the variables $X$ and $Y$ conditionally independent, i.e., $X\perp\!\!\!\perp Y|Z$.\footnote{Note that this is different from the notion of \textit{causal independence}, which refers to the situation where multiple causes contribute independently to a common effect \citep{zhang1996exploiting}.} To identify latent graphs, \cite{KocaogluNEURIPS2020} proposed an iterative algorithm (Algorithm \ref{alg:LatentSearch}) that discovers the trade-off between the entropy of the unmeasured confounder and the conditional mutual information of the observed variables. This trade-off is formally defined as follows: \begin{equation}\label{eq:lossfunc} L = I(X;Y|Z)+\beta H(Z) \end{equation} In fact, \textbf{LatentSearch} (Algorithm \ref{alg:LatentSearch}) sets $q(x,y,z) =q(z|x,y)p(x,y)$ and searches over $q(z|x,y)$ to find the stationary point of the loss function $L$ in Equation (\ref{eq:lossfunc}). For this purpose, \textbf{LatentSearch} returns a joint probability distribution $q(X,Y,Z)$ from which the Shannon entropy of the latent variable $W$, i.e., $H(W)$ can be computed. To verify whether the causal graph $G=(V=\{X,Y\},E)$ is a latent graph or not, \textbf{InferGraph} (Algorithm \ref{alg:InferGraph}) \citep{KocaogluNEURIPS2020} runs \textbf{LatentSearch} multiple times and selects the smallest $H(W)$ discovered by the algorithm among those that ensure the conditional independence of $X$ and $Y$ given $W$, i.e., $I(X;Y|W)\le\theta$ for a practical threshold ( as suggested in \citep{KocaogluNEURIPS2020}, $\theta=0.001$). We refer readers to \citep{KocaogluNEURIPS2020} for more experimental settings. \cite{KocaogluNEURIPS2020} conjecture that, under assumptions \ref{mainassump1}, \ref{mainassump2} and in practice, the Shannon entropy of observed variables $X$ and $Y$ for directed graphs and triangle graphs is lower-bounded by the entropies of $X$ and $Y$, up to a scaling by a constant (as suggested in \citep{KocaogluNEURIPS2020}, $0.8\min\{H(X),H(Y)\}$). For more detailed discussion see \citep{KocaogluNEURIPS2020}. \begin{algorithm}[!ht] \caption{\textbf{InferGraph}: Identifying the Latent Graph \citep{KocaogluNEURIPS2020}}\label{alg:InferGraph} \SetAlgoLined \scriptsize\KwIn{Joint probability distribution $p(x,y)$; Number of iterations $N$; $I(X; Y|Z)$ threshold $T$; $H(Z)$ threshold that is determined by $\theta=\alpha\min(H(X),H(Y))$; $\{\beta_i\}_{i=1}^N$; Support size of $X,Y$, and $Z$, i.e., $r,m$, and $n$, respectively.} \KwOut{"Latent Graph" if $Z$ is an unmeasured confounder for $X$ and $Y$, otherwise, returns "Triangle or Direct Graph".} \For{$i = 1:N$}{ $q_i(x,y,z)\gets \textrm{\textbf{LatentSearch}}(p(x,y), \alpha,\beta_i,r,m,n)$\; Calculate $I_i(X; Y|Z)$ and $H_i(Z)$ from $q_i(x,y,z)$\; } $S=\{i: I_i(X; Y|Z)\le T\}$\; \uIf{$\min(H_i(Z):i\in S)>\theta$ or $S=\text{\O}$ }{ \textbf{return} Triangle or Direct Graph\; }\Else{ \textbf{return} Latent Graph\; } \end{algorithm} \section{Evaluation on Quantum Causal Synthetic Data}\label{sec:evaluation} Since there is no quantum cause-effect repository to verify the validity of our proposed algorithm, we put forward an experimental scheme that can be used to confront our theoretical framework. To show the effectiveness of the proposed approach in section \ref{sec:method}, we use quantum noisy links, where it is validated that the input before noise, as a latent confounder (hidden source), is the cause of the noisy outputs, as shown in Figure \ref{fig:AliceBob}. \begin{figure}[!ht] \centering \includegraphics[width=.5\linewidth]{AliceBob.pdf} \caption{Alice and Bob are connected by a noisy channel (e.g., an optical fiber) with an unknown source of message.} \label{fig:AliceBob} \end{figure} We first apply the proposed approach to a classical setup, as explained in Model \ref{model:classicchannel}, where two bits are transmitted over a binary symmetric channel (to illustrate the case of no confounder), or two bits are transmitted over two separate channels (to illustrate the case of latent confounder). We show that the proposed approach outperforms the classical causal inference in \citep{KocaogluNEURIPS2020} due to the use of quantum density matrix. Finding the optima over a quantum density matrix rather than over a probability distribution provides larger degrees of freedom thus resulting in improved results. Our results indicate that the proposed approach can also be used for classical causal inference with improved results. \begin{model}[Classical Binary Symmetric Channel: Latent and Direct Graph]\label{model:classicchannel} \emph{\textbf{Part I: Latent Graph.}} Assume a 2-bit input $Z\in \{00,01,10,11\}$. Let each bit of $Z$ be in the state 1 with probability $q$ and $1-q$ otherwise, and independent of each other. So, $p(Z=00)=(1-q)^2$, $p(Z=01)=p(Z=10)=q(1-q)$, and $p(Z=11)=q^2$. $Z$ is transmitted over a binary symmetric channel with independent bit error probability of $p_1$, and is denoted $X$. A cloned version of $Z$ is transmitted over a binary symmetric channel with independent bit error probability of $p_2$, and is denoted $Y$. The joint probability distribution of $X, Y$, and $Z$, where $Z$ is the cause of $X$ and $Y$, i.e., $X\gets Z\to Y$ can be computed as $p(X,Y,Z)=p(Z)p(X|Z)p(y|Z)$. For example, $p(01,10,00)= (1-q)q*p_1p_2*(1-p_1)p_2$. Then we marginalize out $Z$ to obtain the joint probability distribution for the latent graph $X\leftrightarrow Y$. Note that the corresponding joint density matrix $\rho_{XY}$ is a diagonal matrix that its diagonal entries come from the joint probability distribution $p(X,Y)$. The key reason of constructing $\rho_{XY}$ as the diagonal matrix from $p(X,Y)$ is to have the mixed states, so that the von-Neuman entropy of $\rho_{XY}$ is the same as the Shannon entropy of $p(X,Y)$. \begin{figure} \centering \includegraphics[width=.25\linewidth]{NBSC.pdf} \caption{2-bit non-Binary symmetric channel.} \label{fig:NBSC} \end{figure} Now, we apply {\sc \textbf{\texttt{QInferGraph}}}~(Algorithm \ref{alg:qInferGraph}) on $\rho_{XY}$ to verify that $X$ and $Y$ are confounded by $Z$. For this purpose, we use {\sc \textbf{\texttt{QLatentSearch}}}~(Algorithm \ref{alg:qLatentSearch}) on 100 different values of $\beta$, uniformly spaced in the interval $(0.7, 0.8)$. A discussion regarding the best choice of $\beta$ can be found in appendices \ref{eval:moreGQSC} and \ref{eval:moreDepolar}. We run {\sc \textbf{\texttt{QLatentSearch}}}~for 500 iterations each time. We use the conditional mutual information threshold of 0.05. In other words, of the algorithm outputs for the 100 $\beta$ values used, we pick the smallest entropy $W$ discovered by the algorithm among those that ensure $I(X; Y |W) \le 0.05$. Table \ref{t:LatentQSC} summarizes the results for different $S(W)$ threshold that is determined by $\theta=\alpha\min\{S(X),S(Y)\}$. For different values of $\alpha=0.7,0.8,0.9,1$, the results are the same, and are given in Table \ref{t:LQNBSC}. We let $q=0.4$. In the Table, $\textcolor{blue}{T}$ means that {\sc \textbf{\texttt{QInferGraph}}}~(Algorithm \ref{alg:qInferGraph}) identifies the latent graph correctly. But, $\textcolor{red}{F}$ means that the algorithm fails to identify the latent graph. For very small or very large $p_i$'s the case of latent confounder or direct graph are hard to separate, while the proposed algorithm works well in most other cases. \begin{table} \caption{Validation of Latent Graph in Model \ref{model:classicchannel} (Part I) for $\alpha=0.7,0.8,0.9,1$, and $\beta\in (0.7,0.8)$ via {\sc \textbf{\texttt{QInferGraph}}}. }\label{t:LQNBSC} \centering \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p_2$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.01& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.1& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.2& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.3& \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ &0.4& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ $p_1$&0.5& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.6& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.7& \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F}\\ &0.8& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.9& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.99& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ \cline{2-13} \end{tabular} \end{table} Now, if we apply \textbf{InferGraph} (Algorithm \ref{alg:InferGraph}) on $p(X,Y)$ with $\alpha=0.8$, as suggested in \citep{KocaogluNEURIPS2020}, and two different $\beta$ parameters: (1) the same as one that we used in Table \ref{t:LQNBSC} in {\sc \textbf{\texttt{QInferGraph}}}, i.e., $\beta\in (0.7,0.8)$, and (2) the same as one that suggested in \citep{KocaogluNEURIPS2020}, i.e., $\beta\in (0,0.1)$, we obtain the results summarized in Table \ref{t:LNBSCclassic}. \begin{table}[!htb] \caption{Validation of Latent Graph in Model \ref{model:classicchannel} (Part I) via classical causal inference (Algorithm \ref{alg:InferGraph}), and $\alpha=0.8$.}\label{t:LNBSCclassic} \begin{subtable}{.5\linewidth} \caption{$\beta\in (0,0.1)$} \centering \resizebox{\columnwidth}{!}{% \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p_2$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.01& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.1& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.2& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.3& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.4& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ $p_1$&0.5& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.6& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.7& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.8& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.9& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.99& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ \cline{2-13} \end{tabular}% } \end{subtable}% \begin{subtable}{.5\linewidth} \centering \caption{$\beta\in (0.7,0.8)$} \resizebox{\columnwidth}{!}{% \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p_2$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.01& \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T}\\ &0.1& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.2& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.3& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.4& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ $p_1$&0.5& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.6& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.7& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.8& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.9& \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F}\\ &0.99& \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{red}{F} & \textcolor{blue}{T}\\ \cline{2-13} \end{tabular}% } \end{subtable} \end{table} \noindent\emph{\textbf{Some highlights for results in Part I:}} (1) Note that where the probability of errors i.e., $p_1$ and $p_2$ are very small, the latent confounder $Z$ is hardly distinguishable from $X$ (or $Y$) and {\sc \textbf{\texttt{QInferGraph}}}~fails to discover the latent graph. (2) Note that {\sc \textbf{\texttt{QLatentSearch}}}~tries to find the stationary point(s) of the loss function $L$ in Equation (\ref{eq:qlossfunc}), and there is no guarantee to find the global optimum. However, the performance of {\sc \textbf{\texttt{QInferGraph}}}~in this case is acceptable: true positive rate (recall) = 0.6, false positive rate (fall-out) = 0, false negative rate (miss rate) = 0.4, accuracy = 0.6. (3) The hyperparameter $\alpha$ does not affect significantly on the quality of results in our experimental settings that indicates {\sc \textbf{\texttt{QInferGraph}}}~is not very sensitive to hyperparameters. (4) Although \textbf{InferGraph} (Algorithm \ref{alg:InferGraph}) perfectly identifies latent graphs in Model \ref{model:classicchannel} (Part I), it fails to identify latent graphs in many cases (about 80\%). (5) It seems that the classical causal inference algorithm, i.e., \textbf{InferGraph} (Algorithm \ref{alg:InferGraph}) outperforms {\sc \textbf{\texttt{QInferGraph}}}~in classical data. However, as we will see in Part II of Model \ref{model:classicchannel}, the performance of the classical algorithm is not consistent (no longer outperform {\sc \textbf{\texttt{QInferGraph}}}), where there is no latent confounder. \noindent\emph{\textbf{Part II: Direct Graph.}} Assume that there is a 2-bit symmetric noisy channel, where there is no latent common cause, i.e., there is an input $X$ and an output $Y$, as shown in Figure \ref{fig:NBSC}, with the same properties explained in Part I. The results of applying {\sc \textbf{\texttt{QInferGraph}}}~and \textbf{InferGraph} on $p(X,Y)$ and $\rho_{XY}$ are summarized in Table \ref{t:DQNBSC} and \ref{t:DNBSCclassic}, respectively. $\textcolor{blue}{T}$ means that {\sc \textbf{\texttt{QInferGraph}}}~(Algorithm \ref{alg:qInferGraph}) identifies the direct graph correctly. But, $\textcolor{red}{F}$ means that the algorithm fails to identify the direct graph. \begin{table}[!htbp] \caption{Validation of Direct Graph in Model \ref{model:classicchannel} (Part II) via {\sc \textbf{\texttt{QInferGraph}}}.} \label{t:DQNBSC} \centering \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.7&\textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ &0.8 &\textcolor{blue}{T}& \textcolor{blue}{T} &\textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ $\alpha$&0.9&\textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ &1&\textcolor{blue}{T}& \textcolor{blue}{T} &\textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ \cline{2-13} \end{tabular} \end{table} \begin{table}[!htb] \caption{Validation of Latent Graph in Model \ref{model:classicchannel} (Part II) via classical causal inference (Algorithm \ref{alg:InferGraph}).}\label{t:DNBSCclassic} \begin{subtable}{.5\linewidth} \caption{$\beta\in (0,0.1)$} \centering \resizebox{\columnwidth}{!}{% \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.7&\textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ &0.8 &\textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}\\ $\alpha$&0.9&\textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}\\ &1&\textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}\\ \cline{2-13} \end{tabular}% } \end{subtable}% \begin{subtable}{.5\linewidth} \centering \caption{$\beta\in (0.7,0.8)$} \resizebox{\columnwidth}{!}{% \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.7&\textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{red}{F}\\ &0.8 &\textcolor{red}{F}& \textcolor{blue}{T} &\textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{red}{F}\\ $\alpha$&0.9&\textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{red}{F}\\ &1&\textcolor{red}{F}& \textcolor{blue}{T} &\textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{red}{F}\\ \cline{2-13} \end{tabular}% } \end{subtable} \end{table} \noindent\emph{\textbf{Some highlights for results in Part II:}} (1) Note that when $p=0.5$, $X$ and $Y$ are uncorrelated and then $X$ and $Y$ are not cause and effect. In this case, \textcolor{blue}{T} means that both quantum and classical algorithms identify that $X$ and $Y$ are not cause and effect, as we expected. (2) Although the performance of the classical algorithm (Algorithm \ref{alg:InferGraph}), where $\beta\in (0,0.1)$, is perfect for Model \ref{model:classicchannel} (Part I), its performance for Model \ref{model:classicchannel} (Part II), where there is no confounder, is not acceptable. In conclusion, results from Part I and II, indicate that {\sc \textbf{\texttt{QInferGraph}}}~is a more consistent and less sensitive to the change of parameters than its counterpart in the classical causal inference, even for the classical data. \end{model} Next, we apply our proposed approach on a quantum (non-classical) model, as explained in Model \ref{model:depolar}, where mixed entangled quantum subsystems are used for which subsystems are communicated over noisy channels (e.g., optical fiber) to create a coexisting set of quantum systems. \begin{model}[Depolarizing Quantum Channel: Latent Graph and Direct Graph]\label{model:depolar}~\\ \emph{\textbf{Part I: Latent Graph.}} Assume that there are real numbers $\gamma_1$, $\gamma_2$, $\lambda_1$, and $\lambda_2$ such that $\gamma_1^2+\lambda_1^2=1$ and $\gamma_2^2+\lambda_2^2=1$. We consider a joint entangled system (of three qubits) as the mixture of the following pure density matrices: \noindent\resizebox{\linewidth}{!}{% $\displaystyle \left\{ \begin{array}{lll} [(\gamma_1 |0\rangle + \lambda_1|1\rangle)(\gamma_1 |0\rangle + \lambda_1|1\rangle)(\gamma_1 |0\rangle + \lambda_1|1\rangle)] [(\gamma_1 |0\rangle + \lambda_1|1\rangle)(\gamma_1 |0\rangle + \lambda_1|1\rangle)(\gamma_1 |0\rangle + \lambda_1|1\rangle)]^\dagger & & q\\ [(\gamma_2 |0\rangle + \lambda_2|1\rangle)(\gamma_2 |0\rangle + \lambda_2|1\rangle)(\gamma_2 |0\rangle + \lambda_2|1\rangle)] [(\gamma_2 |0\rangle + \lambda_2|1\rangle)(\gamma_2 |0\rangle + \lambda_2|1\rangle)(\gamma_2 |0\rangle + \lambda_2|1\rangle)]^\dagger & & 1-q \end{array} \right. $} \vspace{\baselineskip} In other words, the system considered has density matrix $q[(\gamma_1 |0\rangle + \lambda_1|1\rangle)(\gamma_1 |0\rangle + \lambda_1|1\rangle)(\gamma_1 |0\rangle + \lambda_1|1\rangle)] [(\gamma_1 |0\rangle + \lambda_1|1\rangle)(\gamma_1 |0\rangle + \lambda_1|1\rangle)(\gamma_1 |0\rangle + \lambda_1|1\rangle)]^\dagger + (1-q) [(\gamma_2 |0\rangle + \lambda_2|1\rangle)(\gamma_2 |0\rangle + \lambda_2|1\rangle)(\gamma_2 |0\rangle + \lambda_2|1\rangle)] [(\gamma_2 |0\rangle + \lambda_2|1\rangle)(\gamma_2 |0\rangle + \lambda_2|1\rangle)(\gamma_2 |0\rangle + \lambda_2|1\rangle)]^\dagger $. The system is a mixture of two pure density matrices. This quantum system has entanglement among the three quantum bits. Let the second quantum bit is transmitted over a \emph{quantum depolarizing channel} with error probability $p_1$, and the third quantum bit is transmitted over a \emph{quantum depolarizing channel} with error probability $p_2$. Note that the depolarizing channel with error probability $p$ has no error with probability $1-p$, and each of the phase-flip, bit-flip, or the combination of phase-flip and bit-flip errors with probability $p/3$ \citep{nielsen2002quantum}. With this setup, the joint density matrix is given as $\rho_{ZXY} = q \rho_{ZXY}^{\gamma_1,\lambda_1} + (1-q) \rho_{ZXY}^{\gamma_2,\lambda_2}$, where $\rho_{ZXY}^{\gamma,\lambda}$ is given as the mixture of the following pure density matrices: \vspace{\baselineskip} \noindent\resizebox{\linewidth}{!}{% $\displaystyle \left\{ \begin{array}{lll} [(\gamma |0\rangle + \lambda|1\rangle)(\gamma |0\rangle + \lambda|1\rangle)(\gamma |0\rangle + \lambda|1\rangle)] [(\gamma |0\rangle + \lambda|1\rangle)(\gamma |0\rangle + \lambda|1\rangle)(\gamma |0\rangle + \lambda|1\rangle)]^\dagger & & (1-p_1)(1-p_2)\\ [(\gamma |0\rangle + \lambda|1\rangle)(\gamma |0\rangle + \lambda|1\rangle)(\gamma |0\rangle - \lambda|1\rangle)] [(\gamma |0\rangle + \lambda|1\rangle)(\gamma |0\rangle + \lambda|1\rangle)(\gamma |0\rangle - \lambda|1\rangle)]^\dagger & & (1-p_1)(p_2/3) \\ [(\gamma |0\rangle + \lambda|1\rangle)(\gamma |0\rangle + \lambda|1\rangle)(\lambda |0\rangle +\gamma |1\rangle)] [(\gamma |0\rangle + \lambda|1\rangle)(\gamma |0\rangle + \lambda|1\rangle)(\lambda |0\rangle +\gamma |1\rangle)]^\dagger & & (1-p_1)(p_2/3) \\ [(\gamma |0\rangle + \lambda|1\rangle)(\gamma |0\rangle + \lambda|1\rangle)(-\lambda |0\rangle +\gamma |1\rangle)] [(\gamma |0\rangle + \lambda|1\rangle)(\gamma |0\rangle + \lambda|1\rangle)(-\lambda |0\rangle +\gamma |1\rangle)]^\dagger & & (1-p_1)(p_2/3) \\ [(\gamma |0\rangle + \lambda|1\rangle)(-\lambda |0\rangle +\gamma |1\rangle)(\gamma |0\rangle + \lambda|1\rangle)] [(\gamma |0\rangle + \lambda|1\rangle)(-\lambda |0\rangle +\gamma |1\rangle)(\gamma |0\rangle + \lambda|1\rangle)]^\dagger & & (p_1/3)(1-p_2)\\ [(\gamma |0\rangle + \lambda|1\rangle)(-\lambda |0\rangle +\gamma |1\rangle)(\lambda |0\rangle +\gamma |1\rangle)] [(\gamma |0\rangle + \lambda|1\rangle)(-\lambda |0\rangle +\gamma |1\rangle)(\lambda |0\rangle +\gamma |1\rangle)]^\dagger & & (p_1/3)(p_2/3) \\ [(\gamma |0\rangle + \lambda|1\rangle)(-\lambda |0\rangle +\gamma |1\rangle)(-\lambda |0\rangle +\gamma |1\rangle)] [(\gamma |0\rangle + \lambda|1\rangle)(-\lambda |0\rangle +\gamma |1\rangle)(-\lambda |0\rangle +\gamma |1\rangle)]^\dagger & & (p_1/3)(p_2/3) \\ [(\gamma |0\rangle + \lambda|1\rangle)(-\lambda |0\rangle +\gamma |1\rangle)(\gamma |0\rangle - \lambda|1\rangle)] [(\gamma |0\rangle + \lambda|1\rangle)(-\lambda |0\rangle +\gamma |1\rangle)(\gamma |0\rangle - \lambda|1\rangle)]^\dagger & & (p_1/3)(p_2/3) \\ [(\gamma |0\rangle + \lambda|1\rangle)(\lambda |0\rangle +\gamma |1\rangle)(\gamma |0\rangle + \lambda|1\rangle)] [(\gamma |0\rangle + \lambda|1\rangle)(\lambda |0\rangle +\gamma |1\rangle)(\gamma |0\rangle + \lambda|1\rangle)]^\dagger & & (p_1/3)(1-p_2)\\ [(\gamma |0\rangle + \lambda|1\rangle)(\lambda |0\rangle +\gamma |1\rangle)(\lambda |0\rangle +\gamma |1\rangle)] [(\gamma |0\rangle + \lambda|1\rangle)(\lambda |0\rangle +\gamma |1\rangle)(\lambda |0\rangle +\gamma |1\rangle)]^\dagger & & (p_1/3)(p_2/3) \\ [(\gamma |0\rangle + \lambda|1\rangle)(\lambda |0\rangle +\gamma |1\rangle)(-\lambda |0\rangle +\gamma |1\rangle)] [(\gamma |0\rangle + \lambda|1\rangle)(\lambda |0\rangle +\gamma |1\rangle)(-\lambda |0\rangle +\gamma |1\rangle)]^\dagger & & (p_1/3)(p_2/3) \\ [(\gamma |0\rangle + \lambda|1\rangle)(\lambda |0\rangle +\gamma |1\rangle)(\gamma |0\rangle - \lambda|1\rangle)] [(\gamma |0\rangle + \lambda|1\rangle)(\lambda |0\rangle +\gamma |1\rangle)(\gamma |0\rangle - \lambda|1\rangle)]^\dagger & & (p_1/3)(p_2/3) \\ [(\gamma |0\rangle + \lambda|1\rangle)(\gamma |0\rangle - \lambda|1\rangle)(\gamma |0\rangle + \lambda|1\rangle)] [(\gamma |0\rangle + \lambda|1\rangle)(\gamma |0\rangle - \lambda|1\rangle)(\gamma |0\rangle + \lambda|1\rangle)]^\dagger & & (p_1/3)(1-p_2)\\ [(\gamma |0\rangle + \lambda|1\rangle)(\gamma |0\rangle - \lambda|1\rangle)(\lambda |0\rangle +\gamma |1\rangle)] [(\gamma |0\rangle + \lambda|1\rangle)(\gamma |0\rangle - \lambda|1\rangle)(\lambda |0\rangle +\gamma |1\rangle)]^\dagger & & (p_1/3)(p_2/3) \\ [(\gamma |0\rangle + \lambda|1\rangle)(\gamma |0\rangle - \lambda|1\rangle)(-\lambda |0\rangle +\gamma |1\rangle)] [(\gamma |0\rangle + \lambda|1\rangle)(\gamma |0\rangle - \lambda|1\rangle)(-\lambda |0\rangle +\gamma |1\rangle)]^\dagger & & (p_1/3)(p_2/3) \\ [(\gamma |0\rangle + \lambda|1\rangle)(\gamma |0\rangle - \lambda|1\rangle)(\gamma |0\rangle - \lambda|1\rangle)] [(\gamma |0\rangle + \lambda|1\rangle)(\gamma |0\rangle - \lambda|1\rangle)(\gamma |0\rangle - \lambda|1\rangle)]^\dagger & & (p_1/3)(p_2/3) \end{array} \right. $} \vspace{\baselineskip} We note that $X$ and $Y$ coexist, thus we can find joint density matrix of $X$ and $Y$ by tracing out $Z$ in $\rho_{ZXY}$. Then, we apply {\sc \textbf{\texttt{QInferGraph}}}~(Algorithm \ref{alg:qInferGraph}) on $\rho_{XY}$ to verify that $X$ and $Y$ are confounded by a latent variable. For this purpose, we use the same parameters specification as explained in Model \ref{ex:generalqsc}, and $q=0.4$. Table \ref{t:Ldepolar} summarizes the results for different entropy threshold of the latent confounder that is determined by $\theta=\alpha\min\{S(X),S(Y)\}$, where $\alpha=0.7,0.8,0.9,1$. $\textcolor{blue}{T}$ means that {\sc \textbf{\texttt{QInferGraph}}}~(Algorithm \ref{alg:qInferGraph}) identifies the latent graph correctly. But, $\textcolor{red}{F}$ means that the algorithm fails to identify the latent graph. The results confirm our observations that we made in Model \ref{model:classicchannel} (Part I). However, in this case {\sc \textbf{\texttt{QInferGraph}}}~has a higher performance quality. For example, for $\alpha=0.7,0.8,0.9,1$ we have: true positive rate (recall) = 1, false positive rate (fall-out) = 0, false negative rate (miss rate) = 0, accuracy = 1. \begin{table}[!htb] \caption{Validation of Latent Graph in Model \ref{model:depolar} (Part I) for $\alpha=0.7,0.8,0.9,1$, and $\beta\in (0.7,0.8)$ via {\sc \textbf{\texttt{QInferGraph}}}, and with the density matrix obtained from $0.6\rho_{ZXY}^{1/\sqrt{2},1/\sqrt{2}}+0.4\rho_{ZXY}^{0.6,0.8}$ via tracing out $Z$.}\label{t:Ldepolar} \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p_2$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.01& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.1& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.2& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.3& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.4& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ $p_1$&0.5& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.6& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.7& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.8& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.9& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ &0.99& \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T} & \textcolor{blue}{T}\\ \cline{2-13} \end{tabular} \end{table} \emph{\textbf{Part II: Direct Graph.}} Assume that there are real numbers $\gamma_1$, $\gamma_2$, $\lambda_1$, and $\lambda_2$ such that $\gamma_1^2+\lambda_1^2=1$ and $\gamma_2^2+\lambda_2^2=1$. We consider a joint entangled system (of two qubits) as the mixture of the following pure density matrices: $$\displaystyle \left\{ \begin{array}{lll} \gamma_1^2|00\rangle+\gamma_1\lambda_1|01\rangle+\gamma_1\lambda_1|10\rangle+\lambda_1^2|11\rangle & & q\\ \gamma_2^2|00\rangle+\gamma_2\lambda_2|01\rangle+\gamma_2\lambda_2|10\rangle+\lambda_2^2|11\rangle & & 1-q \end{array} \right. $ The system is a mixture of two pure density matrices. This quantum system has entanglement among the two quantum bits. Let the second quantum bit is transmitted over a \emph{quantum depolarizing channel} with error probability $p$. With this setup, the joint density matrix is given as $\rho_{XY} = q \rho_{XY}^{\gamma_1,\lambda_1} + (1-q) \rho_{XY}^{\gamma_2,\lambda_2}$, where $\rho_{XY}^{\gamma,\lambda}$ is given as the mixture of the following pure density matrices: \begin{equation*} \left\{ \begin{array}{lll} \gamma^2|00\rangle+\gamma\lambda|01\rangle+\gamma\lambda|10\rangle+\lambda^2|11\rangle & & 1-p\\ \gamma^2|00\rangle-\gamma\lambda|01\rangle+\gamma\lambda|10\rangle-\lambda^2|11\rangle & & p/3\\ \gamma\lambda|00\rangle+\gamma^2|01\rangle+\lambda^2|10\rangle+\gamma\lambda|11\rangle & & p/3\\ -\gamma\lambda|00\rangle+\gamma^2|01\rangle-\lambda^2|10\rangle+\gamma\lambda|11\rangle & & p/3\\ \end{array} \right. \end{equation*} We note that $X$ and $Y$ coexist in the quantum system, and thus the joint density matrix has been obtained. We already know that $X$ is the cause of $Y$ in this scenario, i.e., $X\to Y$ is the corresponding directed graph. To verify this, we use Algorithm \ref{alg:qLatentSearch} and \ref{alg:qInferGraph} as we explained earlier in this model. The results are summarized in Table \ref{t:DGdepolar}. $\textcolor{blue}{T}$ means that {\sc \textbf{\texttt{QInferGraph}}}~(Algorithm \ref{alg:qInferGraph}) identifies the direct graph correctly. But, $\textcolor{red}{F}$ means that the algorithm fails to identify the direct graph. In all cases the probability of $X$ be in state $X_1$ is $q=0.4$. Results show that the best performance belongs to $\alpha=0.9$ with only one false positive case. In general, in quantum noisy channels with very small probability of errors, {\sc \textbf{\texttt{QInferGraph}}}~very likely fails to draw the right conclusion about the identification of latent confounders. \begin{table}[!htbp] \caption{Validation of Direct Graph in Model \ref{model:depolar} (Part II) with joint density matrix $\rho_{XY}=0.4*\rho_{XY}^{0.6,0.8}+0.6*\rho_{XY}^{1/\sqrt{2},1/\sqrt{2}}$. } \label{t:DGdepolar} \centering \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.7&\textcolor{red}{F}& \textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ &0.8 & \textcolor{red}{F}& \textcolor{red}{F} &\textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ $\alpha$&0.9&\textcolor{red}{F}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ &1&\textcolor{red}{F}& \textcolor{red}{F} &\textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}& \textcolor{blue}{T}\\ \cline{2-13} \end{tabular} \end{table} \end{model} \section{Introduction}\label{intro} Causal inference lies at the heart of science \citep{Pearl09,pearl2018book}: the conclusions drawn from scientific studies almost always involve extracting causation (cause and effect relationships) from association, even if researchers often refrain from explicitly acknowledging the causal goal of research projects \citep{hernan2018c,hernan2019second}. However, causal inference from observational data is an ambitious and difficult task. Identifying cause and effect relationships from observational data is even more challenging in the presence of \textit{hidden common causes} (\textit{latent confounders}) \citep{heckerman2019toward}. The broad impact of this phenomena has been studied in multiple domains of science such as epidemiologic studies \citep{lipsitch2010negative}, biology and medicine \citep{skelly2012assessing,meinshausen2016methods}, experiential education \citep{ewert2009creating,Kallus2018Removing}, economics and marketing \citep{varian2016causal,hunermund2019causal}, among others. A similar concept is increasingly appreciated among quantum physicists, namely the inference of \textit{quantum common causes} \citep{wolfe2020quantifying,allen2017quantum,ried2015quantum,Chaves2014UAI,chaves2014causal,chaves2015information,hofer1999reichenbach}. It has been used to provide a satisfactory causal explanation (i.e., non-fine-tuned) of Bell inequality violations \citep{allen2017quantum,hofer1999reichenbach}. This also has led to a formalization of quantum causal models \citep{costa2016quantum,barrett2019quantum,chiribella2019quantum,shrapnel2019discovering}. As shown in \citep{Chaves2014UAI,chaves2014causal,chaves2015information}, \textit{in some cases}, (hidden) common causes can be distinguished from direct causation using information theoretical generalization of Bell’s inequalities and causal directed acyclic graphs (DAGs). Also, as shown in \citep{fitzsimons2015quantum,ried2015quantum}, observed quantum correlations alone are \textit{sometimes} enough to imply causation. However, the proposed approach in \citep{fitzsimons2015quantum,ried2015quantum} depends on the precise knowledge of the physical system and the measurement apparatuses \citep{gachechiladze2020quantifying}. In this paper, we propose the first tractable algorithmic approach to distinguish between a hidden common cause and direct causal influences among two observed quantum systems without any interventional data. To show the difficulty of causal structure discovery task even in the simplest classical case, where our observation consists of only two jointly-distributed random variables $X$ and $Y$ that are statically correlated, we recall Reichenbach's common cause principle \citep{reichenbach1991direction}: If two random variables $X$ and $Y$ are statistically dependent, then there exists a third variable $Z$ that causally affects both. As a special case, $Z$ may coincide with either $X$ or $Y$. Furthermore, this variable $Z$ makes $X$ and $Y$ conditionally independent, i.e., $X\perp\!\!\!\perp Y|Z$. So, possible candidates for representing causal relationships between $X$ and $Y$ are: $X\to Y$, $X\gets Y$, and $X\gets Z\to Y$, and there is no easy way to determine which one is the right structure based on the observational data alone. The variable $Z$ in the case $X\gets Z\to Y$ is called \textit{unmeasured (latent) confounder} or \textit{unmeasured (latent) common cause}. So, one of the fundamental questions in causality is to determine how cause-effect relationships can be inferred from statistical information, encoded as a joint probability distribution, obtained under normal, intervention-free experiments. To discover the true cause-effect relationships, scientists normally perform randomized experiments where a sample of units drawn from the population of interest is subjected to the specified manipulation directly. In many cases, however, such a direct approach is not possible due to expense or ethical considerations. Instead, investigators have to rely on observational studies to infer causality. This task is even more challenging in quantum context due to quantum superpositions and entanglement relations. In this work, we are interested in quantum generalizations of causal structures in the presence of latent common causes. These structures can be shown as a directed acyclic graph (DAG), where nodes are quantum systems, and edges are quantum operations\footnote{In the context of quantum computation \citep{hogg1996quantum}, a quantum operation is called a quantum channel.}. However, the key theoretical distinction between an entirely classical causal structure and a quantum casual structure is the concept of \textit{coexisting}. Because of the impossibility of cloning, the outcomes and the quantum systems that led to them do not exist simultaneously. If a system $X$ is measured to produce $Y$, then $\rho_{XY}$ is not defined and hence neither is the entropy $S(\rho_{XY})$ \citep{weilenmann2017analysing}. For a given causal structure, a \textit{coexisting set} of systems is one for which a joint state can be defined \citep{chaves2015information,weilenmann2017analysing, weilenmann2020analysing}. If we pick a coexisting set of nodes (e.g., a classical system, or a set of nodes that are created at the same instance of time, i.e., they do enjoy a joint density operator), then we can investigate the identification of quantum causal structures in the presence of latent confounders. \begin{wrapfigure}{r}{0.49\textwidth} \vspace{-.2in} \centering \includegraphics[width=0.55\columnwidth]{HiddenCausal.pdf} \caption{We develop a theoretical framework for solving the problem of quantum common cause.} \label{fig:quantumcommoncause} \vspace{-.25in} \end{wrapfigure} In this paper, we consider causality between two coexisting quantum subsystems. As a part of the evaluation framework, we provide a model of such a coexisting system, where two entangled qubits are used, and one of the qubit is transmitted over a quantum channel. Similarly, three entangled qubits are used, and two of them are transmitted over two separate quantum channels. The models can be further generalized, while note that the subsystems which are being considered for quantum causality relationships have to coexist, unlike in the classical case where it is not necessary for the sub-systems to coexist. To address this problem, we introduce a theoretical framework to merge quantum information science with causal inference using entropic principles. Classically, it has been proposed and tested that minimization of the trade-off between the entropy of the (hidden) common cause $Z$ (i.e., $H(Z)$) and the conditional mutual information of observed variables $X$ and $Y$ given $Z$ (i.e., $I(X;Y|Z)$) can be used to distinguish the \textit{latent graph} $X\gets Z\to Y$ ($Z$ is an unmeasured confounder) from the directed graphs $X\to Y$ and $X\gets Y$ based on observational data alone under certain assumptions \citep{KocaogluNEURIPS2020} (a brief review is given in Section \ref{sec:classic}). We will provide the first generalization of this approach to the quantum domain (Figure \ref{fig:quantumcommoncause}). Even though the paper considers an approach for quantum causal inference, we also apply the proposed approach to a classical setup, where two bits are transmitted over a binary symmetric channel (to illustrate the case of no confounder), or two bits are transmitted over two separate channels (to illustrate the case of latent confounder). We show that the proposed approach outperforms the classical causal inference in \citep{KocaogluNEURIPS2020} due to the use of quantum density matrix. We note that finding the optima over a quantum density matrix rather than over the probability distribution function provides larger degrees of freedom thus resulting in improved results. This demonstrates that the proposed approach can also be used for classical causal inference with improved results. Our main contributions are as follows: \noindent$\bullet$ Inferring causality in the presence of latent confounders from observational data alone is one of the most important and challenging problems in statistical inference. We propose an iterative algorithm, called {\sc \textbf{\texttt{QInferGraph}}}, for identifying \textit{latent confounders} in Section \ref{sec:method}. Our method leverages the concept of quantum conditional matrices to unify the solution for classical and quantum (latent) common cause problem in a principled way.\\ \noindent$\bullet$ We evaluate the proposed approach for classical causal inference. By leveraging optimization over density matrices, the proposed approach is shown to outperform the results of classical causal inference in \citep{KocaogluNEURIPS2020}. \noindent$\bullet$ We put forth an experimental scheme that can be used to confront our theoretical framework. We consider a minimalistic model of an unknown message (possibly encrypted) with unknown origin in a two-node quantum network with the possibility of the presence of a latent common cause, where nodes are a \textit{coexisting set} of quantum systems for which a joint density matrix can be defined. Entangled quantum subsystems are used, where subsystems are communicated over noisy channels (e.g., optical fiber) to create such coexisting set of quantum systems. We prove that only using the joint density matrix of the observed two quantum system, we can identify the originator of the message (i.e., the sub-system that did not encounter the noisy channel). To verify the validation of {\sc \textbf{\texttt{QInferGraph}}}, we use realistic quantum noisy links such as quantum symmetric channel and depolarizing channel (valid for quantum networking and quantum communications) (Section \ref{sec:evaluation}). Moreover, we show that finding the joint probability distribution and using the classical common entropy technique may result in erroneous outcomes, as shown in Section \ref{sec:mapping}, thus showing that the classical approaches cannot be directly used on quantum systems. This specific approach can lay the foundations of identifying originators of malicious activity on multi-node quantum networks. The rest of the paper is organized as follows. In Section \ref{sec:classic}, we review the classical causal inference approach proposed in \citep{KocaogluNEURIPS2020} for the identification of causal structures in the presence of hidden common causes. In Section \ref{sec:method}, we generalized the classical approach to the quantum domain. In Section \ref{sec:evaluation}, we put forward an experimental scheme that can be used to validate our proposed approach using a minimalistic model of an unknown message (possibly encrypted) with unknown origin in a two-node/three-node quantum network. In Section \ref{sec:mapping}, we explain and show why should we not map quantum to classical directly. In Appendix \ref{eval:moreGQSC} and \ref{eval:moreDepolar}, we provide details on the best choice of hyper-parameters of our proposed algorithm. \subsubsection*{\bibname}} \newtheorem{assumption}{Assumption} \newtheorem{model}{Model} \newcommand{{\sc \texttt{QECI}}}{{\sc \texttt{QECI}}} \newcommand{{\sc \textbf{\texttt{QLatentSearch}}}}{{\sc \textbf{\texttt{QLatentSearch}}}} \newcommand{{\sc \textbf{\texttt{QInferGraph}}}}{{\sc \textbf{\texttt{QInferGraph}}}} \newcommand{{\cal D}}{{\cal D}} \newcommand{\fracpartial}[2]{\frac{\partial #1}{\partial #2}} \jmlrheading{1}{2021}{1-??}{4/00}{10/00}{meila00a}{Mohammad Ali Javidian and Vaneet Aggarwal and Zubin Jacob} \ShortHeadings{Quantum Causality: an Entropic Approach}{Javidian and Aggarwal and Jacob} \firstpageno{1} \begin{document} \title{Quantum Causal Inference in the Presence of Hidden Common Causes: an Entropic Approach} \author{\name Mohammad Ali Javidian \email mjavidia@purdue.edu \\ \name Vaneet Aggarwal \email vaneet@purdue.edu \\ \name Zubin Jacob \email zjacob@purdue.edu \\ \addr School of Electrical and Computer Engineering\\ Purdue University\\ West Lafayette, IN 47907, USA } \editor{} \maketitle \begin{abstract Quantum causality is an emerging field of study which has the potential to greatly advance our understanding of quantum systems. One of the most important problems in quantum causality is linked to this prominent aphorism that states \textit{correlation does not mean causation}. A direct generalization of the existing causal inference techniques to the quantum domain is not possible due to superposition and entanglement. We put forth a new theoretical framework for merging quantum information science and causal inference by exploiting entropic principles. For this purpose, we leverage the concept of conditional density matrices to develop a scalable algorithmic approach for inferring causality in the presence of latent confounders (common causes) in quantum systems. We apply our proposed framework to an experimentally relevant scenario of identifying message senders on quantum noisy links, where it is validated that the input before noise as a latent confounder is the cause of the noisy outputs. We also demonstrate that the proposed approach outperforms the results of classical causal inference even when the variables are classical by exploiting quantum dependence between variables through density matrices rather than joint probability distributions. Thus, the proposed approach unifies classical and quantum causal inference in a principled way. This successful inference on a synthetic quantum dataset can lay the foundations of identifying originators of malicious activity on future multi-node quantum networks. \end{abstract} \begin{keywords} Structure learning, Confounder, Common Cause, Optimization, Quantum causality \end{keywords} \input{intro} \input{classic} \input{method} \input{evaluate} \input{mapping} \section*{Conclusions and Future Work} This paper provides a new approach for quantum entropic causal inference in the presence of hidden common causes. As a part of the approach, an iterative algorithmic solution is provided for the optimization problem that deals with the trade-off between the entropy of the latent quantum system and the quantum conditional mutual information of the observed quantum systems. The approach is validated on quantum noisy link, where the approach detects the expected causal relation. The extension of the problem to general quantum causality graph relations between multiple variables is an open problem for the future. \acks{The authors would like to thank Xueji Wang for generating Figure \ref{fig:quantumcommoncause}. This research was supported by the Defense Advanced Research Projects Agency (DARPA) Quantum Causality [Grant No. HR00112010008]. } \newpage \section{Why Should We Not Map Quantum to Classical Directly?}\label{sec:mapping} Here, we show why classical common entropy approach do not directly apply to the quantum case. We emphasize that although a joint density operator (matrix) can be converted to a joint probability distribution (as explained in Example \ref{ex:counterex}), we \textit{lose} some \textit{quantum information} due to the loss of entanglement. We give an example that shows converting a joint density matrix $\rho_{XY}$ directly to a joint probability distribution $p(X,Y)$, and then applying classical common entropy approach on $p(X,Y)$ will not lead to the correct results. \begin{algorithm}[!ht] \caption{Rotational procedure for computing the joint probability distribution of a joint density matrix}\label{alg:rotate} \SetAlgoLined \scriptsize\KwIn{Joint density matrix of quantum systems $X$ and $Y$ i.e., $\rho_{XY}$.} \KwOut{Joint probability distribution $p(X,Y)$ corresponding to the joint density matrix $\rho_{XY}$.} \tcc{\textcolor{blue}{Compute eigenvalues and eigenvectors of $\rho_X.$}} $[V_1, D_1] = eig(\rho_X)$\; \tcc{\textcolor{blue}{Compute eigenvalues and eigenvectors of $\rho_Y.$}} $[V_2, D_2] = eig(\rho_Y)$\; \tcc{\textcolor{blue}{Rotational procedure}} $U= V_1\otimes V_2$\; $\rho'_{XY}= U^\dagger \rho_{XY}U$\; \textbf{return} $p(X,Y)$ as the entries on the main diagonal of $\rho'_{XY}$. \end{algorithm} \begin{example}[Counter Example]\label{ex:counterex} Assume the depolarizing channel as described in Model \ref{model:depolar}, Part II. We already know that $X$ causes $Y$ in this model. To convert the joint density matrix $\rho_{XY}$, we use a rotational procedure explained as follows: Assume that $\rho_{XY}$ is rotated using a unitary matrix $U$. Let us say $\rho_{XY} = U \rho'_{XY} U^\dagger$. So, the joint density matrix $\rho'_{XY}$ is computed as $\rho'_{XY}= U^\dagger \rho_{XY}U$. To compute the unitary matrix $U$ for a given $\rho_{XY}$ we use the eigenspaces of $\rho_X$ and $\rho_Y$, where $\rho_X=\textbf{\textrm{Tr}}_Y(\rho_{XY})$ and $\rho_Y=\textbf{\textrm{Tr}}_X(\rho_{XY})$ are computed by tracing out $Y$ and $X$, respectively. This simple observation enables us to design a procedure that converts a joint density matrix $\rho_{XY}$ to a joint probability distribution $p(X,Y)$ in a way that it takes into account the rotation. This procedure is formally described in Algorithm \ref{alg:rotate}. By converting the joint density matrix $\rho_{XY}$ directly to a joint probability distribution $p(X,Y)$, using Algorithm \ref{alg:rotate}, and then applying classical entropic causal inference, i.e., Algorithm \ref{alg:InferGraph} on $p(X,Y)$ we obtain the results represented in Table \ref{t:mapping} which are opposite to the expected results in all cases. This confirms that classical statistics are not adequate for identification of cause–effect relations in quantum systems due to accessibility of a richer spectrum of causal relations in quantum scenarios. \begin{table}[!htbp] \caption{Classical Approach to Identify Direct Graph for Model \ref{model:depolar} does not work.} \label{t:mapping} \centering \begin{tabular}{c|c|ccccccccccc|} \multicolumn{2}{c}{} & \multicolumn{10}{c}{\text{$p$}}&\multicolumn{1}{c}{} \\ \cline{3-13} \multicolumn{1}{c}{}&& 0.01 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 0.99\\ \cline{2-13} &0.7&\textcolor{red}{F}& \textcolor{red}{F} &\textcolor{red}{F}& \textcolor{red}{F} &\textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F} &\textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}\\ &0.8 & \textcolor{red}{F}& \textcolor{red}{F} &\textcolor{red}{F}& \textcolor{red}{F} &\textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F} &\textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}\\ $\alpha$&0.9&\textcolor{red}{F}& \textcolor{red}{F} &\textcolor{red}{F}& \textcolor{red}{F} &\textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F} &\textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}\\ &1&\textcolor{red}{F}& \textcolor{red}{F} &\textcolor{red}{F}& \textcolor{red}{F} &\textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F} &\textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}& \textcolor{red}{F}\\ \cline{2-13} \end{tabular} \end{table} \end{example} \section{Proposed Entropic Approach for Confounder Discovery in Quantum Systems}\label{sec:method} In this section, we provide an approach for identifying latent graphs in quantum systems, where we assume the Assumptions \ref{mainassump1} and \ref{mainassump2}, with the entropy replaced by the von-Neumann entropy, $S(X) = - {\sf tr}(\rho_X\log \rho_X)$. We first briefly review the formalism of \textit{quantum conditional density matrices}, which provides a solid framework for adapting classical iterative algorithms (Algorithm \ref{alg:LatentSearch} and \ref{alg:InferGraph}) to the quantum domain. Then, the proposed algorithm to identify latent graphs is described which uses the concept of the quantum conditional density matrix. \subsection{Conditional Density Matrix}\label{sec:conditionaldensity} Quantum theory can be understood as a non-commutative generalization of classical probability theory wherein probability measures are replaced by density operators \citep{leifer2013towards}. Analogies between the classical theory of Bayesian inference and the conditional states formalism for quantum theory are listed in Table \ref{table:analogies}. \begin{table*}[!ht] \caption{Analogies between classical and quantum formalism \centering \resizebox{\textwidth}{!}{\begin{tabular}{|l c l|} \hline \textbf{Classical Probability} & & \textbf{Quantum Theory}\\ \hline \hline probability distribution $p(X)$ & & density operator (matrix) $\rho_X$\\ \hline joint distribution $p(X,Y)$& & joint density $\rho_{XY}$\\ \hline marginal distribution $p(X)=\sum_Yp(X,Y)$&& partial trace $\rho_X=Tr_Y(\rho_{XY})$\\ \hline conditional probability&& conditional density \citep{leifer2007conditional,leifer2013towards} \\ $p(X|Y)=p(X,Y)/p(Y)$&&$\rho_{X|Y}=(I_X\otimes \rho_Y^{-1/2})\rho_{XY}(I_X\otimes \rho_Y^{-1/2})$\\ \hline instance conditional probability && instance conditional density matrix \citep{javidian2021quantum}\\ $p(X|Y=y)=\frac{p(X,Y=y)}{\sum_x p(X,Y=y)}$&& $\rho_{X|Y=|y\rangle}=\frac{Tr_Y\{\rho_{XY}\star |y\rangle\langle y|\}}{trace\{Tr_Y\{\rho_{XY}\star |y\rangle\langle y|\}\}}$, where\\ &&$\rho_{XY}\star |y\rangle\langle y|=(I\otimes (|y\rangle\langle y|)^{1/2})*\rho_{XY}* (I\otimes (|y\rangle\langle y|)^{1/2})$\\ \hline \end{tabular}} \label{table:analogies} \end{table*} Quantum conditional densities are a generalization of classical conditional probability distributions. However, to generalize conditional probabilities to the quantum case, several approaches have been proposed in the literature. The three following generalizations are the best known in the literature of quantum information: (1) quantum conditional expectation \citep{umegaki1962conditional}, (2) quantum conditional amplitude operator \citep{cerf1997negative,cerf1999quantum}, and (3) quantum conditional states \citep{leifer2007conditional,leifer2013towards}. Arguably, quantum conditional states are the most useful generalization of conditional probability from the point of view of practical applications. For example, quantum conditional states have been used in \citep{leifer2013towards} to build a quantum theory of Bayesian inference. Since quantum conditional states provides a closer analogy between quantum theory and classical probability theory, we choose this formalism to define quantum conditional density matrices. We will see that this formalism plays a significant role in the design and success of our entropic quantum causal inference algorithm. \subsection{{\sc \textbf{\texttt{QLatentSearch}}}: An Algorithm for Computing Exact Quantum Common Entropy} In this section, we propose an iterative algorithm (Algorithm \ref{alg:qLatentSearch}) that discovers the trade-off between the entropy of the unmeasured confounder and the quantum conditional mutual information of two observed quantum systems given the unmeasured confounder, which is fundamental for designing an algorithm for the identification of latent confounders in quantum systems as we show in the next subsection. This trade-off is formally defined as follows: \begin{equation}\label{eq:qlossfunc} L = I_Q(X;Y|Z)+\beta S(Z) \end{equation} Note that $I_Q(X;Y|Z)=0$ implies that the quantum conditional independence of $X$ and $Y$ given $Z$ \citep[Theorem 3]{allen2017quantum}. Having low von Neumann entropy of hidden common cause $Z$, i.e., $S(Z)$ under the quantum version of Assumption \ref{mainassump1} and \ref{mainassump2} enable us to identify latent graphs from direct/mediator graphs in practice, as we show in Section \ref{sec:evaluation}. For this purpose, rather than searching over $\rho_{XYZ}$ and enforcing the constraint $\rho_{XY}=Tr_Z(\rho_{XYZ})$, we can search over $\rho(Z|X,Y)$ and set $\rho_{XYZ} = (\rho_{XY}^{-1/2}\otimes I_Z)\rho(Z|X,Y)(\rho_{XY}^{-1/2}\otimes I_Z)$ because: \begin{equation*} \label{eq1} \begin{split} L &= I_Q(X;Y|Z)+\beta S(Z)\\ & = S(XZ) + S(YZ) - S(Z) - S(XYZ) + \beta S(Z)\\ &= S(XZ) + S(YZ) - S(XYZ) + (\beta-1) S(Z)\\ &= S(X) + S(Z|X)+S(Y)+S(Z|Y)-S(XY)-S(Z|X,Y)+(\beta-1) S(Z)\\ &= S(Z|X)+S(Z|Y)-S(Z|X,Y)+(\beta-1) S(Z)+I_Q(X;Y) \end{split} \end{equation*} Note that $\rho(Z|Y)=Tr_X((\rho^{1/2}(X|Y)\otimes I_Z)\rho(Z|X,Y)(\rho^{1/2}(X|Y)\otimes I_Z))$, $\rho(Z|X)=Tr_Y((\rho^{1/2}(Y|X)\otimes I_Z)\rho(Z|X,Y)(\rho^{1/2}(Y|X)\otimes I_Z))$, and $\rho_Z=Tr_{X,Y}((\rho^{1/2}_{XY}\otimes I_Z)\rho(Z|X,Y)(\rho^{1/2}_{XY}\otimes I_Z))$. So, we have $L=L(\rho(Z|X,Y))$, which is the counterpart of the classical loss function in Equation \ref{eq:lossfunc} with the following differences: (i) rather than using (conditional) probability distributions, we use (conditional) density matrices, and (ii) rather than using R\'{e}nyi entropy, we use the von Neumann entropy. We aim to optimize the objective $L$ over $\rho(Z|X,Y)$. Although first order methods (e.g., gradient descent) or genetic algorithm (GA), which is a metaheuristic method inspired by the process of natural selection, can be used to find a stationary point of the optimization problem in (\ref{eq:qlossfunc}), as we empirically observe the convergence is unattainable/slow and the performance is very sensitive to the tuning parameters such as step size and the mutation probability. This optimization problem is difficult to perform numerically because the boundary of the space of positive semidefinite matrices is hard to compute. In order to provide a scalable algorithm for this optimization, we extend the iterative algorithm that was proposed for classical version of the problem in \citep{KocaogluNEURIPS2020}. The proposed iterative algorithm for the optimization of $L$ is described in Algorithm \ref{alg:qLatentSearch}, and is called {\sc \textbf{\texttt{QLatentSearch}}}. This algorithm starts from a random initialization $\rho_1(Z|X,Y)$, and then at each iteration $i$ does the following two phases to update $\rho_{i+1}(Z|X,Y)$ from $\rho_i(Z|X,Y)$ to finally minimize the loss function $L$ in (\ref{eq:qlossfunc}): \begin{itemize} \item \textbf{Calculate Phase:} In this phase we use partial trace to get $\rho_i(Z|X)$ (line 3-5), $\rho_i(Z|Y)$ (line 6-8), and $\rho_Z^i$ (line 9) from $\rho_{XYZ}^i$. \item \textbf{Update Phase:} In this phase we update $\rho_{i+1}(Z|X,Y)$ to get $\rho_{XYZ}^{i+1}$ (line 10) for the next iteration. \end{itemize} \begin{algorithm}[!ht] \caption{\textbf{{\sc \textbf{\texttt{QLatentSearch}}}}, An Iterative Algorithm for Computing Exact Quantum Common Entropy}\label{alg:qLatentSearch} \SetAlgoLined \scriptsize\KwIn{Joint density matrix $\rho_{XY}$; Number of iterations $N$; $\beta$ parameter in the loss function $L=I_Q(X; Y|Z)+\beta S(Z)$, Initialization of $\rho_1(Z|X,Y)$.} \KwOut{Joint density matrix $\rho_{XYZ}$.} \For{$i = 1:N$}{ \tcc{Form the joint density matrix:} $\rho_{XYZ}^i = (\rho_{XY}^{1/2}\otimes I_Z)\rho_i(Z|X,Y)(\rho_{XY}^{1/2}\otimes I_Z)$\; \tcc{\textcolor{blue}{Calculate Phase:}} \tcc{\textcolor{red}{(i) Calculate $\rho_i(Z|X)$:}} $\rho_{XZ}^i=Tr_Y(\rho_{XYZ}^i)$ \tcp{Then, compute $\rho_{XI_YZ}^i$ by reordering the entries of $\rho_{XZ}^i$} $\rho_{X}^i=Tr_Z(\rho_{XZ}^i)$\; $\rho_i(Z|X)\gets ((\rho_{X}^i)^{-1/2}\otimes I_{YZ})\rho_{XI_YZ}^i((\rho_{X}^i)^{-1/2}\otimes I_{YZ})$\; \tcc{\textcolor{red}{(ii) Calculate $\rho_i(Z|Y)$:}} $\rho_{YZ}^i=Tr_X(\rho_{XYZ}^i)$ \tcp{Then, compute $\rho_{I_XYZ}^i=I_X\otimes\rho_{YZ}^i$} $\rho_{Y}^i=Tr_Z(\rho_{YZ}^i)$\; $\rho_i(Z|Y)\gets (I_{X}\otimes(\rho_{Y}^i)^{-1/2}\otimes I_{Z})\rho_{I_XYZ}^i(I_{X}\otimes(\rho_{Y}^i)^{-1/2}\otimes I_{Z})$\; \tcc{\textcolor{red}{(iii) Calculate $\rho_Z^i$:}} $\rho_{Z}^i=Tr_{XY}(\rho_{XYZ}^i)$\; \tcc{\textcolor{blue}{Update Phase:}} $\rho_{i+1}(Z|X,Y)\gets (I_{XY}\otimes(\rho_{Z}^i)^{\beta-1})\rho_i(Z|X)\rho_i(Z|Y)$\; } \textbf{return} $\rho_{XYZ}:=(\rho_{XY}^{1/2}\otimes I_Z)\rho_{N+1}(Z|X,Y)(\rho_{XY}^{1/2}\otimes I_Z)$. \end{algorithm} \subsection{\textbf{{\sc \textbf{\texttt{QInferGraph}}}:} An Algorithm for the Identification of Latent Confounders} In this section, we propose a quantum entropic approach to causal inference that can discern the difference between causation and correlation. Specifically, under the assumption that \textit{there are no low-entropy mediators} (Assumption \ref{mainassump2}), Algorithm \ref{alg:qLatentSearch} can be used to distinguish causation from spurious correlation between two observed quantum systems. This enables us to distinguish latent graph in Figure \ref{fig:possibilities}(a) from the triangle or direct graphs in Figure \ref{fig:possibilities}(b)-(c). Our main assumption is that the latent confounders, if they exist, have small von Neumann entropy. In other words, in Figure \ref{fig:possibilities}(a), $S(Z)\le \theta$ for some $\theta$. Similar to the classical version of this problem, we conjecture that $\theta=\alpha\min\{S(X),S(Y)\}$ for some $\alpha < 1$. Considering Assumption \ref{mainassump1} and Assumption \ref{mainassump2} along with {\sc \textbf{\texttt{QLatentSearch}}}~(Algorithm \ref{alg:qLatentSearch}), we propose an algorithm, called {\sc \textbf{\texttt{QInferGraph}}}~(Algorithm \ref{alg:qInferGraph}), to identify latent graphs. \begin{algorithm}[!ht] \caption{\textbf{{\sc \textbf{\texttt{QInferGraph}}}}: Identifying the Latent Graph}\label{alg:qInferGraph} \SetAlgoLined \scriptsize\KwIn{Joint density matrix $\rho_{XY}$; Number of iterations $N$; $I_Q(X; Y|Z)$ threshold $T$; $S(Z)$ threshold that is determined by $\theta=\alpha\min(S(X),S(Y))$; $\{\beta_i\}_{i=1}^N$; The number of rows (or equivalently, columns) of $X,Y$, and $Z$, i.e., $r,m$, and $n$, respectively.} \KwOut{"Latent Graph" if $Z$ is an unmeasured confounder for $X$ and $Y$, otherwise, returns "Triangle or Direct Graph".} \For{$i = 1:N$}{ $\rho_{XYZ}^i\gets \textrm{\textbf{{\sc \textbf{\texttt{QLatentSearch}}}}}(\rho_{XY}, \alpha,\beta_i,r,m,n)$\; Calculate $I_Q^i(X; Y|Z)$ and $S_i(Z)$ from $\rho_{XYZ}^i$\; } $S=\{i: I_Q^i(X; Y|Z)\le T\}$\; \uIf{$\min(S_i(Z):i\in S)>\theta$ or $S=\text{\O}$ }{ \textbf{return} Triangle or Direct Graph\; }\Else{ \textbf{return} Latent Graph\; } \end{algorithm} In short, {\sc \textbf{\texttt{QInferGraph}}}~calls {\sc \textbf{\texttt{QLatentSearch}}}~$N$ times to figure out if there exist a $W$, for which $I_Q(X;Y|W)<T$, i.e., $W$ makes $X$ and $Y$ conditionally independent. Also, the von Neumann entropy of $W$ is enough small such that $S(W)<\alpha\min\{S(X),S(Y)\}$ for some $\alpha$ in practice. If there exist such a $W$, the algorithm declares $W$ is a latent confounder. In other words, latent graph represents correlation without causation relationship between observed quantum systems $X$ and $Y$. Otherwise, very likely such a $W$ that minimizes the loss function $L$ does not exist, and {\sc \textbf{\texttt{QInferGraph}}}~declares that a triangle graph or a direct graph represents the connection between $X$ and $Y$ better than a latent graph in this case. In the next section we conduct experiments to verify this procedure in practice. \if 0 \begin{remark} Although first order methods (e.g., gradient descent) or genetic algorithm (GA), which is a metaheuristic method inspired by the process of natural selection, can be used to find a stationary point of the optimization problem in (\ref{eq:qlossfunc}), as we empirically observe the convergence is unattainable/slow and the performance is very sensitive to the tuning parameters such as step size and the mutation probability. This optimization problem is difficult to perform numerically because the boundary of the space of positive semidefinite matrices is hard to compute. \end{remark} \fi

98aec874fcd99e0613617be8c5c82cb4d00259b7

\section{Introduction} \label{sec:introduction} Quantum computing is a promising new paradigm for computer science. Quantum algorithms have proven themselves superior to classical ones in some classes of problems. The major examples are Shor’s algorithm for solving the hidden subgroup problem and Grover’s algorithm for the unstructured search problem, but quantum computing is not limited to them\cite{Montanaro:2016}. Simulating complex atomic systems using a quantum computer\cite{quantum-simulation} could profoundly impact physics research since a quantum computer can naturally simulate the effects of quantum mechanics. Quantum Computing has also been impactful in the machine learning field. Basic linear algebra subroutines, such as Fourier transform, finding eigenvectors and eigenvalues, and solving linear equations, for example, exhibit exponential quantum speedups over their best known classical counterparts\cite{QML} Also, quantum machine learning aims to implement machine learning algorithms in quantum systems, by using the quantum properties such as superposition and entanglement to solve these problems efficiently\cite{QML_review}. However, practical limitations still hinder the development of a universal quantum computer that could use such algorithms. Current quantum computers are referred to as NISQ (Noisy Intermediate-Scale Quantum) computers \cite{NISQ} because of qubits imperfections and their available limited number, currently between 50 and 100. These numbers are not sufficient to execute error-correcting codes once 9 qubits are necessary to make 1 qubit fault-tolerant \cite{Devitt_2013}. Another limiting factor of NISQ computers is qubit connectivity. Not all qubits are directly connected, which limits the available two-qubit gate operations. This problem can be solved by the insertion of SWAP gates to change the qubit mapping\cite{Nisq-map}, but due to imperfections in NISQ quantum gates, the remap of qubits introduces more errors into the system. While we do not have the necessary resources to make an efficient quantum computer, an alternative solution is to make the best of the currently available resources through the development of techniques for quantum circuit synthesis and quantum circuit optimization. This paper focuses on quantum circuit synthesis. Quantum circuit synthesis is mainly made through the synthesis of reversible boolean circuits\cite{AIG-MIG}\cite{ESOP}\cite{BDD}\cite{DFS-BFS}. These techniques focus on generating a classical reversible circuit and then matching it with an equivalent quantum circuit. This is possible because quantum computing is naturally reversible. Hence, it is easy to convert a reversible boolean circuit to a quantum circuit. Techniques to synthesize quantum circuits directly such as genetic programming\cite{genetic09}\cite{GP-quantum} or swarm algorithm\cite{swarm} have also been proposed. They are an interesting alternative once it enables the optimization of the circuit over the synthesis process. Reinforcement learning also presented good results in a correlated problem. In \cite{Zeilinger:2018} a reinforcement learning technique called Projective Simulation \cite{PS-original} was used to develop an artificial agent capable to discover new optical quantum experiments by learning new placements for the optical elements. In this sense, this paper proposes the application of the Projective Simulation reinforcement learning technique as a new method for quantum circuit synthesis. We demonstrate that reinforcement learning could be a valuable tool for quantum circuit synthesis: the agent could learn alternative circuit layouts to perform computations. However, we first present how to model the problem of quantum circuit synthesis in a way an artificial agent could learn it. The paper is structured as follows. Section \ref{sec:background} explains the basics of quantum computing and projective simulation. Section \ref{sec:Method} describes the agent's model and the simulation environment parameters, detailing our contributions. Section \ref{sec:results} shows the results obtained to different circuit configurations. In Section \ref{sec:conclusion} we discuss the results and propose future developments of this work. \section{Background} \label{sec:background} This section presents a short introduction to quantum computing and quantum circuits. We focus on Bell states as the target circuits we want to synthesize. In Section \ref{ssec:Projective} we introduce the reinforcement learning technique used in our proposal. \subsection{Quantum Computing and Quantum Circuits} \label{ssec:quantumC} A quantum bit (qubit) is the basic unit of quantum information. A qubit is represented by a unit vector in a Hilbert space. Let $\ket{0}$ and $\ket{1}$ the computational basis of quantum computing. They are represented by the vectors \[ \ket{0} = \begin{bmatrix} 1 \\ 0 \end{bmatrix} \hspace{2em} \ket{1} = \begin{bmatrix} 0 \\ 1 \end{bmatrix}. \] A qubit can take the value of the computational basis and also any intermediate value between $\ket{0}$ and $\ket{1}$. The qubit is then described as a linear combination of the vectors forming the basis: \[ \ket{\psi} = \alpha\ket{0} + \beta\ket{1} \] where $\alpha$ and $\beta$ are complex numbers. These intermediate states are called \textit{superpositions}. A multiple qubit system follows the notation \[\ket{00} = \ket{0}\ket{0} = \ket{0} \otimes \ket{0},\] where $\otimes$ is the tensor product operation. Tensor product is a way to put qubits together and form larger vector spaces. When a system of two qubits cannot be described as the tensor product between the qubits that make it, the qubits are \textit{entangled} \cite{horodecki2009quantum}. Figure \ref{fig:Bell} describes the entangled states of a 2-qubit system, called Bell's states, and of a 3-qubit system, called $GHZ$-state \cite{greenberger1990bell,greenberger1989going}. Entanglement and superposition are properties explored by quantum algorithms to achieve speedup over classical ones.\cite[Chapter 1.3.7, 1.4]{Nielsen:2010} \begin{figure}[h] \[ \frac{\ket{00} + \ket{11}}{\sqrt{2}} = \ket{\beta_{00}}, \] \[ \frac{\ket{01} + \ket{10}}{\sqrt{2}} = \ket{\beta_{01}}, \] \[ \frac{\ket{00} - \ket{11}}{\sqrt{2}} = \ket{\beta_{10}}, \] \[ \frac{\ket{01} - \ket{10}}{\sqrt{2}} = \ket{\beta_{11}}, \] \[ \frac{\ket{000} + \ket{111}}{\sqrt{2}} = GHZ\text{-state}. \] \caption{A set of entangled states: Bell's states are obtained by the entanglement of 2 qubits and $GHZ$-state is obtained by 3 qubits.} \label{fig:Bell} \end{figure} Operations in quantum computing are described by $2^n \times 2^n$ dimension unitary complex-valued matrices where $n$ denotes the number of qubits in the quantum system. A unitary matrix follows the rule $U^{\dagger}U = UU^{\dagger} = I$, whereas $U^\dagger$ is the Hermitian conjugate of U and I is the identity matrix. These operations are known as quantum logic gates. Figure \ref{fig:pauli} shows some examples of common single-qubit gates. Gates $X$, $Y$, and $Z$ rotate the qubit around the axis $x$, $y$, and $z$, respectively and are known as Pauli Matrices. The $X$ gate is also known as the quantum NOT gate, because it sends the qubit from the state $\ket{0}$ to the state $\ket{1}$ and vice-versa. The $H$ gate is the Hadamard gate. It sends the qubit from the computational basis to a state ``half the way'' from $\ket{0}$ to $\ket{1}$. All these gates are single-qubit gates. The $CNOT$ gate is the controlled-not gate. \begin{figure}[ht!] \centering $ X = \begin{bmatrix} 0 & 1 \\ 1 & 0 \end{bmatrix} \hspace{2em} Y = \begin{bmatrix} 0 & -i \\ i & 0 \end{bmatrix} \hspace{2em} Z = \begin{bmatrix} 1 & 0 \\ 0 & -1 \end{bmatrix} $ \bigskip $ H = \begin{bmatrix} 1 & -1 \\ 1 & 1 \end{bmatrix} \hspace{2em} CNOT = \begin{bmatrix} 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 0 & 1 \\ 0 & 0 & 1 & 0 \end{bmatrix} $ \caption{Matrix representation of some basic quantum gates, notably Pauli (X,Y,Z), Hadamard and CNOT.} \label{fig:pauli} \end{figure} All quantum logic gates can be transformed into a controlled-gate. Controlled gates have two sets of qubits, target qubits and control qubits. Once control qubits are in the correct value - $\ket{1}$ for positive controls or $\ket{0}$ for negative controls - the gate is applied on the target qubits. When drawing a circuit, a control qubit is represented by a bullet connected by a line to the gate it is related to. In a circuit, a qubit is represented by a horizontal line. Quantum logic gates are represented by labeled boxes, with few exceptions. The computation is performed from left to right, by the application of the gates' operation over the qubits. The maximum number of sequential gates is called the circuit depth. Figure \ref{fig:circuit-example} shows an example of a quantum circuit composed of 3 qubits and 4 quantum logic gates, 2 single-qubit gates, and 2 positive controlled-gates. \begin{figure}[h!] \centerline{ \Qcircuit @C=1em @R=0.7em { \lstick{\ket{0}} & \gate{H} & \qw & \ctrl{1} & \qw \\ \lstick{\ket{0}} & \gate{X} & \ctrl{1} & \gate{Z} & \qw \\ \lstick{\ket{0}} & \qw & \targ & \qw & \qw } \hspace{1cm} } \caption{A generic quantum circuit composed of two single qubit gates, a Hadamard (H) and a X gate, and two controlled gates, a CNOT and a controlled-Z gate.} \label{fig:circuit-example} \end{figure} The circuit starts by applying a Hadamard gate on the first qubit and an X gate on the second qubit. Then, there is a CNOT operation between the second and third qubits, written in its most common form. The last operation is a controlled-$Z$ operation between the first and the second qubit. Current quantum circuits have physical problems that should be considered. Qubits have a maximum lifespan before they degenerate into noise, called \textit{decoherence time}, and quantum logic gates have an error rate caused by imperfections in implementation. When these limitations are considered, a reliable quantum circuit should have gates with minimum error and shorter depths. \subsection{Projective Simulation} \label{ssec:Projective} \textit{Projective Simulation} (PS) is a learning model based on stochastic processing of an episodic memory that can be applied in reinforcement learning solutions. This technique was introduced by \cite{PS-original} and was further developed by \cite{PS-definition}. The PS agent is an entity located in a partially unknown environment, that receives perceptions through its sensors. The agent acts in response to the environment and its actions are rewarded by the environment. The agent has a specific type of memory named Episodic and Compositional Memory (ECM). The ECM is represented by a weighted directed graph, where nodes are called \emph{clips}. Clips are the basic memory units corresponding to single episodic experiences, as perceptions (perception-clip), action (action-clip), or a combination of both (perception-action-clip). A clip can be excited when the agent receives the actual state of the environment. The excitation hops to adjacent clips with some probability related to the weight of the edge connecting those clips. Thus each coming perception from the environment starts a random walk through the ECM. The random walk ends when an action-clip is reached, which generates a real action in the environment. The learning process of PS is defined through modifications in the agent's ECM. These modifications can be the addition or the deletion of a clip or the edge weight update between clips. At the beginning of the simulation, the ECM is a \textit{tabula rasa}, i.e. the agent has no preferences towards a specific kind of action. According to the received reward from the environment, the agent's ECM is updated following a predefined set of rules. The changes in the agent's ECM result in a new decision making, that leads to a new reward from the environment, which changes the structure of the ECM. Since the PS agent has no explicit model of the environment which predicts the next state or reward, it is considered model-free. The agent has the probability $P^{(t)}(a|s)$ of making the action \textit{a} given the perception \textit{s}. However, a complete description of the agent connects the probability $P^{(t)}(a|s)$ with its memory internal state at time \textit{t}, and specifies how the memory is updated as the agent interacts with the environment. In order to clarify the main concepts of the PS model, we list some formal terms used in PS: \begin{itemize} \item \textbf{Episodic and Compositional Memory (ECM):} is the main element of PS, defined as a weighted directed graph. Each node is called a clip. \item \textbf{Clip:} represents short episodic experiences. Clips are defined by L-tuples $c = (c_1, c_2, ..., c_L)$. Each $c_i$ is either a perception ($s \in S$) or an action ($a \in A$), where both are defined below. \item \textbf{Perceptions:} are defined as $N$-tuples $s = (s_1, s_2,...,s_N) \in S \equiv S_1 \times S_2 \times ... \times S_N, \> s_i \in {1,...,\card{S}}$, where the number of possible perceptions is given by $S \equiv \card{S} \equiv \card{S_1}...\card{S_N}$. \item \textbf{Actions:} are defined as $M$-tuples $a = (a_1, a_2,...,a_M) \in A \equiv A_1 \times A_2 \times ... \times A_M, \> a_i \in {1,...,\card{A}}$, where the number of possible actions is given by $A \equiv \card{A} \equiv \card{A_1}...\card{A_M}$. \item \textbf{Edges:} Each edge connecting clip $c_i$ to clip $c_j$ has a dynamic weigth $h^{(t)}(c_i,c_j)$, called $h$-value, which changes over time. \item \textbf{Hopping Probability:} The probability of an excitation hopping from clip $c_i$ to clip $c_j$ is given by \[p^{(t)}(c_i,c_j) = \frac{h^{(t)}(c_i,c_j)}{\sum_k h^{(t)}(c_i,c_k)}\] which represents the sum of the $h$-values of all edges connecting the clips $c_k$ to $c_i$ for all clip $k$. \item \textbf{Interface:} The interface between the PS agent and the environment is made with its sensors and actuators and their connection to memory. An external perception $s$ excites a certain perception-clip $c$ according to the function $\mathcal{I}(c|s)$. Similarly, an action-clip couples out to perform a real action $a$ according to the function $\mathcal{O}(a|c)$. The input function $\mathcal{I}(c|s)$ and the output function $\mathcal{O}(a|c)$ connect the internal random walk, described by the hopping probability $p^{(t)}(c_j|c_i)$, with the external behaviour of the agent, described by $P^{(t)}(a|s)$. \item \textbf{Damping Parameter:} PS admits an optional parameter $\gamma$ called damping parameter, usually $(0\leq\gamma\leq 1)$. This parameter can be seen as the agent ``forgetting'' over time. This value weakens the connections between clips, allowing the agent to adapt to changing environments. \end{itemize} The process behind PS is stochastic. Each step $t$ begins with a perception from the environment exciting a memory clip $c_i$ according to function $\mathcal{I}(c|s)$. Then, the excitation hops from clip $c_i$ to an adjacent clip $c_j$, with the probability $p^{(t)}(c_j|c_i)$. This process continues as a random walk, allowing the excitation to move through the clip network. The walk ends when an action-clip is reached and generates an equivalent action in the environment according to function $\mathcal{O}(a|c)$. Each action receives from the environment a reward $\lambda \ge 0$ that is used, together with damping parameter $\gamma$, to update all the $h$-values as follows: \begin{enumerate}[(i)] \item For activated edges, i.e. the edges that were crossed in the last random walk \[ h^{(t+1)}(c_i,c_j) = h^{(t)}(c_i,c_j) - \gamma (h^{(t)}(c_i,c_j) -1) + \lambda \] where $t$ is the current step, $\gamma$ $(0\leq\gamma\leq 1)$ is the damping parameter and $\lambda$ is the reward given by the environment. \item For other edges the $h$-values are damped \[ h^{(t+1)}(c_i,c_j) = h^{(t)}(c_i,c_j) - \gamma (h^{(t)}(c_i,c_j) -1) \] \end{enumerate} When the agent earns a positive reward, the edges that were crossed during the random walk that ended in the correct action are strengthened. These edges then have a greater probability of being chosen again in future steps. Still, if a wrong action is taken and no reward is earned, i.e. $\lambda=0$, all edges are damped, including those that were crossed in the last random walk. \begin{subsubsection}{Temporal Correlation} In some reinforcement learning problems, the reward is not related to just the last action made by the agent, but also to previous actions. This case is referred to as Temporal Correlation. PS allows the use of temporal correlation through a new parameter in the reward function. This parameter is called edge glow and represents the level of excitement of an edge. For example, every time an edge is chosen, the agent sets the edge glow to the maximum value, and for each step that it is not chosen its glow fades. Thus an edge that was not chosen in the last random walk can still be partially rewarded. This effect is referred to as \emph{afterglow}. Formally, the afterglow is implemented by adding a new parameter $g$ called glow, which is related to the ECM's edges. Initially, $g=0$ for all edges. So, anytime an edge is crossed in the random walk it sets $g=1$. Every subsequent step $g$ decays by a rate $\eta$ until reaches 0, following the rule: \[ g^{(t+1)} = g^{(t)} - \eta g^{(t)}, 0 \le \eta \le 1, \] and the reward function chages to: \[ \begin{array}{lll} h^{(t+1)}(c_1,c_2) = & h^{(t)}(c_1,c_2) - \gamma(h^{(t)}(c_1,c_2)-1)\\ & + \lambda g^{(t)}(c_1,c_2) \end{array} \] \noindent where $g^{(t)}(c_1,c_2)$ represents the value $g$ of the edge connecting $c_1$ and $c_2$ at time $t$. As before, only the excited edges are strengthened by the reward $\lambda$. The difference is that the edges that were partially excited can also be partially rewarded. \end{subsubsection} \begin{subsubsection}{Composition} The compositional aspect of the ECM is a dynamic process that allows structural changes in its inner connections. In particular, it allows the spontaneous creation of new clips inside the ECM, through the combination or variation of existing clips. The new clips can represent fictitious episodes that were not tried before, thus expanding the diversity of conceivable events and actions that exists inside the ECM. As a result, the memory is less connected with the agent's real past. This effectively enables the agent to create alternative options that it has already found previously, thus making it more capable and flexible. There are many possibilities to merge clips. In \cite{PS-definition} a merging mechanism is defined for $M$-dimension action clips, considering the following cases: \begin{itemize} \item Two action clips $c_a = (a_1, a_2, ..., a_M)$ and $c_b = (b_1, b_2, ..., b_M)$ are merged in a new clip if and only if: \begin{enumerate}[(a)] \item Both corresponding actions were sufficiently rewarded through the same perception; \item The action clips $c_a$ and $c_b$ differ exactly in two components. \end{enumerate} \item When action clips $c_a$ and $c_b$ differ in theirs $i$-th and $j$-th components, the merge results in two new composed clips: $$c^{new}_1 = (a_1, a_2, ..., b_i, ..., a_j, ..., a_M)$$ \noindent and $$c^{new}_2 = (a_1, a_2, ..., a_i, ..., b_j, ..., a_M)$$ \item A new clip is only created if it is not already in the agent's ECM. \item New action clips are connected to corresponding perception clip $c_0$ with $h$-value equals to the sum of the $h$-values of the originals action clips. $$h(c_0, c^{new}_1) = h(c_0, c^{new}_2) = h(c_0, c_a) + h(c_0, c_b)$$ \noindent Furthermore, new action clips are connected to all other perception clips with initial $h$-value equals to $1$. \end{itemize} \end{subsubsection} \section{Synthesizing Quantum Circuits with Projective Simulation} \label{sec:Method} Inspiring by the good results obtained by \cite{Zeilinger:2018} in optical quantum experiments, we used Projective Simulation to develop a quantum circuits synthesizer. Our first intent was to explore the capability of reinforcement learning to synthesize circuits capable of generating entangled states. We use PS with the optional damping parameter $\gamma$ as well as temporal correlation. We defined the environment as a quantum circuit, following the architecture Tenerife from IBM Quantum Experience \footnote{https://quantum-computing.ibm.com/}. The circuit has a predetermined number of qubits (varying from $2$ until $5$) and at the beginning of the learning process, it does not have any gate. The agent is described by its episodic and compositional memory (ECM) whose initial perception-clip is the tensor product of the qubits, i.e, the environment's initial state. We set the action pool as gates and the agent's action-clip is the application of one gate in the circuit, i.e. the agent can choose an action from the set of gates composed by Hadamard, $X$, $Y$, $Z$ and $CNOT$ gates. The agent can place the gate at any qubit in the circuit, always after all the other gates already placed. We chose this architecture to limit the possibilities of placement for the CNOT gate. Every time the agent places a new gate in the circuit the next state is generated by multiplying the gate unitary matrix to the tensor product of the qubits. The agent receives the qubits as a new perception and looks for a corresponding perception-clip. If none is found a new clip is created that matches with the perception. For practical reasons, if the episode ends without a reward, the agent deletes every recently created perception-clip in the episode. Otherwise, the consumption of memory would spike fastly. We chose to apply the following reward function: \begin{equation} \lambda = base\_value - \sum e_g * \frac{d_{min}}{d_i} \end{equation} \noindent where $e_g$ is the error of each gate put in the circuit, $d_{min}$ is the depth of the minimum circuit found, $d_i$ is the depth of the circuit being rewarded and $base\_value$ is a predetermined fixed value that changes with the number of qubits. We set $d_{min}$ to starts equal to the maximum depth allowed by the circuit and it is updated dynamically during the learning process. The $base\_value$ was set to $100$ for $2$-qubits circuits and it is increased by $50$ for each new qubit, for a total of $250$ for $5$ qubits. The reward was based on the reward used by \cite{Zeilinger:2018} to learn new quantum experiments. They used a reward value $\lambda = 100$ without any penalties being calculated. We also added penalties in the reward function to make the agent consider physical limitations and prioritize better circuits. One circuit is better if the used gates have minimum error and if the depth is shorter. Thereby, the operation made by the circuit will be less affected by qubit decoherence time and errors resulted from technological implementations. We made 4 simulations, generating circuits from 2-qubits to 5-qubits. The used parameters of each simulation are summarized in Table \ref{tab:parameters}. \begin{table*}[th!] \centering \begin{tabular}{cccccc} \toprule Parameter & Simulation 1 & Simulation 2 & Simulation 3 & Simulation 4 \\ \midrule Number of Qubits & 2 & 3 & 4 & 5 \\ Damping Parameter $\gamma$ & 0.1 & 0.1 & 0.1 & 0.1 \\ Parameter $\eta$ (afterglow) & 0.1 & 0.1 & 0.1 & 0.1 \\ Maximum Circuit Depth & 4 & 5 & 6 & 7 \\ Base Value of Reward $\lambda$ & 100 & 150 & 200 & 250 \\ \bottomrule \end{tabular} \caption{Tunning for each simulation} \label{tab:parameters} \end{table*} We set the damping parameters $\gamma = 0.1$ and $\eta = 0.1$ for all performed simulations. We defined a maximum circuit depth for the circuits to limit the size of each episode. The maximum depth was 4 gates, for circuits with 2-qubits, 5 gates for circuits with 3-qubits, 6 gates for circuits with 4-qubits, and 7 gates for 5-qubits. The values were chosen based on the goal of the simulation, described below. As allowing greater depths would not make it easier for the agent to learn, it could actually hinder its progress, we decided to add a margin of two gates for the agent to learn the right circuit. \begin{figure}[b] \centerline{ \Qcircuit @C=1em @R=.7em { \lstick{\ket{0}} & \gate{H} & \ctrl{1} & \qw & \qw & \multimeasureD{2}{\frac{\ket{000}+\ket{111}}{\sqrt{2}}} \\ \lstick{\ket{0}} & \qw & \targ & \ctrl{1} & \qw & \ghost{\frac{\ket{000}+\ket{111}}{\sqrt{2}}} \\ \lstick{\ket{0}} & \qw & \qw & \targ & \qw & \ghost{\frac{\ket{000}+\ket{111}}{\sqrt{2}}} } } \caption{Basic circuit for generating a 3-qubit GHZ state.} \label{fig:ghz_circuit} \end{figure} The goal of each simulation was to find a circuit capable of generating entangled states. The goal states were the Bell state $\ket{\beta_{00}}$ and the $GHZ$-states from 3 to 5 qubits. Figure \ref{fig:ghz_circuit} shows an example of a $GHZ$-state of 3 qubits and its generator circuit. The circuit has a well-defined structure. It starts with a Hadamard gate followed by CNOT gates connecting all subsequent qubits. These states were chosen because of their relevance to quantum computing and because their circuits are well-known in the literature. We performed each simulation with a different number of episodes. The total number of episodes was chosen experimentally, increasing the value until enough successful results were obtained as explained in the next section. \section{Results and Analysis} \label{sec:results} As said in the previous section, we made 4 simulations in total. Our results are summarized in Table \ref{tab:results}. The first simulation had the 2-qubit state $\ket{\beta_{00}}$ as a goal. After 1000 episodes the agent found 26 different circuits capable of generating $\ket{\beta_{00}}$. The number of circuits found is not equal to the total number of successful circuits that the agent synthesized. Since the agent was able to exploit the same answer for its reward, we counted only the distinguished successful circuits. \begin{table*} \centering \begin{tabular}{ccc} \toprule \textbf{Goal State} & \textbf{Circuits Found} & \textbf{Number of Episodes} \\ \midrule $\ket{\beta_{00}}$ & 26 & 1000\\ GHZ 3 qubits & 31 & 5000\\ GHZ 4 qubits & 7 & 20000\\ GHZ 5 qubits & 1 & 30000\\ \bottomrule \end{tabular} \caption{Results obtained by the Projective Simulation Agent. For the goal state $\beta_{00}$ composed of two qubits, the agent was able to find 26 distinct circuits after 1000 episodes. For the 3 qubit GHZ state, the agent was able to found 31 circuits after 5000 episodes. The circuits obtained dropped significantly for the GHZ states with 4 and 5 qubits, only 7 and 1 distinct circuits were find respectively, even after longer runs.} \label{tab:results} \end{table*} Figure \ref{fig:2q-cf} shows a sample of the circuits found by the agent. The circuit at the top follows the same structure as the basic circuit in Figure \ref{fig:circuit-example}, which is the minimal circuit possible. The agent has also found equivalent circuits but they were not as efficient. The Z gate in the middle circuits does not have any effect when applied to a qubit with value $\ket{0}$, and both Y gates at the bottom circuit cancel each other. These redundant gates can be easily spotted by a later optimization operation. \begin{figure}[h!] \centerline{ \Qcircuit @C=1em @R=.7em { \lstick{\ket{0}} & \qw & \targ & \qw & \multimeasureD{1}{\frac{\ket{00}+\ket{11}}{\sqrt{2}}} \\ \lstick{\ket{0}} & \gate{H} & \ctrl{-1} & \qw & \ghost{\frac{\ket{00}+\ket{11}}{\sqrt{2}}} } } \bigskip \centerline{ \Qcircuit @C=1em @R=.7em { \lstick{\ket{0}} & \gate{Z} & \targ & \qw & \multimeasureD{1}{\frac{\ket{00}+\ket{11}}{\sqrt{2}}} \\ \lstick{\ket{0}} & \gate{H} & \ctrl{-1} & \qw & \ghost{\frac{\ket{00}+\ket{11}}{\sqrt{2}}}} } \bigskip \centerline{ \Qcircuit @C=1em @R=.7em { \lstick{\ket{0}} & \qw & \qw & \targ & \qw & \multimeasureD{1}{\frac{\ket{00}+\ket{11}}{\sqrt{2}}}\\ \lstick{\ket{0}} & \gate{Z} & \gate{H} &\ctrl{-1} & \qw & \ghost{\frac{\ket{00}+\ket{11}}{\sqrt{2}}}} } \bigskip \centerline{ \Qcircuit @C=1em @R=.7em { \lstick{\ket{0}} & \qw & \qw & \qw & \targ & \qw & \multimeasureD{1}{\frac{\ket{00}+\ket{11}}{\sqrt{2}}}\\ \lstick{\ket{0}} & \gate{Y} & \gate{Y} & \gate{H} & \ctrl{-1} & \qw & \ghost{\frac{\ket{00}+\ket{11}}{\sqrt{2}}}} } \caption{2-qubits circuits found by the synthesizer.} \label{fig:2q-cf} \end{figure} The other simulations had as a goal the $GHZ$-states from 3 to 5 qubits. The second simulation found 31 circuits in 5000 episodes. The agent was able to find the minimum circuit showed in Figure \ref{fig:circuit-example}. The third simulation found 7 circuits capable of generating the 4-qubit $GHZ$-state in 20000 episodes. The fourth simulation found only one circuit in 30000 episodes. The 5-qubit circuit showed in Figure \ref{fig:5q-cf}, demonstrates that the agent was able to adapt well to the environment's architecture. Depending on the architecture, some qubits may not be directed connected or have only a one-way connection. This fact limits the possible placements for controlled-gates. Since the example circuit was not allowed by the architecture, the agent had to find another sequence of CNOT operations that could achieve the goal state. \begin{figure}[h] \centerline{ \Qcircuit @C=1em @R=.7em { \lstick{\ket{0}} & \qw & \qw & \qw & \targ & \qw & \qw & \multimeasureD{4}{\frac{\ket{00000}+\ket{11111}}{\sqrt{2}}} \\ \lstick{\ket{0}} & \gate{Z} & \gate{Y} & \gate{Y} & \qw & \targ & \qw & \ghost{\frac{\ket{00000}+\ket{11111}}{\sqrt{2}}} \\ \lstick{\ket{0}} & \qw & \qw & \targ & \ctrl{-2} & \ctrl{-1} & \qw & \ghost{\frac{\ket{00000}+\ket{11111}}{\sqrt{2}}} \\ \lstick{\ket{0}} & \gate{H} & \ctrl{1} & \qw & \qw & \qw & \qw & \ghost{\frac{\ket{00000}+\ket{11111}}{\sqrt{2}}} \\ \lstick{\ket{0}} & \gate{Z} & \targ & \ctrl{-2} & \qw & \qw & \qw & \ghost{\frac{\ket{00000}+\ket{11111}}{\sqrt{2}}} } \hspace{2em} \Qcircuit @C=1em @R=.7em { \lstick{\ket{0}} & \gate{H} & \ctrl{1} & \qw & \qw & \qw & \qw & \multimeasureD{4}{\frac{\ket{00000}+\ket{11111}}{\sqrt{2}}} \\ \lstick{\ket{0}} & \qw & \targ & \ctrl{1} & \qw & \qw & \qw & \ghost{\frac{\ket{00000}+\ket{11111}}{\sqrt{2}}} \\ \lstick{\ket{0}} & \qw & \qw & \targ & \ctrl{1} & \qw & \qw & \ghost{\frac{\ket{00000}+\ket{11111}}{\sqrt{2}}} \\ \lstick{\ket{0}} & \qw & \qw & \qw & \targ & \ctrl{1} & \qw & \ghost{\frac{\ket{00000}+\ket{11111}}{\sqrt{2}}} \\ \lstick{\ket{0}} & \qw & \qw & \qw & \qw & \targ & \qw & \ghost{\frac{\ket{00000}+\ket{11111}}{\sqrt{2}}} } } \caption{5-qubit circuit found by the synthesizer (left). Basic circuit for the 5-qubit GHZ-state (right)} \label{fig:5q-cf} \end{figure} The learning process of the agent required it to first test possible solutions randomly, and as it finds correct answers it starts to learn the right structure of the circuit. As shown by the graphs in Figure \ref{fig:graph-cf}, the agent had good results for 2-qubits circuits and 3-qubits circuits, but as the number of qubits increased it took more time for the agent to find a correct solution to work with. In the gaps of the graphs, where the agent has not found new circuits, the damping parameter was still active. This fact made the reinforcement that the edges got from the successful circuits fade, so that the learning process almost returned to the initial random search. Such behavior is intensified as the search space increase. \begin{figure*}[h!] \centering \includegraphics[width=0.45\columnwidth, keepaspectratio]{images/2-qubits.png} \includegraphics[width=0.45\columnwidth, keepaspectratio]{images/3-qubits.png} \includegraphics[width=0.45\columnwidth, keepaspectratio]{images/4-qubits.png} \includegraphics[width=0.45\columnwidth, keepaspectratio]{images/5-qubits.png} \caption{Circuits found by synthesizer.} \label{fig:graph-cf} \end{figure*} A possible way to overcome the lack of results could be to start the agent with an ECM that already has some information learned to prevent it to search blindly for a first successful circuit. For example, for circuits with a well-defined structure, i.e. W states or GHZ states, the trained ECM of the agents for circuits with less qubits could be loaded and altered with the necessary additional clips. This way, the new agent could start searching by using the previous structure already learned by its predecessor. Also, a deeper analysis could show that a lower damping parameter $\gamma$ and a higher reward would help the agent to learn faster after a few successful results are found. The ideal values for the parameters $\gamma$ and $\eta$ can be learned through the process of meta-learning, as showed in \cite{meta-PS}. \section{Conclusions} \label{sec:conclusion} We proposed the use of reinforcement learning to approach quantum circuit synthesis. We used projective simulation as the model of our work because it presented good results in discovering new quantum experiments, as shown in \cite{Zeilinger:2018}. At this moment, only a subset of functions from projective simulation was implemented, once we intend to verify the usability of this technique in quantum circuit synthesis. To ascertain the viability of our model we ran 4 simulations where our agent had to learn how to generate an entangled state with a different number of qubits. As intended, the model learned to build a quantum circuit that was equivalent to the circuit known in the literature. However, the agent found less solutions as the number of qubits increased. We propose two ways to overcome this problem: the use of better parameters $\gamma$ and $\eta$ and to start the agent with an ECM that already has some information learned. The agent also found alternatives circuits that were different from the basic circuit known in the literature. However, due to the agent's limited action-pool and the simplicity of our model, the alternatives were not efficient, because they had redundant gates. Still, the alternative circuits could be easily minimized by a later optimization process. For future works, there are more functions available in projective simulation that could be implemented in our agent. For example, the action-composition function could help with more complex circuits, as it allows to find subroutines and to create action clips that serve as a shortcut to them. Our agent can also be used to optimize circuits. Since only the perceptions that resulted in a rewarded circuit were added definitely in the agent's memory, it is possible to analyze the strongest edges of the memory to find an optimal circuit. As all the circuits that reached the agent's goal may share a sequence of gates that creates the right answer, this sequence would be rewarded in all circuits. As this sequence is likely to not have any unnecessary gates, finding it may optimize the results found by the agent. We expect reinforcement learning become an important tool in quantum circuit synthesis. The ability to learn new circuit layouts demonstrated that it can be successfully used in order to synthesize more efficient circuits, as seen in \cite{GHZ-W-synth}. Besides that, projective simulation presented here considers the limitations of qubit connectivity when designing a circuit and the quality of quantum gates, which is a factor to still be considered in current research \cite{Mosca_2020}. \section*{Acknowledgements} EID acknowledges the financial support by Brazilian agencies CAPES, CNPq, and INCT-IQ (National Institute of Science and Technology for Quantum Information). \bibliographystyle{plain}

828b9ed4f258c2fa1b7d97d9af7352550d9fd821

\section{Introduction} The class of coherent theories is one that has fairly nice model-theoretic properties (investigated under the name of positive logic), as well as it admits a useful categorical description, hence the toolkit of category theory can also be applied. Moreover this class is suitably general, as any first-order theory can be replaced by a coherent one (with the same class of $\mathbf{Set}$-models). This process is known as Morleyization in the literature. The categorical viewpoint has a long history, the main developments are the syntactic approach (identifying theories and categories, see e.g.~\cite{makkai}), the model theoretic one (characterising the categories of models in terms of accessibility, as it is given in \cite{rosicky}), the geometric one (lifting the syntactic approach to the level of topos theory, discussed in \cite{sheaves}) and the computational one (type theory). Our work belongs to the first topic, some of whose results are summarised in the first section. We will denote the 2-category of coherent categories, coherent functors and all natural transformations by $\mathbf{Coh}$, while $\mathbf{Coh_{\sim }}$ will stand for the (2,1)-category, whose 2-cells are the natural isomorphisms. Our goal is to understand the 2-categorical structure of the latter, in particular to prove \begin{manualtheorem}{4.12} $\mathbf{Coh_{\sim}}$ is 2-complete and 2-cocomplete. \end{manualtheorem} This result, together with some properties of the underlying 1-category will imply a 2-categorical version of the small object argument for $\mathbf{Coh_{\sim}}$, which is hoped to form the first step towards a 2-dimensional model structure on the category of small coherent categories. Finally we make the general remark, that most of our results apply in many other contexts, we restrict our attention to $\mathbf{Coh}$ for simplicity. First of all, everything could be stated for a lower level of logical complexity, which would result the same theorems for lex, regular, etc.~categories, with essentially the same (or simpler) proofs. Secondly, the requirement of strictness in the discussion of $(2,1)$-categories could be omitted, at the price that one has to triangulate each diagrams and resist using the terms "(homotopy) commutative square", "pentagon", etc. \section{Some notions of categorical logic} In this section we summarise the connection between coherent (also called positive) logic and coherent categories. Everything is taken from \cite{makkai}. \begin{definition} A formula is \emph{coherent} if it is built up from atomic formulas using $\wedge , \vee $ and $\exists $. \end{definition} \begin{definition} A formula of the form $\forall x_1 \dots \forall x_n (\varphi \to \psi ) $ is also written as $\varphi \Rightarrow \psi $ and it is called a \emph{sequent}. If $\varphi $ and $\psi $ are coherent formulas then the sequent is said to be coherent. \end{definition} \begin{definition} A \emph{coherent theory} is a set of coherent sequents. \end{definition} \begin{definition} Given a signature $L=\langle S,\mathbf{R},\mathbf{F} \rangle$, where $S$ is the set of sorts, $\mathbf{R}$ is the set of relation symbols (i.e.~has elements of the form $R\subseteq s_1\times \dots \times s_n$ with $s_i$'s being the related sorts), and $\mathbf{F}$ is the set of function symbols (e.g.~$f:s_1\times \dots \times s_n \to s$), an \emph{$L$-structure} in a category $\mathcal{C}$ associates to each sort $s$ an object $M(s)$ of $\mathcal{C}$, to each relation symbol $R$ a subobject $M(R) \leq M(s_1)\times \dots \times M(s_n)$, and to a function symbol $f$ a morphism $M(f):M(s_1)\times \dots \times M(s_n)\to M(s)$. \end{definition} Our next goal is to interpret first-order formulas in structures (inside a fixed category $\mathcal{C}$). Depending on the complexity of our formula, this will require some extra structure on $\mathcal{C}$. \begin{definition} Let $M$ be an $L$-structure in a category $\mathcal{C}$. The \emph{interpretation} of a formula is given by the following steps: \begin{itemize} \item If $\vec{x}=(x_1,\dots x_n)$ is a finite sequence of free variables, $x_i$ is of sort $s_i$, then $M(\vec{x})=M(s_1)\times \dots \times M(s_n)$. \item If $t$ is a term (of sort $s$) whose free variables are from $\vec{x}$, then $M_{\vec{x}}(t)$ will be an arrow $M(\vec{x})\to M(s)$ in the following way: \begin{itemize} \item If $t=x_i$, then $M_{\vec{x}}(t)$ is the projection map $M(\vec{x})\to M(x_i)=M(s)$. \item If $t=f(t_1,\dots t_n)$, then $M_{\vec{x}}(t)$ is the composite $M(\vec{x})\xrightarrow {\langle M_{\vec{x}}(t_1),\dots \rangle} \prod M(s_i) \xrightarrow{M(f)} M(s)$ \end{itemize} When $\mathcal{C}=\mathbf{Set}$, these are the functions which for a possible evaluation of $\vec{x}$ assign the induced value of $t$. \item If $\varphi $ is a formula, whose free variables are along $\vec{x}=(x_1,\dots x_n)$, then its interpretation in the context $\vec{x}$ (if it exists) is a subobject $M_{\vec{x}}(\varphi )\leq M(\vec{x})$. It should be readily checked that in the case of $\mathbf{Set}$-models this gives precisely the set of evaluations of $\vec{x}$ which make $\varphi $ valid in $M$. \begin{itemize} \item \begin{tikzpicture} \node (1) at (0,0) {$M_{\vec{x}}(t_1\approx t_2) $}; \node[anchor=base,right=10mm of 1] (2) {$M(\vec{x})$}; \node[anchor=base,right=10mm of 2] (3) {$M(s)$}; \draw[right hook ->] (1)--(2) node [midway,above] {\small{$e$}}; \draw[postaction={transform canvas={yshift=-1mm},draw}] [->] (2) -- (3) node [midway,above] {\small{$M_{\vec{x}}(t_1)$}} node [midway,below] {\small{$M_{\vec{x}}(t_2)$}}; \end{tikzpicture} is an equalizer. \item \begin{tikzpicture} \node (3) at (0,0) {$M_{\vec{x}}(R(t_1,\dots t_n))$}; \node[anchor=base,right=10mm of 3] (2a) {$M(\vec{x})$}; \node[anchor=base,below=10mm of 3] (2b) {$M(R)$}; \node[anchor=base,below=10mm of 2a] (1) {$\prod _{i=1}^n M(s_i)$}; \draw[right hook->] (3)--(2a); \draw[->] (3)--(2b); \draw[->] (2a)--(1) node [midway,right] {$\langle M_{\vec{x}}(t_1),\dots \rangle$}; \draw[right hook->] (2b)--(1) node [midway,below] {$M(i)$}; \end{tikzpicture} is a pullback (with $M(i)$ being the subobject $M(R)\hookrightarrow \prod _{i=1}^n M(S_i)$). \item $M_{\vec{x}}(\bigwedge \Theta)=\bigwedge\{M_{\vec{x}}(\theta):\theta\in \Theta\}$ is the infimum (pullback) of the subobjects $M_{\vec{x}}(\theta)$. \item $M_{\vec{x}}(\bigvee \Theta)=\bigvee\{M_{\vec{x}}(\theta):\theta\in \Theta\}$ is the supremum of the subobjects $M_{\vec{x}}(\theta)$. \item $M_{\vec{x}}(\exists y \varphi)$ (where $y$ is not in $\vec{x}$) is the surjective-mono factorisation: \begin{tikzpicture} \node (3) at (0,0) {$M_{\vec{x},y}(\varphi)$}; \node[anchor=base,right=10mm of 3] (2a) {$M(\vec{x},y)$}; \node[anchor=base,right=10mm of 2a] (2b) {$M(\vec{x})$}; \node[anchor=base,below=10mm of 2a] (1) {$M_{\vec{x}}(\exists y \varphi)$}; \draw[right hook->] (3)--(2a); \draw[->] (2a)--(2b) node [midway,above] {$\pi _{\vec{x}}$}; \draw[->>] (3)--(1); \draw[right hook->] (1)--(2b); \end{tikzpicture} (An arrow $f:A\to B$ of $\mathcal{C}$ is surjective iff whenever it factors through a subobject $i:B'\hookrightarrow B$, we get that $i$ is an isomorphism.) \item $M_{\vec{x}}(\neg \varphi)$ is the biggest (i.e.~contains every other such) subobject $A$ of $M(\vec{x})$, such that $A\wedge M_{\vec{x}}(\varphi) \leq 0_{M(\vec{x})}$, where $0_{M(\vec{x})}$ is the smallest subobject of $M(\vec{x})$. \item $M_{\vec{x}}(\varphi\to \psi )$ is the biggest subobject $A$ of $M(\vec{x})$ such that $A\wedge M_{\vec{x}}(\varphi) \leq M_{\vec{x}}(\psi )$. \item $M_{\vec{x}}(\forall y \varphi)$ is the biggest subobject $A$ of $M(\vec{x})$ such that $\pi _{\vec{x}}^{-1}(A)\leq M_{\vec{x},y}(\varphi )$. $\pi _{\vec{x}}^{-1}(A)$ denotes the pullback of $A$ along the projection $\pi _{\vec{x}}:M(\vec{x},y)\to M(\vec{x})$ \end{itemize} \end{itemize} \label{intpret} \end{definition} \begin{definition} The sequent $\varphi \Rightarrow \psi$ is \emph{valid} in the structure $M$ (in symbols: $M\models \varphi \Rightarrow \psi $), iff $ M_{\vec{x}}(\varphi ) \leq M_{\vec{x}}(\psi ) $ (where $\vec{x}$ is the collection of all free variables in $\varphi \Rightarrow \psi $). $M$ is a \emph{model} of the theory $T$ iff all the sequents from $T$ (have interpretation and) are valid in $M$. A \emph{homomorphism} $\alpha :M\to M'$ of $T$-models consists of an arrow $\alpha _s: M(s)\to M'(s)$ for each sort $s$, for which the square \[\begin{tikzcd} {M(s_1)\times \dots \times M(s_n)} && {M(s)} \\ \\ {M'(s_1)\times \dots \times M'(s_n)} && {M'(s)} \arrow["{M(f)}", from=1-1, to=1-3] \arrow["{\alpha _{s_1} \times \dots \times \alpha _{s_n}}"', from=1-1, to=3-1] \arrow["{\alpha _s}", from=1-3, to=3-3] \arrow["{M'(f)}", from=3-1, to=3-3] \end{tikzcd}\] commutes and the dashed arrow in \[\begin{tikzcd} {M(R)} && {M(s_1)\times \dots \times M(s_n)} \\ {\alpha (M(R))} \\ {M'(R)} && {M'(s_1)\times \dots \times M'(s_n)} \arrow[hook, from=1-1, to=1-3] \arrow["{\alpha _{s_1} \times \dots \times \alpha _{s_n}}", from=1-3, to=3-3] \arrow[hook, from=3-1, to=3-3] \arrow[two heads, from=1-1, to=2-1] \arrow[hook, from=2-1, to=3-3] \arrow[dashed, from=2-1, to=3-1] \end{tikzcd}\] exists. The category of $T$-models and homomorphisms in a category $\mathcal{C}$ is denoted by $T$-$mod(\mathcal{C})$. \label{modelincat} \end{definition} \begin{remark} For the interpretation of coherent logic, it is enough to assume that $\mathcal{C}$ has finite limits, finite sups and surjective-mono factorisation. \end{remark} This observation motivates the following definitions: \begin{definition} An arrow $f:X\to Y$ in a category $\mathcal{C}$ is an \emph{effective epimorphism} if the pullback \begin{center} \begin{tikzpicture} \node (3) at (0,0) {$X\times _{Y} X$}; \node[anchor=base,right=10mm of 3] (2a) {$X$}; \node[anchor=base,below=10mm of 3] (2b) {$X$}; \node[anchor=base,below=10mm of 2a] (1) {$Y$}; \draw[->] (3)--(2a) node [midway,above] {$\pi$}; \draw[->] (3)--(2b) node [midway,left] {$\pi '$}; \draw[->] (2a)--(1) node [midway,right] {$f$}; \draw[->] (2b)--(1) node [midway,below] {$f$}; \end{tikzpicture} \end{center} exists and \begin{center} \begin{tikzpicture} \node (1) at (0,0) {$X\times _{Y} X$}; \node[anchor=base,right=10mm of 1] (2) {$X$}; \node[anchor=base,right=10mm of 2] (3) {$Y$}; \draw[->] (2)--(3) node [midway,above] {\small{$f$}}; \draw[postaction={transform canvas={yshift=-1mm},draw}] [->] (1) -- (2) node [midway,above] {\small{$\pi$}} node [midway,below] {\small{$\pi '$}}; \end{tikzpicture} \end{center} is a coequalizer. \label{effepi} \end{definition} \begin{definition} A \emph{category} $\mathcal{C}$ is \emph{coherent}, if it \begin{itemize} \item has finite limits, \item has images, i.e.~every morphism can be factored as an effective epimorphism followed by a monomorphism, \item for any object $X$, the poset of its subobjects $Sub(X)$ is a lattice, \item effective epimorphisms are pullback-stable, \item for any map $f:X\to Y$, the induced map $f^{-1}:Sub(Y)\to Sub(X)$ is a lattice homomorphism. \end{itemize} A \emph{functor} $F:\mathcal{C}\to \mathcal{D}$ is \emph{coherent} if it \begin{itemize} \item preserves finite limits, \item preserves effective epimorphisms, \item the induced map $Sub(X)\to Sub(F(X))$ is a lattice homomorphism. \end{itemize} The 2-category of small coherent categories, coherent functors and all natural transformations is denoted by $\mathbf{Coh}$. \end{definition} \begin{remark} As $F$ preserves finite limits, it follows that monomorphisms are also preserved (since $f:X\to Y$ is mono iff \begin{center} \begin{tikzpicture} \node (3) at (0,0) {$X$}; \node[anchor=base,right=10mm of 3] (2a) {$X$}; \node[anchor=base,below=10mm of 3] (2b) {$X$}; \node[anchor=base,below=10mm of 2a] (1) {$Y$}; \draw[->] (3)--(2a) node [midway,above] {$1_X$}; \draw[->] (3)--(2b) node [midway,left] {$1_X$}; \draw[->] (2a)--(1) node [midway,right] {$f$}; \draw[->] (2b)--(1) node [midway,below] {$f$}; \end{tikzpicture} \end{center} is a pullback), so the map $Sub(X)\to Sub(F(X))$ makes sense. \end{remark} \begin{remark} An arrow $f:X\to Y$ in a coherent category is surjective iff it is an effective epimorphism. \end{remark} The first part of the proposed correspondence is to replace categories with theories: \begin{definition} The \emph{canonical language} of the category $\mathcal{C}$ has the signature $L=L_{\mathcal{C}}$ which contains a sort $\bar{A}$ for every object $A$ of $\mathcal{C}$, and a function symbol $\bar{f}:\bar{A}\to \bar{B}$ for every such arrow $f$ of $\mathcal{C}$ (and nothing else). Then $\mathcal{C}$ is naturally an $L$-structure by the identical interpretation of $L$ (i.e.~sending $\bar{A}$ to $A$ and $\bar{f}$ to $f$). More generally; each functor $F:\mathcal{C}\to \mathcal{D}$ creates an $L$-structure in $\mathcal{D}$. \end{definition} The following theorem (2.4.5. in \cite{makkai}) says, that from inside, $\mathcal{C}$ looks similar to $\mathbf{Set}$. \begin{theorem} Assume, that $\mathcal{C}$ has finite limits. Then the following diagrams in $\mathcal{C}$ have the stated properties, iff the sequents on their right side (have interpretation and) are valid (in $\mathcal{C}$, as a structure over its canonical language, with the identical interpretation for the signature). \begin{tabularx}{\textwidth}{cXX} 1. & $A\xrightarrow{f} A$ is the identity on $A$ & $\Rightarrow f(a)\approx a$ \\ \hline 2. & \begin{tikzpicture} \node (3) at (0,0) {$A$}; \node[anchor=base,right=5mm of 3] (2a) {$C$}; \node[anchor=base,below=5mm of 3] (2b) {$B$}; \draw[->] (3)--(2a) node [midway,above] {$h$}; \draw[->] (3)--(2b) node [midway,left] {$f$}; \draw[->] (2b)--(2a) node [midway,below] {$g$}; \end{tikzpicture} is commutative & $\Rightarrow gf(a)\approx h(a)$\\ \hline 3. & $A\xrightarrow{f} B$ is mono & $f(a)\approx f(a') \Rightarrow a\approx a'$ \\ \hline 4. & $A\xrightarrow{f} B$ is surjective & $\Rightarrow \exists a:f(a)\approx b$ \\ \hline 5. & $A$ is the terminal object & $\Rightarrow a\approx a'$ \\ & & $\Rightarrow \exists a: a\approx a$ \\ \hline 6. & $A$ is the initial object & $a\approx a \Rightarrow $ \\ \hline 7. & $A\xleftarrow{f} C\xrightarrow{g} B$ is a product diagram & \small{$f(c)\approx f(c') \wedge g(c)\approx g(c') \Rightarrow c\approx c'$} \\ & & $\Rightarrow \exists c (f(c)\approx a \wedge g(c)\approx b) $ \\ \hline 8. & \begin{tikzpicture} \node (1) at (0,0) {$E$}; \node[anchor=base,right=5mm of 1] (2) {$A$}; \node[anchor=base,right=5mm of 2] (3) {$B$}; \draw[right hook->] (1)--(2) node [midway,above] {$\epsilon$}; \draw[postaction={transform canvas={yshift=-1mm},draw}] [->] (2) -- (3) node [midway,above] {$f$} node [midway,below] {$g$}; \end{tikzpicture} is an equalizer & $f(a)\approx g(a) \Leftrightarrow \exists e: \epsilon (e)\approx a$ \\ \hline 9. & \makecell[l]{$B\xhookrightarrow{g} X\xhookleftarrow{f_i} A_i$ ($i\in I$). \\ $B$ is the sup of $A_i$-s} & \small{$\bigvee _{i\in I} \exists a_i: f_i(a_i)\approx x \Leftrightarrow \exists b: g(b)\approx x$}\\ \hline 10. & \makecell[l]{$B\xhookrightarrow{g} X\xhookleftarrow{f_i} A_i$ ($i\in I$). \\ $B$ is the inf of $A_i$-s} & \small{$\bigwedge _{i\in I} \exists a_i: f_i(a_i)\approx x \Leftrightarrow \exists b: g(b)\approx x$} \\ \end{tabularx} \label{diagram} \end{theorem} \begin{definition} Let $\mathcal{C}$ be a coherent category. Its (coherent) \emph{internal theory} $T_{\mathcal{C}}$ (or $Th(\mathcal{C})$) over the signature $L_{\mathcal{C}}$ consists of those sequents which refer to identities, commutative triangles, finite limits, surjective arrows and finite unions (as it is described above). \end{definition} \begin{theorem} The categories $T_{\mathcal{C}}$-$mod(\mathcal{E})$ and $\mathbf{Coh}(\mathcal{C},\mathcal{E})$ are isomorphic. \end{theorem} Now we replace theories with categories. The notion of derivability ($\vdash $) refers to a deduction system which is sound wrt.~every coherent category and which is complete wrt.~Boolean-valued $\mathbf{Set}$-models, see \cite{makkai} for the details. \begin{definition} Let $T$ be a coherent theory. Its \emph{syntactic category} $\mathcal{C}_T$ is defined as follows: \begin{itemize} \item The objects are equivalence classes of coherent formulas (in context) over the given signature, where $\varphi (\vec{x})\sim \psi (\vec{y})$, iff $\psi (\vec{y})=\varphi (\vec{y}/ \vec{x})$. Note that $[\varphi (\vec{x})]$ and $[\varphi (\vec{x},\vec{y})]$ (with $y$ being an extra variable not present in $\varphi$) corresponds to different objects. This technicality is not essential, as $[\varphi (\vec{x},y)]$ turns out to be isomorphic with $[\varphi (\vec{x}) \wedge y\approx y]$, hence if we require all variables in the context $\vec{x}$ to appear freely in $\varphi $ we get an equivalent category. \item An arrow $[\varphi (\vec{x})]\xrightarrow{[\theta (\vec{x},\vec{y})]} [\psi (\vec{y})]$ is an equivalence class of formulas, having the following properties: \begin{itemize} \item $\vec{x}$ and $\vec{y}$ are disjoint (this can always be assumed, as we can find such representatives of the objects), \item $T\vdash \theta(\vec{x},\vec{y}) \Rightarrow \varphi (\vec{x}) \wedge \psi (\vec{y})$, \item $T\vdash \varphi (\vec{x}) \Rightarrow \exists \vec{y} \theta(\vec{x},\vec{y})$, \item $T\vdash \theta(\vec{x},\vec{y}) \wedge \theta(\vec{x},\vec{y'}) \Rightarrow \vec{y}=\vec{y'}$. \end{itemize} $\theta(\vec{x},\vec{y}) \sim \theta'(\vec{x'},\vec{y'})$, iff $T\vdash \theta(\vec{x},\vec{y}) \Leftrightarrow \theta'(\vec{x},\vec{y})$. \end{itemize} \end{definition} \begin{remark} The required properties for $\theta $ are often referred as being "T-provably functional". This is because these are exactly the conditions which can guarantee, that the interpretation of $\theta $ in a model $M$ is not merely a subobject of $M(\varphi )\times M(\psi )\leq M(\vec{x})\times M(\vec{y})$, but the graph of an arrow from $M(\varphi )$ to $M(\psi )$. \end{remark} \begin{theorem} Given a coherent theory $T$, its syntactic category $\mathcal{C}_T$ is a well-defined coherent category. The categories $T$-$mod(\mathcal{E})$ and $\mathbf{Coh}(\mathcal{C}_T,\mathcal{E})$ are equivalent. If $\mathcal{C}$ is coherent then $\mathcal{C}_{Th(\mathcal{C})}$ and $\mathcal{C}$ are equivalent. \end{theorem} \section{Limits and weak colimits} The main ingredient for the existence of weak colimits is the following fact: \begin{proposition} Given an arbitrary subcategory $i:\mathcal{C}\hookrightarrow \mathcal{E}$ of a coherent category $\mathcal{E}$, it is included in a coherent subcategory $\tilde{\mathcal{C}}\hookrightarrow \mathcal{E}$ (with coherent embedding), such that $|\tilde{\mathcal{C}}|\leq \aleph _0 \cdot |\mathcal{C}|$. (Let's say $|\mathcal{C}|$ is defined to be $|Arr(\mathcal{C})|$.) \label{solution} \end{proposition} \begin{proof} Form the theory $T$ over the signature $L_{\mathcal{C}}$ which consists of the sequents for identities and commutative triangles. Then $i$ corresponds to a $T$-model in $\mathcal{E}$, which induces a coherent functor $\mathcal{C}_T\to \mathcal{E}$, whose image has a size smaller or equal to $\aleph _0 \cdot |L_{\mathcal{C}}|=\aleph _0\cdot |\mathcal{C}|$ and contains $\mathcal{C}$ as a subcategory. \end{proof} We start with the existence of limits. \begin{proposition} $\mathbf{Coh}$ has pullbacks along isofibrations, and the forgetful functor $U:\mathbf{Coh}\to \mathbf{Cat}$ preserves them. \label{cohpb} \end{proposition} \begin{lemma} Take a diagram of the form $\mathcal{C}\xrightarrow{F}\mathcal{D}\xleftarrow{F'}\mathcal{C'}$ in $\mathbf{Cat}$, where either $F$ or $F'$ is an isofibration. If all three categories have certain types of limits or colimits, and these are preserved by both functors, then $\mathcal{C}\times _{\mathcal{D}}\mathcal{C'}$ will also admit these (co)limits, and they will be preserved by the projection maps. Moreover they are reflected by the pair of the projections. \label{pblim} \end{lemma} \begin{proof} Take a diagram of that fixed type (i.e.~a functor from the fixed index category) in $\mathcal{C}\times _{\mathcal{D}}\mathcal{C'}$. It is sent to the same type of diagrams in $\mathcal{C}$ and $\mathcal{C'}$, so by assumption, we can take a (co)limiting (co)cone over them, which are preserved by $F$ and $F'$. As (co)limits are unique up to unique isomorphism, there is an isomorphism $(i)$ in $\mathcal{D}$, that connects the tips of the images and makes everything commute. As (let's say) $F$ was an isofibration, we can modify our (co)cone in $\mathcal{C}$ by the composition of a preimage of $i$, hence we can assume, that the images at $F$ and at $F'$ are the same. By the pullback-construction, there is a (co)cone in $\mathcal{C}\times _{\mathcal{D}}\mathcal{C'}$, that is mapped to the chosen (co)limiting ones by the projections. A similar argument shows that it must be a (co)limiting one, the preservation (and the reflection by the pair) follows from the construction. \end{proof} \begin{lemma} The coequalizer of kernel pairs (as in Definition \ref{effepi}) exists in every coherent category, and it is preserved by every coherent functor. \label{coeq} \end{lemma} \begin{proof} Using the notation of Definition \ref{effepi}, factor $f$ as $X\xrightarrow{e} \bar{X} \xrightarrow{i} Y$, where $e$ is an effective epi and $i$ is mono. It follows that \begin{center} \begin{tikzpicture} \node (3) at (0,0) {$X\times _Y X$}; \node[anchor=base,right=10mm of 3] (2a) {$X$}; \node[anchor=base,below=10mm of 3] (2b) {$X$}; \node[anchor=base,below=10mm of 2a] (1) {$\bar{X}$}; \draw[->] (3)--(2a) node [midway,above] {$\pi$}; \draw[->] (3)--(2b) node [midway,left] {$\pi '$}; \draw[->] (2a)--(1) node [midway,right] {$e$}; \draw[->] (2b)--(1) node [midway,below] {$e$}; \end{tikzpicture} \end{center} is also a pullback, hence $e=coeq(\pi ,\pi ')$, as by definition $e$ is the coequalizer of its kernel pair. This also proves the uniqueness (up to unique isomorphism) of such factorisations. \end{proof} \begin{proof}[Proof of Proposition \ref{cohpb}] We will use the notation of Lemma \ref{pblim}, and check that $\mathcal{C}\times _{\mathcal{D}} \mathcal{C'}$ is coherent: \begin{itemize} \item finite limits: Lemma \ref{pblim} \item images: Using that $F$ was an isofibration, we get a factorisation in \\ $\mathcal{C}\times _{\mathcal{D}}\mathcal{C'}$ from the factorisations in $\mathcal{C}$ and in $\mathcal{C'}$. By the previous lemmas, the projections $\pi $ and $\pi '$ preserve, and together reflect effective epimorphisms and monomorphisms. \item joins: By the existence of finite limits, it is enough to see that finite (possibly empty) joins of subobjects exist. This follows from a similar argument to the one in the proof of Lemma \ref{pblim}. \item pullback-stability: $\pi $ and $\pi '$ preserve, and together reflect pullbacks. \end{itemize} \end{proof} \begin{corollary} $\mathbf{Coh}$ has finite products. \label{finprod} \end{corollary} \begin{proof} The unique map to the terminal object is an isofibration. \end{proof} \begin{remark} Since limits, subobjects, unions, composition of arrows, etc. are defined coordinate-wise, it is equally easy to see the existence of arbitrary products. \end{remark} \begin{theorem} $\mathbf{Coh}$ has weak colimits. \label{complete} \end{theorem} \begin{proof} Fix a diagram $d_{\bullet }:\mathcal{I}\to \mathbf{Coh}$, and let $F:\mathbf{Coh}\to \mathbf{Set}$ be the functor $lim _{\mathcal{I}^{op}} \mathbf{Coh}(d_i, -)$. The existence of a weak colimit is equivalent to the existence of a weak initial object in the category $*\downarrow F$ (see: \cite{maclane} Theorem V.6.3.). As $\mathbf{Coh}$ has products and $F$ preserves them, $*\downarrow F$ has products, so it is enough to find a weakly initial family, as in this case the product of its elements is a weak initial object. An element of $lim _{\mathcal{I}^{op}} \mathbf{Coh}(d_i, C)$ is a compatible family of functors $\{f_i:d_i\to C\}$. To prove the existence of a weakly initial family, set $C'$ to be the subcategory with objects $\bigcup Ob(Im\ (f_i))$ and with arrows the composites of arrows from $\bigcup Arr(Im\ (f_i))$. The size of $C'$ is bounded by the size of $\sum |d_i|$, but it is not necessarily coherent. Using Corollary \ref{solution}, the inclusion $C'\to C$ factors through a coherent subcategory $g: \tilde{C}\to C$, where the size of $\tilde{C}$ is still limited by $\sum |d_i|$. Now it follows that each $f_i$ factors as $g\circ \tilde{f_i}$, and in this case $\tilde{f_i}:d_i\to \tilde{C}$ must also be coherent. This shows that for some fixed $\kappa \geq \sum |d_i|$ the set of coherent categories of cardinality $\leq \kappa$ (one from each isomorphism class) is a solution set, namely the set of all cocones (over the fixed diagram) with top element having cardinality $\leq \kappa$ is a weakly initial family. \end{proof} \section{Filtered colimits} First recall the following basic result on the construction of general colimits (cf.~\cite{maclane} Theorem V.2.1.). \begin{theorem} Let $d_\bullet :\mathcal{I}\to \mathcal{C}$ be a diagram, where $\mathcal{C}$ is a cocomplete category. Its colimit can be computed as the coequalizer \begin{center} \begin{tikzpicture} \node (1) at (0,0) {$\bigsqcup _{f\in \mathcal{I}_1} dom(f)$}; \node[anchor=base,right=15mm of 1] (2) {$\bigsqcup _{A\in \mathcal{I}_0} A$}; \node[anchor=base,right=10mm of 2] (3) {$colim\ d_\bullet $}; \draw[->] (2)--(3) node [midway,above] {\small{$r$}}; \draw[postaction={transform canvas={yshift=-1mm},draw}] [->] (1) -- (2) node [midway,above] { \small{$\sqcup f$}} node [midway,below] { \small{$\sqcup 1_{dom(f)}$}}; \end{tikzpicture} \end{center} where the edges of the cocone are given by $A\xrightarrow{j_A} \bigsqcup _{A\in \mathcal{I}_0} A \xrightarrow{r} colim\ d_\bullet $. \end{theorem} As in the category $\mathbf{Set}$ we have a good understanding of coproducts (disjoint union) and coequalizers (factorisation by the equivalence relation generated by the pairs $(f(x),g(x))$), we can describe filtered colimits explicitly: \begin{theorem} Let $d_\bullet :\mathcal{I}\to \mathbf{Set}$ be a diagram, where $\mathcal{I}$ is a filtered category. Its colimit is the set $\faktor{\{(x,i):x\in d_i \}}{\sim}$, where $(x,i)\sim (y,j)$ iff $\exists \varphi _{ik}:i\to k $ $ \exists \varphi _{jk}:j\to k \ d(f)(x)=d(g)(y)$. The $i$-th coprojection is given by $d_i\ni x\mapsto [(x,i)]$. \end{theorem} The proof consists of the simple observation that the axioms of filtered categories are precisely the ones that force $\sim $ to be an equivalence relation. (In what follows we will not distinguish between objects and arrows in $\mathcal{I}$ and their image at $d_\bullet $.) Now we claim that the situation for $\mathbf{Cat}$ is essentially the same, in other words: \begin{proposition} The forgetful functor $V:\mathbf{Cat}\to \mathbf{Set}$ (which assigns to every category the set of its arrows) preserves filtered colimits. The identity arrow on an object $A$ of $colim\ d_\bullet$ is represented by $(1_A,i)$, where $A$ is in the image of the $i$-th coprojection. Similarly, the composition of $(f,i):(A,i)\to (B,i)$ and $(g,j):(B,i)=(C,j)\to (D,j)$ is $(\varphi _{jk} (g)\circ \varphi _{ik}(f), k)$, where $\varphi _{i/j,k}$ is a map $\mathcal{C}_{i/j}\to \mathcal{C}_k$ in the diagram, such that $\varphi _{ik}(B) =\varphi _{jk}(C)$. \label{catset} \end{proposition} \begin{proof} The above construction of identities and composition is well-defined, since each function in the diagram is a functor (hence we can choose any suitable $k$). Each coprojection map is also a functor. Given a cocone $(\mu _i)_{i\in \mathcal{I}_0}$ with all edges being functors, the induced (unique) function must also be a functor: any composable pair exists at some stage $\mathcal{C}_i$, hence it is mapped to the composition by $\mu _i$, and also by the induced universal map by commutativity. The same holds for identities. \end{proof} The final step is to prove \begin{theorem} $\mathbf{Coh}$ has filtered colimits and the forgetful functor $U:\mathbf{Coh}\to \mathbf{Cat}$ preserves them. \label{cohcat} \end{theorem} \begin{proof} First we have to prove that if all $\mathcal{C}_i$'s and all $\varphi _{ij}$'s are coherent then so is the colimit, with coherent coprojections. \begin{itemize} \item finite limits: Given a finite diagram $\Delta \xrightarrow{\delta _\bullet} colim\ d_\bullet $, it factors through some $\mathcal{C}_i$. One can take the limit cone here, whose image at $\mu _i$ will be the limit of $\delta _\bullet $. I.e.~given another cone with top object $[(y,j)]$ (assuming the whole diagram with the two cones factors through $\mathcal{C}_j$), we can find an index $k$ and maps $\varphi _{ik}$, $\varphi _{jk}$ such that the image of $\delta _\bullet $ at $\mathcal{C}_i$ and at $\mathcal{C}_j$ is identified by them. Since $\varphi _{ik}$ preserves finite limits, we have an induced map $!:(x,k)\to (y,k)$, whose image at $\mu _k$ makes everything commute. Uniqueness is proved similarly. It follows that $\mu _i$'s preserve finite limits. \item image factorisation: By the previous argument monomorphisms are preserved by the coprojection maps and the image of the kernel pair of $f$ at $\mu _i$ is the kernel pair of $\mu _i(f)$. Since kernel pairs have coequalizers in all $\mathcal{C}_i$'s, and they are preserved by $\varphi _{ij}$'s, the above argument shows that they are also preserved by the coprojections, hence effective epimorphisms are preserved. Given an arrow in the colimit one can take any of its preimages, and by the previous argument its factorisation will be mapped to an image factorisation in $colim\ d_\bullet$. \item unions: Take an object $[(x,i)]$ in $colim\ d_\bullet$ and two monomorphism $[(m_1,i)]$, $[(m_2,i)]$ into it (again: we can assume that these $i$'s are the same). In $\mathcal{C}_i$ we see the same diagram, but we can not guarantee that $m_1$ and $m_2$ are monos. Instead we can factor them into an effective epi followed by a mono, and by the uniqueness of these factorisations (up to unique isomorphism), we see that the monic part will go to the same subobject which was represented by $[(m_1,i)]$ (resp.~$[(m_2,i)]$). Now we can take the union in $\mathcal{C}_i$, and check that its image at $\mu _i$ is the union in the colimit. \item pullback-stability: Every effective epimorphism (resp.~pullback square) comes from an effective epi (resp.~pullback) in some $\mathcal{C}_i$, hence we are done. The same works for unions. \end{itemize} \end{proof} Our next goal is to prove that for every coherent category $\mathcal{C}$, the hom-functor $\mathbf{Coh}(\mathcal{C},-)$ commutes with $|\mathcal{C}|^+$-filtered colimits. This is known for the category $\mathbf{Set}$: \begin{proposition} For any set $A$ the functor $\mathbf{Set}(A,-)$ preserves $|A|^+$-filtered colimits. \end{proposition} \begin{proof} The induced map $!: colim\ \mathbf{Set}(A,d_i) \to \mathbf{Set}(A,colim\ d_\bullet)$ has as domain the set of equivalence classes $[(f:A\to d_i)]$, where $f:A\to d_i$ is equivalent to $g:A\to d_j$ if there are maps $\varphi _{ik}$, $\varphi _{jk}$ in the image of $d_\bullet$, such that $\varphi _{ik}f=\varphi _{jk}g$. $!$ takes $[(f:A\to d_i)]$ to $A\xrightarrow{f} d_i \xrightarrow{\mu _i} colim\ d_\bullet$, and by the construction of an $|A|^+$-filtered colimit, it is automatically injective (i.e.~for each element $a$ there is a suitable $k$, and maps $\varphi _{ik}$, $\varphi _{jk}$, such that $\varphi _{ik}f(a)=\varphi _{jk}g(a)$, and these have a common upper-bound). Given a function $f:A\to colim\ d_\bullet$, each $a\in A$ is included in the image of some $\mu _i$, by $|A|^+$-filteredness these $d_i$-s have a common extension $d_k$, hence $f$ factors through $A_k$. Therefore $!$ is also surjective. \end{proof} Now it follows easily for $\mathbf{Cat}$: \begin{proposition} For any category $\mathcal{C}$ the functor $\mathbf{Cat}(\mathcal{C},-)$ preserves $|\mathcal{C}|^+$-filtered colimits. \end{proposition} \begin{proof} By Proposition \ref{catset} we can compute filtered colimits in $\mathbf{Set}$. If the map $!:colim\ \mathbf{Cat}(\mathcal{C},d_i) \to \mathbf{Cat}(\mathcal{C},colim\ d_\bullet)$ would take two elements to the same functor, then $!':colim\ \mathbf{Set}(V(\mathcal{C}),V(d_i)) \to \mathbf{Set}(V(\mathcal{C}),colim\ V\circ d_\bullet)$ would not be injective. Given a functor $F:\mathcal{C}\to colim\ d_\bullet$, it factors through some $d_i$ as a(n arrow) function (e.g.~$F=\mathcal{C}\xrightarrow{F'}d_i \xrightarrow{\mu _i} colim\ d_\bullet$). Now we just count how many things can go wrong. For each commutative triangle and each identity arrow it might happen that $F'$ does not preserve it, but in this case there is an index $k$ and an arrow $\varphi _{ik}$ in the diagram, such that $\varphi _{ik}F'$ corrects that mistake. Since there are at most $|\mathcal{C}|$-many commutative triangles and identities in $\mathcal{C}$ (or finitely many if $\mathcal{C}$ is finite), and our diagram is $|\mathcal{C}|^+$-filtered (or $\aleph _0$-filtered in the finite case), we are done. \end{proof} This shows how to proceed when the category $\mathbf{Coh}$ is considered: \begin{proposition} For any coherent category $\mathcal{C}$ the functor $\mathbf{Coh}(\mathcal{C},-)$ preserves $|\mathcal{C}|^+$-filtered colimits. \label{filtered} \end{proposition} \begin{proof} Again, $!$ is injective and given a coherent functor $F:\mathcal{C}\to colim\ d_\bullet$, it factors through some $d_i$ as $\mathcal{C}\xrightarrow{F'}d_i \xrightarrow{\mu _i} colim\ d_\bullet$, where $F'$ is a (not necessarily coherent) functor. That is, for each finite diagram, effective epimorphism, and pair of subobjects it might happen, that $F'$ does not preserve the limit, the effective epi and the union, but some $\varphi _{ij}$ corrigates one of these mistakes. Since there are at most $ |\mathcal{C}|$-many (or finitely many) such diagrams in $\mathcal{C}$, we can find a $\varphi_{ik}$ for which $\varphi _{ik}F'$ is coherent. \end{proof} \begin{theorem} $\mathbf{Coh}$ is $\aleph _1$-accessible. \end{theorem} \begin{proof} Proposition \ref{filtered} shows that every coherent category $\mathcal{C}$ is $|\mathcal{C}|$-presentable, and Proposition \ref{solution} implies that every coherent category is the union of its countable subcategories. Clearly the poset of countable subcategories is $\aleph _1$-filtered since $\aleph _0 \cdot \aleph _0 =\aleph _0$. \end{proof} \section{2-categorical aspects} If an accessible category has all limits then it is cocomplete (see Corollary 2.47. of \cite{rosicky}). Hence if $\mathbf{Coh}$ had pullbacks we could derive that $\mathbf{Coh}$ is complete and cocomplete. This is not the case since e.g. the two inclusions $* \hookrightarrow (*\leftrightarrow *)$ have empty intersection and the map from the 2-element Boolean algebra $\mathbf{2}$ to a coherent category is unique only up to unique natural isomorphism. However these examples show that we can hope for completeness and cocompleteness in a 2-categorical sense. In this section we will show that this idea is right. \begin{definition} Given a small 2-diagram $d_\bullet :\mathcal{I}\to \mathcal{C}$ where $\mathcal{C}$ is an arbitrary (2,1)-category, its \emph{homotopy} (or 2-) \emph{limit} is a cone $ $\\ \adjustbox{scale=0.8,center}{ \begin{tikzcd} &&& d \\ \\ \\ \\ {d_i} &&&&&& {d_k} \\ \\ &&&& {d_j} \arrow["{f}"', from=5-1, to=7-5] \arrow["{g}"', from=7-5, to=5-7] \arrow["{h}"', from=5-1, to=5-7] \arrow["{p_i}"', from=1-4, to=5-1] \arrow["{p_j}", from=1-4, to=7-5] \arrow["{p_k}", from=1-4, to=5-7] \arrow["{\eta _f}"{description}, shift left=5, shorten <=34pt, shorten >=34pt, Rightarrow, from=5-1, to=7-5] \arrow["{\eta _g}"{description}, shift left=5, shorten <=17pt, shorten >=17pt, Rightarrow, from=7-5, to=5-7] \arrow["{\eta _h}"{description}, shift left=5, shorten <=53pt, shorten >=53pt, Rightarrow, from=5-1, to=5-7] \end{tikzcd} } such that for each 2-cell $g\circ f \Rightarrow h$ in the diagram the above tetrahedron is filled by the identical 3-cell. Moreover it has the following universal property: given another such cone there is a map $r:e\to d$, unique up to unique natural isomorphism, together with 2-isomorphisms $\alpha _i : p_i r \Rightarrow q_i$ $ $\\ \adjustbox{scale=0.8,center}{ \begin{tikzcd} &&&&& {} \\ &&& e \\ \\ && {} && {} \\ && {} & d \\ &&&&& {} \\ &&&& {} &&&&& {} \\ \\ {d_i} &&&&&& {d_k} \\ \\ &&&& {d_j} \arrow["{f}"', from=9-1, to=11-5] \arrow["{g}"', from=11-5, to=9-7] \arrow["{h}"', from=9-1, to=9-7] \arrow["{p_i}"', from=5-4, to=9-1] \arrow["{p_j}", from=5-4, to=11-5] \arrow["{p_k}", from=5-4, to=9-7] \arrow["{\eta _{f}}"{description}, shift left=5, shorten <=34pt, shorten >=34pt, Rightarrow, from=9-1, to=11-5] \arrow["{q_i}"', curve={height=12pt}, from=2-4, to=9-1] \arrow["{q_k}", curve={height=-12pt}, from=2-4, to=9-7] \arrow[""{name=0, anchor=center, inner sep=0}, "{q_j}"{pos=0.2}, curve={height=-12pt}, from=2-4, to=11-5] \arrow["r"', dashed, from=2-4, to=5-4] \arrow["{\alpha _i}"{description}, Rightarrow, from=5-4, to=4-3] \arrow["{\alpha _j}"'{pos=0.1}, shorten >=22pt, Rightarrow, from=5-4, to=4-5] \arrow["{\nu _{f}}"{description}, shift right=5, shorten >=11pt, Rightarrow, from=5-3, to=0] \end{tikzcd} } $ $\\ such that the composite of the 2-cells $\alpha _i: p_i r\Rightarrow q_i $, $\alpha _j : p_j r \Rightarrow q_j$ and $\eta _f: f p_i \Rightarrow p_j$ is $\nu _f : f q_i \Rightarrow q_j$ (for each arrow $f$ of the diagram). \end{definition} \begin{remark} If $\mathcal{C}$ is a strict (2,1)-category then we can give a simpler description for certain homotopy limits. E.g.~a 2-pullback is a square \[\begin{tikzcd} {C^*} && C \\ \\ B && A \arrow["{g'}"', from=1-1, to=3-1] \arrow["f"', from=3-1, to=3-3] \arrow["{f'}", from=1-1, to=1-3] \arrow["g", from=1-3, to=3-3] \arrow["\eta"{description}, shorten <=11pt, shorten >=11pt, Rightarrow, from=3-1, to=1-3] \end{tikzcd}\] such that given an outer square with natural isomorphism $\nu :fh_1 \Rightarrow gh_2$ there is a map $r: D\to C^*$ (unique up to unique natural isomorphism) and isomorphisms $\alpha _1: g'r\Rightarrow h_1$, $\alpha _2: f'r\to h_2$ \[\begin{tikzcd} D \\ \\ && {C^*} && C \\ \\ && B && A \arrow["{g'}"', from=3-3, to=5-3] \arrow["f"', from=5-3, to=5-5] \arrow["{f'}", from=3-3, to=3-5] \arrow["g", from=3-5, to=5-5] \arrow["\eta"{description}, shorten <=11pt, shorten >=11pt, Rightarrow, from=5-3, to=3-5] \arrow[""{name=0, anchor=center, inner sep=0}, "{h_2}", curve={height=-12pt}, from=1-1, to=3-5] \arrow[""{name=1, anchor=center, inner sep=0}, "{h_1}"', curve={height=12pt}, from=1-1, to=5-3] \arrow["r", dashed, from=1-1, to=3-3] \arrow["{\alpha _1}"{description}, shorten >=5pt, Rightarrow, from=3-3, to=1] \arrow["{\alpha _2}"{description}, shorten >=5pt, Rightarrow, from=3-3, to=0] \end{tikzcd}\] such that the composition of the 2-cells $\alpha _1^{-1}$, $\alpha _2$ and $\eta $ is $\nu$. \end{remark} \begin{remark} Our notion of a 2-(co)limit coincides with that of \cite{lurietopos} when $\mathcal{C}$ is regarded as an $(\infty ,1)$-category where the lifting in the diagram \[\begin{tikzcd} {\Lambda _i^n} && {\mathcal{C}} \\ \\ {\Delta ^n} \arrow[hook, from=1-1, to=3-1] \arrow[dashed, from=3-1, to=1-3] \arrow[from=1-1, to=1-3] \end{tikzcd}\] is unique for $n\geq 3$ and $0<i<n$. \end{remark} \begin{theorem} $\mathbf{Coh_{\sim}}$ has all homotopy products and pullbacks. \end{theorem} \begin{proof} First observe, that the (1-categorical) product is a homotopy limit. I.e.~as there are no arrows in the diagram, the notion of a(n ordinary) cone coincides with the one in the 2-categorical sense, hence it suffices to prove that given two maps $f,g:\mathcal{D}\to \prod \mathcal{C}_i$, whose projections $p_i\circ f$, $p_i \circ g$ are naturally isomorphic (shown by $\eta _i$), the original maps $f$, $g$ are also isomorphic (just take $\eta _i$ in the $i^{\text{th}}$ coordinate) and the isomorphism is uniquely determined by the $\eta _i$-s. Now we show the existence of all homotopy pullbacks. By the existence of the Joyal model structure on $\mathbf{Cat}$ we can factor any functor $f:\mathcal{C}\to \mathcal{D}$ as $\mathcal{C}\xrightarrow{j}\mathcal{C}'\xrightarrow{f'}\mathcal{D}$ where $j$ is an equivalence (and it is injective on objects), and $f'$ is an isofibration (see e.g.~Theorem 6.2. in \cite{joyal}). Obviously each equivalence with coherent domain is a coherent functor with coherent codomain, and if $f$, $\mathcal{C}$, $\mathcal{D}$ are coherent, then so are $\mathcal{C}'$, $j$ and $f'$. As the (1-)pullback along an isofibration exists (see \ref{cohpb}), we have a natural candidate for a homotopy pullback. Assume that we are given the coherent functors $f:\mathcal{B}\to \mathcal{A}$, $g^*:\mathcal{C}\to \mathcal{A}$, we would like to form their 2-pullback. Factor $g^*$ as $\mathcal{C}\xrightarrow{\varphi } \mathcal{C}'\xrightarrow{g} \mathcal{A}$, where $\varphi $ is an equivalence and $g$ is an isofibration. Now form the 1-pullback of $f$ and $g$ in $\mathbf{Coh}$, then precompose $f'$ to the quasi-inverse $\varphi ^*$ of $\varphi$ to get the edge of a cone, whose codomain is $\mathcal{C}$. We claim that this construction results a 2-pullback in the 2-category $\mathbf{Coh_{\sim}}$. \begin{center} \begin{tikzcd} D &&&&&&& C \\ \\ &&&& {C^*} &&& {C'} \\ \\ \\ &&&& B &&& A \arrow["{f'}"', from=3-5, to=3-8] \arrow["{g'}"', from=3-5, to=6-5] \arrow["f"', from=6-5, to=6-8] \arrow["g", from=3-8, to=6-8] \arrow["\varphi", from=1-8, to=3-8] \arrow["{f^*}"'{pos=0.6}, from=3-5, to=1-8] \arrow["{\varphi ^*}", curve={height=-12pt}, dashed, from=3-8, to=1-8] \arrow["{h_1}"', curve={height=12pt}, from=1-1, to=6-5] \arrow["{h_2}", from=1-1, to=1-8] \arrow["r", dashed, from=1-1, to=3-5] \arrow["h", curve={height=-12pt}, from=1-1, to=3-8] \arrow[shorten <=37pt, shorten >=37pt, Rightarrow, no head, from=6-5, to=3-8] \arrow["\alpha", shift left=5, shorten <=16pt, shorten >=25pt, Rightarrow, from=3-5, to=1-8] \end{tikzcd} \end{center} First, it is clear that $g\varphi f^* \cong fg'$ and the components of this isomorphism ($\eta $) are given by: $$\eta _x: fg'(x)=gf'(x)\xrightarrow{g(\chi f')}g\varphi \varphi ^* f' (x)= g\varphi f^*(x)$$ where $\chi : 1_{C'}\to \varphi \varphi ^*$ is the unit of the (adjoint) equivalence $\varphi$. Assume that we are given the maps $h_1: D\to B$, $h_2:D\to C$ and an isomorphism $\nu : fh_1\Rightarrow g\varphi h_2$. The isomorphisms $\nu _d:fh_1(d)\to g\varphi h_2$ can be lifted to $C'$ (as $g$ is an isofibration), hence there are isomorphisms $\mu _d: c'_d\to \varphi h_2 (d)$ with $g(\mu _d)=\nu _d$. We define a functor $h:D\to C'$: it takes an object $d$ to $c'_d$ and an arrow $i:d\to d'$ to $\mu _{d'}^{-1} \circ \varphi h_2(i) \circ \mu _d$. \[\begin{tikzcd} {c'_d} && {\varphi h_2(d)} \\ \\ {c'_{d'}} && {\varphi h_2(d')} \arrow["{\mu_d}", from=1-1, to=1-3] \arrow[from=1-1, to=3-1] \arrow["{\varphi h_2(i)}", from=1-3, to=3-3] \arrow["{\mu _{d'}}"', from=3-1, to=3-3] \end{tikzcd}\] It is clear that $\mu $ is a natural isomorphism from $h$ to $\varphi h_2$. It also follows that $gh=fg'$ (just take the image of the above square at $g$). By the universal property of the pullback we have an arrow $r:D\to C^*$ with $g'r=h_1$ and $f'r=h$. We need to find a natural isomorphism $\alpha : f^*r\Rightarrow h_2$ such that the composite $$fh_1=fg'r\xRightarrow{\eta r} g\varphi f^* r \xRightarrow{g\varphi (\alpha )} g\varphi h_2$$ gives $\nu$. We will take $$\alpha :=f^*r = \varphi ^*f'r =\varphi ^* h \xRightarrow{\varphi ^*(\mu )}\varphi ^* \varphi h_2 \xRightarrow{\chi 'h_2}h_2$$ (where $\chi ': \varphi ^*\varphi \to 1_C$ is the counit) and then check the above property. The solution is shipped by \[\begin{tikzcd} {f'r(d)} && {\varphi \varphi ^*f'r(d)} &&& {\varphi \varphi ^*\varphi h_2(d)} && {\varphi h_2(d)} \\ \\ && {f'r(d)} &&& {\varphi h_2(d)} \\ && {} \arrow["{\chi ^{-1}_{f'r(d)}}", from=1-3, to=3-3] \arrow["{\varphi \varphi ^*(\mu _d)}", from=1-3, to=1-6] \arrow["{\mu _d}"', from=3-3, to=3-6] \arrow["{\chi ^{-1}_{\varphi h_2(d)}}"', from=1-6, to=3-6] \arrow["{\chi _{f'r(d)}}", from=1-1, to=1-3] \arrow[Rightarrow, no head, from=1-1, to=3-3] \arrow["{\varphi (\chi '_{h_2(d)})}", from=1-6, to=1-8] \arrow[Rightarrow, no head, from=3-6, to=1-8] \end{tikzcd}\] where the right triangle commutes as $\chi $ and $\chi '$ satisfy the triangle identities. It remains to prove that $r$ is unique up to unique compatible natural isomorphism. Assume that we are given \[\begin{tikzcd} D \\ && {C^*} && C \\ &&&& {C'} \\ && B && A \arrow["{g'}", from=2-3, to=4-3] \arrow["f"', from=4-3, to=4-5] \arrow["{f^*}"', from=2-3, to=2-5] \arrow["{r_2}", shift left=2, from=1-1, to=2-3] \arrow["{r_1}"', shift right=2, from=1-1, to=2-3] \arrow[""{name=0, anchor=center, inner sep=0}, "{h_2}", curve={height=-18pt}, from=1-1, to=2-5] \arrow[""{name=1, anchor=center, inner sep=0}, "{h_1}"', curve={height=12pt}, from=1-1, to=4-3] \arrow["g", from=3-5, to=4-5] \arrow["\varphi", from=2-5, to=3-5] \arrow["{\eta }", shorten <=11pt, shorten >=11pt, Rightarrow, from=4-3, to=2-5] \arrow["{\alpha '_1,\alpha '_2}"', shorten >=3pt, Rightarrow, from=2-3, to=0] \arrow["{\beta _1, \beta _2}", shorten >=7pt, Rightarrow, from=2-3, to=1] \end{tikzcd}\] such that the composition of the 2-cells $\alpha '_i$, $\beta _i^{-1}$ and $\eta $ is $\nu$. We first show that this implies the case \[\begin{tikzcd} &&&& C \\ D \\ && {C^*} && {C'} \\ \\ && B && A \arrow["{f'}"', from=3-3, to=3-5] \arrow["g", from=3-5, to=5-5] \arrow["{g'}", from=3-3, to=5-3] \arrow["f"', from=5-3, to=5-5] \arrow["{r_1,r_2}"{description, pos=0.4}, from=2-1, to=3-3] \arrow[""{name=0, anchor=center, inner sep=0}, "{h_1}"', curve={height=6pt}, from=2-1, to=5-3] \arrow[""{name=1, anchor=center, inner sep=0}, "h", curve={height=-12pt}, from=2-1, to=3-5] \arrow["{f^*}"'{pos=0.7}, from=3-3, to=1-5] \arrow["{\varphi }", from=1-5, to=3-5] \arrow["{h_2}", from=2-1, to=1-5] \arrow[shorten <=22pt, shorten >=22pt, Rightarrow, no head, from=5-3, to=3-5] \arrow["{\varphi ^*}"', curve={height=18pt}, dashed, from=3-5, to=1-5] \arrow["{\beta _1,\beta _2}"', shorten >=6pt, Rightarrow, from=3-3, to=0] \arrow["{\alpha _1,\alpha _2}", shorten >=3pt, Rightarrow, from=3-3, to=1] \end{tikzcd}\] with $f(\beta _i)=g(\alpha _i)$. We take $$\alpha _i:= f'r_i\xRightarrow{\chi _{f'r_i}} \varphi \varphi ^*f'r_i \xRightarrow{\varphi (\alpha _i')} \varphi h_2 \xRightarrow{\mu ^{-1}} h$$. The fact that $\eta$, $\beta _i^{-1}$ and $\alpha _i$ glue together and form $\nu$ can be expressed by the commutative square: \[\begin{tikzcd} {fg'r_i(d)} && {g\varphi f^*r_i(d)} \\ \\ {fh_1(d)} && {g\varphi h_2(d)} \arrow["{\eta _{r_i(d)}}", from=1-1, to=1-3] \arrow["{g\varphi ((\alpha _i')_d)}", from=1-3, to=3-3] \arrow["{f((\beta_i)_d)}"', from=1-1, to=3-1] \arrow["{\nu _d}"', from=3-1, to=3-3] \end{tikzcd}\] hence $f((\beta _i)_d)$ is $g(\mu _d^{-1})\circ g\varphi ((\alpha '_i)_d) \circ g(\chi _{f'r(d)})=g((\alpha _i)_d)$. Now if we take $$\alpha :=\alpha _2^{-1}\alpha _1: f'r_1 \Rightarrow f'r_2$$ and $$\beta := \beta _2^{-1}\beta _1: g'r_1\Rightarrow g'r_2$$, then we have $f(\beta )=g(\alpha )$. Recall that the (coherent) category $C^*$ can be explicitly described as the one with objects $\{(b,c'): b\in Ob(B), c'\in Ob(C'), f(b)=g(c')\}$ and similarly for arrows. By the above property $(\beta , \alpha )$ is a well-defined natural isomorphism $r_1=(g'r_1,f'r_1)\Rightarrow (g'r_2,f'r_2)=r_2$ and it is uniquely determined by the 2-cells $\alpha _i$, $\beta _i$, $\eta $ and $\nu $. But the latter is equivalent to the datum of $\alpha '_i$, $\beta _i$, $\eta $ and $\nu $. \end{proof} Recall the following Proposition (4.4.2.6.) together with its dual from \cite{lurietopos}. \begin{theorem}[Lurie] If $\mathcal{C}$ is an $(\infty , 1)$-category and it has (homotopy) pushouts and $\kappa$-small (homotopy) coproducts then $\mathcal{C}$ has all $\kappa$-small (homotopy) colimits. \end{theorem} \begin{corollary} $\mathbf{Coh_{\sim}}$ has all homotopy limits. \end{corollary} Now we would like to prove the existence of homotopy colimits. First we need the basic fact that equalizers are monic: \begin{proposition} Given a diagram formed by a set of paralel arrows and some natural isomorphisms between them, its homotopy limit (equalizer) \[\begin{tikzcd} && v \\ \\ w \\ &&& {w'} \arrow[""{name=0, anchor=center, inner sep=0}, "{f_i}"', shift right=3, from=3-1, to=4-4] \arrow[""{name=1, anchor=center, inner sep=0}, "{f_j}", shift left=2, shorten <=8pt, shorten >=8pt, from=3-1, to=4-4] \arrow[""{name=2, anchor=center, inner sep=0}, "e"', from=1-3, to=3-1] \arrow[""{name=3, anchor=center, inner sep=0}, "{e'}", from=1-3, to=4-4] \arrow[shorten <=1pt, shorten >=1pt, Rightarrow, from=0, to=1] \arrow["{\eta _i}"{description}, curve={height=6pt}, shorten <=10pt, shorten >=10pt, Rightarrow, from=2, to=3] \arrow["{\eta _j}"{description}, curve={height=-6pt}, shorten <=10pt, shorten >=10pt, Rightarrow, from=2, to=3] \end{tikzcd}\] has the property, that for any natural isomorphism $\alpha : eg\Rightarrow eh$ there is a unique natural isomorphism $\gamma : g\Rightarrow h$ such that $e\gamma =\alpha $. \label{eqmono} \end{proposition} \begin{proof} Take $\beta _i$ to be $e'g\xRightarrow{\eta _i g} f_ieg \xRightarrow{f_i \alpha} f_ieh \xRightarrow{\eta _i^{-1}h} e'h$. Then \[\begin{tikzcd} & u &&& u \\ & v &&& v \\ w && {w'} & w && {w'} \arrow[""{name=0, anchor=center, inner sep=0}, "e"{description}, from=2-2, to=3-1] \arrow[""{name=1, anchor=center, inner sep=0}, "{e'}"{description}, from=2-2, to=3-3] \arrow["{f_i}"', from=3-1, to=3-3] \arrow[""{name=2, anchor=center, inner sep=0}, "eg"{description}, curve={height=12pt}, from=1-2, to=3-1] \arrow[""{name=3, anchor=center, inner sep=0}, "{e'h}"{description}, curve={height=-12pt}, from=1-2, to=3-3] \arrow["h"', from=1-2, to=2-2] \arrow["{f_i}"', from=3-4, to=3-6] \arrow[""{name=4, anchor=center, inner sep=0}, "e"{description}, from=2-5, to=3-4] \arrow[""{name=5, anchor=center, inner sep=0}, "{e'}"{description}, from=2-5, to=3-6] \arrow[""{name=6, anchor=center, inner sep=0}, "eg"{description}, curve={height=12pt}, from=1-5, to=3-4] \arrow[""{name=7, anchor=center, inner sep=0}, "{e'h}"{description}, curve={height=-12pt}, from=1-5, to=3-6] \arrow["g"', from=1-5, to=2-5] \arrow["{\eta _i}"', shorten <=6pt, shorten >=6pt, Rightarrow, from=0, to=1] \arrow["\alpha", shorten <=6pt, Rightarrow, from=2, to=2-2] \arrow[shorten <=4pt, shorten >=8pt, Rightarrow, no head, from=2-2, to=3] \arrow["{\eta _i}"', shorten <=6pt, shorten >=6pt, Rightarrow, from=4, to=5] \arrow[shorten <=8pt, shorten >=4pt, Rightarrow, no head, from=6, to=2-5] \arrow["\beta _i", shorten <=2pt, shorten >=6pt, Rightarrow, from=2-5, to=7] \end{tikzcd}\] are both splittings of the 2-cells $f_ieg \xRightarrow{f_i\alpha} f_ieh \xRightarrow{\eta _i h} e'h$, hence there is a unique natural isomorphism $\gamma :g\Rightarrow h$ for which $\alpha =e\gamma $ and $\beta _i =e'\gamma$. The latter is easily proved to be redundant. \end{proof} The following is the 2-categorical analogue of Theorem V.6.1. in \cite{maclane}. \begin{proposition} Let $\mathcal{C}$ be a locally small 2-complete strict (2,1)-category. Assume that there exists a small set $X\subset Ob(\mathcal{C})$ such that for each $c\in Ob(\mathcal{C})$ there is an element $x\in X$ and an arrow $x\to c$. Then $\mathcal{C}$ has a homotopy initial object. \end{proposition} \begin{proof} The product $w=\prod _{x\in X} x$ is a weak initial object, i.e.~given any other object $c$ there is at least one map $w\to c$ (e.g.~the assumed one composed with the suitable projection). By assumption the class $Hom(w,w)$ is a set, hence we can take its joint 2-equalizer: the homotopy limit of this 1-dimensional diagram: \[\begin{tikzcd} & v \\ \\ w && w \arrow[from=3-1, to=3-3] \arrow[shift left=1, from=3-1, to=3-3] \arrow["{f_i}"', shift right=1, from=3-1, to=3-3] \arrow[""{name=0, anchor=center, inner sep=0}, "e"', from=1-2, to=3-1] \arrow[""{name=1, anchor=center, inner sep=0}, "{e'}", from=1-2, to=3-3] \arrow[shorten <=6pt, shorten >=6pt, Rightarrow, from=0, to=1] \arrow["{\eta _i}"', shift right=2, shorten <=6pt, shorten >=6pt, Rightarrow, from=0, to=1] \end{tikzcd}\] We claim that $v$ is homotopy initial. It is clear, that $v$ is a weakly initial object. Let $g,h:v\to c$ be two maps. By assumption the natural isomorphisms between $f$ and $g$ form a set, and we can take the equalizer of this (now 2-dimensional) diagram. \[\begin{tikzcd} u \\ && v && c \\ \\ w && w && w \arrow[""{name=0, anchor=center, inner sep=0}, "h"', shift right=3, from=2-3, to=2-5] \arrow[""{name=1, anchor=center, inner sep=0}, "g", shift left=3, from=2-3, to=2-5] \arrow[shift right=2, from=4-3, to=4-5] \arrow[shift left=2, from=4-3, to=4-5] \arrow[from=4-3, to=4-5] \arrow["e"', from=2-3, to=4-3] \arrow[""{name=2, anchor=center, inner sep=0}, "{e^*}"{pos=0.3}, from=1-1, to=2-3] \arrow[""{name=3, anchor=center, inner sep=0}, "{e^{**}}", curve={height=-24pt}, from=1-1, to=2-5] \arrow["s", from=4-1, to=1-1] \arrow[shorten <=2pt, shorten >=2pt, Rightarrow, from=1, to=0] \arrow["{\nu _g}"{description}, shorten <=4pt, shorten >=4pt, Rightarrow, from=2, to=3] \arrow["{\nu _h}"{description}, shift right=4, shorten <=8pt, Rightarrow, from=2, to=3] \end{tikzcd}\] As $w$ was weakly initial there is a map $s:w\to u$. Since $e$ was the equalizer of all homomorphisms from $w$ to $w$, there is a natural isomorphism: $ee^*se \xRightarrow{\eta _{ee^*s}} e' \xRightarrow{\eta _{1_w}^{-1}} e$, and by Proposition \ref{eqmono} there is a unique isomorphism $\gamma : e^*(se)\Rightarrow 1_v$ such that $e\gamma =\eta _{1_w}^{-1}\eta _{ee^*s}$ (but this fact will not be used). We can construct a natural isomorphism: $$g\xRightarrow{g\gamma ^{-1}} ge^*(se) \xRightarrow{\nu _g(se)} e^{**}(se) \xRightarrow{\nu _h^{-1}(se)} he^*(se)\xRightarrow{h\gamma }h $$ For uniqueness we have to prove that the pentagon \[\begin{tikzcd} && {e^{**}(se)} \\ {ge^*(se)} &&&& {he^*(se)} \\ g &&&& h \arrow["\chi"', Rightarrow, from=3-1, to=3-5] \arrow["{\chi (e^*se)}"', Rightarrow, from=2-1, to=2-5] \arrow["g\gamma", Rightarrow, from=2-1, to=3-1] \arrow["h\gamma"', Rightarrow, from=2-5, to=3-5] \arrow["{\nu_g (se)}", Rightarrow, from=2-1, to=1-3] \arrow["{\nu_h^{-1}(se)}", Rightarrow, from=1-3, to=2-5] \end{tikzcd}\] commutes for arbitrary $\chi$. The first floor commutes by the interchange law for strict 2-categories (as both composites must be equal to the horizontal composite of $\gamma $ and $\chi $), the roof commutes as $(u,e^*,e^{**})$ form a homotopy cone, in particular $\chi e^*=\nu _h^{-1}\nu _g$. \end{proof} Recall that given a map $p:K\to \mathcal{C}$ of simplicial sets (where $\mathcal{C}$ is an $(\infty ,1)$-category), its (homotopy) colimit is an initial object of $\mathcal{C}_{p/}$ (the infinity category of homotopy cocones, or the undercategory) (See: \cite{lurietopos}). Therefore we would like to use our previous statement for $\mathcal{C}_{p/}$. The following theorem is due to Pál Zsámboki. \begin{theorem} Let $\mathcal{C}$ be a complete $(\infty ,1)$-category and let $p:K\to \mathcal{C}$ be a map of simplicial sets. Then the undercategory $\mathcal{C}_{p/}$ is also complete. \end{theorem} \begin{proof} We will use that the construction $K\star L$ gives a simplicial set with the property that given a simplicial map $q:L\to \mathcal{C}$, the simplicial maps of the form $K\to \mathcal{C}_{/q}$ are the same as those maps $K\star L \to \mathcal{C}$ whose restriction to $L$ gives $q$ (and the dual property holds for $\mathcal{C}_{p/}$). Let $L$ be a simplicial set and $L\xrightarrow{q} \mathcal{C}_{p/}$ be a diagram. Then $q$ is a diagram $K\star L \to \mathcal{C}$. Let $q_0:L\to \mathcal{C}$ be its restriction and $\bar{q_0}:\Delta _0 \star L\to \mathcal{C}$ be the limit for $q_0$. That is, the restriction map $\mathcal{C}_{/\bar{q_0}} \to \mathcal{C}_{/q_0}$ is a trivial fibration, in particular there is a lift in \[\begin{tikzcd} \emptyset && {\mathcal{C}_{/\bar{q_0}}} \\ \\ K && {\mathcal{C}_{/q_0}} \arrow[from=1-1, to=3-1] \arrow[from=1-1, to=1-3] \arrow[from=1-3, to=3-3] \arrow["q", from=3-1, to=3-3] \arrow["{\bar{q}}"{description}, dashed, from=3-1, to=1-3] \end{tikzcd}\] Then $\bar{q}$ corresponds to a map $K\star \Delta _0 \star L \to \mathcal{C}$, i.e.~to a map $\Delta _0 \star L \to \mathcal{C}_{p/}$. We claim that $\bar{q}$ is the limit of $q$, that is the restriction map $(\mathcal{C} _{p/})_{/\bar{q_0}} \to (\mathcal{C} _{p/})_{/q_0}$ is a trivial fibration. Let $X\hookrightarrow X'$ be an inclusion of simplicial sets. The lifting problem \[\begin{tikzcd} X && {(\mathcal{C} _{p/})_{/\bar{q_0}}} \\ \\ {X'} && {(\mathcal{C} _{p/})_{/q_0}} \arrow[hook, from=1-1, to=3-1] \arrow[from=3-1, to=3-3] \arrow[from=1-3, to=3-3] \arrow[from=1-1, to=1-3] \arrow[dashed, from=3-1, to=1-3] \end{tikzcd}\] is the same as the lifting problem \[\begin{tikzcd} {K\star X} && {\mathcal{C} _{/\bar{q_0}}} \\ \\ {K\star X'} && {\mathcal{C}_{/q_0}} \arrow[hook, from=1-1, to=3-1] \arrow[from=3-1, to=3-3] \arrow[from=1-3, to=3-3] \arrow[from=1-1, to=1-3] \arrow[dashed, from=3-1, to=1-3] \end{tikzcd}\] and thus has a solution as $\bar{q_0}$ was the limit of $q_0$. \end{proof} \begin{corollary} The locally small strict (2,1)-category $(\mathbf{Coh_{\sim}})_{p/}$ is 2-complete. (In particular it is non-empty by the existence of a terminal object.) \end{corollary} It remains to find a weakly initial family of homotopy cocones over $p$. It is a straightforward consequence of Proposition \ref{solution}: given a homotopy cocone (with top object $c$), the joint image of the edges $p(i)\to c$ ($i\in K_0$) is included in some coherent subcategory with cardinality $\leq \aleph _0 \cdot \prod _{i\in K_0} |p(i)|$, hence the set of all cocones with top object having at most this cardinality is a solution set. We proved: \begin{theorem} The (2,1)-category $\mathbf{Coh_{\sim}}$ is 2-complete and 2-cocomplete. \end{theorem} \section{Small object argument} In this section we will generalise the classical small object argument for locally small 2-cocomplete strict (2,1)-categories, which we will typically denote by $\mathbf{C}$. The proof follows the one given in \cite{hovey} for the 1-categorical setting. \begin{definition} Given a 2-colimit preserving diagram $\lambda \to \mathbf{C}$ with homotopy colimit $\mathcal{X}$ \[\begin{tikzcd} && {\mathcal{X}} \\ {\mathcal{X}_0} & {\mathcal{X}_1} & {\mathcal{X}_2} & \dots \arrow["{f_0}"', from=2-1, to=2-2] \arrow["{f_1}"', from=2-2, to=2-3] \arrow["{f_2}"', from=2-3, to=2-4] \arrow["f", from=2-1, to=1-3] \arrow[from=2-2, to=1-3] \arrow[from=2-3, to=1-3] \end{tikzcd}\] the coprojection map $f:\mathcal{X}_0\to \mathcal{X}$ is called the \emph{transfinite composition} of the $\lambda $-sequence $(f_i)_{i<\lambda }$. \end{definition} \begin{definition} Let $I\subset Arr(\mathbf{C})$ be a set. \emph{$I$-cell} is the class of maps that can be written as the transfinite composition of 2-pushouts from $I$. \emph{$I$-inj} is the class whose members ($f$) have the following right lifting property: given a square \[\begin{tikzcd} {\mathcal{C}} && {\mathcal{X}} \\ \\ {\mathcal{D}} && {\mathcal{Y}} \arrow["g"', from=1-1, to=3-1] \arrow["{h'}", from=3-1, to=3-3] \arrow["f", from=1-3, to=3-3] \arrow["h", from=1-1, to=1-3] \arrow["\eta"{description}, shorten <=11pt, shorten >=11pt, Rightarrow, from=3-1, to=1-3] \end{tikzcd}\] with $g\in I$, there is a lifting \[\begin{tikzcd} {\mathcal{C}} && {\mathcal{X}} \\ \\ {\mathcal{D}} && {\mathcal{Y}} \arrow["g"', from=1-1, to=3-1] \arrow[""{name=0, anchor=center, inner sep=0}, "{h'}", from=3-1, to=3-3] \arrow["f", from=1-3, to=3-3] \arrow[""{name=1, anchor=center, inner sep=0}, "h", from=1-1, to=1-3] \arrow["k"{description}, from=3-1, to=1-3] \arrow["{\nu_1}"{description}, shorten <=13pt, shorten >=9pt, Rightarrow, from=3-1, to=1] \arrow["{\nu_2}"{description}, shorten <=13pt, shorten >=9pt, Rightarrow, from=0, to=1-3] \end{tikzcd}\] such that the composition of $\nu _1$ and $\nu _2$ is $\eta$. \emph{$I$-proj} is the class whose members have the left lifting property wrt.~$I$. As usual \emph{$I$-cof}=($I$-inj)-proj, and \emph{$I$-fib}=($I$-proj)-inj. \end{definition} \begin{proposition} $I$-cell $\subseteq I$-cof. \end{proposition} \begin{proof} Clearly $I\subseteq I$-cof, hence it suffices to prove that $I$-cof is closed under pushouts and transfinite compositions. First we show that if $f$ has the left lifting property wrt. $m$, then its 2-pushout $f'$ has also. $l_1$ is induced by the lifting property of $f$ and $l_2$ by the universality of the 2-pushout. \[\begin{tikzcd} \bullet && \bullet && \bullet \\ \\ \bullet && \bullet && \bullet \arrow[""{name=0, anchor=center, inner sep=0}, "f"{description}, from=1-1, to=3-1] \arrow["{g'}"{description}, from=3-1, to=3-3] \arrow[""{name=1, anchor=center, inner sep=0}, "g"{description}, from=1-1, to=1-3] \arrow["{f'}"{description}, from=1-3, to=3-3] \arrow["h"{description}, from=1-3, to=1-5] \arrow[""{name=2, anchor=center, inner sep=0}, "m"{description}, from=1-5, to=3-5] \arrow[""{name=3, anchor=center, inner sep=0}, "k"{description}, from=3-3, to=3-5] \arrow["{l_1}"{description, pos=0.7}, curve={height=12pt}, dashed, from=3-1, to=1-5] \arrow["{l_2}"{description, pos=0.7}, curve={height=-12pt}, dashed, from=3-3, to=1-5] \arrow["\alpha"{description}, shorten <=7pt, shorten >=7pt, Rightarrow, from=3, to=2] \arrow["\eta"{description}, shorten <=7pt, shorten >=7pt, Rightarrow, from=0, to=1] \end{tikzcd}\] The properties that $l_1$ and $l_2$ are splittings of the related 2-cells can be written as $\eta +\alpha =\gamma + \beta ^{-1}$ and $\eta +\nu +\mu =\gamma $. We should prove that $l_2$ is a splitting of $\alpha $. It is enough to see that in \[\begin{tikzcd} \bullet && \bullet &&&& \bullet && \bullet \\ \\ \bullet && \bullet &&&& \bullet && \bullet && \bullet \\ &&& \bullet \\ &&&& \bullet &&&& \bullet && \bullet \arrow["g"{description}, from=1-1, to=1-3] \arrow["{f'}"{description}, from=1-3, to=3-3] \arrow["f"{description}, from=1-1, to=3-1] \arrow["{g'}"{description}, from=3-1, to=3-3] \arrow[""{name=0, anchor=center, inner sep=0}, "h"{description}, curve={height=-12pt}, from=1-3, to=4-4] \arrow[""{name=1, anchor=center, inner sep=0}, "{l_1}"{description}, curve={height=12pt}, from=3-1, to=4-4] \arrow["m"{description}, from=4-4, to=5-5] \arrow[""{name=2, anchor=center, inner sep=0}, "{l_2}"{description}, from=3-3, to=4-4] \arrow["\eta"{description}, shorten <=17pt, shorten >=17pt, Rightarrow, from=3-1, to=1-3] \arrow["f"{description}, from=1-7, to=3-7] \arrow["{g'}"{description}, from=3-7, to=3-9] \arrow["g"{description}, from=1-7, to=1-9] \arrow["{f'}"{description}, from=1-9, to=3-9] \arrow["m"{description}, from=3-11, to=5-11] \arrow[""{name=3, anchor=center, inner sep=0}, "h"{description}, from=1-9, to=3-11] \arrow[""{name=4, anchor=center, inner sep=0}, "k"{description}, from=3-9, to=5-11] \arrow[""{name=5, anchor=center, inner sep=0}, "{l_1}"{description}, from=3-7, to=5-9] \arrow["m"{description}, from=5-9, to=5-11] \arrow["\eta"{description}, shorten <=17pt, shorten >=17pt, Rightarrow, from=3-7, to=1-9] \arrow["\mu", shorten <=5pt, shorten >=5pt, Rightarrow, from=2, to=0] \arrow["\nu", shorten <=7pt, shorten >=7pt, Rightarrow, from=1, to=2] \arrow["\alpha"{description}, shorten <=9pt, shorten >=9pt, Rightarrow, from=4, to=3] \arrow["\beta"{description}, shorten <=13pt, shorten >=13pt, Rightarrow, from=5, to=4] \end{tikzcd}\] the 2-cells filling the boundaries are identical as in this case both $ml_2$ and $k$ are suitable splittings, hence there is a unique natural isomorphism $\delta:k \Rightarrow ml_2$ for which $\beta +\delta =m\nu $ and $\mu +\delta =\alpha$. This follows from the identities observed above. Now assume that each $f_i$ ($i<\lambda $) has left lifting property wrt.~$m$ (and that $f_i$-s form a (co)continuous sequence). We have to prove that its transfinite composition $f$ has the same lifting property. The proof is similar to the previous one and it is pictured as \[\begin{tikzcd} && \bullet \\ \\ \bullet &&&& \bullet \\ \bullet \\ \bullet && \bullet \\ \dots \arrow["{f_0}"', from=3-1, to=4-1] \arrow["{f_1}"', from=4-1, to=5-1] \arrow["{f_2}"', from=5-1, to=6-1] \arrow["f"{description}, from=3-1, to=5-3] \arrow[from=4-1, to=5-3] \arrow[from=5-1, to=5-3] \arrow["h"{description}, from=3-1, to=1-3] \arrow["m"{description}, from=1-3, to=3-5] \arrow["k"{description}, from=5-3, to=3-5] \arrow[dashed, from=4-1, to=1-3] \arrow[dashed, from=5-1, to=1-3] \arrow[squiggly, from=5-3, to=1-3] \end{tikzcd}\] \end{proof} It is worth to write out explicitly: \begin{proposition} (Homotopy) left lifting properties are preserved by (homotopy) pushouts and transfinite compositions. Dually, right lifting properties are preserved by pullbacks and transfinite cocompositions (homotopy limit of the reversed sequence). In particular $I$-$inj$ and $I$-$proj$ are subcategories. \end{proposition} \begin{proposition} $I$-cell is closed under transfinite composition. \end{proposition} \begin{proof} We need to prove that "the transfinite composition of transfinite compositions is a transfinite composition", i.e.~that if we have a sequential (2-)diagram then its colimit can be computed as the colimit of any cofinal subsequence. This is Proposition 4.1.1.8. in \cite{lurietopos}. \end{proof} \begin{proposition} The homotopy pushout of a coproduct of maps from $I$ is in $I$-cell. \end{proposition} \begin{proof} Let $g_j$ $(j\in J)$ be a family of arrows from $I$. Their coproduct is the induced map: \[\begin{tikzcd} {\mathcal{C}_j} && {\cup _{j\in J} \mathcal{C}_j} \\ \\ {\mathcal{D}_j} && {\cup _{j\in J}\mathcal{D}_j} \arrow["{g_j}", from=1-1, to=3-1] \arrow[from=3-1, to=3-3] \arrow[from=1-1, to=1-3] \arrow["{\cup g_j}", dashed, from=1-3, to=3-3] \arrow["{\chi _j}"{description}, shorten <=11pt, shorten >=11pt, Rightarrow, from=1-3, to=3-1] \end{tikzcd}\] Now take the 2-pushout: \[\begin{tikzcd} {\cup _j\mathcal{C}_j} && {\mathcal{X}} \\ \\ {\cup _j\mathcal{D}_j} && {\mathcal{Y}} \arrow["{\cup g_j}"', from=1-1, to=3-1] \arrow["{h_1}", from=3-1, to=3-3] \arrow["f", from=1-3, to=3-3] \arrow["{h_0}", from=1-1, to=1-3] \arrow["\eta"{description}, shorten <=11pt, shorten >=11pt, Rightarrow, from=1-3, to=3-1] \end{tikzcd}\] We will proceed by transfinite recursion and take: $X_0=X$, $\rho _0=f$ and $i_{0,0}=1_X$. In the successor step we form the 2-pushout of $g_j:\mathcal{C}_j\to \mathcal{D}_j$ and $\mathcal{C}_j\to \cup \mathcal{C}_j \xrightarrow{h_0} \mathcal{X} \xrightarrow{i_{0,j}}\mathcal{X}_j$ to get $X_{j+1}$ and induce $\rho _{j+1}$ by the universal property of the square. Hence we get a commutative cube (where the faces are filled with the obvious 2-cells): \[\begin{tikzcd} {\mathcal{C}_j} &&&& {\mathcal{X}_j} \\ & {\cup \mathcal{C}_j} & {\mathcal{X}} & {\mathcal{X}_j} \\ \\ & {\cup \mathcal{D}_j} && {\mathcal{Y}} \\ {\mathcal{D}_j} &&&& {\mathcal{X}_{j+1}} \arrow["{h_0}", from=2-2, to=2-3] \arrow["{i_{0,j}}", from=2-3, to=2-4] \arrow["{\rho _j}"{description}, from=2-4, to=4-4] \arrow["{\cup g_j}"{description}, from=2-2, to=4-2] \arrow["{h_1}"{description}, from=4-2, to=4-4] \arrow[from=1-1, to=2-2] \arrow[Rightarrow, no head, from=1-5, to=2-4] \arrow[from=5-1, to=4-2] \arrow[dashed, from=5-5, to=4-4] \arrow[from=1-5, to=5-5] \arrow[from=1-1, to=1-5] \arrow["{g_j}"{description}, from=1-1, to=5-1] \arrow[from=5-1, to=5-5] \arrow["f"{description}, from=2-3, to=4-4] \end{tikzcd}\] When $j$ is a limit ordinal $\mathcal{X}_j$ is given by the transfinite composition \[\begin{tikzcd} &&& {\mathcal{Y}} \\ &&& {\mathcal{X}_j} \\ {\mathcal{X}_0} && {\mathcal{X}_1} && \dots \arrow["{i_{0,1}}"', from=3-1, to=3-3] \arrow["{i_{1,2}}"', from=3-3, to=3-5] \arrow["{i_{0,j}}"{description}, from=3-1, to=2-4] \arrow["{i_{1,j}}"'{pos=0.7}, from=3-3, to=2-4] \arrow[dashed, from=2-4, to=1-4] \arrow["{\rho_1}"{description}, curve={height=-6pt}, from=3-3, to=1-4] \arrow["{\rho _0}"{description}, curve={height=-6pt}, from=3-1, to=1-4] \end{tikzcd}\] (the 3-cells are filled). We claim that with $\lambda =|J|$ the map $\mathcal{X}\to \mathcal{X}_{\lambda}$ is also a homotopy pushout for $\cup g_j$ along $h_0$. To see this we should find some 2-cells for \[\begin{tikzcd} {\cup \mathcal{C}_j} &&& {\mathcal{X}} \\ \\ {\cup \mathcal{D}_j} &&& {\mathcal{X}_\lambda} \\ &&&& {\mathcal{Y}} \arrow["{h_0}"{description}, from=1-1, to=1-4] \arrow["{i_{0,\lambda}}"{description}, from=1-4, to=3-4] \arrow["{\cup g_j}"{description}, from=1-1, to=3-1] \arrow["{\small{\cup \{\mathcal{D}_j\to \mathcal{X}_{j+1}\to \mathcal{X}_\lambda\}}}", from=3-1, to=3-4] \arrow["{\rho _{\lambda}}", from=3-4, to=4-5] \arrow["f"{description}, curve={height=-12pt}, from=1-4, to=4-5] \arrow["{h_1}"{description}, curve={height=12pt}, from=3-1, to=4-5] \end{tikzcd}\] whose composite is $\eta$. They can be found on the surface of the commutative 3-simplicial set \[\begin{tikzcd} {\mathcal{C}_j} &&&& {\mathcal{X}_j} \\ & {\cup \mathcal{C}_j} & {\mathcal{X}} & {\mathcal{X}_j} \\ \\ & {\cup \mathcal{D}_j} && {\mathcal{Y}} \\ {\mathcal{D}_j} &&&& {\mathcal{X}_{j+1}} \\ & {\cup\mathcal{D}_j} && {\mathcal{X}_{\lambda}} \arrow["{h_0}", from=2-2, to=2-3] \arrow["{i_{0,j}}", from=2-3, to=2-4] \arrow["{\rho _j}"{description}, from=2-4, to=4-4] \arrow["{\cup g_j}"{description}, from=2-2, to=4-2] \arrow["{h_1}"{description}, from=4-2, to=4-4] \arrow[from=1-1, to=2-2] \arrow[Rightarrow, no head, from=1-5, to=2-4] \arrow[from=5-1, to=4-2] \arrow[dashed, from=5-5, to=4-4] \arrow[from=1-5, to=5-5] \arrow[from=1-1, to=1-5] \arrow["{g_j}"{description}, from=1-1, to=5-1] \arrow[from=5-1, to=5-5] \arrow["f"{description}, from=2-3, to=4-4] \arrow[Rightarrow, no head, from=4-2, to=6-2] \arrow[from=5-1, to=6-2] \arrow["{\rho _\lambda}"{description, pos=0.3}, from=6-4, to=4-4] \arrow[from=5-5, to=6-4] \arrow[from=6-2, to=6-4] \end{tikzcd}\] \end{proof} \begin{definition} An object $\mathcal{X}$ of $\mathbf{C}$ is $\lambda $-small wrt.~a subcategory $J$ if $\mathbf{C}(\mathcal{X},-)$ commutes with $\lambda $-filtered sequential 2-colimits formed in $J$. $\mathcal{X}$ is small if it is $\lambda $-small for some $\lambda $. \end{definition} \begin{theorem}[Small object argument] Let $I\subset Arr(\mathbf{C})$ be a set, and assume that domains of $I$ are small relative to $I$-cell. Then for any map $f:\mathcal{X}\to \mathcal{Y}$ there are arrows $\mathcal{X}\xrightarrow{f'}\mathcal{Z}\xrightarrow{f''}\mathcal{Y}$ such that $f'\in I$-cell, $f''\in I$-inj and $f'' \circ f'$ is isomorphic to $f$. \label{smallob} \end{theorem} \begin{proof} We proceed by transfinite recursion and take $\mathcal{Z}_0=\mathcal{X}$, $\rho _0 =f$ and $i_{0,0}=1_{\mathcal{X}}$. For successor ordinal $j+1$ collect all squares \[\begin{tikzcd} {\mathcal{A}_s} && {\mathcal{Z}_j} \\ \\ {\mathcal{B}_s} && {\mathcal{Y}} \arrow["{g_s}"{description}, from=1-1, to=3-1] \arrow["{h_s}"{description}, from=1-1, to=1-3] \arrow["{\rho _j}"{description}, from=1-3, to=3-3] \arrow["{k_s}"{description}, from=3-1, to=3-3] \arrow["{\eta _s}"{description}, shorten <=11pt, shorten >=11pt, Rightarrow, from=1-3, to=3-1] \end{tikzcd}\] with $g_s\in I$ to an $S$-indexed set, then form the 2-pushout of $\sqcup g_s$ and $\sqcup h_s$ and induce $\rho _{j+1}$: \[\begin{tikzcd} {\sqcup \mathcal{A}_s} && {\mathcal{Z}_j} \\ \\ {\sqcup \mathcal{B}_s} && {\mathcal{Z}_{j+1}} \\ &&& {\mathcal{Y}} \arrow["{\sqcup h_s}", from=1-1, to=1-3] \arrow["{\sqcup g_s}"', from=1-1, to=3-1] \arrow[from=3-1, to=3-3] \arrow["{i_{j,j+1}}"', from=1-3, to=3-3] \arrow[""{name=0, anchor=center, inner sep=0}, "{\rho _j}", curve={height=-12pt}, from=1-3, to=4-4] \arrow[""{name=1, anchor=center, inner sep=0}, "{\sqcup k_s}"', curve={height=12pt}, from=3-1, to=4-4] \arrow["{\rho _{j+1}}", dashed, from=3-3, to=4-4] \arrow[shorten <=17pt, shorten >=17pt, Rightarrow, from=1-3, to=3-1] \arrow[shorten <=4pt, shorten >=2pt, Rightarrow, from=0, to=3-3] \arrow[shorten <=2pt, shorten >=3pt, Rightarrow, from=3-3, to=1] \end{tikzcd}\] Note that the composition of the three 2-cells is the natural isomorphism induced by $\{ \eta _s : s\in S \}$. (*) When $j$ is a limit ordinal we form the transfinite composition \[\begin{tikzcd} &&&& {\mathcal{Y}} \\ &&&& {\mathcal{Z}_j} \\ {\mathcal{Z}_0} && {\mathcal{Z}_1} && \dots \arrow["{i_{0,1}}"{description}, from=3-1, to=3-3] \arrow["{i_{1,2}}"{description}, from=3-3, to=3-5] \arrow["{i_{0,j}}"{description, pos=0.4}, from=3-1, to=2-5] \arrow["{i_{1,j}}"{description, pos=0.4}, from=3-3, to=2-5] \arrow["{\rho _0}"{description}, curve={height=-12pt}, from=3-1, to=1-5] \arrow["{\rho_1}"{description}, curve={height=-6pt}, from=3-3, to=1-5] \arrow["{\rho _j}"', dashed, from=2-5, to=1-5] \end{tikzcd}\] Let $\lambda $ be a cardinal, such that domains of $I$ are $\lambda $-small. The composition $\mathcal{X} \xrightarrow{i_{0,\lambda}} \mathcal{Z}_{\lambda} \xrightarrow{\rho _{\lambda}} \mathcal{Y}$ is isomorphic to $f$ and $i_{0,\lambda} \in I$-cell by the previous propositions. It remains to prove that $\rho _\lambda \in I$-inj. Take a square \[\begin{tikzcd} {\mathcal{A}} && {\mathcal{Z}_\lambda} \\ \\ {\mathcal{B}} && {\mathcal{Y}} \arrow["h"{description}, from=1-1, to=1-3] \arrow["{\rho _\lambda}"{description}, from=1-3, to=3-3] \arrow["g"{description}, from=1-1, to=3-1] \arrow["k"{description}, from=3-1, to=3-3] \arrow["\eta"{description}, shorten <=11pt, shorten >=11pt, Rightarrow, from=1-3, to=3-1] \end{tikzcd}\] As $\mathcal{A}$ is $\lambda $-small, $h$ factors through some stage $\mathcal{Z}_j$ (up to isomorphism). This means, that the back face of the left cube in \[\begin{tikzcd} {\mathcal{A}} &&& {\mathcal{Z}_{j}} && {\mathcal{Z}_{\lambda}} \\ & {\sqcup \mathcal{A}_s} & {\mathcal{Z}_{j}} \\ & {\sqcup \mathcal{B}_s} & {\mathcal{Z}_{j+1}} \\ {\mathcal{B}} &&& {\mathcal{Y}} \arrow["{h'}"{description}, from=1-1, to=1-4] \arrow["{\rho _j}"{description, pos=0.7}, from=1-4, to=4-4] \arrow["g"{description}, from=1-1, to=4-1] \arrow["k"{description}, from=4-1, to=4-4] \arrow["{\sqcup g_s}"', from=2-2, to=3-2] \arrow[from=3-2, to=3-3] \arrow[from=2-3, to=3-3] \arrow["{\sqcup h_s}", from=2-2, to=2-3] \arrow[from=1-1, to=2-2] \arrow[from=4-1, to=3-2] \arrow[from=3-3, to=4-4] \arrow[Rightarrow, no head, from=2-3, to=1-4] \arrow["{i_{j,\lambda }}"{description}, from=1-4, to=1-6] \arrow["{i_{j,\lambda }}"{description, pos=0.6}, from=2-3, to=1-6] \arrow["{i_{j+1,\lambda }}"{description}, from=3-3, to=1-6] \arrow["{\rho _{\lambda}}"{description, pos=0.7}, from=1-6, to=4-4] \arrow["{\rho _j}"{description}, color={rgb,255:red,110;green,110;blue,110}, curve={height=-6pt}, from=2-3, to=4-4] \arrow["{\sqcup k_s}"{description}, color={rgb,255:red,110;green,110;blue,110}, curve={height=6pt}, from=3-2, to=4-4] \arrow["h"{description}, curve={height=-12pt}, from=1-1, to=1-6] \end{tikzcd}\] was considered in the formation of $\mathcal{Z}_{j+1}$. This face is just the gluing of \[\begin{tikzcd} {\mathcal{A}} &&& {\mathcal{Z}_{j}} \\ & {\mathcal{Z}_{\lambda}} \\ {\mathcal{B}} &&& {\mathcal{Y}} \arrow["{h'}"{description}, from=1-1, to=1-4] \arrow["g"{description}, from=1-1, to=3-1] \arrow["k"{description}, from=3-1, to=3-4] \arrow[""{name=0, anchor=center, inner sep=0}, "{\rho _j}"{description}, from=1-4, to=3-4] \arrow[""{name=1, anchor=center, inner sep=0}, "h"{description}, from=1-1, to=2-2] \arrow["{i_{j,\lambda}}"{description}, from=1-4, to=2-2] \arrow["{\rho _\lambda}"{description}, from=2-2, to=3-4] \arrow[shorten <=4pt, shorten >=8pt, Rightarrow, from=2-2, to=3-1] \arrow[shorten <=17pt, shorten >=22pt, Rightarrow, from=0, to=2-2] \arrow[shorten <=36pt, shorten >=22pt, Rightarrow, from=1-4, to=1] \end{tikzcd}\] By (*) the left cube is a commutative (identical) 3-cell, and so is the cone over the $\mathcal{Z}_n$-s. Hence the lift $\mathcal{B}\to \sqcup \mathcal{B}_s \to \mathcal{Z}_{j+1} \to \mathcal{Z}_\lambda $ is a splitting of $\eta $. \end{proof} In \cite{filtered} there is an explicit description for filtered 2-colimits in $\mathbf{Cat}$. As a special case we get the following description for the homotopy colimit of the sequence $\mathcal{C}_0\xrightarrow{F_{0,1}} \mathcal{C}_1 \xrightarrow{F_{1,2}} \dots$. Its class of objects is the disjoint union of that of the $\mathcal{C}_i$'s, and an arrow from $(x,i)$ to $(y,j)$ (with $x\in \mathcal{C}_i$ and $y\in \mathcal{C}_j$) is the equivalence class of an arrow $F_{i,k}(x)\xrightarrow{f} F_{j,k}(y)$, where $f$ and $f':F_{i,k'}(x)\to F_{j,k'}(y)$ are equivalent if (assuming $k<k'$) we have $F_{k,k'}(f)=f'$. The induced map in \[\begin{tikzcd} &&& {\mathcal{D}} & {G_k(F_{i,k}(x))} & {G_k(F_{j,k}(y))} \\ &&&& {G_i(x)} & {G_j(y)} \\ &&& {\mathcal{C}} & {(x,i)} & {(y,j)} \\ {\mathcal{C}_0} && {\mathcal{C}_1} && \dots \arrow["{F_{0,1}}", from=4-1, to=4-3] \arrow["{F_{1,2}}", from=4-3, to=4-5] \arrow[from=4-1, to=3-4] \arrow[from=4-3, to=3-4] \arrow[""{name=0, anchor=center, inner sep=0}, "{G_0}", curve={height=-12pt}, from=4-1, to=1-4] \arrow[""{name=1, anchor=center, inner sep=0}, "{G_1}"{pos=0.4}, curve={height=-6pt}, from=4-3, to=1-4] \arrow[dashed, from=3-4, to=1-4] \arrow[""{name=2, anchor=center, inner sep=0}, "{[f]}"', from=3-5, to=3-6] \arrow["{(\eta _{i,k})_x}", from=2-5, to=1-5] \arrow["f", from=1-5, to=1-6] \arrow["{(\eta _{j,k}^{-1})_y}", from=1-6, to=2-6] \arrow[""{name=3, anchor=center, inner sep=0}, dashed, from=2-5, to=2-6] \arrow["{\eta_{0,1}}", shift left=5, shorten <=14pt, shorten >=3pt, Rightarrow, from=0, to=1] \arrow[shorten <=4pt, shorten >=4pt, maps to, from=2, to=3] \end{tikzcd}\] makes the diagram strictly commute when $G_i$'s form a strict cocone, hence we got that this $\mathcal{C}$ is isomorphic to the 1-categorical colimit described in section 3. As it was proved to be coherent, we have that transfinite compositions (of strict sequences) in the 2-categorical sense can be chosen to be 1-categorical colimits. Therefore in the inductive proof of Theorem \ref{smallob} the sequence $\mathcal{Z}_0 \xrightarrow{i_{0,1}} \dots $ can be chosen to be strict, so any $\lambda $ with $cf(\lambda )> sup \{|dom(f)| :f\in I\}$ works. Finally we proved: \begin{theorem} Let $I$ be a small set of coherent functors. Given a coherent functor $\mathcal{C}\xrightarrow{m} \mathcal{E}$ it is isomorphic to a composition $\mathcal{C}\xrightarrow{f}\mathcal{D}\xrightarrow{g}\mathcal{E}$ where $f\in I$-cell and $g\in I$-inj. In particular $f\in I$-cof. \end{theorem} \printbibliography \end{document} \[\begin{tikzcd} &&&& {\mathcal{Y}} \\ \\ &&&& {\mathcal{Z}_j} \\ \\ {\mathcal{Z}_0} && {\mathcal{Z}_1} && \dots \arrow["{i_{0,1}}"{description}, from=5-1, to=5-3] \arrow["{i_{1,2}}"{description}, from=5-3, to=5-5] \arrow["{i_{0,j}}"{description, pos=0.4}, from=5-1, to=3-5] \arrow["{i_{1,j}}"{description, pos=0.4}, from=5-3, to=3-5] \arrow["{\rho _0}"{description}, curve={height=-12pt}, from=5-1, to=1-5] \arrow["{\rho_1}"{description}, curve={height=-6pt}, from=5-3, to=1-5] \arrow["{\rho _j}"{description}, dashed, from=3-5, to=1-5] \end{tikzcd}\]

0f3b8f9133a40efe57ed24da1ed968a7ee2ab7b7

\section{Measurement of double occupancies} \label{sec:PA} We measure double occupancies (doublons) in the 3D lattice exploiting a one-color photoassociation (PA) process which transfers the atoms in doubly-occupied lattice sites into highly excited molecular states \cite{jordens2008}. The photoassociated atoms are rapidly lost from the lattice due to the fast decay of the molecular states, leaving a sample consisting only of singly-occupied sites. The number of double occupancies is thus inferred by first measuring the total number of atoms in the sample and then subtracting the number of atoms in singly occupied lattice sites remaining after the PA pulse. For the regime of atom number considered in this work we expect the number of triply-occupied sites to be negligible. Our measurement procedure starts with a rapid freeze of the atomic density distribution that we realize increasing the lattice depth up to $V_0=25 E_{rec}$ ($E_{rec}=h^2/8md^2$, where $h$ is the Planck constant, $m$ the atomic mass and $d$ the lattice spacing) in a time 1.5 ms. At the final lattice depth tunnelling between lattice sites is completely negligible. Two different schemes are then employed for the PA excitation depending on the specific measurement that we want to perform. In order to measure the total number of doublons independently on their spin composition (results of Fig. \ref{fig2} of the main text), a 5 ms long, $\sigma^-$ polarized PA pulse is shone on the atomic sample with an intensity $I_{PA}=90$ mW/mm$^2$. The PA pulse excites a molecular line that is 796.2 MHz red-detuned with respect to the ${^1S_0}(F=5/2) \rightarrow {^3P_1}(F=7/2)$ atomic transition \cite{taie2012}. During the process, atoms are subjected to a low magnetic field $B$ = 3 G which defines a quantization axis for the spin, but is low enough not to unveil the Zeeman substructure of the molecular line, as it is shown in the spectrum reported in Fig. \ref{fig:photoassociation}a. In the presence of a higher external magnetic field, the PA transition frequency is shifted accordingly to the projection $M_T$ of the molecular total angular momentum on the quantization axis \cite{Han2018}. $M_T$ is a quantum number conserved during the PA process, with a value given by the relation $M_T=m_1+m _2+\sigma$, where $m_{1,2}$ are the nuclear spin projections of the the two atoms composing the molecule and $\sigma=-1$ is the angular momentum transferred by the PA photon. As long as the sum $m_1+m_2$ unambiguously characterizes an atomic pair, the molecular Zeeman substructure can thus be employed as a tool to distinguish between doublons with different spin flavours. \begin{figure} \includegraphics[width=\linewidth]{figs1bis} \caption{{\bf Detection of double occupancies.} Photoassociation spectroscopy on an SU(3) atomic sample (a) and on different SU(2) mixtures (b). The plots report the number of atoms remaining in the sample after a 5 ms long PA pulse at low laser intensity (unsaturated regime).} \label{fig:photoassociation} \end{figure} Taking this into consideration, in order to detect only doublons with a particular spin composition $\{m_1,m_2\}$, we apply a magnetic field $B=66$ G, which induces a Zeeman shift of the order of several MHz between molecular lines with different $M_T$ projection. In this case the spectrum is complicated by the presence of a plethora of features originated by the Zeeman splitting of different transitions, which makes it difficult to label individual lines. The association of the PA lines to a particular spin flavour has been realized preparing different SU(2) mixtures and acquiring individual spectra for each of them, as it is shown in Fig. \ref{fig:photoassociation}b. In particular, at this magnetic field, we identify three strong PA lines at 778.3 MHz, 792.8 MHz and 802.0 MHz which are associated to doublons with spin compositions $|12\rangle=\{+5/2, +1/2\}$, $|23\rangle=\{+1/2, -5/2\}$ and $|31\rangle=\{-5/2, +5/2\}$, respectively. For the measurements reported in Fig. \ref{fig4} of the main text, $\gamma(12)$ is calculated from a spin-resolved measurement of doublons in the $|12\rangle$ and $|31\rangle$ channels as $\gamma(12)=N_d(12)/(N_d(12)+2N_d(31))$, taking into account $N_d(23)=N_d(31)$. This assumption is justified by the equal population of the two flavours $|1\rangle$ and $|2\rangle$ after the adiabatic preparation sequence described in Sec. \ref{sec:adiabaticpreparation}, as shown in Fig. \ref{fig:dressedstatestability}. \section{Data analysis} The number of double occupancies is obtained from a direct measurements of the total number of atoms, with and without photoassociation (PA), obtained through standard time-of-flight absorption imaging (see Sec. \ref{sec:PA}). Measurements with and without PA are alternated in time and the number of double occupancies $N_d$ is obtained by the difference between close pairs, in order to compensate for fluctuations in the initial number of atoms $N$ (without PA) due to slow environmental drifts. For the same set of parameters $\left\{U/D,\Omega/D\right\}$ we acquire from 40 to 80 $(N,N_d)$ pairs. The experimental points in Figs. \ref{fig2} and \ref{fig4} are obtained by a bootstrapping method in which the raw data are resampled uniformly over a chosen subrange of $N$, in bins of width $\delta N = 4\times10^3$ for Fig. \ref{fig2}a, and over an extended range $N=(5\ldots 35)\times10^3$ for Fig. \ref{fig2}b,c and Fig. \ref{fig4}. The error bars are obtained as the standard deviation of the mean values for different resamplings of the same $N$ interval. \section{Unsupervised machine-learning analysis} In addition to the data analysis presented in the main text and detailed in S.II, we have performed an additional analysis of the flavour-selective double occupancies, based on an unsupervised machine-learning approach. The raw datasets are large, with at least 200 acquisitions of pairs $(N, N_d)$ for each set of experimental parameters. Following the well-established \textit{K-Means method} \cite{andrewNG}, we consider the first component of each pair, i.e. the number of atoms $N$, and we set (as the only constraint) the number $K$ of groups to divide them into. Once the groups (clusters) are defined, the mean values for both the number of atoms $N$ and the number of doublons $N_d$ for each cluster are evaluated. The error bars are estimated with a bootstrapping procedure. \begin{figure}[t!] \begin{centering} \includegraphics[width=0.46\textwidth ]{unsup_1.pdf} \caption{\textbf{ Comparison between different data analysis methods.} The number of doublons in the $|12\rangle$ channel is shown as a function of the total atom number $N$ for $U/D = 2.6$ and $\Omega/D = 0.8$, for different data analysis methods. Small points are the raw $(N,N_d)$ data, with different colours representing the clusters of attribution after the application of the unsupervised K-Means method with $K = 10$ groups. Pink points represent the mean values for the different K-Means clusters. Blue points are obtained after the analysis explained in S.II with fixed bin widths.} \label{fig:unComp} \end{centering} \end{figure} In Fig. \ref{fig:unComp} we compare the points obtained with the analysis presented in the main text and those given by the unsupervised K-Means analysis, taking the same number of clusters. The good agreement between them ensures that the specific choice of data analysis procedure does not introduced biases on the results. We note that the K-Means method implies only a constraint on the number of clusters, allowing for a different width among them, differently from the main analysis procedure, where the bin size was fixed. As the datasets for the flavour-selective double occupancies are rather large, we can increase the number of groups $K$ to pinpoint the specific range of atom numbers above which doublons form. The flavour-selective averages obtained with the K-Means analysis are fitted with the piecewise function: \begin{equation}\label{eq:piecewise} N_d= \begin{cases} 0 & N \leq N_0 \\ A (N-N_0) & N>N_0 \end{cases} \end{equation} where $N_0$ defines the threshold atom number below which atoms are localised. In Fig. \ref{fig:piecewise} we show the K-Means averages for the coupled flavours $\lvert 1 2 \rangle$ and the uncoupled ones $\lvert 2 3 \rangle$, together with the results of the fit. The difference between the fitted thresholds for the two doublon flavours $\Delta N_0 = N_0(12)-N_0(23)$ is shown in the inset of Fig. \ref{fig:piecewise} for different values of $\Omega/D$. \begin{figure}[b] \begin{centering} \includegraphics[width=0.47\textwidth ]{unsup2} \caption{\textbf{Flavour-selective localization.} The number of doublons in different flavour channels is shown as a function of $N$ for $U/D = 2.6$ and $\Omega/D = 0.8$, after a K-Means analysis with $K = 18$ clusters. Pink points show the average number of doublons formed by the coupled flavours $\lvert 1 2 \rangle$, while purple points refer to the uncoupled flavours $\lvert 2 3 \rangle$. Solid lines are the fits performed with the piecewise function in Eq. \ref{eq:piecewise}. In the inset we show the difference between the fitted thresholds $\Delta N_0 = N_0(12)-N_0(23)$ as a function of the coupling strength $\Omega$.} \label{fig:piecewise} \end{centering} \end{figure} This analysis demonstrates that, by increasing the Raman coupling intensity (i.e. lifting the flavour degeneracy more), the threshold for $|12\rangle$ doublons moves to a higher number of atoms, showing an increased localisation in the centre of the trap for the coupled flavours. This result, derived from the unsupervised machine-learning analysis, is consistent with the measurements reported in Fig. \ref{fig4}, and provides another strong indication in support of flavour-selective localization for the symmetry-broken SU(3) Hubbard Hamiltonian. \section{Implementation of Raman coupling} The coherent coupling between states $|1\rangle$ ($m=5/2$) and $|2\rangle$ ($m=1/2$) is realized by exploiting a two-photon $\sigma^+ / \sigma^-$ Raman transition. The Raman coupling is implemented with two co-propagating $\lambda = 556$ nm laser beams characterized respectively by angular frequencies $\omega$ and $\omega+\delta\omega$. In order to reduce the inelastic photon scattering rate, the two beams are 1.754 GHz blue-detuned with respect to the $^1S_0$ $\rightarrow$ ${^3P_1}\,(F=7/2)$ intercombination transition. A 150 G magnetic field is used to define a quantization axis and to remove the degeneracy between the six states of the $^{173}$Yb ground-state manifold, which are split according to their nuclear spin $m$ by $207 \times m$ Hz/G. The $\sigma^+/\sigma^-$ coupling between $m=+5/2$ and $m=+1/2$ is obtained by setting the polarization of the two beams to be orthogonal with respect to the quantization axis and adjusting the frequency difference $\delta\omega/2\pi$ in order to compensate the Zeeman splitting and the residual Raman light shift between the two states. As explained in detail in the supplementary material of Ref. \onlinecite{Mancini2015}, this choice for the polarization makes the Raman light shifts largely spin-dependent, thus making the $m=+1/2 \leftrightarrow m=-3/2$ coupling nonresonant and effectively excluding state $m=-3/2$ from the dynamics. For a similar reason, also the $m=-5/2 \leftrightarrow m=-1/2$ coupling is nonresonant and state $|3\rangle$ ($m=-5/2$) does not participate to the Raman dynamics. On the basis of the discussion above, the Raman coupling on the basis formed by the three states $\left\{|1\rangle, |2\rangle, |3\rangle \right\}$ can be described by the following 3x3 rotating-wave-approximated Hamiltonian \begin{equation} \label{Raman_Hamiltonian} \hat{H}_R=\frac{\hbar}{2} \begin{pmatrix} \delta & \Omega & 0\\ \Omega& -\delta & 0\\ 0 & 0 & 0 \end{pmatrix} \end{equation} where $\Omega$ is the angular Rabi frequency associated to the Raman coupling and $\delta$ is the two-photon detuning with respect to the resonance frequency for the $|1\rangle \leftrightarrow |2\rangle$ transition. No detuning is indicated for state $|3\rangle$ as it is decoupled from the Raman process. In the resonant case $\delta=0$, $\hat{H}_R$ corresponds to the third term in the many-body Hubbard Hamiltonian in Eq. (\ref{eq:hamiltonian}) of the main text. The Rabi frequency $\Omega$ can be experimentally changed by adjusting the intensities $I_1$ and $I_2$ of the two Raman beams, according to the relation $\Omega \propto \sqrt{I_1 I_2}$. In order to assess the value of $\Omega$ for given values of the Raman beam intensities $I_1$ and $I_2$, we induce resonant Rabi oscillations between states $|1\rangle$ and $|2\rangle$ and determine $\Omega$ from a sinusoidal fit of the experimental data, as shown in Fig. \ref{fig:raman}. \begin{figure} \includegraphics[width=\linewidth]{figs2} \caption{{\bf Coherent Rabi dynamics.} Population dynamics induced by the sudden activation of a resonant Raman coupling between states $|1\rangle$ and $|2\rangle$ for localized particles in a deep optical lattice (no tunnelling). The points show the fractional population of the states $|1\rangle$ (blue), $|2\rangle$ (green) and $|3\rangle$ (red) as a function of time. The resonant Raman coupling is switched on at time $t=0$, when the initial fractional populations for the three states are $2/3$, $0$ and $1/3$, respectively (as in the adiabatic state preparation procedure described in Sec. \ref{sec:adiabaticpreparation}). The solid lines are sinusoidal fits of the experimental data.} \label{fig:raman} \end{figure} \section{Adiabatic state preparation} \label{sec:adiabaticpreparation} A degenerate Fermi gas of $^{173}$Yb atoms, initially confined in a crossed optical dipole trap with harmonic trapping frequencies $\omega_{x,y,z} = 2\pi\times\{55, 96, 73\}$ Hz, is transferred in a simple-cubic 3D optical lattice using a two-step 3s-long adiabatic loading procedure \cite{Mikio2017}. The optical lattice is described by a potential energy $V(x,y,z)=V_0 \left[ \sin^2(\pi x/d)+\sin^2(\pi y/d)+\sin^2(\pi z/d)\right]$, where $d=\lambda/2$ is the lattice spacing and $V_0$ is the lattice energy depth along each of the three principal axes. In the first 2 seconds, the lattice intensity is ramped up with a first spline ramp from $V_0=0$ to $V_0=4 E_{rec}$ ($E_{rec}=h^2/8md^2$ is the recoil energy). After this first step, the lattice intensity is further increased with a 1s-long spline ramp to the final value $V_0$ ranging from $4E_{rec}$ to $15E_{rec}$. During the lattice loading, the depth of the crossed dipole trap is progressively reduced in such a way to keep the harmonic trapping frequencies constant along the three principal axes, independently from the value of $V_0$. For the measurements at $\Omega=0$ the loading procedure described above is applied to a balanced 3-component mixture of $^{173}$Yb atoms in the three spin states $|1\rangle$, $|2\rangle$ and $|3\rangle$. The mixture is prepared before the lattice ramp-up procedure with a sequence of optical pumping pulses on the $^1S_0$ $\rightarrow$ $^3P_1$ transition, following the techniques discussed in Ref. \onlinecite{pagano2014}. The populations of the three spin states are all equal, $N_1=N_2=N_3=N/3$, with experimental imperfections on the order of a few $\%$ at most. For the measurements at $\Omega \ne 0$ the loading procedure starts with a 2-component unbalanced mixture of atoms in states $|1\rangle$ and $|3\rangle$. With a proper choice of optical pumping pulses we adjust the initial populations to be $N_1=2N/3$ and $N_3=N/3$. After the lattice loading, we switch on the Raman beams, initially far detuned from any two-photon transition, and perform an adiatic frequency sweep to bring them resonant with the $|1\rangle \leftrightarrow |2\rangle$ transition. The resonant condition is reached by means of an exponential frequency sweep of the form \begin{equation} \label{Raman_detuning} \delta(t)=\delta_{0} ( e^{-t/\tau} - e^{-T/\tau})/(1 - e^{-T/\tau}) \end{equation} that reduces the two-photon detuning from $\delta_0 \gg \Omega$ to $\delta=0$ in a time $T$ and time constant $\tau$ (see an example in Fig. \ref{fig:adiabloading}a). This procedure is an adiabatic passage that brings an atom in state $|1\rangle$ to the lowest-energy dressed state of the Raman-coupled system $|+\rangle = \left(|1\rangle+|2\rangle\right)/\sqrt{2}$ (see also the sketch in Fig. \ref{fig1} of the main text). We note that, due to the initial spin unbalance and because of the equal-weighted admixture of states $|1\rangle$ and $|2\rangle$ in the Raman-dressed states, at the end of the detuning ramp the population is equally distributed among the spins, $N_1=N_2=N_3=N/3$, as it was natively in the loading protocol employed for $\Omega=0$. \begin{figure}[b!] \includegraphics[width= \linewidth]{figs3} \caption{{\bf State population after adiabatic state preparation}. Fractional state population as a function of time after the adiabatic state preparation sequence. Points are the mean value of 3 independent measurements. Shaded regions represent the standard deviation with respect to the mean value. The slight excess of $|3\rangle$ atoms in this dataset has to be attributed to an imperfect optical pumping before the adiabatic state preparation procedure.} \label{fig:dressedstatestability} \end{figure} The ramp parameters $\delta_0$, $T$ and $\tau$ are chosen according to a numerical analysis in which we solve the time-dependent Schr\"odinger equation associated to the Raman Hamiltonian in Eq. (\ref{Raman_Hamiltonian}), verifying that at the end of the ramp the initial state is effectively projected onto the Hamiltonian lowest energy eigenstate. Experimentally, we check the adiabaticity of this procedure by verifying that the initial unbalanced $|1\rangle$-$|3\rangle$ mixture can be recovered with a reversed detuning ramp following that in Eq. (\ref{Raman_detuning}). To further assess the loading fidelity, we verify the time stability of the spin populations at the end of the loading ramp, as shown in Fig. \ref{fig:dressedstatestability}. In both the cases, we measure population differences of a few $\%$ at most, validating the effectiveness of this loading procedure. \section{Numerical validation of the loading procedure} In order to assess the validity of the adiabatic state preparation at the many-body level, we have developed a numerical simulation based on the exact diagonalization of Eq. (\ref{eq:hamiltonian}) on a reduced-scale version of our system. We consider $N$ three-component fermions in a 1D lattice with $N_s$ sites, retaining all the relevant terms of Eq. (\ref{eq:hamiltonian}): hopping, repulsive interactions, Raman coupling between states $|1\rangle$ and $|2\rangle$. We work in the occupation number basis in which the Hilbert space is constituted by the Fock vectors \begin{equation} \label{Fock_vector} |\psi\rangle = |n_{1,1},...,n_{\alpha,i},...,n_{3,N_s}\rangle \end{equation} where $n_{\alpha,n} =\{0,1\}$ is the occupation number for a particle in internal state $\alpha=\{1,2,3\}$ in the lattice site $i=\{1,...,N_s\}$. In this basis we can write the many-body Hamiltonian as \begin{equation} \label{toy_model_hamiltonian} \begin{aligned} \hat{H} = & -t \sum_{i,\alpha} \left(\hat{c}^{\dag}_{\alpha,i} \hat{c}_{\alpha,i+1} + \mathrm{h.c.}\right) + U \sum_{i, \alpha,\beta \neq \alpha} \hat{n}_{\alpha i} \hat{n}_{\beta i} + \\ & +\frac{\Omega}{2} \sum_{i}\left( \hat{c}^{\dag}_{1,i} \hat{c}_{2,i} + \mathrm{h.c.}\right)+ \frac{\delta}{2} \sum_{i} \left( \hat{n}_{1,i} - \hat{n}_{2,i} \right) \end{aligned} \end{equation} where we have added the last term to take into account the detuning of the Raman coupling (already introduced in the rotated-wave-approximated Hamiltonian of Eq. (\ref{Raman_Hamiltonian})), that is crucial in the state preparation protocol. \begin{figure*}[t!] \includegraphics[width= \linewidth]{figadiab} \caption{{\bf Numerical analysis of adiabatic state preparation.} {\bf a)} Raman detuning ramp used for the preparation of the initial state. The parameters refer to the actual ramp used in the experiment for the preparation of the system for $U=2.6D$ and $\Omega=0.35D$. {\bf b)} Calculated fidelity $|\langle\Psi_f|\Psi(t)\rangle|^2$ of the time-evolved state $|\Psi(t)\rangle$ with respect to the target ground state at the end of the ramp $|\Psi_f\rangle$. The calculation has been performed for a small system of 3 particles in 3 sites for $U=2.6D$, $\Omega=0.35D$ and the same detuning ramp shown in panel a). {\bf c)} Full spectrum of the system for $U=2.6D$ and $\Omega=0.35D$ and different Raman detunings (see text fore more details).} \label{fig:adiabloading} \end{figure*} We simulate the effect of our loading procedure for a system composed by $N=3$ particles in a lattice with $N_P=3$ sites. In this case, the dimension of the Hilbert space associated to the system is $\binom{3\,N_S}{N}=84$. Scaling the problem to bigger lattices is possible but computationally expensive. In order to simulate the loading scheme with $\Omega \ne 0$, we fix the initial populations to be $\{N_1,N_2,N_3\}=\{2,0,1\}$, where $N_{\alpha}$ refers to the total number of particles in state $\alpha$. We determine the ground state $|\Psi_0\rangle$ before the Raman loading as the lowest-energy eigenstate of the Hamiltonian in Eq. (\ref{toy_model_hamiltonian}), with $\Omega$=0, that is compatible with the constraint on the populations (i.e. $\{2,0,1\}$). The Raman loading is then simulated by solving the time-dependent Schr{\"o}dinger equation in the presence of all the terms of Eq. (\ref{toy_model_hamiltonian}), where $\Omega$ is kept fixed at the final value and $\delta$ is swept from $\delta \gg \Omega$ to $\delta=0$ according to Eq. (\ref{Raman_detuning}), as discussed in Sec. \ref{sec:adiabaticpreparation}. In order to validate the loading procedure we calculate the fidelity between the time-evolved state $|\Psi(t)\rangle=e^{-i \hat{H} t/\hbar} |\Psi_0\rangle$ and the target wavefunction, that is defined as the lowest-energy eigenstate $|\Psi_f\rangle$ of the final Hamiltonian with $\delta=0$ that is compatible with the constraint on the populations $N_1+N_2=2$ and $N_3$=1. An example is shown in Fig. \ref{fig:adiabloading}b for the actual detuning ramp used in the experiment to prepare the state at $U=2.6D$ and $\Omega=0.35D$. Fig. \ref{fig:adiabloading}c shows the full spectrum of the system as a function of the Raman detuning. There it is evident that the ground state of the system at large detuning $|\Psi_0\rangle$ is adiabatically connected with the ground state $|\Psi_f\rangle$ on resonance. We note that, even if the system initially starts with an unbalanced mixture of two species only, after this procedure it behaves as a mixtures of three species $N_1=N_2=N_3=1$ all interacting among themselves: in particular, in the presence of tunnelling and interactions between the atoms, the state rotation induced by the detuning sweep does not happen uniformly in all the sites of the lattice, making the atoms initially in state $|1\rangle$ distinguishable and therefore, interacting, at the end of the state preparation sequence. \section{Dynamical Mean-Field Theory Calculations} We solve the model in Eq. (\ref{eq:hamiltonian}) using Dynamical Mean-Field theory (DMFT)\cite{rev_dmft}, a non-perturbative theoretical method which maps a lattice model onto an effective impurity model. The interaction of the site with the rest of the lattice is approximated by an effective dynamical bath which is determined self-consistently. This allows to fully capture the quantum dynamics while freezing spatial fluctuations beyond mean-field. We solve the model in a Bethe lattice of bandwidth $W$, which is known to account correctly for the physics of three-dimensional lattices. The impurity model is solved using an exact-diagonalization solver developed at SISSA\cite{capone_prbed} which requires to describe the bath in terms of a finite number of levels $N_b$ that we fix to 9 in the calculations reported in this work. DMFT is naturally formulated in a grandcanonical ensemble, where the total density is fixed by a chemical potential $\mu$ and the occupations of the various flavours are not fixed. In order to {enforce the experimental population constraint} $N_1 + N_2 = 2 N_3$ in the presence of the coupling $\Omega$, we have to include an external field which counterbalances the tendency to favour (for total occupation of 1) the occupation of the coupled flavours $|1\rangle$ and $|2\rangle$. Therefore the Hamiltonian is supplemented by the terms \begin{equation} \hat{H}_{fields} = \sum_i \left( -\mu\sum_{\alpha} \hat{n}_{\alpha i} +h \left( \hat{n}_{1i} + \hat{n}_{2i} -2\hat{n}_{3i} \right) \right), \label{hfields} \end{equation} where $\mu$ and $h$ have to be determined self-consistently to reach a total occupation of 1 fermion per site. For a uniform infinite lattice, we have drawn a phase diagram based on the flavour-resolved quasiparticle weight $Z_{\alpha}$, which is 1 for a non-interacting fermion and it is reduced by the interactions. The global Mott transition is reached when all the $Z_{\alpha}$ vanish, while a sharp crossover is observed between a standard metal and a region of selective correlation where the quasiparticle weight of the coupled species rapidly drops to a value smaller than 0.05. Besides the calculations for a uniform lattice, we also have used DMFT to estimate the effect of the harmonic trapping potential \begin{equation} \hat{V}_T=\kappa \sum_{i,\alpha} R_i^2 \hat{n}_{\alpha i}, \label{hfields} \end{equation} where $R_i$ is the distance of the site i from the center of the trap. In order to reduce the computational cost, we have employed a local-density approximation (LDA), which assumes that the local properties of the system can be computed from a uniform model with the local values of the physical parameters, in the present case of the potential given by the harmonic trap. This amounts to have a local value of the chemical potential $\mu' = \mu -\kappa \sum_{i,\alpha} R_i^2$. In order to build the global observables for the trapped system, we have solved the model given by Eqs. (\ref{eq:hamiltonian}) and (\ref{hfields}) for a wide range of values of $\mu'$ spanning the whole range of densities from 0 to 3. For every value of $\mu'$ we have found the value of $h$ which gives $N_1 = N_2 = N_3$. The global observables (total number of particles, doublons, flavour-selective doublons) are then obtained integrating the local counterparts over the whole system. In this way we obtain the plots of doublons as a function of the total number of particles shown in the manuscript. \section{Density Matrix Renormalization Group Calculations} We also solve the model in Eq. (\ref{eq:hamiltonian}) by means of a Density Matrix Renormalization Group (DMRG) calculation powered by the ITensor library \cite{itensor}. We fix the size of the system as $L=30$ and the total number of particles as $N=21$. We simulate the internal states as effective physical sites with the redefinition of proper tunnelling and interaction couplings, which results in a total of $N_{L}=90$ lattice sites. We find the ground state of the Hamiltonian in Eq. (\ref{eq:hamiltonian}) by performing $35-50$ DMRG sweeps, with a maximum bond dimension in the interval $(20,200)$. To enforce particle number conservation in the Raman-coupled case, we introduce an additional chemical potential defined as the $h$ term in Eq. (\ref{hfields}). The number of doubly occupied sites shown in Fig. \ref{fig3} also includes the number of triple occupancies (giving a maximum contribution of $\sim 10\%$ at small values of $U$, $\Omega$).

9e9ccc3ad1b5df81c012bded535e3b6c0f21af6c

\section{Introduction} Effectively immune sets, introduced by Smullyan in 1964 \cite{MR180485}, are well-known in computability as one of the incarnations of diagonal non-computability, first made famous by Arslanov's completeness criterion. A set $A\subseteq\omega$ is \emph{effectively immune} if there is a computable function $h$ such that $\abs{W_e}\le h(e)$ whenever $W_e\subseteq A$, where $\{W_e\}_{e\in\omega}$ is a standard enumeration of the computably enumerable (c.e.) sets. There is a more obvious effectivization of immunity (the lack of infinite computable subsets), however: \emph{constructive immunity}, introduced by Xiang Li \cite{MR723334} who actually (and inconveniently) called it ``effective immunity''. \begin{definition} A set $A$ is \emph{constructively immune} if there exists a partial recursive $\psi$ such that for all $x$, if $W_x$ is infinite then $\psi(x)\downarrow$ and $\psi(x)\in W_x\setminus A$. \end{definition} The Turing degrees of constructively immune sets and the related $\Sigma^0_1$-dense sets have not been considered before in the literature, except that Xiang Li implicitly showed that they include all c.e.~degrees. We prove in \Cref{sec:prevalent} that the Turing degrees of $\Sigma^0_1$-dense sets include all non-$\Delta^0_2$ degrees, all high degrees, and all c.e.~degrees. We do not know whether they include \emph{all} Turing degrees. The history of the study of constructive immunity seems to be easily summarized. After Xiang Li's 1983 paper, Odifreddi's 1989 textbook \cite{MR982269} included Li's results as exercises, and Calude's 1994 monograph \cite{MR1323429} showed that the set $RAND^C_t=\{x: C(x)\ge \abs{x}-t\}$ is constructively immune, where $C$ is Kolmogorov complexity. Schafer 1997 \cite{MR1654312} further developed an example involving minimal indices, and Brattka 2002 \cite{MR2059846} gave one example in a more general setting than Cantor space. Finally in 2008 Ferbus-Zanda and Grigorieff proved an equivalence with constructive $\Sigma^0_1$-density. \begin{definition}[Ferbus-Zanda and Grigorieff \cite{ferbuszanda2008refinment}]\label{def_eff_susc} A set $A\subseteq\omega$ is \emph{$\Sigma^0_1$-dense} if for every infinite c.e. set $C$, there exists an infinite c.e. set $D$ such that $D \subseteq C$ and $D \subseteq A$. If there is a computable function $f:\omega\to\omega$ such that for each $W_e$, $W_{f(e)}\subseteq A\cap W_e$, and $W_{f(e)}$ is infinite if $W_e$ is infinite, then $A$ is \emph{constructively $\Sigma^0_1$-dense}. \end{definition} We should note that while the various flavors of immune sets are always infinite by definition, Ferbus-Zanda and Grigorieff do not require $\Sigma^0_1$-dense sets to be co-infinite. The $\Sigma^0_1$-dense sets form a natural $\Pi^0_4$ class in $2^\omega$ that coincides with the simple sets on $\Delta^0_2$ but is prevalent (in fact exists in every Turing degree) outside of $\Delta^0_2$ by \Cref{prevalent} below. \section{$\Sigma^0_1$-density} To show that there exists a set that is $\Sigma^0_1$-dense, but not constructively so, we use Mathias forcing. A detailed treatment of the computability theory of Mathias forcing can be found in \cite{MR3210076}. \begin{definition} A \emph{Mathias condition} is a pair $(d, E)$ where $d, E \subseteq \omega$, $d$ is a finite set, $E$ is an infinite computable set, and $\max(d) < \min(E)$. A condition $(d_2, E_2)$ \emph{extends} a condition $(d_1, E_1)$ if \begin{itemize} \item $d_1=d_2\cap(\max d_1+1)$, i.e., $d_1$ is an initial segment of $d_2$, \item $E_2$ is a subset of $E_1$, and \item $d_2$ is contained in $d_1 \cup E_1$. \end{itemize} A set $A$ is \emph{Mathias generic} if it is generic for Mathias forcing. \end{definition} \begin{theorem}\label{mathias} If $A$ is Mathias generic, then \begin{enumerate} \item $\omega\setminus A$ is $\Sigma^0_1$-dense. \item $\omega\setminus A$ is not constructively $\Sigma^0_1$-dense. \end{enumerate} \end{theorem} \begin{proof} \noindent 1. Let $W_e$ be an infinite c.e.~set. Let $(d,E)$ be a Mathias condition. \indent Case (i): $E\cap W_e$ is finite. Then for any Mathias generic $A$ extending the condition $(d,E)$, $\omega\setminus A$ contains an infinite subset of $W_e$, in fact a set of the form $W_e\setminus F$ where $F$ is finite. \indent Case (ii): $E\cap W_e$ is infinite. Then $E\cap W_e$ is c.e., hence has an infinite computable subset $D$. Write $D=D_1\cup D_2$ where $D_1, D_2$ are disjoint infinite c.e.~sets. The condition $(d,D_1)$ extends $(d,E)$ and forces a Mathias generic $A$ extending it to be such that $\omega\setminus A$ has an infinite subset in common with $W_e$, namely $D_2$. We have shown that for each infinite c.e.~set $W_e$, each Mathias condition has an extension forcing the statement that a Matias generic $A$ satisfies \[ \text{$\omega\setminus A$ has an infinite c.e.~subset in common with $W_e$.}\tag{*} \] Thus by standard forcing theory it follows that each Mathias generic satisfies ($*$). \noindent 2. Let $f$ be a computable function. It suffices to show that for each Mathias generic $A$, there exists an $i$ such that $W_i$ is infinite and $W_{f(i)}$ is either finite, or not a subset of $W_i$, or not a subset of $\overline A$. For this, as in (1) above it suffices to show that for each condition $(d,D)$ there exists a condition $(d',E')$ extending $(d,E)$ and an $i$ such that $W_i$ is infinite and $W_{f(i)}$ is either finite, or not a subset of $W_i$, or not a subset of $\overline A$ for any $A$ extending $(d',D')$. Let $(d,E)$ be a Mathias condition and write $D=W_i$. If $W_{f(i)}$ is finite or not a subset of $W_i$ then we are done. Otherwise there exists a condition $(d',E')$ extending $(d,E)$ such that $E'\cap W_{f(i)}$ is nonempty. This can be done by a finite extension (making only finitely many changes to the condition). \end{proof} \begin{theorem}[{\cite[Proposition 3.3]{ferbuszanda2008refinment}}]\label{p:immune} A set $Z\subseteq\omega$ is constructively immune if and only if it is infinite and $\omega\setminus Z$ is constructively $\Sigma^0_1$-dense. \end{theorem} Since Ferbus-Zanda and Grigorieff's paper has not gone through peer review, we provide the proof. \begin{proof} $\Leftarrow$: Let the function $g$ witness that $\omega\setminus Z$ is constructively $\Sigma^0_1$-dense. Define a partial recursive function $\varphi$ by stipulating that $\varphi(i)$ is the first number in the enumeration of $W_{g(i)}$, if any. $\Rightarrow$: Define a partial recursive function $\mu(i,n)$ by \begin{itemize} \item $\mu(i,0)=\varphi(i)$; \item $\mu(i,n+1)=\varphi(i_n)$, where $i_n$ is such that $W_{i_n}=W_i\setminus\{\mu(i,m) : m\leq n\}$. \end{itemize} Let $g$ be total recursive so that $W_{g(i)}=\{\mu(i,m) : m\in\omega\}$. If $W_i$ is infinite then all $\mu(i,m)$'s are defined and distinct and belong to $W_i\cap Z$. Thus, $W_{g(i)}$ is an infinite subset of $W_i\cap Z$. \end{proof} Recall that a c.e.~set is \emph{simple} if it is co-immune. \begin{theorem}[Xiang Li \cite{MR723334}]\label{xiang} Let $A$ be a set and let $\{\phi_x\}_{x\in\omega}$ be a standard enumeration of the partial computable functions. \begin{enumerate} \item\label{xiang-1} If $A$ is constructively immune then $A$ is immune and $\overline A$ is not immune. \item\label{xiang-2} If $A$ is simple then $\overline A$ is constructively immune. \item\label{xiang-3} $\{x: (\forall y)(\phi_x=\phi_y \to x \le y)\}$ is constructively immune. \end{enumerate} \end{theorem} \subsection{Numberings} A \emph{numbering} of a countable set $\mathcal A$ is an onto function $\nu:\omega\to\mathcal A$. The theory of numberings has a long history \cite{MR1720731}. Numberings of the set of rational numbers $\mathbb Q$ provide an application area for $\Sigma^0_1$-density. Rosenstein \cite[Section 16.2: Looking at $\mathbb Q$ effectively]{MR662564} discusses computable dense subsets of $\mathbb Q$. Here we are mainly concerned with noncomputable sets. \begin{proposition}\label{co-immune} Let $A \subseteq \omega$. The following are equivalent: \begin{enumerate} \item $\nu(A)$ is dense for every injective computable numbering $\nu$ of $\mathbb{Q}$; \item $A$ is co-immune. \end{enumerate} \end{proposition} \begin{proof} \indent (1)$\implies$(2): We prove the contrapositive. Suppose $\overline{A}$ contains an infinite c.e. set $W_e$. Consider a computable numbering $\nu$ that maps $W_e$ onto $[0,1]\cap\mathbb Q$. Then $\nu(A)$ is disjoint from $[0,1]$ and hence not dense. \indent (2)$\implies$(1): We again prove the contrapositive. Assume that $\nu(A)$ is not dense for a certain computable $\nu$. Let $\{x_n:n\in\omega\}$ be a converging infinite sequence of rationals disjoint from $\nu(A)$. Then $\{\nu^{-1}(x_n):n\in\omega\}$ is an infinite c.e.~subset of $\overline{A}$. \end{proof} \begin{definition} A subset $A$ of $\mathbb{Q}$ is \emph{co-nowhere dense} if for each interval $[a,b] \subseteq \mathbb{Q}$, $[a', b'] \subseteq A$ for some $[a', b'] \subseteq [a,b]$. \end{definition} \begin{proposition}\label{co-finite} A set is co-nowhere dense under every numbering iff it is co-finite. \end{proposition} \begin{proof} Only the forward direction needs to be proven; the other direction is immediate. Let $A$ be a co-infinite set, and define $\nu$ by letting $\nu$ map $\omega \setminus A$ onto $[0,1]$. Then $A$ is not co-nowhere dense. \end{proof} \begin{proposition}\label{lax-non-immune} $A$ is infinite and non-immune iff there exists a computable numbering with respect to which $A$ is co-nowhere dense. \end{proposition} \begin{proof} Let $A$ be infinite and not immune. Thus, there is an infinite $W_e \subseteq A$ for some $e$. Let $\nu$ be a computable numbering that maps $W_e$ onto $\mathbb{Q} \setminus \omega$. Then $A$ is co-nowhere dense under $\nu$. Conversely, let $A$ be co-nowhere dense under some computable numbering $\nu$. Then $\nu^{-1}([a,b])$ is an infinite c.e. subset of $A$ for some suitable $a,b$. \end{proof} A set $D\subseteq\mathbb Q$ is \emph{effectively dense} if there is a computable function $f(a,b)$ giving an element of $D\cap (a,b)$ for $a<b\in \mathbb Q$. \begin{proposition}\label{thm:numbering} A set $A$ is constructively $\Sigma^0_1$-dense iff it is effectively dense for all computable numberings. \end{proposition} \begin{proof} By \Cref{p:immune}, $A$ is constructively $\Sigma^0_1$-dense iff it is infinite and $\omega\setminus A$ is constructively immune. Constructive immunity of $\omega\setminus A$ implies effective density of $A$ since the witnessing function for constructive immunity can be be used to witness effective density. For the converse we exploit the assumption that we get to choose a suitable $\nu$. \end{proof} Let $A$ and $B$ be sets, with $B$ computable. We say that $A$ is \emph{co-immune within $B$} if there is no infinite computable subset of $A^c\cap B$. The following diagram includes some claims not proved in the paper, whose proof (or disproof) may be considered enjoyable exercises. The quantifiers $\exists \nu$, $\forall \nu$ range over computable numberings of $\mathbb Q$. \[ \xymatrix{ \makecell{\textsf{co-finite (\Cref{co-finite})}\\ \text{(eff.) co-nowhere dense }\forall\nu}\ar[r] & \makecell{\text{constructively}\\ \text{$\Sigma^0_1$-dense (\Cref{thm:numbering})}\\ \textsf{constr.~co-immune}\\\text{eff.~dense }\forall\nu}\ar_{\text{strict: \Cref{mathias}}}[d] & \\ &\text{$\Sigma^0_1$-dense}\ar_{\text{strict: }\omega\oplus\emptyset}[dl]\ar^{\text{strict: any bi-immune}}[d]& \\ \makecell{\textsf{infinite \& non-immune (\Cref{lax-non-immune})}\\ \text{(eff.) co-nowhere dense }\exists\nu\\ \text{eff.~dense }\exists\nu \ar[dr]} &\makecell{\textsf{co-immune (\Cref{co-immune})}\\\text{dense }\forall\nu}\ar[d]& \\ \text{dense }\exists\nu&\makecell{\text{co-immune within}\\\text{some infinite computable set}}\ar[l] } \] \section{Prevalence of $\Sigma^0_1$-density}\label{sec:prevalent} In this section we investigate the existence of $\Sigma^0_1$-density in the Turing degrees at large. \subsection{Closure properties and $\Sigma^0_1$-density} \begin{proposition}\label{jan23} 1. The intersection of two $\Sigma^0_1$-dense sets is $\Sigma^0_1$-dense. 2. The intersection of two constructively $\Sigma^0_1$-dense sets is constructively $\Sigma^0_1$-dense. \end{proposition} \begin{proof} Let $A$ and $B$ be $\Sigma^0_1$-dense sets. Let $W_e$ be an infinite c.e.~set. Since $A$ is $\Sigma^0_1$-dense, there exists an infinite c.e.~set $W_d\subseteq A\cap W_e$. Since $B$ is $\Sigma^0_1$-dense, there exists an infinite c.e.~set $W_a\subseteq B \cap W_d$. Then $W_a\subseteq (A \cap B) \cap W_e$, as desired. This proves (1). To prove (2), let $f$ and $g$ witness the effective $\Sigma^0_1$-density of $A$ and $B$, respectively. Given $W_e$, we have $W_{f(e)}\subseteq A\cap W_e$ and then \[ W_{g(f(e))}\subseteq B\cap W_{f(e)} \subseteq A\cap B\cap W_e. \] In other words, $g\circ f$ witnesses the effective $\Sigma^0_1$-density of $A\cap B$. \end{proof} \begin{corollary} Bi-$\Sigma^0_1$-dense sets do not exist. \end{corollary} \begin{proof} If $A$ and $A^c$ are both $\Sigma^0_1$-dense then by \Cref{jan23}, $A\cap A^c$ is $\Sigma^0_1$-dense, which is a contradiction. \end{proof} For sets $A$ and $B$, $A\subseteq^* B$ means that $A\setminus B$ is a finite set. \begin{proposition}\label{upclosed} \begin{enumerate} \item\label{one} If $A$ is $\Sigma^0_1$-dense and $A \subseteq^* B$, then $B$ is $\Sigma^0_1$-dense. \item\label{two} If $A$ is constructively $\Sigma^0_1$-dense and $A \subseteq^* B$, then $B$ is constructively $\Sigma^0_1$-dense. \end{enumerate} \end{proposition} \begin{proof} Let $W_e$ be an infinite c.e.~set. Since $A$ is $\Sigma^0_1$-dense, there exists an infinite c.e.~set $W_d$ such that $W_d\subseteq A\cap W_e$. Let $W_c=W_d\setminus (A\setminus B)$. Since $A\setminus B$ is finite, $W_c$ is an infinite c.e.~set. Since $W_d\subseteq A$, we have $W_c = W_d \cap (B\cup A^c) = W_d \cap B$. Then, since $W_d\subseteq W_e$, we have \( W_c \subseteq B \cap W_e, \) and we conclude that $B$ is $\Sigma^0_1$-dense. This proves (1). To prove (2), if $f$ witnesses that $A$ is constructively $\Sigma^0_1$-dense then a function $g$ with $W_{g(e)}=W_{f(e)}\setminus (A\setminus B)$ witnesses that $B$ is constructively $\Sigma^0_1$-dense. \end{proof} \begin{proposition}\label{obvious} Let $B$ be a co-finite set. Then $B$ is constructively $\Sigma^0_1$-dense. \end{proposition} \begin{proof} The set $\omega$ is constructively $\Sigma^0_1$-dense as witnessed by the identity function $f(e)=e$. Thus by \Cref{two} of \Cref{upclosed}, $B$ is as well. \end{proof} As usual we write $A\oplus B = \{2x \mid x \in A\} \cup \{2x+1 \mid x \in B\}$. \begin{proposition}\label{bjoernJan23} 1. If $X_0$ and $X_1$ are $\Sigma^0_1$-dense sets then so is $X_0\oplus X_1$. 2. If $X_0$ and $X_1$ are constructively $\Sigma^0_1$-dense sets then so is $X_0\oplus X_1$. \end{proposition} \begin{proof} Let $W_e=W_{c_0}\oplus W_{c_1}$ be an infinite c.e.~set. For $i=0,1$, since $X_i$ is $\Sigma^0_1$-dense there exists $W_{d_i}\subseteq X_i\cap W_{c_i}$ such that $W_{d_i}$ is infinite if $W_{c_i}$ is infinite. Then $W_{d_0}\oplus W_{d_1}$ is an infinite c.e.~subset of $(X_0\oplus X_1) \cap W_e$. This proves (1). To prove (2), if $d_i$ are now functions witnessing the effective $\Sigma^0_1$-density of $X_i$ then $W_{d_i(c_i)}\subseteq X_i\cap W_{c_i}$, and $W_{d_0(c_0)}\oplus W_{d_1(c_1)}$ is an infinite c.e.~subset of $(X_0\oplus X_1) \cap W_e$. Thus a function $g$ satisfying \[ W_{g(e)}=W_{d_0(c_0)}\oplus W_{d_1(c_1)}, \] where $W_e=W_{c_0}\oplus W_{c_1}$, witnesses the effective $\Sigma^0_1$-density of $X_0\oplus X_1$. \end{proof} \begin{theorem}\label{bjoern3:53} There is no $\Sigma^0_1$-dense set $A$ such that all $\Sigma^0_1$-dense sets $B$ satisfy $A\subseteq^* B$. \end{theorem} \begin{proof} Suppose there is such a set $A$. Let $W_d$ be an infinite computable subset of $A$. Let $G$ be a Mathias generic with $G\cap W_d^c=\emptyset$, i.e., $G\subseteq W_d$. Then $B:=G^c$ is $\Sigma^0_1$-dense by \Cref{mathias}. Thus $A\cap G^c$ is also $\Sigma^0_1$-dense by \Cref{jan23}. And $G \subseteq W_d \subseteq A$ and by assumption $A\subseteq^* G^c$ so we get $G\subseteq^* G^c$, a contradiction. \end{proof} These results show that the $\Sigma^0_1$-dense sets under $\subseteq^*$ form a non-principal filter whose Turing degrees form a join semi-lattice. \begin{theorem}\label{thm:simple} Let $A$ be a c.e.~set. The following are equivalent: \begin{enumerate} \item $A$ is co-infinite and constructively $\Sigma^0_1$-dense. \item $A$ is co-infinite and $\Sigma^0_1$-dense. \item $A$ is co-immune. \end{enumerate} \end{theorem} \begin{proof} $1\implies 2\implies 3$ is immediate from the definitions, and $3\implies 1$ is immediate from \Cref{p:immune} and \Cref{xiang}. \end{proof} \begin{theorem} Every c.e.~Turing degree contains a constructively $\Sigma^0_1$-dense set. \end{theorem} \begin{proof} Let $\mathbf a$ be a c.e.~degree. If $\mathbf a>\mathbf 0$ then $\mathbf a$ contains a simple set $A$, see, e.g., \cite{Soare}, so \Cref{thm:simple} finishes this case. The degree $\mathbf 0$ contains all the co-finite sets, which are constructively $\Sigma^0_1$-dense by \Cref{obvious}. \end{proof} \subsection{Cofinality in the Turing degrees of constructive $\Sigma^0_1$-density} \begin{definition}\label{def_ve} For $k\ge 0$, let $I_k$ be intervals of length $k+2$ such that $\min(I_0)=0$ and $\max(I_k)+1=\min(I_{k+1})$. Let $V_e=\bigcup_{s\in\omega}V_{e,s}$ be a subset of $W_e$ defined by the condition that $x\in I_k$ enters $V_e$ at a stage $s$ where $x$ enters $W_e$ if this makes $\abs{V_{e,s}\cap I_k}\le 1$, and for all $j>k$, $V_{j,s} \cap I_k=\emptyset$. \end{definition} \begin{lemma}\label{dec31} There exists a c.e., co-infinite, constructively $\Sigma^0_1$-dense, and effectively co-immune set. \end{lemma} \begin{proof} Let $A=\bigcup_{e\in\omega} V_e$. $V_e$ is c.e. by construction, and if $W_e$ is infinite, $V_e$ is also infinite. So $V_e=W_{f(e)}$ is the set witnessing that $A$ is constructively $\Sigma^0_1$-dense. Moreover $A$ is coinfinite since $\abs{A\cap I_k}\le k+1 < k+2 = \abs{I_k}$ gives $I_k\not\subseteq A$ for each $k$ and \[ \abs{\omega\setminus A} =\left| \left(\bigcup_{k\in\omega}I_k\right) \setminus A \right| = \left| \bigcup_{k\in\omega} (I_k\setminus A)\right| = \sum_{k\in\omega} \abs{I_k\setminus A} \ge \sum_{k\in\omega} 1 = \infty. \] The set $A$ is effectively co-immune because if $W_e$ is disjoint from $A$ then since as soon as a number in $I_k$ for $k\ge e$ enters $W_e$ then that number is put into $A$, $W_e \subseteq \bigcup_{k<e}I_k$ so $\abs{W_e}\le \sum_{k<e} (k+2) = \sum_{k\le e+1}k = \frac{(e+1)(e+2)}2$. \end{proof} \begin{theorem} For each set $R$ there exists a constructively $\Sigma^0_1$-dense, effectively co-immune set $S$ with $R\le_T S$. \end{theorem} \begin{proof} Let $R$ be any set, which we may assume is co-infinite. Let $A$ be as in the proof of \Cref{dec31}. Let $S\supseteq A$ be defined by \[ S = A \cup \bigcup_{k\in R} I_k. \] Since $A\subseteq S$ and $S$ is co-infinite, $S$ is constructively $\Sigma^0_1$-dense and effectively co-immune. Since $k\in R\iff I_k\subseteq S$, we have $R\le_T S$. \end{proof} \subsection{Non-$\Delta^0_2$ degrees} \begin{lemma}\label{jan26pigeon} Suppose that $T\subseteq 2^{<\omega}$ is a tree with only one infinite path. Then for each length $n$ there exists a length $m>n$ such that exactly one string of length $n$ has an extension of length $m$ in $T$. \end{lemma} \begin{proof} Suppose not, i.e., there is a length $n$ such that for all $m>n$ there are at least two strings $\sigma_m,\tau_m$ of length $n$ with extensions of length $m$ in $T$. By the pigeonhole principle there is a pair $(\sigma,\tau)$ that is a choice of $(\sigma_m,\tau_m)$ for infinitely many $m$. Then by compactness both $\sigma$ and $\tau$ must be extendible to infinite paths of $T$. \end{proof} \begin{lemma}\label{jan26} Suppose that $T\subseteq 2^{<\omega}$ is a tree with only one infinite path $A$, and that $T$ is a c.e.~set of strings. Then $A$ is $\Delta^0_2$. \end{lemma} \begin{proof} By \Cref{jan26pigeon}, for each length $n$ there exists a length $m>n$ such that exactly one string of length $n$ has an extension of length $m$ in $T$. Using $0'$ as an oracle we can find that $m$ and define $A\upharpoonright n$ by looking for such a string. In fact, $T\le_T 0'$ and so its unique path $A\le_T 0'$ as well. \end{proof} \begin{theorem}\label{prevalent} Given $A \in 2^{\omega}$, let $\hat{A} := \left\{\sigma \in 2^{<\omega} \mid \sigma \prec A \right\}$ be the set of finite prefixes of $A$. If $A$ is not $\Delta^0_2$ then $\hat{A}$ is co-$\Sigma^0_1$-dense. \end{theorem} \begin{proof} Let $A^*$ be the complement of $\hat A$. Let $W_e\subseteq 2^{<\omega}$ be an infinite c.e.~set of strings. Let $T$ be the set of all prefixes of elements of $W_e$. Then $T$ is an infinite tree, hence by compactness it has at least one infinite path. That is, there is at least one real $B$ such that all its prefixes are in $T$. \indent Case 1: The only such real is $B=A$. Then by \Cref{jan26}, $A$ is $\Delta^0_2$. \indent Case 2: There is a $B\ne A$ such that all its prefixes are in $T$. Let $\sigma$ be a prefix of $B$ that is not a prefix of $A$. Let $W_d=[\sigma]\cap W_e$. Since all prefixes of $B$ are prefixes of elements of $W_e$, there are infinitely many extensions of $\sigma$ that are prefixes of elements of $W_e$. Consequently $W_d$ is infinite. Thus, $W_d$ is our desired infinite subset of $A^*\cap W_e$. \end{proof} \subsection{High degrees} \begin{definition} A set $A$ is \emph{co-r-cohesive} if its complement is r-cohesive. This means that for each computable (recursive) set $W_d$, either $W_d\subseteq^* A$ or $W_d^c\subseteq^* A$. \end{definition} \begin{definition}[{Odifreddi \cite[Exercise III.4.8]{MR982269}, Jockusch and Stephan \cite{MR1270396}}] A set $A$ is strongly hyperhyperimmune (s.h.h.i.) if for each computable $f:\omega\to\omega$ for which the sets $W_{f(e)}$ are disjoint, there is an $e$ with $W_{f(e)}\subseteq \omega\setminus A$. A set $A$ is strongly hyperimmune (s.h.i.) if for each computable $f:\omega\to\omega$ for which the sets $W_{f(e)}$ are disjoint and computable, with $\bigcup_{e\in\omega}W_{f(e)}$ also computable, there is an $e$ with $W_{f(e)}\subseteq \omega\setminus A$. \end{definition} \begin{proposition}\label{thm:shi} Every s.h.i.~set is co-$\Sigma^0_1$-dense. \end{proposition} \begin{proof} Let $A$ be s.h.i. Let $W_e$ be an infinite c.e.~set. Let $W_d$ be an infinite computable subset of $W_e$. Effectively decompose $W_d$ into infinitely many disjoint infinite computable sets, \[ W_d = \bigcup_{i\in\omega} W_{g(d,i)}. \] For instance, if $W_d=\{a_0 < a_1 < \dots\}$ then we may let $W_{g(e,i)}=\{a_n: n=2^i(2k+1), i\ge 0, k\ge 0\}$. Since $A$ is s.h.i., there exists some $i_e$ such that $W_{g(d,i_e)}\subseteq A^c$. The sets $W_{g(d,i_e)}$ witness that $A^c$ is $\Sigma^0_1$-dense. \end{proof} Clearly r-cohesive implies s.h.i., and s.h.h.i. implies s.h.i. It was shown by Jockusch and Stephan \cite[Corollary 2.4]{MR1270396} that the cohesive degrees coincide with the r–cohesive degrees and (Corollary 3.10) that the s.h.i.~and s.h.h.i.~degrees coincide. \begin{proposition} Every high degree contains a $\Sigma^0_1$-dense set. \end{proposition} \begin{proof} Let $\mathbf h$ be a Turing degree. If $\mathbf h\not\le \mathbf 0'$, then $\mathbf h$ contains a $\Sigma^0_1$-dense set by Theorem \ref{prevalent}. If $\mathbf h\le \mathbf 0'$ and $\mathbf h$ is high then since the strongly hyperhyperimmune and cohesive degrees coincide, and are exactly the high degrees \cite{MR360240}, $\mathbf h$ contains a strongly hyperimmune set. Hence by Theorem \ref{thm:shi}, $\mathbf h$ contains a $\Sigma^0_1$-dense set. \end{proof} \subsection{Progressive approximations} \begin{definition} Let $A$ be a $\Delta^0_2$ set. A computable approximation $\{\sigma_t\}_{t\in\omega}$ of $A$, where each $\sigma_t$ is a finite string and $\lim_{t\to\infty}\sigma_t=A$, is \emph{progressive} if for each $t$, \begin{itemize} \item if $\abs{\sigma_t}\le\abs{\sigma_{t-1}}$ then $\sigma_t\upharpoonright(\abs{\sigma_t}-1) = \sigma_{t-1}\upharpoonright(\abs{\sigma_t}-1)$ (the last bit of $\sigma_t$ is the only difference with $\sigma_{t-1}$); \item if $\abs{\sigma_t} > \abs{\sigma_{t-1}}$ then $\sigma_{t-1}\prec\sigma_t$; and \item if $\sigma_t\not\prec\sigma_s$ for some $s>t$ then $\sigma_t\not\prec\sigma_{s'}$ for all $s'\ge s$ (once an approximation looks wrong, it never looks right again). \end{itemize} If $A$ has a progressive approximation then we say that $A$ is \emph{progressively approximable}. \end{definition} Note that a progressively approximable set must be $h$-c.e.~where $h(n)=2^n$. \begin{theorem}\label{January 31} Let $A$ be a progressively approximable and noncomputable set. Let $\{\sigma_t\}_{t\in\omega}$ be a progressive approximation of $A$. Then $\{t: \sigma_t \prec A\}$ is constructively immune. \end{theorem} \begin{proof} Let $W_e$ be an infinite c.e.~set and let $T$ be an infinite computable subset of $W_e$. Since $A$ is noncomputable, we do not have $T\subseteq\{t:\sigma_t\prec A\}$. Since the approximation $\{\sigma_t\}_{t\in\omega}$ is progressive, once we observe a $t$ for which $\sigma_t\not\prec\sigma_s$, for some $s>t$, then we know that $\sigma_t\not\prec A$. Then we define $\varphi(e)=t$, and $\varphi$ witnesses that $\{t: \sigma_t \prec A\}$ is constructively immune. \end{proof} A direction for future work may be to find new Turing degrees of progressively approximable sets. \bibliographystyle{plain}

482222046506965dc953bb6a7d53d37ae09217ef

\section{Plastic scintillators at the Pierre Auger Observatory} \label{sec::intro} Cosmic rays with energies above $10^{15}$\,eV can only be indirectly detected by measuring the extensive air showers produced after they interact with nucleons of the Earth's atmosphere. The multiplicative process following the first interaction is approximately explained by the Heitler-Matthews model of hadronic cascades~\cite{Matthews:2005sd}: secondary particles are produced by the decay or interaction of products from the first interaction. In this way, we can identify two distinguished components in the shower, namely the hadronic and the electromagnetic components. The latter is mainly composed of electrons, positrons, and photons, the former by predominantly pions and kaons; the decay of kaons and electrical charged pions constitutes the main contribution to the air shower muons. Even though the Heitler-Matthews model is a simplified picture of the real process involved in the extensive air shower development, it is accurate enough to replicate three major features: (i) the atmospheric depth of the maximum development of the cascade is linearly dependent on the logarithm of the energy, $E$, of the primary cosmic ray, (ii) the number of muons in the shower, $N_\upmu$, grows as $(E/\xi_\mathrm{c})^\beta$ and (iii) the mass number $A$ of the primary particle scales as $N_\upmu\propto\ A^{\beta-1} (E/\xi_\mathrm{c})^\beta$ where $\mathrm{\xi_\mathrm{c}}$ is the critical energy at which charged pions decay into muons\footnote{Kaons decay into charged pions, which in turn, decay into muons.} rather than interacting with nucleons, and $\mathrm{\beta\approx 0.9}$ is an effective parameter of the model~\cite{AlvarezMuniz:2002ne}. In the Heitler-Matthews model, the muon content of an extensive air shower is directly related to the mass of the primary particle, thus, direct measurements of these muons with ground-based arrays have become increasingly relevant to unlock the riddles concerning cosmic rays with energy above $10^{15}$\,eV, the so-called high- and ultra-high-energy ($\geq 10^{18}$\,eV) cosmic rays. Both the consistency of hadronic interaction models for these energies tuned with the Large Hadron Collider data~\cite{MuonDeficit}, as well as the astrophysical scenarios still compatible with observations~\cite{AugerSpectrum}, can only be explored knowing in more detail the mass composition of the cosmic ray flux. Precise and direct measurements of muons at the ground are, undoubtedly, of utmost importance to forward our understanding of the physics of high- and ultra-high-energy cosmic rays. The implementation of plastic scintillator arrays is a common approach to measure the lateral profile of air shower particles. Their versatility allows the deployment of these detectors either on the ground or buried underground. On-ground, they are optimal to measure the most abundant electromagnetic component of the cascades, whereas underground, they are ideal to measure directly the shower muons as electrons, positrons, and photons are easily shielded by the soil overburden. To improve the sensitivity in large-area scintillator detectors, optical fibers are typically used to collect the light produced in the plastic and conduct it to a photo-sensor that might be meters apart from the impinging position. In this way, it is possible to use large scintillator surfaces achieving better control of fluctuations in the measurements of the particle density in air showers. At the Pierre Auger Observatory~\cite{PierreAugerObservatory}, cosmic rays with energies above $10^{16.5}$\,eV are measured with a hybrid detection technique. Arrays of water-Cherenkov detectors covering different areas and different spacing are used to measure the lateral spread of shower particles at the ground. These arrays, commonly referred to as the surface detector, are deployed in triangular grids with spacings of 1500\,m (SD-1500), 750\,m (SD-750) and 433\,m (SD-433) for energy thresholds of about $10^{18.5}$\,eV, $10^{17.5}$\,eV, and $10^{16.5}$\,eV respectively. Additionally to the surface detector, which has a 100\% duty cycle, the Pierre Auger Observatory is also equipped with a fluorescence detector consisting of 27 telescopes to measure the fluorescence light produced with the longitudinal development of the air showers in the atmosphere. It allows for a measurement of the calorimetric energy of the primary particle and it is optimal to determine the depth of the maximum development of the extensive air shower on an event-by-event basis. Nevertheless, as it can be operated only during moonless nights, it has a reduced duty cycle of around 15\%. As part of its upgrade, dubbed AugerPrime~\cite{AugerPrime}, arrays of plastic scintillator detectors are being deployed at the Pierre Auger Observatory on top of each water-Cherenkov detector of the SD-1500 to measure air shower muons and electromagnetic particles for the highest-energy cosmic rays. The surface scintillator detector (SSD) will provide a complimentary response from which the different components of the signals, roughly distinguished by their electromagnetic or hadronic origin, can be deconvolved. In addition, direct access to the muonic content of air showers at energies up to the ankle region (around $10^{18.7}$\,eV conserving high statistics)\footnote{This upper limit is a consequence of the flux and the area covered by the SD-750 and SD-433.}, will be attained with arrays of plastic scintillator detectors deployed at 2.3\,m underground and at each position of the SD-750 and SD-433~\cite{MuonsWithAMIGA}. The underground muon detector (UMD) will provide clean and precise measurement of the muons in the extensive air shower improving the cosmic ray mass identification in the second-knee and ankle regions of the energy spectrum~\cite{spectrum}. Each SSD station consists of 48 plastic-scintillator strips of (160\,$\times$\,5\,$\times$\,1)\,cm with wavelength-shifting optical fibers deployed in an aluminum container constituting two modules with a total active area of 3.8\,m$^2$~\cite{SSD}. The read-out electronics consist mainly of a photomultiplier tube with low- and high-gain outputs. On the other hand, each UMD station consists of three 10\,m$^2$ modules segmented into 64 plastic-scintillator strips of (400\,$\times$\,4\,$\times$\,1)\,cm also with wavelength-shifting optical fibers that conduct the light from the scintillators to an array of 64 silicon photomultipliers (SiPMs)~\cite{AMIGAPrototype, AMIGASIPM}. Pictures of a UMD module construction (left and middle) and of a station deployment (right) are presented in Fig.~\ref{fig:deployment}. We paint in black the optical-fiber end that is not coupled to the SiPM to avoid photon reflections and, in turn, the arrival of delayed photons to the SiPM. \begin{figure} [t] \centering \includegraphics[trim={0.0cm 0.0cm 0.0cm 0.0cm},clip,width=1.0\textwidth]{./img/setup/deploy3.jpg} \caption{UMD module under construction (left and middle) and a UMD station after deployment and before being overburden with soil (right). \label{fig:deployment}} \end{figure} In this paper, we present a detailed description of the simulation sequence of one scintillator strip with the optoelectronics system of the UMD based on a thorough characterization of signals measured in the laboratory with individual strips. The setup built in the laboratory includes a complete replica of the UMD acquisition chain under controlled conditions which has allowed us to both fine-tune the parameters used in the simulation and to validate their outcome. The structure of this article is as follows: section~\ref{sec:setup} is an introduction of the UMD while section~\ref{sec:simulations} is devoted to the explanation of the simulation rationale; in section~\ref{sec:validation} the characterization of single-muon signals and the validation of the simulation in terms of the efficiency and dynamic range of the detector are described; finally, we summarize and conclude in section~\ref{sec:conclusion}. \section{The underground muon detector and its simulation} \label{sec:setup} When a muon impinges on a scintillator strip, photons are generated and propagated through the optical fiber to a SiPM\footnote{In the UMD, the photo-detectors are 1584-cell SiPM, Hamamatsu S13361-2050.} producing a single-muon signal. The SiPM output is then processed with two different read-outs, corresponding to each of the acquisition modes of the UMD~\cite{UMDFrontEnd, UMDCalibration, UMDComms}: the {\it binary} and {\it ADC} modes implemented to probe muon densities ranging from one up to tens of muons per $\mathrm{m^2}$. The binary acquisition is designed to measure low particle densities, by attaining a high efficiency for single particles and reducing the signal fluctuations in this range. It handles independently the 64-SiPM signals through a pre-amplifier, fast-shaper, and a discriminator, built within each channel of two 32-channel Application-Specific Integrated Circuits (ASIC). The maximum number of simultaneous particles that can be detected with this mode is limited by its own segmentation as it can only detect one muon per strip at the same time. This limit is extended with the ADC mode, in which the 64 SiPM signals are summed and afterward amplified. At the output of a single UMD module, we obtain 64 traces for the binary mode and two for the ADC mode. The memory depth of the UMD electronics stores 6.4\,$\upmu$s; each trace of the binary mode has 2048 bits where a ``1'' or a ``0'' is output if the fast-shaper signal is above or below the discriminator threshold. For the ADC mode, on the other hand, the outcomes are waveforms of 1024 samples. The binary and ADC signals are sampled in two different ways: (a) the discriminator signals of the binary mode with a Field-Programmable Gate Array (FPGA) at a speed of 320\,MHz (3.125\,ns sample time), and (b) the ADC mode with two Analog-to-Digital Converters (ADCs) at a speed of 160\,MHz (6.25\,ns sample time). A schematic flow of the working chain corresponding to one scintillator strip is displayed in the left panel of Fig.~\ref{fig:electronics}. \begin{figure}[t] \centering \begin{subfigure}{1.00\textwidth} \centering \includegraphics[trim={0.0cm 0.0cm 0.0cm 0.0cm},clip,width=1.0\linewidth]{./img/setup/schematic4.pdf} \end{subfigure} \caption{Schematics of the detector working chain (left), simulation chain of the single-muon signal (middle), and layout of the laboratory data used to develop and validate the simulation (right).} \label{fig:electronics} \end{figure} To translate the signal of the UMD modules into a particle density, both binary and ADC traces need to be converted to a number of impinging muons. This implies interpreting the number of photons collected by SiPMs in the form of photo-equivalents\footnote{We refer to as "photo-equivalents" to the number of triggered cells in the SiPM.} (PE) processed by the electronics after a charged particle passes through the scintillator and photons propagate through the optical fiber. The shape of the final traces depends on the arrival time profile of the photons, and in turn on the timing of the excitation and de-excitation processes of both scintillator and optical fiber molecules. In the ADC acquisition mode, we use the signal charge to infer the number of impinging muons, and therefore the PE timing is not relevant in the estimation of the muon density. However, when we count and store signals above a given amplitude, as in the binary mode, the timing of PEs plays a key role, since the amplitude of the signal is determined by the superposition of the PE pulses according to their arrival time profile. PE pulses are not only the outcome of incoming light; they are also produced by thermal fluctuations within the SiPM cells when electrons spontaneously jump to the silicon conduction band and trigger an avalanche identical to that generated by an impinging photon. Furthermore, when electrons in a triggered cell (disregarding if it was by incident radiation or thermal noise) are recombined, the photons produced may escape towards a neighboring cell inducing a second and simultaneous avalanche. This effect, known as optical cross-talk, may produce pulses of more than one PE, even though only one cell was triggered either by impinging radiation or thermal noise and is accounted for in the simulation. Each scintillator strip of the UMD is simulated following the steps illustrated in the middle panel of Fig.~\ref{fig:electronics}. The sequence is based on the parameters accessible (and therefore useful to tune and validate the simulation) with laboratory measurements: analog 1\,PE pulses produced by thermal fluctuations in the SiPM, analog single-muon pulses, the binary traces, and the ADC waveforms obtained with the electronics. These last three, produced at fixed distances over the scintillator strips and, thus, giving information on the effect of the light attenuation along with the optical fiber. Taking this into account, the main ingredients to simulate the sequence for single particle consists of: \begin{enumerate} \item a PE pulse generator developed to simulate the signals of individual photons arriving at the SiPM, \item a single-muon generator that sums several PE signals according to a given number of photons arriving at the SiPM with a time profile that follows the physical processes involved in the photon generation and propagation, \item a transfer model corresponding to the read-out electronics of the binary and ADC modes, \item a sampler for the final digitization and discretization of the analog signals for both acquisition modes. \end{enumerate} To acquire the laboratory data used to develop, tune, and validate the simulation, we used the setup illustrated in Fig.~\ref{fig:pipaDesign}: one scintillator strip with its optical fiber coupled to a SiPM, all components identical to those used in UMD modules~\cite{AMIGAPrototype}. To trigger the acquisition, a muon telescope consisting of two (4\,$\times$\,4\,$\times$\,1)\,cm scintillators, matching the width of one scintillator strip, was used. When a muon impinges on the segments of the telescope, a trigger signal for acquisition is produced. Moving the telescope over the length of the strip allowed us to obtain muon signals as a function of the optical-fiber length between the particle impact position and the SiPM and, therefore, to retrieve the information of the light attenuation in the fiber. The SiPM output was processed with two different read-out systems: a customized electronics plus an oscilloscope to obtain analog signals, and standard UMD electronics to obtain digitized signals. \begin{figure} [t] \centering \includegraphics[trim={0.0cm 0.0cm 0.0cm 0.0cm},clip,width=0.9\textwidth]{./img/setup/pipaDesign2.pdf} \caption{Schematics of the laboratory setup used to extract analog and digital single-muon signals. \label{fig:pipaDesign}} \end{figure} The SiPM reverse-bias voltage, $V_\mathrm{{bias}}$, in the laboratory setups was set following the procedure explained in~\cite{UMDCalibration}, so the configuration used in the Pierre Auger Observatory is properly reproduced. The procedure consists of determining the SiPM break-down voltage, $V_\mathrm{{br}}$, by means of the binary channel and then setting the reverse-bias voltage to fix the same over-voltage ($V_\mathrm{{ov}}$\,=\,3.5V) in all SiPMs: $V_\mathrm{{bias}} = V_\mathrm{{br}} + V_\mathrm{{ov}}$. The procedure relies on obtaining the 1\,PE signal amplitude as a function of the $V_\mathrm{{bias}}$ by measuring the rate of dark counts: for each $V_\mathrm{{bias}}$, we count the number of transitions in the binary trace between a ``0'' and a ``1'' for different discriminator thresholds. As a result, for each $V_\mathrm{{bias}}$, we obtain the dark-count rate as a function of the discriminator threshold as presented in Fig.~\ref{fig:calibrationPEGen}. In other words, this is the number of pulses per second at the fast-shaper output crossing the discriminator threshold as a function of this threshold. For each integer number of PE, the signal amplitude is well separated from its consecutive PE, therefore, when the discriminator threshold moves between two PE peaks, the rate of dark counts varies slowly and a plateau is obtained; when the threshold moves close to the mean signal amplitude of an integer number of PEs, the rate varies rapidly and a drop in the rate is obtained. As a consequence, the PE mean amplitude can be easily identified in the transition between two consecutive plateaus. This method, based on measuring the SiPM thermal noise, was developed to set the operating point of the UMD modules to guarantee a uniform response in terms of noise and gain among the whole array. We indicate the positions of the first and second plateaus along with the mean amplitude of the one- and two-PE signals (indicated with dashed-orange lines). \begin{figure}[t] \centering \begin{subfigure}{0.8\textwidth} \centering \includegraphics[trim={0.0cm 0.0cm 0.0cm 0.0cm},clip,width=1.0\linewidth]{./img/setup/curve3.pdf} \end{subfigure} \caption{Dark-count rate in counts per second (cps) as a function of the discriminator threshold acquired with an over-voltage of 3.5\,V. Orange-dashed lines indicate the position of the 1 and 2\,PE peaks. The gray area denotes the region where the baseline is found.} \label{fig:calibrationPEGen} \end{figure} As we mentioned in Fig.~\ref{fig:electronics}, analog data obtained in the laboratory was used to develop both the single-PE and single-muon pulse generator: we obtained two thousand dark-count signals with 1\,PE amplitude, and 16000 single-muon signals using the muon telescope at eight positions on the scintillator strip with steps of 0.5\,m. This latter dataset was also used to validate the muon generator by characterizing the amplitude, charge, and width of the single-muon analog signals. In this same way, digital data was obtained to validate the output of the whole optoelectronic simulation chain. \section{Simulations} \label{sec:simulations} The simulation of the UMD scintillator strip was not developed using models based on first principles or a microscopic analytical assumption of the detector physics, but through a phenomenological approach that describes the macroscopic behavior of the detector components tuned to laboratory data. This methodology allowed for a simulation that accurately represents the signal features of the detector minimizing the computing costs. The steps corresponding to the photon production in the scintillator, propagation through the optical fiber, and SiPM response were developed using analog measurements. In these steps, model parameters were extracted by fitting laboratory data, and used as the input of a pulse generator. Then, the response of the read-out electronics and subsequent sampling were simulated using models for the transfer functions tuned to digital data also acquired in the laboratory. As a result, each step of the simulation sequence was validated by means of the laboratory data to guarantee an accurate and realistic representation of the main features of the UMD traces and, therefore, of the full detector response. We present in the following subsections the simulation of the analog and digital signals of UMD. \subsection{Analog simulation} \label{subsec:simDR} When a muon impinges on a scintillator strip, scintillation photons are produced in the plastic and subsequently propagated through the optical fiber to the SiPM resulting in a signal; this is the addition of the pulses produced by each individual photon. For this reason, the first practical step towards the simulation of the UMD signal, as displayed in Fig.~\ref{fig:electronics}, is the development of a PE generator to reproduce the SiPM analog pulses for individual photons. To this aim, we used PE pulses obtained in the laboratory and presented in the left panel of Fig.~\ref{fig:sipmSignalsRaw}. Since the signal produced by dark counts is indistinguishable from that by impinging photons, this data set was built using thermal noise. In the figure, we display an overlap of all the data, where the effect of the optical cross-talk is apparent: the one-, two-, and three-PE signals are clearly visible. To better illustrate this, we present in the right panel of Fig.~\ref{fig:sipmSignalsRaw} the maximum amplitude of signals as a function of their charge (corresponding to the signal integral), where each population corresponds to a plateau transition previously presented in Fig.~\ref{fig:calibrationPEGen}, and in turn, to each integer number of PEs. It is also worth mentioning that for the 1\,PE population, the amplitude and charge means are (0.798\,$\pm$\,0.001)\,mV and (0.2313\,$\pm$\,0.0001)\,pC with a standard deviation of (0.0327\,$\pm$\,0.0004)\,mV, and (0.0117\,$\pm$\,0.0002)\,pC. The fluctuations of about 5\% contain all constructive differences within each of the 1584 SiPM cells and denote the uniformity in the response between SiPMs. \begin{figure}[t] \hspace{-0.2cm} \centering \begin{subfigure}{.49\textwidth} \centering \includegraphics[trim={0.0cm 0.0cm 0.0cm 0.0cm},clip,width=1.0\linewidth]{./img/SiPMPeaks/output2.png} \end{subfigure} \hspace{0.2cm} \begin{subfigure}{0.49\textwidth} \centering \includegraphics[trim={0.0cm 0.0cm 0.0cm 0.0cm},clip,width=1.0\linewidth]{./img/SiPMPeaks/amplitudeVSCharge.pdf} \end{subfigure} \caption{(Left) overlapped dark-count pulses. We can clearly identify the different PE peaks produced by crosstalk between the SiPM inner-cells. (Right) signal amplitude as a function of its charge (signal integral). The three populations correspond to the one-, two- and three-PE signals.} \label{fig:sipmSignalsRaw} \end{figure} The PE generator was built using a phenomenological model for the signal of individual photons. To fit the dark-count pulses displayed in Fig.\ref{fig:sipmSignalsRaw}, we used a function with one exponential rise time and three exponential fall times~\cite{maranoElectric} \begin{equation} \operatorname{Signal}\left(t\right) = A_1 \left(1-e^{\frac{-t}{\tau_\text{r}}}\right) \left(A_2\,e^{\frac{-t}{\tau_{1}}}+A_3\,e^{\frac{-t}{\tau_{2}}}+e^{\frac{-t}{\tau_{3}}}\right), \label{eq:electricmodel} \end{equation} \noindent where $A_1$, $A_2$, $A_3$ are the coefficients that determine the signal amplitude, $\tau_\text{r}$ is the rise time, and $\tau_{1}$, $\tau_{2}$, and $\tau_{3}$ are the three fall times. The first corresponds to the quenching timing; the second, and slow, decay corresponds to the recovery time of the SiPM cells. The third corresponds to a fast decay which is produced by the coupling of parasitic capacitors to the load impedance of the read-out system~\cite{sipms101, maranoElectric}. The parameters of the model are estimated by fitting the measured 1\,PE pulses, for which signals that not exclusively contained single-PEs, were removed by dismissing all events outside the first population in the right panel of Fig.~\ref{fig:sipmSignalsRaw}. To properly model the signal fluctuations, each measured PE pulse is fitted with Equation~\eqref{eq:electricmodel}, and a histogram is filled for each parameter to extract its corresponding mean and standard deviation. Results are summarized in Table~\ref{table:params}. The fluctuations of the signal baseline were estimated using the first 100 samples of the individual PE signals, obtaining a mean fluctuation of (26.90 $\pm$ 0.04)\,$\upmu$V with a standard deviation of (1.78 $\pm$ 0.03)\,$\upmu$V, and included as a parameter of the simulation. The PE generator is, therefore, a set of random generators with normal distributions whose means and standard deviations are the values presented in Table~\ref{table:params}. \begin{table}[t] \centering \renewcommand{\arraystretch}{1.25} \begin{tabular}{c c c} \toprule Parameter & Mean & Std. Dev.\\ \midrule $A_1$ / mV & 0.29 $\pm$ 0.01 & (1.37 $\pm$ 0.03) $10^{-2}$\\ $A_2$ & 23.22 $\pm$ 0.07 & 2.94 $\pm$ 0.06 \\ $A_3$ & 1.609 $\pm$ 0.001 & 0.054 $\pm$ 0.001\\ $\mathrm{\tau_r}$ / ns & 3.82 $\pm$ 0.01 & 0.62 $\pm$ 0.01 \\ $\mathrm{\tau_{1}}$ / ns & 1.187 $\pm$ 0.002 & 0.070 $\pm$ 0.001\\ $\mathrm{\tau_{2}}$ / ns & 23.44 $\pm$ 0.04 & 1.95 $\pm$ 0.04\\ $\mathrm{\tau_{3}}$ / ns & 0.221 $\pm$ 0.001 & (3.24 $\pm$ 0.04) $10^{-2}$\\ \bottomrule \end{tabular} \caption{Mean and standard deviation for the parameters in Equation~\eqref{eq:electricmodel} obtained by fitting two thousand signals of 1\,PE.} \label{table:params} \end{table} To test the accuracy of the PE generator, we created a data set with two thousand simulated single PEs and compared it with the measured dark counts used to develop the generator. The mean signal for each data set was obtained by averaging all the pulses in the data set, and the results are presented in Fig.~\ref{fig:residuals}, for both laboratory data (black squares) and simulation (red circles). The difference between the two waveforms is displayed in the bottom of Fig.~\ref{fig:residuals}; it varies between $-$5\% and 10\% of the measured signal, having the maximum differences when the signal changes abruptly, i.e.\,at the signal rise and first fall, and the following bump. However, it is worth noting that the difference oscillates around 0, and the average is quite negligible: the difference in the integrated signal charge is less than 1\% and less than 3\% when looking at the maximum amplitude, being these the most relevant parameters of the UMD performance. Each of the simulated PE pulses, presented in the inset of Fig.~\ref{fig:residuals}, was fitted in the same way as in the laboratory data, and the parameter mean and standard deviation were computed. We then obtained a maximum difference of 9\% between laboratory measurements and simulation, with differences mostly below 5\%, which proves the accuracy of the proposed algorithm. \begin{figure}[t] \centering \begin{subfigure}{0.8\textwidth} \centering \includegraphics[trim={0.0cm 0.0cm 0.0cm 0.0cm},clip,width=1.0\linewidth]{./img/SiPMPeaks/residuals.pdf} \end{subfigure} \caption{Average single-PE signal of two thousand pulses for both laboratory data and simulation. At the bottom, we present the difference between the averages at each sample time. In the inset, we display the two thousand simulated pulses obtain with the PE-generator.} \label{fig:residuals} \end{figure} After devising a generator that accurately simulates the single-PE pulses, we can reproduce the muon signal by adding several PEs. When a muon impinges on the detector, photons are produced in the plastic scintillator and then absorbed and re-emitted by the optical fiber; all three processes have their own characteristic times and spectral distributions. The photons re-emitted in the fiber are propagated towards the SiPM and finally detected with an efficiency that depends on their wavelength and is about 40\% at 492\,nm. Due to the emission delays, the incoming photons spread in time, producing a signal with a structure for which the number of PEs in amplitude is not equivalent to the number of PEs in charge, since not all photons add to the maximum signal amplitude. Furthermore, the total number of PEs measured by the SiPM depends on the position where the muon arrived at the scintillator strip due to the optical-fiber attenuation, which follows a double-exponential decay law~\cite{attenuation} \begin{equation} \overline{n_\mathrm{PE}} = n_\mathrm{PE}(0)\left( a\,\mathrm{e}^{-\frac{x}{\lambda_1}}+(1-a)\,\mathrm{e}^{-\frac{x}{\lambda_2}}\right), \label{eq:attenuation2} \end{equation} \noindent where $\lambda_1$, $\lambda_2$ are the attenuation lengths, \textit{a} is a proportionality constant, and $n_\mathrm{PE}(0)$ corresponds to the mean number of PE that would be detected if a muon impinged the scintillator at 0\,m from the SiPM. This model, with a short and long attenuation length, describes well the physical processes involved in the scintillator and optical fiber system. Nevertheless, since measurements were taken only at distances greater than 1\,m in this work, the contribution of the short attenuation length, which is a few centimeters, is negligible. Therefore, we fixed $\lambda_2=3.5$\,cm and considered $n_\mathrm{PE}(0)$, $\lambda_1$ and $a$ as the only free parameters of Equation~\eqref{eq:attenuation2}. We display the number of PEs by (i) dividing the signal maximum amplitude by the maximum amplitude of 1\,PE (blue squares), and (ii) by dividing the signal charge by the mean charge of 1\,PE (orange circles). Fitting results are displayed on tables in the left panel. As previously mentioned, given the time profile of the arriving photons, not all PEs add to the maximum signal amplitude, even though they add to the total signal charge; this explains the difference between the two curves. It is worth noting that the binary mode, as it implements an amplitude threshold, is sensitive to the amplitude attenuation, while the ADC mode is sensitive to that in charge. With the parametrization in Equation~\eqref{eq:attenuation2}, we obtain the mean and variance of the number of PEs for vertical muons at each position on the scintillator strip without requiring a detailed simulation of the energy deposit in the plastic or of the subsequent photo-production and propagation through the optical fiber. Nevertheless, it is important to stress that, when merging the code into the Auger official simulation framework (\Offline)~\cite{offline}, non-vertical muons and the energy deposit fluctuations are simulated using \textsc{Geant4}\xspace~\cite{geant4}. In contrast, the photon production in the scintillator and optical fiber propagation are parametrized. This approach significantly simplifies the complexity of the simulation algorithm and drastically reduces the computing time. Furthermore, the parametrization of the number of PEs already includes the effects of the SiPM photo-detection efficiency and cross-talk, which also aids in improving the simulation performance. \begin{figure}[t] \begin{subfigure}{0.65\textwidth} \includegraphics[trim={0.0cm 0.0cm 0.0cm 0.0cm},clip,width=1.0\linewidth]{./img/muonPeaks/attenuation2.pdf} \end{subfigure} \hspace{1.0cm} \begin{subfigure}{.1\textwidth} \begin{tabular}{c c} \toprule {\color{BurntOrange} {\huge ---}} & Charge\\ \midrule $n_\mathrm{PE}(0)\ /\ \text{PE}$ & 43 $\pm$ 2\\ $a$ & 0.97 $\pm$ 0.01\\ $\lambda_1$ / m & 3.8 $\pm$ 0.3\\ $\lambda_2$ / m & 0.035 (fixed)\\%$\pm$ 0.001\\ \bottomrule \\ \toprule {\color{blue} {\huge ---}} & Amplitude\\ \midrule $n_\mathrm{PE}(0)\ /\ \text{PE}$ & 17.4 $\pm$ 0.7 \\ $a$ & 0.95 $\pm$ 0.01 \\ $\lambda_1$ / m & 4.2 $\pm$ 0.3 \\ $\lambda_2$ / m & 0.035 (fixed) \\%$\pm$ 0.001\\ \bottomrule \end{tabular} \end{subfigure} \caption{Mean number of photo-equivalents in charge and amplitude as a function of the telescope position. Data were fitted with the double-exponential decay function in Equation~(\ref{eq:attenuation2}) with $n_\mathrm{PE}(0)$, $\lambda_1$ and $a$ as the only free parameters; fit results are displayed in the tables on the right.} \label{fig:average} \end{figure} The single-muon generator is, therefore, a random generator following a Poissonian distribution whose parameter is the mean number of PEs measured in a given position on the scintillator strip, extracted from the charge attenuation parametrization presented in Fig.~\ref{fig:average}. Having drawn the number of PEs, we then generate for each a pulse with the PE generator, described previously. The start time of each PE signal is determined using two random generators following exponential decays, with the characteristic decay times of the scintillator and optical fiber. By adding all the PE signals within the corresponding time profile, the single-muon signal is obtained. As an example, we present in the left panel of Fig.~\ref{fig:muonPulse} a simulated muon signal at 2\,m on the scintillator strip (red) with its individual PEs (gray) re-scaled to ease the visualization: the different start times of the PEs determine the time structure of the muon signal. In the same figure, we also present, only for illustrative purposes, a selected signal measured in the laboratory also at 2\,m (blue). \begin{figure}[t] \centering \begin{subfigure}{1.00\textwidth} \centering \includegraphics[trim={0.0cm 0.0cm 0.0cm 0.0cm},clip,width=1.0\linewidth]{./img/muonPeaks/muonExample.pdf} \end{subfigure} \caption{(Left) example of simulated muon signal at 2\,m of the scintillator strips (red), with its single-PE pulses re-scaled to ease the visualization (gray), along with measured signal at the same position (blue). Ticks on the top x-axis represent the start time of the PE pulses. (Right) correlation between the muon signal charge and amplitude for two thousand events of both laboratory data and simulation.} \label{fig:muonPulse} \end{figure} The correlation between the muon signal charge and its amplitude depends on the arriving time of the photons, which in turn, determines the signal correlation between binary and ADC channels. In this sense, it is of utmost importance to provide an accurate simulation of the correlation between amplitude and charge, as we present in the right panel of Fig.~\ref{fig:muonPulse}. The signal charge as a function of its maximum amplitude is presented for measured (blue circles) and simulated (red squares) muon signals at 2\,m on the scintillator strip. Both simulations and measurements are consistent around the means, demonstrating that the amplitude: charge correlation is directly detached from the proper simulation of the photon time profile. A slight difference of less than 5\% in the slopes is identified, along with outliers in laboratory data for high and low amplitudes that are not reproduced in the simulation. Nevertheless, these represent roughly 7\% of the data, and, as it will be presented in Section~\ref{sec:validation}, does not have an impact on the simulation accuracy. Furthermore, these differences are expected since the number of PEs in the simulation is extracted using Equation~(\ref{eq:attenuation2}), and large fluctuations in the energy deposit produced from the landau tail are not considered. The completion of a generator for single-muon signals concludes the analog part of the simulation. The following steps, as presented in the next Subsection, consists of the electronics processing and sampling for which transfer models and digital data were used. \subsection{Digital simulation: binary and ADC acquisition modes} \label{subsec:simeKIT} In the binary mode, the 64 SiPM outputs are handled individually, first with a pre-amplifier and a fast-shaper. To simulate these two steps, we implemented electric models based on simplified circuits~\cite{marano}: the pre-amplifier is modeled as a 10\,MHz active low-pass filter, with an amplification factor of 10x, and the fast-shaper as a practical differentiator with a characteristic time of 15\,ns and a maximum gain of about 18x. The circuit components were selected to match the mentioned amplification and timing features and then tuned to obtain the best representation of the laboratory data used to validate the simulations (see section~\ref{sec:validation}). The fast-shaper output is then processed with a discriminator and sampled with an FPGA into a 2048-bit trace. In a rough approximation, the discriminator outputs a signal of 3.3\,V if the fast-shaper is above the discriminator threshold and 0\,V otherwise. Then, the FPGA outputs a ``1'' if the discriminator signal is above 1.7\,V and a ``0'' if it is below 0.8\,V. In more detail, the discriminator has a typical rise time that affects the signal sampling: if one of the FPGA flip-flops is clocked when the discriminator signal is between 0.8\,V and 1.7\,V, then the FPGA output is undetermined. To properly reproduce this in the simulation, we first estimated the rise time of the discriminator as the time that signal needs to rise between the baseline and 1.7\,V, using two thousand analog single-muon pulses measured at the output of the discriminator. We obtain a mean rise time of (1.51 $\pm$ 0.01)\,ns, and we assumed that if the FPGA samples during this rise time, then a ``0'' is output. On the other hand, if the fast-shaper output is above the discriminator threshold during more than 1.51\,ns when sampling, then a ``1'' is output. The transition time between 0.8\,V and 1.7\,V is rather negligible: different approaches were tested for the FPGA indetermination resulting all in the same outcome. In the left panel of Fig.~\ref{fig:exampleADC} we present an example of a simulated single-muon signal in the binary channel, where the re-scaled SiPM output (solid blue line) along with the output of the fast shaper (dotted-dashed black line) is displayed. The discriminator threshold set at 2.5\,PEs, as in the UMD modules, is presented (dotted pink line) along with the discriminator output (dashed green line) which was re-scaled to fit in the figure. When the fast-shaper output is above (below) the threshold during more than 1.51\,ns, at the output of the discriminator a 3\,(0)\,V signal is obtained and a ``1'' (``0'') is sampled in the binary trace for the corresponding time. \begin{figure} [t] \centering \begin{subfigure}{1.0\textwidth} \centering \includegraphics[trim={0.0cm 0.0cm 0.0cm 0cm},clip,width=1.0\linewidth]{./img/CounterSim/exampleSignal2.pdf} \end{subfigure} \caption{Two simulated single-muon signal at 2\,m for the binary (left) and ADC (right) modes. In the first case, the signal from the SiPM and the discriminator pulse are re-scaled for illustration.} \label{fig:exampleADC} \end{figure} In the ADC mode, the electronics chain starts with two steps of adders that sum all the 64 SiPM analog signals, and an amplification step with a low- and high-gain amplifier follows. The amplifier outputs are sampled with an ADC at a speed of 160\,MHz, with a step of about 0.122\,mV/bit into two waveforms of 1024 samples. If we consider the components cutoff frequency to be the frequency for which the output of the circuit is $-3$\,dB (30\% fall in magnitude) of the nominal passband value, the first step of adders determines the signal timing: they introduce a frequency cutoff of about 12\,MHz~\cite{UMDFrontEnd}, while the other steps of the electronics introduce a frequency cutoff above 100\,MHz, with the bandwidth of the PE signal mostly between 15\,MHz and 2\,GHz. In this sense, the electronics response after the first step of adders is rather flat in frequency, and the whole transfer chain can be simplified to an active low-pass filter with a frequency cutoff of 12\,MHz, and with an amplification equivalent to the net gain of all the steps, $-2.50$($-8.47$)\,dB for the high(low)-gain channel~\cite{UMDFrontEnd}. Similar to the binary mode, these parameters were all tuned to match the laboratory data, and the result is presented in the next section. An important feature in the ADC channel is the baseline fluctuation due to the SiPM dark counts, which has a significant impact on the signal charge estimation and detector resolution. For this reason, we included these fluctuations in the same way as we did for the PE analog signal: we determined the baseline offset and its fluctuations after the analog electronics processing and we input this information into a random generator that injects the fluctuations in the signal following a normal distribution. The baseline mean and standard deviation resulted in (8210.24 $\pm$ 0.02)\,ADC\,counts, and (3.24 $\pm$ 0.02)\,ADC\,counts for the low-gain channel, and in (8242.6 $\pm$ 0.1)\,ADC\,counts and (12.4 $\pm$ 0.1)\,ADC\,counts for the high gain. The final step of the simulation of the ADC mode is the sampling, for which we grouped the amplifier outputs in bins of 6.25\,ns, and we extract the maximum amplitude reached in each bin. We then obtain the ADC\,counts per sample by diving the signal amplitude by the ADC step. An example of a simulated signal in the ADC channel is presented in the right panel of Fig.~\ref{fig:exampleADC}, where the SiPM output for one impinging muon produced at 2\,m is displayed (blue line and left y-axis), along with the output for the low- (orange) and high-gain (red) channels after the ADC sampling (stars and right y-axis). \section{End-to-end validation of simulated data} \label{sec:validation} The UMD simulation must provide an accurate representation of the signal parameters relevant to the estimation of muon densities in air showers. In this section, we discuss the validation performed for both analog and digital signals to guarantee that this requirement is fulfilled. Considering the analog muon signals, the main features to assess are the maximum amplitude which has an impact on the performance of the binary mode, the signal charge, which has an impact on the ADC mode, and the signal width, which has an impact on the general timing of the signal. To verify that the analog simulation achieved in the previous section accurately describes these signal features, we present in Fig.~\ref{fig:muonPulsesSim} the maximum amplitude (top left), charge (top right), and full width at half maximum (bottom) of the single-muon signals as a function of the position where the muon impinged on the scintillator strip for two thousand muon signals at each position obtained with the muon telescope. In each panel, we present the results for the mean (orange up triangles) and standard deviation (pink crosses) measured in the laboratory, along with the mean (blue down triangles) and standard deviation (green plus signs) from the simulation. The average amplitude and charge of about 8\,mV and 5\,pC corresponds to the addition of the amplitude and charge of the PEs previously presented in Fig.~\ref{fig:average}: the amplitude (charge) of the muon signal is about 10\,(22)\,PE, each of which has an amplitude of 0.80\,mV, and a charge of 0.23\,pC. Similarly, the attenuation in charge and amplitude corresponds to the attenuation in the number of PEs presented in Fig.~\ref{fig:average}. The signal width is the result of the convolution between the time profile of the arriving photons and the time response of the SiPM. \begin{figure}[t] \centering \begin{subfigure}{1.00\textwidth} \centering \includegraphics[trim={0.0cm 0.0cm 0.0cm 0.0cm},clip,width=1.0\linewidth]{./img/muonPeaks/AmplitudeChargeFWHA.pdf} \end{subfigure} \caption{Main features of the muon signals at different positions on the scintillator strip. The means and standard deviation for laboratory data and simulations are shown. We display on the top-left panel the muon signal amplitude, in the top-right panel the muon signal charge, and in the bottom, the muon signal full width at half maximum (FWHM). The agreement between the data and the simulations is apparent.} \label{fig:muonPulsesSim} \end{figure} Understanding the signal fluctuations is of utmost importance to assess the performance of the UMD. The variations in charge determine the resolution of the ADC channel to the number of impinging particles, while those in amplitude impact in the binary-mode efficiency, since signals that fluctuate to amplitudes below the discriminator threshold, do not produce ``1''s in the binary trace, and therefore, are not detected. In turn, the fluctuations in the signal width play a critical role when defining the reconstruction and calibration strategy for the UMD~\cite{UMDCalibration}. Therefore, the simulation needs to provide an accurate representation not only of the means of the parameters but also of their variation portrayed with the standard deviations of the distributions. From this stage on, (almost) no further fluctuations are introduced, since most stochastic processes involved in the UMD signal generation are found in the scintillator photo-production, the absorption and re-emission of the optical fiber, and in the SiPM photo-detection. The agreement between laboratory data and simulations is apparent in Fig.~\ref{fig:muonPulsesSim}, which denotes that the muon generator output provides a good description of the main features of the analog muon signals. The main differences between data and simulations are found in the charge fluctuation at distances close to the SiPM. These are produced by a few signals with a very low or large charge that are not reproduced in the simulation, which represents roughly 5\% of the data. Nonetheless, the fluctuations at this stage are compared to assess the general performance of the analog outputs and do not include the fluctuation in the baselines introduced by the SiPM thermal noise, which has a large impact at the end of the simulation chain. These are included at the last step of the simulation to recover the charge fluctuations relevant to the detector performance. The analog output is only the input of the read-out electronics, and as we showed in Fig.~\ref{fig:exampleADC}, after the processing, the traces become significantly different. By implementing the transfer function of the binary mode, we can simulate dark-count rates as in Fig.~\ref{fig:calibrationPEGen}. Using the PE generator, we create pulses with a 2.2\% probability of having a cross-talk avalanche, and we processed the output with the simulation chain of the binary mode. To build the curve, we increase the count of the rate if the resulting binary trace has at least one ``1''. We compare, in Fig.~\ref{fig:counterSignal}, the simulated dark-count curve with the curve measured in the laboratory, for which we normalized the number of counts to the number of pulses generated with the simulation. As a result, we demonstrated that the procedure to obtain the PE amplitude can be accurately simulated obtaining similar values as in the laboratory; this procedure is the ground of the SiPM calibration. To obtain a ``1'' at the FPGA output the fast-shaper signal needs to be over the discriminator threshold during more than 1.51\,ns (see section~\ref{subsec:simeKIT}), due to the effect of the discriminator rise time. To better understand the impact of this effect on the SiPM calibration, we compared the single-PE amplitude obtained in the simulated calibration with that of the fast-shaper analog output. The mean PE amplitude is in this case (35.39 $\pm$ 0.04)\,mV, and for the simulated calibration (24.3 $\pm$ 0.3)\,DAC\,counts. Since the DAC step of the discriminator threshold equals 1.3\,mV, the amplitude obtained is (31.6 $\pm$ 0.4)\,mV. When the fast-shaper output barely reaches the discriminator threshold, the discriminator signal does not have time to rise and no ``1'' is recorded in the binary trace. It is then necessary to lower the threshold below the amplitude of the PE signal to effectively obtain a ``1'' in the binary trace; this dynamic is the reason for the underestimation of the PE amplitude with a negative bias of approximately 10\% when using the full chain of the binary channel. This 10\% bias does not represent a problem for the UMD calibration or performance, since the whole process for determining the optimal operation setup of the modules was performed within this same system~\cite{AMIGASIPM}, so the bias was consequently taken into account during the performance and calibration studies. \begin{figure}[t] \centering \begin{subfigure}{0.45\textwidth} \centering \includegraphics[trim={0.0cm 0.0cm 0.0cm 0.0cm},clip,width=1.0\linewidth]{./img/CounterSim/DRCurve.pdf} \end{subfigure} \caption{Dark-count rate curve used to calibrate the SiPM. We show the curves obtained using both laboratory data and the binary mode simulation. The single-PE amplitude obtained with the simulation is ($24.3 \pm 0.3$)\,DAC\,counts.} \label{fig:counterSignal} \end{figure} To obtain the number of muons using the binary channel, a reconstruction strategy based on the number of ``1'' in the traces needs to be determined to both reject the noise and preserve the detection efficiency as high as possible. To characterize the signal width, we present in the left panel of Fig.~\ref{fig:counterResults} the number of ``1''s as a function of the position on the scintillator strip obtained with the muon telescope. We present the signal mean widths (blue up triangles) and standard deviations (green plus signs), for laboratory measurements (full markers) and simulation (empty markers). Due to the optical-fiber attenuation, the signal mean width decreases when moving the muon telescope towards the end of the strips, which increases the total fluctuation over the whole strip and challenges the selection of the reconstruction strategy. If we consider all the widths within a three\,standard\,deviations from the mean, most of the muon signals in the laboratory have between 12 (37.5\,ns) and 4 (12.5\,ns) ``1''s. Furthermore, pulses produced by SiPM dark counts that reach the 2.5\,PE threshold have a typical width of less than 12.5\,ns~\cite{UMDCalibration,UMDICRC}. According to this, requesting a minimum width of 12.5\,ns effectively rejects 99\% of the SiPM noise without a significant loss of muon signals (about 1\%). With this reconstruction strategy, the probability of over-counting one muon in one event is reduced by a factor of about 2.7 with respect to a strategy in which a minimum width of 3.125\,ns (one sample) is requested to count as a muon. It is worth mentioning, that this estimations are valid in the laboratory where the effects of the soil or of clipping-corner muons are not taking place~\cite{MuonsWithAMIGA}. \vspace{0.5cm} \begin{figure}[t] \centering \begin{subfigure}{1.00\textwidth} \centering \includegraphics[trim={0.0cm 0.0cm 0.0cm 0.0cm},clip,width=1.0\linewidth]{./img/CounterSim/widthCharge.pdf} \end{subfigure} \caption{(Left) mean signal width over sample time and standard deviation in the binary mode as a function of the position on the scintillator strip with both laboratory and simulated data of single muons. The signal widths were obtained also using two different discriminator thresholds: 2.5 and 3.5\,PEs. (Right) Signal charge as a function of the position on the scintillator strip. We show the signal mean charge and standard deviation obtained using laboratory data. We also display the result using the ADC mode simulation.} \label{fig:counterResults} \end{figure} The results shown in the left panel of Fig.~\ref{fig:counterResults} demonstrate that the simulation is accurately reproducing the binary traces measured in the laboratory: we obtained a maximum difference of less than 10\% in the mean width at the end of the strip, having the difference in the mean over the whole strip less than 4\%. To further validate the simulation, we have also performed this measurement with a discriminator threshold of 3.5\,PEs, for which the signal width is reduced on the order of about half a sample with respect to the 2.5\,PE threshold. We then present, also in the left panel of Fig.~\ref{fig:counterResults}, the results for the mean widths (orange down triangle) and standard deviation (pink crosses) of the laboratory data (full markers) and simulations (empty markers). The consistency between simulation and data denotes the potential to assess different configurations that eventually may be considered in the data taking for dedicated studies on the detector performance. It is also worth mentioning that when raising the discriminator threshold to 3.5\,PE, the number of SiPM dark counts reaching the threshold is reduced to 2.2\% with respect to the previous setup. In this sense, a muon can be considered as any ``1'' in the trace without significantly increasing the probability of over-counting a muon because of noise, but only equal to the previous result when considering three ``1''s. The main feature in the ADC mode relevant to the UMD data reconstruction is the signal charge and its fluctuations presented in the right panel of Fig.~\ref{fig:counterResults}: to estimate the number of muons with the ADC channel we compute the signal charge and divide it by the mean charge of a single-muon. The mean signal charge for the low- (red circles) and high-gain (orange squares) channels along with their standard deviations (green crosses and blue plus signs, respectively) are presented as a function on the position of the scintillator strip for laboratory data (full markers) and simulations (empty markers). The signal charge was computed by adding the ADC\,counts in the muon signal of every sample above one standard deviation off of the baseline, and the average is presented for two thousand signals obtained using the muon telescope at each position. It is apparent how the PE attenuation has an important impact on the detector resolution since the total charge of single-muon signals between the beginning and end of the strip differs by a factor of two. The average charge over the whole strip for the low- and high-gain channels is (262 $\pm$ 5)\,ADC\,counts and (1059 $\pm$ 19)\,ADC\,counts with standard deviations of (141 $\pm$ 2)\,ADC\,counts, and (551 $\pm$ 8)\,ADC\,counts respectively. This translates into a resolution of about 60\% for single-muon signals in the ADC channel, which significantly diminishes with the arrival of many muons in the same module, the regime for which the ADC mode was designed. Similar to the binary mode, the simulation of the ADC channel, presented in the right panel of Fig.~\ref{fig:counterResults}, accurately reproduces the signal charge and its fluctuations for both high- and low-gain channels. We obtained a maximum difference in the single-muon charge of less than 15\% in the high-gain channel, and less than 10\% in the low-gain channel for the mean at the end of the strip, being the difference in the average over the whole strip of less than 3\% for both means and standard deviations. To reproduce the UMD performance realistically, the simulation must replicate the muon detection efficiency of the binary mode and the saturation of the ADC mode, which determines the dynamic range of the detector. In the top-left panel of Fig.~\ref{fig:efficiency} the detection efficiency of the binary mode for single muons is shown as a function of the position on the scintillator strip of the impinging particle. The efficiency is defined as the number of detected muons, using the reconstruction strategy previously described, over the total number of triggered events. It was estimated using laboratory data (blue full circles and orange full squares), simulations (blue up triangles and orange down triangles), and a Poissonian prediction (blue empty circles and orange empty squares) based on the measured number of PEs as a function of the position on the strip. This prediction consists of integrating a Poissonian distribution between the discriminator threshold (2.5 or 3.5\,PEs) and infinity, where the mean number of PEs at each position on the strip is extracted accordingly to measurements shown in Fig.~\ref{fig:average}. The Poissonian prediction shows that the efficiency loss, mainly at the back of the strip, is principally produced by muon signals with amplitude fluctuating below the discriminator threshold. Therefore, the agreement between data and prediction denotes that there is not a significant efficiency loss in the signal processing or in the reconstruction analysis, having an integrated efficiency of 98.5\% over the whole scintillator strip\footnote{The discriminator threshold for the UMD is set at 2.5\,PE. Measurements with 3.5\,PE were obtained in the laboratory to validate the simulation with a different operating point, but are not foreseen in the field. The integrated detection efficiency, in this case, is roughly 96\%.}. In addition, it is worth mentioning that the simulation provides an accurate description of the binary mode efficiency for single muons. \begin{figure} [t] \centering \begin{subfigure}{1.00\textwidth} \centering \includegraphics[trim={0.0cm 0.0cm 0.0cm 0.0cm},clip,width=1.0\linewidth]{./img/CounterSim/effiiciencySaturation.pdf} \end{subfigure} \caption{(Top) Efficiency of the binary mode as a function of the position on the scintillator strip estimated with laboratory data, simulations, and a Poissonian prediction using discriminator thresholds of 2.5 and 3.5\,PEs. (Bottom-left) number of triggered segments in the binary mode (red squares) and reconstructed muons (blue circles) after a pile-up correction in the binary mode as a function of the number of injected muons. The identity function is represented with a dashed-gray line. (Bottom-right) Mean signal charge of the low- (blue circles) and high-gain (orange squares) channels of the ADC mode simulation. To ease the interpretation we present in dashed lines the theoretical linear relation between the number of injected muons and signal charge.} \label{fig:efficiency} \end{figure} To assess the saturation in the binary mode, we performed a toy Monte Carlo which consisted of injecting simultaneous particles randomly in a 64-segment detector and counting the number of segments hit by particles. We then determined the number of reconstructed particles by applying a pile-up correction~\cite{AMIGALDF} which acknowledges the probability of having multiple particles hitting the same segment. We present the results in the bottom-right panel of Fig.~\ref{fig:efficiency}: the mean number of triggered segments (red squares) and reconstructed muons (blue circles) as a function of the number of injected particles is shown, along with the identity function (dashed-gray line). The maximum number of segments that can be hit (dotted-gray line) corresponds to the 64 scintillator strips in a UMD module. At about 250 injected particles, the mean number of reconstructed muons begins to differ 5\% from the identity. It is worth mentioning that this toy Monte Carlo does not acknowledge for corner-clipping muons and knock-on electrons~\cite{MuonsWithAMIGA} from the soil that may produce additional hits on the segments. Finally, in the bottom-right panel of Fig.~\ref{fig:efficiency} we show the mean signal charge of the low- (blue circles) and high-gain (orange squares) channels as a function of the number of muons injected simultaneously and simulated with the ADC mode. To ease the interpretation we present in dashed lines the theoretical linear relation between the number of injected muons and signal charge. To perform the simulation, the position of the muon on the scintillator strip was randomly drawn from a uniform distribution between 1\,m, and 5\,m. If we identified the saturation point where the signal charge is 5\% below the linearity with respect to the number of injected muons, the saturation for the high-\,(low-)gain channel occurs, roughly, at 81\,(357)\,muons per module. These results are quite consistent with previous tests in the laboratory~\cite{UMDFrontEnd} (85 and 362 respectively) where a light source was used to emulate the signal of several muons arriving at exactly the same time. \section{Summary} \label{sec:conclusion} \begin{table}[t] \centering \renewcommand{\arraystretch}{1.25} \begin{tabular}{c c c c c} \toprule[1.25pt] \multicolumn{2}{c}{Parameter} & Laboratory & Simulation & Ratio\\ \midrule \multirow{2}{*}{Binary Width / 3.125 ns} & Mean & 7.75 $\pm$ 0.01 & 8.12 $\pm$ 0.03 & 0.95 \\ & Std. Dev. & 1.49 $\pm$ 0.01 & 1.51 $\pm$ 0.02 & 0.99\\ \multirow{2}{*}{Charge LG / ADC counts} & Mean & 262 $\pm$ 5 & 269 $\pm$ 3 & 0.97 \\ & Std. Dev. & 141 $\pm$ 2 & 136 $\pm$ 2 & 1.04 \\ \multicolumn{2}{c}{Binary efficiency } & 99\% & 98\% & 1.01\\ \multicolumn{2}{c}{Saturation LG (in 10 m$^{2}$) / muons} & 362 & 357 & 1.01 \\ \bottomrule \end{tabular} \caption{Comparison of features relevant to the UMD event reconstruction and performance. The results are displayed for laboratory measurements and simulation, along with the ratio between them. } \label{table:summary} \end{table} Currently, the Pierre Auger Observatory is undergoing a major upgrade whose aim, among others, is to determine the composition of primary cosmic rays at the highest energies. The underground muon detector (UMD) is an essential part of this upgrade and is meant to obtain direct measurements of the muonic component in air showers for cosmic rays with energies between $10^{16.5}$ up to the region of the ankle (about $10^{18.7}$\,eV). To determine the number of hadrons in the primary particle using the information on the air shower muons, the UMD data ought to be properly interpreted at high-level physics analysis. To this aim, we needed to develop a reliable simulation tool that accurately reproduces the response of the detector to single particles. In this work, we have presented a detailed description of the simulation sequence developed for the basic unit of the UMD detector: one scintillator strip with optical fiber and its corresponding optoelectronics. The models used in the simulation chain have been tuned and thoroughly validated using laboratory data. We present a summary of the main features relevant to the estimation of the number of impinging muons with the UMD and the detector performance in Table~\ref{table:summary}. The results for laboratory data and simulations, along with the ratio between these, are presented for the single-muon width and detection efficiency in the binary channel, and the signal charge and saturation in the ADC low-gain channel. The ratios show a maximum difference of 5\% proving the accuracy of the simulation at reproducing the relevant features of the signals. We have established that single-muon signals in the binary mode have a typical width between 12.5\,ns and 37.5\,ns. Using this as a condition to select muons, 99\% of the SiPM noise is effectively rejected while keeping an integrated efficiency of $99$\%. For the ADC mode, we have shown that the high-gain channel is capable of measuring up to about 80 muons arriving at exactly the same time, while the low-gain channel extends this number up to about 350 muons per 10\,m$^2$, results consistent with previous studies. Finally, it is worth mentioning that all the algorithms and models discussed in this article have been already implemented in \Offline, the official framework for analysis and simulation of the Pierre Auger Observatory~\cite{offline}. \acknowledgments The authors thank the support of the Pierre Auger Collaboration, in particular its Publication Committee for reviewing this article. They also thank the technical staff of the Instituto en Tecnologías de Detección y Astropartículas for aiding on the built of the setup used in this work. The corresponding author of this article has a post-doctoral grant from the Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET) of Argentina. \bibliographystyle{JHEP}

e29ac232f72e1c7db763719f71c702e6c75716b2

\section{Introduction} \label{sec:intro} Nonequilibrium thermodynamics traditionally studies both the steady-state flow of a system maintained out of equilibrium, e.g., by a temperature gradient, and the detailed approach to equilibrium of transient fluctuations \cite{Chapman_and_Cowling,deGroot,deGroot_and_Mazur}. In this paper we focus on the latter phenomenon, that is, the time evolution of small fluctuations from a certain equilibrium state, defined as the long-time expected or \textit{a priori} most-likely value of a given set of macroscopic variables. One way of quantifying this transient behavior is through the determination of transport coefficients, which may be used to characterize the rate of approach to equilibrium. A well-known example of such transport coefficients is provided by the linear theory of Onsager \cite{Onsager1931a,Onsager1931b} relating macroscopic flows to thermodynamic forces. This relation, known as the Onsager principle, has become a cornerstone of nonequilibrium statistical mechanics and has seen wide application in a diverse range of fields. The purpose of this paper, however, is to establish Onsager's principle on a rigorous basis and in a fairly general context. In doing so, we hope to understand more clearly its meaning and range of validity. The key insight to our approach is the introduction of a generalized free energy function which may be used to define the time-dependent macroscopic variables of interest. This approach has been applied successfully to the broader problem of macroscopic determinism in interacting and noninteracting systems \cite{LaCour_and_Schieve2002,LaCour_and_Schieve2000} and relies chiefly on rigorous results from the mathematical theory of large deviations \cite{LaCour_and_Schieve2002,Dembo_and_Zeitouni,Deuschel_and_Stroock,Ellis}. Under suitable regularity conditions, one may construct a mapping, $\psi_t$, from an initial macrostate $a_0$ at time zero to a predicted macrostate $\psi_t(a_0)$ at time $t$. It may be shown that the actual macroscopic observable, $G_{n,t}$, considered as a phase function for a system with finitely-many degrees of freedom, converges in probability to the value $\psi_t(a_0)$ when conditioned on the initial value $a_0$. The approach used here is simply to expand $\psi_t(a_0)$ in a Taylor series about the equilibrium macrostate and initial time. Assuming the function is smooth enough that this may be done, identification of the equilibrium large deviation rate function with the thermodynamic entropy allows us to derive, for short times, the linear Onsager relation between macroscopic flows and thermodynamics forces. Section~\ref{sec:NET} begins with a brief overview of the Onsager principle in the context of nonequilibrium thermodynamics and discusses prior work in its statistical formulation and derivation. Section~\ref{sec:OLR} contains the main results of this paper, which is a derivation of the Onsager principle, valid in the short-time, near-equilibrium regime, using techniques from large deviation theory. In Sec.~\ref{sec:GKF} a Green-Kubo formula for the Onsager matrix in terms of the covariance matrix of the macroscopic variables is derived and, in the case of time reversal invariance, shown to be symmetric. The forward Onsager matrix is shown to be positive semidefinite as a consequence of measure preservation, and the positivity of the entropy production rate is discussed in relation to the second law of thermodynamics. We end in Sec.~\ref{sec:IGE} with the example of an ideal gas undergoing specular reflection in a closed box. The Onsager coefficient for the fraction of particles on one side of a uniform partition is computed in terms of the \textit{a priori} velocity distribution. Using Maxwellian and uniform velocity distributions, respectively, the time-dependent entropy is computed numerically and the results discussed. Concluding remarks may be found in Sec.~\ref{sec:Discussion} \section{Nonequilibrium Thermodynamics} \label{sec:NET} Before addressing the problem of transport properties, we must first establish a connection between the various mathematical objects we have considered thus far and the more familiar objects of nonequilibrium thermodynamics. The traditional approach takes concepts and relations from equilibrium thermodynamics and assumes a range of validity into the nonequilibrium regime. This extension is largely justified by empirical success, and generally works best when the system is near equilibrium. Far from equilibrium, well-defined dissipative structures may form and evolve in time \cite{Prigogine1980,Prigogine1967}, but such structures are not the object of study here. Consider a system for which we are interested in the time evolution of one or more macroscopic observables. Let $\alpha(t)$ be the vector-valued observable at time $t$. (Later, we shall identify $\alpha(t)$ with the asymptotically most probable macrostate at time $t$.) At equilibrium the entropy, $S$, is of course a well-defined function of the given macroscopic variables. By replacing the equilibrium macrostates with $\alpha(t)$, we obtain a time-dependent entropy $S(\alpha(t))$. If $S$ and $\alpha$ are differentiable, then from the chain rule we have the well-known formula \begin{equation} \dot{S}(\alpha(t)) = X(\alpha(t)) \, J(t), \label{eqn:entropy_rate} \end{equation} where $X = \transpose{\nabla}S$ is the so-called \emph{thermodynamic force} and $J = \dot{\alpha}$ is the time rate of change or \emph{flow} of the observable. (We use $\nabla$ to denote a column vector gradient and $\transpose{\nabla}$ for a row vector gradient.) In what follows, $J(t)$ will be considered a $d$-dimensional column vector, while $X(\alpha(t))$ is considered a row vector. With the aforementioned differentiability assumptions, Eq.\ (\ref{eqn:entropy_rate}) is nothing more than a definition of $X$ and $J$. In the absence of phase transitions, it is indeed reasonable to suppose $S$ is a differentiable function of the macrostates. The differentiability of $\alpha$, however, is less clear, and this difficulty stems largely from the unclear and imprecise meaning of $\alpha(t)$ itself. An obvious candidate is for $\alpha(t)$ is the actual, time-varying observable, $G_{n,t}$. In this case, $J(t)$ is the instantaneous time derivative, if it exists, of a particular realization of the corresponding stochastic process. However, such detailed time variation surely cannot be part of a general macroscopic law. In his 1952 paper, M. Green \cite{Green1952} considered using the ensemble average, $\avg{G_{n,t}}_{a_0}$, of $G_{n,t}$ conditioned on $G_{n,0} = a_0$ as a macroscopic variable, implicitly identifying this with its large-$n$ probabalistic limit. For macroscopic flows, Casimir \cite{Casimir1945} and, later, Mori \cite{Mori1958} suggested that $J(t)$ be defined as the time average rate of change of this ensemble average over a mesoscopic time $\tau$, i.e., one intermediate between the time between collisions and the relaxation time of the system as a whole. More specifically, they proposed that $J(t)$ be defined as \begin{equation} J(t) = \frac{\avg{G_{n,t+\tau}}_{a_0} - \avg{G_{n,t}}_{a_0}}{\tau}, \end{equation} where $\avg{\,\cdot\,}_{a_0}$ is an expectation with respect to the initial conditional distribution for which $G_{n,t_0} = a_0$. The difficulty with such a proposal, however, is that it requires an explicit time scale be incorporated into the basic definition of $J(t)$. Since what is desired of $J(t)$ is a measure of the typical, macroscopic rate of change of the observables, a reasonable and more elegant proposal is to equate $\alpha(t)$ with $\psi_t(a_0)$, the large-$n$ expected value of the observable. This has the twin virtue of remaining faithful to the actual dynamics, since we assume $G_{n,t}$ converges to $\psi_t(a_0)$ in probability, while remaining insensitive to detailed time variations by taking, in effect, an ensemble average. In addition, this choice is free of the need to specify an ad hoc time scale for $J(t)$. Of course, for $J(t)$ to be well defined we must still have that $\psi_t(a_0)$ is differentiable in $t$, but that is a far more reasonable assumption. (Later we shall see that the one-sided time derivatives at the initial time, $t=0$, will usually not agree, so care must be taken here to distinguish between forward and reverse time derivatives.) In 1931, Lars Onsager \cite{Onsager1931a,Onsager1931b} proposed a simple, linear relationship between flows and forces based on a generalization of well-known phenomenological laws such as the thermoelectric effect studied earlier by Thompson. Onsager postulated a relationship of the form \begin{equation} J(t) = L \, \transpose{X(\alpha(t))}, \label{eqn:Onsager_relation} \end{equation} where $L$ is known as the \emph{Onsager matrix}. Using time reversal invariance (or ``microscopic reversibility,'' in his terminology), Onsager showed that $L$ is symmetric for observables which are even functions of the momentum, a result known as the Onsager reciprocity relation. The papers by Casimir \cite{Casimir1945} and Wigner \cite{Wigner1954} give a detailed discussion of this latter derivation. Elements of the Onsager matrix may be identified with various transport coefficients associated with the macroscopic variables of interest. These, in turn, may be derived in terms of the correlations between macrostates at different times. The term ``Green-Kubo formula'' is here refers to any formula giving elements of the Onsager matrix in terms of time-correlations of observables, owing to the early work of Green \cite{Green1950} and Kubo \cite{Kubo1957} in deriving such relations. Recently, Oono \cite{Oono1993} has reconsidered the Casimir-Mori proposal in the context of large deviation theory to obtain the Onsager relation of Eq.~(\ref{eqn:Onsager_relation}). Like Mori, Oono considers a macroscopic time evolution which is coarse-grained in time according to the postulated mesoscopic time scale $\tau$. The approach he takes is to write the macroscopic variable $\alpha(t)$ as a time average over the interval $[t,t+\tau]$ of the observable denoted here by $G_{n,t}$. (The macroscopic flow is defined similarly.) A large deviation principle for the empirical time average (valid for $\tau$ large but $n$ fixed) is then assumed, from which an LDP for the macroscopic flow is obtained via the contraction principle \cite{Dembo_and_Zeitouni}. To get the right initial distribution, Oono appeals to information theory to argue that the most likely macroscopic flow should be in the form a canonical expectation in which, following Mori, the Lagrange multipliers are determined by the given initial flow. (By contrast, this canonical form was derived in \cite{LaCour_and_Schieve2002} using a conditional LDP theorem \cite{LaCour_and_Schieve2002}.) The connection to the traditional thermodynamic quantities is then made in the usual way by equating the large deviation rate function with the thermodynamic entropy using the Einstein fluctuation formula. The result which Oono obtains is \begin{equation} \begin{split} L &= \int_0^{\tau} \avg{\dot{G}_{n,s}\transpose{\dot{G}_{n,0}}} \d s \\ &= \avg{[G_{n,\tau}-G_{n,0}]\dot{G}_{n,0}} \xrightarrow{\tau\to\infty} -\avg{G_{n,0}\transpose{\dot{G}_{n,0}}}, \end{split} \label{eqn:Oono} \end{equation} assuming $\avg{G_{n,0}} = 0$ and the correlation $\avg{G_{n,\tau}\dot{G}_{n,0}}$ vanishes for $\tau$ large. Large deviation theory has also been used recently by Bertini et al.\ \cite{Bertini2002} to describe macroscopic fluctuations in the density field of a Markovian lattice. In the appropriate scaling limit, fluctuations of the asymptotically most likely density field, analogous to our $\psi_t(a_0)$, are shown to satisfy the Onsager-Machlup principle; i.e., the fluctuations are time symmetric. The present work differs in that we consider a general, deterministic microscopic dynamics rather than an underlying Markovian microstate. Our approach, however, will accomodate a fundamentally stochastic system as well as it will a deterministic one. Finally, we note that Gallavotti \cite{Gallavotti1996} has recently derived the Onsager relations and Green Kubo formula for thermostated systems as a consequence of the fluctuation theorem\cite{Gallavotti_and_Cohen1995a,Gallavotti_and_Cohen1995b}. Gallavotti's focus is on small fluctuations from zero forcing in a steady state system whereas ours is on small fluctuations of select observables from equilibrium. Critical to his results is the assumption that the dynamics of the system satisfy the chaotic hypothesis, from which the fluctuation theorem may be inferred. By contrast, the validity of our approach rests solely on the differentiability of the dynamical free energy, as defined in Eq.\ (\ref{eqn:DFE}) below. \section{Onsager Linear Relations} \label{sec:OLR} In this section, the Onsager linear relations will be derived for the short-time, near-equilibrium regime. For a system with $n$ degrees of freedom, let $X_n$ denote the set of microstates along with its associated topology. If, say, $x_0$ is the initial microstate of the system at time zero, then the state at time $t$ is denoted by $\Phi_{n,t}(x_0)$, where $\Phi_{n,t}$ is a Borel measurable transformation on $X_n$. For simplicity we shall restrict attention to cases for which the set $\set{\Phi_{n,t}}_{t\in\Rset}$ forms a group, though many results generalize without this assumption. By $G_n$ we denote the macroscopic observable of interest, which is assumed to be a Borel measurable function from $X_n$ to $\Rset^d$. Given an initial microstate $x_0$, then, $G_n(x_0) = a_0$ is the initial value of all macroscopic variables under consideration. As only $a_0$, not $x_0$, is given, an \textit{a priori} probability measure $P_n$ over the Borel subsets of $X_n$ is assumed to be valid and invariant under $\Phi_{n,t}$. This last point introduces a statistical description of the system. \subsection{Large Deviation Results} The sequence of macroscopic observables $\seq{G_n}_{n\in\Nset}$ is assumed to scale as $v_n$, where $\seq{v_n}_{n\in\Nset}$ is positive and unbounded. (For example, if $G_n$ is a sample mean then $v_n=n$.) For $G_{n,t} \equiv G_{n}\circ\Phi_{n,t}$ and $G_{n,0} = G_n$, we defined in \cite{LaCour_and_Schieve2002} a quantity $\Psi_t$, called the dynamical free energy, where \begin{equation} \label{eqn:DFE} \Psi_{t}(\lambda,\nu) \equiv \lim_{n\to\infty} v_n^{-1}\log\avg{\e^{v_n[\lambda G_{n,0}+\nu G_{n,t}]}}, \end{equation} Here, the values of $G_{n,0}$ and $G_{n,t}$ are considered to be column vectors while the conjugate macrostates, $\lambda$ and $\nu$ are taken as row vectors; the symbol $\avg{\cdot}$ denotes expectation with respect to the \textit{a priori} measure $P_n$. We suppose that $\Psi_{t}$ is everywhere well defined, finite, and differentiable. (The validity of this assumption depends, of course, on the microscopic dynamics, as given by $\Phi_{n,t}$, the macroscopic observables, $G_{n}$, and the scaling parameter $v_n$.) It then follows from the G\"{a}rtner-Ellis theorem \cite{Dembo_and_Zeitouni} that the sequence $\seq{(G_{n,0},G_{n,t})}_{n\in\Nset}$ satisfies a large deviation principle (LDP) whose rate function is given by the Legendre transform of $\Psi_t$. Using a theorem regarding conditional LDPs \cite{LaCour_and_Schieve2002}, it further follows that $G_{n,t}$, when conditioned on a value $a_0$ of $G_{n,0}$, converges in probability as $n\rightarrow\infty$ to the value $\nabla_{\nu}\Psi_t(\lambda_0,0)$, where $\lambda_0$ satisfies $a_0 = \nabla_{\lambda}\Psi_t(\lambda_0,0)$ \cite{LaCour_and_Schieve2002}. The macroscopic map, $\psi_t$, associating an initial macrostate $a_0$ with a final macrostate $\psi_t(a_0)$ may then be defined as \begin{equation} \psi_t(a_0) \equiv \nabla_{\nu}\Psi_t(\lambda_0,0) = \lim_{n\rightarrow\infty} \frac{\avg{G_{n,t}\,\e^{v_n\lambda_0 G_{n}}}}{\avg{\e^{v_n\lambda_0 G_{n}}}}. \end{equation} where $\lambda_0$ is uniquely defined by \begin{equation} a_0 = \nabla\Psi(\lambda_0) = \lim_{n\rightarrow\infty} \frac{\avg{G_{n}\,\e^{v_n\lambda_0 G_{n}}}}{\avg{\e^{v_n\lambda_0 G_{n}}}} \end{equation} and $\Psi(\cdot) \equiv \Psi_t(\cdot,0)$ is the equilibrium free energy. We assume the Jacobian of $\nabla\Psi$ is nowhere zero, so the mapping $\lambda_0 \mapsto a_0$ is invertible. Large deviation theory therefore implies that the most likely future macrostate is given by a canonical expectation. Note that for $a_0$ equal to the equilibrium value $a_*$, where \begin{equation} a_* \equiv \lim_{n\rightarrow\infty}\avg{G_n} = \nabla\Psi(0), \end{equation} the corresponding conjugate macrostate is $\lambda_0 = 0$. \subsection{Linearization} The basic starting assumption is that the macroscopic variable corresponds to the predicted macrostate $\psi_t(a_0)$. Since a one-to-one correspondence between $a_0$ and the conjugate macrostate $\lambda_0$ holds, this macrostate may alternately be written as a function $\alpha$ of $\lambda_0$ and $t$; thus, $\alpha(t,\lambda_0) \equiv \psi_t(a_0)$. Assuming this function to be analytic, a series expansion is then made about equilibrium ($\lambda_0=0$) and the initial time ($t=0$). A linear law is thereby obtained which is valid for short times and near-equilibrium initial conditions. By equating the equilibrium large deviation rate function with the thermodynamic entropy, the flow and force will be identified, thereby establishing the Onsager relation. Throughout the derivation, no \textit{ad hoc} mesoscopic time is introduced. Finally, a expression for the Onsager coefficients is derived in terms of a correlation matrix, which is identified as a Green-Kubo formula, and the reciprocity relations then follow from time reversal invariance. Expanding $\alpha(t,\lambda_0)$ about $(0^{\pm},0)$, where $0^{\pm}$ indicates a right/left-sided time derivative evaluated at zero, we find, to second order, \begin{equation} \begin{split} \alpha(t,\lambda_0) &\approx \left.\alpha\right|_{(0^{\pm},0)} + \left.(t\partial_t + \lambda_0\nabla)\alpha\right|_{(0^{\pm},0)} + \frac{1}{2}\left.(t\partial_t + \lambda_0\nabla)^2\alpha\right|_{(0^{\pm},0)} \\ &= a_* + \left.(\lambda_0\nabla)\alpha\right|_{(0^{\pm},0)} + \left.(t\partial_t\lambda_0\nabla)\alpha\right|_{(0^{\pm},0)} + \frac{1}{2}\left.(\lambda_0\nabla)^2\alpha\right|_{(0^{\pm},0)}. \end{split} \end{equation} The second line in the above display results from the fact that $\partial_t\alpha(0^{\pm},0)$ and $\partial_t^2\alpha(0^{\pm},0)$ are both zero, since $\alpha(t,0) = a_*$. (Recall that $P_n$ is invariant under $\Phi_{n,t}$.) As all terms but the third are independent of $t$ (to second order), the partial time derivative of $\alpha(t,\lambda_0)$ is \begin{equation} \partial_t\alpha(t,\lambda_0) \approx \left.(\lambda_0\,\partial_t\nabla)\alpha\right|_{(0^{\pm},0)}. \end{equation} The left-hand side is identified with the thermodynamic flow, $J(t)$. (This follows from the identification of $\alpha(t,\lambda_0)$ as the macroscopic variable.) To identify $L$ and $X$ on the right hand side, recall that $X = \transpose{\nabla}S$ by definition. If we postulate that $S = -I$, where $I$ is the equilibrium rate function, i.e., the Legendre transform of $\Psi$, then $a_0 = \nabla\Psi(\lambda_0)$ implies that $\lambda_0 = \transpose{\nabla}I(a_0) = -\transpose{\nabla}S(a_0) = -X(a_0)$. Substituting, we find \begin{equation} J(t) \approx \left.(-X(a_0)\,\partial_t\nabla)\alpha\right|_{(0^{\pm},0)}, \label{eqn:flow-force} \end{equation} or, in explicit component form, \begin{equation} J_i(t) \approx \sum_{j} -\left.\frac{\partial^2\alpha_i}{\partial t \partial\lambda_j}\right|_{(0^{\pm},0)} X_j(a_0). \end{equation} Evidently, the Onsager matrix is given by \begin{equation} L^{(\pm)} = \left. \transpose{(-\partial_t\nabla\transpose{\alpha})}\right|_{(0^{\pm},0)}. \label{eqn:Onsager_matrix} \end{equation} For a single (i.e., scalar) observable, a positive $\lambda_0$ corresponds to an initial macrostate above equilibrium, while a negative $\lambda_0$ corresponds to one below equilibrium. Therefore, assuming $L^{(+)}$ is positive, the initial tendency of $\alpha$ is to move towards equilibrium. (For large $n$ this means that the actual observable, $G_{n,t}$, tends towards equilibrium with probability near one.) If $\alpha$ is symmetric in time, due for example to time reversal invariance (see \cite{LaCour_and_Schieve2002}), then the same holds true in the reverse time direction as well. \subsection{Semigroups and Equilibration} The initial tendency towards equilibrium does not necessarily imply a long-time approach to equilibrium unless the family $\set{\psi_t}_{t\in\Rset}$ of macroscopic maps forms a semigroup. As an aside, suppose the family of macroscopic maps does in fact form a semigroup, i.e., that $\psi_{t+s}(a_0) = \psi_s(\psi_t(a_0))$ for $s,t \ge 0$ (or $s,t \le 0$), and suppose further that $\psi_t(a_0)$ is near the equilibrium value, $a_*$. The linear approximation then leads to an exponential time evolution for the macrostate, since \begin{equation} \begin{split} \dot{\psi}_t(a_0)&=\lim_{s\to0}\frac{\psi_{t+s}(a_0)-\psi_t(a_0)}{s}\\ &= \lim_{s\to0}\frac{\psi_{s}(\psi_t(a_0))-\psi_t(a_0)}{s} \approx L \, \nabla S(\psi_t(a_0)) \\ &\approx L\left\{ \nabla S(a_*) + \nabla\transpose{\nabla}S(a_*) \, [\psi_t(a_0)-a_*] \right\} \\ &= L \, \nabla\transpose{\nabla}S(a_*) \, [\psi_t(a_0)-a_*] \end{split} \end{equation} implies \begin{equation} \psi_t(a_0) \approx a_* + (a_0 - a_*) \exp[tL\nabla\transpose{\nabla}S(a_*)]. \end{equation} The essence of the time-averaging approach used by Casimir, Mori, Oono, and others is to suppose that there is a mesoscopic time $\tau$ large enough so that $\psi_{t+\tau}(a_0) \approx \psi_{\tau}(\psi_t(a_0))$ yet small enough so that $\dot{\psi}_t(a_0) \approx [\psi_{t+\tau}(a_0)-\psi_t(a_0)]/\tau$. If this can be shown, then the usual exponential macroscopic law, as derived above, follows. The precise conditions for the validity of this approximation will not be pursued here further, however. \section{Green-Kubo Formula} \label{sec:GKF} Having obtained an explicit expression for the Onsager matrix in terms of $\alpha$, the dynamical free energy may now be used to derive a Green-Kubo formula, giving $L^{(\pm)}$ in terms of the rate of change of the asymptotic covariance between $G_n$ and $G_{n,t}$. We begin by observing that, since $\alpha(t,\lambda) = \nabla_{\nu}\Psi_t(\lambda,0)$, \begin{equation*} \left.(\nabla\transpose{\alpha})\right|_{(t,0)} = \left. \left[ \nabla_{\lambda} \lim_{n\to\infty} \avg{\transpose{G_{n,t}}}_{\lambda} \right] \right|_{\lambda=0}, \end{equation*} where \begin{equation} \avg{G_{n,t}}_{\lambda} \equiv \frac{\avg{G_{n,t}\,\e^{v_n\lambda G_{n}}}}{\avg{\e^{v_n\lambda G_{n}}}}. \end{equation} Convexity of the free energy allows us to take the derivative within the limit, so \begin{equation*} \begin{split} \left.(\nabla\transpose{\alpha})\right|_{(t,0)} &= \lim_{n\to\infty} v_n \!\left.\left[ \avg{G_{n}\transpose{G_{n,t}}}_{\lambda} \!\!- \avg{G_{n}}_{\lambda} \avg{\transpose{G_{n,t}}}_{\lambda} \right]\right|_{\lambda\!=0} \\ &= \lim_{n\to\infty} v_n \left[ \avg{G_{n}\transpose{G_{n,t}}} - \avg{G_{n}} \avg{\transpose{G_{n,t}}} \right]. \end{split} \end{equation*} From the above result and Eq.\ (\ref{eqn:Onsager_matrix}), the Green-Kubo relation now follows: \begin{equation} L^{(\pm)} = - \left.\frac{\d}{\d t} \lim_{n\to\infty} v_n \left[\avg{G_{n,t}^{} \transpose{G_{n}}} - \avg{G_{n,t}^{}}\avg{\transpose{G_{n}}}\right]\right|_{t=0^{\pm}}. \label{eqn:Green-Kubo} \end{equation} This result is equivalent to that of Oono in Eq.~(\ref{eqn:Oono}), assuming $\avg{G_n} = 0$ and $G_{n,t}$ is differentiable almost everywhere. Furthermore, since the \textit{a priori}\ measure is invariant, the Onsager matrix may be written in the following, more explicit form. \begin{equation} L^{(\pm)} = -\lim_{t\to0^{\pm}} \lim_{n\to\infty} \frac{v_n}{t} \left[ \avg{G_{n,t} \transpose{G_{n}}} - \avg{G_{n}} \avg{\transpose{G_{n}}} \right]. \label{eqn:simple_Green-Kubo} \end{equation} \subsection{Time Reversal Invariance} Using Eq.~(\ref{eqn:Green-Kubo}) or (\ref{eqn:simple_Green-Kubo}), the Onsager reciprocal relations can be shown to follow from time reversal invariance and measure preservation. Specifically, if $\Phi_{n,t}$ is time reversal invariant under some map\ $R_{n}$ and both $P_n$ and $G_{n}$ are invariant under $R_{n}$, then $L^{(\pm)}$ is symmetric. To prove this, it suffices to show that $\avg{G_{n,t}^{}\transpose{G_{n}}} = \avg{G_{n}^{}\transpose{G_{n,t}}}$. First note that \begin{equation} \begin{split} \avg{G_{n,t}^{}\transpose{G_{n}}} &= \avg{(G_{n}^{}\circ R_{n}^{}\circ\Phi_{n,t}^{})\,\transpose{G_{n}}} \\ &= \avg{(G_{n}^{}\circ\Phi_{n,-t}^{}\circ R_{n}^{})(\transpose{G_{n}}\circ R_{n}^{})} \\ &= \avg{G_{n,-t}^{}\transpose{G_{n}}}. \end{split} \end{equation} Using this result and the fact that $P_n$ is invariant, we find \begin{equation} \begin{split} \avg{G_{n,-t}^{}\transpose{G_{n}}} &= \avg{(G_{n}^{}\circ\Phi_{n,-t}^{})\,\transpose{G_{n}}} \\ &= \avg{G_{n}^{}\,\transpose{(G_{n}^{}\circ\Phi_{n,t}^{})}} = \avg{G_{n}^{}\transpose{G_{n,t}}}. \end{split} \end{equation} Combining the above two equations then gives the desired result that $L^{(\pm)}$ is symmetric. Note that time reversal invariance is not a necessary condition for the reciprocity relations to hold, as shown in \cite{Gabrielli1996}. Time reversal invariance, as defined above, is known to imply time symmetry for the predicted macroscopic time evolution; i.e., $\alpha(-t,\cdot) = \alpha(t,\cdot)$. It is therefore not surprising that $L^{(+)}$ and $L^{(-)}$ should reflect that symmetry. Indeed, it is clear that $L^{(-)} = -L^{(+)}$ since \begin{equation} \begin{split} L^{(-)} &= -\lim_{t\to0^-} \lim_{n\to\infty} \frac{v_n}{t} \left[ \avg{G_{n,t} \transpose{G_{n}}} - \avg{G_{n}} \avg{\transpose{G_{n}}} \right] \\ &= -\lim_{t\to0^+} \lim_{n\to\infty} \frac{v_n}{-t} \left[ \avg{G_{n,-t} \transpose{G_{n}}} - \avg{G_{n}} \avg{\transpose{G_{n}}} \right] \\ &= \lim_{t\to0^+} \lim_{n\to\infty} \frac{v_n}{t} \left[ \avg{G_{n,t} \transpose{G_{n}}} - \avg{G_{n}} \avg{\transpose{G_{n}}} \right] \\ &= -L^{(+)}. \end{split} \end{equation} This result may be viewed as a statement of the Onsager-Machlup principle. \subsection{Positive Semidefinite Matrix} The forward Onsager matrix, $L^{(+)}$, may be shown to be positive semidefinite under more general conditions, using only the assumption that the \textit{a priori}\ measure is invariant. To see this, observe that for any given row vector $X$ \begin{equation} X L^{(+)} \transpose{X} = -\lim_{t\to0^{\pm}} \lim_{n\to\infty} \frac{v_n}{t} \left[ \avg{XG_{n,t} \transpose{G_{n}}\transpose{X}} - \avg{XG_{n}} \avg{\transpose{G_{n}}\transpose{X}} \right]. \end{equation} Note that $X G_{n,t}^{}$ and $\transpose{G_{n}}\transpose{X} = \transpose{(X G_{n}^{})} = X G_{n}^{}$ are both scalar-valued random variables. The Cauchy-Schwarz inequality therefore implies that \begin{equation} \begin{split} \avg{X G_{n,t}^{} X G_{n}^{}} &\le \avg{(X G_{n,t}^{})^2}^{1/2} \avg{(X G_{n}^{})^2}^{1/2} = \avg{(X G_{n}^{})^2}, \end{split} \end{equation} where the final equality is due to the invariance of the \textit{a priori} measure. Since $\avg{X G_{n,t}^{} \transpose{G_{n}} \transpose{X}} = \avg{X G_{n,t}^{} X G_{n}^{}}$ and $\avg{X G_{n}} \avg{\transpose{G_{n}} \transpose{X}}= \avg{(X G_{n}^{})^2}$, we conclude that $X L^{(+)} \transpose{X} \ge 0$ for any $X$, so $L^{(+)}$ is positive semidefinite. \subsection{Entropy Production Rate} We end this section with a brief discussion of the internal entropy production rate, $\sigma(t,\lambda_0) \equiv \dot{S}(\alpha(t,\lambda_0))$, which is an important measure of irreversibility in nonequilibrium thermodynamics. Its nonnegativity is traditionally derived from an extrapolation of the equilibrium second law to the nonequilibrium regime \cite{deGroot_and_Mazur}; here we derive it directly. From Eqs.~(\ref{eqn:entropy_rate}) and (\ref{eqn:Onsager_relation}), together with Eqs.~(\ref{eqn:flow-force}) and (\ref{eqn:Onsager_matrix}), the initial entropy production rate is \begin{equation} \begin{split} \sigma(0^{\pm},\lambda_0) &\equiv \lim_{t\rightarrow0^{\pm}} [S(\alpha(t,\lambda_0))-S(\alpha(0,\lambda_0))]/t \\ &= X(a_0)J(0^{\pm}) \approx \lambda_0 L^{(\pm)} \transpose{\lambda_0}. \end{split} \end{equation} For a system which is time reversal invariant, $L^{(+)}$, is positive semidefinite and $L^{(-)} = -L^{(+)}$ is negative semidefinite, so the nonequilibrium second law follows: \begin{equation} \sigma(0^+,\lambda_0) \ge 0, \quad \sigma(0^-,\lambda_0) \le 0. \end{equation} Note that the above inequalities are valid only for early times near equilibrium. The first inequality, $\sigma(0^+,\lambda_0) \ge 0$, tells us that the initial tendency of the system is to evolve to a macrostate of higher entropy. Since $S = -I$, this means that the macrostate initially tends to move closer to the minimum of $I$, i.e., the equilibrium value $a_*$. Similarly, the second inequality, $\sigma(0^-,\lambda_0) \le 0$, means that nonequilibrium states tend to arise from states closer to equilibrium in the near past. These interpretations of macroscopic fluctuations are entirely consistent with those of Boltzmann's $H$-theorem as given long ago by P. and T. Ehrenfest \cite{Ehrenfests}, provided $H$ is defined as a function on the phase space rather than as a functional of the probability density. \section{Ideal Gas Example} \label{sec:IGE} As an illustration of the above results, we consider the case of an ideal gas in a rigid box with an imaginary partition in the middle. The particles evolve freely and interact only with the walls of the container via specular reflection. Such a system is represented by a Knundsen gas, for which the mean free path is much larger than the linear dimension of the container, so particle collisions may be ignored \cite{Kogan}. The observable is taken to be simply the fraction of particles on one side of the imaginary partition. The \textit{a priori} spatial distribution of the particles is uniform, whereas that of the velocity is direction independent and given by the symmetric probability density function $\phi_{\sigma}$. In \cite{LaCour_and_Schieve2000} it was shown that the macroscopic map for this observable is \begin{equation} \psi_t(a_0) = a_0 \sum_{n=-\infty}^{\infty}I_n(t) + (1-a_0) \sum_{n=-\infty}^{\infty}\bar{I}_n(t), \end{equation} where \begin{subequations} \begin{equation} I_{n}(t) \equiv 2 \int_{-\infty}^{\infty}\int_{0}^{1/2} \!1_{[0,1/2]}(\xi\!-\!\eta t\!-\!n) \, \phi_{\sigma}(\eta) \, \d\xi \d\eta \end{equation} if $n$ is even and, if $n$ is odd, \begin{equation} I_{n}(t) \equiv 2 \int_{-\infty}^{\infty}\int_{0}^{1/2} \!1_{[0,1/2]}(1\!-\!\xi\!+\!\eta t\!+\!n) \, \phi_{\sigma}(\eta) \, \d\xi \d\eta. \end{equation} \end{subequations} ($1_{[0,1/2]}$ is the indicator function on the interval $[0,1/2]$.) For $\bar{I}_{n}(t)$, similar definitions hold with, however, $1_{[0,1/2]}$ replaced by $1_{[1/2,1]}$. For $t$ positive and small we find that \begin{equation} \psi_t(a_0) = a_0\left[I_0(t)+I_{-1}(t)\right] + (1-a_0)\left[\bar{I}_0(t)+\bar{I}_{-1}(t)\right] + \cdots, \end{equation} which, since $\lambda_0 = \log[a_0/(1-a_0)]$, may be written \begin{equation} \alpha(t,\lambda_0) = \frac{\e^{\lambda_0}\left[I_0(t)+I_{-1}(t)\right] + \left[\bar{I}_0(t)+\bar{I}_{-1}(t)\right]}{1+\e^{\lambda_0}} + \cdots. \end{equation} Expanding $\alpha$ about $(0^{\pm},\lambda_0)$ gives \begin{equation} \alpha(t,\lambda_0) = \frac{\e^{\lambda_0} - 2(\e^{\lambda_0}-1)L\abs{t}}{1+\e^{\lambda_0}} + \cdots, \end{equation} where \begin{equation} L \equiv \int_{0}^{\infty} \eta \, \phi_{\sigma}(\eta) \, \d\eta, \end{equation} so expanding about $\lambda_0 = 0$ finally gives \begin{equation} \alpha(t,\lambda_0) = \frac{1}{2} + \left(\frac{1}{4} - L\abs{t}\right)\lambda_0 + \cdots. \end{equation} By inspection we identify $L$ with the forward Onsager coefficient $L^{(+)}$. Note that, from its definition, $L$ is strictly positive since $\phi_{\sigma}$ was assumed to be symmetric. For an \textit{a priori} Maxwell-Boltzmann velocity distribution, \begin{equation} \phi_{\sigma}(\eta) = (2\pi\sigma^2)^{-1/2}\,\e^{-\eta^2/(2\sigma^2)}, \end{equation} the Onsager coefficient is \begin{equation} L = \frac{\sigma}{\sqrt{2\pi}}=\frac{1}{\ell}\sqrt{\frac{kT}{2\pi m}}, \end{equation} where $T$ is the absolute temperature, $m$ is the particle mass, and $\ell$ is the length of the box. The initial entropy production rate is therefore \begin{equation} \sigma(0^{\pm},\lambda_0) = \pm\frac{1}{\ell}\sqrt{\frac{kT}{2\pi m}} \left[\log\left(\frac{a_0}{1-a_0}\right)\right]^2. \label{eqn:idealgas_entropy_production} \end{equation} Similarly, for an \textit{a priori} velocity distribution which is uniform over the interval $[-\Delta,\Delta]$, we find $L = \Delta/4$. The equilibrium entropy in both cases is \begin{equation*} S(a_0) = - \left[ a_0\log a_0 + (1-a_0)\log(1-a_0) + \log 2 \right], \end{equation*} where $a_0$ is the initial fraction of particles in the interval $[0,\ell/2]$ (i.e., the left side). Note that this is the entropy derived from the large deviation rate function for this particular observable and is not to be confused with the standard Sackur-Tetrode equation of an ideal gas. In Fig.~\ref{fig:entropy} are plotted the equilibrium entropy, $S(a_0)$, and the initial forward entropy production rate, $\sigma(0^+,\lambda_0)$. (The latter has been rescaled to remove the incidental factor of $L$.) These graphs correspond to an ideal gas in a rigid box (or cylinder) with an arbitrary symmetric velocity distribution and independent uniform spatial distribution. The symmetry of the graphs is due to the symmetry of the imaginary box partition. The peak at $a_0=1/2$ in $S$ corresponds to the most likely \textit{a priori} configuration. The entropy rate takes its minimum value of zero at this location, indicating that there is no tendency to deviate from an initial equilibrium macrostate. For $a_0\neq1/2$, the entropy production rate is strictly positive, indicating an initial tendency to approach equilibrium. At the extremal points, $a_0=0\text{ or }1$, the entropy production rate diverges; however, Eq.~(\ref{eqn:idealgas_entropy_production}) is strictly valid only near equilibrium. \begin{figure} \centerline{\scalebox{.5}{\includegraphics{entropy}}} \caption[Equilibrium entropy and entropy rate of an ideal gas]{Plot of the equilibrium entropy, \protect$S(a_0)$, (solid curve) and rescaled initial entropy production rate, \protect$\sigma(0^+,\lambda_0)/L$, (dashed curve) versus the initial fractional occupation, \protect$a_0$, of ideal gas particles occupying one side of a rigid box.} \label{fig:entropy} \end{figure} In Fig.~\ref{fig:maxgas_entropy} we have plotted the dynamic entropy, $S(\alpha(t,\lambda_0))$, and entropy production rate, $\sigma(t,\lambda_0)$, (computed numerically from the former) versus time for an ideal gas with $a_0=1$ and an \textit{a priori} Maxwellian velocity distribution. (The picture is qualitatively the same for other values of $a_0$.) The approach to equilibrium is monotonic, as indicated by the fact that $\sigma(t,\lambda_0) \ge 0$, in accordance with the usual extrapolation of the second law beyond equilibrium. \begin{figure} \centerline{\scalebox{.5}{\includegraphics{maxgas_entropy}}} \caption[Dynamic entropy of an ideal gas with a Maxwellian distribution]{Plot of the dynamic entropy, \protect$S(\alpha(t,\lambda_0))$, (solid curve) and entropy production rate, \protect$\sigma(t,\lambda_0)$, (dashed curve) versus time for an ideal gas with an \textit{a priori} Maxwellian velocity distribution of unit variance. The initial macrostate is \protect$a_0=1 \; (\lambda_0=\infty)$.} \label{fig:maxgas_entropy} \end{figure} The following picture, Fig.~\ref{fig:unifgas_entropy}, shows the same quantities but with an \textit{a priori}\ velocity distribution which is uniform on the interval $[-1,1]$. After an initial monotonic approach to equilibrium, the system exhibits ``anti-thermodynamic'' behavior from $t=1$ to about $t=1.4$. For larger values of $t$ the entropy productions rate continues to oscillate about zero, in step with the oscillations of $\alpha(t,\lambda_0)$ about $1/2$, but converges to zero as $t\to\infty$. The oscillations may be interpreted as an indication of phase mixing, wherein the phase space becomes filibrated due solely to the dispersion of particles traveling at different, yet constant, velocities. An important point is that, despite their anomalous behavior, these graphs are consistent with a positive Onsager coefficient, since the antithermodynamic behavior occurs for $t$ far from zero. A comparison of Figs.~\ref{fig:maxgas_entropy} and~\ref{fig:unifgas_entropy} suggests that a system of nonideal gas particles, e.g.\ rigid spheres, should give rise to a dynamic entropy which is monotonic beyond an initial equilibration phase in which, in accordance with the Boltzmann equation, the particles should attain an approximately Maxwellian velocity distribution. Given the very short time scale over which this local equilibration is expected to occur, it is not surprising that antithermodynamic behavior such as we have described is not normally observed. \begin{figure} \centerline{\scalebox{.5}{\includegraphics{unifgas_entropy}}} \caption[Dynamic entropy of an ideal gas with a uniform distribution]{Plot of the dynamic entropy, \protect$S(\alpha(t,\lambda_0))$, (solid curve) and entropy production rate, \protect$\sigma(t,\lambda_0)$, (dashed curve) versus time for an ideal gas with an \textit{a priori} uniform velocity distribution with support on \protect$[-1,1]$. The initial macrostate is \protect$a_0=1 \; (\lambda_0=\infty)$.} \label{fig:unifgas_entropy} \end{figure} \section{Discussion} \label{sec:Discussion} We have considered the time evolution of a given set of macroscopic variables in a microscopically deterministic, isolated system. At equilibrium, these variables attain their most probable values but over time will exhibit fluctuations away from equilibrium. We have shown that for small fluctuations the most probable evolution of these fluctuations in the near future (and near past) is governed by the Onsager principle, which gives a linear relationship between the flows and thermodynamic forces. The former was interpreted as the time rate of change of the most probable macrostate, while the latter was found to correspond to the Legendre-transform conjugate of the initial, nonequilibrium macrostate. This result was obtained using well-established techniques from large deviation theory and is independent of any \textit{ad hoc} mesoscopic time scale. An explicit expression for the forward and reverse Onsager matrices in terms of the covariance matrix of the macroscopic variables then followed. The forward Onsager matrix was found to be positive semidefinite as a consequence of the assumed invariance of the \textit{a priori} probability measure. The Onsager reciprocity relations were verified in the case of time reversal invarinace and found to imply that the forward and reverse Onsager matrices differ only by a sign. Together, these last two results imply that the second law of thermodynamics, as regards fluctuations, is valid in the aforementioned regime for both the forward and reverse time directions. Thus, given an initial nonequilibrium macrostate, a system's initial tendency is to approach equilibrium (in the forward time direction) or to have arisen from a state closer to equilibrium (in the reverse time direction). This need not imply long-time approach to equilibrium, however, unless, for example, the macroscopic dynamics form a semigroup. At much later or earlier times, antithermodynamic behavior is also possible, as illustrated by the example considered of an ideal gas with a uniform velocity distribution. \section*{Acknowledgments} This work was supported in part by the Engineering Research Program of the Office of Basic Energy Sciences at the U.S.\ Department of Energy, Grant No.\ DE-FG0394ER14465. One of us (B.L.) has also received partial funding from Applied Research Laboratories of the University of Texas at Austin, Independent Research and Development Grant No.\ 926.

795620603132429cc819ec96c3c9151e9602acf2

\section{Introduction} \label{sec:introduction} \begin{figure}[t] \centering \includegraphics[width=0.65\linewidth]{fig1.pdf} \vspace{-5pt} \caption{BAtch Knowledge Ensembling (BAKE) produces soft targets for self-distillation \emph{with a single network} (an encoder and a classifier). For an anchor image $x^\text{anchor}$, the knowledge of the other samples $\{x_1, x_2, x_3, \cdots\}$ in the same batch is weightedly propagated and ensembled to form better soft targets for distillation on-the-fly. Note that $x^\text{anchor}$ and $\{x_1, x_2, x_3, \dots\}$ are fed into the same network. This method enables knowledge ensembling with minimal computational and memory overhead.} \vspace{-10pt} \label{fig:intro} \end{figure} Deep neural networks have achieved impressive success on computer vision tasks, where image classification \cite{he2016deep,Xie2016,howard2017mobilenets,tan2019efficientnet,zhang2020resnest} is considered as one of the most fundamental tasks given the wide range of applications of its learned representations and their transferability to downstream tasks, \textit{e.g.}, detection \cite{fasterrcnn,lin2017feature}, segmentation \cite{zhao2017pyramid,he2017mask}, generation \cite{CycleGAN2017,pix2pix2017,Liu_DivCo,ge2020structured}, retrieval \cite{ge2018fd,ge2020mutual,ge2020self,ge2020selfpaced}, \textit{etc}. There is a tremendous number of methods \cite{bagherinezhad2018label,beyer2020we,xie2020self,guo2020online,yun2020regularizing,yuan2020revisiting,yun2021relabel} working on improving the image classification accuracy, especially on the large-scale ImageNet \cite{deng2009imagenet} dataset. Recent studies \cite{bagherinezhad2018label,beyer2020we,yun2021relabel} have shed light on the limitations of supervised learning of image classification. They have observed that the imperfect learning targets resulted from manually annotated ground-truth labels (one-hot class vectors) turn out to be a key factor that hinders the further improvement of classification accuracy. Thanks to the great success of knowledge distillation \cite{kd}, the soft probability vectors predicted by a teacher network, carrying the learned ``dark knowledge'', can serve as informative training supervisions to enhance a student network. The quality of the teacher's predictions is found critical to the accuracy of the student network. State-of-the-art methods \cite{lan2018one,shen2019meal,tian2019crd,guo2020online,son2021densely} have found that multiple teachers or students could encode complementary knowledge and their ``ensembled'' soft targets are more robust learning objectives. Positive as their results are, they depend on extra networks or branches, undoubtedly increasing the computational and memory cost to a noticeable extent. To produce high-quality soft targets \emph{with minimal cost}, we introduce a self-distillation method with a novel \textbf{BA}tch \textbf{K}nowledge \textbf{E}nsembling (\textbf{BAKE}) scheme, as illustrated in Figure \ref{fig:intro}. Rather than ensembling multiple networks or branches to generate the distillation soft targets, our method only adopts a single network. It achieves knowledge ensembling by on-the-fly aggregating the ``dark knowledge'' from different samples within the same mini-batch, yielding better soft targets. Specifically, given the samples' encoded representations and predictions in a mini-batch, we conduct cross-sample knowledge ensembling under the assumption that visually similar samples with close-by representations should encode consistent class-wise predictions. In practice, for each anchor sample, the other samples' predictions (``dark knowledge'') can be weightedly propagated and ensembled to form a soft target. The knowledge propagation and ensembling are conducted iteratively until convergence, \textit{i.e.}, the soft targets no longer change. In order to perform the proposed batch knowledge ensembling at each training iteration efficiently, we adopt approximate inference to estimate the iterative knowledge ensembling results. After properly ensembling the samples' knowledge in the batch, we are able to create refined soft targets for each sample to improve the training of image classification. We demonstrate the effectiveness of our proposed BAKE via a comprehensive set of evaluations with various architectures and datasets. On ImageNet \cite{deng2009imagenet} classification, BAKE well boosts the top-1 accuracy of ResNet-50 \cite{he2016deep} by significant \textbf{+1.2\%} gains (76.8\%$\to$78.0\%) with a negligible {3.7\%} computational overhead at training. We also evaluate our BAKE on vision transformers, \textit{i.e.}, Swin Transformer \cite{liu2021swin}, and boost the state-of-the-art performance by up to \textbf{+0.7\%} improvements (81.3\%$\to$82.0\%). The network trained by BAKE shows consistent improvements for transfer learning on downstream tasks, including detection \cite{fasterrcnn} and segmentation \cite{he2017mask} on COCO \cite{coco}. BAKE also improves the classification robustness on perturbed datasets \cite{fgsm,hendrycks2019robustness,hendrycks2021nae} at test time. Not only improving the accuracy on ImageNet, BAKE also improves the classification accuracies on CIFAR-100 \cite{cifar}, TinyImageNet, CUB-200-2011 \cite{cub}, MIT67 \cite{mit} and Stanford Dogs \cite{dog}. BAKE substantially outperforms the single-network state-of-the-arts \cite{furlanello2018born,yun2020regularizing,yuan2020revisiting,zhang2019your,xu2019data} on all the benchmarks. The contributions of our method are three-fold. (1) We for the first time introduce to produce ensembled soft targets for self-distillation without using multiple networks or additional network branches. (2) We propose a novel batch knowledge ensembling mechanism to online refine the distillation targets with the cross-sample knowledge, \textit{i.e.}, weightedly aggregating the knowledge from other samples in the same batch. (3) Our method is simple yet consistently effective on improving classification performances of various networks and datasets with minimal computational overhead and zero additional parameters. \section{Related Works} \paragraph{Knowledge distillation.} Knowledge distillation \cite{kd} aims to transfer the ``dark'' knowledge learned from a high-capacity teacher network to a student network via soft labels. The soft labels can be the class probabilities \cite{furlanello2018born,guo2020online,yun2020regularizing} or the feature representations \cite{tung2019similarity,park2019relational} output by the teacher, containing more complete structured information than the one-hot ground-truth labels. The distillation process can be formed by a ``teacher-student'' framework \cite{kd,shen2019meal}, a ``peer-teaching'' framework \cite{lan2018one,zhang2018deep,guo2020online}, or a self-distillation framework \cite{furlanello2018born,zhang2019your,xu2019data,yun2020regularizing,kim2021self}. Our BAKE is mostly related to the self-distillation methods, \textit{i.e.}, teaching a single network using its own knowledge. However, most of them \cite{furlanello2018born,zhang2019your,kim2021self} only considered the knowledge of individual instances, resulting in sub-optimal learning targets. Recent works introduced to preserve the predictive consistency between intra-image (original \textit{v.s.} perturbed) \cite{xu2019data} or intra-class (images out of the same class) \cite{yun2020regularizing} samples. However, they only focused on pairwise images, carrying limited information compared to the ensembled batch knowledge of BAKE. More importantly, they simply defined the positive pairs using constant instance or class IDs, which may incur false supervisions as their visual features might be actually dissimilar, especially after the random crop augmentation \cite{szegedy2015going}. \begin{table}[t] \footnotesize \begin{center} \begin{tabular}{lcc} \toprule & \makecell[c]{No extra \\parameters} & \makecell[c]{Ensembled \\knowledge} \\ \midrule Self-distillation \cite{furlanello2018born,zhang2019your,xu2019data,yun2020regularizing,kim2021self} & \textcolor{ForestGreen}{\ding{52}} & \textcolor{red}{\ding{56}} \\ Ensemble distillation \cite{lan2018one,shen2019meal,tian2019crd,guo2020online,son2021densely} & \textcolor{red}{\ding{56}} & \textcolor{ForestGreen}{\ding{52}} \\ Label refinery \cite{bagherinezhad2018label,beyer2020we,yun2021relabel} & \textcolor{red}{\ding{56}} & \textcolor{red}{\ding{56}} \\ Our \textbf{BAKE} & \textcolor{ForestGreen}{\ding{52}} & \textcolor{ForestGreen}{\ding{52}} \\ \toprule \end{tabular} \end{center} \vspace{-15pt} \caption{Key differences between our method and related works.} \vspace{-10pt} \label{tab:diff} \end{table} \vspace{-10pt} \paragraph{Knowledge ensembling.} It is well-known that an ensemble of multiple networks generally yields better predictions than a single network in the ensemble. The ensembling technologies aim to generate robust supervision signals via aggregating models \cite{meanteacher,he2020momentum,byol} or predictions \cite{temensemble,lan2018one,shen2019meal,tian2019crd,guo2020online,son2021densely}. Several attempts leveraged the spirit of knowledge ensembling in distillation tasks, dubbed ``ensemble distillation'' methods. For example, CRD \cite{tian2019crd} and MEAL \cite{shen2019meal} proposed to enhance the soft targets by ensembling the knowledge of multiple pre-trained teacher networks. KDCL \cite{guo2020online} introduced to aggregate the information from multiple independent students, which are collaboratively training. The idea of knowledge ensembling can be not only applied in supervised tasks, but also employed in semi-supervised \cite{temensemble,meanteacher} and self-supervised \cite{he2020momentum,byol} learning tasks. Note that the way of ``knowledge ensembling'' is not limited to multi-model ensembling, \textit{e.g.}, in semi/self-supervised learning tasks, \cite{meanteacher,he2020momentum,byol} use temporal ensembling with momentum updates to integrate the knowledge of a student model over iterations. Despite a big success, we remark that existing knowledge ensembling techniques all require additional networks or branches, which may be inapplicable in resource-limited environments. \vspace{-10pt} \paragraph{Label refinery for ImageNet classification.} ImageNet \cite{deng2009imagenet} is a widely-acknowledged dataset in computer vision. Although it could well benchmark the performance of image classification methods, some studies \cite{bagherinezhad2018label,beyer2020we,shankar2020evaluating,yun2021relabel} have observed that the manually annotated labels for ImageNet are incomplete. Specifically, ImageNet was annotated under a single-labeling policy, \textit{i.e.}, one label per image, however, there are generally multiple objects in a single image. To overcome the problem, state-of-the-art label refinery methods \cite{bagherinezhad2018label,yun2021relabel} produced multi-labels by an auxiliary annotator. For instance, \cite{bagherinezhad2018label} introduced an iterative training scheme, \textit{i.e.}, the trained network acted as the annotator for the next generation training. \cite{yun2021relabel} relabeled the dataset with a super-strong classifier, which was pre-trained on super-scale datasets. Although they cleaned up the noisy labels to some extent, they heavily depended on the capacity of external networks and required much resources to train a strong annotator, which was inflexible. The key advantages of our BAKE against existing self-distillation, ensemble distillation and label refinery methods are summarized in Table \ref{tab:diff}. \begin{figure*}[t] \centering \includegraphics[width=0.9\linewidth]{fig2.pdf} \vspace{-5pt} \caption{Conceptual comparison of three knowledge ensembling mechanisms. (a) {\bf Multi-teacher ensembling} \cite{shen2019meal,tian2019crd}: the soft targets are produced by ensembling the knowledge from multiple pre-trained high-capacity teachers. (b) {\bf Multi-student ensembling} \cite{guo2020online}: the soft targets are refined by ensembling the knowledge from multiple peer-teaching students. (c) {\bf The proposed BAKE}: generating robust soft targets by ensembling the knowledge of the other samples (\textit{i.e.}, $\{x_1,\cdots,x_N\}$) in the same mini-batch. BAKE enables knowledge ensembling \textit{with a single network}, saving the memory and time costs to a large extent.} \vspace{-5pt} \label{fig:compare} \end{figure*} \section{Method} \subsection{Revisit of Knowledge Distillation} \vspace{-2pt} Knowledge distillation \cite{kd} can be considered as regularizing the training of the student network using soft targets that carry the ``dark knowledge'' of the teacher network. The student network generally consists of a backbone encoder $F$ and a classifier $C$ to perform classification. For each training sample $x$, its logit vector is encoded as ${\bm{z}}=C(F(x))$. The predictive probability vector ${\bm{p}}^\tau$ can be obtained via a softmax function on the logits, \ie, the probability of class $k$ can be formulated as \begin{align} {\bm{p}}^\tau(k)=\frac{\exp({\bm{z}}_k/\tau)}{\sum_{i=1}^K \exp({\bm{z}}_i/\tau)}, \end{align} where $\tau$ is a temperature hyper-parameter, and $K$ is the number of total classes. Let $y\in\{1,\dots,K\}$ denotes the ground truth label and ${\bm{q}}^\tau$ is the soft target produced by the teacher network. The cross-entropy loss and the KL divergence between the predictions and soft targets are minimized jointly to train the student via \begin{align} \label{eq:self-distill} \mathcal{L}_x=-\log {\bm{p}}(y)+\lambda\cdot\tau^2\cdot D_\text{KL}({\bm{q}}^\tau\|{\bm{p}}^\tau), \end{align} where ${\bm{p}}(y)$ denotes the probability normalized without a temperature, and $\lambda$ weights the two terms. Recent works \cite{lan2018one,shen2019meal,tian2019crd,guo2020online} found that ensembling diverse ``dark knowledge'' from multiple teachers or students can form better soft targets, leading to better final performance (see Figure \ref{fig:compare} (a)\&(b) for details). However, this strategy would increase much more computational and memory overhead to enable multiple networks or branches training. To tackle the challenge, we introduce batch knowledge ensembling in a single network via self-distillation, as illustrated in Figure \ref{fig:compare} (c). \subsection{Batch Knowledge Ensembling} \vspace{-2pt} Rather than utilizing multiple pre-trained high-capacity teachers \cite{shen2019meal,tian2019crd} or collaboratively training students \cite{guo2020online}, we propose a novel self-knowledge ensembling solution by exploring how to ensemble knowledge of different samples in the same mini-batch with a single network, \ie, the student network itself. Intuitively, samples with high visual similarities are expected to have more consistent predictions on their predicted class probabilities, regardless of their ground-truth labels. In our solution, similar samples' knowledge is systematically aggregated and ensembled to provide better soft targets. \vspace{-10pt} \paragraph{Batch knowledge propagation and ensembling.} We propose to propagate and ensemble knowledge among samples on-the-fly in terms of their feature similarities. Given a mini-batch of $N$ samples and a network $C\circ F$ under training, we first estimate the samples' pairwise similarities by the dot product of their encoded representations with the current network. Such similarities can be stored in an affinity matrix $A\in\mathbb{R}^{N\times N}$ as \begin{align} A(i,j)=\sigma(F(x_i))^\top \sigma(F(x_j)), \end{align} where $\sigma({\mathbf{f}})={\mathbf{f}}/\|{\mathbf{f}}\|_2$ denotes the $\ell_2$-norm and ${i,j}$ are the indices of samples in a batch. To avoid the self-knowledge reinforcement, we discard the diagonal entries from $A$ by $A=A\odot(1-I)$, where $I$ is an identity matrix and $\odot$ denotes the element-wise multiplication. Subsequently, we normalize each row of the affinity matrix $A$ so that $\sum_{j=1}^N \hat{A}(i,j)=1$ for all $i$, while keeping the diagonal all zeros, \textit{i.e.}, $\hat{A}(i,i)=0$. The normalization can be formulated as a softmax function over each row of the matrix $A$ \begin{align} \hat{A}(i,j) = \frac{\exp(A(i,j))}{\sum_{j\ne i} \exp(A(i,j))},~~\forall i\in\{1,\dots,N\}. \end{align} We denote the predicted probabilities of samples within a batch as $P^\tau=[{\bm{p}}_1^\tau,\dots,{\bm{p}}_N^\tau]^\top \in \mathbb{R}^{N\times K}$, which satisfy $\sum_{k=1}^K {P}^\tau(i,k)=1, \forall i$. For the $i$-th sample in the mini-batch, we would like to weightedly propagate and ensemble the other samples' predictions to create a better soft target for it based on the inter-sample affinities, which can be formulated as \begin{align} \hat{{\bm{p}}}_i^\tau=\sum_{j\ne i}\hat{A}(i,j){\bm{p}}^\tau_j=\hat{A}(i)P^\tau, \end{align} where $\hat{{\bm{p}}}_i^\tau$ is the propagated probability vector for the $i$-th sample and can serve as the refined soft targets. Intuitively, if the $i$-th sample and the $j$-th sample are similar with a high affinity $\hat{A}(i,j)$, the prediction ${\bm{p}}^\tau_j$ would have a larger weight to be propagated to $\hat{{\bm{p}}}_i^\tau$. Propagating the predictions between all the samples in a mini-batch in parallel can be formulated as $\hat{P}^\tau=\hat{A} P^\tau$. To avoid propagating and ensembling noisy predictions too much, we produce the soft learning targets $Q^\tau$ as a weighted sum of the initial probability matrix $P^\tau$ and the propagated one $\hat{A} P^\tau$, \begin{align} \label{eq:once} Q^\tau = \omega \hat{A} P^\tau + (1-\omega)P^\tau, \end{align} where $\omega\in[0,1]$ is the weighting factor and $Q^\tau=[{\bm{q}}_1^\tau,\dots,{\bm{q}}_N^\tau]^\top\in\mathbb{R}^{N\times K}$. With the above formulations, the knowledge of the samples within the same batch can be propagated to each other and ensembled for one iteration. \SetAlCapHSkip{0em} \setlength{\algomargin}{0pt} \begin{algorithm}[t!] \scriptsize \ttfamily \caption{PyTorch-style pseudocode for BAKE.} \label{alg} \algorithmfootnote{\texttt{mm}: matrix multiplication; \texttt{eye}: identity matrix; \texttt{inv}: inverse matrix.} \Comment{w: ensembling weight} \Comment{t: temperature} \Comment{r: loss weight} \fontseries{m}\selectfont for (x, gt\_labels) in loader:\\ ~~~~\Comment{features: N$\times$D, logits: N$\times$K} ~~~~f, logits = net.forward(x)\\ ~~~~\Comment{classification loss with ground-truth labels} ~~~~loss = CrossEntropyLoss(logits, gt\_labels)\\ ~\\ ~~~~\Comment{produce soft targets} ~~~~f = normalize(f)\\ ~~~~A = softmax(mm(f, f.t())-eye(N)*1e9) \Comment{row-wise normalization of affinity matrix with zero diagonal} ~~~~soft\_targets = mm((1-w)$\cdot$inv(eye(N)-w$\cdot$A), softmax(logits/t)) \Comment{approximate inference for propagation and ensembling} ~~~~soft\_targets = soft\_targets.detach() \Comment{no gradient} ~~~~\Comment{distillation loss with soft targets} ~~~~loss += KLDivLoss(log\_softmax(logits/t), soft\_targets)*t$^\texttt{2}$*r\\ ~~~~\Comment{SGD update} ~~~~loss.backward()\\ ~~~~update(net.params)\\ \end{algorithm} \vspace{-10pt} \paragraph{Approximate inference.} The knowledge propagation and ensembling can be conducted for multiple times until convergence for fully fusing their knowledge \begin{align} Q^\tau_{(t)} &= \omega \hat{A} Q^\tau_{(t-1)} + (1-\omega)P^\tau \nonumber\\ &= (\omega \hat{A})^t P^\tau + (1-\omega) \sum_{i=0}^{t-1} (\omega \hat{A})^i P^\tau, \end{align} where $t$ denotes the $t$-th propagation and ensembling iteration. When the number of iterations approaches infinite, we have $\lim_{t\to \infty}(\omega \hat{A})^t=0$, given that $\omega\in[0,1]$. Also, since \begin{align} \lim_{t\to \infty}\sum_{i=0}^{t-1} (\omega\hat{A})^i=(I-\omega\hat{A})^{-1}, \end{align} where $I$ is an identity matrix. We can obtain an approximate inference formulation for the knowledge propagation and ensembling as \begin{align} \label{eq:infty} Q^\tau_{(\infty)}=(1-\omega)(I-\omega\hat{A})^{-1}P^\tau. \end{align} Note that $Q^\tau_{(\infty)}$ naturally satisfies $\sum_{k=1}^K Q^\tau_{(\infty)}(i,k)=1$ for all $i$ without requiring extra normalization, which forms valid soft targets for training. The gradient is not back-propagated through the soft targets $Q^\tau_{(\infty)}$ for stable training. For each training sample, we estimate its soft targets ${\bm{q}}^\tau_{(\infty)} \in Q^\tau_{(\infty)}$ by ensembling the knowledge from other samples in the same batch with the approximate inference formula (Eq. (\ref{eq:infty})). The produced ${\bm{q}}^\tau_{(\infty)}$ is then used as a refined learning target to supervise the self-distillation procedure with Eq. (\ref{eq:self-distill}). The overall training procedure is detailed in Algorithm \ref{alg}. \subsection{Discussion} \label{sec:discuss} \paragraph{Time and memory consumption.} BAKE generates refined soft targets on-the-fly with \textit{minimal} computational and memory overhead compared to existing ensemble distillation methods. BAKE does not need additional network parameters on top of ordinary backbones (\eg, \cite{lan2018one}), and does not require auxiliary networks (\eg, multiple student networks for peer teaching \cite{guo2020online}). Only very few additional GPU memory (+2.2\% for ResNet-50) is required for prediction propagation and ensembling in addition to the conventional classification loss. Furthermore, BAKE does not require pre-trained teachers or annotators (\eg, \cite{tian2019crd,shen2019meal}). Only little additional training time (+3.7\% for ResNet-50) is required for ensembling the knowledge within each batch and optimizing the network via the self-distillation loss. \vspace{-10pt} \paragraph{Data sampling.} BAKE propagates and ensembles the knowledge based on inter-sample affinities. In the experiments, we found that BAKE does not work if the samples in a mini-batch are totally dissimilar to each other, \ie, showing uniformly low affinities. To ensure that similar pairs can always be found in the mini-batch, we introduce a per-class data sampler on top of the common random sampling mechanism. Specifically, for an anchor image in a randomly sampled mini-batch, we randomly select another $M$ images out of the same class. Given the initial batch size of $\hat{N}$ and $M$ intra-class images for each anchor, we would have a total batch size of $N=\hat{N}\times(M+1)$ for training. For example, we use $\hat{N}=256$ and $M=1$ for our experiments on ImageNet, yielding a batch size of $N=512$. \section{Experiments} \subsection{Experimental Details} \paragraph{Datasets.} We study the effectiveness of BAKE mainly on the large-scale ImageNet-1K \cite{deng2009imagenet} (ILSVRC2012), which is considered as one of the most important benchmarks in learning visual representations. We also evaluate BAKE on the other two conventional image classification datasets, CIFAR-100 \cite{cifar} and TinyImageNet, and three fine-grained image classification datasets, CUB-200-2011 \cite{cub}, Stanford Dogs \cite{dog} and MIT67 \cite{mit}. Top-1 and top-5 classification accuracies are calculated for evaluation. \vspace{-10pt} \paragraph{Network architectures.} To demonstrate that our BAKE can consistently improve various architectures on multiple datasets, we study the family of ResNets, including ResNet \cite{he2016deep}, ResNeSt \cite{zhang2020resnest} and ResNeXt \cite{Xie2016}, the lightweight networks, including MobileNet-V2 \cite{howard2017mobilenets} and EfficientNet-B0 \cite{tan2019efficientnet}, and the vision transformers, including Swin-T/S/B \cite{liu2021swin}, on the ImageNet benchmark. We also evaluate ResNet \cite{he2016deep} and DenseNet \cite{huang2017densely} on the other relatively smaller scale datasets. Note that we use the pre-activation blocks \cite{he2016identity} for CIFAR-100 and TinyImageNet datasets following \cite{yun2020regularizing}. \subsection{ImageNet Classification} \paragraph{Training details.} There are three hyper-parameters required by the training of BAKE. In the loss function Eq. (\ref{eq:self-distill}), we set the distillation loss weight $\lambda$ as $1.0$ and the temperature $\tau$ as $4.0$. In the approximate inference function Eq. (\ref{eq:infty}) of knowledge ensembling, we set the ensembling weight $\omega$ as $0.5$. BAKE is not sensitive to these hyper-parameters, which will be discussed next. $M$ is set to $1$ for the per-class data sampling (see Section \ref{sec:discuss}). If not specified, all the experiments on CNNs and Transformers are trained on 8 GPUs for 100 and 300 epochs, respectively. More details can be found at Appendix \ref{app:training_details}. \vspace{-10pt} \paragraph{Improvements on various architectures.} As illustrated in Table \ref{tab:arch}\&\ref{tab:vit}, we verify the effectiveness of BAKE on multiple network architectures. We not only consider the widely-used architectures (\textit{e.g.}, ResNet-50), but also evaluate both the deeper/wider architectures (\textit{e.g.}, ResNet-152, ResNeXt-152) and the lighter architectures (\textit{e.g.}, MobileNet). We also evaluate on the state-of-the-art architecture, namely Swin Transformer \cite{liu2021swin}. BAKE consistently improves the ``vanilla'' setting (training with a cross-entropy loss) by significant margins. Most importantly, BAKE boosts the performance with only negligible computational overhead, \textit{i.e.}, for each training iteration of Swin-T, BAKE takes only extra $+1.5\%$ more time compared to the plain classification network. \begin{table}[t] \footnotesize \begin{center} \begin{tabular}{lccc} \toprule \multicolumn{1}{c}{\multirow{2}{*}{Architecture}} & \multicolumn{2}{c}{Method} & \multicolumn{1}{c}{\multirow{2}{*}{GPU Time}} \\ \multicolumn{1}{c}{} & Vanilla & Our \textbf{BAKE} & \multicolumn{1}{c}{} \\ \midrule ResNet-50 & 76.8 & 78.0 (+1.2) & +3.7\% \\ ResNet-101 & 78.6 & 79.3 (+0.7) & +2.1\% \\ ResNet-152 & 79.1 & 79.6 (+0.5) & +1.1\% \\ \midrule ResNeSt-50 & 78.4 & 79.4 (+1.0) & +2.8\% \\ ResNeSt-101 & 79.6 & 80.4 (+0.8) & +1.3\% \\ \midrule ResNeXt-101 (32x4d) & 78.7 & 79.3 (+0.6) & +1.9\% \\ ResNeXt-152 (32x4d) & 79.3 & 79.7 (+0.4) & +1.7\% \\ \midrule MobileNet-V2 & 71.3 & 72.0 (+0.7) & +1.5\% \\ EfficientNet-B0 & 75.1 & 76.2 (+1.1) & +6.8\% \\ \toprule \end{tabular} \end{center} \vspace{-15pt} \caption{BAKE improves various architectures with minimal computational overhead. We report the top-1 accuracy (\%) on ImageNet. ``Vanilla'' indicates training with a conventional cross-entropy loss. The time consumption is counted on 8 Titan X GPUs.} \vspace{-5pt} \label{tab:arch} \end{table} \begin{table}[t] \footnotesize \begin{center} \begin{tabular}{lcccc} \toprule \multicolumn{1}{c}{\multirow{2}{*}{Architecture}} & \multicolumn{1}{c}{\multirow{2}{*}{Input Size}} & \multicolumn{2}{c}{Method} & \multicolumn{1}{c}{\multirow{2}{*}{GPU Time}} \\ \multicolumn{1}{c}{} & \multicolumn{1}{c}{} & Vanilla & Our \textbf{BAKE} & \multicolumn{1}{c}{} \\ \midrule Swin-T & 224$\times$224 & 81.3 & 82.0 (+0.7) & +1.5\% \\ Swin-S & 224$\times$224 & 83.0 & 83.3 (+0.3) & +1.0\% \\ Swin-B & 224$\times$224 & 83.5 & 83.9 (+0.4) & +0.5\% \\ Swin-B & 384$\times$384 & 84.5 & 84.9 (+0.4) & +0.2\% \\ \toprule \end{tabular} \end{center} \vspace{-15pt} \caption{BAKE improves vision transformers with various scales in terms of the top-1 accuracy (\%) on ImageNet. The time consumption is counted on 8 V100 GPUs.} \vspace{-5pt} \label{tab:vit} \end{table} \vspace{-10pt} \paragraph{Comparison with label regularization methods.} As mentioned by \cite{yuan2020revisiting}, label smoothing regularization is an ad hoc distillation with manually designed soft targets. To indicate that the ensembled soft targets by BAKE are better roles than the ad hoc regularizations, we evaluate both conventional label smoothing \cite{szegedy2016rethinking} regularization and the newly proposed teacher-free $\text{Tf-KD}_{reg}$ \cite{yuan2020revisiting} regularization based on the ResNet-50 classification backbone network, respectively. As shown in Table \ref{tab:self}, the above mentioned label regularization methods can improve the ``vanilla'' network with satisfactory margins, but still show inferior performance than our proposed BAKE. For instance, the state-of-the-art $\text{Tf-KD}_{reg}$ improves the ``vanilla'' by $0.7\%$ in terms of top-1 accuracy but is still $-0.5\%$ lower than our BAKE. \begin{table}[t] \footnotesize \begin{center} \begin{tabular}{lccc} \toprule \multicolumn{2}{c}{\multirow{2}{*}{Method}} & \multicolumn{2}{c}{ImageNet} \\ \multicolumn{2}{c}{} & top-1 acc. & top-5 acc. \\ \midrule \multicolumn{2}{c}{Vanilla ResNet-50} & 76.8 & 93.4 \\ \midrule Label smoothing \cite{szegedy2016rethinking} & CVPR'16 & 77.2 & 93.7 \\ $\text{Tf-KD}_{reg}$ \cite{yuan2020revisiting} & CVPR'20 & 77.5 & 93.7 \\ BAN \cite{furlanello2018born}* & ICML'18 & 77.4 & 93.5 \\ CS-KD \cite{yun2020regularizing} & CVPR'20 & 77.0 & 93.4 \\ $\text{Tf-KD}_{self}$ \cite{yuan2020revisiting} & CVPR'20 & 77.5 & 93.7 \\ PS-KD \cite{kim2021self} & ICCV'21 & 77.2 & 93.5 \\ \midrule \multicolumn{2}{c}{Our \textbf{BAKE}} & $\bm{78.0}$ & $\bm{94.0}$ \\ \toprule \end{tabular} \end{center} \vspace{-15pt} \caption{Comparison with state-of-the-art label regularization and self-distillation methods, both of which are based on single networks. We report the results of ResNet-50 on the ImageNet. All the methods are reproduced on our implementation with identical training settings as BAKE for fair comparisons and their performances surpass the reported results in their original papers. (*) Note that the original BAN \cite{furlanello2018born} performs model ensembling during inference. However, it is unfair for other compared methods which only use a single model for testing. So we discard the test-time model ensembling here.} \vspace{-5pt} \label{tab:self} \end{table} \vspace{-10pt} \paragraph{Comparison with self-distillation methods.} Our BAKE can be categorized into self-distillation methods, \ie, regularizing the network predictions using its own knowledge. State-of-the-art self-distillation methods only considered the knowledge of individual samples \cite{furlanello2018born,yuan2020revisiting,kim2021self} or pairwise samples \cite{yun2020regularizing}, producing inferior soft targets than our BAKE. To verify it, we reproduce these methods on top of our implementation and achieve better performances than the performances reported in their original papers (see Table \ref{tab:self}). There remains an obvious performance gap between their methods and our BAKE, \eg, BAKE surpasses $\text{Tf-KD}_{self}$ \cite{yuan2020revisiting} by $0.5\%$ and PS-KD \cite{kim2021self} by $0.8\%$ in terms of top-1 accuracy. \vspace{-10pt} \paragraph{Comparison with ensemble distillation methods.} Our BAKE is designed following the spirit of knowledge ensembling, which aims to refine the soft targets by aggregating diverse and complementary knowledge. State-of-the-art ensemble distillation methods achieved knowledge ensembling by leveraging multiple teachers \cite{shen2019meal,tian2019crd} or multiple students \cite{guo2020online}, which are illustrated in Figure \ref{fig:compare}. We argue that our BAKE could not only save the memory and time consumption by enabling knowledge ensembling in a single network, but also generate equally robust soft targets by aggregating knowledge from a set of samples in the same mini-batch. To demonstrate it, we compare state-of-the-art MEAL \cite{shen2019meal} and KDCL \cite{guo2020online} in Table \ref{tab:ensemble}. We observe that our BAKE achieves similar results by using a single network for much fewer training epochs. BAKE even beats KDCL with half of the training epochs, where KDCL requires an extra ResNet-18 for ensembling. \begin{table}[t] \footnotesize \begin{center} \begin{tabular}{lccc} \toprule Method & Epochs & Top-1 acc. & Extra Params \\ \midrule MEAL \cite{shen2019meal} & 180 & 78.2 & ResNet-101 \& 152 \\ KDCL \cite{guo2020online} & 200 & 77.8 & ResNet-18 \\ Our \textbf{BAKE} & 100 & 78.0 & {None} \\ \toprule \end{tabular} \end{center} \vspace{-15pt} \caption{Comparison with state-of-the-art ensemble distillation methods. We report the results of ResNet-50 on the ImageNet. The results of MEAL and KDCL are from the original papers. CRD \cite{tian2019crd} and DGKD \cite{son2021densely} did not report on ResNet-50, thus is not compared here. CRD, DGKD, MEAL and KDCL require extra networks, while BAKE does not.} \vspace{-5pt} \label{tab:ensemble} \end{table} \vspace{-10pt} \paragraph{Comparison with label refinery methods.} Recent studies on improving ImageNet classification found that the single labels are noisy as there might exist multiple objects per image. State-of-the-art ReLabel \cite{yun2021relabel} method introduced to use a super-strong EfficientNet-L2 for re-labelling the ImageNet with multiple labels in the image spatial plane. However, it requires super-scale datasets ($\sim$1K times larger than the original ImageNet dataset) and much more time and memory consumption to pre-train an annotator network first, which is inflexible and even inapplicable if we do not have such resources. Despite that direct comparison with ReLabel is actually \textit{unfair} due to its much larger training set and much deeper annotator network, we are glad to find that our BAKE can surpass the ReLabel method when training for 100 epochs by using only ImageNet dataset with the ResNet-50 backbone. As the original ReLabel was trained for 300 epochs, we reproduce it with their official code to fairly compare with BAKE for 100 epochs. As shown in Table \ref{tab:relabel}, ReLabel achieves $77.3\%$ top-1 accuracy when trained for 100 epochs, showing much lower performance than our BAKE's $78.0\%$. BAKE is also well compatible with external training tricks, \textit{e.g.}, CutMix \cite{yun2019cutmix}. BAKE consistently improves the baseline performance by noticeable $+1.2\%$ ($77.4\% \to 78.6\%$) when trained for 100 epochs with CutMix. ``BAKE + CutMix'' also consistently surpasses ``ReLabel + CutMix'' by a $+0.2\%$ gain for 100 training epochs. When trained for longer epochs (\textit{i.e.} 300 epochs), ReLabel achieves slightly better performance due to its extra knowledge from super-scale training sets and extra annotator networks. Note that our BAKE can obtain performance gains in a much more efficient and lightweight manner. \begin{table}[t] \footnotesize \begin{center} \begin{tabular}{lccc} \toprule Method & Epochs & Top-1 acc. & Extra Params \\ \midrule Vanilla & 100 & 76.8 & None \\ ReLabel \cite{yun2021relabel} & 100 & 77.3 & EffNet-L2 \\ Our \textbf{BAKE} & 100 & $\bm{78.0}$ & None \\ \midrule Vanilla + CutMix \cite{yun2019cutmix} & 100 & 77.4 & None \\ ReLabel \cite{yun2021relabel} + CutMix \cite{yun2019cutmix} & 100 & 78.4 & EffNet-L2 \\ Our \textbf{BAKE} + CutMix \cite{yun2019cutmix} & 100 & $\bm{78.6}$ & None \\ \midrule ReLabel \cite{yun2021relabel} + CutMix \cite{yun2019cutmix} & 300 & $\bm{80.2}$ & EffNet-L2 \\ Our \textbf{BAKE} + CutMix \cite{yun2019cutmix} & 300 & 79.4 & None \\ \toprule \end{tabular} \end{center} \vspace{-15pt} \caption{Comparison with state-of-the-art label refinery method on ImageNet. We use ResNet-50 as the backbone. The results of ReLabel for 100 epochs are reproduced with the official code. ReLabel \cite{yun2021relabel} requires a strong annotator network (EfficientNet-L2 \cite{tan2019efficientnet}) pre-trained on super-scale datasets (JFT-300M \cite{sun2017revisiting} or Instagram-1B \cite{mahajan2018exploring}) for re-labelling, while BAKE does not. } \vspace{-5pt} \label{tab:relabel} \end{table} \vspace{-10pt} \paragraph{Transfer learning.} Apart from performing the classification task, the models trained on ImageNet can also be well transferred to the downstream tasks by fine-tuning. To show that the ImageNet pre-trained model by BAKE could achieve better transfer learning results, we evaluate two important downstream tasks on the COCO \cite{coco} dataset. As demonstrated in Table \ref{tab:transfer}, we use Faster-RCNN \cite{fasterrcnn} and Mask-RCNN \cite{he2017mask} with feature pyramid network \cite{lin2017feature} as base models for object detection and instance segmentation, respectively. The baseline results are achieved by fine-tuning the model pre-trained with the ``vanilla'' backbone. We observe that the model pre-trained by BAKE could consistently improve the baseline results by $0.5\% \sim 0.8\%$ in terms of mAP. \begin{table}[t] \footnotesize \begin{center} \begin{tabular}{lccc} \toprule \multicolumn{1}{c}{} & Faster-RCNN & \multicolumn{2}{c}{Mask-RCNN} \\ \multicolumn{1}{c}{} & bbox mAP & bbox mAP & mask mAP \\ \midrule ResNet-50 & 37.7 & 38.5 & 35.0 \\ ~ + Our \textbf{BAKE} & 38.3 (+0.6) & 39.2 (+0.7) & 35.5 (+0.5) \\ \midrule Swin-T ($\times 3$ scheduler) & 45.1 & 46.0 & 41.6 \\ ~ + Our \textbf{BAKE} & 45.9 (+0.8) &46.5 (+0.5) & 42.0 (+0.4) \\ \midrule Swin-S ($\times 3$ scheduler) & 48.1 & 48.5 & 43.3 \\ ~ + Our \textbf{BAKE} & 48.7 (+0.6) & 48.9 (+0.4) &43.5 (+0.2) \\ \toprule \end{tabular} \end{center} \vspace{-15pt} \caption{Transfer learning performances for object detection and instance segmentation on the COCO dataset \cite{coco}.} \vspace{-5pt} \label{tab:transfer} \end{table} \vspace{-10pt} \paragraph{Robustness testing.} Our BAKE does not only improve image classification, but also improves the classification robustness on much harder test sets with either common perturbations or adversarial perturbations. ImageNet-A \cite{hendrycks2021nae} contains difficult testing images sampled from the failure cases of modern classifiers. ImageNet-C \cite{hendrycks2019robustness} consists of 19 different corruptions and perturbations, \eg, blurring, fogging. AutoAttack \cite{croce2020reliable} ensembles diverse parameter-free attacks. As indicated in Table \ref{tab:robust}, BAKE successfully improves the robustness of the trained models against various perturbations. FGSM \cite{fgsm} imposes one-step adversarial perturbations on the input image with a weight of $\epsilon$. As shown in Figure \ref{fig:fgsm}, at $\epsilon=16$, BAKE significantly improves ResNet-50's accuracy from $17.4\%$ to $29.5\%$ though the model is not optimized for adversarial robustness. \begin{table*}[t] \scriptsize \begin{center} \begin{tabular}{clcccccc} \toprule Architecture & \multicolumn{2}{c}{Method} & CIFAR-100 \cite{cifar} & TinyImageNet & CUB-200-2011 \cite{cub} & Stanford Dogs \cite{dog} & MIT67 \cite{mit} \\ \midrule \multirow{6}{*}{ResNet-18} & \multicolumn{2}{c}{Vanilla} & $24.71_{\pm0.24}$ & $43.53_{\pm0.19}$ & $46.00_{\pm1.43}$ & $36.29_{\pm0.32}$ & $44.75_{\pm0.80}$\\ \cmidrule{2-8} & Label smoothing \cite{szegedy2016rethinking} & CVPR'16 & $22.69_{\pm0.28}$ & $43.09_{\pm0.34}$ & $42.99_{\pm0.99}$ & $35.30_{\pm0.66}$ & $44.40_{\pm0.71}$\\ % & DDGSD \cite{xu2019data} & AAAI'19 & $23.85_{\pm1.57}$ & \bm{$41.48_{\pm0.12}$} & $41.17_{\pm1.28}$ & $31.53_{\pm0.54}$ & $41.17_{\pm2.46}$\\ % & BYOT \cite{zhang2019your} & ICCV'19 & $23.81_{\pm0.11}$ & $44.02_{\pm0.57}$ & $40.76_{\pm0.39}$ & $34.02_{\pm0.14}$ & $44.88_{\pm0.46}$\\ % & CS-KD \cite{yun2020regularizing} & CVPR'20 & $21.99_{\pm0.13}$ & $41.62_{\pm0.38}$ & $33.28_{\pm0.99}$ & $30.85_{\pm0.28}$ & $40.45_{\pm0.45}$\\ \cmidrule{2-8} & \multicolumn{2}{c}{Our \textbf{BAKE}} & \bm{$21.28_{\pm0.15}$} & $41.71_{\pm0.21}$ & \bm{$29.74_{\pm0.70}$} & \bm{$30.20_{\pm0.11}$} & \bm{$39.95_{\pm0.20}$} \\ \midrule \multirow{4}{*}{DenseNet-121} & \multicolumn{2}{c}{Vanilla} & $22.23_{\pm0.04}$ & $39.22_{\pm0.27}$ & $42.30_{\pm0.44}$ & $33.39_{\pm0.17}$ & $41.79_{\pm0.19}$\\ \cmidrule{2-8} & Label smoothing \cite{szegedy2016rethinking} & CVPR'16 & $21.88_{\pm0.45}$ & $38.75_{\pm0.18}$ & $40.63_{\pm0.24}$ & $31.39_{\pm0.46}$ & $42.24_{\pm1.23}$\\ % & CS-KD \cite{yun2020regularizing} & CVPR'20 & $21.69_{\pm0.49}$ & $37.96_{\pm0.09}$ & $30.83_{\pm0.39}$ & $27.81_{\pm0.13}$ & $40.02_{\pm0.91}$\\ \cmidrule{2-8} & \multicolumn{2}{c}{Our \textbf{BAKE}} & \bm{$20.74_{\pm0.19}$} & \bm{$37.07_{\pm0.24}$} & \bm{$28.79_{\pm1.30}$} & \bm{$27.66_{\pm0.05}$} & \bm{$39.15_{\pm0.37}$} \\ \toprule \end{tabular} \end{center} \vspace{-15pt} \caption{Top-1 error rates (lower is better) on multiple image classification and fine-grained classification tasks. The performances of state-of-the-art single-network methods (\textit{i.e.}, label regularization methods and self-distillation methods) are reported for comparison. All the experiments are run for three times with different random seeds.} \vspace{-5pt} \label{tab:small} \end{table*} \begin{table}[t] \scriptsize \begin{center} \begin{tabular}{lcccc} \toprule \multicolumn{1}{c}{} & ImageNet & ImageNet-A & ImageNet-C & AutoAttack \\ \multicolumn{1}{c}{} & top-1 acc. $\uparrow$ & top-1 acc. $\uparrow$ & mCE $\downarrow$ & top-1 acc. $\uparrow$ \\ \midrule ResNet-50 & 76.8 & 2.7 & 57.9 & 1.4 \\ ~ + Our \textbf{BAKE} & 78.0 (+1.2) & 4.6 (+1.9) & 57.4 (-0.5) & 1.9 (+0.5) \\ \toprule \end{tabular} \end{center} \vspace{-15pt} \caption{Robustness on ImageNet-A \cite{hendrycks2021nae}, ImageNet-C \cite{hendrycks2019robustness} and AutoAttack (Linf-norm, $\epsilon=4/255$) \cite{croce2020reliable} test sets. Note that mCE is the weighted average of top-1 error rates with different corruptions (lower is better).} \vspace{-5pt} \label{tab:robust} \end{table} \begin{figure}[t] \centering \includegraphics[width=0.65\linewidth]{fgsm.png} \vspace{-10pt} \caption{Our BAKE improves adversarial robustness against the FGSM \cite{fgsm} attack. The improvements are more significant as $\epsilon$ increases. The results are reported based on ResNet-50.} \vspace{-5pt} \label{fig:fgsm} \end{figure} \vspace{-10pt} \paragraph{Ablation studies on per-class data sampling.} As described in Section \ref{sec:discuss}, we adopt a per-class data sampling strategy to ensure that the sample affinities are not uniformly low in a mini-batch. We would like to claim that (1) The per-class data sampling strategy is critical to the success of BAKE. BAKE would fail when a conventional random sampling scheme is used. (2) The gains of BAKE derive from the refined soft targets by ensembled knowledge rather than the data sampling strategies. To demonstrate our first claim, we conduct experiments when removing the proposed per-class sampling strategy. As shown in Table \ref{tab:ins}, we use ``$M=0$'' to indicate the conventional random sampling without per-class selection, and we observe that BAKE hardly improves the baseline performances. When adopting the introduced per-class sampling with $M=1$ or $M=3$, BAKE stably boosts the baseline results by up to $+1.8\%$ gains. As BAKE achieves similar improvements when $M>0$, we set $M$ as $1$ for brevity. To further verify the second claim, we evaluate ``vanilla'' when using different values of $M$. From Table \ref{tab:ins}, we observe that the performance of ``vanilla'' settings would decrease when using $M>0$, indicating that the per-class sampling would even hurt the model training if not used with BAKE. The reason might be that the per-class data sampling actually decreases the data variations within each batch. As the ``vanilla'' experiment achieves the optimal performance when $M=0$, we do not use the per-class sampling for all the ``vanilla'' experiments throughout the paper. \begin{table}[t] \scriptsize \begin{center} \begin{tabular}{lcccc} \toprule Method & $M$ & $\hat{N}$ & \# Batch & ImageNet top-1 acc. \\ \midrule \textbf{ResNet-50} & \textbf{0} & \textbf{256} & \textbf{256} & \textbf{76.8} \\ ~ + Our {BAKE} & 0 & 256 & 256 & 76.8 (+0.0) \\ \midrule ResNet-50 & 1 & 256 & 512 & 76.3 \\ \textbf{~ + Our {BAKE}} & \textbf{1} & \textbf{256} & \textbf{512} & \textbf{78.0 (+1.7)} \\ \midrule ResNet-50 & 3 & 256 & 1024 & 76.1 \\ ~ + Our {BAKE} & 3 & 256 & 1024 & 77.9 (+1.8) \\ \toprule \end{tabular} \end{center} \vspace{-15pt} \caption{Ablation studies on the value of $M$ in the per-class data sampling scheme, where $\text{batch\_size}=\hat{N}\times(M+1)$.} \vspace{-5pt} \label{tab:ins} \end{table} \begin{table}[t] \scriptsize \begin{center} \begin{tabular}{ccc} \toprule Soft targets & $\omega$ & ImageNet top-1 acc. \\ \midrule Equal to raw predictions & 0.0 & 76.3 \\ Refined by BAKE & 0.1 & 77.7 \\ Refined by BAKE & 0.3 & 77.8 \\ Refined by BAKE & 0.5 & \bm{$78.0$} \\ Refined by BAKE & 0.7 & 77.8 \\ Refined by BAKE & 0.9 & \bm{$78.0$} \\ Refined by BAKE & 1.0 & 77.9 \\ \toprule \end{tabular} \end{center} \vspace{-15pt} \caption{Ablation studies on the value of the ensembling weight $\omega$ in Eq. (\ref{eq:infty}). We report the results of ResNet-50 on the ImageNet.} \vspace{-5pt} \label{tab:ensemble_weights} \end{table} \vspace{-10pt} \paragraph{Ablation studies on the ensembling weight $\omega$.} The ensembling weight $\omega\in[0,1]$ is adopted to weigh the original knowledge of the anchor sample and the propagated knowledge of other samples within the same batch, as formulated in Eq. (\ref{eq:infty}). The refined soft targets $Q^\tau_{(\infty)}$ ensemble more knowledge from the other samples as $\omega$ gets larger. As shown in Table \ref{tab:ensemble_weights}, when $\omega$ approaches $0$, the soft targets is equal to the vanilla predictions of the anchor sample, \textit{i.e.}, $Q^\tau_{(\infty)}=P^\tau$, the distillation loss then becomes useless. When $\omega$ approaches $1$, the soft targets are totally produced with propagated knowledge from the other samples in the same mini-batch. To avoid the soft targets becoming all zeros after propagating infinite iterations, we adopt only one iteration for $\omega=1$. The performance of BAKE is robust when changing the value of $\omega$ within a large interval of $[0.1,1.0]$. \subsection{Small-scale Dataset Classification} \paragraph{Training details.} We adopt almost the same settings as those used for training the ImageNet. We use $\lambda=1.0$ and $\tau=4.0$ for Eq. (\ref{eq:self-distill}). The ensembling weight $\omega$ (Eq. (\ref{eq:infty})) is chosen from $\{0.5,0.9\}$, according to the results in Table \ref{tab:ensemble_weights}. $M$ is chosen from $\{1,3\}$, according to the results in Table \ref{tab:ins}. All the experiments are trained with only 1 GPU. More details can be found at Appendix \ref{app:training_details}. \vspace{-10pt} \paragraph{Improvements on various architectures and datasets.} As shown in Table \ref{tab:small}, we study the effectiveness of BAKE with a lightweight ResNet-18 and a deep DenseNet-121 on multiple classification datasets. BAKE consistently improves the baseline results (``vanilla'') by significant margins, \eg, on the fine-grained dataset CUB-200-2011, BAKE boosts the baseline by $16.26\%$ with ResNet-18. Also, on the widely-used CIFAR-100 dataset, the performance of ResNet-18 is improved by $3.43\%$ by BAKE. \vspace{-10pt} \paragraph{Comparison with self-distillation methods.} Following the benchmark used by \cite{yun2020regularizing}, we compare with state-of-the-art self-distillation methods, DDGSD \cite{xu2019data}, BYOT \cite{zhang2019your} and CS-KD \cite{yun2020regularizing}. BAKE stably surpasses all the methods except for the experiments of ResNet-18 on TinyImageNet, which shows a slight drop from DDGSD \cite{xu2019data}. The positive results of BAKE are enough to demonstrate its effectiveness and superiority over existing self-distillation methods. \section{Limitations and Conclusions} \vspace{-5pt} In this work, we introduce a novel batch knowledge ensembling method, dubbed BAKE, to produce refined soft targets for self-distillation. BAKE improves the image classification performance with minimal computational and memory overhead, and outperforms state-of-the-art single-network methods on all the tested benchmarks. Beyond the classification tasks, the spirit of batch knowledge ensembling has great potential on other tasks, \eg, image retrieval, segmentation and detection. BAKE does not work if the samples in a mini-batch are totally dissimilar to each other, so we introduce a per-class data sampler to solve this problem. However, the data sampler would increase the CPU time when the training dataset is large-scale since it needs to reorganize the data loader before each epoch.

12b96bbdd4a87e0447febead76251b86dc92dbbd

\section{Introduction} \IEEEPARstart{T}{he} healthcare sector was hit hard in the years 2020 and 2021 due to the spread of COVID-19 worldwide, causing a global pandemic. Hospital beds were completely full in certain cities, and intensive care units (ICU) were overcrowded. The grim scenario has led governments, industries, institutions, and researchers to rethink and adopt new innovative solutions to prevent, control, and combat such infectious diseases through new healthcare models driven by emerging technological innovations. Although some efforts were introduced that relied heavily on the use of IoT devices for symptom monitoring, such as connected Unmanned Ariel Vehicles (UAVs) and device positioning systems, those solutions are simply temporal and do not provide ongoing sustainability, disease preventiveness, patient data privacy and overall system reliability. Therefore, a novel solution that considers the concepts of distributivity (but also provides secure data sharing), self-learnability (but also provides explainability features), and autonomy needs to be considered to revolutionize healthcare systems and achieve new levels of quality healthcare for citizens globally. Recent developments in wireless communication, Machine Learning (ML), and decentralized transaction processing techniques like Blockchain, have led to significant changes in all governmental and industrial sectors. The emergence of Beyond 5G (B5G) networks is now leading such sectors to allow for intelligent and on-demand service provisioning. Today, AI is already being adopted in a multitude of healthcare applications, aiding specialists in early-stage disease detection, such as diabetes and cancer. For instance, several medical devices combined with AI have been applied to oversee early-stage heart disease, enabling specialists to better monitor and detect any potential life-threatening cases at earlier and more treatable stages. \begin{figure*}[ht] \centering \includegraphics[scale=0.5]{Figures/Health4.pdf} \caption{An Overview of a healthcare industry 4.0 scenario that integrates intelligent edge devices, blockchain technology, and machine learning with next generation networking.} \label{fig:health4} \end{figure*} Although such AI-supported solutions have been adapted in different areas of the health sector, this integration is very limited and would require significant efforts to enable a real shift towards Healthcare Industry 4.0 \cite{health4.0}, illustrated in Figure \ref{fig:health4}. Such shift will enable pre-hospital emergency care in both urban and geographically remote areas. With that said, a huge amount of data will be generated through these applications, allowing for more in-depth training and analysis. But, communicating and sharing this data requires utmost care due to its sensitivity and privacy. Patients need to feel more secure in using applications that would reveal very sensitive and private information and at the same time be able to understand the autonomous diagnosis that may be determined by AI-capable applications. Moreover, the exchange of information across multiple distributed entities will in no doubt lead to better quality level healthcare, but at the same time, such frameworks need to ensure data privacy and integrity. \begin{table*}[ht] \centering \caption{Comparing centralized and federated learning solutions when incorporated in a healthcare framework. \label{tab:table2} \resizebox{\linewidth}{!}{% {\begin{tabular}{|>{\hspace{0pt}}m{0.119\linewidth}|>{\hspace{0pt}}m{0.355\linewidth}|>{\hspace{0pt}}m{0.466\linewidth}|} \hline & \textbf{Centralized Learning} & \textbf{Federated Learning} \\ \hline \textbf{Data Ownership} & May remain with the user/source agency.\par{}May be transferred to the centralized entity.\par{}Shared for data control. & Remains at the user/source agency level.\par{}No need for shared data control. \\ \hline \textbf{Privacy/Security} & All data must be communicated with the central entity. & Model updates only.\par{}Data is not communicated with other entities. \\ \hline \textbf{Data Availability} & Frequent data retrieval to ensure accurate model. & Access to data is always available which ensures accurate local model. \\ \hline \textbf{Data Quality} & Collected data from local devices is validated.\par{}More reliable data. & Dependent on the local device.\par{}Some data used in the learning process may be unreliable. \\ \hline \textbf{Performance} & Load is placed on the centralized entity.\par{}Poor performance when considering scalability. & Load is distributed among local devices.\par{}Can accommodate for larger set of user/source agencies. \\ \hline \end{tabular} }} \end{table*} To ensure reliable early-stage warning, prevention, control and combatant of infectious diseases, rejuvenated health-care systems must consider efficient and reliable storage, processing, and sharing of massive real-time data being collected at the edge. Moreover, when dealing with complex data processing and sharing, such systems must consider and fulfill the diversified security and privacy requirements, which may be enforced either by citizens or government authorities. Given the advances and sophistication of today's edge devices, most of the training and learning can actually be accomplished in a distributed and decentralized manner, namely, Federated Learning (FL) and plug-and-play artificial intelligence (PNP-AI). Such a technique enables the training of a common model to occur with minimal aid of a centralized entity, whilst maintaining all the sensitive data to reside locally (i.e. local health institutions or patient devices). Furthermore, with the aid of Blockchain, secure data sharing and identity security are maintained, and greater transparency for autonomous diagnosis solutions can also be attained using explainable-AI. \section{Epidemic Disease Management through Intelligent Learning} \label{related} Adopting ML in healthcare systems enables machines to learn various parameters such as behaviours, symptoms and pathological variables related to patients and diseases. AI-enhanced diagnostics and treatments have been adopted in various healthcare-steered applications. In \cite{10.3390/ijerph15081596}, a framework was developed to optimize the parameters of Deep Learning (DL) techniques in order to predict infectious diseases whilst considering big data. The solution adopted a deep neural network (DNN) along with long-short term memory (LSTM) learning models for predicting three infectious diseases a week ahead of time. In \cite{Kaissis}, a solution for medical imaging applications was developed using next-generation methods for federated, secure and privacy-preserving artificial intelligence. A predictive analysis method for epidemic diseases to identify and locate infected areas by using a back-propagation method was introduced in \cite{doi:10.1063/1.5005397}. The solution-focused on classifying the epidemic disease spreading factors as the elements for weight adjustment on the prediction of epidemic disease occurrences. Furthermore, some research works have focused on analyzing the epidemic trends of Severe Acute Respiratory Syndrome (SARS) to validate their solutions. For instance, in \cite{jia2020prediction}, the authors introduced three mathematical models, namely, the Logistic model, Bertalanffy model and Gompertz model. The results have then been utilized to analyze the situation of COVID-19. The authors in \cite{libin2020deep} investigated a deep reinforcement learning mechanism to automatically learn prevention strategies in the context of pandemic influenza. Their work proves that deep reinforcement learning can be an efficient solution for mitigation strategies in complex epidemiological models with a large state space. \subsection{Federated Learning (FL)} In FL, a shared ML model is built using distributed data from various devices in which the end-devices train the model using their local data, and subsequently share the model parameters with the coordinating entity, which in essence arranges the model training and gathers the contributions from all devices, without sharing its own local data. A centralized server then builds the global model by aggregating and averaging all collected parameters and broadcasts the new updated model to all end-devices. Each device will then upload its own local model to the server and download the global model. The concept behind decentralized learning is to replace centralized communication with peer-to-peer (P2P) communication. FL is very crucial for modernized healthcare systems. The most important advantage of using FL in critical healthcare applications is that patient data is not shared with other devices and nodes, hence, remaining at the patient's device. This will guarantee the desired level of patient data privacy. Furthermore, the use of FL is more convenient and cheaper than traditional centralized learning approaches, due to the minimal need for data updates and the no latency issues given that the model is at the end-users’ devices, allowing for real-time inference. A comparison between the centralized learning approach and that of the federated learning approach when incorporated in a healthcare framework is summarized in Table \ref{tab:table2}. \subsubsection{Improving FL Efficiency and Effectiveness} Various tactics can be adapted to improve the effectiveness of FL on the healthcare framework and its efficiency, such as adopting various models to various users, acquiring better optimization algorithms, and upgrading the communication efficiency. \begin{itemize} \item Viral infectious diseases, such as COVID-19, are characterized as viruses that mutate, resulting in different variants and strains. As such, the idea of adapting personalized solutions for different types of end-users becomes very powerful and efficient. The use of different and various models for different end-users would outperform the use of a shared global model. \textit{Personalization through featurization} is a example of this technique, where different end-users run different model gradients (i.e. parameters). Furthermore, \textit{multi-task learning} is another approach that can be accomplished by considering each client's local issue as a distinct task (instead of as a shared one). Any multi-task learning algorithm will be applicable in that case. Additionally, \textit{fine-tuning} is a technique that starts with FL as a single model and applies that model to all end-users. The main difference in fine-tuning is that before the shared model is utilized to make predictions on the end-devices, an additional final training process is considered in order to personalize the model to the users' local data-set. \item The goal of FL is to identify a global model that dismisses the risk function over the training data-set, which is, the federation of the data among all end-users. Since the total number of end-devices in FL is significant, the necessity for \textit{client-sampling} algorithms (that necessitate a handful of clients to contribute to each round) is essential. Furthermore, each end-device is expected to participate only once in the model training, which raises the need for state-less algorithms. On the other hand, the optimization algorithms have to be \textit{composable} with other techniques. This is very crucial for a healthcare framework used to combat infectious diseases. Up-to-date data must be gathered to create local models and then shared globally to create accurate models that reflect the situation at different distributed ends. By adapting client-sampling algorithms, faster decisions can be taken to avoid the further spread of disease. \item Since wireless edge communication for end-devices may be unreliable and operates at low data rates at certain times, wireless communication could be a crucial bottleneck for FL. This drawback has led to substantial interest in lowering the communication bandwidth, in which a trade-off between communication and accuracy exists in FL. With that being said, with today's 5G wireless communication technology, such a bottleneck will be eliminated to provide enhanced and improved edge communication to support the FL process. \end{itemize} Table \ref{tab:DLvsFL} summarizes some of the benefits and drawbacks of adapting FL to preventive healthcare systems over a distributed learning approach. \begin{table*}[ht] \centering \caption{Comparing federated learning and distributed learning in terms of their advantages and disadvantages when adapted to critical infrastructures like healthcare frameworks.} \label{tab:DLvsFL} \resizebox{\linewidth}{!}{% {\begin{tabular}{|>{\hspace{0pt}}m{0.167\linewidth}|>{\hspace{0pt}}m{0.363\linewidth}|>{\hspace{0pt}}m{0.413\linewidth}|} \hline & \textbf{Distributed Learning} & \textbf{Federated Learning} \\ \hline \textbf{Data Distribution and Accessibility} & Centralized storage of patient data.\par{}Patient data is accessible by end-users. & Patient data is generated and preserved locally.\par{}Decentralized storage of patient data.\par{}Patients do not have access to data stored at other end-devices (either local or shared). \\ \hline \textbf{Data Availability} & All participating patients are assumed to be available. & Only a fraction of patients must be available at any point of time. \\ \hline \textbf{Model Setting} & Training is conducted on a large data-set.\par{}End-users are nodes in a cluster or data-center. & Training is conducted using a local dataset on patients' end-devices. \\ \hline \textbf{Communication} & Communication cost is low.\par{}No wide-area communication. & Communication cost is considered high (wide-area communication).\par{}Consistent connection with end-users. \\ \hline \textbf{Drawbacks} & Computation is processed at the data-center.\par{}Global model is not highly accurate since patient data is not retrieved frequently. & Diversified end-device capabilities may cause issues with computation and communication. \\ \hline \textbf{Advantages} & Few end-user failures. & Local models are constantly accurate, leading to a more accurate global model.\par{}Highly effective for infectious disease scenarios. \\ \hline \end{tabular} }} \end{table*} \begin{comment} \begin{table*}[ht] \caption{Comparing federated learning and distributed learning in terms of their advantages and disadvantages when adapted to critical infrastructures like healthcare frameworks.} \label{tab:DLvsFL} \resizebox{\textwidth}{!}{% \begin{tabular}{|c| >{\columncolor[HTML]{FFFFFF}}l | >{\columncolor[HTML]{FFFFFF}}l |} \hline \multicolumn{1}{|l|}{} & \multicolumn{1}{c|}{{\color[HTML]{000000} \textit{\textbf{Distributed Learning}}}} & \multicolumn{1}{c|}{{\color[HTML]{000000} \textit{\textbf{Federated Learning}}}} \\ \hline {\color[HTML]{000000} \textit{\textbf{Data Distribution}}} & {\color[HTML]{000000} \begin{tabular}[c]{@{}l@{}}- Centralized storage of data.\\ - Data is accessible by end-users.\end{tabular}} & {\color[HTML]{000000} \begin{tabular}[c]{@{}l@{}}- Data is generated locally.\\ - Decentralized storage of data.\\ - Each end-user preserves its own data.\\ - End users do not have access to data stored at \\ other end-devices.\end{tabular}} \\ \hline {\color[HTML]{000000} \textit{\textbf{Model Setting}}} & {\color[HTML]{000000} \begin{tabular}[c]{@{}l@{}}- Training the model on a large data-set.\\ - End-users are nodes in a cluster or data-center.\end{tabular}} & {\color[HTML]{000000} \begin{tabular}[c]{@{}l@{}}- Training is conducted locally on end-devices.\\ - End-users are considered IoT or mobile end-devices.\end{tabular}} \\ \hline {\color[HTML]{000000} \textit{\textbf{Data Availability}}} & {\color[HTML]{000000} - All users are assumed to be available.} & {\color[HTML]{000000} \begin{tabular}[c]{@{}l@{}}- All end-users might be available.\\ - A fraction of users are available at any point of time\\ in certain FL scenarios.\end{tabular}} \\ \hline {\color[HTML]{000000} \textit{\textbf{Main Drawback}}} & {\color[HTML]{000000} \begin{tabular}[c]{@{}l@{}} - Computation is processed at the data-center\\ (i.e. single entity). \end{tabular}} & {\color[HTML]{000000} \begin{tabular}[c]{@{}l@{}}- Diversified end-device capabilities\\ which may cause issues with computation and\\ communication.\end{tabular}} \\ \hline {\color[HTML]{000000} \textit{\textbf{Addressability}}} & {\color[HTML]{000000} \begin{tabular}[c]{@{}l@{}}- Each end-user has its own identity that allows\\ the system to reach it.\end{tabular}} & {\color[HTML]{000000} \begin{tabular}[c]{@{}l@{}}- Each end-user has its own identity or \\ name that allows the system to reach it.\end{tabular}} \\ \hline {\color[HTML]{000000} \textit{\textbf{End-users' Reliability}}} & {\color[HTML]{000000} - Few failures.} & {\color[HTML]{000000}\begin{tabular}[c]{@{}l@{}} - Few to moderate levels of failures,\\ depending on device capabilities. \end{tabular}} \\ \hline {\color[HTML]{000000} \textit{\textbf{Wide-area communication}}} & {\color[HTML]{000000} \begin{tabular}[c]{@{}l@{}}- No wide-area Communication.\\ - Users are grouped into one cluster.\end{tabular}} & {\color[HTML]{000000} \begin{tabular}[c]{@{}l@{}}- A coordinating service provider is required \\ along with consistent connection with end-users.\end{tabular}} \\ \hline \end{tabular}% } \end{table*} \end{comment} \subsubsection{Maintaining Patient Data Privacy} The major issue that faces any healthcare framework is the patient data privacy policy enforced by governments. FL provides a novel technique to address privacy and data governance challenges. It enables ML to occur on non-co-located data. In FL settings, the owner of the local data controls and defines its own governance processes and the associated privacy policies. It also controls entities that may have access to the data and have the ability to revoke accessibility at any moment. With that said, different governmental jurisdictions can enforce their rules and policies on the healthcare framework in a distributed and decentralized manner. End-user FL participants will need to abide by the local governed rules in regards to data sharing. By moving the model to the data and not vice versa, allows for high-dimensional storage-intense data to remain stored locally without the need for data communication and duplication. If for instance, certain regional government policies dictate that medical data be stored locally within the health region, the local training of the FL process can also occur at the regional level. As such, the proposed healthcare framework is very flexible and can accommodate real-world healthcare scenarios, especially those that require fast and immediate attention, such as in the prevention and control of epidemics. Some threats may still arise with the use of FL. For instance, an adversary may try to stop a model from being learned or might bias a model under training to generate the adversary's preferred inferences. Therefore, FL requires rigorous protection against such threats. Threat models might arise from the end users' devices or even from the centralized entities. Table \ref{threats_examples} presents some examples of such threats. \begin{table*}[ht] \centering \caption{FL-based threats that may arise as a result of malicious attacks on the healthcare framework.} \label{threats_examples} \resizebox{\linewidth}{!}{% {\begin{tabular}{|>{\centering\hspace{0pt}}m{0.102\linewidth}|>{\hspace{0pt}}m{0.211\linewidth}|>{\hspace{0pt}}m{0.628\linewidth}|} \hline \multicolumn{1}{|>{\hspace{0pt}}m{0.102\linewidth}|}{\textbf{Targeted Entity}} & \multicolumn{1}{>{\centering\hspace{0pt}}m{0.211\linewidth}|}{\textbf{Data Accessibility}} & \multicolumn{1}{>{\centering\arraybackslash\hspace{0pt}}m{0.628\linewidth}|}{\textbf{Threat Scenario}} \\ \hline \textbf{End-users} & Authorized entities that have access to the end-devices. & Decentralized storage of patient local data may lead to the exploitation of model parameters and training data by attackers.\par{}Malicious devices might have unauthorized access to the learned models.\par{}Some end-users may leverage the benefits of the FL healthcare framework by accessing global models without contributing to the training process.\par{}Unavailability of participant end-users~may yield inefficient results in training the global model. \\ \hline \textbf{Central Server} & Authorized entities that have access to the server. & Attacks on the central server may lead to vulnerability of~initial model parameters, aggregated local models and shared global models.\par{} \\ \hline \textbf{Communication Channel} & Anyone & FL requires~significant amount of communication over the network, where a non-secure communication channel is open to vulnerability through eavesdropping attacks. \\ \hline \textbf{Trained Models} & Developers, data scientists, and software architects. & Malicious end-users may use patient data poisoning attacks on the global model by manipulating the training process.\par{}Malicious end-users may~modify the updated local models before sending them to the central server. \\ \hline \textbf{Global Models} & Anyone & Data reconstruction attacks may lead to global data vulnerability.\par{}Abnormality in local model updates from suspicious end-users may go unnoticed, making the global model vulnerable. \\ \hline \end{tabular} }} \end{table*} \begin{comment} \begin{table*}[!b] \caption{FL-based threats that may arise as a result of malicious attacks on patients' data.} \label{threats_examples} \resizebox{\textwidth}{!}{% \begin{tabular}{| >{\columncolor[HTML]{FFFFFF}}c | >{\columncolor[HTML]{FFFFFF}}c | >{\columncolor[HTML]{FFFFFF}}l |} \hline \multicolumn{1}{|l|}{\cellcolor[HTML]{FFFFFF}{\color[HTML]{000000} \textit{\textbf{Targeted Entity}}}} & {\color[HTML]{000000} \textit{\textbf{Malicious}}} & \multicolumn{1}{c|}{\cellcolor[HTML]{FFFFFF}{\color[HTML]{000000} \textit{\textbf{Threat Scenario}}}} \\ \hline {\color[HTML]{000000} \textit{\textbf{End-Users}}} & \multicolumn{1}{l|}{\cellcolor[HTML]{FFFFFF}{\color[HTML]{000000} Authorized entities that have access to the end-devices.}} & {\color[HTML]{000000} \begin{tabular}[c]{@{}l@{}}- Can interfere in the training procedure.\\ - Can eavesdrop on all servers' messages.\end{tabular}} \\ \hline {\color[HTML]{000000} \textit{\textbf{Centralized Server}}} & \multicolumn{1}{l|}{\cellcolor[HTML]{FFFFFF}{\color[HTML]{000000} Authorized entities that have access to the server.}} & {\color[HTML]{000000} \begin{tabular}[c]{@{}l@{}}- Can interfere in the training procedure.\\ - Can eavesdrop on all messages sent from the end devices, \\ such as the parameter changes.\end{tabular}} \\ \hline {\color[HTML]{000000} \textit{\textbf{Trained Models}}} & {\color[HTML]{000000} Data scientists} & {\color[HTML]{000000} \begin{tabular}[c]{@{}l@{}}- Data scientists or engineers might gain access to \\ numerous outputs, such as, the results of different \\ models' training runs.\end{tabular}} \\ \hline \multicolumn{1}{|l|}{\cellcolor[HTML]{FFFFFF}{\color[HTML]{000000} \textit{\textbf{Finalized Models}}}} & {\color[HTML]{000000} Anyone} & {\color[HTML]{000000} \begin{tabular}[c]{@{}l@{}}- The finalized models of the FL might be adopted\\ by hundreds of millions of end-devices. \\- Malicious devices might have unauthorized access\\ to the learned models.\end{tabular}} \\ \hline \end{tabular}% } \end{table*} \end{comment} \section{Secure Data Sharing through Blockchain} A distributed Healthcare framework, where data is collected from patients, filtered, trained, analyzed and processed either locally or through neighbouring nodes, will require a secure and private communication mechanism, that not only is secure but also reliable to ensure data and transaction authenticity and integrity. Additionally, with such a cooperative healthcare framework, in which local and centralized healthcare entities may require real-time access to a collection of patients' records in order to provide timely response to ensure responsive actions towards suspicious infectious disease. Hence, privacy-preserving schemes are essential to ensure proper system integration of all healthcare entities. Data access, the tracking of users' identity, and raw data disclosure are also considered mandatory aspects for any state-of-the-art distributed and intelligent healthcare framework. Blockchain is highly essential for such cooperative and distributed frameworks, where the healthcare system will be comprised of a significant number of organizations and users having different roles (e.g. patients, healthcare workers, pharmacies, health insurance companies, etc.). It provides a secure mechanism to exchange and store sensitive medical data in a distributed and decentralized manner. Blockchain transactions are considered fast and can provide a mechanism to aggregate policies of different and diversified healthcare entities. This will allow for such diversified entities to cooperate and share a unique model among different healthcare sectors. The adopted blockchain consensus algorithm will make it nearly impossible to modify transaction records, since most of the participants need to verify each block of data \cite{consensus}. This will in essence eliminate cyber-attacks that may occur on data storage sites in a traditional centralized entity scenario. Many researchers, companies and agencies across the globe started developing blockchain-supported applications to aid in countering the COVID-19 pandemic disease. Such applications can validate the continuously changing data, which is quite effective in handling the rapidly escalating COVID-19 situations. In \cite{app1}, the authors launched a Blockchain-supported application, namely Civitas, which is considered a safety system for controlling the impact of COVID-19 by associating people's IDs with Blockchain records in order to verify whether a person in self-quarantining or not. By adopting blockchain to this application, it ensures the security and privacy of patients' data. Another blockchain-supported application, MiPasa, has been introduced in \cite{app2}, which is a data streaming platform that is able to facilitate the sharing of verified health information between people, agencies, and hospitals. It works by gathering the information offered by medical organizations. \subsection{Current Trends in Blockchain-Based Healthcare Systems} Numerous attempts to secure distributed health-care systems using blockchain have been considered lately. Such attempts consider applying the use of blockchain at either the health-care institutions' side (such as hospitals and clinics) or at the patient side. Institution-based blockchain solutions do not involve patients in the secure data sharing and transaction processing aspect, but rather allow them to interact with health institutions to acquire services that have been authorized and authenticated. On the contrary, patient-based blockchain-solutions allow patients to participate in both data sharing and processing transactions. As such, patients are directly involved in the healthcare service provisioning process. Such architectures would require the use of both public and private blockchains. We would assume that involving patients (or even users that contribute and share collected health-related data) requires the use of public blockchains to offer such a decentralized framework. On the contrary, institution-based blockchain solutions would adopt a private blockchain framework solution to allow only authenticated or verified individuals to get involved in data sharing and processing transactions. \subsubsection{Scalability Issues} Current blockchain-based healthcare system trends face major issues in regards to the storage of patient health information and learnt models. For instance, in \cite{scalabilityIssues}, the authors proposed a framework for blockchain-based personal health information sharing among healthcare institutions for faster and more reliable disease diagnosis. They integrate two blockchains in their solution, namely, private and consortium blockchains. The private blockchain of healthcare institutions stores raw encrypted patient health information, while the consortium blockchain stores records of the secure indices of patients' health information. Proof of conformance is used for the blockchains as the consensus mechanism. The solution limits the collection and sharing of data among healthcare service providers only and requires patients' approval. Such a technique although is promising, would create storage scalability issues as more data is collected from patients. \subsubsection{Latency Issues} Other issues that may face blockchain-based healthcare solutions is the excessive amount of time needed for health authorities to respond to critical infectious disease events through data collected at the patients' side, then shared and communicated to other participants at the edge through public blockchains. By the time this information is shared with the healthcare providers and government authorities, epidemics might have gone out of control. For instance, in \cite{latencyIssues}, the authors tackled the security issues that may arise as a result of transferring and logging data transactions in healthcare systems. Permissioned, consortium-based blockchains are used to execute smart contracts to evaluate information collected from patients. Alerts are triggered for patients and healthcare providers under critical conditions. Although such a solution might provide a mechanism to collect data from patients in real-time, then authenticate this data and communicate it to the health authorities to take action, especially in epidemic related cases, the time needed to filter and analyze the collected information at the healthcare provider side would result in unsatisfactory outcomes. \subsection{Hybrid Blockchain-Based Healthcare Framework} Hybrid solutions which would assign certain roles to institutions and other roles to users would provide a more optimal strategy towards data collection, sharing, and transaction processing. We envision a healthcare environment that relies heavily on data collection, aggregation, filtering and partial analysis (i.e. Federated Learning) that is achieved using a public blockchain. Data storage and service composition may also be achieved at the edge. Moreover, the healthcare framework would also rely on private blockchains to link various medical parties for data sharing to achieve faster and wider service responsiveness, especially for epidemic outbreak scenarios. Figure \ref{fig:layering} depicts a solution that adopts such a hybrid blockchain solution. This approach will not only link the different users at the edge layer, or the institution layer separately, but it will also allow for blockchain system links to occur between patients and doctors. This can be achieved through customized smart contracts, allowing for patients to have on-demand medical data when moving between different healthcare institutions \cite{smartContract}. Adopting FL in the Blockchain-based healthcare system helps in parallelizing the computation capacity of numerous parameters' updates over numerous end-devices. Additionally, this learning environment is an asynchronous learning style where each end-user is independently involved at random occasions which eliminates the need for a global synchronization entity while improving scalability. \begin{figure}[ht] \centering \includegraphics[scale=0.29]{Figures/Layering.pdf} \caption{An Overview of a hybrid blockchain solution. The solution adopts a decentralized data collection and federated learning aspect which is applied to a public blockchain for secure sharing and transaction authentication. Healthcare institutions use the collected information and epidemic alerts to provide healthcare services to concerned communities and share data securely with other institutions through private blockchain.} \label{fig:layering} \end{figure} There have been a number of recent attempts to address issues related to scalability and latency for blockchain-based healthcare systems. Shamim et al. \cite{shamim} developed a mass surveillance strategy at the network edge to monitor human social distancing, mask-wearing, and body temperature readings. The data is analyzed at the edge by relying on 5G network edge servers. Additionally, data collected in hospitals such as chest x-rays, CT scans, and ultrasound, is also analyzed at the edge. Both data collected from users and hospitals is trained on the cloud using deep learning (DL) models, namely, ResNet50, deep tree, and Inception v3. The trained models are downloaded to the edge during off-peak hours. New data collected either from users or hospitals is sent to the DL models available at the edge. But prior to the data being sent, it is applied on a blockchain to both authenticate and secure the data. Such an approach ensures that the system is scalable in terms of both storage and the number of participants, in addition to providing fast responsiveness for infectious disease diagnosis, results and alerts. \section{Future Insights} We envision that in addition to its critical role in patient data exchange among different health entities, blockchain could be used today for setting check-up stations for testing individuals who show symptoms related to an infectious disease. The number of conducted tests in the stations can be managed over the blockchain which can assist in getting reliable reports on the number of conducted tests as well as the number of positive cases. It can help health agencies fight disease spread in areas that have a high number of positive cases. Blockchain can be trusted by healthcare agencies due to its aspect of being immutable. Furthermore, blockchain along with federated learning could be used to store the patients' records, who tested positive to an infectious disease, to contain all their details (e.g. gender, region, age, close contact, etc.) as individual instances, where federated learning could be used to learn and better understand the disease expansion pattern along with facilitating the prediction property. \subsection{Explainable and Plug-and-Play AI} Given the diversity and abundant numbers in both edge and core network devices having advanced processing, storage, and communication capabilities, and given that FL is an optimal choice for decentralized and secure health-care systems capable of adapting to the changing health conditions at different health districts, the adoption of a generalized AI technique that is adaptable to a wide range of devices and applications is highly necessary, namely plug-and-play (PnP) \cite{PnP-AI}. The vision of a simplified and generalized learning solution that may be adapted to any problem and domain, is one that may accelerate the technological advancements in AI and will have a high impact on healthcare provisioning. Generalized AI refers to a system that can autonomously decide the type of ML algorithm, the training and the processing of the data-set, as well as the reasoning of the data-set selection for optimal feature extraction. Moreover, generalized AI is able to autonomously perform the training and fine-tuning in order to provide an optimal level of generality at the end-devices. Given that all participating end-devices are part of the learning process, PnP-AI would simplify and speed up the FL process. The goal of the PnP-AI solution is to be adaptable to a wide range of edge and IoT devices. PnP-AI is more beneficial when adapted to resource-limited IoT devices, as it gives support to edge devices. The solution is adaptable in the FL context and is highly beneficial, where the learning process is offloaded to end-devices. In health-related scenarios, the type of diagnosis mechanism needs to first be identified before applying a particular ML algorithm. For instance, in a case where a patient is showing flu-like symptoms, such as high fever, coughing, and low levels of oxygen, which are identified either through body sensors or the use of cooperative devices distributed in a given environment, the type of problem and the input data to be used for training purposes needs to be determined. Such determination of the problem and input data needs to be achieved without user intervention. Then, the training on the data-set, whether performed locally (i.e. distributed) or through FL, the labelling and models are all selected and performed autonomously. Such an approach will require two layers of learning, namely, one that focuses on accurate identification of the configurations and learning mechanisms for the given health-related problem to diagnose the disease, and the other focuses on the problem domain to determine the treatments and actions to be taken, whether locally or by the local healthcare authorities. With autonomy, comes some disadvantages in terms of system transparency. PnP-AI will further widen the gap between the end-user and the mechanisms involved in decision-making. Both authorities and patients are reluctant to involve the use of AI in medical applications. For AI to be rolled out at a much faster pace in the health-care sector, and due to the sensitivity of the information and decisions/actions made, providing detailed insight into the training, learning, and decision-making process of AI systems is necessary. Explainable AI is a mechanism that would enable such transparency, allowing for AI techniques to be defined, implemented, and interpreted to improve user trust in ML. To achieve a high standard of ethics in AI, explanations of the systems' decisions made by ML algorithms need to be provided to the users. Explainability needs to be adapted at different layers of the system and interpreted for different users according to their expertise. This will ensure user verification, safety and trust of AI results. For instance, decisions made in regards to infectious diseases need to be explained to different categories of healthcare workers and patients. \begin{figure*}[ht] \centering \includegraphics[scale=0.5]{Figures/SolutionOverview.pdf} \caption{An Overview of an integrated explainable and PnP-AI solution for distributed healthcare frameworks in a cooperative environment to prevent and support fast control and combat of infectious diseases.} \label{fig:solutionOverview} \end{figure*} Figure \ref{fig:solutionOverview} provides an overview of a healthcare framework that adapts both an explainable and PnP-AI solution. Explainable models need to be created to match the problem domain (e.g. infectious diseases) and the expertise level of the user (e.g. infectious disease specialist, family doctor, nurse, patient, etc.) to describe the data labelling and model selection processes. Since this stage is very critical to ensure accurate diagnosis and actions, the set of users need to be accurately defined. The selection of the ML algorithm will also need to be justified to the user with reasoning to provide more of a glass-box solution style, thus enabling transparency. This will ensure that results presented to the end-user are justified to overcome any safety and health concerns. \subsection{B5G Network Support} With the advances in 5G and beyond networks in terms of their mobility and scalability support, mass surveillance and data acquisition at the edge will enable rapid and on-demand health decisions and actions to be considered frequently. Real-time data analysis achieved with the support of roaming end-devices and ultra-fast wireless communication will provide reliable health-decision outcomes in both regional and global avenues \cite{mobility}. Large multi-dimensional datasets and/or locally trained models generated in areas considered highly infectious can be exchanged across 5G networks to generate global models or used by DNNs across different governmental jurisdictions for model training. Medical staff will be able to effectively and promptly respond to urgent cases and have access to regional medical information while roaming across different networks. This will minimize the risk of coming into contact with those that have been infected and will be able to dispatch medical support remotely. Vertical training, where cooperative machine learning can occur at both the edge and cloud can be conducted more efficiently (in terms of reduced latency and accurate modeling) with the support of B5G. Edge servers with the support of edge caches and edge nodes can provide fast decision-making and monitoring near regional health facilities. Trained models at the edge can be shared with the cloud for further and more accurate global DL. Results can then be rapidly deployed to different jurisdictions globally to ensure synchronized community-based health cooperation. Drug and vaccine manufacturers can determine the frequency and number of vaccines that need to be rolled out to different communities. \section{Lessons Learned and Concluding Remarks}\label{concluding} The COVID-19 pandemic has greatly affected and temporally suppressed the majority of healthcare systems globally. This has influenced researchers from different disciplines to develop new techniques and solutions that would enable epidemic discovery ahead of time, remote monitoring, and fast health-authority response. In this article, we envisioned a distributed and decentralized healthcare framework for epidemic diseases' screening, monitoring and rapid reaction. The framework relies on Federated Learning (FL) and Blockchain technology which enable a cooperative and distributed healthcare system. We proposed a PnP-AI solution that is capable of adapting to changing device specifications and health conditions. Additionally, we proposed that explainability be adapted to health systems to ensure system decision transparency to all users according to their expertise level. For FL-enabled healthcare frameworks to be implemented efficiently, the training process should occur in a distributed manner (i.e. local device training). As such, the adaptation of a PnP-AI solution would ensure proper training and learning to occur at the edge. The implementation of such a solution is indeed complex and is only at its early stages. The concept of achieving a PnP-AI solution would in itself require two layers of learning. One that targets the problem domain and another which targets the management aspect of machine learning. This is considered resource-consuming. As such, the management layer needs to be decentralized and distributed, which would require collaboration between different health entities. Furthermore, the application of blockchain at two levels, namely, user and institution, will provide an opportunity for Electronic Health Records (EHR) to be shared only among those that are given access to the intended data with limited on-chain data storage. The hash of the data is also stored on-chain, thus preserving patient data security. Although such a system is considered complex, we believe that such a proposal will create novel and efficient solutions that are more convenient not only for health authorities but also for patients and citizens. \begin{comment} \begin{table}[] \caption{} \label{tab:my-table} \resizebox{\textwidth}{!}{% \begin{tabular}{|c| >{\columncolor[HTML]{FFFFFF}}l | >{\columncolor[HTML]{FFFFFF}}l |} \hline \multicolumn{1}{|l|}{} & \multicolumn{1}{c|}{\cellcolor[HTML]{EFEFEF}{\color[HTML]{000000} \textit{\textbf{}}}} & \multicolumn{1}{c|}{\cellcolor[HTML]{EFEFEF}{\color[HTML]{000000} \textit{\textbf{}}}} \\ \hline \cellcolor[HTML]{EFEFEF}{\color[HTML]{000000} \textit{\textbf{}}} & {\color[HTML]{000000} } & {\color[HTML]{000000} } \\ \hline \cellcolor[HTML]{EFEFEF}{\color[HTML]{000000} \textit{\textbf{}}} & {\color[HTML]{000000} } & {\color[HTML]{000000} } \\ \hline \cellcolor[HTML]{EFEFEF}{\color[HTML]{000000} \textit{\textbf{}}} & {\color[HTML]{000000} } & {\color[HTML]{000000} } \\ \hline \end{tabular}% } \end{table} \end{comment} \section*{Acknowledgment} This research was supported by the Faculty of Technological Innovation, Zayed University (ZU), under grant number STG-046. \ifCLASSOPTIONcaptionsoff \newpage \fi \balance \bibliographystyle{IEEEbib}

07b2eaceae5b0fccf2a281320dc519d53841817b

\section{Introduction} The familiar classes of conjugated unsaturated hydrocarbon molecules, such as benzenoids, coronoids, helicenes and more general fusenes, may all be regarded in a mathematical sense as sets of graphs equipped with additional properties. In the simplest case, the hexagonal rings of such molecules may be considered as faces of a map on the plane. In this note we extend this notion by retaining the local properties of benzenoids but relaxing global planarity. Since the first isolation of benzene almost 200 years ago, benzenoids and their derivatives have had a significant, if not always benign, presence in the mainstream of organic chemistry and its applications. Mathematical study of benzenoids also has a long history, with central ideas \cite{Crocker1922,Armit1925} contributed by pioneering experimental chemists such as Kekul{\'e} \cite{Kekule1865}, Fries \cite{Fries1927} and Clar \cite{clar1964,clar1964a,Clar1972} feeding into an enormous primary literature codified in influential textbooks \cite{cyvin_1988, cyvin1991,cyvin1994,gutman1989,trinajstic1992}. The present paper is dedicated to another major figure, Milan Randi\'{c}, whose ideas on the use of conjugated circuits \cite{Randic1976Ccar} for the description of resonance energy have been influencing thinking in this area for nearly half a century \cite{RaNiTr1987,ElRa1987,Ra1991,Ra1993,PiRa2000,GuRa2001,VuRaBa2004,BaRa2005,BaRa2007,BaRaVu2008,RaNoPl2012}. Most recently, his simple but insightful picture of ring-current aromaticity of benzenoids has revived interest in ways of modelling and, especially, of interpreting molecular currents \cite{Gomes1979,Randic2010Gtat, Rand2011,Rand2012A,Rand2012B,Rand2012C,Cies2009,Mandado2009,PWF2011,PWF2016,PWF2020}. The simplest structures, with many applications in chemistry, are \emph{benzenoids}. Graphene may be viewed as an infinite benzenoid. In this paper we are interested only in finite structures. There are several ways to describe a finite benzenoid: boundary-edges codes \cite{guo2002,BasFowPis2017,deza1990}, inner duals \cite{Nikolic1990,Mallion1983}, flag-graphs \cite{kovic_2014,balaban2012}, or through the coordinates of the hexagons in the infinite hexagonal tesselation of the plane \cite{Bas2017}. Benzenoids having no inner vertices (i.e.\ no vertices common to three hexagons) are called \emph{catacondensed}, whilst those having inner vertices are called \emph{pericondensed}. Among catacondensed benzenoids we distinguish \emph{branched} and \emph{unbranched} benzenoids. The simplest unbranched benzenoids are \emph{linear benzenoids} or \emph{linear polyacenes}. If the structures are allowed to spiral we call them\emph{ helicenes}. These are still planar, in the graph theoretical sense, and simply connected but no longer fit onto a hexagonal grid without overlap. The term \emph{fusenes} covers both benzenoids and helicenes. Note that the boundary does not determine uniquely a general fusene; see for instance work by Brinkmann \cite{Brinkmann2002,Brinkmann2007}. Benzenoids with \emph{holes} (i.e.\ those that are not simply connected) are \emph{coronoids}. Again, those that have no internal vertices are catacondensed coronoids (or perhaps more simply, \emph{catacoronoids} to correspond to \emph{catabenzenoids}). Benzenoids and coronoids have both been considered as maps on a surface with boundary \cite{kovic_2014,balaban2012,Basic2016}. Fusenes can be further generalised to allow for structures that are not necessarily simply connected. In the literature, various generalizations to surfaces of higher genus have been made. For instance, \emph{torusenes} (also called toroidal polyhexes or torenes) have been considered \cite{KiMaPo1993,MaPi2000,KiPi2007}. Since we may tile the Klein bottle by hexagons \cite{Deza2000}, we may also speak of \emph{kleinbottlenes}. There is a whole menagerie of proposed finite and infinite theoretical carbon nanostructures, such as M\"{o}biusenes, tubulenes, hexagonal systems, hexagonal animals, toroidal benzenoids, Schwarzites, Haeckelites, etc.\ \cite{Gu1984,Sa1984, SaHaZh1996,TrZi2015,TrZi2015a,Tr2017,Terrones1993,Terrones2000}. The theory of maps \cite{Gross1987,PiPo2004} offers a toolbox for a general treatment of these diverse structures. Note that each map on a surface determines a graph, called the \emph{skeleton} of the map, that is obtained by discarding the faces of the map and retaining the vertices and edges. Whilst the skeleton is uniquely determined by the map, the converse is not true. A given graph may be a skeleton of several non-isomorphic maps. This fact has long been known to geometers: it was already Johannes Kepler who presented non-convex regular polyhedra \cite{Kepler1997}. For instance, the great dodecahedron has the same skeleton as the icosahedron. Another example is the skeleton of the tetrahedron, which is the complete graph $K_4$. The graph $K_4$ is also the skeleton of the hemihexahedron (also called hemicube), a map with three quadrilateral faces in the projective plane \cite{Coxeter1973,McMullen2002}. In mathematical chemistry this problem is relevant when counting the number of distinct toroidal polyhexes. One has to choose whether to count graphs or maps. Pisanski and Randić \cite{PiRa2000} give the example of the cube graph ($Q_3$), which has two non-equivalent hexagonal embeddings in the torus; see also Figure~\ref{fig:threeembeddings} below. In the next section, we present a flexible language for describing benzenoids and their many generalisations. \section{Polygonal complex} \subsection{Scheme} Following Ringel \cite{Ri1974}, one can describe a cellular embedding of a graph in a closed surface by a \emph{scheme}. Here we generalise Ringel's approach in two directions. If we do not insist that each symbol appears exactly twice, we may use such schemes to describe the combinatorial structure of more general polygonal complexes in the sense of Schulte \emph{et al.}~\cite{PeSc2010,PeSc2013,PeSc2014,ScWe2017}. On the other hand, if we allow symbols with a single appearance, we may describe chemical structures, such as benzenoids as graphs embedded in a surface with a boundary. Assume we are given a finite alphabet $A$. To each symbol $a \in A$ assign two literals $a^+, a^-$. We say that $a^+$ is inverse of $a^-$ and that $a^-$ is inverse of $a^+$. Hence, if alphabet $A$ has $n$ symbols, there are $2n$ literals. When there is no ambiguity, we will write $a$ for $a^+$. A \emph{word} over literals denotes an oriented polygon. A sequence of words, also called a \emph{scheme}, denotes a \emph{polygonal complex}, i.e.\ collection of polygons, some glued along their edges. A double appearance of a symbol represents the gluing. If the symbols appear in the same literal, the gluing is \emph{parallel}; otherwise it is \emph{antiparallel}. This terminology is used in the description of polyhedral self-assembly in synthetic biology \cite{FiPiRu2014,Koetal2016}. Usually, we present a scheme in a tabular form, where each row corresponds to a word. Ringel \cite{Ri1974} defines some operations on schemes that induce an equivalence relation such that two equivalent schemes define the same polygonal complex. Two schemes are \emph{equivalent} if one can be obtained from the other by a sequence of transformations of the following types: \begin{enumerate}[label=(T\arabic*)] \item \label{T:1} Permute the rows of a scheme (since we may always reorder the list of polygons); \item \label{T:2} Make a cyclic permutation of a row (since we may always start following the edges of a polygon from any of its vertices); \item \label{T:3} Replace any symbol by an unused symbol while keeping the exponents (since we may always relabel the edges of the polygonal complex); \item \label{T:4} Pick a symbol and replace each occurrence of a literal by its inverse (since we may always reverse the direction of any edge); \item \label{T:5} Reverse the row and simultaneously replace each literal by its inverse (since we may always reverse the orientation of any polygon). \end{enumerate} A scheme may satisfy some additional properties. For example: \begin{enumerate}[label=(S\arabic*)] \item \label{prop:connected} A scheme is \emph{connected} if it cannot be divided into two disjoint sub-schemes that have no symbol in common; \item \label{prop:flat} A scheme is \emph{flat} if each symbol appears at most twice in the scheme; \item \label{prop:closed} A scheme is \emph{closed} if each symbol appears at least twice in the scheme; \item \label{prop:linear} A scheme is \emph{linear} if each word that contains exactly two symbols that appear multiple times in the scheme has them in antipodal positions; \item \label{prop:chemical} A scheme is \emph{chemical} if, whenever $ab$ appears in the scheme such that $a$ and $b$ both have multiple appearance, then there exists a literal $c$ (different from $a$ and $b$) such that $b^-c$ (or, alternatively, $ c^-b $) and $c^-a $ (or, alternatively, $ a^-c $) appear in the same scheme; \item \label{prop:catacondensed} A scheme is \emph{catacondensed} if, whenever $ ab $ appears in the scheme, then at least one of the symbols $a$ and $b$ appears only once; \item \label{prop:unbranched} A scheme is \emph{unbranched} if every word of the scheme contains at most two symbols that appear more than once in the scheme and if there are two in a given word, they are non-adjacent. A catacondensed scheme is called \emph{branched} whenever it is not unbranched; \item \label{prop:hexagonal} A scheme is \emph{hexagonal} if each word contains six literals; \item \label{prop:oriented} A scheme is \emph{oriented} if no literal appears in it twice (i.e.\ no symbol appears twice with the same exponent). It is \emph{orientable} if it is equivalent to an oriented scheme. A scheme that is not orientable is \emph{nonorientable}. \end{enumerate} Note that \ref{prop:linear} implies \ref{prop:unbranched} and \ref{prop:unbranched} implies \ref{prop:catacondensed}. Also, every orientable scheme is flat. All properties \ref{prop:connected} -- \ref{prop:hexagonal} are preserved under the aforementioned transformations \ref{T:1} -- \ref{T:5} and hence also apply to polygonal complexes. The property \ref{prop:chemical} is equivalent to requiring that the skeleton graph is a chemical graph (i.e.\ has maximum degree less than or equal to $3$). The property `oriented' \ref{prop:oriented} is not preserved under \ref{T:5}, though it is still preserved under \ref{T:1} -- \ref{T:4}. However, properties `orientable' and `nonorientable' are preserved under \ref{T:1} -- \ref{T:5}. We know that a connected flat scheme represents a compact surface with a boundary. The boundary is determined by symbols that appear only once in the scheme. If a connected flat scheme is also closed then the surface itself is closed, i.e.\ it has no boundary. With these definitions, a fullerene is a case of a closed chemical complex that is not hexagonal, since it has $12$ pentagonal faces \cite{atlas2007}. \begin{example} \label{exam:1} A typical example of a polygonal complex in the sense of Schulte \emph{et al.}~\cite{PeSc2010,PeSc2013,PeSc2014,ScWe2017} is a $2$-dimensional skeleton of the tesseract (the $4$-dimensional cube). This skeleton is composed of $16$ vertices, $32$ edges and $24$ quadrilateral faces. \begin{figure}[!ht] \centering \includegraphics[scale=0.4]{figs/tesseract_labeled.pdf} \caption{The tesseract showing edge labelling as in the scheme presented in Example \ref{exam:1}.} \label{fig:tesseract} \end{figure} The eight facets of the tesseract (which are all cubes) are discarded. A scheme describing the skeleton is given here (split into two columns for convenience): \begin{align} & \begin{matrix*}[l] a & f & i^- & c^- \\ a & e & k^- & d^- \\ a & g & u^- & b^- \\ b & v & j^- & c^- \\ b & w & m^- & d^- \\ c & h & l^- & d^- \\ e & n & r & f^- \\ e & o & x^- & g^- \\ f & t & z^- & g^- \\ h & p & r & i^- \\ h & q & 4 & j^- \\ i & t & 6^- & j^- \end{matrix*} & & \begin{matrix*}[l] k & n & p^- & l^- \\ k & o & y^- & m^- \\ l & q & 1^- & m^- \\ n & s & 2^- & o^- \\ p & s & 3^- & q^- \\ r & t & 5^- & s^- \\ u & x & y^- & w^- \\ u & z & 6^- & v^- \\ v & 4^- & 1^- & w^- \\ z & 5^- & 2^- & x^- \\ 3 & 5 & 6^- & 4^- \\ 1 & 3 & 2^- & y^- \end{matrix*} \end{align} As we can see, each symbol appears three times, because each edge lies on the boundary of three quadrilaterals. Since the scheme is not flat it is nonorientable. Note that the $1$-skeleton, i.e. the skeleton graph of the tesseract, is the $4$-hypercube graph, $Q_4$. \hfill $\Diamond$ \end{example} From now on, we will only consider flat polygonal complexes. Originally the term polygonal complex was reserved for flat polygonal complexes. See for instance chapter by Pisanski and Potočnik in \cite{PiPo2004}. More information about maps can be obtained from \cite{GrYe2004}. The following example shows how one can distinguish between a tetrahedron and a tetrahedron with one face missing. \begin{figure}[!htb] \centering \begin{tikzpicture} \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw,decoration={markings,mark=at position 0.55 with {\arrow{Stealth[scale=1.5]}}},postaction={decorate}] \node[label=-90:$A$] (a) at (0, 0) {}; \node[label=90:$D$] (d) at (90:2) {}; \node[label=0:$C$] (c) at (-30:2) {}; \node[label=180:$B$] (b) at (210:2) {}; \path[edge] (b) -- (c); \path[edge] (c) -- (d); \path[edge] (d) -- (b); \path[edge] (a) -- (d); \path[edge] (b) -- (a); \path[edge] (c) -- (a); \node[draw=none,fill=none] at (0, -1.4) {$f$}; \node[draw=none,fill=none] at (1.2, 0.6) {$d$}; \node[draw=none,fill=none] at (-1.2, 0.6) {$c$}; \node[draw=none,fill=none] at (-0.7, -0.15) {$a$}; \node[draw=none,fill=none] at (0.7, -0.15) {$e$}; \node[draw=none,fill=none] at (0.25, 0.8) {$b$}; \end{tikzpicture} \caption{Open and closed tetrahedra. The Schlegel diagram with the infinite face BCD included represents the closed tetrahedron. In the open tetrahedron the final triangular face $cfd$ ($BCD$) has been removed.} \label{fig:tetrahedra} \end{figure} \begin{example} Consider the scheme: \begin{equation} \Theta = \; \begin{matrix*}[l] a & b & c \\ a^- & f & e \\ b^- & e^- & f \\ f^- & c^- & d^- \end{matrix*} \end{equation} This represents a tetrahedron. There are six edges and each row corresponds to a triangular face. All symbols in $\Theta$ appear twice, and hence the corresponding surface is closed (has no boundary). The surface in this case is a sphere. By removing a face, for instance the last one, we obtain a connected scheme: \begin{equation} \Theta' = \; \begin{matrix*}[l] a & b & c \\ a^- & e & e \\ b^- & e^- & f \end{matrix*} \end{equation} that represents a tetrahedron with one face missing. The symbols $c,d,f$ each occur only once and the corresponding surface is a disk. \hfill $\Diamond$ \end{example} \begin{figure}[!htb] \centering \begin{minipage}[b]{\linewidth} \centering \begin{tikzpicture}[scale=1.0] \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw,decoration={markings,mark=at position 0.55 with {\arrow{Stealth[scale=1.5]}}},postaction={decorate}] \node[] (a) at (0, 0) {}; \node[] (b) at (1, 0) {}; \node[] (c) at (1, 1) {}; \node[] (d) at (0, 1) {}; \node[] (a2) at (-1, -1) {}; \node[] (b2) at (2, -1) {}; \node[] (c2) at (2, 2) {}; \node[] (d2) at (-1, 2) {}; \path[edge] (a) -- (b); \path[edge] (b) -- (c); \path[edge] (c) -- (d); \path[edge] (d) -- (a); \path[edge] (a2) -- (b2); \path[edge] (b2) -- (c2); \path[edge] (c2) -- (d2); \path[edge] (d2) -- (a2); \path[edge] (a2) -- (a); \path[edge] (b2) -- (b); \path[edge] (c2) -- (c); \path[edge] (d2) -- (d); \node[draw=none,fill=none] at (0.5, -1.2) {$a$}; \node[draw=none,fill=none] at (0.5, 0.2) {$e$}; \node[draw=none,fill=none] at (0.5, 1.2) {$g$}; \node[draw=none,fill=none] at (0.5, 2.2) {$c$}; \node[draw=none,fill=none] at (-1.3, 0.5) {$d$}; \node[draw=none,fill=none] at (-0.2, 0.5) {$h$}; \node[draw=none,fill=none] at (1.25, 0.5) {$f$}; \node[draw=none,fill=none] at (2.25, 0.5) {$b$}; \end{tikzpicture} \subcaption{$\Sigma$}\label{fig:threeembeddings1} \end{minipage} \begin{minipage}[b]{0.45\linewidth} \centering \begin{tikzpicture}[scale=1.3] \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw,decoration={markings,mark=at position 0.6 with {\arrow{Stealth[scale=1.5]}}},postaction={decorate}] \tikzstyle{torus1}=[draw,decoration={markings,mark=at position 0.5 with {\arrow{>[scale=1.5]}}, ,mark=at position 0.55 with {\arrow{>[scale=1.5]}}},postaction={decorate}] \tikzstyle{torus2}=[draw,decoration={markings,mark=at position 0.5 with {\arrow{>[scale=1.5]}}},postaction={decorate}] \node[] (a) at (0, 0) {}; \node[] (b) at (1, 0) {}; \node[] (c) at (1, 1) {}; \node[] (d) at (0, 1) {}; \node[] (a2) at (-1, 0) {}; \node[] (b2) at (2, 0) {}; \node[] (c2) at (2, 1) {}; \node[] (d2) at (-1, 1) {}; \path[edge] (a2) -- (a); \path[edge] (a) -- (b); \path[edge] (b) -- (b2); \path[edge] (d2) -- (d); \path[edge] (d) -- (c); \path[edge] (c) -- (c2); \path[edge] (c) -- (b); \path[edge] (d2) -- (a2); \path[edge] (-1.8, 0) -- (a2); \path[edge] (b2) -- (2.8, 0); \path[edge] (c2) -- (2.8, 1); \path[edge] (-1.8, 1) -- (d2); \path[edge] (d) -- (0, 1.7); \path[edge] (0, -0.7) -- (a); % \path[edge] (c2) -- (2, 1.7); \path[edge] (2, -0.7) -- (b2); \path[torus1] (-1.8, -0.7) -- (-1.8, 1.7); \path[torus1] (2.8, -0.7) -- (2.8, 1.7); \path[torus2] (-1.8, -0.7) -- (2.8, -0.7); \path[torus2] (-1.8, 1.7) -- (2.8, 1.7); \node[draw=none,fill=none] at (2.5, 0.2) {$h$}; \node[draw=none,fill=none] at (1.5, 0.2) {$g$}; \node[draw=none,fill=none] at (0.5, 0.2) {$f$}; \node[draw=none,fill=none] at (-0.5, 0.2) {$e$}; \node[draw=none,fill=none] at (-1.5, 0.2) {$h$}; \node[draw=none,fill=none] at (2.5, 0.8) {$d$}; \node[draw=none,fill=none] at (1.5, 0.8) {$c$}; \node[draw=none,fill=none] at (0.5, 0.8) {$b$}; \node[draw=none,fill=none] at (-0.5, 0.8) {$a$}; \node[draw=none,fill=none] at (-1.5, 0.8) {$d$}; \node[draw=none,fill=none] at (1.2, 0.5) {$k$}; \node[draw=none,fill=none] at (-0.8, 0.5) {$i$}; \node[draw=none,fill=none] at (2.2, 1.5) {$l$}; \node[draw=none,fill=none] at (0.2, 1.5) {$j$}; \node[draw=none,fill=none] at (2.2, -0.5) {$l$}; \node[draw=none,fill=none] at (0.2, -0.5) {$j$}; \end{tikzpicture} \subcaption{$\Sigma'$}\label{fig:threeembeddings2} \end{minipage} \begin{minipage}[b]{0.45\linewidth} \centering \begin{tikzpicture}[scale=1.3] \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw,decoration={markings,mark=at position 0.6 with {\arrow{Stealth[scale=1.5]}}},postaction={decorate}] \tikzstyle{torus1}=[draw,decoration={markings,mark=at position 0.5 with {\arrow{>[scale=1.5]}}, ,mark=at position 0.55 with {\arrow{>[scale=1.5]}}},postaction={decorate}] \tikzstyle{torus2}=[draw,decoration={markings,mark=at position 0.5 with {\arrow{>[scale=1.5]}}},postaction={decorate}] \node[] (a) at (0, 0) {}; \node[] (b) at (1, 0) {}; \node[] (c) at (1, 1) {}; \node[] (d) at (0, 1) {}; \node[] (a2) at (-1, 0) {}; \node[] (b2) at (2, 0) {}; \node[] (c2) at (2, 1) {}; \node[] (d2) at (-1, 1) {}; \path[edge] (a) -- (a2); \path[edge] (b) -- (a); \path[edge] (b2) -- (b); \path[edge] (d2) -- (d); \path[edge] (d) -- (c); \path[edge] (c) -- (c2); \path[edge] (c2) -- (b); \path[edge] (d) -- (a2); \path[edge] (a2) -- (-1.8, 0); \path[edge] (2.8, 0) -- (b2); \path[edge] (c2) -- (2.8, 1); \path[edge] (-1.8, 1) -- (d2); \path[edge] (d2) -- (-0.5, 1.7); \path[edge] (-0.5, -0.7) -- (a); % \path[edge] (c) -- (1.5, 1.7); \path[edge] (1.5, -0.7) -- (b2); \path[torus1] (-1.8, -0.7) -- (-1.8, 1.7); \path[torus1] (2.8, -0.7) -- (2.8, 1.7); \path[torus2] (-1.8, -0.7) -- (2.8, -0.7); \path[torus2] (-1.8, 1.7) -- (2.8, 1.7); \node[draw=none,fill=none] at (2.5, -0.2) {$e$}; \node[draw=none,fill=none] at (1.3, -0.2) {$f$}; \node[draw=none,fill=none] at (0.5, -0.2) {$g$}; \node[draw=none,fill=none] at (-0.7, -0.2) {$h$}; \node[draw=none,fill=none] at (-1.5, -0.2) {$e$}; \node[draw=none,fill=none] at (2.5, 0.8) {$c$}; \node[draw=none,fill=none] at (1.3, 0.8) {$b$}; \node[draw=none,fill=none] at (0.5, 0.8) {$a$}; \node[draw=none,fill=none] at (-0.7, 0.8) {$d$}; \node[draw=none,fill=none] at (-1.5, 0.8) {$c$}; \node[draw=none,fill=none] at (1.7, 0.5) {$k$}; \node[draw=none,fill=none] at (-0.3, 0.5) {$i$}; \node[draw=none,fill=none] at (1.6, 1.5) {$j$}; \node[draw=none,fill=none] at (-0.4, 1.5) {$l$}; \node[draw=none,fill=none] at (1.9, -0.5) {$j$}; \node[draw=none,fill=none] at (-0.1, -0.5) {$l$}; \end{tikzpicture} \subcaption{$\Sigma''$}\label{fig:threeembeddings3} \end{minipage} \caption{Three embeddings of the cube graph $Q_3$.} \label{fig:threeembeddings} \end{figure} \begin{example} Here are three maps that all have the same skeleton, namely the cube graph, $Q_3$: \begin{equation} \Sigma = \; \begin{matrix*}[l] a & b & c & d\\ h^- & g^- & f^- & e^- \\ a^- & i & e & j^- \\ j & f & k^- & b^- \\ g & l^- & c^- & k \\ d^- & l & h & i^- \end{matrix*} \end{equation} \begin{equation} \Sigma' = \; \begin{matrix} a & b & k&f^-&e^-&i^-\\ c & d &i & h^- & g^- & k^-\\ f & g & l^- & c^- & b^- & j\\ h & e & j^- & a^- & d^- & l \end{matrix} \end{equation} \begin{equation} \Sigma'' = \; \begin{matrix} a & b & k&g&h&i^-\\ c & d &i & e & f & k^-\\ l & g^- & f^- & j^- & a^- & d^- \\ b & c & l & h & e & j^- \end{matrix} \end{equation} Note that $\Sigma$ describes the usual hexahedron, i.e.\ the surface of the cube. $\Sigma'$ and $\Sigma''$ describe two non-equivalent toroidal polyhexes. All three maps share the same underlying skeleton, the cube graph $Q_3$. However, the embeddings $\Sigma'$ and $\Sigma''$ are clearly distinct. $\Sigma''$ has the property that each pair of faces intersects in exactly two edges which are antipodal in each face, whereas $\Sigma'$ does not. $\Sigma''$ is a regular map \cite{Coxeter1973,McMullen2002}, a generalisation of Platonic polyhedra. \hfill $\Diamond$ \end{example} The above example raises an interesting question: Which toroidal polyhexes are completely determined by their skeleta? \section{Equilinear catacondensed chemical hexagonal complexes} Catacondensed chemical hexagonal complexes are characterized by the following rules: they are connected \ref{prop:connected}, flat \ref{prop:flat}, hexagonal \ref{prop:hexagonal} and catacondensed \ref{prop:catacondensed}. These rules imply that such a complex is also chemical \ref{prop:chemical}. We may view such a complex as a collection of branching hexagons that are connected by chains of hexagons. If each hexagonal chain is linear (property \ref{prop:linear} holds for the complex), we say that such a structure is a \emph{linear catacondensed chemical hexagonal complex}. Moreover, if all linear hexagonal chains are of the same length, i.e.\ contain the same number of hexagons, these structures are called \emph{equilinear}. We denote by $l$ the common length of these linear chains. In the unbranched case (where \ref{prop:unbranched} holds), for a given $l$, there are only three catacondensed structures: $P_l, C_l$ and $M_l$, as illustrated in Figure~\ref{fig:thethreenonbranched}. \begin{figure}[!ht] \centering \includegraphics[scale=0.6]{figs/complexes240321.pdf} \caption{Unbranched catacondensed chemical hexagonal complexes: (a) linear polyacene $P_6$, (b) untwisted cyclacene $C_6$, and (c) M{\"o}bius cyclacene $M_6$. For the two cyclacene cases, the arrowed left and right edges are to be identified.} \label{fig:thethreenonbranched} \end{figure} In the branched case, there are three types of hexagon: \emph{branching} (attached to three hexagons, i.e.\ of type $A_3$ in the notation of \cite{gutman1989}), \emph{connecting} (attached to two hexagons, i.e.\ of type $A_2$ or $L_2$), and \emph{terminal} (attached to a single hexagon, i.e.\ of type $L_1$). Such a structure defines a labelled 1-3 map (i.e.\ vertices are of degrees $1$ and $3$), called the \emph{blueprint map}. A map is a graph together with a rotation projection \cite{Gross1987}. The underlying graph is called the \emph{blueprint graph}. In the map, each arc is labelled by a pair $(w, \sigma)$, where $w$ is a word over the alphabet $\{L, R, S\}$ and $\sigma \in \{+, -\}$. The reverse of a label $(w, \sigma)$ is the label $(w^\rho, \sigma)$, where $w^\rho$ is the reverse of the word in which symbols $L$ and $R$ are interchanged. Labels are assigned to arcs (half-edges) and two opposite arcs are assigned reverse labels. For instance, reverse of the label $(RLSR, +)$ is the label $(LSRL, +)$. Degree $3$ vertices correspond to branching hexagons, while vertices of degree $1$ correspond to terminal hexagons. The connecting hexagons are implicitly described by the labels; see Figure~\ref{fig:words}. \begin{figure}[!htbp] \centering \includegraphics[scale=0.6]{figs/wordfig.pdf} \caption{Arc labelling in the blueprint map. In the untwisted (a) and the twisted (b) the word $w$ is $RLSR$ taken in the direction from bottom to top or $LSRL$ from top to bottom, whilst $\sigma = +$ and $\sigma = -$, respectively.} \label{fig:words} \end{figure} In the equilinear case, the description given above can be simplified. Words in labels on arcs comprise only the letter $S$. Therefore each word can be described by giving its length, and then only one integer parameter is needed. Figure~\ref{fig:blue} shows rotation projections for all blueprint maps on up to two trivalent vertices. Note that in the case of zero vertices of degree $3$, two cases are anomalous, as they are free loops with no vertices at all. \begin{figure}[!htb] \centering \begin{minipage}[b]{\linewidth} \centering \subcaption{Zero vertices of degree $3$:}\label{fig:blue0} \begin{tikzpicture}[scale=0.7] \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \node (a) at (0, 1) {}; \node (b) at (0, 0) {}; \path[edge] (a) -- (b); \end{tikzpicture} $\mathcal{M}_{1,0}^{0}$ \hspace{1cm} \begin{tikzpicture}[scale=0.7] \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.5 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \path[edge] (0, -1-0.6) circle (0.6); \end{tikzpicture} $\mathcal{M}_{2,0}^{0}$ \hspace{1cm} \begin{tikzpicture}[scale=0.7] \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.5 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \path[medge] (0, -1-0.6) circle (0.6); \end{tikzpicture} $\mathcal{M}_{2,1}^{0}$ \end{minipage} \vspace{\baselineskip} \begin{minipage}[b]{\linewidth} \centering \subcaption{One vertex of degree $3$:}\label{fig:blue1} \begin{tikzpicture}[scale=0.7] \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.25 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \tikzstyle{medge2}=[draw,decoration={markings,mark=at position 0.75 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \node (a) at (0, 1) {}; \node (b) at (0, 0) {}; \node (c1) at ($ (0, 1) + (30:1) $) {}; \node (c2) at ($ (0, 1) + (150:1) $) {}; \path[edge] (a) -- (b); \path[edge] (c1) -- (a) -- (c2); \end{tikzpicture} $\mathcal{M}_{1,0}^{1}$ \hspace{1cm} \begin{tikzpicture}[scale=0.7] \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.25 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \tikzstyle{medge2}=[draw,decoration={markings,mark=at position 0.75 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \path[edge] (0, 1+0.6) circle (0.6); \node (a) at (0, 1) {}; \node (b) at (0, 0) {}; \path[edge] (a) -- (b); \end{tikzpicture} $\mathcal{M}_{2,0}^{1}$ \hspace{1cm} \begin{tikzpicture}[scale=0.7] \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.25 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \tikzstyle{medge2}=[draw,decoration={markings,mark=at position 0.75 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \path[medge] (0, 1+0.6) circle (0.6); \node (a) at (0, 1) {}; \node (b) at (0, 0) {}; \path[edge] (a) -- (b); \end{tikzpicture} $\mathcal{M}_{2,1}^{1}$ \end{minipage} \vspace{\baselineskip} \begin{minipage}[b]{\linewidth} \centering \subcaption{Two vertices of degree $3$, one connecting edge:}\label{fig:blue21} \begin{tikzpicture}[scale=0.7] \useasboundingbox (-0.7,-2.2-0.15) rectangle (0.7,+2.2+0.15); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.25 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \node (a) at (0, 1) {}; \node (b) at (0, -1) {}; \path[edge] (a) -- (b); \node (b1) at (-0.5, -2) {}; \node (b2) at (0.5, -2) {}; \path[edge] (b) -- (b1); \path[edge] (b) -- (b2); \node (a1) at (-0.5, 2) {}; \node (a2) at (0.5, 2) {}; \path[edge] (a) -- (a1); \path[edge] (a) -- (a2); \end{tikzpicture} $\mathcal{M}_{1,0}^{2}$ \hspace{0.5cm} \begin{tikzpicture}[scale=0.7] \useasboundingbox (-0.7,-2.2-0.15) rectangle (0.7,+2.2+0.15); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.25 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \path[edge] (0, 1+0.6) circle (0.6); \node (a) at (0, 1) {}; \node (b) at (0, -1) {}; \path[edge] (a) -- (b); \node (b1) at (-0.5, -2) {}; \node (b2) at (0.5, -2) {}; \path[edge] (b) -- (b1); \path[edge] (b) -- (b2); \end{tikzpicture} $\mathcal{M}_{2,0}^{2}$ \hspace{0.5cm} \begin{tikzpicture}[scale=0.7] \useasboundingbox (-0.7,-2.2-0.15) rectangle (0.7,+2.2+0.15); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.25 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \path[medge] (0, 1+0.6) circle (0.6); \node (a) at (0, 1) {}; \node (b) at (0, -1) {}; \path[edge] (a) -- (b); \node (b1) at (-0.5, -2) {}; \node (b2) at (0.5, -2) {}; \path[edge] (b) -- (b1); \path[edge] (b) -- (b2); \end{tikzpicture} $\mathcal{M}_{2,1}^{2}$ \hspace{0.5cm} \begin{tikzpicture}[scale=0.7] \useasboundingbox (-0.7,-2.2-0.15) rectangle (0.7,+2.2+0.15); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.25 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \path[edge] (0, 1+0.6) circle (0.6); \path[edge] (0, -1-0.6) circle (0.6); \node (a) at (0, 1) {}; \node (b) at (0, -1) {}; \path[edge] (a) -- (b); \end{tikzpicture} $\mathcal{M}_{3,0}^{2}$ \hspace{0.5cm} \begin{tikzpicture}[scale=0.7,] \useasboundingbox (-0.7,-2.2-0.15) rectangle (0.7,+2.2+0.15); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.25 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \path[medge] (0, 1+0.6) circle (0.6); \path[edge] (0, -1-0.6) circle (0.6); \node (a) at (0, 1) {}; \node (b) at (0, -1) {}; \path[edge] (a) -- (b); \end{tikzpicture} $\mathcal{M}_{3,1}^{2}$ \hspace{0.5cm} \begin{tikzpicture}[scale=0.7] \useasboundingbox (-0.7,-2.2-0.15) rectangle (0.7,+2.2+0.15); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.25 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \tikzstyle{medge2}=[draw,decoration={markings,mark=at position 0.75 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \path[medge] (0, 1+0.6) circle (0.6); \path[medge2] (0, -1-0.6) circle (0.6); \node (a) at (0, 1) {}; \node (b) at (0, -1) {}; \path[edge] (a) -- (b); \end{tikzpicture} $\mathcal{M}_{3,2}^{2}$ \end{minipage} \vspace{\baselineskip} \begin{minipage}[b]{\linewidth} \centering \subcaption{Two vertices of degree $3$, two connecting edges:}\label{fig:blue22} \begin{tikzpicture}[scale=0.7] \useasboundingbox (-0.7,-2-0.15) rectangle (0.7,+2+0.15); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.5 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \node (a) at (0, 1) {}; \node (a2) at (0, 2) {}; \node (b) at (0, -1) {}; \node (b2) at (0, -2) {}; \path[edge] (a) -- (a2); \path[edge] (b) -- (b2); \path[edge] (a) .. controls (-0.7, 0.5) and (-0.7, -0.5) .. (b); \path[edge] (a) .. controls (0.7, 0.5) and (0.7, -0.5) .. (b); \end{tikzpicture} $\mathcal{M}_{4,0}^{2}$ \hspace{1cm} \begin{tikzpicture}[scale=0.7] \useasboundingbox (-0.7,-2-0.15) rectangle (0.7,+2+0.15); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.5 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \node (a) at (0, 1) {}; \node (a2) at (0, 2) {}; \node (b) at (0, -1) {}; \node (b2) at (0, -2) {}; \path[edge] (a) -- (a2); \path[edge] (b) -- (b2); \path[medge] (a) .. controls (-0.7, 0.5) and (-0.7, -0.5) .. (b); \path[edge] (a) .. controls (0.7, 0.5) and (0.7, -0.5) .. (b); \end{tikzpicture} $\mathcal{M}_{4,1}^{2}$ \hspace{1cm} \begin{tikzpicture}[scale=0.7] \useasboundingbox (-0.7,-2-0.15) rectangle (0.7,+2+0.15); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.5 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \node (a) at (0, 1) {}; \node (a2) at (0, 2) {}; \node (b) at (0, -1) {}; \node (b2) at (0, 0) {}; \path[edge] (a) -- (a2); \path[edge] (b) -- (b2); \path[edge] (a) .. controls (-0.7, 0.5) and (-0.7, -0.5) .. (b); \path[edge] (a) .. controls (0.7, 0.5) and (0.7, -0.5) .. (b); \end{tikzpicture} $\mathcal{M}_{5,0}^{2}$ \hspace{1cm} \begin{tikzpicture}[scale=0.7] \useasboundingbox (-0.7,-2-0.15) rectangle (0.7,+2+0.15); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.5 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \node (a) at (0, 1) {}; \node (a2) at (0, 2) {}; \node (b) at (0, -1) {}; \node (b2) at (0, 0) {}; \path[edge] (a) -- (a2); \path[edge] (b) -- (b2); \path[medge] (a) .. controls (-0.7, 0.5) and (-0.7, -0.5) .. (b); \path[edge] (a) .. controls (0.7, 0.5) and (0.7, -0.5) .. (b); \end{tikzpicture} $\mathcal{M}_{5,1}^{2}$ \end{minipage} \vspace{\baselineskip} \begin{minipage}[b]{\linewidth} \centering \subcaption{Two vertices of degree $3$, three connecting edges:}\label{fig:blue22} \begin{tikzpicture}[scale=1] \useasboundingbox (-0.7,-1-0.15) rectangle (0.7,+1+0.15); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.6 with {\arrow{Stealth[scale=1.5]}}},postaction={decorate}] \node (a) at (0, 1) {}; \node (b) at (0, -1) {}; \path[edge] (a) -- (b); \path[edge] (a) .. controls (-0.7, 0.5) and (-0.7, -0.5) .. (b); \path[edge] (a) .. controls (0.7, 0.5) and (0.7, -0.5) .. (b); \end{tikzpicture} $\mathcal{M}_{6,0}^{2}$ \hspace{1cm} \begin{tikzpicture}[scale=1] \useasboundingbox (-0.7,-1-0.15) rectangle (0.7,+1+0.15); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.5 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \node (a) at (0, 1) {}; \node (b) at (0, -1) {}; \path[edge] (a) -- (b); \path[medge] (a) .. controls (-0.7, 0.5) and (-0.7, -0.5) .. (b); \path[edge] (a) .. controls (0.7, 0.5) and (0.7, -0.5) .. (b); \end{tikzpicture} $\mathcal{M}_{6,1}^{2}$ \hspace{1cm} \begin{tikzpicture}[scale=1] \useasboundingbox (-0.7,-1-0.15) rectangle (0.7,+1+0.15); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.5 with {\draw(1, 1) -- (-1, -1);}},postaction={decorate}] \node (a) at (0, 1) {}; \node (b) at (0, -1) {}; \path[edge] (a) .. controls (-0.7, 0.5) and (-0.7, -0.5) .. (b); \path[name path=p1] (a) .. controls (1.2, 0) and (0.0, 0.2) .. (b); \path[name path=p2] (b) .. controls (1.2, 0) and (0.0, -0.2) .. (a); \path [name intersections={of=p1 and p2,by=x}]; \path[edge] (a) .. controls (1.2, 0) and (0.0, 0.2) .. (b); \node[circle,fill=white,draw=none,inner sep=2pt] at (x) {}; \path[edge] (b) .. controls (1.2, 0) and (0.0, -0.2) .. (a); \end{tikzpicture} $\mathcal{M}_{7,0}^{2}$ \hspace{1cm} \begin{tikzpicture}[scale=1] \useasboundingbox (-0.7,-1-0.15) rectangle (0.7,+1+0.15); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.5 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \node (a) at (0, 1) {}; \node (b) at (0, -1) {}; \path[medge] (a) .. controls (-0.7, 0.5) and (-0.7, -0.5) .. (b); \path[name path=p1] (a) .. controls (1.2, 0) and (0.0, 0.2) .. (b); \path[name path=p2] (b) .. controls (1.2, 0) and (0.0, -0.2) .. (a); \path [name intersections={of=p1 and p2,by=x}]; \path[edge] (a) .. controls (1.2, 0) and (0.0, 0.2) .. (b); \node[circle,fill=white,draw=none,inner sep=2pt] at (x) {}; \path[edge] (b) .. controls (1.2, 0) and (0.0, -0.2) .. (a); \end{tikzpicture} $\mathcal{M}_{7,1}^{2}$ \end{minipage} \caption{The blueprint maps with $n \leq 2$ trivalent vertices. The symbol \tikzCross on an edge represents a half-twist. In maps $\mathcal{M}_{7,0}^{2}$ and $\mathcal{M}_{7,1}^{2}$ the crossing edges are necessary, because the order of edges around a vertex is significant and cannot be represented in a planar drawing. } \label{fig:blue} \end{figure} \section{Kekul\'{e} structures in catacondensed coronoid complexes} A natural question to ask is which flat hexagonal complexes admit a Kekul\'{e} structure. It is well known that all catacondensed benzenoids are Kekulean \cite{gutman1989}. \begin{theorem} Any catacondensed flat hexagonal complex $B$ is Kekulean. \end{theorem} \noindent Note that any catacondensed flat hexagonal complex is also a chemical hexagonal complex. \begin{proof} The proof is constructive -- we construct a perfect matching $\mathcal{M}$. A catacondensed flat hexagonal complex contains only hexagons of type $L_1$, $L_2$, $A_2$ and $A_3$ (see \cite[p.~21]{gutman1989}). In the first step we remove all hexagons of type $A_3$ from $B$. Edges labelled $a, b$ and $c$ in Figure~\ref{fig:types} will be single bonds (they are not in the matching $\mathcal{M}$). By removing a hexagon of type $A_3$ we mean deleting edges $a$, $b$ and $c$. \begin{figure}[!h] \centering \includegraphics[scale=0.6]{figs/hexagons.pdf} \caption{Types of hexagons in catafused flat hexagonal complex $B$.} \label{fig:types} \end{figure} In the second step we remove all hexagons of type $A_2$ from $B$. Edges labelled $a, b$ and $c$ in Figure~\ref{fig:types} will be single bonds, whilst the edge $d$ will be double (we add it to $M$). \begin{figure}[!h] \centering \includegraphics[scale=0.6]{figs/complex.pdf} \caption{The complex $B$ after deletion of hexagons of types $A_3$ and $A_2$.} \label{fig:leftovers} \end{figure} What remains is a disjoint union of $k$ linear polyacene chains $P^i$ (see Figure~\ref{fig:leftovers}(a)) which may be of different lengths, $p$ untwisted cyclacenes $C^i$ (see Figure~\ref{fig:leftovers}(a)), $q$ twisted cyclacenes $T^i$ (see Figure~\ref{fig:leftovers}(b)), and $m$ isolated $K_2$ fragments: $$ \{P^1, P^2, \ldots, P^k\} \cup \{C^1, C^2, \ldots, C^p \} \cup \{ M^1, M^2, \ldots, M^q \} \cup \{ K_2^1, K_2^2, \ldots, K_2^r \}. $$ We add all isolated $K_2$ fragments to the perfect matching $\mathcal{M}$. All linear chains and cyclacenes are Kekulean. For each chain we may pick any of $l_i + 1$ perfect matchings. An untwisted cyclacene has $4$ perfect matchings, and a twisted cyclacene has $2$ perfect matchings, hence $ K(B) \geq 4^p 2^q \prod_{i=1}^{k} (l_i + 1). $ \end{proof} \begin{figure}[!b] \centering \includegraphics[width=0.9\textwidth]{figs/f4.pdf} \caption{Basic Rules for perfect matchings of hexagonal complexes derived with linear polyacene strips (illustrated for strips of length $3$). They are: (a) the \textit{linear forcing} rule; (b) the {\it crossover} rule; (c) the {\it pairing rule}. Hexagons A and B are derived from cubic vertices. Fixing the illustrated {\it endo} bonds of hexagon A (denoted $d$ or $s$ for double or single) either forces (rules (a) and (b)) or rules out ((c)) given pairings of the corresponding {\it endo} bonds in hexagon B. The panels show (left) the untwisted chain and (right) the chain with a M{\"o}bius half-twist.} \label{fig:PWF4} \end{figure} This reasoning can be used as the basis of a simple procedure for counting the perfect matchings for a flat hexagonal complex with a given rotation scheme. For the linear polyacene motif, this leads to polynomial functions in $l + 1$. The detailed form of these expressions, the powers that appear and the coefficients that multiply them can be rationalised by thinking about a set of forcing rules for replacements of edges of the cubic graph by linear chains of polyacenes. In this case, three rules apply to allowed combinations of pairs of edges in the two branching hexagons (see Figure~\ref{fig:PWF4}). Rule (a) is the \emph{linear forcing} rule, by which two double bonds in A \emph{force} a fixed matching in the chain and two single bonds in B. Rule (b) is the \emph{crossover rule}, by which a single/double pair in A \emph{forces} a fixed matching in the chain and a double/single pair in B. Rule (c) is the \emph{pairing rule}, by which a pair of single bonds in A is \emph{compatible with} either a pair of single bonds or a pair of double bonds in B. The pair of single bonds in B results from taking any of the $(l+1)$ perfect matchings of the intervening hexagons. The pair of double bonds of B arises from reversal of the linear forcing rule. (A single/double pair in B is ruled out by the crossover rule.) Note that if we make a complex from a cubic graph with $m$ edges by using a straight chain of length $l$ on every edge, there is a term $(l+1)^m$ in the Kekul\'{e} count. Kinks in the chains will increase this leading term \cite{Balaban1989,Fowler2014}. E.g.\ fibonacene chains of length $l$ would lead to a term $(F_{l+2})^m$, where $F_{l+2}$ is the $(l+2)$-th Fibonacci number. Fibonacene chains also allow favourable perfect matchings in which there are many hexagonal rings containing three double bonds, thus conforming to classical models of stability based on the ideas of Fries \cite{Fries1927} and Clar \cite{Clar1972}; see Figure~\ref{fig:PWF7}. \begin{figure}[!htbp] \centering \includegraphics[scale=0.6]{figs/fries.pdf} \caption{A fully Fries hexagonal complex, i.e.\ one in which every hexagonal face includes three matched edges, can be constructed by using either an odd or or an even zig-zag fibonacene to inflate all edges of a cubic graph. Several attachment isomers are possible: for example, a fully Fries attachment isomer could be built using any external double bond in each terminal hexagon. } \label{fig:PWF7} \end{figure} Some explicit formulas for Kekul{\' e} counts of complexes built from cubic graphs and linear polyacenes are: \begin{figure}[!htb] \begin{tikzpicture}[scale=1] \useasboundingbox (-1.5,-0.9) rectangle (1.5,1.6); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.5 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \node (s) at (0, 0) {}; \node (b) at (90:1.5) {}; \node (a) at (-30:1.5) {}; \node (c) at (210:1.5) {}; \path[edge] (a) -- (s); \path[edge] (a) -- (c); \path[edge] (s) -- (b); \path[edge] (s) -- (c); \path[edge] (a) -- (b); \path[edge] (b) -- (c); \end{tikzpicture}$\mathcal{T}_{0,0}$ \begin{tikzpicture}[scale=1] \useasboundingbox (-1.5,-0.9) rectangle (1.5,1.6); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.5 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \node (s) at (0, 0) {}; \node (b) at (90:1.5) {}; \node (a) at (-30:1.5) {}; \node (c) at (210:1.5) {}; \path[edge] (a) -- (s); \path[edge] (a) -- (c); \path[edge] (s) -- (b); \path[edge] (s) -- (c); \path[edge] (a) -- (b); \path[medge] (b) -- (c); \end{tikzpicture}$\mathcal{T}_{0,1}$ \begin{tikzpicture}[scale=1] \useasboundingbox (-1.5,-0.9) rectangle (1.5,1.6); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.5 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \node (s) at (0, 0) {}; \node (b) at (90:1.5) {}; \node (a) at (-30:1.5) {}; \node (c) at (210:1.5) {}; \path[edge] (a) -- (s); \path[edge] (a) -- (c); \path[edge] (s) -- (b); \path[edge] (s) -- (c); \path[medge] (a) -- (b); \path[medge] (b) -- (c); \end{tikzpicture}$\mathcal{T}_{0,2}$ \begin{tikzpicture}[scale=1] \useasboundingbox (-1.5,-0.9) rectangle (1.5,1.6); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.5 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \node (s) at (0, 0) {}; \node (b) at (90:1.5) {}; \node (a) at (-30:1.5) {}; \node (c) at (210:1.5) {}; \path[edge] (a) -- (s); \path[medge] (a) -- (c); \path[edge] (s) -- (b); \path[edge] (s) -- (c); \path[medge] (a) -- (b); \path[medge] (b) -- (c); \end{tikzpicture}$\mathcal{T}_{0,3}$ \vspace{0.5\baselineskip} \begin{tikzpicture}[scale=1] \useasboundingbox (-1.5,-1.3) rectangle (1.5,1.6); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.5 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \node (s) at (0, 0) {}; \node (b) at (90:1.5) {}; \node (a) at (-30:1.5) {}; \node (c) at (210:1.5) {}; \path[name path=p1] (a) -- (c); \path[name path=p2] (s) .. controls ($ (s) + (-30:0.5) $) and ($ (a) + (230:1.5) $) .. (a); \path [name intersections={of=p1 and p2,by=x}]; \path[draw] (s) .. controls ($ (s) + (-30:0.5) $) and ($ (a) + (230:1.5) $) .. (a); \node[circle,fill=white,draw=none,inner sep=2pt] at (x) {}; \path[edge] (a) -- (c); \path[edge] (s) -- (b); \path[edge] (s) -- (c); \path[edge] (a) -- (b); \path[edge] (b) -- (c); \end{tikzpicture}$\mathcal{T}_{1,0}$ \begin{tikzpicture}[scale=1] \useasboundingbox (-1.5,-1.3) rectangle (1.5,1.6); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.5 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \node (s) at (0, 0) {}; \node (b) at (90:1.5) {}; \node (a) at (-30:1.5) {}; \node (c) at (210:1.5) {}; \path[name path=p1] (a) -- (c); \path[name path=p2] (s) .. controls ($ (s) + (-30:0.5) $) and ($ (a) + (230:1.5) $) .. (a); \path [name intersections={of=p1 and p2,by=x}]; \path[draw] (s) .. controls ($ (s) + (-30:0.5) $) and ($ (a) + (230:1.5) $) .. (a); \node[circle,fill=white,draw=none,inner sep=2pt] at (x) {}; \path[edge] (a) -- (c); \path[edge] (s) -- (b); \path[edge] (s) -- (c); \path[edge] (a) -- (b); \path[medge] (b) -- (c); \end{tikzpicture}$\mathcal{T}_{1,1}$ \begin{tikzpicture}[scale=1] \useasboundingbox (-1.5,-1.3) rectangle (1.5,1.6); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.5 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \node (s) at (0, 0) {}; \node (b) at (90:1.5) {}; \node (a) at (-30:1.5) {}; \node (c) at (210:1.5) {}; \path[name path=p1] (a) -- (c); \path[name path=p2] (s) .. controls ($ (s) + (-30:0.5) $) and ($ (a) + (230:1.5) $) .. (a); \path [name intersections={of=p1 and p2,by=x}]; \path[draw] (s) .. controls ($ (s) + (-30:0.5) $) and ($ (a) + (230:1.5) $) .. (a); \node[circle,fill=white,draw=none,inner sep=2pt] at (x) {}; \path[edge] (a) -- (c); \path[edge] (s) -- (b); \path[edge] (s) -- (c); \path[medge] (a) -- (b); \path[medge] (b) -- (c); \end{tikzpicture}$\mathcal{T}_{1,2}$ \begin{tikzpicture}[scale=1] \useasboundingbox (-1.5,-1.3) rectangle (1.5,1.6); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.5 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \node (s) at (0, 0) {}; \node (b) at (90:1.5) {}; \node (a) at (-30:1.5) {}; \node (c) at (210:1.5) {}; \path[name path=p1] (a) -- (c); \path[name path=p2] (s) .. controls ($ (s) + (-30:0.5) $) and ($ (a) + (230:1.5) $) .. (a); \path [name intersections={of=p1 and p2,by=x}]; \path[draw] (s) .. controls ($ (s) + (-30:0.5) $) and ($ (a) + (230:1.5) $) .. (a); \node[circle,fill=white,draw=none,inner sep=2pt] at (x) {}; \path[medge] (a) -- (c); \path[edge] (s) -- (b); \path[edge] (s) -- (c); \path[medge] (a) -- (b); \path[medge] (b) -- (c); \end{tikzpicture}$\mathcal{T}_{1,3}$ \vspace{0.5\baselineskip} \begin{tikzpicture}[scale=1] \useasboundingbox (-1.5,-1.3) rectangle (1.5,1.6); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.5 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \node (s) at (0, 0) {}; \node (b) at (90:1.5) {}; \node (a) at (-30:1.5) {}; \node (c) at (210:1.5) {}; \path[name path=p1] (a) -- (c); \path[name path=p2] (s) .. controls ($ (s) + (-30:0.5) $) and ($ (a) + (230:1.5) $) .. (a); \path [name intersections={of=p1 and p2,by=x}]; \path[draw] (s) .. controls ($ (s) + (-30:0.5) $) and ($ (a) + (230:1.5) $) .. (a); \node[circle,fill=white,draw=none,inner sep=2pt] at (x) {}; \path[edge] (a) -- (c); \path[name path=pa1] (a) -- (b); \path[name path=pa2] (s) .. controls ($ (s) + (90:0.5) $) and ($ (b) + (-10:1.5) $) .. (b); \path [name intersections={of=pa1 and pa2,by=xa}]; \path[draw] (s) .. controls ($ (s) + (90:0.5) $) and ($ (b) + (-10:1.5) $) .. (b); \node[circle,fill=white,draw=none,inner sep=2pt] at (xa) {}; \path[draw] (a) -- (b); \path[edge] (s) -- (c); \path[edge] (b) -- (c); \end{tikzpicture}$\mathcal{T}_{2,0}$ \begin{tikzpicture}[scale=1] \useasboundingbox (-1.5,-1.3) rectangle (1.5,1.6); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.5 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \node (s) at (0, 0) {}; \node (b) at (90:1.5) {}; \node (a) at (-30:1.5) {}; \node (c) at (210:1.5) {}; \path[name path=p1] (a) -- (c); \path[name path=p2] (s) .. controls ($ (s) + (-30:0.5) $) and ($ (a) + (230:1.5) $) .. (a); \path [name intersections={of=p1 and p2,by=x}]; \path[draw] (s) .. controls ($ (s) + (-30:0.5) $) and ($ (a) + (230:1.5) $) .. (a); \node[circle,fill=white,draw=none,inner sep=2pt] at (x) {}; \path[edge] (a) -- (c); \path[name path=pa1] (a) -- (b); \path[name path=pa2] (s) .. controls ($ (s) + (90:0.5) $) and ($ (b) + (-10:1.5) $) .. (b); \path [name intersections={of=pa1 and pa2,by=xa}]; \path[draw] (s) .. controls ($ (s) + (90:0.5) $) and ($ (b) + (-10:1.5) $) .. (b); \node[circle,fill=white,draw=none,inner sep=2pt] at (xa) {}; \path[draw] (a) -- (b); \path[edge] (s) -- (c); \path[medge] (b) -- (c); \end{tikzpicture}$\mathcal{T}_{2,1}$ \begin{tikzpicture}[scale=1] \useasboundingbox (-1.5,-1.3) rectangle (1.5,1.6); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.5 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \node (s) at (0, 0) {}; \node (b) at (90:1.5) {}; \node (a) at (-30:1.5) {}; \node (c) at (210:1.5) {}; \path[name path=p1] (a) -- (c); \path[name path=p2] (s) .. controls ($ (s) + (-30:0.5) $) and ($ (a) + (230:1.5) $) .. (a); \path [name intersections={of=p1 and p2,by=x}]; \path[draw] (s) .. controls ($ (s) + (-30:0.5) $) and ($ (a) + (230:1.5) $) .. (a); \node[circle,fill=white,draw=none,inner sep=2pt] at (x) {}; \path[edge] (a) -- (c); \path[name path=pa1] (a) -- (b); \path[name path=pa2] (s) .. controls ($ (s) + (90:0.5) $) and ($ (b) + (-10:1.5) $) .. (b); \path [name intersections={of=pa1 and pa2,by=xa}]; \path[draw] (s) .. controls ($ (s) + (90:0.5) $) and ($ (b) + (-10:1.5) $) .. (b); \node[circle,fill=white,draw=none,inner sep=2pt] at (xa) {}; \path[medge] (a) -- (b); \path[edge] (s) -- (c); \path[medge] (b) -- (c); \end{tikzpicture}$\mathcal{T}_{2,2}$ \begin{tikzpicture}[scale=1] \useasboundingbox (-1.5,-1.3) rectangle (1.5,1.6); \tikzstyle{every node}=[draw,circle,inner sep=1.6pt,fill=green!60!black] \tikzstyle{edge}=[draw] \tikzstyle{medge}=[draw,decoration={markings,mark=at position 0.5 with {\draw(.15, .15) -- (-.15, -.15);\draw(-.15, .15) -- (.15, -.15);}},postaction={decorate}] \node (s) at (0, 0) {}; \node (b) at (90:1.5) {}; \node (a) at (-30:1.5) {}; \node (c) at (210:1.5) {}; \path[name path=p1] (a) -- (c); \path[name path=p2] (s) .. controls ($ (s) + (-30:0.5) $) and ($ (a) + (230:1.5) $) .. (a); \path [name intersections={of=p1 and p2,by=x}]; \path[draw] (s) .. controls ($ (s) + (-30:0.5) $) and ($ (a) + (230:1.5) $) .. (a); \node[circle,fill=white,draw=none,inner sep=2pt] at (x) {}; \path[medge] (a) -- (c); \path[name path=pa1] (a) -- (b); \path[name path=pa2] (s) .. controls ($ (s) + (90:0.5) $) and ($ (b) + (-10:1.5) $) .. (b); \path [name intersections={of=pa1 and pa2,by=xa}]; \path[draw] (s) .. controls ($ (s) + (90:0.5) $) and ($ (b) + (-10:1.5) $) .. (b); \node[circle,fill=white,draw=none,inner sep=2pt] at (xa) {}; \path[medge] (a) -- (b); \path[edge] (s) -- (c); \path[medge] (b) -- (c); \end{tikzpicture}$\mathcal{T}_{2,3}$ \caption{Maps derived from the embeddings of the graph $K_4$, with and without twists. Only $9$ of the $12$ drawings shown here are distinct, as $(\mathcal{T}_{0,3}, \mathcal{T}_{2,3})$, $(\mathcal{T}_{1,1}, \mathcal{T}_{2,2})$, and $(\mathcal{T}_{1,2}, \mathcal{T}_{2,1})$ are isomorphic pairs. Conventions for edges as in Figure~\ref{fig:blue}. } \end{figure} \begin{enumerate}[label=(\roman*)] \item For the theta graph for all distinct embeddings and sets of twists: \begin{align} K(\mathcal{M}_{6,0}^{2}; l) & = {\left(l + 1\right)}^{3} + 8 \\ K(\mathcal{M}_{6,1}^{2}; l) & = {\left(l + 1\right)}^{3} + 4 \\ K(\mathcal{M}_{7,0}^{2}; l) & = {\left(l + 1\right)}^{3} + 3 \, (l + 1) + 2 \\ K(\mathcal{M}_{7,1}^{2}; l) & = {\left(l + 1\right)}^{3} + (l + 1) + 2 \end{align} \item For the tetrahedron for all distinct embeddings and sets of twists: \begin{align} K(\mathcal{T}_{0,0}; l) & = {\left(l + 1\right)}^{6} + 4 \, {\left(l + 1\right)}^{3} + 3 \, {\left(l + 1\right)}^{2} \\ K(\mathcal{T}_{0,1}; l) & = {\left(l + 1\right)}^{6} + 4 \, {\left(l + 1\right)}^{3} + 3 \, {\left(l + 1\right)}^{2} + 2 \, (l + 1) \\ K(\mathcal{T}_{0,2}; l) & = {\left(l + 1\right)}^{6} + 4 \, {\left(l + 1\right)}^{3} + 2 \, {\left(l + 1\right)}^{2} + 3 \, (l + 1) \\ K(\mathcal{T}_{0,3}; l) = K(\mathcal{T}_{2,3}; l) & = {\left(l + 1\right)}^{6} + 5 \, {\left(l + 1\right)}^{3} + 3 \, {\left(l + 1\right)}^{2} + 6 \, (l + 1) + 2 \\ K(\mathcal{T}_{1,0}; l) & = {\left(l + 1\right)}^{6} + 4 \, {\left(l + 1\right)}^{3} + 3 \, (l + 1) \\ K(\mathcal{T}_{1,1}; l) = K(\mathcal{T}_{2,2}; l) & = {\left(l + 1\right)}^{6} + 5 \, {\left(l + 1\right)}^{3} + {\left(l + 1\right)}^{2} + 2 \, (l + 1) + 2 \\ K(\mathcal{T}_{1,2}; l) = K(\mathcal{T}_{2,1}; l) & = {\left(l + 1\right)}^{6} + 4 \, {\left(l + 1\right)}^{3} + 2 \, {\left(l + 1\right)}^{2} + 2 \, (l + 1) + 2 \\ K(\mathcal{T}_{1,3}; l) & = {\left(l + 1\right)}^{6} + 4 \, {\left(l + 1\right)}^{3} + 5 \, {\left(l + 1\right)}^{2} + 4 \, (l + 1) + 4 \\ K(\mathcal{T}_{2,0}; l) & = {\left(l + 1\right)}^{6} + 4 \, {\left(l + 1\right)}^{3} + 4 \end{align} \item For the cube in the usual embedding, $\Sigma$ (see Figure~\ref{fig:threeembeddings}(a)), on the sphere, with no twist and one twisted edge, and for the untwisted toroidal embeddings, $\Sigma'$ and $\Sigma''$ (see Figures~\ref{fig:threeembeddings}(b) and \ref{fig:threeembeddings}(c)), respectively: \begin{align} K(\Sigma_0; l) & = {\left(l + 1\right)}^{12} + 8 \, {\left(l + 1\right)}^{9} + 32 \, {\left(l + 1\right)}^{6} + 64 \, {\left(l + 1\right)}^{3} + 64 \\ K(\Sigma_1; l) & = {\left(l + 1\right)}^{12} + 8 \, {\left(l + 1\right)}^{9} + 26 \, {\left(l + 1\right)}^{6} + 40 \, {\left(l + 1\right)}^{3} + 32 \\ K(\Sigma'_0; l) & = {\left(l + 1\right)}^{12} + 8 \, {\left(l + 1\right)}^{9} + 2 \, {\left(l + 1\right)}^{8} + 24 \, {\left(l + 1\right)}^{6} + 8 \, {\left(l + 1\right)}^{5} \\ & \quad {} + 17 \, {\left(l + 1\right)}^{4} + 32 \, {\left(l + 1\right)}^{3} + 8 \, {\left(l + 1\right)}^{2} + 16 \notag \\ K(\Sigma''_0; l) & = {\left(l + 1\right)}^{12} + 8 \, {\left(l + 1\right)}^{9} + 6 \, {\left(l + 1\right)}^{8} + 16 \, {\left(l + 1\right)}^{6} + 24 \, {\left(l + 1\right)}^{5} \\ & \quad {} + 9 \, {\left(l + 1\right)}^{4} + 16 \, {\left(l + 1\right)}^{3} + 24 \, {\left(l + 1\right)}^{2} + 16 \notag \end{align} \item For the dodecahedron on the sphere, with no twist and one twisted edge, respectively: \begin{align} K(\mathcal{D}_0; l) & = {\left(l + 1\right)}^{30} + 20 \, {\left(l + 1\right)}^{27} + 160 \, {\left(l + 1\right)}^{24} + 660 \, {\left(l + 1\right)}^{21} + 36 \, {\left(l + 1\right)}^{20} \\ & \quad {} + 1510 \, {\left(l + 1\right)}^{18} + 360 \, {\left(l + 1\right)}^{17} + 1972 \, {\left(l + 1\right)}^{15} + 1260 \, {\left(l + 1\right)}^{14} \notag \\ & \quad {} + 120 \, {\left(l + 1\right)}^{13} + 1560 \, {\left(l + 1\right)}^{12} + 1800 \, {\left(l + 1\right)}^{11} + 636 \, {\left(l + 1\right)}^{10} \notag \\ & \quad {} + 660 \, {\left(l + 1\right)}^{9} + 1020 \, {\left(l + 1\right)}^{8} + 600 \, {\left(l + 1\right)}^{7} + 125 \, {\left(l + 1\right)}^{6} \notag \\ K(\mathcal{D}_1; l) & = {\left(l + 1\right)}^{30} + 20 \, {\left(l + 1\right)}^{27} + 160 \, {\left(l + 1\right)}^{24} + 2 \, {\left(l + 1\right)}^{23} + 2 \, {\left(l + 1\right)}^{22} \\ & \quad {} + 660 \, {\left(l + 1\right)}^{21} + 52 \, {\left(l + 1\right)}^{20} + 24 \, {\left(l + 1\right)}^{19} + 1512 \, {\left(l + 1\right)}^{18} \notag \\ & \quad{} + 394 \, {\left(l + 1\right)}^{17} + 124 \, {\left(l + 1\right)}^{16} + 1984 \, {\left(l + 1\right)}^{15} + 1250 \, {\left(l + 1\right)}^{14} \notag \\ & \quad{} + 428 \, {\left(l + 1\right)}^{13} + 1608 \, {\left(l + 1\right)}^{12} + 1738 \, {\left(l + 1\right)}^{11} + 936 \, {\left(l + 1\right)}^{10} \notag \\ & \quad{} + 848 \, {\left(l + 1\right)}^{9} + 984 \, {\left(l + 1\right)}^{8} + 708 \, {\left(l + 1\right)}^{7} + 285 \, {\left(l + 1\right)}^{6} + 50 \, {\left(l + 1\right)}^{5} \notag \end{align} We note in passing that the corresponding formula for the untwisted embedding of the Petersen graph in the projective plane, which is admittedly of less chemical interest, is \begin{equation} K(\text{Petersen} ;l) = {\left(l + 1\right)}^{15} + 10 \, {\left(l + 1\right)}^{12} + 30 \, {\left(l + 1\right)}^{9} + 55 \, {\left(l + 1\right)}^{6} + 55 \, {\left(l + 1\right)}^{3} \end{equation} \item For the $k$-prism $\mathcal{R}^k$: For $k$ odd, the prism with linear polyacene motifs, embedded on the sphere without a twist has \begin{equation} K(\mathcal{R}^k_0;l) = ((l+1)\sp3 + 2)\sp{k} - 2\sp{k} \end{equation} and for $k$ even it has \begin{equation} K(\mathcal{R}^k_0; l) = [ ((l+1)\sp3 + 2)\sp{k/2} + 2\sp{k/2} ]\sp2 \end{equation} \end{enumerate} We note that as the rules (a) to (c) apply without change to the limiting case of $l = 0$ hexagons in the linear polyene chain, the formulas for the untwisted chains apply to the leapfrog \cite{PWFTP1994} of an embedded cubic graph and hence give the Kekul\'{e} count of the leapfrog by summation of coefficients of all powers of $(l+1)$. For example, the formula for the untwisted dodecahedron $\mathcal{D}_0$ gives the number of Kekul\'{e} structures of the icosahedral C$_{60}$ fullerene as $K(\mathcal{D}_0; 0) = 12500$. Simple results are also found for the numbers of perfect matchings of leapfrog prisms, namely $3\sp{k}-2\sp{k}$ and $(3\sp{k/2}+2\sp{k/2})\sp2$. The prism itself has Kekul{\'e} count given by sequence A068397 \cite{oeisA068397}, i.e.\ $K$ for the $k$-prism is $F(k+1) + F(k-1) + (-1)^k + 1$. \section{Structural calculations} This section connects the foregoing mathematical development with possible realisations of new unsturated hydrocarbon frameworks. Graph theoretical considerations based on Kekulé counts, HOMO-LUMO gap and $\pi$ energy can be valuable indicators of stability of a $\pi$ system. In particular, it is useful to know if a $\pi$ system is predicted to have a closed shell (in which all electrons are paired) and to characterise such shells in terms of whether this electron configuration is {\it properly closed} (having all bonding orbitals occupied and all antibonding orbitals empty). HOMO-LUMO maps \cite{FoPi2010,JaFoPi2012} can give a useful picture of trends in frontier-orbital energies and shell types for families of molecules. However, for a more reliable estimate of the prospects for overall stability of an unsaturated hydrocarbon C$_x$H$_y$ it is necessary to take into account the full range of steric and electronic effects arising from both $\sigma$ and $\pi$ electronic subsystems. This section reports a selection of preliminary all-electron structural calculations using standard quantum chemical methods for various examples of chemical hexagonal complexes. They serve to show that this generalisation of benzenoids is chemically as well as mathematically plausible, and give clues to some of the factors that can influence absolute and relative stabilities. The systems chosen for study are linear polyacene expansions of the three cubic Platonic polyhedra, and a wider choice of isomeric expansions of the simplest cubic graph, the theta graph (see Figure~\ref{fig:PWF1}). All these structures correspond to the standard embedding on the sphere; twisted systems and alternative embeddings were left for future investigations. \begin{figure}[!htb] \centering \includegraphics[scale=0.8]{figs/f1.pdf} \caption{Cubic graphs used as the basis for flat hexagonal complexes: (a) the theta graph, (b) the tetrahedron, (c) the cube, (d) the dodecahedron. In cases (b) to (d), each edge is decorated with an anthracene chain; for case (a) see Figures~\ref{fig:PWF2}, \ref{fig:PWF3} and \ref{fig:PWF5}.} \label{fig:PWF1} \end{figure} In each case, the structure was optimised at the DFT level using the B3LYP functional and, in all but one, the 6-31G* basis for C and 6-31G for H. (In the case of the dodecahedral complex, C$_{420}$H$_{180}$, the basis was reduced to 6-31G for all atoms, on grounds of computational cost.) Candidate minima were checked in most cases in the usual way, by diagonalization of the Hessian. (For the largest cases, of the expanded cube and dodecahedron, stability of the candidate minimum structure was checked by relaxation of several nearby unsymmetrically perturbed structures.) Calculations were carried out with the QChem and Gaussian~16 packages \cite{qchem4,g16}. Energies and lowest harmonic frequencies are reported in Tables~\ref{tab:PWF1} and \ref{tab:PWF2}, together with geometric parameters and three graph invariants (Kekul{\'e} count, H{\"u}ckel binding energy per carbon atom, and H{\"u}ckel HOMO-LUMO gap). Snapshots of some optimised structures are shown in Figure~\ref{fig:PWF6}. \begin{table}[!t] \centering \setlength{\tabcolsep}{3pt} \begin{tabular}{lclrcrrc} Base graph & $l$ & Formula & $K$ & $(E_\pi/n)\beta$ & $\Delta_{HL}\beta$& $E$/eV & $\tilde{\nu}/\text{cm}^{-1}$ \\ \hline Theta graph & $3$ & C$_{ 42}$H$_{ 18}$ & 72 & $1.439933$ & $0.54778$ & $-43836.415$ & $76$ \\ & $4$ & C$_{ 54}$H$_{ 24}$ & 133 & $1.431856$ & $0.39425$ & $-56381.449$ & $52$ \\ & $5$ & C$_{ 66}$H$_{ 30}$ & 224 & $1.426646$ & $0.29932$ &$-68925.208$ & $34$ \\ Tetrahedron & $3$ & C$_{ 84}$H$_{ 36}$ & 4356 & $1.439017$ & $0.56885$ & $-175391.522$ & $39$ \\ Cube & $3$ & C$_{168}$H$_{ 72}$ & 19009600 & $1.443904$ & $0.54778$ & $-175390.812$ & --- \\ Dodecahedron & $3$ & C$_{420}$H$_{180}$ & 1561300213688815616 & $1.439049$ & $0.55627$ &$-438398.859$ & --- \end{tabular} \caption{Hexagonal chemical complexes based on inflation of cubic graphs with linear polyacene chains along edges. All isomers are based on linear annelation to alternate edges of the branching hexagons corresponding to vertices of the cubic graph. $l$ is the length of the chain motif, $K$ is the number of Kekul{\'e} structures, $(E_\pi/n)$ is the H{\"u}ckel $\pi$ energy per carbon centre, in units of the $\beta$ resonance parameter, $\Delta_{HL}$ is the H{\"u}ckel HOMO-LUMO gap, in the same units, $E$ is the total all-electron energy in eV (see text for the level of theory), and $\tilde{\nu}$ is the wavenumber in units of cm$^{-1}$ of the vibrational mode of lowest energy.} \label{tab:PWF1} \end{table} \begin{table}[!t] \centering \begin{tabular}{llrcclrr} Formula & Isomer & $K$ & $(E_\pi/n)\beta$ & $\Delta_{HL}\beta$ & $E$/eV & $\Delta E$/eV & $\tilde{\nu}/\text{cm}^{-1}$ \\ \hline C$_{42}$H$_{18}$ & Aa & 72 & $1.439933$ & $0.54778$ & $-43836.415$ & 1.789 & $76$ \\ & Ab & 108 & $1.445495$ & $0.81078$ & $-43837.062$ & 1.142 & $117$ \\ & Ac & 144 & $1.450217$ & $0.85153$ & $-43836.830$ & 1.373 & $ 70$ \\ & Ad & 144 & $1.450269$ & $0.85153$ & $-43836.782$ & 1.422 & $ 118$ \\ & Pa & 208 & $1.455866$ & $1.09287$ & $-43834.363$ & 3.840& $ 74$ \\ & Pb & 160 & $1.450380$ & $1.09287$ & $-43838.204$ & 0.000 & $ 117$ \\ & Pc & 208 & $1.456138$ & $1.09287$ & $-43836.842$ & 1.362& $ 124$ \\ & Pd & 176 & $1.452603$ & $1.09287$ & $-43836.495$ & 1.708& $ 102$ \\ & Pe & 208 & $1.455940$ & $1.09287$ & $-43835.822$ & 2.382& $ 136$ \\ & Pf & 176 & $1.452827$ & $1.03801$ & $-43837.000$ & 1.204& $ 102$ \\ C$_{66}$H$_{30}$ & Fa & 224 & $1.426646$ & $0.29932$ & $-68925.208$ & 2.610& $34$ \\ & Fb & 1088 & $1.436307$ & $0.75569$ & $-68927.219$ & 0.599 & $52$ \\ & Fc & 2500 & $1.444639$ & $0.91215$ & $-68927.818$ & 0.000& $50$ \end{tabular} \caption{Attachment isomers of hexagonal chemical complexes based on inflation of the theta graph with catafusenes composed of $3$ and $5$ hexagons (for formulas C$_{42}$H$_{18}$ and C$_{66}$H$_{30}$, respectively). The isomers are depicted in Figures~\ref{fig:PWF3} and \ref{fig:PWF5}. $K$ is the number of Kekul{\'e} structures, $(E_\pi/n)$ is the H{\"u}ckel $\pi$ energy per carbon centre, in units of the $\beta$ resonance parameter, $\Delta_{HL}$ is the H{\"u}ckel HOMO-LUMO gap, in the same units, $E$ is the total all-electron energy in eV (see text for the level of theory), and $\tilde{\nu}$ is the wavenumber in units of cm$^{-1}$ of the vibrational mode of lowest energy. } \label{tab:PWF2} \end{table} Structures based on the theta graph (Figure \ref{fig:PWF1}(a)) with all edges inflated to linear polyacene chains of $l$ hexagons were optimised successively for chains of length $l = 3, 4, 5$. These correspond to molecular formulas C$_{42}$H$_{18}$, C$_{54}$H$_{24}$, and C$_{66}$H$_{30}$, respectively. All were found to occupy minima on the potential energy surface within the model chemistry. The optimised structures are barrels, with CC bond lengths in the expected range for polycyclic aromatic systems in this model chemistry, \begin{figure}[!p] \centering \includegraphics[scale=0.7]{figs/f2.pdf} \caption{A {\lq polymer\rq} molecular notation for (a) anthracene and (b) phenanthrene expansions of the theta graph into flat hexagonal complexes of formula C$_{42}$H$_{18}$. Each monomer is to be repeated twice more to give a cyclic molecular structure with three-fold symmetry, topped and tailed by a hexagonal ring. Only graph vertices corresponding to carbon atoms are shown: vertices of degree two each carry a single H atom, and the graph is filled out with an appropriate Kekulé system of double bonds. The illustrated isomers are those denoted Aa and Pf, respectively, in the notation of Figure~\ref{fig:PWF3}.} \label{fig:PWF2} \end{figure} \begin{figure}[!p] \centering \includegraphics[scale=0.7]{figs/f3.pdf} \caption{Schematic notation for attachment isomers of three-hexagon catafusene expansions of the theta graph. Each vertical block represents a possible strip in a three-fold symmetric isomer based on either anthracene (A) or phenanthrene (P). Notation: a black circle denotes a C centre that has three C neighbours; a white circle denotes a C centre that has two C neighbours and one H neighbour. For simplicity, the catafusene strip is shown as vertical; in the molecule the strip must bend in order to keep parallel the median planes of the hexagons centred on the C$_3$ axis of the hexagonal complex. Double bonds can be added, for example in any way consistent with internal Kekulé structures of the parent catafusene.} \label{fig:PWF3} \end{figure} and low-frequency vibrational modes consistent with the flexibility expected of their open cage structures. Attempts to optimise the putative molecule with $l = 2$, C$_{30}$H$_{12}$, failed to yield converged structures that corresponded to the initial molecular graph. The equivalent inflations using $l = 3$ with the three cubic polyhedra (Figure \ref{fig:PWF1}) were also used to generate optimised molecular structures with formulas C$_{84}$H$_{36}$, C$_{168}$H$_{72}$, and C$_{420}$H$_{180}$, respectively (see Table \ref{tab:PWF1} and Figure \ref{fig:PWF1}). \begin{figure}[!b] \centering \includegraphics[scale=0.7]{figs/f5.pdf} \caption{Two isomers of hexagonal complexes based on decoration of the theta graph with five-hexagon catafusenes. The diagrams represent three-fold symmetric decorations of the graph with (a) a singly kinked polyacene chain (Fb), and (b) a zig-zag fibonacene chain (Fc). Double bonds can be filled in {\it ad lib} to correspond with internal perfect matchings of the respective catafusenes. Both isomers share molecular formula C$_{66}$H$_{30}$ with the straight-chain isomer~(Fa).} \label{fig:PWF5} \end{figure} Whilst these results are already encouraging, it is not to be expected that the use of the mathematically simple linear polyacene fragments to inflate graph edges will automatically lead to the isomer of the hexagonal complex that has the highest chemical stability. There are at least two variations to the construction recipe that might be expected on electronic and/or steric grounds to improve stability. The more obvious of these is that for $l \ge 3$ we have choice for the isomer of the catafused benzenoid to be used in the inflation procedure. Considered as isolated molecules, bent catafusenes are typically more stable than their linear counterparts. Figure \ref{fig:PWF2} illustrates this degree of freedom for the $l = 3$ expansions of the theta graph, where phenanthrene offers a plausible alternative to anthracene. However, even once we fix on a given catafusene as our favoured structural motif, there are still different possibilities for its mode of annelation to the branch-point hexagons (of type $A_3$) that represent the vertices of the original cubic graph. We can, for example, construct {\lq attachment isomers\rq} by choosing any contiguous pair $-$CH$-$CH$-$ on the catafusene perimeter as the site of the shared connection with the branch-point hexagonal ring. Figure \ref{fig:PWF3} illustrates the variety of possible attachment isomers for the case of three-hexagon chains, with the added constraint of threefold rotational symmetry around the branching hexagons. Optimisation shows that both variations on the basic recipe for construction are significant (see Table~\ref{tab:PWF2}). Compared to direct linear annelation, the linear anthracene fragment gives a more stable isomer when attached to the branching hexagons in non-linear fashion (Ab). However, a further improvement in total energy comes from switching to the phenanthrene motif in the construction of the complex, again with a significant energetic preference for one particular attachment mode (Pb). \begin{figure}[!b] \centering \begin{minipage}[b]{.5\linewidth} \centering \includegraphics[scale=1]{figs/BallAndStick1.png} \subcaption{}\label{fig:1a} \end{minipage}% \begin{minipage}[b]{.5\linewidth} \centering \includegraphics[scale=0.36]{figs/BallAndStick2.png} \subcaption{}\label{fig:1b} \end{minipage} \vspace{0.5cm} \begin{minipage}[b]{.5\linewidth} \centering \includegraphics[scale=1]{figs/BallAndStick3.png} \vspace{0.7cm} \subcaption{}\label{fig:1b} \end{minipage}% \begin{minipage}[b]{.5\linewidth} \centering \includegraphics[scale=1]{figs/BallAndStick4.png} \subcaption{}\label{fig:1b} \end{minipage} \caption{Ball-and-stick representations of some optimised molecular structures based on hexagonal complexes. (a) Isomer of C$_{42}$H$_{18}$ based on the anthracene expansion of the theta graph shown in Figure \ref{fig:PWF2}(a); (b) Isomer of C$_{42}$H$_{18}$ based on the phenanthrene expansion Pb of the theta graph (see Figure~\ref{fig:PWF3}(b); (c) Top view of C$_{66}$H$_{30}$ isomer based on the fibonacene expansion of the theta graph shown in Figure \ref{fig:PWF5}(b); (d) View down the two-fold axis of a hexagonal complex, C$_{84}$H$_{36}$, based on the anthracene expansion of the tetrahedron.} \label{fig:PWF6} \end{figure} Preference for a phenanthrene over an anthracene motif is consistent with the relative stabilities of the isomeric C$_{14}$H$_{6}$ compounds \cite{JChemThermo1979}. Higher stability of the bent polyacene system is attributed in part to $\pi$ resonance effects, although these are hard to quantify uniquely \cite{CiesielskiArkadiusz2008WAtK,Cyra}, and are offset by steric effects, such as H-H repulsion in the bay region. The computed relative energy of $1.142$~eV ($\sim 110$~kJ~mol$^{-1}$) of isomers Ab and Pb, each containing three copies of the respective motifs, is compatible with the differential stability estimated from the difference of $23$~kJ~mol$\sp{-1}$ in the standard formation enthalpies of the pure compounds \cite{JChemThermo1979}, but points to a significant role for relief of steric crowding in the more open structure of the phenanthrene hexagonal complex Pb (see Figure~\ref{fig:PWF3} (b)). In support of this hypothesis of a major role for steric effects, we note that the indicators of pure $\pi$ electronic stability do not show any clear correlation with the computed relative energies of attachment isomers (Table~\ref{tab:PWF2}, Aa-Ad, Pa-Pf). For example, whilst it is true that the isomer Pb, which has the lowest all-electron energy, has a higher Kekul{\'e} count, larger $\pi$ energy per electron and bigger HOMO-LUMO gap than its nearest competitor, Ab, it does not stand out on any of qualitative measures from the mass of the phenanthrenoid and anthracenoid isomers. Again, Pb has a large but not maximum Fries number. An overall trend to lower $\pi$ energy per electron and smaller HOMO-LUMO gap, countered by a rapidly increasing Kekul{\'e} count, is evident for CHCCs with longer linear polyacene motifs, and frequency calculations suggest increasing flexibility in larger cages with longer catafusene motifs. These considerations may also be promising for the prospects of larger hexagonal complexes based on cubic polyhedra, where the face sizes are typically larger, and there should be more room for avoidance of steric clashes. With larger faces and longer chains, the complexes with twisted M{\"o}bius catafusenes along polyhedral edges may also become less sterically disfavoured. There are clearly many possibilities to be explored. Although by no means complete, this short survey has shown that at least some generalised hexagonal complexes survive the initial test of chemical plausibility in that they occupy minima on the potential surface. Synthetic accessibility is of course another matter. \section*{Acknowledgements} TP is supported in part by the Slovenian Research Agency (research program P1-0294 and research projects J1-2481, N1-0032, J1-9187, J1-1690 and N1-0140), and in part by H2020 Teaming InnoRenew CoE. NB is supported in part by the Slovenian Research Agency (research program P1-0294 and research projects J1-2481, J1-9187, J1-1691, and N1-0140.). CSA is supported by the U.S. Department of Energy, Office of Science, Basic Energy Sciences, under Award \#DE-SC0019394, as part of the Computational Chemical Sciences Program. CSA performed calculations using resources of the National Energy Research Scientific Computing Center (NERSC), a U.S.~Department of Energy Office of Science User Facility operated under Contract No.~DE-AC02-05CH11231. CSA would also like to thank J. R. R. Verlet for access to computing facilities. \bibliographystyle{plain}

25596b5466d5e78567036d2177bc72a7c2629e97

\section{Introduction} \subsection{Background} The safety and security of the passengers in vehicles in the face of cyber attacks is a key element in the automotive industry, especially with the emergence of the Advanced Driver Assistance Systems (ADAS) and the vast improvement in Autonomous Vehicles (AVs)~\cite{ziebinski2016survey}. The main sensors available on modern self-driving cars are Ultrasonic sensors, mmWave (Millimeter Wave) radars, cameras, and LiDAR~\cite{wang2019multi,fung2017sensor}. These sensors are used for sensing the physical environment and for safety-critical decisions, such as collision avoidance and intersection management. Adversarial sensor attacks manipulate the input signals to the sensors in order to produce incorrect environment views with the goal of causing unsafe actions. These attacks can be organized into two main categories: Spoofing and Jamming~\cite{yan2016can}. While the latter attack can be easily detected, mitigating a spoofing attack usually requires more intelligent countermeasures~\cite{kapoor2018detecting}. Spoofing makes it very difficult for the sensor system to recognize that it is under attack, as it provides the victim sensor with seemingly legitimate but actually false data. In this work we focus on spoofing a vehicle's mmWave radar. \subsection{Related Work} Various sensors of autonomous vehicle are vulnerable to spoofing attacks. Prior work mainly focused on the study of security in perception sensors, primarily the LiDAR and camera. Physical world camera attacks were demonstrated by fooling Deep Neural Networks (DNNs), making objects mislabeled or ignored~\cite{song2018physical,eykholt2018robust}. For example, in~\cite{eykholt2018robust}, the researchers caused a ``STOP'' sign disappear in the eyes of the detector, by adding adversarial stickers onto the sign. Attacks against LiDAR-based perception systems in AVs appear in~\cite{cao2019adversarial,changalvala2019lidar}. The authors found that simply spoofing the LiDAR using common methods is not sufficient due to the use of machine learning object detection models. To mount an attack they leveraged an optimization method to generate adversarial examples to fool a real world working LiDAR. Some work has also been done to attack and spoof ranging systems, such as Radar and Ultrasonic sensors~\cite{thing2016autonomous}. Due to the fact that the mmWave radar is crucial in bad weather conditions, an attack to fool its measured distance is a critical physical security threat~\cite{goodin2019predicting,ranganathan2012physical}. Specifically, a distance-only spoofing attack~\cite{miura2019low,nashimoto2021low} was presented on a mmWave Frequency Modulated Continuous Wave (FMCW) radar. The system used a weak Arduino platform against an FMCW radar which used the less-common triangular waveform. The attackers synchronised to the victim via half-chirp modulation, and were able to control the delay measured by a victim radar to spoof the range. However, beyond measuring range, FMCW radars are typically used to measure velocity, which they can achieve with high precision~\cite{fung2017sensor}. The attack method in~\cite{miura2019low} does not spoof the velocity measured by the victim at all. Recently, the researches in~\cite{sun2020control} constructed several attack scenarios in order to spoof a mmWave radar of a Lincoln MKZ-based AV testbed, using a Software Defined Radio (SDR) transceiver system from National Instruments. The attack strategies involved synchronised attack radars on both sides of the road, using a complex setup. Furthermore, the method the attackers used to spoof velocity measurements was by spoofing the distance at subsequent time intervals. Since modern mmWave radars have the ability to measure velocity independently from distance, even a simple intra-radar sensor fusion approach can mitigate the attacks described, since the measured range and velocity are not coherent~\cite{jagielski2018threat}. \subsection{Our Contribution} In this work we demonstrate an adversarial radar based on a SDR, which can simultaneously manipulate the range and velocity measured by a victim radar used by an ADAS or AV platform. The adversarial radar we designed utilizes the advantages of a complex baseband FMCW radar: beyond controlling the delay to spoof the range, we are able to manipulate the signal phase received at the victim radar's Rx antenna, to spoof the velocity. Unlike~\cite{miura2019low,nashimoto2021low}, which focused on radars using a triangular chirp waveform, our system addresses the modern saw-tooth waveform (so called ``fast chirps''), as this chirp waveform is frequently found in contemporary FMCW radars in the AV industry~\cite{lutz2014fast,tong2015fast}. And unlike~\cite{sun2020control}, we only require a single rogue radar within a vehicle in front of the victim. After developing the attack theory, we demonstrate the spoofing attack by building a proof-of-concept hardware-based system, using a bladeRF Software Defined Radio. Using the ability to manipulate the velocity and the range measured by the victim radar, we demonstrate two realistic automotive attack scenarios: spoofing a phantom emergency break, and spoofing a phantom acceleration. In both cases, the range and velocity measured by the victim radar are coherent and fit the laws of physics governing vehicular motion. \section{The FMCW Radar} Generally, Radars emit electromagnetic waves and receive the reflection to measure the time of flight. Specifically, in FMCW radar solutions, the transmitted signal is a linear frequency modulated continuous wave (FMCW) chirp~\cite{mitomo201077}. FMCW radars are common in a variety of industries and applications, such as naval navigation radars, smart ammunition sensors, industrial radars, and in particular automotive radars~\cite{stove1992linear,lin2016design,komissarov2019partially}. Since the frequency component of the chirp is non constant in time, it is usually described as a frequency over time spectrogram, as described in Figure~\ref{fig:Range Measurement}(a). The description of the chirp development in time is given in Figure~\ref{fig:chirp tx and rx}(a). Modern systems use a saw-tooth waveform chirp with slope $S = \frac{df}{dt}$ and bandwidth parameters depending on the distance and velocity of the object of interest~\cite{winkler2007range,tong2015fast}. The state of the art FMCW radars integrated in commercial vehicles use the 76-81 GHz frequency band~\cite{gao2019experiments}. Typical FMCW Radar implementations include a Voltage Controlled Oscillator (VCO) which is used to create the necessary chirp waveform. The output of the VCO is amplified by the Power Amplifier (PA), and then transmitted through the Antenna. In order to improve the precision of the physical measurements, a series of N chirps (a frame) is averaged. The number of chirps $N$ is usually in the range of 50 to 1000~\cite{lutz2014fast}. \subsubsection{Measuring Range Using FMCW Radar} \begin{figure}[t] \centering \includegraphics[scale=0.4]{fmcw_spectrogram.png} \caption{Range measurement using FMCW radar: (a) Spectrogram of a frame of saw-tooth chirps with a chirp duration $T_C$. (b) The spectrogram reflected from the object of interest with returned chirps showing a delay of $t_d$. (c) The spectrogram of the IF signal at the receiver, with a beat frequency of $f_b$. (d) Demonstration of range-FFT in order to find the range corresponding to the beat frequency. } \label{fig:Range Measurement} \end{figure} The transmitted FMCW signal is reflected by the object of interest, and gets back to the Rx antenna of the radar after a total delay time $t_d$, as seen in Figure~\ref{fig:Range Measurement}(b). The received chirp is then mixed with the currently transmitted signal. The product of the mixing process is called an IF (Intermediate Frequency) signal, and it exhibits a single beat frequency of $f_{b}$ (Figures~\ref{fig:Range Measurement}(c),(d)). Finally, the IF signal is sampled by the ADC and further processed by a DSP or MCU. The range $d$ of the object of interest is related to the beat frequency $f_{b}$ by the following equation: \[d = \frac{c\cdot f_{b}}{2S}\] where S is the slope of the Tx chirp and $c$ is the speed of light. There are several ways to extract the beat frequency of the IF signal. A common way is to use a fast Fourier transform (Range-FFT), to convert the time domain signal into a spectrum with a peak at $f_{b}$, as depicted in Figure~\ref{fig:Range Measurement}(d). Typical IF frequencies are in the range of a few hundred kHz for objects within 100m. Thus, there is no requirement for a relatively high speed ADC. \subsubsection{Measuring Velocity Using FMCW Radar} \begin{figure}[t] \centering \includegraphics[scale=0.105]{vel_meas_1-1.png} \caption{Velocity measurement using FMCW radar: (a) Time representation of a chirp. The chirp is transmitted with an initial phase $\varphi _0$. (b) Two consecutive returned chirps $R_1(t)$, $R_2(t)$. (c) Two consecutive IF signals $B_1(t)$, $B_2(t)$ with the same beat frequency. The velocity of the object is proportional to the phase difference $\Delta\varphi _{21}$ between the phases of the IF signals.} \label{fig:chirp tx and rx} \end{figure} To measure velocity, the radar utilizes pairs of chirps $T(t)$ that are transmitted consecutively, as depicted in Figure~\ref{fig:chirp tx and rx}(b). The calculation of the velocity relies on the fact that the change in the beat frequency between adjacent IF signals in the frame is negligible compared to the phase shift of the IF signal. This assumption holds since the phase of the mmWave signal is more sensitive to object movement, in contrast to the return delay, which is practically constant between the chirps in the same frame, due to the Doppler effect and the fast chirp modulation. Similarly to the range measurement, the IF signal $B_i(t)$ (Figure~\ref{fig:chirp tx and rx}(c)) corresponding to each returned chirp $R_i(t) = T(t-t_d)$ is sampled and further processed, this time in order to calculate its phase, again using range-FFT. The relative velocity $v$ of the moving object to the radar can be derived from the phase difference $\Delta\varphi_{i+1,i}$ between adjacent chirps' IF signals $B_{i+1},B_i$:\[v = \frac{\lambda\Delta\varphi_{i+1,i}}{4\pi T_{c}}\] where $T_{c}$ and $\lambda$ are the chirp duration and the radar signal wavelength, respectively (see~\cite{iovescu2017fundamentals}). \section{The Adversarial FMCW Radar Model} \begin{figure}[t] \centering \includegraphics[scale=0.12]{attack_scenario-1.png} \caption{The attack scenario. The victim vehicle is behind the attacker's vehicle, on the same traffic lane. The attacker sends a spoofing signal to the victim in order to perform the attack.} \label{fig:Attack Scenario} \end{figure} \begin{figure*}[t] \centering \includegraphics[scale=0.25]{FMCW_adversarial_radar_2_no_background.png} \caption{ The Adversarial FMCW System: The original radar signal is received from the victim. Then, it is mixed with an internal synchronised chirp, to calculate the phase differences $\Delta\varphi_c$ and $\Delta\varphi_t$ between each consecutive pair of IF signals. The Adversary transmitter uses a quadrature modulator to control the transmitted phase and timing of each adversary chirp, to manipulate the victims measured velocity and range simultaneously. It sets the delay to $t_a$ and sets the phase to $\varphi_a$ compensating for $\Delta\varphi_c$ and $\Delta\varphi_t$ so the victim radar measures distance $\hat{d}$ and velocity $\hat{v}$.} \label{fig:System Summary} \end{figure*} \subsection{Attack Model} In our scenario the victim vehicle is behind the attacker's vehicle and has an FMCW radar installed, facing forward, as depicted in Figure~\ref{fig:Attack Scenario}. We assume the attacker has a modified radar system in his vehicle facing back. The attacker can send a much more powerful signal compared to the real reflected signal arriving at the victim. This assumption easily holds, even without extra amplification, as the attacker's signal only needs to traverse half of the distance, compared to the victim radar's signal. We assume that the victim radar parameters are known to the attacker. Moreover, the victim's true relative distance $d$ and relative velocity $v$ from the attacker are already measured by the attacker's radar, in order to be able to perform the attack. \subsection{Attack System Architecture}The adversarial system is depicted in Figure~\ref{fig:System Summary}. The adversarial radar is a modified version of the victim radar: essentially the only change is the additional delay and phase control logic. Our system is a complex baseband (quadrature) radar that simultaneously controls the delay (to spoof range measurements) and phase (to spoof the velocity). At a high level, the adversarial radar starts by generating internal chirps with the same parameters as the victim, but without transmitting them. Once it receives the chirps transmitted by the victim radar, it mixes the first $n$ chirps out of the frame with its internal chirps to produce its own IF signal. The attack system utilizes this IF signal to synchronise to the victim, and tune the attack parameters. Once it is synchronised and tuned, the attack radar ignores the victim's remaining chirps in the frame, and instead actively transmits $N-n$ spoofed return signals. The synchronization and tuning techniques we use are explained below. \subsection{Range Manipulation} By precisely controlling the time the spoofed signal is received at the victim radar, the attacker can manipulate the victim radar to measure a delay $t_a$, which is different than the original delay $t_d$, leading to an error in the calculated distance. In order to spoof the victim radar to measure an arbitrary distance $\hat{d}$, the attacker should know the exact transmit time of the current frame of chirps. To achieve synchronisation with the victim's transmit time, the attacker estimates the arrival time to his Rx antenna of the first $n$ chirps in the frame. Since the first $n$ chirps are used in order to synchronise to the victim, only the remaining $N-n$ chirps in the frame are used for the range and velocity manipulation itself. Due to the fact that $N\gg n$, the chirps used for synchronisation have little effect on the manipulated range and velocity measured by the victim, since it is done by averaging the measurements of each frame, or by using two dimensional range-FFT~\cite{lutz2014fast}. The transmition time can be estimated by down-converting the arriving signal and estimating the TOA (Time of Arrival) using a relatively high speed ADC, or by using HCM (Half Chirp Modulation)~\cite{goodin2019predicting}. Considering that the chirp duration $T_c$, the true distance $d$ and also the true delay $t_d$ are known, the attacker knows when the next chirp will be transmitted, and thus can decide when to emit an adversarial set of chirps, which will modify the delay measured by the victim radar. The attacker can force the victim to measure a distance $\hat{d}$ corresponding to a specific adversary delay $t_a$, by delaying or preceding the returned signal the victim senses. Since the adversarial radar is actively transmitting, its spoofed signals will overpower the legitimate echos reflected from the adversarial vehicle. \subsection{Manipulating The Velocity} As depicted in Figure~\ref{fig:chirp tx and rx}, in normal operation, the phase difference the victim radar measures between two consecutive chirps is given by:\[\Delta\varphi _{21}\ = \angle{B_2(t)} -\angle{B_1(t)}\] \[=(\varphi _2 - \varphi _0) - (\varphi _1 - \varphi _0) = \varphi _2 -\varphi _1\] where $\varphi_2$ and $\varphi_1$ are the phases $\angle{B_1(t)|_{t=t_1}}$ and $\angle{B_2(t)|_{t=t_2}}$ of two consecutive IF signals from two consecutive chirps, and $t_1$, $t_2$ are the delays corresponding to each chirp. As seen in Figure~\ref{fig:chirp tx and rx}(b), the victim FMCW radar expects to receive a delayed cloned version of the signal it sends, i.e., with the same phase $\varphi _0$, when it hits the Rx antenna. The method for velocity manipulation we propose takes advantage of this fact. To manipulate the velocity, the attacker transmits a frame of chirps with \emph{changing} phases. The difference $\Delta\varphi_{21}$ is known to the attacker because the true velocity of the victim is known. Therefore, by changing the phase $\varphi _0$ of the spoofed return signal separately for each chirp, it is possible for the attacker to control the velocity the victim measures. This can be done regardless of the victim radar's current transmitting chirp phase: the victim's measured phase difference depends only on the phase difference between consecutive attacker-generated returned chirps, and does not depend on the victim radar itself. \emph{Controlling The Phase:} As demonstrated in Figure~\ref{fig:System Summary}, the attacker should send a sequence of chirps with changing phases. To do so, the attacker utilizes the quadrature manner of the complex baseband radar in order to control the phase of the transmitted chirp, similarly to PSK (Phase Shift Keying) techniques. In order to make the victim radar measure a specific velocity $\hat{v}$, the phase of each chirp of the signal transmitted by the attacker should change with each consecutive chirp. The manipulated phase between consecutive chirps $\Delta\varphi _{i+1,i}$ is given by: \[\Delta\varphi _{i+1,i} = \frac{4\pi \hat{v}T_{c}}{\lambda}\] \section{The Process of Attacking a Victim Radar} \subsection{Range Manipulation} Beyond Figure~\ref{fig:System Summary}, Figure~\ref{fig:vel-spoof} shows the evolution of the attack signal. First, the adversarial radar enters a ``waiting for trigger'' mode. When it receives the first chirp, it synchronises with the victim radar using estimation of TOA of the signal, or using an internal HCM. This synchronisation occurs only once in the whole process of position and velocity manipulation. As seen in Figure~\ref{fig:System Summary}, when the attacker is synchronised to the victim, it can generate an IF signal $B_i(t)$ by mixing an internal synchronised version of the victim radar chirp with the incoming signal. This IF signal is depicted in Figure~\ref{fig:vel-spoof}(a). Now, it can manipulate the range the victim radar measures by precisely controlling the adversarial delay $t_a = t_d + t_m$, where $t_d$ is the actual time of flight (TOF) and $t_m$ is the manipulated delay the attacker adds. Since the legitimate round trip is $t_d$, the attacker should transmit the spoofing signal at $t_d/2 + t_m$, taking into account that the spoofing signal should still traverse the return path, which adds a delay of $t_d/2$. This is depicted in Figure~\ref{fig:vel-spoof}(c). \subsection{Velocity Manipulation} Although the victim radar and the adversarial radar are both using the same central frequency and sweep bandwidth, there is some difference between them, because of the TCXO (Temperature Controlled Crystal Oscillator) frequency tolerance. Therefore, there is an accumulated phase $\Delta\varphi_c$ with each internal IF signal $B_i(t-nT_c)$ in the attackers' radar. Furthermore, there is another phase accumulation $\Delta\varphi _t$, due to the true velocity difference $v$ between the victim and the adversary. The effect of the phases $\Delta\varphi_c$ and $\Delta\varphi _t$ on the internal IF signals $B_i(t-nT_c)$ is depicted in Figure~\ref{fig:vel-spoof} (b). The attacker takes these phases into account, and compensates them when manipulating the velocity. The total adversarial phase increment $\Delta\varphi_a = \Delta\varphi_c + \Delta\varphi_t + \varphi_m$ of each chirp the attacker sends back consists the compensation of the phases $\Delta\varphi_c$ and $\Delta\varphi_t$, and the manipulating phase $\varphi_m$. The manipulating phase $\varphi_m$, is the phase determining the current measurement change in relative velocity the victim will measure. By changing this phase with each chirp that the adversary transmits, it can precisely control the velocity $\hat{v}$ measured by the victim, as described in Figure~\ref{fig:vel-spoof}(d). Velocity and distance spoofing can occur independently, as the phase and the timing of the returned chirps are not related to each other. Therefore, the adversarial radar can manipulate both of them simultaneously. The adversary signal $S_{Adv}$ transmitted by the attacker in order to manipulate both range and velocity is the concatenation (sum) of the $N-n$ manipulated chirps: \[S_{Adv} = \sum_{i=n+1}^{N}T(t-(\frac{t_d}{2} + t_{m} + i\cdot T_c))\cdot e^{j\cdot i\Delta\varphi_a}\] Where $T(t)$ is the chirp transmitted by the victim radar, and $i$ is an index iterating over the $N-n$ manipulated chirps. \begin{figure*}[t] \centering \includegraphics[scale=0.4]{vel_pos_spoofing_nobackground.png} \caption{The Process of Range and Velocity Spoofing: (a) The received chirp at the adversary radar (orange) and the synchronised internal chirp (blue). (b) The changing phase between each internal measured IF signal in the adversary radar. $\Delta\varphi_c$ is the phase accumulated because of frequency difference from the victim, and $\Delta\varphi_t$ is the phase change caused by the relative velocity. (c) The transmitted adversary signal spectrogram, with the controlled delay $t_a=t_d/2+t_m$ corresponding to the spoofed distance $\hat{d}$. (d) The initial phase of each adversarial chirp emitted from the attacker as function of time. The phase is composed of the compensation for the phases $\Delta\varphi_c$ and $\Delta\varphi_t$, and the controlled phase $\varphi_m$ corresponding to the spoofed velocity $\hat{v}$.} \label{fig:vel-spoof} \end{figure*} \begin{figure}[t] \centering \includegraphics[scale=0.3]{system_photo2.png} \caption{The attack radar (xA9) below and the victim radar (xA4) above. The two radars are connected by 15m RF cables beyond the photo's right margin. } \label{fig:system_photo} \end{figure} \section{The Proof of Concept Setup} The setup we implemented (see Figure~\ref{fig:system_photo}) consists of two Software Defined Radios (SDR): one is the attack radar, and the other represents the victim. Both SDRs are models of bladeRF 2.0 micro. The bladeRF 2.0 micro is a 2x2 MIMO, 47MHz to 6 GHz frequency range, USB 3.0 Software Defined Radio~\cite{nuand20163}. It provides a powerful waveform development platform. We used the xA9 model to build the attack radar, and for the victim FMCW radar we used the xA4 model. The libbladeRF open source library was used for communicating with the devices, from a Windows machine. The interface we selected is the bladeRF Synchronous Interface, which allows transmitting bursts of samples at a specified timestamp. The timestamp is a free-running counter in the FPGA, incremented at the sample rate defined, with each outgoing sample. For the ease of prototyping, the victim Tx and Rx were connected the attacker's Rx and Tx respectively by 15m-long RF cables. To simulate greater and more realistic distances between vehicles with these cables, a one-time calibration process was carried out. In this process, the victim radar transmitted a set of chirps and the attack radar immediately returned the signal it received. Then, a fixed delay was added to the response time of the attacker, so the victim radar measured a distance $d = 60m$. Due to the fact that the length of the cable imitating the distance was fixed, the relative velocity $v$ simulated between the victim radar and the attacker was fixed to zero. To add an arbitrary phase shift $\Delta\varphi$ to the transmitted spoofed signal, we used the quadrature modulator of the attacking SDR. The attacker controls the transmitted chirps' phases by utilizing Euler's formula \[ A(t)\cdot e^{j\cdot\Delta\varphi} = A(t)\cdot\cos\Delta\varphi + jA(t)\cdot\sin\Delta\varphi \] and multiplying the in-phase and quadrature-phase parts of the modulator with the controlled amplitudes. In order to synchronise to the victim, we estimated the TOA, by sensing the first sample of the IF signal which is above an arbitrary threshold of noise. The center frequency $f_c$ used in the setup was 1GHz, the full bandwidth of the radar was 28MHz, the chirp duration $T_c$ was set to 1ms, and $n$ was set to 2 chirps. \begin{figure*}[t] \centering \includegraphics[scale=0.2]{simul_vel_pos_spoof.png} \caption{Results of Simultaneous Spoofing of Distance and Velocity: The green line shows the true range and velocity. The orange curve shows the desired manipulated range and velocity and the blue dots show the average values measured by the victim. The time between the measurements is 250ms. The shaded area shows the errors (+/- one standard deviation). (a) Spoofing an emergency brake. (b) Spoofing a scenario where an obstacle appears to be moving away from the victim radar.} \label{fig:sim_vel_pos} \end{figure*} \section{Results and Discussion} We simulated two scenarios: a sudden phantom stop and a sudden phantom acceleration. Each scenario was simulated 15 times. Simultaneous spoofing of distance and velocity is demonstrated in Figure~\ref{fig:sim_vel_pos}. Figure~\ref{fig:sim_vel_pos}(a) describes a scenario where the attacking radar simultaneously spoofs the distance and velocity representing a phantom emergency break in the adversary's vehicle in front of it: the victim's measured relative velocity $\hat{v}$ drops at a constant deceleration of $-10~m/sec^2$ spoofing a loss of $120~km/h$, and the measured relative distance drops at a quadratic rate from $60m$ to $0$ over about 3.5 sec: This rate of deceleration represents an emergency brake with ABS assistance \cite{kudarauskas2007analysis}. Note that the errors are in the order of the bandwidth resolution limit of FMCW, and only limited by the SDR relatively low bandwidth capabilities. Such a scenario will probably trigger an emergency break by the victim. Figure~\ref{fig:sim_vel_pos}(b) demonstrates another security hazard: the vehicle in front of the victim is made to appear accelerating, when it actually remains at a fixed distance $d = 60m$ and a fixed relative velocity $v = 0$. In such a case, the victim car might decide it can accelerate, leading to an accident and injuries to the driver and passengers inside of the car. The figure shows that the victim measures a velocity that mimics a fixed acceleration of $10~m/sec^2$, and the range measurements increase quadratically, matching the desired curve. Note that in both scenarios the desired spoofed orange curves are within the measured error range, and in particular the measured velocities are very close to the intended values. These attacks are very difficult for detection, since they are indistinct from a real scenario from the perspective of the victim vehicle. They cannot be mitigated by intra-radar sensor-fusion, since the range and velocity measurements are coherent with each other and consistent with the laws of physics. \section{Countermeasures} Our experiments showed the feasibility of simultaneous spoofing of the distance and velocity, and the security threats to the FMCW system. There are several countermeasures that can be used: \subsubsection{Sensor fusion with LiDAR/camera} Since modern vehicles are equipped with several range and velocity sensors, sensor fusion with non-radar sensors may mitigate the attack. I.e., when the mmWave sensor warns against impending collision, other sensors should agree. For example, in~\cite{yang2021secure}, the authors are proposing a sensor fusion framework that uses multiple sensors for measuring the same physical variable to create redundancy. Then it is used to estimate the real measurements and mitigate a physical level attack on sensor level. However, there are scenarios, such as bad weather conditions, where the only reliable sensor in the car is the mmWave radar. In such cases, sensor fusion can mislead the vehicle to believe that it has no danger ahead, while it is actually about to collide with the car in front of it, or, conversely, to halt when no emergency exists. \subsubsection{Phase Randomization} The velocity spoofing attack relies on the fact that the victim radar is expecting the returned signal to arrive with the same phase it was transmitted at. To break this assumption, one can send chirp signals with a randomized phases, which are known to the victim radar. Since the phases are randomized, the adversarial radar can't compensate the phase difference. This approach requires more complex circuity, e.g., another quadrature mixer, and therefore increases the cost of the radar. Moreover, sending another phase with each chirp adds another phase error to an environment that is already very error-prone. \subsubsection{Frequency Randomization} Another possible method to countermeasure the proposed attacks is to use random frequency hoping. This approach will make the attack radar measure a high beat frequency when implementing the proposed attack, and therefore will make it hard to synchronise to the victim radar. The BlueFMCW for example is a novel frequency hoping radar, that uses frequency randomization in order to countermeasure attacks~\cite{moon2020bluefmcw}. In~\cite{nashimoto2021low}, the authors also propose a random-chirp modulation based on randomization of frequency. Both mitigation systems require additional hardware in order to cope with the attack. The authors of~\cite{moon2020bluefmcw,nashimoto2021low} provide only numeric simulation level results for the countermeasures, and still require further work on a real equipment in order to convince their effectiveness. Furthermore, even in non-adversarial environments, the large number of vehicles on the same road already leads to frequency spectrum density. Therefore, frequent changes in frequency may help not only countermeasure the described attacks, but can also offer interference mitigation~\cite{bourdoux2017phenomenology}. \subsubsection{RSSI Measurements} Since the attacker assumes it will overpower the legitimate echos reflected from his vehicle with his adversarial signal, the victim can use the Received Signal Strength Indicator (RSSI) to countermeasure the attack. For instance, the victim should look for a pattern of $n$ weak chirps out of the total $N$ chirps in the frame. This approach may fail, when there is a legitimate change in the power of the signal received, due to weather conditions or a fading channel. \section{Conclusions} The security of the ADAS and autonomous vehicles sensors in the face of cyber attacks is crucial for the future of the automotive industry. In this paper, we proposed a system to attack an automotive FMCW mmWave radar, that uses fast chirp modulation. Our attack system is capable of spoofing the distance and velocity measured by the victim vehicle simultaneously, presenting phantom measurements coherent with the laws of physics governing vehicle motion. The attacking radar controls the delay in order to spoof its distance, and uses phase compensation and control in order to spoof its velocity. We demonstrated the spoofing attack by building a proof-of-concept hardware-based system, using a Software Defined Radio. We successfully demonstrated two real world scenarios, in which the victim radar is spoofed to detect either a phantom emergency stop or a phantom acceleration, measuring coherent range and velocity. We believe that with the proliferation of vehicular FMCW radars, the demonstrated attack system could pose a threat to AV and ADAS safety, and we propose several possible mitigations. \bibliographystyle{IEEEtranS}

95204adac1f446e987172d7d49bbf44ebcaf491b

\section{Introduction} \label{s:intro} The \textit{Spectrum-Roentgen-Gamma} (\textit{SRG}) orbital observatory\footnote{\url{http://srg.cosmos.ru}} was launched into a halo orbit around the Lagrange L2 point of the Sun--Earth system on July 13, 2019, from the Baikonur Cosmodrome by a Russian Proton rocket with a DM-03 booster (Fig.~\ref{fig:launch}). The Navigator platform (total mass 2712\,kg), developed by NPO Lavochkin (NPOL) in Khimki near Moscow, carries a scientific payload (total mass 1170\,kg) consisting of two X-ray telescopes with grazing-incidence optics (Figs.~\ref{fig:srg_npol} and \ref{fig:flight}): extended Roentgen Survey with an Imaging Telescope Array (eROSITA, \citealt{Predehl_2020}), developed in the Max Planck Institute for extraterrestrial Physics (MPE), Germany, and the Mikhail Pavlinsky ART-XC (Astronomical Roentgen Telescope -- X-ray Concentrator), developed in Russia \citep{pavlinsky21}. This scheme is implemented in accordance with a memorandum signed in 2007 between the Russian Space Agency (at that time) Roscosmos and the German Aerospace Agency, Deutsches Zentrum f\"ur Luft- und Raumfahrt (DLR). The eROSITA\ telescope is sensitive to X-rays in the energy range from {200\,eV} to 8\,keV, and ART-XC is sensitive at 4--30\,keV. The Navigator platform has been developed as a universal medium-class platform for scientific and meteorological missions to be launched into various orbits. Since January 2011, the Navigator platform has been used in the three {\it Elekro-L} meteorological satellite missions, as well as in the scientific {\it Spektr-R} mission ({\it RadioAstron}, \citealt{Kardashev2013}), which was launched in 2011 and operated until 2018. \begin{figure} \centering \includegraphics[width=0.98\columnwidth,clip]{srg_launch.png} \caption{Baikonur launch site (Kazakhstan): Proton rocket and the DM-03 upper stage with the \textit{SRG}\ spacecraft.} \label{fig:launch} \end{figure} \begin{figure} \centering \includegraphics[width=0.98\columnwidth,clip]{srg_npol.png} \caption{\textit{SRG}\ orbital observatory with the folded solar panels in NPO Lavochkin's assembly hall before shipment to Baikonur.} \label{fig:srg_npol} \end{figure} \begin{figure} \centering \includegraphics[width=0.98\columnwidth,clip]{SRG_in_flight.png} \caption{\textit{SRG}\ observatory in flight (artist's impression). Each X-ray telescope consists of seven independent mirror modules.}\label{fig:flight} \end{figure} \begin{figure} \centering \includegraphics[width=0.98\columnwidth,clip]{SRG_Fig-2p0_Ant_corr} \caption{{Russian Deep Space Network for the \textit{SRG}~ mission.} } \label{fig:map_with_antennae} \end{figure} The \textit{SRG}\ spacecraft is operated by NPOL's Control Center, while the responsibility for operation of the X-ray telescopes lies with the MPE in Garching near Munich (eROSITA) and the Space Research Institute (IKI) of the Russian Academy of Sciences in Moscow (ART-XC). { The downlink and uplink of spacecraft telemetry and commanding data is performed by the Centers of Deep Space Communications in Bear Lakes (64 m diameter antenna) near Moscow and Ussuriysk (70 m diameter antenna) in the Russian Far East as well as Baikonur (12 m antenna) placed in Kazakhstan Fig.~\ref{fig:map_with_antennae}.} The daily command uplink and scientific data downlink take about one hour for ART-XC\ and from 2.5 to 4 hours for eROSITA\ (including health checks of the telescopes' detectors and subsystems). The uplink of commands to the spacecraft and telescopes is performed using the antennas mentioned above or additional 12m diameter antennas in Baikonur and Bear Lakes. The \textit{SRG}\ orbit allows observations to be conducted around the clock with just short breaks for uploading and implementation of operating commands for the Navigator platform and the telescopes. The Navigator platform and both telescopes have onboard mass memory, allowing them to store the data accumulated over several days and dump them to the receiving stations during daily communication sessions. During downlink, the telescopes can (in most cases) continue observations and accumulation of data. The eROSITA\ and ART-XC\ telescopes each have seven independent mirror systems and seven independent positionally sensitive detectors in their focal planes (see Fig.~\ref{fig:artxc} for ART-XC\ and Fig.~\ref{fig:erosita} for eROSITA), six of which are coaligned on a hexagon surrounding an identical central mirror system and detector. To focus X-rays, both telescopes use grazing incidence optics \citep{Wolter1952a,Wolter1952b}. Figure~\ref{fig:wolter} demonstrates the Wolter I optical scheme of each of the seven ART-XC\ telescope modules, consisting of 28 pairs of coaxial parabolic and hyperbolic mirror shells, and a detector at the focus of each. During the 100-day flight of \textit{SRG}\ to the L2 point, {a thorough check-up of subsystems of the Navigator platform and both telescopes as well as calibrations of the telescopes} were carried out. An extensive program of observations of known point and diffuse X-ray sources with various spectral and timing characteristics was carried out during the flight to L2 to {calibrate the alignment of the optical axes of the 14 mirror systems of the telescopes (and their alignment with the axes of the satellite, of the optical star trackers, and of the eROSITA\ and ART-XC\ telescopes)} and to calibrate the imaging and spectral characteristics of the detectors as well as parameters relevant for timing studies of X-ray sources. The main operation regime of the \textit{SRG}\ orbital observatory is to scan the entire sky in X-rays with the goal of constructing all-sky X-ray maps in several energy bands. A detailed catalog of compact and diffuse X-ray sources in our Galaxy and the extragalactic Universe is created thereby. It was decided several years prior to launch that the first four years of the \textit{SRG}\ mission after the satellite's arrival at the operational orbit near L2 will be devoted to an all-sky survey through eight consecutive scans of the entire sky. This will allow the observatory to monitor the time variability of numerous X-ray sources: its telescopes will for each sufficiently bright source {provide} a light curve consisting of eight data points separated by half a year. The total number of these bright sources may reach many tens of thousands at energies between 300\,eV and 2\,keV and several hundreds at energies above 4\,keV. Among the most interesting X-ray transients expected to be found by \textit{SRG}\ in large numbers (for the first time) are tidal disruptions of stars by supermassive and intermediate-mass black holes \citep{Khabibullin_2014,Malyali2019,Jonker2020}. Furthermore, the telescopes of the \textit{SRG}\ observatory can trace the ``fast'' variability of X-ray sources during the all-sky survey because each of the eight data points in the long-term light curves will in fact consist of six measurements separated by four hours. This capability should be useful, in particular, for observations of afterglows of gamma-ray bursts (GRBs, \citealt{Khabibullin_2012,Ghirlanda2015,Ascenzi2020}), even if a GRB is directed away from the Sun or obscured by Earth from the view of near-Earth spacecraft (e.g., the {\it Fermi} observatory). It is planned that after four years of continuous scanning of the sky and the creation of eight independent X-ray maps of the entire sky, the \textit{SRG}\ observatory will switch for {at least two years} to a mode of detailed observation of the most interesting X-ray sources in triaxial stabilization. At this stage, long-term scanning of extended X-ray sources and selected regions in the sky with an area of up to 150\,square degrees will also be possible. Table~\ref{tab:srg} summarizes the main parameters of the \textit{SRG}\ observatory and its telescopes. Further parameters of the ART-XC\ and eROSITA\ telescopes are provided in Table~\ref{tab:art_keypar} in \S\ref{s:art} and Tables~\ref{tab:erosita} and \ref{tab:erosita_sens} in \S\ref{s:erosita}, respectively. \begin{table} \caption{Overview parameters of the \textit{SRG}\ observatory and its telescopes.} \label{tab:srg} \begin{tabular}{lcc} \hline \multicolumn{3}{c}{Observatory} \\ \hline Orbit & \multicolumn{2}{c}{Halo orbit around L2 point} \\ & \multicolumn{2}{c}{of the Sun--Earth system} \\ Total mass & \multicolumn{2}{c}{2712\,kg} \\ Power consumption & \multicolumn{2}{c}{1700\,W} \\ Telemetry rate & \multicolumn{2}{c}{512\,Kbit\,s$^{-1}$} \\ {Pointing accuracy} & \multicolumn{2}{c}{$4\hbox{$^{\prime\prime}$}$} \\ Observation modes: & \multicolumn{2}{c}{a) all-sky survey} \\ & \multicolumn{2}{c}{b) scanning (up to 150\,sq. deg)} \\ & \multicolumn{2}{c}{c) triaxial pointing} \\ Planned mission duration & \multicolumn{2}{c}{4 years all-sky survey and} \\ & \multicolumn{2}{c}{$>2$ years pointed/scanning obs.} \\ \hline Telescope & eROSITA & ART-XC \\ \hline Energy range & 0.2--8\,keV & 4--30\,keV \\ Energy resolution & $\sim 80\,{\rm eV}$ at 1.5\,keV & 9\% at {13.9\,keV} \\ FoV (FWZI) & $\diameter=62\hbox{$^\prime$}$ & $\diameter=36\hbox{$^\prime$}$ \\ Angular resolution & $30$\hbox{$^{\prime\prime}$}\,HEW & 53\hbox{$^{\prime\prime}$}\,FWHM \\ (during survey) & & \\ Pixel angular size & 9.6\hbox{$^{\prime\prime}$}\ & 45\hbox{$^{\prime\prime}$}\ \\ Sensitivity & $<\mu{\rm Crab}$ & {$\sim 100\mu{\rm Crab}$} \\ (4-year all-sky survey) & & \\ \hline \end{tabular} Notes: FWZI -- full width at zero intensity, HEW -- half-energy width, FWHM -- full width at half maximum. \end{table} The paper is organized as follows: in Section 2 we present the orbit and operation regimes of the \textit{SRG}\ observatory. In Section 3 we demonstrate the sensitivities and capabilities of the two X-ray telescopes of the \textit{SRG}\ and discuss the heterogeneity of the X-ray maps obtained during the first two scans of the entire sky. In Section 4 we describe the scientific goals and first results of the mission. In Sections 5, 6, and 7 we describe the key properties of the Navigator platform and of the ART-XC\ and eROSITA\ telescopes. In Section 8 we briefly describe the history of the \textit{SRG}\ project in Russia and the responsibilities of the German and Russian consortia of scientists in the processing and scientific analysis of the \textit{SRG}/eROSITA\ data coming from two complementary hemispheres of the sky. We finally draw our conclusions. \section{Orbit and operation regimes of the SRG observatory} \label{s:orbit} Figure~\ref{fig:orbit} shows the projection of the \textit{SRG}\ orbit onto the ecliptic plane and the scheme of the flight to the L2 point \citep{Eismont_2020}. Orbital corrections that were implemented before and after entry into the halo orbit around L2 are also indicated. \begin{figure} \centering \includegraphics[width=0.95\columnwidth,clip]{xy-crop_v2} \includegraphics[width=0.95\columnwidth,clip]{xz-crop_v2} \includegraphics[width=0.95\columnwidth,clip]{yz-crop_v2} \caption{Three projections of the SRG orbit onto the ecliptic planes and the scheme of the flight to the L2 point. The dotted line shows the lunar orbit. The points TCM1 and TCM2 denote major trajectory correction maneuvers of the orbit (see §5.4). The point TCM3 denotes the insertion into the quasi-stationary orbit around L2. The green and blue points indicate subsequent minor corrections of the orbit. Courtesy of the Keldysh Institute of Applied Mathematics of the Russian Academy of Sciences.} \label{fig:orbit} \end{figure} During the all-sky survey, the satellite rotates around an axis close to the direction toward the Sun with a period of four hours. The rotation axis {gradually} shifts by approximately one degree per day following the motion of the Sun (see Fig.~\ref{fig:rotation}). As a result, the eROSITA\ telescope (with its 1 degree field of view, FoV) observes each point source in the sky six times for 30--40\,s over a day, but typically only once every six months. The full FoV of ART-XC\ is 36 {arcmin}, so that a given celestial source is {typically exposed four} times per day for $\sim 20$\,s, also every six months. The variability of the sources in the vicinity of the ecliptic poles could be monitored much longer. The \textit{SRG}\ observatory observes the entire sky and builds its full map every six months (Fig.~\ref{fig:early}). Star trackers are used for the precise orientation of the \textit{SRG}\ observatory. Both telescopes use star tracker data to refine their pointing accuracy. Fig.~\ref{fig:srg_spacecraft} shows the positions of the main star trackers that are installed on the two telescopes of the observatory, as well as the position of the radio antennas on the Navigator platform. \begin{figure} \centering \includegraphics[width=0.98\columnwidth,clip]{SRG_scanning_corr.png} \caption{Rotation of \textit{SRG}\ with a period of four hours around the axis pointed at the Sun. This causes scans of large narrow circles in the sky. The survey plane moves slowly, following the Sun at approximately 1 degree per day.} \label{fig:rotation} \end{figure} \subsection{Inhomogeneity of the sky coverage during the survey} \begin{figure} \centering \includegraphics[width=0.98\columnwidth,clip]{srg_erosita_skysur_beg_nob.png} \caption{Beginning of the eROSITA\ all-sky survey. Initial test scans are visible, and a dark stripe shows when there were no observations. Individual scans cross at the north and south ecliptic poles. The inset shows a $2^\circ\times 2^\circ$ region within the deep field near the north ecliptic pole, containing the cluster of galaxies A2255 and a bright quasar.} \label{fig:early} \end{figure} The adopted strategy of the \textit{SRG}\ survey leads to the appearance of deep fields around the north and south ecliptic poles in the sky map (see Fig.~\ref{fig:early}), where the large circles of all individual scans cross. As a result, the exposure time depends on the ecliptic latitude $\theta$ as $1/\cos\theta$, that is, the exposure is shortest at the ecliptic equator and longest at the ecliptic poles. \begin{figure} \centering \includegraphics[width=0.98\columnwidth,clip]{scheme_SC_high_rez_reduced_corr.png} \caption{\textit{SRG}\ spacecraft. Positions of the key star trackers, two omnidirectional radio antennas and a medium-gain antenna for transmitting scientific data.} \label{fig:srg_spacecraft} \end{figure} \begin{figure*} \centering \includegraphics[width=0.95\textwidth,clip]{Fig9b.png} \caption{Rotation speed of the Z-axis of the \textit{SRG}\ spacecraft in the ecliptic plane during the first two sky surveys. The spacecraft spins around the Z-axis with a period of four hours, thus performing the all-sky survey. Due to limitations on orientation of the spacecraft, the angle between the Z-axis direction and the Sun and Earth must not exceed 13 and 24 degrees, respectively. This condition defines the speed with which the Z-axis has to follow the Sun and Earth at any point in the halo orbit around L2. Variation in the speed of the Z-axis causes variations in the exposure of the survey. Higher speed leads to lower exposures, and lower speed allows higher exposures of the corresponding sky regions. The vertical bars mark the dates of the gaps in the all-sky survey that are associated with orbit corrections and calibrations.} \label{fig:srg_scan_speed} \end{figure*} \begin{figure} \centering \includegraphics[width=0.98\columnwidth,clip]{art_exposure_1st_survey_3_scale_2} \caption{Exposure map (in Galactic coordinates) of the first all-sky survey by the ART-XC\ telescope. Exposure time is given in seconds (see the color scale shown on the right). The white spots in the map correspond to the ecliptic poles. Exposure times for eROSITA\ are approximately three times longer than for ART-XC,\ in accordance with the instruments' FoV (36\hbox{$^\prime$}\ and 1.03\hbox{$^\circ$}, respectively).} \label{fig:exposure} \end{figure} \begin{figure} \centering \includegraphics[width=0.98\columnwidth,clip]{ero_sens_nogrid.pdf} \caption{Distribution of eROSITA\ exposure (upper panel) and sensitivity in the 0.3--2.2\,keV energy band (lower panel) after the first three all-sky surveys. The solid curve in the upper panel shows the area of the sky in which the eROSITA\ exposure exceeded the given value. The solid curve in the lower panel shows the area of the sky in the hemisphere analyzed by the Russian consortium in which the $3\sigma$ sensitivity is better than the given value.} \label{fig:erosita_exposure} \end{figure} During the first two all-sky surveys, additional inhomogeneity in the sky exposure results from the \textit{SRG}\ elongated (in the ecliptic plane) halo orbit around L2. The distance to the Sun varies significantly over six months (see Fig.~\ref{fig:orbit}). The satellite moves ahead of Earth for three months and lags Earth for the next three months, which causes a significant variation in drifting rate of the scanning plane, from 0.7\,deg/day {at the} closest to the Earth segment of the orbit to 1.6\,deg/day at its farthest segment (see Fig.~\ref{fig:srg_scan_speed}). Figure~\ref{fig:exposure} shows the actual exposure map for the first \textit{SRG}\ all-sky survey based on ART-XC\ data. Figure~\ref{fig:erosita_exposure} (upper panel) shows the sky area covered by eROSITA\ with a given exposure. After the first three all-sky surveys, half of the sky has been covered with an exposure of at least 600\,s, while a total area of $\sim 3$\,square degrees around the ecliptic poles has been covered with an exposure exceeding 50~ks, so that source confusion is already high in these regions. The lower panel of Fig.~\ref{fig:erosita_exposure} shows the corresponding sensitivity achieved by eROSITA\ in the 0.3--2.2\,keV energy band after the first three all-sky surveys. \subsection{Deep scanning of sky fields with a size up to 150\,square degrees} \begin{figure} \centering \includegraphics[width=0.95\columnwidth,clip]{Fig12b.png} \caption{Sky visualization of a typical route of the X-axis in scanning mode.} \label{fig:scanmode} \end{figure} During the final stages of the flight to the L2 point, long observations of a large number of point sources and extended Galactic and extragalactic sources were carried out for calibration (Cal) and performance verification (PV) of the \textit{SRG} telescopes. In addition, scanning deep-survey observations were performed of a number of extended extragalactic fields, a region of the Galactic X-ray Ridge, and fields in the direction of nearby molecular clouds (to reveal X-ray diffuse emission between them and us). The targets for the PV phase have been chosen by the ART-XC\ science team and the science working groups of the German and Russian eROSITA\ consortia (see examples of the maps obtained during these observations in \S\ref{s:survey},\ref{s:deep} below). To perform deep surveys of extended (up to 150\,square degrees) fields, NPOL and IKI proposed and implemented a novel fruitful method of scanning observations, described in \S\ref{s:spacecraft} below and illustrated in Fig.~\ref{fig:scanmode}. {This method provides a much more uniform coverage of the observed field compared to the standard method, which is based on a grid of pointed observations, and enables obtaining high-quality maps of extended astrophysical objects.} \subsection{Pointed observations} The Navigator platform also allows observations of chosen targets to be done with the triaxial orientation of the observatory. This regime was successfully tested during flight calibration and PV observations. The pointing precision is described in \S\ref{s:spacecraft} below. \subsection{Instrumental background at L2} The particle background recorded by both instruments at L2 has been notably stable during the first 1.5 years of operations, when the solar activity was very modest. This is illustrated in Fig.~\ref{fig:bgatl2}, where the eROSITA\ count rate in the 7--9\,keV band is shown. In this energy range, the internal detector background dominates the astrophysical background. The horizontal dashed red line shows the typical quiescent background. There are rare spikes of counts with a modest amplitude, which are most likely caused by low-energy charged particles reaching the CCD through the thin entrance filter. \begin{figure} \centering \includegraphics[trim=2cm 5cm 0mm 9cm,width=1\columnwidth]{lc_erosita_s12time.pdf} \caption{eROSITA\ count rate (per single detector) in the 7--9\,keV energy band during the first ten months of the all-sky survey in 6000\,s bins. In this energy band, the detector background dominates astrophysical sources. The horizontal dashed red line shows the typical quiescent background, which is globally very stable. Short spikes, whose amplitude rarely exceeds 10-20\%, are plausibly associated with low-energy charged particles, which reach the {CCDs} through the thin entrance filters.} \label{fig:bgatl2} \end{figure} \begin{figure} \centering \includegraphics[trim=0cm 5cm 0mm 2cm,width=0.98\columnwidth,clip]{srg_ero_art_backgrounds.pdf} \caption{Sketch of ART-XC\ (dashed lines) and eROSITA\ (solid lines) astrophysical and internal detector backgrounds. The blue lines show the estimated sky background using the models of \citet{2002A&A...389...93L} and \citet{1999ApJ...520..124G}, convolved with the telescopes' responses averaged over the FoV. The red lines show the level of the detector internal background (fluorescent lines are not shown). For eROSITA, the astrophysical background dominates below $\sim 2$\,keV, while at higher energies, the detector background exceeds the sky background by an order of magnitude.} \label{fig:srg_backgrounds} \end{figure} The spectrum of the eROSITA\ detector background is described in \citet{Predehl_2020}. Here, we show a simplified version of the background model for both instruments (see Fig.~\ref{fig:srg_backgrounds}). In this figure, the blue lines show the (slightly modified) model of sky astrophysical background \citep{1999ApJ...520..124G,2002A&A...389...93L} convolved with the telescopes' responses. For comparison, the red lines illustrate the level of the internal background of the detectors (excluding fluorescent lines). For eROSITA, the astrophysical background dominates below $\sim 2$\,keV the astrophysical background, while at higher energies, the detector background is the factor that affects the sensitivity of the telescope. Figure~\ref{fig:srg_backgrounds} also demonstrates that the particle background in both instruments is comparable. \begin{figure} \centering \includegraphics[trim=0cm 0cm 0mm 0cm,width=0.95\columnwidth,clip]{cygx1_e04_4_stray} \caption{eROSITA\ image of the region around the extremely bright Galactic black hole binary Cygnus X-1 in the 0.4--4\,keV energy band. The image was saturated in the core to make very faint diffuse structures visible. In particular, the extended halo that is visible up to $\sim 3$ degrees from the source is due to X-ray photons scattered by mirrors only once (stray light). The same effect is largely responsible for a halo around Sco X-1 that is visible in the all-sky map. A special baffle mounted on top of eROSITA\ mirrors strongly reduces the magnitude of stray light, so that it is visible only around extremely bright objects. The radial rays that are clearly seen in the image are due to the mirror support structure.} \label{fig:stray} \end{figure} \subsection{Role of stray light} The eROSITA\ telescope optics (Wolter-I scheme) focuses photons that are scattered two times: first by parabolic shells, and then by hyperbolic shells. However, some photons can reach the detector after only one scattering or after two scatterings involving ``wrong scatterings'' , for instance, when they are scattered by the outer surface of a shell. All these photons are collectively called ``stray light''. A special baffle was introduced to reduce the magnitude of the stray light by an order of magnitude \citep{2014SPIE.9144E..4RF,Predehl_2020}. Nevertheless, a faint and extended halo associated with stray light can be seen when extremely bright sources are observed, such as the X-ray binary Cygnus X-1 (Fig.~\ref{fig:stray}). The role of stray light is more significant in the case of ART-XC\ (see \S\ref{s:art_char} below and \citealt{pavlinsky21} for details). \section{Sensitivity of the telescopes of the SRG observatory} \label{s:sens} Figure~\ref{fig:effarea} compares the on-axis effective areas of the eROSITA\ and ART-XC\ telescopes as a function of the energy of registered photons. The eROSITA\ and ART-XC\ curves intersect near 5\,keV. The telescopes nicely complement each other, detecting X-ray photons in adjacent energy bands. The results of the first all-sky survey have confirmed the preflight estimates of the effective area of the \textit{SRG}\ telescopes. \begin{figure} \centering \includegraphics[width=0.9\columnwidth,clip]{effarea_onaxis_v2.pdf} \caption{On-axis effective area of the {eROSITA\ (red) and ART-XC\ (blue)} telescopes of the \textit{SRG}\ observatory. These curves are based on the pre-flight beam tests and models, which are broadly consistent with the in-flight calibration results.} \label{fig:effarea} \end{figure} Figure~\ref{fig:sensitivity} shows the sensitivity of the eROSITA\ and ART-XC\ telescopes that was achieved during the first half-year survey as a function of energy. The sensitivity has been evaluated {assuming an unabsorbed power-law spectrum with a photon index of 1.8} in the energy interval $\Delta E\sim E$. The in-flight background spectrum (both the intrinsic detector background and unresolved astrophysical sky emission, see \citealt{Freyberg2020} for details) is integrated over this energy interval to predict the image surface brightness. Using this value, we can predict the expected distribution of fluxes (due to Poisson fluctuations of the number of counts) at a given position of an image convolved with the telescope point spread function (PSF). The sensitivity is then estimated as the flux associated with peaks that for a Gaussian distribution would correspond to 5$\,\sigma$ deviations. The corresponding curves are labeled S1. In the survey, the exposure is shortest near the ecliptic equator, where it is 200 and 60\,s for eROSITA\ and ART-XC, respectively. The difference in exposure is due to the difference in the solid angles subtended by the two telescopes at any given moment. The horizontal bars show the broadband sensitivities in the first survey, approximately converted into units of flux density. \begin{figure} \centering \includegraphics[trim=0cm 5cm 0mm 2cm,width=0.98\columnwidth,clip]{sense6_v2.pdf} \caption{Estimated sensitivities of the ART-XC\ (red) and eROSITA\ (blue) telescopes for a continuum spectrum in the energy interval $\Delta E \sim E$ as a function of energy. The curves show the sensitivity in units of energy flux density, i.e., ${\rm erg\,s^{-1}\,cm^{-2}\,keV^{-1}}$, evaluated for an interval of energies between $E_1=E/f$ and $E_2=E\times f$, where $f=1.6$. The top curves, labeled S1, correspond to the sensitivity achieved during the first all-sky survey close to the ecliptic equator, where the exposure time is shortest. The difference in the effective time (200\,s vs 60\,s for eROSITA\ and ART-XC,\, respectively) reflects the difference in the FoVs of the two instruments. Also shown are the sensitivity curves for 10 and 100 times deeper exposures. The horizontal bars show the broadband (0.2--2.3\,keV and 2.3--8\,keV for eROSITA\ and 4--12\,keV for ART-XC) sensitivities in the first survey, approximately converted into units of flux density. The spectrum of the Crab nebula scaled to a flux of 1\,mCrab and 1\,$\mu$Crab is shown for comparison. } \label{fig:sensitivity} \end{figure} \section{Science goals and first results of the SRG observatory} \label{s:tasks} The main scientific objective of the \textit{SRG}\ observatory is the construction of detailed X-ray maps of the sky and catalogs of point sources and extended X-ray sources in different energy bands from 0.3 to 12\,keV. The ART-XC\ and eROSITA\ telescopes were developed specifically for performing these tasks. It is expected that during the eight all-sky surveys spanning a period of four years, up to four to five million compact X-ray sources will be discovered: about three million active galactic nuclei (AGN), nearly one hundred thousand rich clusters of galaxies, up to a million stars with bright coronae (mostly M dwarfs), and tens of thousands of other Galactic objects, including cooling neutron stars and radiopulsars, accreting neutron stars (in particular, X-ray pulsars and bursters), black holes, and numerous white dwarfs in binary stellar systems. Also of interest is the exploration of extended X-ray sources, including supernova remnants (SNRs), pulsar wind nebulae, the hot gas filling most of the interstellar volume near the Galactic plane, rarefied gas in the halo of the Galaxy, and gas fountains in nearby galaxies. In addition, there is interest in the Local Bubble, comets in the Solar System, and in a search for traces of a shock wave at the boundary of the heliosphere, where the solar wind is halted by the interstellar gas surrounding the Solar System. It is also possible to continue the investigation that was begun by {\it ROSAT} \citep{Snowden_1997} of the origin of the Galactic soft X-ray background, which depends on the distribution of molecular and atomic gas and dust in the Galaxy, which absorb the soft X-ray emission. As demonstrated by the highly successful {\it ROSAT} spacecraft (all-sky X-ray survey in 1990, \citealt{Voges_1999}), these data are eagerly anticipated and will be widely used by the global astronomical community. Furthermore, the availability of eight all-sky surveys, each lasting six months, should help the \textit{SRG}\ observatory to discover a huge number of variable compact sources of Galactic and extragalactic origin. \subsection{X-ray maps obtained during the first all-sky survey and the CalPV phase} \label{s:survey} \begin{figure*} \centering \includegraphics[width=0.95\textwidth,clip]{art_sky_new.png} \caption{Positions (in Galactic coordinates) of the X-ray sources detected by \textit{SRG}/ART-XC\ during its first all-sky survey (typical exposure per point is just $\sim 60$\,s). Nearly 600 sources have been detected in the 4--12\,keV energy band, $\sim 60$\% of which are Galactic (black holes, neutron stars, white dwarfs, coronally active stars, SNRs, etc.) and $\sim 40$\% are extragalactic (AGN and a few dozen massive clusters of galaxies), as well as a number of newly discovered sources. The symbol size reflects X-ray brightness of the sources.} \label{fig:art_sky} \end{figure*} \begin{figure} \centering \includegraphics[trim=0cm 0cm 0cm 0cm,width=0.48\textwidth,clip]{erosita_4_8keV_v6} \caption{Positions (in Galactic coordinates) of X-ray sources detected by \textit{SRG}/eROSITA\ in the 4--8\,keV energy band during the first two all-sky surveys on that half of the sky in which the Russian Consortium is responsible for the data analysis (the Galactic center is on the right side of the image). The image includes $\sim$600 sources, and some of the brightest are labeled. } \label{fig:erosita_48} \end{figure} Figure~\ref{fig:art_sky} shows the distribution over the sky of the $\sim 600$ sources detected by ART-XC\ in the 4--12\,keV energy band during its first sky survey (December 12, 2019 -- June 10, 2020). Recently, an updated version of this map has been presented by \cite{Pavlinsky2021}, which is based on the sum of the first two sky surveys (December 2019 -- Dec. 2020). A total of 867 sources (821 point sources and 46 extended sources) have been detected by ART-XC. The 750 sources of known or suspected origin in this catalog include 56\% extragalactic sources (mostly AGN and also 52 rich low-redshift clusters of galaxies), and the rest are Galactic (X-ray binaries, cataclysmic variables (CVs), SNRs, etc.). {From} 114 sources, ART-XC\ has detected X-rays for the first time. Although the majority of them ($\sim 80$) are expected to be spurious given the chosen detection threshold, there are expected to be $\sim 35$ newly discovered astrophysical objects. An ongoing program of optical follow-up observations of these sources has already led to the identification of several new AGN and CVs \citep{zaznobin21,Zaznobin2021b}. The achieved sensitivity to point sources after the first year of the ART-XC\ all-sky survey varies between $\sim 4\times 10^{-12}$\,erg\,s$^{-1}$\,cm$^{-2}$ near the ecliptic plane and $\sim 8\times 10^{-13}$\,erg\,s$^{-1}$\,cm$^{-2}$ (4--12\,keV) near the ecliptic poles. The typical depth of the ART-XC\ survey is already comparable to that achieved in a similar energy band (4--10\,keV) in the recent {\it MAXI} ({\it Monitor of All-sky X-ray Image}) all-sky survey \citep{Kawamuro_2018} and is just slightly poorer than the sensitivity of the {\it XMM-Newton} Slew Survey in the 2--12\,keV band \citep{Saxton_2008}. However, the ART-XC\ survey greatly improves on the former in terms of angular resolution and provides full and regular sky coverage in contrast to the latter. The ART-XC\ survey sensitivity will be growing during the mission as a result of increasing exposure and probably also due to a decreasing flux of Galactic cosmic rays {as the next solar maximum approaches}. As many as $\sim 5000$ sources can be found by ART-XC\ by the end of the four-year all-sky survey. As noted above, the eROSITA\ telescope is noticeably inferior in sensitivity to the ART-XC\ telescope at energies above 5--6\,keV, but between 4 and 5\,keV, eROSITA\ is more sensitive. As shown in Fig.~\ref{fig:erosita_48}, during the first two surveys of the sky, eROSITA\ detected over 600 X-ray sources in one hemisphere in the range from 4 to 8\,keV. This number is comparable to the number of sources detected by ART-XC\ over the entire sky during one survey in the 4--12\,keV range. \begin{figure*} \centering \includegraphics[width=1.9\columnwidth,clip]{srg_map_annotated_reduced_corr.png} \caption{Annotated version of the \textit{SRG}/eROSITA\ first all-sky image. Several prominent X-ray features are marked, ranging from distant galaxy clusters (Coma, Virgo, Fornax, and Perseus) to extended sources such as SNRs and nebulae to bright point sources, e.g., Sco X-1 (the first known extrasolar X-ray source). The map of the whole sky is constructed by the two scientific consortia of \textit{SRG}/eROSITA\ in Germany and Russia. Each consortium created their images for one half of the sky. The map is an RGB map, where photons of different energies are shown in different colors: from 300 to 600\,eV in red, from 600\,eV to 1.0\,keV in green, and from 1 to 2.3\,keV in blue. The colors on this map, obtained from about 400~million photons registered by eROSITA\ during six months of the survey, allow one to immediately judge the temperature of radiating gas, ranging from 3 to more than 10~million\,K. At the very center of the map, the supermassive black hole Sgr A* is located with a mass of about 4 million solar masses. It is a rather weak X-ray source and is nearly invisible on this map. In the middle plane of the picture lies the disk of the Milky Way. It looks dark because the molecular gas and dust in the plane of the Galaxy absorb X-rays. The blue dots located in this region reveal a large number of bright and powerful X-ray sources in the Milky Way: X-ray pulsars, accreting black holes, neutron stars, white dwarfs in binary stellar systems, and remnants of supernova explosions.} \label{fig:erosita_skylabels} \end{figure*} \begin{figure} \centering \includegraphics[width=0.98\columnwidth,clip]{5Re_plotspec_SRGEJ170245p2+130107.pdf} \caption{Optical spectrum of quasar SRGE\,J170245.2+130107 (discovered by \textit{SRG}/eROSITA) obtained with the {BTA 6m telescope}. The light gray and dark gray lines show the spectra obtained on August 17, 2020, and September 13, 2020, respectively, while the combined spectrum is shown by the black line. The red dots show the source flux density in the Pan-STARRS $r, i, z, y$ filters. The vertical dashed lines show the expected positions of the peaks of the emission lines of the quasar at $z=5.466$. Adapted from \citet{khor21}.} \label{fig:qso55} \end{figure} The eROSITA\ telescope has obtained the best map of the sky in the history of X-ray astronomy in the 0.3--2.3\,keV energy band (Fig.~\ref{fig:erosita_skylabels}) even after its first all-sky survey. The good angular resolution and high sensitivity of eROSITA\ allowed it to detect over a million compact sources and map about 20 thousand extended sources. This huge number of sources cannot be displayed in a single image; only the brightest of them are visible on the map as dots. After scanning the sky for just six months, eROSITA\ has constructed a map that is about four times more sensitive and contains almost eight times more sources than the previously best all-sky map that was obtained in 1990 by the {\it ROSAT} satellite. eROSITA\ has already nearly doubled the total number of sources detected by all orbital observatories over the $\sim 60$ years of X-ray astronomy. The eROSITA\ map reveals spectacular giant bubbles of hot gas with temperatures up to 10 million K expelled from the plane of the Galaxy. These are the outcome of hundreds of thousands of supernova explosions and/or the intermittent activity of the supermassive black hole in the center of our Galaxy \citep{predehl20}. These X-ray bubbles are seen above and below the midplane of the image that encloses the well-known {\it Fermi} bubbles, which are associated with gamma-rays emitted through the interaction of cosmic rays with the ambient gas \citep{ackermann14}. About three-quarters of the objects in the \textit{SRG}/eROSITA\ map are AGN that are powered by the accretion of matter onto supermassive black holes residing in their centers. They are located far beyond the Milky Way. The quasar CFHQS~J142952+544717 at $z=6.2$ (corresponding to the age of the Universe of 900~million years) is particularly interesting. It was detected in X-rays for the first time by eROSITA\ and proved to have the highest X-ray luminosity ($\sim 3\times 10^{46}$\,erg~s$^{-1}$, or $\sim 10^{13}$ bolometric luminosities of the Sun) of the quasars at $z>6$ \citep{Medvedev_2020,medvedev21}. During the calibration and performance verification (CalPV) phase and in the course of scanning the entire sky, \textit{SRG}/eROSITA\ has discovered a number of extremely luminous quasars at redshift $z > 5$ and a noticeable number at $z > 4$ \citep{khorunzhev20,dodin20,bikmaev20,wolf20}. In Fig.~\ref{fig:qso55} we show the spectrum of {such a quasar} discovered by \textit{SRG}/eROSITA\ at $z = 5.5$. It was obtained with the 6-meter BTA telescope in the North Caucasus \citep{khor21}. The majority of the $\sim 20000$ extended objects in the eROSITA\ map are clusters of galaxies filled with dark matter and hot intergalactic gas, which shines in X-rays. Fewer than 50\% of these clusters were previously known from optical surveys or due to detection of the Sunyaev--Zeldovich effect \citep{sunyaev80} in their direction by the {\it Planck} spacecraft \citep{Planck_2016}, the Atacama Cosmology Telescope (ACT, \citealt{Hilton_2020}), or the South Pole Telescope (SPT, \citealt{Bleem_2015}) sky surveys. There will be much synergy and competition between different methods of observation of galaxy clusters over the next years through the continuation of sky surveys in different spectral bands, including the \textit{SRG}\ survey in X-rays and ground-based surveys by ACT and SPT in microwaves. Not only will the results of these surveys be valuable for cosmological studies, but we should also expect the discovery of thousands of strong gravitational lenses due to the deep gravitational potential of clusters of galaxies. Some two hundred thousand fairly close stars with coronae much more powerful than the corona of the Sun also contribute to the emission from the zones of relatively low temperature in the eROSITA\ sky map (Fig.~\ref{fig:erosita_skylabels}). Interestingly, eROSITA\ has detected X-ray emission from 150 stars with known exoplanets. This amounts to some 10\% of all nearby stars with known planetary systems (excluding the more distant stars with exoplanets in the field that were explored by the {\it Kepler} satellite). \begin{figure*} \centering \includegraphics[width=0.95\textwidth,clip]{erosita_ru.png} \caption{\textit{SRG}/eROSITA\ maps of one half of the sky in the 0.3--0.7 and 0.7--2.3\,keV energy bands.} \label{fig:erosita_ru} \end{figure*} Figure~\ref{fig:erosita_ru} demonstrates the dramatic difference between images of the sky (one hemisphere) in the 0.3--0.7\,keV and 0.7--2.3\,keV energy bands. In the 0.3--0.7\,keV band, the emission is fairly homogeneous, bright, and diffuse. This might be the background radiation of a huge number of soft X-ray sources located at cosmological distances and/or emission from the hot ($10^5<T_{\rm e}<10^7$~K) gas in the halo of our Galaxy. An additional significant contribution to the sky brightness at these energies might be provided by relatively nearby sources such the hypothetical Local Bubble that was created by supernova explosions in the relative vicinity of the Solar System. The strong absorption of soft X-rays by the cold atomic and molecular gas and dust in the Galactic plane is also noteworthy. A very different picture is observed in the 0.7--2.3\,keV energy band. Here the image is dominated by many hundreds of thousands of extragalactic sources. The main contribution to the source counts and to the diffuse background is provided by AGN. The absorption by cold gas and dust in the Galactic plane is strongly reduced by the rapidly decreasing photoabsorption cross-section with increasing photon energy. Bright Galactic sources in the region of active star formation in the Cygnus constellation are clearly visible in both maps. \begin{figure} \centering \includegraphics[width=\columnwidth,clip]{art_gc.jpeg} \caption{Fragment ($\sim 3^\circ \times 2^\circ$) of an image of the central region of the Galaxy obtained by ART-XC\ in the 4--12\,keV energy band during the CalPV phase.} \label{fig:art_gc} \end{figure} \begin{figure} \centering \includegraphics[width=0.9\columnwidth,clip]{art_snr.png} \caption{Image ($\sim 40$\,arcmin on a side) of the SNR RX~J1713.7$-$3946 obtained by ART-XC\ in the 4--12\,keV energy band during the CalPV phase.} \label{fig:art_snr} \end{figure} \subsection{Examples of the results of deep surveys during the PV phase} \label{s:deep} \subsubsection{Galactic center} During the CalPV phase, ART-XC\ performed deep surveys of a number of extended fields in the sky, in particular, of a large ($\sim 40$\,square degrees) region in the center of our Galaxy. Figure~\ref{fig:art_gc} shows a fragment of the image obtained in the 4--12\,keV energy band, which reveals the high quality and richness of the data. All in-flight characteristics of ART-XC\ have proved to be close to preflight expectations (see more details in \S\ref{s:art_flight}). In particular, the PSF averaged over the FoV is better than 1~arcmin (half-power diameter, HPD) in survey mode. The good angular resolution of the telescope is evident from the image (Fig.~\ref{fig:art_snr}) of the SNR RX~J1713.7$-$3946 that is clearly resolved by ART-XC. \subsubsection{Galactic Ridge} Figure~\ref{fig:ridge} demonstrates a large variety of astrophysical objects of our Galaxy that are accessible for observation by the \textit{SRG}\ observatory. During scans of the Galactic Ridge (in a field $\sim 20$~degrees away from the Galactic center), eROSITA\ has detected stars with active X-ray emitting coronae, star-forming regions and clusters of young stars, X-ray pulsars (rapidly rotating magnetized neutron stars), and SNRs. In the latter, X-ray photons are emitted by gas that is compressed in shocks where the material of the exploded star collides with the surrounding interstellar matter. As is well known, hot gas occupies 80--90\% of the volume near the Galactic plane, and just 10--20\% of the volume is filled by dense clouds of cold molecular and atomic hydrogen. \begin{figure*} \centering \includegraphics[width=0.9\textwidth,clip]{ridge.png} \caption{Annotated X-ray RGB map of a 25\,square degrees region in the disk of the Milky Way (the so-called Galactic Ridge) obtained by eROSITA\ in October 2019. In this field, thousands of Galactic X-ray sources are detected, as well as a number of quasars observed through the Galactic disk. In addition to many individual sources, the map also shows unresolved X-ray emission from hot gas and from a multitude of faint unresolved sources. Blue and green correspond to high photon energies emitted by a gas with a temperature of tens of millions of degrees, while red regions reveal colder gas of lower temperature. Adopted from Gilfanov, Medvedev, Sunyaev et al. 2022 (in preparation). Absorption in the cold interstellar gas near the Galactic plane attenuates soft X-ray radiation, but allows hard X-ray photons to leak through because their absorption cross-section is smaller. A significant contribution to the apparently diffuse emission from the Galactic disk and bulge is provided by the superposition of emission from numerous accreting white dwarfs, hot coronae of low-mass stars, and flares on them \citep{Revnivtsev2009}. eROSITA\ cannot resolve individual contributions of these sources because of their great number and low luminosities. } \label{fig:ridge} \end{figure*} \subsubsection{Lockman Hole} \begin{figure} \centering \includegraphics[width=0.95\columnwidth,clip]{lockman.png} \caption{X-ray image of the Lockman Hole obtained by eROSITA: More than 8,500 X-ray sources in 18\,square degrees. Adopted from Gilfanov, Burenin, Sunyaev et al (in preparation).} \label{fig:lockman} \end{figure} Figure~\ref{fig:lockman} demonstrates the richness of the extragalactic X-ray sky revealed during a long scan (about 8 ks per pixel) of the Lockman Hole zone. In this unique region, the absorption of X-rays by the interstellar medium of the Galaxy is close to its minimum in the entire sky. This allows studies of distant quasars and clusters of galaxies in unprecedented detail. In the $\approx 20$\,square degrees field, eROSITA\ has detected over 8,500 point X-ray sources, that is, $\sim 400$ sources per square degree. This number corresponds to $\sim 16$ million objects when extrapolated to the whole sky. The vast majority of these sources are AGN. According to photometric redshift estimates, the most distant of the quasars detected by eROSITA\ in the Lockman Hole are located at $z\sim 5$. Some $\sim 200$ clusters and rich groups of galaxies filled with hot gas were detected. In addition, several hundred active Galactic stars are also seen in the Lockman Hole field. \subsubsection{Coma cluster} \begin{figure*} \centering \includegraphics[width=0.99\columnwidth,clip]{coma_calpv_5deg_r200_v2.jpeg} \includegraphics[trim=0cm 0cm 0mm 0cm,width=0.935\columnwidth,clip]{srg_art-xc_coma_4_12} \includegraphics[trim=0cm -2.3cm 0mm 0cm,width=0.99\columnwidth,clip]{x_flattened_30beta_b08_3p5x2p4deg_lab} \includegraphics[trim=0cm 0cm 0mm 0cm,width=0.935\columnwidth,clip]{Planck_Coma_large_annot.jpg} \caption{{\it } X-ray image of the Coma cluster \citep{churazov21} in the 0.4--2\,keV band obtained by eROSITA\ in the course of the CalPV program (top left). The image is $\sim$6 degrees on a side, corresponding to 10\,Mpc at the distance of the cluster, with the logarithmic color-code spanning five orders of magnitude. The main cluster is in the process of merging with the NGC\,4839 group (a bright blob in the bottom right corner from the Coma cluster). Flattened X-ray image of the Coma cluster field (bottom left). The labels schematically mark some of the features that are presumably associated with the merger with the NGC\,4839 group. The dashed blue line is the suggested trajectory of the group, which enters the Coma cluster from the northeast direction, and is currently close to apocenter. The presumed positions of two shocks driven by the NGC\,4839 group are shown with the red and purple curves. The shock closer to the center is driven by the displaced gas that settles back to hydrostatic equilibrium. This is the most salient feature that can be directly seen in the image at the surface brightness edge. The green line shows the faint X-ray bridge connecting NGC\,4839 and the main cluster, which is a possible trace of the group passage through the Coma cluster. The yellow line shows the contact discontinuity, which is an interface between cold and hot gas patches with the same pressure. See \citet{churazov21} for details. ART-XC image of the Coma Cluster in the {4--12\,keV} energy band (top right).{\it } \textit{Planck} spacecraft $y$-parameter map of the Coma cluster, based on the data in the microwave spectral band (bottom right). The morphology of this map, which reflects the distribution of the hot gas electron pressure, is strikingly similar to the eROSITA\ image in the X-ray band. } \label{fig:coma} \end{figure*} \begin{figure} \centering \includegraphics[width=0.9\columnwidth,clip]{logN-logS.png} \caption{Preliminary $\log(N)-\log(S)$ distributions of compact sources in the extragalactic part of the hemisphere, where the Russian consortium of scientists is responsible for the data analysis and interpretation. The distribution of sources with and without Gaia stars among their optical counterparts are shown by blue and red histograms and are marked in the plot legend as GAIA stars and AGN, respectively. The distributions are derived from partial data of the first two sky surveys and do not use the full potential of the first-year data. No incompleteness correction was applied at the faint flux end. The coefficient $A$ on the y-axis incorporates approximations and inaccuracies of the preliminary analysis. It is reasonably close to unity and may weakly depend on flux, which is not necessarily the same for GAIA stars and AGN. } \label{fig:logn-logs} \end{figure} One of the main targets of the CalPV phase was the Coma cluster (Fig.~\ref{fig:coma}). \textit{SRG}\ observations in the scanning mode are particularly well suited for a detailed mapping of this massive and nearby cluster well beyond its virial radius. The obtained ART-XC\ data enable measuring the temperature of the hot gas and estimating the contribution of cosmic rays to the observed X-ray emission, while eROSITA\ images provide detailed information about the merger of the Coma cluster with its less massive companion, the galaxy group NGC4839. Fig.~\ref{fig:coma} clearly shows the rich substructure that is generated by the merger, including shock waves and contact discontinuities extending over a few megaparsec. The data from the {\it Planck} observatory (distribution of brightness in the microwave band due to the SZ effect) provide additional information about the distribution of pressure in the space between galaxies \citep{Planck_2013}. The comparison of the data from the three telescopes provides a detailed picture of the distribution of hot gas and dark matter in the cluster and its outskirts. It is important to note that in the Coma cluster field, eROSITA\ has unveiled many dozen massive clusters of galaxies and thousands of quasars located at cosmological distances far beyond the Coma cluster. This is a consequence of the deep scanning of a field with a size of $3^\circ\times 3^\circ$. \subsection{Source catalog} The source catalogs are maintained by the Russian (RU) and German (DE) consortia for the respective sky hemispheres. A preliminary version of the $\log(N)-\log(S)$ distribution of compact sources in the extragalactic part of the RU sky is shown in Fig.~\ref{fig:logn-logs} \subsection{X-ray spectroscopy} During the CalPV phase, ART-XC\ performed long pointed observations of a number of bright X-ray sources, including the well-known nearby AGN Circinus galaxy. The obtained spectrum (Fig.~\ref{fig:circinus}) shows a complex continuum in the 4--20\,keV energy band along with a strong Fe-K$\alpha$ emission line. The estimated energy resolution of $\sim 1.3$\,keV at 6\,keV matches the preflight estimates. \begin{figure} \centering \includegraphics[width=0.7\columnwidth,clip]{art_circinus.pdf} \caption{X-ray spectrum of the Circinus galaxy measured by ART-XC. The Fe-K$\alpha$ line is prominent on top of the complex continuum. The exposure is about 50~ks. Flux units are arbitrary.} \label{fig:circinus} \end{figure} \begin{figure} \centering \includegraphics[width=0.95\columnwidth,clip]{cygloop_e0420.png} \includegraphics[trim=0cm 4cm 0mm 6cm,width=1\columnwidth,clip]{cygloop_spec_s12.pdf} \caption{0.4--2\,keV X-ray image of the SNR Cygnus Loop obtained by eROSITA\ during the first two \textit{SRG}\ all-sky surveys (upper panel). The longest exposure time (per point) across the image is $\sim$400\,s. The axes are labeled in degrees. X-ray spectrum of the entire Cygnus Loop (lower panel). The ions contributing to some of the brightest emission lines are indicated. The spectrum is normalized per square arcminute.} \label{fig:cygloop} \end{figure} \begin{figure} \centering \includegraphics[,bb=40 180 540 650,width=0.9\columnwidth]{srge_w50_ss433pvx57_bjet2refllabb.pdf} \caption{{Spectrum of the Galactic microquasar SS 433 obtained during the \textit{SRG}/eROSITA\ PV phase. The data are well fit by the model of baryonic multitemperature jets \citep{2016MNRAS.455.1414K}, with contributions of the approaching (blueshifted) and receding (redshifted) jets shown in blue and red, respectively. The positions of the the lines of hydrogen- and helium-like silicon, sulphur, iron, and nickel are marked.}} \label{fig:ss433pv} \end{figure} The excellent capabilities of the eROSITA\ telescope in spectroscopy of hot astrophysical plasmas (temperatures of some million degrees) in bright X-ray sources during the all-sky survey are clearly demonstrated (Fig.~\ref{fig:cygloop}) by the spectrum of the central zone of the Cygnus Loop SNR obtained using an exposure of just 200\,s. X-ray emission lines of a number of ions of various chemical elements are clearly seen. The spectroscopic capabilities of eROSITA\ in the 3--9\,keV band are demonstrated in Fig.~\ref{fig:ss433pv}, which shows the spectrum of the Galactic microquasar SS 433 obtained during the \textit{SRG}/eROSITA PV phase. Emission lines of highly ionized atoms (silicon, sulphur, iron, and nickel) corresponding to the approaching (blueshifted) and receding (redshifted) baryonic multitemperature jets are clearly resolved. The line positions are fully consistent with the expectations based on the kinematic precession model with the bulk velocity of the jets equal to a quarter of the speed of light. \subsection{X-ray timing} \begin{figure} \centering \includegraphics[width=0.98\columnwidth,clip]{art_crab.pdf} \caption{Crab phase histograms reconstructed in several energy bands. The ART-XC exposure is about 45~ks.} \label{fig:fold} \end{figure} \begin{figure} \centering \includegraphics[width=0.98\columnwidth,clip]{SRGEJ213527_spec.pdf} \caption{eROSITA\ spectrum of tidal disruption event SRGe J213527.3$-$181634/ZTF20abgbdpr. The best-fit multicolor accretion disk emission model with the inner disk temperature of $\sim 100$\,eV is shown by the dashed line \citep{gilfanov20}.} \label{fig:tde_spec} \end{figure} The ART-XC\ telescope has an excellent time resolution of 23~$\mu s$, which is determined by the resolution of time stamps assigned to incoming events by the reading electronics. This permits a detailed study of the time behavior of millisecond pulsars in pointing mode. The timing capabilities of the telescope were tested during observations of the Crab pulsar. Figure~\ref{fig:fold} shows the light curve of this pulsar in several energy bands as measured by ART-XC\ and folded with the Jodrell Bank radio ephemeris. The $\sim33$~ms pulsations are clearly detected up to 30\,keV, and the energy dependence of the pulse morphology is as expected. The nominal integration time of the eROSITA\ CCDs is 50~ms. This integration time is set by the requirement that smearing of X-ray images in the all-sky survey is to be avoided. With the nominal rotation rate (360 degrees in four hours), the sources in the telescope FoV shift by $3.5"$ in 50~ms, which is $\text{about three}$ times smaller than the CCD pixel size. While this integration time precludes a systematic study of the variability on millisecond scales, the stability of the background and instrument characteristics provides an excellent opportunity for monitoring objects on longer timescales in three important intervals: 0.05--40 s (for sources crossing the FoV during a single scan), four hours to one day (during six intra-day scans of the same object), and on scales from six months to four years (over the entire \textit{SRG}\ all-sky survey). In addition, the variability of sources located close to the intersections of individual scans near the ecliptic poles can be tracked nearly continuously on scales longer than four hours. \subsection{ART-XC Galactic transients} In its daily scans (exposure of $\sim 60$\,s per source), the ART-XC\ telescope reaches a sensitivity ($\sim 5\sigma$) of $\sim 8\times10^{-12}$\,erg~cm$^{-2}$~s$^{-1}$ in the 4--12\,keV energy band (about 0.6\,mCrab). By surveying about $1\%$ of the sky every day, ART-XC\ provides rapid alerts about new X-ray transients (e.g., \citealt{atel13571,atel14206,atel14219}) or new outbursts from known or poorly studied historical sources (e.g., \citealt{atel13606,atel14051}). These events are typically also detected by eROSITA, which provides a more accurate localization for newly discovered sources and detailed spectral information below $\sim 2$--8\,keV for a broadband spectral analysis. The ART-XC telescope provides the unique possibility of studying the population of faint transients that would be otherwise missed because they are too weak for wide FoV telescopes and all-sky monitors [such as the {\it INTErnational Gamma-Ray Astrophysics Laboratory} ({\it INTEGRAL})/IBIS, {\it Swift}/BAT, and {\it MAXI}]. The relatively hard X-ray band of the ART-XC\ telescope also makes it less dependent on the source intrinsic or Galactic absorption and allows the detection of highly absorbed sources, which could be missed by soft X-ray instruments. For example, ART-XC\ transients often appear quite unremarkable in eROSITA\ data. Follow-up campaigns have already enabled establishing the nature of several ART-XC\ sources, such as the new microquasar SRGA\,J043520.9+552226/AT2019wey \citep{yao21a,yao21b,Mereminskiy2021}, the new nova-like CV SRGt\,J062340.2$-$265715 \citep{Schwope2021}, several new Be systems \citep{doroshenko21,Lutovinov2021}, and a few others. \subsection{Extragalactic transients and stellar flares} The approach of repeated sky surveys when every (typical) location on the sky is visited every six months has proven to be an efficient tool for studying the long-term variability of sources and for discovering various types of Galactic and extragalactic transients. Every 24 hours, eROSITA\ detects between half a dozen and a dozen sources that have changed their luminosity by more than an order of magnitude compared to the previous visit six months earlier (e.g., \citealt{gilfanov20,sazonov20}; Medvedev et al. in preparation). About half of these sources are associated with {\it Gaia} stars, and the remaining half are presumably of extragalactic origin. The requirement of at least a tenfold flux increase corresponds to an effective transient detection threshold of $\sim 2\times 10^{-13}$\,erg/s/cm$^2$. As every scan path crosses the ecliptic poles, sources in these regions are repeatedly scanned every four hours. For these sources, we have already accumulated light curves covering over 16 months, which opens unique prospects of detailed variability studies of stars and AGN on timescales from several hours to a few years (Medvedev et al. in preparation). The sky scans are planned to proceed in an unchanged manner at least until the end of the third sky survey. Tidal disruption events (TDEs) are being actively searched for among extragalactic eROSITA\ transients based on the optical and infrared properties of their hosts and on the shape of their X-ray spectra (Fig.~\ref{fig:tde_spec}). The Russian consortium identifies about one relatively bright TDE candidate every week \citep{khabibullin20,khabibullin20b,gilfanov20,Gilfanov2021}. They are followed up with optical spectroscopy on various telescopes in Russia (the 6 m BTA telescope in the Caucasus; the 1.5m Russian-Turkish telescope, RTT-150, in Turkey; the 1.6m AZT-33IK telescope of the Sayan Observatory; and the 2.5m telescope of the Caucasus Mountain Observatory of the Sternberg Astronomical Institute of the Moscow State University). An active collaboration with the Zwicky Transient Facility (ZTF) team facilitates timely classification of detected transients and their optical follow-up on Palomar and Keck telescopes. During the period from June 10 to December 14, 2020, spanned by the second \textit{SRG}/eROSITA\ all-sky survey, 16 TDEs have already been detected and optically confirmed in the $0<l<180^\circ$ hemisphere (\citealt{Sazonov2021}, Gilfanov et al., in preparation). The first eROSITA\ TDEs have also been reported in the other half of the sky \citep{Liu2021}. The \textit{SRG} scanning strategy means that more rapid transients may be found that vary on a timescale of $\text{about one}$ day, in particular, X-ray afterglows of gamma-ray bursts even without an observable gamma-ray trigger \citep{Khabibullin_2012}. A systematic search for these events is continuously ongoing \citep[e.g.,][]{Wilms_2020}. \section{SRG spacecraft} \label{s:spacecraft} Design of the structure and configuration of the {\it Spektr-RG} spacecraft continued until 2008 (see \S\ref{s:history} below about the history of the \textit{SRG}\ project). At that time, the concept design was defined and accepted. \subsection{{\it Navigator-SRG} platform} \label{s:platform} The spacecraft (Figs.~\ref{fig:srg_npol}, \ref{fig:flight}, and \ref{fig:srg_spacecraft}) is based on the {\it Navigator-SRG} space platform (Fig.~\ref{fig:platform}), which has been developed by NPOL for spacecraft of hydro-meteorological and scientific designation. By the time the \textit{SRG}\ design was accepted, the flight models of the Navigator platform for the {\it Spektr-R} ({\it Radioastron}) and {\it Electro-L} No. 1 spacecraft had been produced and tested. In 2011 the platform started its flight qualification. This meant an opportunity to gather invaluable experience of the platform operation, which was considered during the creation of {\it Spektr-RG}. Modifications aimed at enhancing the mission reliability and achieving the required technical parameters of the spacecraft as a whole were made. A principal change made to the Navigator platform was the installation of a new X-band radiocomplex, induced by the necessity of maintaining ground contact with the spacecraft at large distances. \begin{figure} \centering \includegraphics[width=0.98\columnwidth,clip]{Navigator_platform.jpeg} \caption{Navigator platform. Credit: NPO Lavochkin, see also http://www.russianspaceweb.com/navigator.html (credit: A.Zak).} \label{fig:platform} \end{figure} The Navigator-SRG platform includes the following units: an onboard control system, an onboard radiocomplex, solar panels with the orientation mechanism, a propulsion unit, a thermal control system, a telemetry system, an antenna-feeder system, a cable network, construction, a power supply system, and miscellaneous supplementary systems and units. The onboard control system performs the following basic tasks: it controls operation of the scientific payload and service systems of the spacecraft, it controls the motion of the spacecraft around the center of mass and the motion of the center of mass, it controls the spacecraft orientation by choosing the correct omnidirectional antenna for linkage with ground stations, it controls the technical state and diagnostics of onboard systems with subsequent transfer of information to Earth via the onboard radiocomplex, it manages contingency cases, and performs automatic transition of the spacecraft into stand-by mode in contingency cases that cannot be managed by the onboard systems. The onboard control system includes the following elements: an onboard computer, power automation units, a gyroscopic angular velocity sensor, two SDP-1 solar position sensors, two POS 347K solar orientation devices, two SED26 star trackers, and four Agat-15M reaction wheel sets. The main orientation control actuators of the onboard control system are reaction wheels (RWs). Three RWs are in operation, and one RW is in cold reserve. A set of three RWs ensures an angular velocity up to 0.07\,deg/s. For RWs desaturation (momentum dump), 16 stabilization thrusters are used (8 operating and 8 in cold reserve, 0.5~N each). These thrusters are also used during the initial flight phase to establish the solar orientation and in contingency cases. In addition to the stabilization thrusters, the propulsion unit also includes larger correction thrusters (5~N each) that are used for trajectory correction. All the thrusters are powered by hydrazine contained in two tanks, each of which is filled with 180\,kg of fuel. Electric power supply for the onboard service systems and scientific instruments is provided by solar panels. The total power consumption of all onboard systems does not exceed 1700~W, which is less than the expected capacity of the solar panels at the end of their service lifetime (1870~W). The power supply system also includes a 55~Ah battery that provides power for onboard systems during the initial phase after separation of the spacecraft from the launch vehicle or in contingency cases. Thermal control of the spacecraft is established by axial and contour heat pipes, as well as by electrical heaters that are controlled by the onboard control system. In addition, both telescopes have their own thermal control systems. The eROSITA\ and ART-XC\ telescopes are placed on a geometrically stable frame mounted on the top joint of the spacecraft base module. A number of electronic units of ART-XC\ and the gyroscopic angular velocity sensor device of the onboard control system are installed on a thermo-stabilized platform. \subsection{Orientation and operation regimes of the observatory} Except during the initial flight phase, the onboard control system operates in inertial orientation mode, which ensures triple-axis stabilization and spacecraft rotation into a programmed orientation with respect to the inertial frame. Orientation data from the star trackers are continuously coupled with the gyroscopic sensor data, resulting in an estimated orientation quaternion. In the inertial orientation mode, the following operations can be performed: spacecraft rotation for pointing telescopes (\textit{SRG}\ +X~axis) to a predefined target on the celestial sphere and triple-axis stabilization with respect to the inertial frame; spacecraft rotation at a constant angular velocity (typically 0.025\,deg/s) around the \textit{SRG}\ +Z axis, in which the rotation axis is constantly rotated at a low angular velocity ($\sim 1$\,deg/day) to ensure a smooth all-sky survey; and spacecraft triple-axis stabilization during trajectory-correction impulses. Several consequent spacecraft rotations can be combined into series to scan fields on a celestial sphere by predefined routes. In this mode, the onboard control system points the \textit{SRG}\ X-axis to the required starting point of a scan field (with respect to the inertial frame) and performs a program of relative rotations around the Y- and Z-axes of the spacecraft. During the relative rotations, the coupling of the star trackers and gyro data does not interrupt the precise route tracking with respect to the inertial frame. The scan field size can be up to $12.5^\circ\times 12.5^\circ$. A typical scan sequence consists of several S-turns (Fig.~\ref{fig:srg_scan}). The main scanning turns are performed around the Z-axis and smaller auxiliary turns (steps) are performed around the Y-axis. Step rotation can be arbitrarily low and is usually chosen from the 4--12~arcmin range. The highest angular velocity of the scanning turns may be as high as 0.04\,deg/s. \begin{figure} \centering \includegraphics[width=0.98\columnwidth,clip]{SRG_Scan_Scheme.png} \caption{Typical route of the X-axis in scan mode.} \label{fig:srg_scan} \end{figure} The spacecraft attitude-control error has been confirmed to be less than 10~arcsec. The angular velocity stabilization error does not exceed 0.72~arcsec/s. Figure~\ref{fig:main_restrictions} illustrates the main orientation restrictions on pointing and scanning observations of the \textit{SRG}\ observatory. \begin{figure} \centering \includegraphics[width=0.98\columnwidth,clip]{angle_restrictions.png} \caption{Main orientation restrictions for observations during scans and pointings.} \label{fig:main_restrictions} \end{figure} \subsection{Radiocomplex and data transmission} In addition to receiving commands and performing trajectory measurements, the onboard radiocomplex is also capable of combining scientific data and service telemetry into one data stream. The nominal rate of transmission of service telemetry data is 16~Kbit/s (real-time transmission of TM-data) and 32~Kbit/s (stored TM-data dump). The nominal rate of scientific data transmission is 512~Kbit/s, but it can also be performed at 64, 128, and 256~Kbit/s. Telecommands from ground stations are uploaded at a rate of 500~bits/s (typical) or 125~bits/s (contingency cases). The radiocomplex receives telecommands through an omnidirectional antenna. It transfers data through a medium-gain antenna with a beam width of $\pm 24^\circ$. The onboard telemetry system collects service telemetry sensor data and digital telemetry data from the onboard control system and writes it into the internal memory device. During ground contact, it is possible to dump the stored telemetry data or run live telemetry transfer mode. \subsection{Launch and operational orbit insertion} \label{s:launch} \begin{table*} \caption{Trajectory-correction schedule for the \textit{SRG}\ flight to the vicinity of the L2 point.} \label{tab:trajectory_flight} \begin{tabular}{c|c|c|c|c|c} \hline No. & Date/time & Days in flight & Number of PU ignitions & Total characteristic speed & Consumed propellant \\ & (Moscow) & & & m/s & kg \\ \hline 1 & Jul. 22, 2019 & 10 & 2 & 13.59 & 15.98\\ & 17:30:00.000 & & & & \\ 2 & Aug. 06, 2019 & 25 & 2 & 3.49 & 4.09\\ & 17:30:00.000 & & & & \\ \hline \end{tabular} \end{table*} The \textit{SRG}\ observatory operates in a quasi-periodic orbit around the L2 libration point of the Sun--Earth system. This orbit has a number of clear advantages for the mission related to the remoteness (1.5 million km, corresponding to a light travel time of 5\,s) of the L2 point from Earth in the direction away from the Sun, in particular, long visibility periods from ground stations for control and receipt of scientific data, and the prevention of periodical entries of the spacecraft into Earth’s radiation belts. The location of the Russian ground-control stations in the northern hemisphere required choosing a launch date that would ensure maximum daily radio visibility of the spacecraft. In addition, the choice of an optimal operational orbit had to minimize the characteristic velocity required to keep the spacecraft near the L2 point. The formal solution based on these two criteria implied the necessity of preparing separate flight plans for the upper stage for each possible launch date. Taking into account the anticipated time frame of the readiness of the spacecraft segments, two possible launch windows were eventually chosen: June 21-22, 2019, and July 12-13, 2019. The launch of the \textit{SRG}\ spacecraft from the Baikonur Cosmodrome successfully occurred on July 13, 2019, at 15:30:57 Moscow time. The {\it Proton-M} launch vehicle placed the orbital unit (OU), including the upper stage (US), the adapter system, and the spacecraft, in a falling trajectory with an apogee altitude of $\sim 200$~km. The first ignition of the propulsion unit (PU) of the US placed it in a transfer orbit $\sim 168\times 2013$~km. The second ignition placed the OU in the nominal flight orbit toward the L2 point $\sim 500\times 1450000$~km. The spacecraft later separated from the US, followed by the deorbiting of the US. The placement into orbit took two hours. During this process, the US spent 34~minutes in the Earth’s shadow. The flight toward L2 lasted approximately 100 days without any shadow intervals. To ensure placement of the spacecraft in the nominal orbit, three flight trajectory corrections on the way to the L2 point had been planned: on days 10, 20, and 40 of the mission. The corresponding reserve dates were also defined: on days 15, 25, and 45. A characteristic velocity budget of 100~m/s was allocated for these corrections. The first correction was executed in accordance with the schedule. The second correction was moved to the reserve date to acquire more orbit measurements after the first correction. By the time the third correction was due, the calculations indicated that it was not needed, so that there was no correction on day 40 (see Table~\ref{tab:trajectory_flight}). The operational orbit of the \textit{SRG}\ observatory was defined by the following constraints: the maximum distance of the spacecraft from the L2 point in the ecliptic plane $\sim 920000$~km, the exit from the ecliptic plane $\sim 700000$~km and $\sim 550000$~km toward the north and south ecliptic poles, respectively. After placement of the spacecraft into the nominal orbit near the L2 point (on day 100), orbit corrections are executed approximately every 50 days during the mission (see \S\ref{s:orbcorr} below). \begin{figure} \centering \includegraphics[width=0.95\columnwidth,clip]{ZRV_without_cor_corr.png} \caption{Minimum visibility without corrections.} \label{fig:vis_nocorr} \end{figure} \begin{figure} \centering \includegraphics[width=0.95\columnwidth,clip]{ZRV_large_cor_corr.png} \caption{Minimum visibility with large corrections.} \label{fig:vis_corr} \end{figure} \subsection{Spacecraft control} Spacecraft operation planning is performed on three timescales (long term, medium term, and short term). The degree of detail is to be increased when proceeding to the next stage. Long-term mission planning is determined by science tasks formulated by the scientific community. At the medium-term level, operations of the spacecraft and ground stations are planned for the following calendar month. Approximately two weeks before its initiation, NPOL performs a priori calculations of the spacecraft trajectory for this month, schedules preliminary intervals for spacecraft control sessions, and, if necessary, schedules technological operations for the spacecraft (which usually include only trajectory correction). NPOL delivers these data to the science operation and data center of IKI, where the month plan is completed with scientific operations in accordance with the long-term plan. The plan produced at this stage determines the sequence and parameters of the spacecraft orientation modes. The plan undergoes preliminary verification of consistency with the spacecraft and ground-stations operation limitations both on the side of IKI and NPOL. The spacecraft orientation limitations are the main verification criteria. In case of changes in the ground stations or in the onboard systems' condition, a monthly plan can be adjusted even during its implementation. In these cases, secondary verification and decision making usually takes no longer than three days. At the short-term level, exact operations of the ground station and spacecraft during an upcoming ground-contact session are planned. The main task of short-term planning is to ensure necessary spacecraft operations starting from the beginning of a session until at least the next session. Operations scheduled in a month plan are detailed in the form of a ground-contact session program, which is a sequence of commands for ground stations, spacecraft service systems, and the telescopes. To compose a ground-contact session program, NPOL collects requests for commands from supervisors of the corresponding systems. In particular, a sequence of commands to ensure the telescopes' operation (with specific command parameters and issue times or intervals) is generated at IKI based on requests of the ART-XC\ and eROSITA\ teams. Commands for the spacecraft and scientific instruments may be executed immediately during a ground-contact session or stored in the onboard control system memory for execution at a preset time (time-tagged commands). NPOL performs a simulation of a prepared ground-contact session on the onboard control system's logical testing bench. It provides means of precise verification of a program consistency with spacecraft operation limitations, and detects logical planning errors and incorrect commands. Simulation is performed for an interval from the start of a planned control session until the start of the next session. This technology ensures end-to-end simulation during the whole mission, which increases control reliability. Provided the simulation results are positive and all the parties endorse a control session plan, it is executed. The execution of the commands is controlled using telemetry data. If necessary, the program may be adjusted during a session, including issuance of commands initially not planned by the ART-XC\ and eROSITA\ teams. Ground-contact sessions take place daily. Each session lasts approximately 4--5 hours, which makes it possible to dump all scientific and service telemetry data gathered throughout the day, conduct trajectory measurements by at least two ground stations, check the status of onboard systems and scientific instruments, and upload a time-tagged command sequence for a subsequent period of up to several days. During a control session, the full real-time data stream from the spacecraft is received by NPOL through ground stations and transmitted to IKI. \subsection{Mission ground-control complex} Spacecraft control is provided by the Russian ground-control complex of the \textit{SRG}\ spacecraft, including the NPOL mission control center, ground stations, ballistic centers, and means of communication. The central authority responsible for the flight tests and spacecraft control is NPOL, which provides a platform for the cooperation of specialists participating in flight tests and the implementation of the scientific program. Radio communication with the spacecraft is provided by radio-technical ground complexes located at Baikonur (TNA-57 antenna of 12~m diameter), Bear Lakes (TNA-1500 antenna of 64~m diameter), and Ussuriysk (P-2500 antenna of 70~m diameter). Currently, only the Bear Lake and Ussuriysk ground complexes are used for receiving scientific data. The radio-technical ground complexes had been modernized for the implementation of the \textit{SRG}\ project and currently fully ensure the mission requirements. During periods of short daily radiovisibility intervals from the Russian Deep Space Network antennas (which usually happen in April and May), the Malargue, Cebreros, and New Norcia stations of the ESTRACK network are involved when necessary in receiving scientific data, according to the agreement between Roskosmos and ESA. Ballistic maintenance of the mission is provided by two ballistic centers located at the Keldysh Institute of Applied Mathematics of the Russian Academy of Sciences and the Central Research Institute of Machine Building (TsNIIMash). All elements of the ground-control complex are united into one data system by means of a high-bandwidth data network, which ensures data exchange between the elements of the ground-control complex and the scientific ground complex. \subsection{Large corrections of the SRG orbit} \label{s:orbcorr} \begin{table*} \caption{Trajectory-correction schedule for the orbiting of \textit{SRG}\ near the L2 point.} \label{tab:trajectory_l2} \begin{tabular}{c|c|c|c|c|c} \hline No. & Date/time & Days in flight & Number of PU ignitions & Total characteristic speed & Consumed propellant \\ & (Moscow) & & & m/s & kg \\ \hline 1 & Oct. 21, 2019 & 100 & 1 & 0.21 & 0.25\\ & 19:00:00.000 & & & & \\ 2 & Dec. 10, 2019 & 150 & 1 & 0.18 & 0.21\\ & 19:00:00.000 & & & & \\ 3 & Jan. 30, 2020 & 201 & 1 & 0.27 & 0.32\\ & 19:00:00.000 & & & & \\ 4 & Apr. 01,2020 & 263 & 1 & 1.00 & 1.13\\ & 19:00:00.000 & & & & \\ 5 & Jun. 16, 2020 & 339 & 1 & 0.95 & 1.07\\ & 21:00:00.000 & & & & \\ 6 & Aug. 05, 2020 & 389 & 1 & 0.94 & 1.08\\ & 20:00:00.000 & & & & \\ 7 & Oct. 05, 2020 & 450 & 1 & 3.01 & 3.38\\ & 19:00:00.000 & & & & \\ 8 & Nov. 23, 2020 & 499 & 1 & 6.23 & 6.99\\ & 19:00:00.000 & & & & \\ 9 & Feb. 28, 2020 & 588 & 1 & 6.24 & 6.94\\ & 17:00:00.000 & & & & \\ \hline \multicolumn{6}{c}{Remaining propellant: 319.63\,kg} \\ \hline \end{tabular} \end{table*} After placement of the spacecraft into the nominal orbit near the L2 point (on day 100), station keeping corrections were executed approximately every 50 days until September 2020 (see Table~\ref{tab:trajectory_flight}). Because the launch date was chosen using many criteria, there is a gap of about a month every Spring in the radiovisibility of \textit{SRG}\ in its nominal operational orbit from the Russian ground-control stations (see Fig.~\ref{fig:vis_nocorr}). In the Fall of 2020, a strategy of large manoeuvres began to be implemented (see Table~\ref{tab:trajectory_l2}) to alleviate this problem. This strategy was developed in accordance with \citet{canalias04} and consists of one test manoeuvre (3~m/s) and ten manoeuvres (6~m/s each). As demonstrated by calculations (see Fig.~\ref{fig:vis_corr}), it should enhance the visibility of \textit{SRG}\ from the Russian Deep Space Network antennas during Spring in the next years. This will significantly improve the implementation of control operations with the Navigator platform and telescopes as well as the dumping of science data. \section{Mikhail Pavlinsky ART-XC} \label{s:art} ART-XC\ on board the \textit{SRG}\ spacecraft is an X-ray grazing-incidence mirror telescope array. It was developed by the Space Research Institute (IKI) and the All-Russian Scientific Research Institute for Experimental Physics (VNIIEF). The NASA Marshall Space Flight Center (MSFC) produced the flight modules of the X-ray mirror systems. ART-XC\ is designed for conducting an all-sky survey in the 4--12\,keV energy band and pointed observations of selected astrophysical objects in the 4--30\,keV energy band. \subsection{Design} ART-XC\ consists of seven identical mirror systems (MS) paired with the unit of the Rontgen detector (URD) (Fig.~\ref{fig:artxc}). Each MS plus URD pair forms a telescope. All seven of them are pointing in the same direction. \begin{figure} \centering \includegraphics[width=.98\columnwidth,clip]{art_structure.png} \caption{Internal arrangement of the ART-XC\ telescope: Cone-shaped carbon fiber tube with seven identical X-ray mirror systems at the top. The mirror systems focus X-rays onto seven detector units.} \label{fig:artxc} \end{figure} The basic structure of ART-XC\ is a cone-shaped carbon fiber tube with a height of three meters. The MSs are mounted on top of this tube and focus X-ray photons onto their respective URDs. A sunshield protects the MSs from direct sunlight. The upper part of this tube is covered with copper shielding to reduce the stray-light background in the detectors. Each URD is equipped with a collimator to reduce stray-light background. The collimator includes a block of calibration X-ray sources ($^{55}$Fe + $^{241}$Am) for in-flight calibration. Heat pipes and a radiator are used to maintain the operating temperature of the detector at about {$-20\deg$C}. The star tracker is mounted near the MSs. \begin{figure} \centering \includegraphics[width=0.8\columnwidth,clip]{artxc_wolter.png} \caption{{Wolter-I optical scheme of ART-XC. $Dfr_{28}$ and $Dfr_1$ are the entrance apertures of the outermost and innermost shells, respectively, $l_h$ and $l_p$ are the heights of the paraboloids and hyperboloids, respectively, and $f_W$ is the focal length of the mirror system. Adapted from \citet{Pavlinsky2020}.} } \label{fig:wolter} \end{figure} The ART-XC\ MSs were produced and calibrated by MSFC \citep{Gubarev_2012,Gubarev_2014,Krivonos_2017}. Each MS contains 28 Wolter-I (Fig.~\ref{fig:wolter}) nested-mirror shells. The shells were fabricated using an electroformed-nickel-replication technique and coated with a $\sim10$~nm layer of 90\% bulk density iridium. The shell thickness varies slightly with radius: the thickness of the outer shells is larger than the nominal thickness to make them stiffer and thus to improve the angular resolution of the module. The upper ends of the shells are glued to the supporting spider. The nominal focal length of the MSs is 2700~mm. They were defocused by 7~mm during installation on the telescope to provide a more uniform angular resolution across the FoV. The detector system of ART-XC\ consists of seven URDs, two electronic modules, and a serial interface connection module. The position-sensitive X-ray detector for ART-XC\ was developed by IKI \citep{Levin_2014,Levin_2016}. The sensitive element is a double-sided strip detector (DSSD) based on a CdTe die with dimensions of $29.953\times29.953\times1$~mm$^3$. The high-quality CdTe dies were manufactured specifically for IKI by Acrorad Co. Ltd. (Japan). Coordinate resolution is provided by two mutually perpendicular sets of 48 strips on the two sides of the crystal. For signal acquisition, two VA64TA1 ASICs per detector are used, one on each side. The ASICs were manufactured by Ideas (Norway). More than 30 URDs were produced and tested at IKI. Seven of them were installed in the flight model of the ART-XC\ telescope. \subsection{Characteristics} \label{s:art_char} \begin{figure} \centering \includegraphics[width=\columnwidth,clip]{artxc_effar_comp.pdf} \includegraphics[width=\columnwidth,clip]{artxc_foveffar_2r_1r.pdf} \caption{ART-XC\ effective area for doubly reflected events (blue) and singly reflected events (red) at 8.1\,keV as a function of offset angle (left). ART-XC\ FoV-averaged effective area for doubly (blue) and singly reflected events (red) in the 4--35\,keV energy band as a function of energy (right). The graphs for the effective area are based on the effective area of the simulated mirror system and on the efficiency of the spare detector.} \label{fig:art_effar} \end{figure} The ART-XC\ FoV is $\sim36$ {arcmin} ($\sim0.3$\,sq.~deg), within which the MSs provide an angular resolution better than 1\hbox{$^\prime$}. The effective area (on-axis) is substantial up to energies $\sim 30$\,keV. In addition, there are photons that are only once reflected from the MSs, and as a result, they can fall onto the detector with offset angles up to $\sim 50'$. Figure~\ref{fig:art_effar} illustrates the effect of these singly reflected photons on the ART-XC\ FoV and their contribution to the FoV-averaged effective area. For this type of events, the ART-XC\ \ FoV is $\sim2$\,square~degrees. Although true imaging cannot be done in this extended FoV and singly reflected photons generally cause an increase in the background, they can be used to measure the X-ray fluxes of bright sources. Therefore, the ART-XC\ telescope can also be exploited as a concentrator. In survey mode, ART-XC\ can therefore monitor bright transient sources for at least 28--32~hours \citep{Pavlinsky_2019b}. {The URD can register photons with energies up to $\sim 100$\, keV, with an energy resolution of 9\% at 13.9 keV.} The detector strip size corresponds to an angular size of 45\hbox{$^{\prime\prime}$}. During a calibration campaign at the IKI X-ray test facility \citep{Pavlinsky_2018,Pavlinsky_2019a,Pavlinsky_2019b}, it was determined that the efficiency of the ART-XC\ spare URD reaches 50\% in the 4.47--4.76\,keV energy range and 90\% in the 9.43--10.04\,keV energy range. Based on the model for the MS effective area and on the telescope ground calibrations, the on-axis effective area, vignetting, and grasp of ART-XC\ have been estimated. The ART-XC\ on-axis effective area at 8.1\,keV is 385\,cm$^2$. The grasp at 8.1\,keV is 43.8\,cm$^2$~deg$^2$. Table~\ref{tab:art_keypar} summarizes the performance parameters of the ART-XC. \begin{table}[hb] \caption{ART-XC characteristics.} \label{tab:art_keypar} \resizebox{0.95\columnwidth}{!}{ \centerline{\begin{tabular}{l|c} \hline Telescope mass & 350\,kg \\ Dimensions & 3.5~m $\times \diameter 0.9$~m \\ Power & 150~watts \\ Number of modules & 7 \\ Nominal focal length & 2700~mm \\ Operational energy range & 4--30\,keV \\ FoV & 0.3\,sq. deg \\ Eff. area for pointed observations & 385\,cm$^2$ @ 8.1\,keV \\ Grasp & 43.8\,cm$^2$~deg$^2$ @ 8.1\,keV \\ Ang. resolution (FWHM) in survey & $53\hbox{$^{\prime\prime}$}$ \\ Energy resolution & 9\% @ 13.9\,keV \\ Time resolution & 23\,$\mu$s \\ \hline \end{tabular}} } \end{table} \subsection{Scientific goals} \label{s:art_goals} The main goal of ART-XC\ is to survey the whole sky in the broad X-ray energy range of 4--30\,keV with a sensitivity of $\sim 10^{-12}$\,erg~s$^{-1}$~cm$^{-2}$ ($\sim 10^{-13}$\,erg~s$^{-1}$~cm$^{-2}$ near the ecliptic poles) in the 4--12\,keV band and an angular resolution better than an arcminute. The ART-XC\ survey will thus be the most sensitive all-sky survey ever conducted at these energies. The ART-XC\ survey complements the all-sky survey performed by eROSITA\ in the overlapping energy band of 0.3--8\,keV (with the highest sensitivity below 2\,keV). The harder energy band of the ART-XC\ survey is particularly valuable for studying populations of heavily obscured astrophysical objects. Preliminary estimates show that during the four-year all-sky survey, ART-XC\ will detect $\sim 5000$ X-ray sources, mostly AGN, including heavily obscured ones (with hydrogen column densities $N_{\rm H}\gtrsim 10^{23}$~cm$^{-2}$). This will provide a rich database for explorations of the AGN population at $z\lesssim 0.3$. ART-XC\ will also provide valuable information about the temperature of the intracluster gas in rich low-redshift clusters of galaxies. This will help tighten constraints on the cosmological parameters inferred from the eROSITA\ all-sky survey. The ART-XC\ all-sky survey can make breakthroughs in studying various classes of Galactic X-ray sources, such as X-ray binaries and cataclysmic variables (CVs). As many as $\sim 10^3$ CVs can be found, compared to $\lesssim 100$ known from previous X-ray surveys. With its unique combination of broad energy coverage, good angular resolution, and wide FoV, ART-XC\ should make major advances in the study of SNRs and the Galactic Ridge X-ray emission. Finally, ART-XC\ is well suited for discovering and monitoring transient and variable X-ray sources, such as X-ray and gamma-ray bursts, Galactic X-ray transients, and AGN. Although the probability of catching GRBs in the ART-XC\ FoV is not high, bright GRBs can penetrate the shielding material of the telescope and induce a signal on the detectors. For these events, ART-XC\ can provide precise timing information that can be used to localize them through triangulation with other space observatories. \subsection{In-flight performance} \label{s:art_flight} ART-XC\ has been operating in orbit for almost two years at the time of writing. The results obtained during the CalPV phase and the all-sky survey so far fully confirm the expected unique capabilities of the instrument. In particular, in order to calibrate the ART-XC\ effective area, a series of observations of the Crab nebula was performed during the CalPV phase. The effective area proved to agree well with the results of simulations and ground calibrations \citep{Pavlinsky_2018,Pavlinsky_2019a,Pavlinsky_2019b}. \section{eROSITA} \label{s:erosita} \subsection{History} \label{s:erosita_history} The concept of the eROSITA\ telescope is based on a long series of previous scientific and technological developments dating back to the very successful German, US, and UK \textit{ROSAT} mission \citep[1990-1999;][]{Truemper1982}, which was developed and managed under the leadership of the Max Planck Institute for Extraterrestrial Physics (MPE). \textit{ROSAT} carried out the first complete survey of the sky with an imaging X-ray telescope in the energy range between 0.1 and 2.4\,keV, and it performed tens of thousands of pointed observations. The following generation of large X-ray telescopes, the NASA-led \textit{Chandra} Observatory and the ESA mission \textit{XMM-Newton}, required the development of mirror systems with longer focal length (7.5-10 meters) in order to bring harder X-ray radiation into focus. These observatories, both launched in 1999, were only able to perform pointed observations, {however, their} limited FoV meant that it was not possible to perform large-area surveys. This led to the proposal of a hard X-ray imaging telescope with all-sky surveying capabilities. The Astrophysikalisches Institut Potsdam (AIP), the MPE, and the University of T\"ubingen (IAAT) therefore proposed the mission called A BRoad-band Imaging X-ray All-sky Survey (ABRIXAS). From the very beginning, the \textit{ABRIXAS} mission concept was developed through the coherent adaptation of mirror and detector technologies developed for \textit{XMM-Newton} for a national, small-scale satellite mission. Consequently, the development timescale for the project was relatively short ($\text{about three}$ years) and the overall costs moderate. The seven mirror modules shared in the focal plane an identical copy of the pn-CCD camera developed for \textit{XMM-Newton}, and with their small focal length of 1.6 meters, they were ideal for a small satellite. Because of a design failure in the power supply, however, the satellite lost its main battery soon after launch, in April 1999, and could never be used for scientific purposes. The hard- and software developments enabled by the \textit{ABRIXAS} mission have been extremely useful in the framework of a number of subsequent projects, however. \begin{figure*} \centering \includegraphics [width = 0.45\textwidth,clip]{10_201612_front.png} \includegraphics [width = 0.45\textwidth,clip]{201612_IABG_7103_crop.jpg} \caption{Pictures of the fully integrated eROSITA\ telescope taken during the final space qualification campaign in December 2016. Front view of the telescope (with open cover) with all seven mirror assemblies installed (left). Rear view of the telescope with all seven camera assemblies installed (right).} \label{fig:erosita} \end{figure*} Despite the unfortunate outcome of the \textit{ABRIXAS} mission, the appeal of the original scientific goal of an imaging hard X-ray all-sky survey remained high, with no other planned astrophysical mission with similar objectives on the horizon. With the help of the Semiconductor Laboratory (HLL) for the production of high-sensitivity detectors of the MPE, a new project to further develop the highly successful \textit{XMM-Newton} pn-CCD technology was launched. In March 2002, all participating institutes proposed to ESA to accommodate the ROentgen Survey with an Imaging Telescope Array (ROSITA) telescope on an external platform of the International Space Station (ISS). The seven mirror modules would have been built identically to those of \textit{ABRIXAS,} but each of them would have been equipped with its own newly developed, frame-store pn-CCD in the focal plane. In September 2002, this proposal was supported by ESA with the highest scientific priority, and was recommended for phase A study. Launch would only have been possible from 2011 onward, however, because the external platforms of the ISS was occupied before then. Subsequently, it emerged that the ISS would not have been a viable option for ROSITA, first of all because NASA decided to terminate its shuttle flights to the ISS in 2010, and furthermore, because a contamination experiment on the ISS revealed that its environment was not safe from contamination for the sensitive X-ray mirrors and detectors of ROSITA. At the turn of the millennium, the observations of supernovae type Ia by two independent groups (awarded with the Nobel Prize for Physics in 2011 for their discoveries) revealed that the expansion of the Universe is accelerating, a fact that may suggest the existence of a cosmological constant. The subsequent measures, taken by Boomerang and \textit{WMAP}, of the tiny temperature fluctuations of the microwave background radiation out of which galaxies, clusters of galaxies, and the overall large-scale structure of the Universe originated, pointed toward a flat Universe, whereby the total matter plus energy content of the cosmos attains almost exactly the critical value. In fact, studies of the baryon fraction in X-ray selected (mainly from \textit{ROSAT}) clusters of galaxies throughout the 1990s had already shown compelling evidence that the matter density was lower than unity \citep{2003Schuecker}. The fact that a very large sample of clusters of galaxies can be particularly useful for precision cosmology stimulated many different groups to conceive dedicated large-area cluster surveys. In April 2003, members of the ROSITA team participated in the proposal of DUO, a NASA SMEX-mission, based on a modification of the ROSITA telescope design. Together with another four missions, DUO was selected among 36 competing proposals for a phase-A study, carried out in 2004. DUO was to survey an area of the sky of about {6,000} deg$^2$, overlapping that explored by the optical Sloan Digital Sky Survey (SDSS). In this way, about {10,000} clusters of galaxies could have been detected, providing constraints on the dark energy density to an accuracy of within 10\%. NASA did not select the DUO-project for further development, however, and the only mission of the five that was finally executed within the SMEX program was the hard X-ray focusing telescope \textit{NuSTAR}, which was launched in 2012. \begin{table*} \caption{Basic eROSITA instrument parameters in launch configuration} \label{tab:erosita} \centering \begin{tabular}{|c c|c c|c c|} \hline \multicolumn{2}{|c|}{Instrument} & \multicolumn{2}{|c|}{7 Mirror Assemblies} & \multicolumn{2}{|c|}{7 Camera Assemblies}\\ \hline Size & 1.9 $\diameter$ $\times$ 3.5\,m & Diam. of outer shell & 358\,mm & CCD image & $2.88 \times 2.88$ cm$^2$ [$1\fdg{}03 \times 1\fdg{}03$] \\ Mass & 808\,kg & Number of shells & 54 & pixelsize & $75\mu\mathrm{m}\times 75\mu\mathrm{m}$ [$9\farcs{}6 \times 9\farcs{}6$]\\ Power & 522\,W max. & focal length & 1600\,mm & Time Resol. & 50 msec \\ {Datarate} & 600\,MB/day max. & HEW on axis (1.5\,keV) & $18''$ & Energy Resol & 70\,eV [1\,keV] \\ & & HEW FoV averaged & $26''$ & Qu-Efficiency & 95\% \\ \hline \end{tabular} \end{table*} In February 2005, the Astronomy and Astrophysics Advisory Committee (AAAC), founded by the National Science Foundation (NSF), NASA, and the Department of Energy (DOE), and the High Energy Physics Advisory Panel (HEPAP), founded by NSF and DOE, established a Dark Energy Task Force (DETF) with the task of advising NSF, NASA, and DOE on the optimal strategies for the future of dark energy research. In particular, the DETF evaluated and compared various ground- and space-based instrumentation and observational methods. A dedicated white paper \citep{Haiman_2005} reached the conclusion that within available technologies at the time, it would have been possible to gather a sample of about 100,000 X-ray selected clusters of galaxies that would have provided very stringent constraints on the fundamental parameters of the cosmological model of the Universe. The construction of a sample of galaxy clusters of this magnitude is the primary objective of the eROSITA\ telescope. To achieve this goal, the \textit{ABRIXAS} mirror design had to be modified to include 27 additional shells, doubling the diameter of each telescope module and increasing the effective area at low energies by a factor of five. With this fundamental rescope, eROSITA\ will likely be the first stage IV dark energy probe to be realized and will outperform the DUO-like stage IV probe originally considered by the DETF. In June 2006, the funding proposal for eROSITA\ was submitted to the German Space Agency (DLR). Five German institutes (MPE, IAAT, AIP, the Hamburg Observatory, and the Dr. Remeis-Sternwarte in Bamberg, the Astronomical Institute of the Erlangen-N\"urnberg University) agreed to work together to develop, build, and organize the scientific exploitation of the instrument and formed the German eROSITA\ Consortium. They were later joined by three more institutes: the Argelander-Institut f\"ur Astronomie at the University of Bonn, the Max-Planck-Institut f\"ur Astrophysik (MPA), and the Universit\"ats-Sternwarte M\"unchen (LMU). In March 2007, DLR approved the project and funded the eROSITA\ telescope, while a Memorandum of Understanding was signed between DLR and the Russian Space Agency (Roskosmos) to ensure that eROSITA\ would be launched in the framework of the {\it Spektrum-Roentgen-Gamma} mission. Soon afterward, the construction of the telescope began. In September 2008, Roskosmos came to a final decision on the payload orbit and launcher. eROSITA\ would be launched by a Zenit-Fregat launcher (later in 2016 to be replaced by a Proton-M and Block-DM combination) together with the Russian hard-X-ray telescope ART-XC\ on an L2 orbit to ensure maximum efficiency for a survey instrument. Eventually, additional funds from the Max Planck Society and DLR were secured for eROSITA\ in July 2009, to finance the cost-driving L2 orbit. A detailed agreement between Roskosmos and DLR was signed the following month. In the following years, the teams in Germany and Russia were intensively kept busy in the development and construction of the instrument. The mirror modules, smaller versions of the \textit{XMM-Newton} mirrors, were supposed to be built using the same technology and by the same industrial consortium. However, there were significant problems with the shorter focal length and the smaller mirror shells, as figuring errors have a significantly greater effect on the image quality. The mirror engineering model had only a 42 arcsec HEW, almost three times lower than required. Fixing these issues took more development than expected, but was ultimately successful. Another major challenge was the development of the electronics: after the decision for the L2 orbit, the basically completed design had to be changed to use radiation hard components. All qualification tests were carried out in the MPE test facilities or in the facilities of companies in the Munich area. eROSITA\ was finally delivered to Russia in January 2017. A series of tests was then conducted at NPOL before integration of the instruments into \textit{SRG} could begin. This was then completed in April 2019 with the transport to the Baikonur Cosmodrome. \begin{figure} \centering \includegraphics [width =0.95\columnwidth,clip]{A3391_95_clumps_blue.png} \caption{eROSITA\ 0.3--2.3\,keV wavelet-filtered X-ray image of the Abell 3391/95 system, also showing several clumps of diffuse gas at the same distance as well as several background galaxy clusters. One degree corresponds to 3.9\,Mpc at the redshift of A3391. From \citet{reiprich21}.} \label{fig:a3391} \end{figure} \subsection{Instrument} \label{s:erosita_instrument} eROSITA\ (Fig.~\ref{fig:erosita}) consists of seven identical and coaligned X-ray telescopes housed in a common optical bench in a hexagonal shape. A system of carbon fibre honeycomb panels connects the seven mirror assemblies on the front side with the associated seven camera assemblies on the focal plane side. The optical bench is connected to the S/C bus through a hexapod structure. Each of the mirrors comprises 54 paraboloid or hyperboloid mirror shells in a Wolter-I geometry, with an outer diameter of 360 mm and a common focal length of 1\,600 mm \citep{Friedrich2008,Arcangeli2017}. The average on-axis resolution of the seven mirror modules as measured during the on-ground calibration is $16.1''$ half-energy width (HEW) at 1.5\,keV. The unavoidable off-axis blurring typical of Wolter-I optics is compensated for by a 0.4~mm shift of the cameras toward the mirrors. This places each telescope slightly out of focus and leads to a slight degradation of the on-axis performance ($18''$), but improves the angular resolution averaged over the FoV. Each mirror assembly has a CCD camera in its focus \citep{Meidinger2014}. The eROSITA\ CCDs are advanced versions of the EPICpn CCDs on \textit{XMM-Newton}. They have $384\times384$ pixels in an image area of $28.8\,\mathrm{mm} \times 28.8\,\mathrm{mm}$, yielding a square FoV of $1\fdg{}03 \times 1\fdg{}03$. Each pixel corresponds to a sky area of $9\farcs{}6 \times 9\farcs{}6$. The nominal integration time for all eROSITA\ CCDs is 50\,msec. The additional presence of a frame-store area in the CCD substantially reduces the amount of so-called out-of-time events, which are recorded during the CCD read-out, a significant improvement with respect to the PN camera on \textit{XMM-Newton} \citep{Strueder2001}. For optimal performance during operations, the CCDs are cooled down to about $-85^\circ$ by means of passive elements \citep{Fuermetz2008}. For calibration purposes, each camera has its own filter wheel with a radioactive $^{55}$Fe source and an {aluminum--titanium} target. This provides three spectral lines at 5.9\,keV (Mn-K$\alpha$), 4.5\,keV (Ti-K$\alpha$), and 1.5\,keV (Al K$\alpha$). The electronics for onboard-processing of the camera data is provided by seven sets of camera electronics (CE), each one mounted and interfaced to the cameras. Each of the CEs provides the proper voltage control and readout timing of the associated camera and performs the onboard data processing within the time constraints of the camera integration time. \begin{figure*} \centering \includegraphics [width = 0.45\textwidth,clip]{erass1_rgb_3x3deg.png} \includegraphics [width = 0.45\textwidth,clip]{erass2_rgb_3x3deg.png} \caption{False-colour images of the {the first eROSITA\ all-sky survey (eRASS1, left) and the second eROSITA\ all-sky survey (eRASS2, right)} observations in the energy bands 0.2--0.6\,keV (red), 0.6--1.0\,keV (green), 1.0--2.3\,keV (blue), adaptively smoothed. The size of the images is 3 degrees $\times$ 3 degrees. The circular structure is a dust scattering ring, echoing the outburst of the X-ray binary MAXI J1348-630. From \citep{lamer21}.} \label{fig:ring} \end{figure*} The Interface and Thermal Controller (ITC) receives the telemetry generated by each CE and stores it in the mass memory, commands each of the CEs, and controls the power distribution and the temperatures of the mirrors and cameras. Given its criticality, the ITC is a cold redundant unit \citep{Coutinho2018}. Finally, two (redundant) star trackers are mounted on eROSITA\ for the determination of an accurate boresight. The dimensions of the telescope structure are approximately 1.9\,m (diameter) $\times$ 3.2\,m (height in launch configuration, with closed front cover). The total weight of eROSITA\ is 808\,kg. Table~\ref{tab:erosita} summarizes the basic eROSITA\ instrument parameters. \begin{table*} \caption{Summary of performance characteristics of the eROSITA\ telescope and its survey sensitivity. The background counts are based on the first all-sky survey data. For eRASS:1 and the PV eFEDS 140 deg$^2$ survey { \citep{2021arXiv210614517B}}, the flux sensitivity in each band has been computed by taking all sources detected above a likelihood of 8 (soft band) or 10 (hard band), and measuring the flux below which the logarithmic number counts start declining. For eRASS:8 the predictions are based on detailed simulations that include all instrumental effects and particle background intensity consistent with that measured at L2. For each field or region, we quote the total (unvignetted) exposure in seconds. The corresponding effective (vignetted) exposures can be computed by dividing the total exposure by 1.8 and {3.31} for the soft and hard band, respectively. } \label{tab:erosita_sens} \renewcommand{\arraystretch}{1.3} \begin{tabular}{|c|c|c|c|} \hline \multicolumn{2}{|c|}{} & \multicolumn{2}{c|}{Energy Range} \\ \cline{3-4} \multicolumn{2}{|c|}{} & Soft Band & Hard Band \\ \hline \multicolumn{2}{|c|}{} & 0.2--2.3\,keV & 2.3--8\,keV \\ \hline \multicolumn{2}{|c|}{FoV averaged effective area [cm$^2$]} & 1,365 @ 1keV & 139 @ 5\,keV \\ \hline \multicolumn{2}{|c|}{Total Background [10$^{-3}$ cts/s/arcmin$^2$]} & $\approx$ 3.7 & $\approx$ 2.1 \\ \hline \multicolumn{4}{l}{Point source sensitivity eRASS:1} \\ \hline Ecliptic Equatorial region & Total exposure [s] = 200 & $5 \times 10^{-14}$\,erg/s/cm$^2$ & $7 \times 10^{-13}$\,erg/s/cm$^2$ \\ \hline Ecliptic Polar region & Total exposure [s] = 4000 & $7 \times 10^{-15}$\,erg/s/cm$^2$ & $9 \times 10^{-14}$\,erg/s/cm$^2$ \\ \hline \multicolumn{4}{l}{Point source sensitivity eFEDS} \\ \hline eFEDS field & Total exposure [s] = 2500 & $9 \times 10^{-15}$\,erg/s/cm$^2$ & $1.3 \times 10^{-13}$\,erg/s/cm$^2$ \\ \hline \multicolumn{4}{l}{Point source sensitivity eRASS:8 (predicted)} \\ \hline Ecliptic Equatorial region & Total/Effective exposure [s]= 1600 & $1.1 \times 10^{-14}$\,erg/s/cm$^2$ & $2.5 \times 10^{-13}$\,erg/s/cm$^2$ \\ \hline Ecliptic Polar region & Total exposure [s] = 30000 & $2.5 \times 10^{-15}$\,erg/s/cm$^2$ & $4 \times 10^{-14}$\,erg/s/cm$^2$ \\ \hline \end{tabular} \end{table*} \subsection{Early operations, first results, and performance} The commissioning phase of eROSITA\ had the objective of switching on all subsystems to verify they were functional following the launch and that they performed as expected to fulfill the scientific objectives. This phase served not only to test and commission the complete eROSITA\ telescope, but also for the teams in Khimky and Moscow (NPOL, IKI) and Garching (MPE) to learn and update the procedures on how to safely operate the spacecraft and the telescopes in space. The first mission-critical event was the switch-on of the ITC, which had to happen less than four hours after launch, to enable the thermal control of mirrors and electronics. The second event was the opening of the telescope cover, which occurred on July 22. Camera switch-on, instead, had to wait several days before being activated to avoid excess contamination from the first two spacecraft burns, which occurred on days 10 and 25 of the mission. In addition, camera cooling could not be performed without a minimum of 21 days of out-gassing following the cover opening. The commissioning of the cameras lasted about two months, including time necessary to perform a series of tests of the functionality of the camera electronics and thermal balance system. All seven eROSITA\ X-ray telescope modules had been observing the sky simultaneously since October 13. Over the following eight weeks, eROSITA\ collected its first-light images and performed a series of observations designed to accurately calibrate the instruments and verify that the performance of the telescope met pre-launch expectations. The Russian and German science teams jointly defined this CalPV program, which included a combination of pointed observations, field scans, and full-circle scan tests. Figures~\ref{fig:a3391} and \ref{fig:ring} show two examples of images captured by eROSITA\ during the CalPV and early survey phase. They highlight the main properties of this unique X-ray telescope, namely, its ability to take deep images, which are highly sensitive to both point-like and diffuse emission, over very large areas of the sky. As a preview of the eROSITA\ capabilities, a mini-survey called eROSITA\ Final Equatorial Depth Survey (eFEDS) was devised as part of the PV plan, imaging a 140 deg$^2$ patch of the sky to a depth comparable to that expected at the end of the all-sky survey (see Table~\ref{tab:erosita_sens}). These data confirm the sensitivity of the X-ray telescope to its main target classes with high accuracy. This mini-survey revealed more than 20,000 point-like X-ray sources, about 80\% of which are distant AGN harboring growing supermassive black holes. Most of the remainder are X-ray active stars. Finally, on December 13, 2019, the all-sky survey began. It was completed on June 12, 2020, after 182 days of almost continuous scanning of the sky. Eight all-sky surveys are planned in total, each delivering an average exposure with eROSITA\ of about $200 \mathrm{s}/cos(lat)$, where $lat$ is the ecliptic latitude, while the area of 1\,square degree around the ecliptic poles is revisited every four hours. This accumulates an exposure of about 30~ks per survey. During the course of the all-sky survey, the spacecraft has been rotating continuously with a scan rate of 90 deg hr$^{-1}$, giving a four-hour period per revolution. The rotation axis is oriented to the neighborhood of the sun, with an average progression in ecliptic longitude of about 1 deg day$^{-1}$, thus completing one all-sky survey in half a year. The scan speed guarantees that the angular resolution is not degraded by smearing of photons during the 50~ms CCD read-out cycle, and it provides sufficient overlap between individual scans to enable source variability analysis and homogeneous survey exposure. A preliminary analysis indicates that more than one million X-ray point sources and about 20,000 extended sources are detected in the survey. This is comparable to and may indeed exceed the total number of X-ray sources known before eROSITA. About 80\% of the point sources are distant AGN (comprising ~80\% of all known blazars, among others), and 20\% are coronally active stars in the Milky Way. In summary, during its first year of operations in space, most technical, operational, and scientific design characteristics of the eROSITA\ instrument onboard \textit{SRG}\ have been validated. Table~\ref{tab:erosita_sens} describes the main performance characteristics of eROSITA\ based on the data collected in this period during the PV phase and the all-sky survey. Compared to the pre-launch estimates of \citet[table 4.4.1 there]{Merloni2012}, the performance closely matches the expectations in the soft energy band, while it is slightly poorer in the hard band, mainly because the level of particle background is higher. \section{History of the SRG project in Russia} \label{s:history} \subsection{International X-ray astronomy projects in Russia} The history of international X-ray astronomy space projects in the USSR and Russia began with the {\it Roentgen} observatory on the {\it Kvant} module of the {\it Mir} space station. This project had been proposed by the Space Research Institute (IKI) of the USSR Academy of Sciences for implementation in the framework of the Soviet Intercosmos program. A number of European institutions were invited to this project: (i) the University of Birmingham (UK) and the SRON Netherlands Institute for Space Research, which built the TTM coded-mask-aperture X-ray telescope, sensitive in the 2--25\,keV energy band, (ii) the MPE (Germany), which provided the HEXE hard X-ray spectrometer, sensitive in the 20--120\,keV energy band, (iii) the European Space Research and Technology Center (ESTEC) of ESA, which built the GSPS high-pressure-gas spectrometer. IKI constructed the Pulsar X-1 hard X-ray detector for this project. The {\it Kvant} module was put into orbit by a Proton rocket and successfully docked with the space station in April 1987. The main scientific result of the {\it Roentgen} observatory was the discovery in August 1987 of hard X-ray emission with an unusual spectrum from the very bright supernova SN 1987A, which had exploded five months earlier in the Large Magellanic Cloud \citep{Sunyaev_1987a,Sunyaev_1987b}. The detected photons were emitted as gamma-ray lines associated with the decay of radioactive cobalt 56. During the passage through the optically thick shell of the supernova, these photons experienced multiple Compton scatterings off relatively cold electrons and lost their energy due to recoil. As a result, an extremely hard power-law X-ray spectrum formed, and this spectrum was detected for several months by HEXE and Pulsar X-1. At energies below 20\,keV, photoabsorption of photons by ions of heavy elements, first of all iron and cobalt, came into play. This explained the absence of a detectable signal by TTM. In 1987--1995, TTM discovered many transient and persistent sources, which now wear the names ``KS'' ({\it Kvant} source, \citealt{sunyaev91a}). The availability of instruments sensitive in different bands of the X-ray spectrum allowed broadband spectra in the 2--200\,keV energy range to be constructed for very many bright transient and persistent sources in binary stellar systems, including black holes and neutron stars (X-ray pulsars and neutron stars with weak magnetic fields) \citep{sunyaev91b,sunyaev94}. It was demonstrated for the first time how strongly the spectra of X-ray binaries depend on the nature of the accreting object. TTM provided excellent timing observations of X-ray pulsars \citep{gilfanov89} and high-quality X-ray images of the Galactic center region. The second X-ray orbital observatory of IKI within the Intercosmos program was the {\it GRANAT} satellite, built by NPO Lavochkin and put into an elongated elliptical orbit around Earth with a four-day period by a Proton rocket. It operated in orbit from December 1989 to May 1999. The payload of {\it GRANAT} included the SIGMA hard X-ray and gamma-ray telescope with a germanium spectrometer, the PHEBUS gamma-ray burst detectors developed in France, the WATCH all-sky monitor (Denmark), and the ART-P instrument constructed by IKI and its subsidiary in Frunze (now Bishkek, Kyrgyzstan), sensitive in the 2--25\,keV energy band. Both SIGMA and the ART-P telescopes had a coded mask and position-sensitive detectors, permitting construction of the images. The main results of the {\it GRANAT} mission include the maps of the central region of the galaxy in hard X-rays by the SIGMA (40--100\,keV) and ART-P (3--30\,keV) telescopes \citep[e.g.,][]{Churazov_1994,1993ApJ...418..844G,Pavlinsky_1994}. WATCH discovered the extremely interesting black hole X-ray binary GRS~1915$-$105 (``GRS'' means that it is a {\it GRANAT} source, \citealt{Castro_1992}). The spatial correlation of the 8--22\,keV X-ray emission observed by ART-P with the distribution of molecular gas in the Galactic center region became the first indication of the X-ray echo of the past Sgr~A* X-ray activity \citep{1993ApJ...407..606S} that is scattered by H$_2$ molecules in the dense gas surrounding it, according to the prediction by \cite{1980SvAL....6..353V}. PHEBUS observations helped to clearly demonstrate \citep{Tkachenko_2002} the existence of two types of gamma-ray bursts: short and normal, which, as we know now, is closely related to the parameters of the stellar objects that are responsible for their appearance (merger of neutron stars in the case of short bursts and collapse of massive stars in the case of normal bursts). In 2002, according to the agreement between ESA and Roscosmos, a Proton rocket placed the {\it International Gamma-ray Laboratory} ({\it INTEGRAL}) into a high-apogee orbit. {\it INTEGRAL} continues to provide scientific data to this day. Russian scientists are granted 25\% of the observing time of this mission. The prominent results of {\it INTEGRAL} include the detailed spectroscopy of the electron-positron annihilation radiation \citep{churazov05} and the first detection of gamma-ray lines from the type Ia supernova SN2014J \citep{churazov14}. \subsection{Beginning of the \textit{SRG} project in the USSR} In 1987, the thirtieth anniversary of the launch of the first artificial satellite of Earth ({\it Sputnik}) was celebrated. On this occasion, a large international meeting was held at IKI, where Intercosmos announced a competition for the payload of two orbital observatories. This resulted in a decision to support projects {\it Radioastron} and {\it Spectrum-Roentgen-Gamma} (\textit{SRG}). The \textit{SRG}\ project was supported by a number of renowned Soviet physicists, including Ya.B. Zeldovich, A.D. Sakharov, and the IKI director R.Z. Sagdeev. In addition to the High Energy Astrophysics Department of IKI, which proposed the \textit{SRG}\ project, it was joined by Denmark, the UK, Italy, MPE (Germany), NASA (USA), Finland, Switzerland, Israel, and Turkey. The work on the project was in full swing (see Fig.~\ref{fig:old_srg}), but huge changes in the USSR led to a slowdown in the works on the spacecraft and instruments in 2001--2002, after which the project was terminated. \begin{figure} \centering \includegraphics[width=0.98\columnwidth,clip]{old_srg.png} \caption{Former \textit{SRG}\ project.} \label{fig:old_srg} \end{figure} In 2003, the Space Council of the Russian Academy of Sciences considered and supported a more modest project with a launch by a Soyuz rocket (instead of a Proton), a smaller number of instruments, and with a shift in the scientific goals toward cosmology, fine X-ray spectroscopy, and monitoring the whole sky in X-rays. The idea of attracting international cooperation to the project was also supported. The project was then considered by Roscosmos, and a key decision was made to elaborate it further. It was also decided to retain (for the sake of continuity) the former project's name \textit{SRG}, although the new project did not foresee the presence of instruments sensitive in the gamma-ray range. Among the first invited to discuss a possible participation in the project were scientists from the MPE in Germany, with whom the High Energy Astrophysics Department of IKI has closely worked during the {\it Roentgen} mission on {\it Mir}/{\it Kvant} and on the first version of {\it SRG}. Moreover, MPE had developed and built the X-ray telescope for the extremely successful {\it ROSAT} satellite, which obtained wonderful X-ray maps of the sky in 1990. MPE proposed using a modification of the telescope that was developed by MPE for the unfortunately failed {\it ABRIXAS} mission for observations of clusters of galaxies (the most massive objects in the Universe, which are of interest to the Russian side as well). This telescope and its detector were designed for exploring the sky in a significantly harder energy band (from 0.5 to 10\,keV) compared to {\it ROSAT} (see a more detailed discussion in \S\ref{s:erosita_history}). By this time, the number of massive clusters of galaxies required for the detection of baryonic acoustic oscillations in their spatial distribution had been calculated at the Max Planck Institute for Astrophysics (MPA) and was proved to be close to 100,000 \citep{Hutsi_2006}. This would require finding virtually all massive clusters of galaxies in the observable Universe. This question was actively discussed at MPA in connection with searches for the baryonic acoustic oscillations predicted by theorists \citep{Peebles_1970,Sunyaev_1970} and with emerging hopes to find a large number of galaxy clusters through the Sunyaev--Zeldovich effect using ground-based telescopes and spacecraft operating in the microwave band. Therefore the question arose about a telescope with a higher sensitivity and working in a softer X-ray energy band compared to the existing modifications of the instrument for the {\it ABRIXAS} satellite, which could not provide the necessary sensitivity to fulfill this task, nor the majority of other tasks that were of interest for cosmologists at the time. After long discussions, the Russian side agreed on drastic increases in the size, mass, (800\,kg) and power consumption of the German instrument to fulfil the requirements (according to elaborations by G. Hasinger and P. Predehl) posed by the planned scientific results of the sky survey. The parameters of the Russian ART-XC instrument were also growing in parallel. On March 23, 2007, Roscosmos and DLR signed a memorandum about the inclusion of the eROSITA\ instrument into the payload of the \textit{SRG}\ orbital observatory. This opened the way to intensive work on the project and the telescope. \subsection{Data sharing between the participants of the \textit{SRG}\ project} All data obtained by the ART-XC\ telescope belong to the scientists of IKI, who have developed the telescope. The data within an area of 200\,square~degrees around the north ecliptic pole (where the sensitivity of the all-sky survey is highest) are processed together by scientists at IKI and the MSFC, in exchange for the delivery by the latter of a subset of the ART-XC\ X-ray mirrors. According to the memorandum of 2007 between Roscosmos and DLR, the \textit{SRG}/eROSITA\ data belong in equal parts to the German and Russian scientists. Because the primary goal of \textit{SRG}/eROSITA\ consists of constructing all-sky X-ray maps and creating X-ray source catalogs, a decision has been made that the German scientists are responsible for the processing and publication of data in one hemisphere of the sky, and the Russian scientists provide this in the other hemisphere (see Fig.~\ref{fig:division}). \begin{figure} \centering \includegraphics[width=0.98\columnwidth,clip]{sky_division.pdf} \caption{Division of the \textit{SRG}/eROSITA\ data between the German and Russian consortia (denoted MPE and IKI, respectively). The border vertically follows in Galactic coordinates through the central supermassive black hole Sgr~A*.} \label{fig:division} \end{figure} \section{Conclusion} The \textit{SRG}\ observatory has been operating in orbit for almost two years at the time of writing. The ART-XC\ and eROSITA\ telescopes have already scanned 40\% of the entire sky for the fourth time. The experience of 25 months of operation of \textit{SRG}\ allows us to say that the initial plans of the observatory are being successfully implemented and the number of extragalactic objects being discovered corresponds to the expectations before the launch. Using the new data obtained in the fourth scan of the sky, it is possible to search for sources whose X-ray brightness has strongly changed over the period of 6--18 months that passed since their previous appearances in the telescopes' FoV during the first three all-sky surveys. Every day, the observatory discovers about$\text{ } \text{five to ten}$ objects within a $360^\circ \times 1^\circ$ strip (i.e., less than 1\%) of the sky whose brightness has increased more than ten times over 6 months. The depth of the X-ray maps and the number of detected sources continues to increase with increasing exposure. The observatory is continuing its round-the-clock mission. \section{Acknowledgements} This work is based on observations with the eROSITA\ and ART-XC\ telescopes aboard the \textit{SRG}\ observatory. The \textit{SRG}\ observatory was built by Roskosmos in the interests of the Russian Academy of Sciences represented by its Space Research Institute (IKI) in the framework of the Russian Federal Space Program, with the participation of the Deutsches Zentrum f\"ur Luft- und Raumfahrt (DLR). The \textit{SRG}/eROSITA\ X-ray telescope was built by a consortium of German Institutes led by MPE, and supported by DLR. The \textit{SRG}\ spacecraft was designed, built, launched, and is operated by the Lavochkin Association and its subcontractors. The science data are downlinked via the Deep Space Network antennas in Bear Lakes, and Ussurijsk, funded by Roskosmos. The eROSITA\ data used in this work were processed using the eSASS software system developed by the German eROSITA\ consortium and proprietary data reduction and analysis software developed by the Russian eROSITA\ Consortium. \bibliographystyle{aa}

80bea9e3e35a42cd291a3f923a55abe759b70ef9

\section*{Measurement details} The filtering was performed using \myq{4}{nF}/\myq{0.2}{\mu H} $\pi$-filters at room temperature, followed by the N12-50F-257-0 flexible stainless steel coaxial cables, thermalized on the $4\text{K}$ and still plates, with total resistance of \myq{300}{Ohm} down to the coldplate. At coldplate, the coaxial cables were connected to the cinch connector on closed copper box which contained the chip. Inside the box, the sample was connected via twisted copper pairs additionally thermalized with a silver paste to a massive rod. Together with grounding~\myq{10}{nF} capacitors near the chip, the stainless coaxial cables also served as low-pass RC-filters. \section{Device Fabrication} For patterning of the mesa we used e-beam lithography followed by wet mesa etching in an aqueous solution of KI:I$_2$:HBr. Further, Ti/Al contacts were evaporated right after \textit{in situ} cap removal with the Ar gun. The \myq{70}{nm} SiO$_2$ gate dielectric is then magnetron sputtered followed by Ti/Al gate electrode deposition. During fabrication the heating of the substrate was carefully controlled with the highest temperature of $80^\circ$C utilized for resist baking. \newpage \section*{Nonlocal measurements} \begin{figure}[h] \begin{center} \vspace{0mm} \includegraphics[width=6in]{fig2_sup1_width=6in.pdf} \end{center} \caption{Nonlocal transport measurements at~\myvareq{T}{}{4.2}{K}. Voltage probed at the terminals located aside from the bulk current path. (a)~\vg tuned to CNP region, so the bulk is insulating. Then terminals probe electrical potential of edge states, visualizing current flowing along the boundary of the sample. (b)~At \vgeq{0} 2D electron gas dominates the transport making voltage measured on terminals 1-5 negligible.} \label{figSM1} \end{figure} \begin{figure}[h] \begin{center} \vspace{0mm} \includegraphics[width=3.1in]{fig1_sup1_width=3.1in.pdf} \end{center} \caption{False colored SEM micrograph of the device~D2. Studied edges are indicated by red lines captioned the same way they were mentioned in the text. } \label{figSM_edges} \end{figure} \clearpage \section*{Localization} \begin{figure}[h] \begin{center} \vspace{0mm} \includegraphics[width=6in]{fig3_sup1_width=6in.pdf} \end{center} \caption{Temperature dependence of edge transport. (a-e)~$R(V_\text{g})$ plots for five edges of different lengths. (f)~Temperature dependence of the mean conductance in the CNP region. For 5~$\mu$m edge black dashed line represents activation-like behavior with $\Delta=16~\mu\text{eV}$ exponent.} \label{figSM2} \end{figure} \clearpage \section*{Differential resistance} \begin{figure}[h] \begin{center} \vspace{0mm} \includegraphics[width=4.5in]{fig4_sup6_width=4.5in.pdf} \end{center} \caption{Differential resistance of two short edges obtained at different \vg values inside CNP region at 0.5~K. (a)~\myq{0.5}{\mu m} and (b)~corner edges of D2 device.} \label{figSMIV} \end{figure} \clearpage \section*{Magnetotransport data} \begin{figure}[h] \begin{center} \vspace{0mm} \includegraphics[width=6.5in]{fig4_sup5_width=6.5in.pdf} \end{center} \caption{Magnetic field dependence of the edge resistance at 4.2~K. (a)~\myq{0.5}{\mu m}, (b)~corner and (c)~$3\|3\,\mathrm{\mu m}$ edges of D2~device.} \label{figSM7} \end{figure} \begin{figure}[h] \begin{center} \vspace{0mm} \includegraphics[width=4in]{fig4_sup4_width=4in.pdf} \end{center} \caption{Influence of the external magnetic field on the conductance of the long edges at~\myvareq{T}{}{0.5}{K}. Conductance of the $20\,\mathrm{\mu m}$ edge (device D2) and of the $32\,\mathrm{\mu m}$ edge (two-terminal device D3 processed similarly to~D1 and~D2) measured at fixed~$V_{\text{g}}$ in the CNP region, as a function of the external magnetic field~$B$.} \label{figSMRG} \end{figure} \begin{figure}[h] \begin{center} \vspace{0mm} \includegraphics[width=5in]{fig4_sup3_width=5in.pdf} \end{center} \caption{Influence of the external magnetic field on the edge differential resistance at the CNP at~\myvareq{T}{}{0.5}{K}. Edges of different lengths of D2~device \myq{0.5}{\mu m}, $2\|2\,\mathrm{\mu m}$, $3\|3\,\mathrm{\mu m}$ and $20\|20\,\mathrm{\mu m}$ show qualitatively the same behavior with mobility gap opening on the order of hundreds $\mu\text{eV}$ (several mV) for shot (long) edges in $\sim 1$~T magnetic field.} \label{figSM4} \end{figure} \begin{figure}[h] \begin{center} \vspace{0mm} \includegraphics[width=4in]{fig4_sup2_width=4in.pdf} \end{center} \caption{Four consecutive $R(V_\text{g})$ measurements of \myq{0.5}{\mu m} edge of D2~device at 0.5~K in 1.7~T magnetic field. Mesoscopic oscillations are highly reproducible provided \vg sweep range is not too wide.}\label{figSM3} \end{figure} \begin{figure}[h] \begin{center} \vspace{0mm} \includegraphics[width=5in]{fig4_sup1_width=5in.pdf} \end{center} \caption{Reproducibility of $G_{\text{typ}}(B)$ obtained at different \vg values. $R(V_\text{g})$ measurements at 0.5~K and without external magnetic field applied (a)~\myq{0.5}{\mu m} and (b)~corner edges of D2~device. (c)~Magnetotransport data of \myq{0.5}{\mu m} edge normalized over $G_0 \equiv G_{\text{typ}}(0)$. Symbols on~(a) denote position respective to $R(V_\text{g})$ curve at which $G_{\text{typ}}(B)$ was measured. While low-$B$ data differs due to its strong sensitivity to measurement accuracy of $G_0$, high-$B$ data coincides for two \vg values. (d)~The similar data for corner edge. Here the importance of proper choice of \vg is clear: CNP should determined by the linear-response resistance peak in $B=0$ $G_{\text{typ}}^{-3.16}(0) > G_{\text{typ}}^{-3.36}(0)$, while in $B\gg0$ situation can be changed $G_{\text{typ}}^{-3.16}(1.7~\text{T}) \ll G_{\text{typ}}^{-3.36}(1.7~\text{T})$ ($4.5\cdot10^{-8}$~$\Omega^{-1}$ and $1.2\cdot10^{-6}$~$\Omega^{-1}$ respectively).} \label{figSM5} \end{figure} \end{document}

9d7b98972fe9397a2ffaf2376c9392e773ab6d91

\section{Introduction} Finite-dimensional observer-based controllers for parabolic systems were designed by the modal decomposition approach in \cite{balas1988finite, christofides2001, curtain1982finite,harkort2011finite

978c9c39688afe822c51040ba6a03c0132cefcef

\section{Introduction}\label{ntro} The transfers of Hecke elements play a fundamental role to study automorphic representations in Langlands correspondences. Thanks to many people including Waldspurger, Ngo among others, the existence of a transfer for any Hecke element is now guaranteed (cf. \cite{NgoICM}). However, to describe the transfer explicitly is a different matter. In this article we prove an explicit transfer theorem for the characteristic function of any principal congruence subgroup for any cyclic base change in the general linear group $GL_n$ over local fields of odd residual characteristic with a mild condition. The unramified case in the unit elements of Hecke algebras associated to hyperspecial open compact subgroups is already done by Kottwitz (see p.239 of \cite{Ko}) in more general setting and therefore, the ramified case has been remained to be open though some experts might have known. Let us set up some notation to explain our result. Let $p$ be an odd prime number and $F$ a finite extension of $\Q_p$. Let $E$ be a cyclic extension of $F$ with the generator $\theta$ of the Galois group ${\rm Gal}(E/F)$. Put $d=[E:F]$. Let $G={\rm Res}_{E/F}(GL_n/E)$ be the Weil restriction of the general linear group $GL_n/E$ to $F$. The action of $\theta$ is naturally extended to $G(F)=GL_n(E)$ and we also denote it by $\theta$ again. Let $\O_E$ (resp. $\O_F$) be the ring of integers of $E$ (resp. $F$) with a uniformizer $\varpi_E$ (resp. $\varpi_F$). We denote by $e=e(E/F)$ is the ramification index so that $\varpi^e_E$ and $\varpi_F$ differ by a unit in $\O_E$. For a positive integer $m$, put $K_E(m):=I_n+\varpi^m_E M_n(\O_E)$ where $I_n$ stands for the identity matrix of size $n$. Similarly, put $K_F(m):=I_n+\varpi^m_F M_n(\O_F)$. Let us fix Haar measures $dg$ on $G(F)$ and $dh$ on $H(F)$ respectively. Then we will prove the following: \begin{thm}\label{main1}Keep the notation as above. Let $m$ be a positive integer. Suppose that the residual characteristic $p$ of $F$ is odd and prime to $d=[E:F]$. The Hecke element ${\rm vol}^{-1}(K_E(m)_\theta)\textbf{1}_{K_E(m)_\theta}$ in $C^\infty_c(GL_n(F))$ is a transfer $($in the sense of Definition \ref{moi}$)$ of the Hecke element ${\rm vol}^{-1}(K_E(m))\textbf{1}_{K_E(m)}$ in $C^\infty_c(G(F))=C^\infty_c(GL_n(E))$ for the base change lift in $(G,GL_n/F)$. Here $\textbf{1}_C$ means the characteristic function of $C$, $K_E(m)_\theta$ is the $\theta$-fixed part of $K_E(m)$, and ${\rm vol}$ stands for the volume with respect to the fixed measure. \end{thm} This will be applied to the endoscopic classification of Siegel cusp forms (see Section 5 of \cite{KWY}) to make a detour considering quasi-split but not split orthogonal groups in Arthur's classification. \begin{remark} For a positive integer $m$, we write $m=ke+r$ with unique integers $k\ge 0$ and $0\le r<e$. Then we see easily that $$K_E(m)_\theta= \left\{\begin{array}{cl} K_F(k+1) & (\text{if $r\neq 0$}) \\ K_F(k) & (\text{if $r=0$}) \end{array}\right..$$ \end{remark} The following theorem is a more finer version of matching orbital integrals which is an extension of Proposition 3.1, p. 20 of \cite{AC} for the Hecke elements in Theorem \ref{main1}. \begin{thm}\label{main2} Keep the notation and the assumption on $F$ and $E/F$ in Theorem \ref{main1}. Let $m$ be a positive integer. Suppose that the residual characteristic $p$ of $F$ is odd and prime to $d=[E:F]$. Put $f^H:= {\rm vol}^{-1}(K_E(m)_\theta)\textbf{1}_{K_E(m)_\theta}$ in $C^\infty_c(H(F))= C^\infty_c(GL_n(F))$ and $f^G:={\rm vol}^{-1}(K_E(m))\textbf{1}_{K_E(m)}$ in $C^\infty_c(G(F))=C^\infty_c(GL_n(E))$. Put $\wf^G={\rm vol}^{-1}(K_E(m))\textbf{1}_{K_E(m)\rtimes\theta}$ in $C^\infty_c(\tG(F))$ which corresponds to $\tf^G$ under $G(F)\stackrel{\sim}{\lra} \tG(F),\ g\mapsto g\rtimes \theta ($see Section \ref{tc} for the twisted space $\tG)$. For a semisimple element $\gamma\in H(F)$ and a semisimple element $\delta\in G(F)$, we have the matching of normalized orbital integrals $$ I_{\gamma}(f^H,dh,dh_\gamma)= \left\{\begin{array}{cl} I_{\delta\rtimes \theta}(\wf^G,dg,dg_{\delta\rtimes\theta}) & \text{if $\gamma$ is a norm of $\delta$} \\ 0 & \text{if $\gamma$ is not a norm} \end{array}\right.. $$ See $($\ref{noiG1}$)$, $($\ref{noiG2}$)$, and $($\ref{noiH}$)$ for the normalized orbital integrals. \end{thm} \begin{remark}\label{on-assumption} Our assumption on the residual characteristic $p$ is crucial in the following points: \begin{enumerate} \item being odd is necessary to guarantee (\ref{homeo}); \item being prime to $d=[E:F]$ is necessary in the computation done in Section \ref{tc} through Section \ref{sdt} around topologically unipotent elements. In particular Theorem \ref{mainSSD} is important among others. \end{enumerate} \end{remark} \begin{remark}\label{current-situation-for-BC} \begin{enumerate} \item Thanks to works \cite{HI}, \cite{CR} and ours, an explicit transfer theorem for any base change lifts and automorphic inductions has been developed in great general so that we can apply these results to several arithmetic problems. \item Explicit transfer theorems for various settings of twisted endoscopy data, based on Lie-algebra calculation, are now available due to important works \cite{Fe}, \cite{GV}, and \cite{Oi} which embody Waldspurger's philosophy in \cite{Wal-book} saying ``Twisted endoscopy is not so twisted''. \end{enumerate} \end{remark} We give a global application. Let $K$ be a number field and $L/K$ be a cyclic extension of degree $d$. By abusing notation, let $\theta$ be a generator of ${\rm Gal}(L/K)$. Let $\pi=\pi_f\otimes \pi_\infty$ be a cuspidal automorphic representation of $GL_n(\A_K)$ where $\pi_f$ stands for the finite part of $\pi$. We denote by $\Pi:={\rm BC}(\pi)=\Pi_f\otimes \Pi_\infty$ the base change lift of $\pi$ to $GL_n(\A_L)$ (see Theorem 5.1 of \cite{AC}). Let $S_K$ be the set of all finite places $v$ of $K$ such that $L/K$ is ramified at $v$. Let $S_L$ be the set of all finite places $w$ of $L$ lying over some place in $S_K$. Let $S$ be a finite set of finite places of $L$ including $S_L$. For each $\underline{n}=(n_w)_{w\in S\setminus S_L}\in \Z^{|S\setminus S_L|}_{\ge 1}$ and each $\underline{m}=(m_w)_{w\in S_L}\in \Z^{|S_L|}_{\ge 1}$, put $$U_{S}(\underline{n},\underline{m}):= \ds\prod_{w\in S\setminus S_L}K_{E_w}(n_w)\times \prod_{w\in S_L}K_{L_w}(m_w)\times \prod_{w\not\in S}GL_n(\mathcal{O}_{L_w})$$ which is an open compact subgroup of $GL_n(\A^{(\infty)}_L)$ where $\A^{(\infty)}_L$ stands for the ring of finite adeles of $L$. Let $U_{S}(\underline{n},\underline{m})_\theta$ be the $\theta$-fixed part of $U_{S}(\underline{n},\underline{m})$ which can be regarded as an open compact subgroup of $GL_n(\A^{(\infty)}_K)$. \begin{thm}\label{main3} Assume that for each $v\in S_K$ the residue characteristic $p_v$ is odd and prime to $d_v=[L_w:K_v]$ for any finite place $w$ of $L$ lying over $v$. It holds that $${\rm dim}(\pi^{U_{S}(\underline{n},\underline{m})_\theta}_f) ={\rm dim }(\Pi^{U_{S}(\underline{n},\underline{m})}_f).$$ \end{thm} This paper is organized as follows. In Section \ref{tes}, according to the notation of \cite{KS}, we interpret base change lifts in terms of the twisted endoscopic lifts which has been itself well-known and not new at all. Since the author could not find any suitable reference, we state it more than necessary. In Section \ref{Norm} we recall the norm map (in Section 3 of \cite{KS}) for the base change lifts, equivalently, for our twisted endoscopic data. The author would find some explanation in Section 1 of \cite{AC}. However, we fit into the notation in \cite{KS} as much as possible. In Section \ref{tc} through Section \ref{oi} we study basic facts about (normalized) orbital integrals and recall the definition of semi-simple descend. We largely follow the argument in Ganapathy-Varma's paper \cite{GV}. In Section \ref{tf}, we compute the transfer factor $\Delta^{{\rm IV}}$ with ``upper IV''. The author really owe Professor Waldspurger a lot with email correspondences in the computation of $\Delta^{{\rm IV}}$. A key in our matching theorem is the semi-simple descend theorem whose proof is devoted to Section \ref{sdt}. In Section \ref{proof}, we give proofs of main theorems. \textbf{Acknowledgments.} This work has been raised toward an application to equidistribution theorems for $Sp_{2n}$ in \cite{KWY}. I really appreciate my collaborators (H-H .Kim ans S. Wakatsuki) to encourage me to prove the transfer theorem here. The author would also like to thank professors M. Oi, J-L. Waldspurger, and S. Yamana for helpful discussions. Professor Oi explained me about the importance of Ganapathy-Varma's paper \cite{GV} and also how the work of Waldspurger in the book \cite{Wal-book} should be understood. The special thanks are given to Professor Waldspurger again for explaining the author about how to compute the transfer factors in our setting. \section{A twisted endoscopic subgroup of $G={\rm Res}_{E/F}(GL_n/E)$}\label{tes} Keep the notation in the previous section. We follow the notation in Chapter 2 of \cite{KS}. For any algebraic group $\mathcal{G}$ over a field, we denote by $\mathcal{G}^0$ the connected component of $\mathcal{G}$. Recall the connected reductive group $G={\rm Res}_{E/F}(GL_n/E)$ over $F$ which is quasi-split over $F$ since ${\rm Res}_{E/F}(B/E)$ is a Borel subgroup defined over $F$ where $B$ stands for the upper Borel subgroup of $GL_n/F$ and $B/E$ does for its base change to $E$. Note that it is an outer form of $(GL_n/F)^d$. We fix the quasi split datum ${\rm spl}_G$ associated to ${\rm Res}_{E/F}(B/E)$ and its maximal torus $T$. We fix an isomorphism $\iota:{\rm Gal}(E/F)\stackrel{\sim}{\lra} \Z/d\Z,\ \sigma\mapsto \iota(\sigma)$ and extend it to the Weil group $W_F$ of $F$ via the natural projection $W_F\lra W_F/W_E\simeq {\rm Gal}(E/F)$. We also denote it by $\iota$ again. The Langlands dual group of $G$ is given by $\widehat{G}=\ds\prod_{i\in \Z/d\Z}GL_n(\C)$ with the action of $W_F$: \begin{equation}\label{lga} \sigma(g)=(g_{i+\iota(\sigma)})_{i\in \Z/d\Z},\ \sigma\in W_F,\ g=(g_i)_{i\in \Z/d\Z}\in \widehat{G}. \end{equation} It makes up the $L$-group ${}^L G:=\widehat{G}\rtimes W_F$. Let $\theta$ be a generator of $ {\rm Gal}(E/F)$. For any $F$-algebra $R$, the action of $\theta$ on $E\otimes_FR$ is given by $\theta(e\otimes r)=\theta(e)\otimes r$. It naturally induces an automorphism over $F$ of $G$ which is denoted by $\theta$ again. We denote by $\widehat{\theta}$ the automorphism of $\widehat{G}$ induced from the action of $\theta$ on $\widehat{G}$. The automorphism $\widehat{\theta}$ is acting on $\widehat{G}$ by $\widehat{\theta}((g_i)_{i\in\Z/d\Z})=(g_{i+\iota(\theta)})_{i\in \Z/d\Z}$ as in (\ref{lga}). Put ${}^L\theta=\widehat{\theta}\rtimes {\rm id}_{W_F}$. We consider the endoscopic datum $(G,\theta,\textbf{1}_{{\rm triv}})$ where $\textbf{1}_{{\rm triv}}$ is the trivial class in the Galois cohomology $H^1(W_F,Z(\widehat{G}))\simeq H^1(W_E,\C^\times)= {\rm Hom}_{{\rm ct}}(W_E,\C^\times)$. Here the subscript ``ct'' means continuous. The trivial class $\textbf{1}_{{\rm triv}}$ can be regarded as the trivial character $W_F\lra Z(\widehat{G})\subset \widehat{G}\subset {}^L\widehat{G}$ via the transfer map $W^{{\rm ab}}_F\lra W^{{\rm ab}}_E$. Further we extend it to $\mathcal{H}$ via the natural projection $\mathcal{H}\lra W_F$. We write $\textbf{1}_{{\rm triv}}:\mathcal{H}\lra {}^L\widehat{G}$ for the resulting trivial character. In view of the base change, the endoscopic datum $(H,\mathcal{H},s,\xi)$ for $(G,\theta,\textbf{1}_{{\rm triv}})$ consists of \begin{itemize} \item $H=GL_n/F$ with the Langlands dual $\widehat{H}=GL_n(\C)$; \item $\mathcal{H}=GL_n(\C)\rtimes W_F$ with the trivial action of $W_F$ on $GL_n(\C)$; \item $s=(1)_{i\in \Z/d\Z}\in \widehat{G}$ where $1$ stands for the unit element in $GL_n(\C)$; \item $\xi:\mathcal{H}\lra {}^L G,\ h\rtimes w\mapsto \diag(h)\rtimes w$ where $\diag:GL_n(\C)\lra (GL_n(\C))^d$ is the diagonal embedding. \end{itemize} Clearly, for each $h\rtimes w\in \mathcal{H}$, $${\rm Int}(s)\circ {}^L\theta\circ \xi(h\rtimes w)=\widehat{\theta}(\diag(h))\rtimes w =\diag(h)\rtimes w=\xi(h\rtimes w)=\textbf{1}_{{\rm triv}}(h\rtimes w)\cdot\xi(h\rtimes w).$$ The computation $${\rm Cent}_{\widehat{\theta}}(s,\widehat{G}):=\{g\in\widehat{G}\ |\ gs\widehat{\theta}(g)^{-1}=s \}= \{g\in\widehat{G}\ |\ \widehat{\theta}(g)=g \}=\{\diag(h)\ |\ h\in GL_n(\C)\}$$ shows $\xi|_{\widehat{H}}:\widehat{H}\stackrel{\sim}{\lra}{\rm Cent}_{\widehat{\theta}}(s,\widehat{G})$. Therefore, (2.1.4a) and (2.1.4b) in p.17, Chapter 2 of \cite{KS} are satisfied. Further, $\xi(Z(\widehat{H})^{W_F})^0=\{\diag(a\cdot 1)\ |\ a\in \C^\times\}$ is a subset of $Z(\widehat{G})=\{(a_1\cdot 1,\ldots,a_d\cdot 1)\in \widehat{G}\ |\ a_1,\ldots,a_d\in \C^\times\}$. In conclusion, $(H,\mathcal{H},s,\xi)$ is an elliptic, twisted endoscopic datum for $(G,\theta,\textbf{1}_{{\rm triv}})$ and in fact, it is easy to check that any elliptic, twisted endoscopic datum for $(G,\theta,\textbf{1}_{{\rm triv}})$ is isomorphic to our datum. Throughout this paper, we consider only this endoscopic datum $(H,\mathcal{H},s,\xi)$. In view of a global application, being elliptic is related to the discrete spectrum of the $L^2$-space of automorphic forms in \cite{AC}. \section{The norm map} \label{Norm} Let us keep the notation in the previous section. In this section we recall the norm map for our endoscopic datum $(H,\mathcal{H},s,\xi)$ for $(G,\theta,\textbf{1}_{{\rm triv}})$. We follow the notation in Chapter 3 of \cite{KS}. Let $Cl_{{\rm ss}}(H)$ be the set of semisimple conjugacy classes in $H(\oF)=GL_n(\oF)$ and write its element by $[\gamma],\ \gamma\in H(\oF)$. Two elements $g_1,g_2\in G(F)=GL_n(E)$ are said to be $\theta$-conjugate if there exists an element $g\in G(F)$ such that $g_1=g^{-1}g_2\theta(g)$. Let $Cl_{{\rm ss}}(G,\theta)$ be the set of $\theta$-semisimple $\theta$-conjugacy classes in $G(\oF)= GL_n(E\otimes_F\oF)$ (see Definition \ref{image}-(2)). We shall describe the canonical map $\mathcal A_{H/G}:Cl_{{\rm ss}}(H)\lra Cl_{{\rm ss}}(G,\theta)$ in Theorem 3.3.A, p.27 of \cite{KS}. Let $(B_H,T_H)$ be a pair of the upper Borel subgroup of $H$ and the diagonal maximal torus $T_H$ in $B_H$. Let $\Omega_H=N_{H}(T_H)/T_H$ be the Weyl group of $H$ with respect to $(B_H,T_H)$ and it acts on $T\simeq (GL_1/F)^n$ as the permutation of the entries via the natural identification $\Omega_H\simeq \mathfrak{S}_n$. Pick a class $C\in Cl_{{\rm ss}}(H)$. By taking the Jordan normal form, we see that $C=[t_C]$ for some $t_C\in T_H(\oF)$. Therefore, we have a bijection \begin{equation}\label{h-conj} Cl_{{\rm ss}}(H)\lra T_H(\oF)/\Omega_H,\ C\mapsto \widetilde{t}_C \end{equation} where $\widetilde{t}_C$ is the class of $t_C$ in $T_H(\oF)/\Omega_H$. Let $T={\rm Res}_{E/F}(T_H/E)$ be the maximal torus in the Borel subgroup $B={\rm Res}_{E/F}(B_H/E)$ in $G$. Recall that $\theta$ acting on $G$ is quasi-semisimple, hence, $T$ and $B$ are stable under $\theta$. We define the map $1-\theta:T\lra T,\ t\mapsto t\cdot \theta(t)^{-1}$. Then we have an isomorphism \begin{equation}\label{tt} T_\theta(\oF):=T(\oF)/(1-\theta)T(\oF)\lra T_H(\oF),\ (t_1,\ldots,t_d)\mapsto t_1\cdots t_d. \end{equation} where we use a fixed identification $T(\oF)\simeq (T_H(\oF))^d$ so that $\theta$ acts on $(T_H(\oF))^d$ by the shift operator $(t_1,t_2,\ldots,t_d)\mapsto (t_2,\ldots,t_d,t_1)$. Let $\Omega=N_G(T)/T$ be the Weyl group of $G$ with respect to $(B,T)$. As mentioned before, $\theta$ acts on $T$. If we fix an identification $\Omega\simeq \mathfrak{S}^d_n$, then the $\theta$-fixed part $\Omega^\theta$ is identified with the diagonal part $\Delta \mathfrak{S}_n:=\{(\sigma,\ldots,\sigma)\ |\ \sigma\in \mathfrak S_n \}$ of $\mathfrak{S}^d_n$. The map (\ref{tt}) naturally yields an isomorphism \begin{equation}\label{nn} \widetilde{N}_{E/F}:T_\theta(\oF)/\Omega^\theta\lra T_H(\oF)/\Omega_H. \end{equation} Pick an element $C$ in $Cl_{{\rm ss}}(G,\theta)$. By Lemma 3.2.A of \cite{KS}, the image of $C\cap T(\oF)$ in $T_\theta(\oF)$ is a single oribit of $\Omega^\theta$ and we write such an orbit as $\Omega^\theta \widetilde{s}_C,\ \widetilde{s}_C\in T_\theta(\oF)$ for $C$. Thus we have a bijection \begin{equation}\label{g-conj} Cl_{{\rm ss}}(G,\theta)\lra T_\theta/\Omega^\theta,\ C\mapsto \widetilde{s}_C. \end{equation} Putting everything together, we have \begin{equation}\label{norm} \mathcal{A}_{H/G}:Cl_{{\rm ss}}(H)\stackrel{(\ref{h-conj}) \atop\sim}{\lra} T_H(\oF)/\Omega_H \stackrel{\widetilde{N}^{-1}_{E/F}}{\lra} T_\theta(\oF)/\Omega^\theta \stackrel{(\ref{g-conj}) \atop\sim}{\lla} Cl_{{\rm ss}}(G,\theta). \end{equation} \begin{Def}\label{image} \begin{enumerate}\item For a semisimple element $\gamma\in H(F)=GL_n(F)$ we say $\gamma$ is a norm of $\delta\in G(F)=GL_n(E)$ if $\delta$ lies in $\mathcal A_{H/G}([\gamma])\cap G(F)$. Otherwise, we say $\gamma$ is not a norm. \item We say $\delta\in G(F)$ is $\theta$-semisimple if there exists $g\in G(F)$ such that $g^{-1}\delta \theta(g)$ lies in $T(F)=T_H(E)$. This is equivalent to enjoying the condition explained in Section 1.3, p.15 of \cite{KS}. \item We say a $\theta$-semisimple element is $\theta$-regular if $G_{\delta\rtimes \theta}^0={\rm Cent}_{\theta}(\delta,G)^\circ$ is a torus. and strongly $\theta$-regular if ${\rm Cent}_{\theta}(\delta,G)$ is abelian. In the latter case, the centralizer of $G_{\delta\rtimes \theta}$ in $G$ is a maximal torus stable under ${\rm Int}(\delta)\circ \theta$. Here $G_{\delta\rtimes \theta}$ is an $F$-algebraic group defined by the functor from $F$-algebras to sets; $$\underline{G}_{\delta\rtimes \theta}:\{F-algebras\}\lra Sets$$ $$R\mapsto \underline{G}_{\delta\rtimes \theta}(R)= G_{\delta\rtimes \theta}(R):= \{g\in G(R)=GL_n(E\otimes_F R)\ |\ g^{-1}\delta \theta(g)=\delta\}$$ where the action of $\theta$ on $g$ is induced by the action of $\theta$ on $E\otimes_F R$ given by $\theta(e\otimes r)=\theta(e)\otimes r$. \item We say a semisimple element $\gamma\in H(F)=GL_n(F)$ is strongly $G$-regular if it is a norm of a strongly $\theta$-regular element of $G(F)$. \end{enumerate} \end{Def} \begin{remark}\label{diagonal} Fix an isomorphism $T(\oF)\simeq (T_H(\oF))^d$ as before. If $\diag(s_1,\ldots,s_n)\in T_H(F)\subset GL_n(F)$ is a norm of $\delta=\diag(t_1,\ldots,t_g)\in T(F)=T_H(E)\subset GL_n(E)$, then as a multi-set $$\{N_{E/F}(t_1),\ldots,N_{E/F}(t_n)\}=\{s_1,\ldots,s_n\}$$ where $N_{E/F}:E\lra F$ is the usual norm map. \end{remark} \section{The twisted space and the twisted conjugation map }\label{tc} In this section we follow the notation in Section 3 and 4 of \cite{GV}. Let $p$ be the residue characteristic of $F$. For any reductive group $\mathcal{G}$ over $F$, we say an element of $g$ in $\mathcal{G}(F)$ is topologically unipotent if $\ds\lim_{n\to\infty}g^{p^n}=1$. Similarly, for an element $X$ in the Lie algebra ${\rm Lie}(\mathcal{G}(F))$ of $\mathcal{G}(F)$, we say $X$ is topologically nilpotent if $\ds\lim_{n\to\infty}X^{p^n}=0$. We denote by $\mathcal{G}(F)_{{\rm tu}}$ (resp. ${\rm Lie}(\mathcal{G}(F))_{{\rm tn}}$) the set of all topologically unipotent (resp. nilpotent) elements in $\mathcal{G}(F)$ (resp. ${\rm Lie}(\mathcal{G}(F))$). Let us return to our setting. Define the twisted space $\tG=G\rtimes \theta$ which is an $F$-variety. We have $G\stackrel{\sim}{\lra}\tG,\ g\mapsto g\rtimes \theta$ as an $F$-variety. We also define the twisted conjugation map: \begin{equation}\label{twcm} {\rm tc}:G(F)\times H(F)\lra \tG(F),\ (g,h)\mapsto g(h\rtimes \theta) g^{-1}=(gh\theta(g)^{-1})\rtimes\theta \end{equation} and put $\mathcal{U}:={\rm tc}(G(F)\times H(F)_{{\rm tu}})$. \begin{lem}\label{bijection}Keep the notation being as above. Assume the residue characteristic $p$ of $F$ is prime to $d:=[E:F]$. The map $H(F)_{{\rm tu}}\lra \mathcal{U},\ h\mapsto h\rtimes \theta$ induces a bijection from \begin{itemize} \item the set of $H(F)$-conjugacy classes in $H(F)_{{\rm tu}}$ to \item the set of $G(F)$-conjugacy classes in $\mathcal{U}$. \end{itemize} \end{lem} \begin{proof} We mimic the proof of Lemma 4.0.1, p.996 of \cite{GV}. We prove only the first claim and the second claim similarly follows. The surjectivity is clear since $H(F)=GL_n(F)\subset G(F)=GL_n(E)$. Therefore we prove the injectivity. Take two elements $h_1$ and $h_2$ in $H(F)_{{\rm tu}}$ such that $h_1\rtimes \theta$ is $G(F)$-conjugate to $h_2\rtimes \theta$. Hence, there exists $g\in G(F)$ such that $h_2\rtimes \theta=g^{-1}(h_1\rtimes \theta)g=g^{-1}h_1\theta(g)\rtimes \theta$. Therefore, $h_2=g^{-1}h_1\theta(g)$. Since $h_1$ and $h_2$ are $\theta$-fixed, we have $g^{-1}h^2_1\theta(g)=h_2$. By repeating this, we have $$h^d_2=g^{-1}h^d_1\theta^d(g)=g^{-1}h^d_1 g=(g^{-1}h_1g)^d$$ since $\theta^d=1$. Pick $a\in \Z_{>0}$ such that $ap\equiv -1$ mod $d$ so that for any odd positive integer $n$, we have $(ap)^n+1\equiv 0$ mod $d$. It follows from this that $(g^{-1}h_1g)^{(ap)^n+1}=h^{(ap)^n+1}_2$. By taking the limit $\ds\lim_{n\to \infty\atop n:{\rm odd}}$ on both sides, we have $g^{-1}h_1g=h_2$ as desired. \end{proof} \section{The normalizing factors of orbital integrals}\label{factors} In this section we collect some facts on the normalizing factors of orbital integrals which will be used to define the normalized orbital integrals. Let us fix a $p$-adic norm $|\cdot|_F$ on $F$. We denote by $\frak g$ (resp, $\frak h$) the Lie algebra of $G(F)$ (resp. $H(F)$). Similarly for $\delta=g\rtimes \theta$ with an element $g$ in $G(F)$ , we define the Lie algebra $\frak g_{\delta}$ for $G_\delta(F)$ where $G_\delta$ is the stabilizer of $\delta$ in $G$ which is also given by the $\theta$-twisted stabilizer of $g$ in $G$. For $x\in G(F)$ we see ${\rm Int}(\delta)(x):=\delta x\delta^{-1}={\rm Int}(g)\circ \theta(x)$. Hence ${\rm Int}(\delta)={\rm Int}(g)\circ \theta$ in ${\rm Aut}(G)$. It also yields ${\rm Ad}(\delta)={\rm Ad}(g)\circ d\theta$ in ${\rm Aut}(\frak g)$ where $d\theta:\frak g\lra \frak g$ is the derivative of $\theta$ at the origin. Here ${\rm Ad}(\delta)$ stands for the adjoint action of $\delta$ on $\frak g$ and ${\rm Ad}(g)$ as well. The map $d\theta$ is nothing but the natural (Galois) action of $\theta$ on $\frak g=M_n(E)$. For $\delta=g\rtimes \theta$ with a semisimple element $g$ in $G(F)$ we define the normalizing factor at $\delta$ as follows: \begin{equation}\label{nf1} D_{\tG}(\delta):=|\det(({\rm Ad}(\delta)-1)|\frak g/\frak g_\delta)|_F \end{equation} Similarly, for an element $\gamma$ in $H(F)$ we define \begin{equation}\label{nf2} D_{H}(\gamma):=|\det((1-{\rm Ad}(\gamma))|\frak h/\frak h_\gamma)|_F =|\det(({\rm Ad}(\gamma)-1)|\frak h/\frak h_\gamma)|_F \end{equation} where $\frak h_\gamma$ is the Lie algebra of $H_\gamma(F)$. \begin{lem} Assume the residue characteristic $p$ of $F$ is prime to $d:=[E:F]$. If $\gamma\in H(F)\subset G(F)$ is topologically unipotent, then $$D_{\tG}(\gamma\rtimes \theta)=D_{H}(\gamma)$$ \end{lem} \begin{proof} Clearly, $\frak h=\frak g^{d\theta=+1}=\frak g^\theta$. We decompose $\frak g=\frak h\oplus \frak g_1$ where $\frak g_1$ is a $d\theta$-stable subspace of $\frak g$ such that any eigenvalue of $d\theta|_{\frak g_1}$ is different from $1$. Therefore, the all eigenvalues of ${\rm Ad}(\gamma)\circ d\theta-1$ on $\frak g_1$ belong to $\mathcal{O}^\times_{\oF}$ since $\gamma$ is topologically unipotent and the residue characteristic $p$ of $F$ is prime to $d$. It is also clear that $d\theta-1=0$ on $\frak h$ and all eigenvalues of ${\rm Ad}(\gamma)$ on $\frak g$ belong to $\mathcal{O}^\times_{\oF}$ since $\gamma$ is topologically unipotent. Note that ${\rm Ad}(\gamma\rtimes \theta)-1={\rm Ad}(\gamma)\circ d\theta-1= {\rm Ad}(\gamma)(d\theta-{\rm Ad}(\gamma^{-1}))$ and also $\frak g_{\gamma\rtimes \theta}=\frak h_\gamma$ (since $\frak g^\theta=\frak h$). Then we have \begin{eqnarray} D_{\tG}(\gamma\rtimes \theta)&=& |\det({\rm Ad}(\gamma)|\frak g/\frak h_\gamma)|_F\times |\det(d\theta-{\rm Ad}(\gamma^{-1})|\frak g/\frak h_\gamma)|_F \nonumber \\ &=& |\det(d\theta-{\rm Ad}(\gamma^{-1})|\frak g/\frak h_\gamma)|_F \nonumber \\ &=&|\det(d\theta-1+1-{\rm Ad}(\gamma^{-1})|\frak h/\frak h_\gamma)|_F \times |\det(d\theta-{\rm Ad}(\gamma^{-1})|\frak g_1/\frak h_\gamma)|_F \nonumber\\ &=&|\det(1-{\rm Ad}(\gamma^{-1})|\frak h/\frak h_\gamma)|_F =|\det({\rm Ad}(\gamma)-1|\frak h/\frak h_\gamma)|_F\nonumber \\ &=&D_H(\gamma). \nonumber \end{eqnarray} \end{proof} We define the Cayley transform \begin{equation}\label{cayley} \frak c:\frak g\lra G(F),\ X\mapsto \Big(1+\frac{X}{2}\Big)\Big(1-\frac{X}{2}\Big)^{-1} \end{equation} which is birationally defined (hence, not defined on the whole space). By Lemma 3.2.3 of \cite{GV}, $\frak c$ induces \begin{equation}\label{homeo} \frak g_{{\rm tn}}\stackrel{\sim}{\lra} G(F)_{{\rm tu}},\ \frak h_{{\rm tn}}\stackrel{\sim}{\lra} H(F)_{{\rm tu}} \end{equation} and it satisfies nice properties \begin{equation}\label{nice} \frak c(-X)=\frak c(X)^{-1}\ {\rm and}\ {\rm Int}(J)\circ \frak c=\frak c\circ {\rm Ad}(J) \end{equation} for any $X\in \frak g_{{\rm tn}}$ and $J\in G(F)$ (cf. Remark 3.2.4 of \cite{GV}). It also holds that $\frak c \circ d\theta=\theta \circ \frak c$. The following lemma is proved as in the proof of Lemma 4.1.3 of \cite{GV} and therefore we omit giving a proof. \begin{lem}\label{ctf}For each $\gamma\in H(F)_{{\rm tu}}$ it holds that $$D_{\tG}(\gamma\rtimes \theta)=D_H(\gamma).$$ \end{lem} Finally, we end this section with the following lemma. \begin{lem}\label{cprime}The residual characteristic $p$ of $F$ is odd and prime to $d=[E:F]$. Then, the map $H(F)_{{\rm tu}}\lra H(F)_{{\rm tu}},\ \gamma'\mapsto \gamma'^d$ is a homeomorphism. \end{lem} \begin{proof}It follows from the argument in the proof of Lemma 3.2.7 of \cite{GV}. \end{proof} \section{Normalized orbital integrals and semi-simple descend}\label{oi} For any topological space $X$ we denote by $C^\infty_c(X)$ the space of locally constant functions on $X$ whose supports are compact. Recall our twisted space $\tG$. For $\delta=\delta_1\rtimes \theta \in \tG(F)$ with a semisimple element $\delta_1\in G(F)$ and $\widetilde{f}\in C^\infty_c(\tG(F))$, we define the normalized orbital integral of $\widetilde{f}$ at $\delta$ as follows: \begin{equation}\label{noiG1} I_\delta(\wf,dg,dg_{\delta})=D_{\tG}(\delta)^{\frac{1}{2}}\int_{G_\delta(F)\bs G(F)} \wf(g^{-1}\delta g) dg/dg_\delta \end{equation} with respect to the quotient measure $dg/dg_\delta$ of measures $dg$ on $G(F)$ and $dg_\delta$ on $G_\delta(F)$. The normalizing factor plays an important role when we view the above integral $I_\ast(\wf,dg,dg_{\delta})$ as the Schwartz space of $G(F)$ but we do not use this point. Let $f$ be the element in $C^\infty_c(G(F))$ corresponding to $\wf$ under the isomorphism $G(F)\stackrel{\sim}{\lra}\tG(F),\ g\mapsto g\rtimes \theta$. Then, we see that \begin{equation}\label{noiG2} I_\delta(\wf,dg,dg_{\delta})=D_{\tG}(\delta)^{\frac{1}{2}}\int_{G_\delta(F)\bs G(F)} f(g^{-1}\delta_1 \theta(g)) dg/dg_\delta. \end{equation} Similarly, for a semisimple element $\gamma \in H(F)(\subset G(F))$ and $f^H\in C^\infty_c(H(F))$, we define \begin{equation}\label{noiH} I_\gamma (f^H,dh,dh_\gamma ):= D_{H}(\gamma )^{\frac{1}{2}}\int_{H_\gamma (F)\bs H(F)} f^H(h^{-1}\gamma h) dh/dh_\gamma . \end{equation} with respect to the quotient measure $dh/dh_\gamma$ of measures $dh$ on $G(F)$ and $dh_\gamma$ on $H_\gamma(F)$. Recall $\mathcal{U}=t_{{\rm tc}}(G(F)\times H(F)_{{\rm tu}})$. Since $\mathcal U$ is open in $\tG(F)$ (cf. Lemma 4.0.5 of \cite{GV}), any element in $ C^\infty_c(\mathcal U)$ can be regarded with an element in $ C^\infty_c(\tG(F))$. Hence we may assume $ C^\infty_c(\mathcal U)\subset C^\infty_c(\tG(F))$. \begin{Def}$($Semisimple descend$)$Let $\wf\in C^\infty_c(\mathcal U)\subset C^\infty_c(\tG(F))$. We say an element $f^H$ in $C^\infty_c(H(F)_{{\rm tu}})$ can be obtained from $\wf$ by semi-simple descend at $\theta$ with respect to $dg$ and $dh$ if for all semi-simple element $\gamma $ in $H(F)_{{\rm tu}}$ \begin{equation}\label{SSD} I_\gamma (f^H,dh,dh_\gamma )=I_{\gamma \rtimes\theta}(\wf,dg,dg_{\gamma \rtimes \theta}). \end{equation} Notice that $G_{\gamma \rtimes \theta}=H_\gamma $. Therefore, we choose the same measure on $G_{\gamma \rtimes\theta}$ and $H_\gamma$ so that $dg_{\gamma \rtimes\theta}=dh_\gamma$. Therefore, the semisimple descend depends on the choice of the measures $dg$ and $dh$. \end{Def} Next we recall about the transfer matching orbital integrals (see (5.5.1), p.71 of \cite{KS}). As explained below, we fix a $\theta$-stable Whittaker datum defined in Section 5.3 of \cite{KS} for our $\theta$-splitting datum ${\rm spl}_G$ introduced at the beginning of Section \ref{tes}. This is necessary to normalize transfer factors in Section 5.3 of \cite{KS}. \begin{Def}\label{moi}$($Matching of orbital integrals and transfer of Hecke elements$)$ We say $\wf\in C^\infty_c(\tG(F))$ and $f^H \in C^\infty_c(H(F))$ have matching orbital integrals if, for every strongly $G$-regular element $\gamma \in H(F)\subset G(F)$, \begin{equation}\label{transfer} I_{\gamma }(f^H,dh,dh_\gamma )=\sum_{\delta}\Delta^{{\rm IV}}(\gamma,\delta) I_{\delta\rtimes \theta}(\wf,dg,dg_{\delta\rtimes\theta}) \end{equation} where the sum is taken over the set of complete representatives of $\theta$-conjugacy classes of strongly $G$-regular elements $\delta$ in $G(F)$ such that $\gamma$ is a norm of $\delta$ and the transfer factor $\Delta^{{\rm IV}}(\gamma,\delta)$ is defined in Section 5 of \cite{KS} with the collection of \cite{KSc}. The matching $($\ref{transfer}$)$ depends on the choice of the measures $dg, dh, dg_{\delta\rtimes\theta}$, and $dh_\gamma$ (see Section 3.10 of \cite{Wal-book} for measures in the matching of orbital integrals). \end{Def} \begin{remark}The symbol $H_1$ in \cite{KS} means a z-pair of $H=GL_n/F$ and obviously $H_1=H=GL_n/F$. Since $H=GL_n/F$, for elements in $H(F)=GL_n(F)$, the notation of strongly conjugacy classes is the same as one of conjugacy classes. \end{remark} \section{Transfer factors for our endoscopic datum}\label{tf} Let us keep the notation in Definition \ref{moi}. Then we will check the following: \begin{thm}\label{tf-triv}It holds that $$\Delta^{{\rm IV}}(\gamma,\delta)=1.$$ \end{thm} \begin{proof}In what follows, we drop $(\gamma,\delta)$ from the transfer factors. Let us recall our twisted endoscopic datum $(H,\mathcal{H},s,\xi)$ for $(G,\theta,\textbf{1}_{{\rm triv}})$ in Section \ref{tes}. Note that $\widehat{H}=(\widehat{G}^{\widehat{\theta}})^0$ in the notation of \cite{KS}. Because $s=1$, the factor $\Delta_{{\rm I}}$, defined page 33 at line 11 of \cite{KS}, is trivial. The factor $\Delta_{{\rm II}}$ is trivial. The factor $\Delta_{{\rm II}}$ in Lemma 4.3.A in page 36 of \cite{KS} is rewrote in Section 7.3, p.88 of \cite{Wal-book} and the description there is much easier to understand $\Delta_{{\rm II}}$ and in fact, the factor $\Delta_{{\rm II}}=1$ because $s = 1$. The factor $\Delta_{{\rm III}}$ defined page 38 of \cite{KS} is trivial because $\textbf{A} = 1$ since $s=1$. The term $\textbf{A} $ measures the difference between $\mathcal{H}$ and $\widehat{G}^{\widehat{\theta},0}\rtimes W_F$ . There is no difference in our setting because, indeed, $\mathcal{H}= \widehat{G}^{\widehat{\theta},0}\rtimes W_F$. The transfer factor $\Delta^{{\rm IV}}$ does not contain the terms $\Delta_{{\rm IV}}$. But in fact, even this term is 1 because its definition in Lemma 4.5.A, page 46 of \cite{KS}, it is as $\Delta_{{\rm II}}$ a product over an empty set. Summing up, we have the claim. \end{proof} \begin{remark}\label{correction} As Waldspurger pointed out, there is some mistakes in \cite{Wal-book} because he has used the definition of Kottwitz-Shelstad in \cite{KS} but there is a mistake in this definition, cf. Kottwitz-Shelstad ArXiv 2012 \cite{KSc}. However, this does not affect really the proofs in \cite{Wal-book}. \end{remark} \section{Semisimple descend theorem}\label{sdt} In this section we will prove the semisimple descend theorem for some Hecke elements. Recall the Lie algebra $\frak g=M_n(E)$ of $G(F)$ and the derivative $d\theta$ of $\theta$ at the origin acts there by the usual Galois action of $\theta$. We view it as a $\mathcal{O}_F$-module. Recall the open subspace $\mathcal{U}\subset \tG(F)$ defined by the twisted conjugation map (\ref{twcm}). The following theorem is an analogue of Lemma 4.2.4 of \cite{GV}. \begin{thm}\label{mainSSD} Assume the residue characteristic $p$ of $F$ satisfies $p>2$ and $p\nmid d=[E:F]$. Let $L\subset \frak g_{{\rm tn}}$ be a $d\theta$-invariant $\mathcal{O}_F$-lattice satisfying $L\cdot L\subset \varpi_F L$. Assume that $L=L_1\oplus L_2$ where $L_1$ is a $\O_F$-lattice of $\frak g^\theta$ and $L_2$ is a $\O_F$-lattice of $(1-d\theta)\frak g$ $($this is always possible by assumption on the residue characteristic$)$. Let $K_L:=\frak c(L)$ and $K_{L,\theta}=\frak c(L)\cap H(F)$ so that $K_{L,\theta}=K^\theta_L$ by $\frak c\circ d\theta=\theta\circ \frak c$. Then it holds that \begin{enumerate} \item $K_L\subset G(F)_{{\rm tu}}$ $($resp. $K_{L,\theta}\subset H(F)_{{\rm tu}})$ is a compact open subgroup of $G(F)$ $($resp. $H(F))$; \item $K_L\rtimes \theta={\rm tc}(K_L,K_{L,\theta})$; \item Let $C_H\subset H(F)_{{\rm tu}}$ be an open compact subset and $K\subset G(F)$ be an open compact subgroup. Put $C={\rm tc}(K,C_H)$ which is an open compact subset of $\mathcal{U}\subset \tG(F)$. Then the element ${\rm vol}^{-1}(K\cap H(F))\textbf{1}_{C_H}$ of $C^\infty_c(H(F))$ can be obtained from the element ${\rm vol}^{-1}(K)\textbf{1}_{C}$ of $C^\infty_c(\mathcal{U})\subset C^\infty_c(\tG(F))$ by semisimple descend at $\theta$. In particular, ${\rm vol}^{-1}(K_{L,\theta})\textbf{1}_{K_{L,\theta}}$ can be obtained from the element ${\rm vol}^{-1}(K_L)\textbf{1}_{K_L\rtimes \theta}$ by semisimple descend at $\theta$. \end{enumerate} \end{thm} \begin{proof} Let us follow the proof of Lemma 4.2.4 of \cite{GV} verbatim with a minor modification. The first claim follows from Lemma 4.2.3, p.998 of \cite{GV}. Next let us prove the second claim. Obviously, ${\rm tc}(K_L,K_{L,\theta})$ is contained in $K_L\rtimes \theta$. As explained in the proof of Lemma 4.2.4-(i) of \cite{GV}, ${\rm tc}$ is an open map. Since $L_1$ is compact, there exists $m_0\in \Z_{\ge 0}$ such that ${\rm tc}(K_L,K_{L,\theta})\supset \frak c(L_1+\varpi^{m_0}_F L_2)\rtimes \theta$. By reverse induction, for any $m\in \Z_{\ge 0}$, each element of $\frak c(L_1+\varpi^{m}_F L_2)$ can be written as $g^{-1}\delta \theta(g)$ for some $g\in K_L$ and $\delta\in \frak c(L_1+\varpi^{m+1}_F L_2)$. Since $\frak c(L)=1+L$ (cf. Lemma 4.2.3 of \cite{GV}), any element $\frak c(L_1+\varpi^{m}_F L_2)$ is written by $1+X$ for some $X=X_1+X_2$ with $X_1\in L_1$ and $X_2\in \varpi^m_F L_2$. By assumption on $p$, $F(\zeta_d)/F$ is unramified. Therefore we may suppose that $F$ contains a primitive $d$-th root of unity in the argument below. Let us write $X_2=\ds\sum_{i}X_{2,i}$ where $\{X_{2,i}\}_i$ are eigenvectors of $d\theta$ with $d\theta X_{2,i}=\zeta^{n_i}_d X_{2,i},\ \zeta^{n_i}_d\neq 1$. Since $d\theta$ has the distinct eigenvalues, the $\zeta^{n_i}_d$'s are different each other. Further, $1-\zeta^{n_i}_d$ is a unit in $\O_F$ by assumption on $p$ and $d$. Thus, each $X_{2,i}$ belongs to $\varpi^m_F L_2$. Put $Y=\sum_{i}a_i X_{2,i}\in \varpi^m_F L_2$ with $a_i=(1-\zeta^{n_i}_d)^{-1}$. As explained in the mid of the proof of Lemma 4.2.4 of \cite{GV}, $\frak c(Y)\equiv 1+Y\ {\rm mod}\ \varpi^{m+1}_F$. Thus, using the assumption $L\cdot L\subset \varpi_F L$, \begin{eqnarray} \frak c (Y)^{-1}(1+X)\theta(\frak c(Y))&=& \frak c (Y)^{-1}(1+X)\frak c(d\theta(Y)) \nonumber \\ &\equiv & (1-Y)(1+X)(1+\sum_{i}\zeta^{n_i}_d a_i X_{2,i}) \ {\rm mod}\ \varpi^{m+1}_F \nonumber \\ &\equiv & 1-Y+X_1+X_2+\sum_{i}\zeta^{n_i}_d a_i X_{2,i} \ {\rm mod}\ \varpi^{m+1}_F \nonumber \\ &=& 1+X_1 \ {\rm mod}\ \varpi^{m+1}_F. \nonumber \end{eqnarray} By using $\frak c(L)=1+L$ again, we have the claim. Finally let us prove the third claim, namely, the semisimple descend. Let $\gamma\in H(F)_{{\rm tu}}$. By Lemma \ref{ctf}, $D_{\tG}(\gamma\rtimes\delta)=D_H(\gamma)$. If the characteristic polynomial of $\gamma$ has multiple roots, then $D_H(\gamma)=0$. Therefore, we may assume that the eigenvalues of $\gamma$ are different each other. In particular, $G_{m\rtimes \delta}=H_\gamma$ is a torus. In particular $H_\gamma(F)$ is unimodular and so are $G(F)$ and $H(F)$. What we need to check is, by definition, the equality without normalizing factors: $${\rm vol}^{-1}(K\cap H(F))\int_{H_\gamma(F)\bs H(F)} \textbf{1}_{C_H}(h^{-1}\gamma h) dh/dh_\gamma$$ $$={\rm vol}^{-1}(K)\int_{G_{\gamma\rtimes\theta}(F)\bs G(F)} \textbf{1}_{C}(g^{-1}(\gamma\rtimes \theta) g) dg/dg_{\gamma\rtimes \theta}.$$ However, applying Lemma 4.2.5 of \cite{GV} to $G_1=G,\ H_1=H,\ I_1=H_\gamma$, we have the claim. \end{proof} \section{Proofs of the main theorems}\label{proof} We first prove Theorem \ref{main1} and Theorem \ref{main2} simultaneously, and then Theorem \ref{main3} after that. \begin{proof}(A proof of Theorem \ref{main1} and Theorem \ref{main2}) Recall $f^H={\rm vol}(K_E(m)_\theta)\textbf{1}_{K_E(m)_\theta}$ and $f^G={\rm vol}(K_E(m))\textbf{1}_{K_E(m)}$. Put $\wf^G={\rm vol}(K_E(m))\textbf{1}_{K_E(m)\rtimes \theta}$ corresponding to $f^G$ under $G(F)\stackrel{\sim}{\lra} \tG(F),\ g\mapsto g\rtimes\theta$. In view of matching, if a semisimple element $\gamma$ in $H(F)$ is not regular, then $\delta\in G(F)$ is not $\theta$-regular for which $\gamma$ is a norm of $\delta$. In fact, it follows from the proof of Lemma 1.1 of \cite{AC}, $G_{\delta\rtimes \theta}$ is an inner form of $H_\gamma$. The claim follows from this. Therefore, the both sides in Theorem \ref{main2} vanish because of the normalizing factors which vanish on such elements. Thus we may assume that $\gamma\in H(F)$ is regular, then $H_\gamma$ is a torus and it is isomorphic to $G_{\delta\rtimes \theta}={\rm Cent}_\theta(\delta,G)$. In particular, it is abelian. Therefore, $\delta$ is a strongly $\theta$-regular element. The sum of the right hand side in (\ref{transfer}) runs over a set of complete representatives of $G(F)$-conjugacy classes of regular elements $\delta$ such that $\gamma$ is a norm of $\delta$. Plugging this into Theorem \ref{tf-triv}, \begin{equation}\label{eq1} \sum_{\delta}\Delta^{{\rm IV}}(\gamma,\delta) I_{\delta\rtimes \theta}(\wf^G,dg,dg_{\gamma\rtimes\theta})= \sum_{\delta:{\rm regular} \atop \text{$\gamma$ is a norm of $\delta$}}I_{\delta\rtimes \theta}(\wf^G,dg,dg_{\delta\rtimes\theta}). \end{equation} Suppose that $\gamma\not\in H(F)_{{\rm tu}}$. For any $\delta$ on the right in the above sum, it holds that $\delta\not\in G(F)_{{\rm tu}}$ by Remark \ref{diagonal}. Since $f^H$ and $\wf^G$ are supported on topologically unipotent elements, the equality in the claim trivially holds. Thus, we may assume that $\gamma\in H(F)_{{\rm tu}}$ and $\delta\in G(F)_{{\rm tu}}$ for $\gamma$ and $\delta$ in question. Let $\delta_1,\delta_2$ be elements in $\delta\in G(F)_{{\rm tu}}$ such that $\gamma$ is a norm of both of them. By Lemma 1.1-(ii) of \cite{AC}, $\delta_1$ is $\theta$-conjugate to $\delta_2$. Therefore, (\ref{eq1}) becomes \begin{equation}\label{eq2} \sum_{\delta}\Delta^{{\rm IV}}(\gamma,\delta) I_{\delta\rtimes \theta}(\wf^G,dg,dg_{\delta\rtimes\theta})= I_{\delta\rtimes \theta}(\wf^G,dg,dg_{\delta\rtimes\theta}) \end{equation} for any such $\delta$. By Lemma \ref{cprime}, there exists an element $\gamma'\in H(F)_{{\rm tu}}$ such that $\gamma'^d=\gamma$. Thus $N_{E/F}(\gamma')=\gamma'^d=\gamma$. By Lemma 1.1-(ii) of \cite{AC} and Lemma \ref{bijection}, we may assume that $\delta=\gamma'$. To be more precise, \begin{equation}\label{eq3} I_{\delta\rtimes \theta}(\wf^G,dg,dg_{\delta\rtimes\theta}) =I_{\gamma'\rtimes \theta}(\wf^G,dg,dg_{\gamma'\rtimes\theta}) \end{equation} for any element $\gamma'$ of $H(F)$ which is $H(F)$-conjugate to an element whose usual norm is $\gamma$. Since $f^G$ is supported on $K_E(m)$ and the homeomorphism in Lemma \ref{cprime} induces a homeomorphism $K_E(m)\stackrel{\sim}{\lra}K_E(m)$. Thus, \begin{equation}\label{eq4} I_{\gamma'\rtimes \theta}(\wf^G,dg,dg_{\gamma'\rtimes\theta})= I_{\gamma\rtimes \theta}(\wf^G,dg,dg_{\gamma\rtimes\theta}). \end{equation} Further, it follows from Theorem \ref{mainSSD}-(3) that \begin{equation}\label{eq5} I_{\gamma\rtimes \theta}(\wf^G,dg,dg_{\gamma\rtimes\theta})= I_{\gamma}(f^H,dh,dh_{\gamma}). \end{equation} Theorem \ref{main1} follows from the equations (\ref{eq1}) through (\ref{eq5}). Theorem \ref{main2} follows from the equations (\ref{eq3}) through (\ref{eq5}). \end{proof} \begin{proof} (A proof of Theorem \ref{main3}) Since $\pi$ is cuspidal, $\Pi$ is a constituent of the discrete spectrum of $L^2(GL_n(L)\bs GL_n(\A_L))$. Plugging our Hecke elements $f_1:={\rm Vol}(U_{S}(\underline{n},\underline{m})_\theta)^{-1} \textbf{1}_{U_{S}(\underline{n},\underline{m})_\theta}$ and $f_2:={\rm Vol}(U_{S}(\underline{n},\underline{m}))^{-1} \textbf{1}_{U_{S}(\underline{n},\underline{m})}$ into the discrete part of the trace formula in \cite{AC} (see the introduction of \cite{AC} and also \cite{J} as a reader's friendly reference) with Theorem \ref{main2}, we have $${\rm tr}(\pi(f_1))={\rm tr}(\Pi(f_2))$$ which gives us the claim. \end{proof}

d8e59566c659de424806abd117c371810f958b39

\section{Introduction} Factuality is defined as a measure of ``whether eventualities are characterized as corresponding to facts, possibilities, or situations that do not hold in the world'' \citep{sauri2008factuality, sauri-event-factuality}. In summarization, this ``world'' is the article, which is taken as ground-truth, and the output summary must be faithful to the article's facts. Despite advancements in neural abstractive summarization \citep{xsum, liu-lapata-2019-text, bart}, $\sim$30\% of summaries have factual inconsistencies \citep{fact-aware}. With summarization being an integral component of information consumption, this highlights a need for ensuring summarization systems are factually consistent and developing methods for evaluating them. Common evaluation metrics for summarization based on n-gram overlap -- BLEU, ROUGE, and METEOR \citep{Bleu, rouge, meteor} -- are insufficient to measure the factual correctness of summaries and fail to correlate with the human judgements of factuality \citep{falke-etal-2019-ranking, sumcriticaleval}. More recent metrics proposed to improve the evaluation of summarization factuality \citep{factcc, feqa, factqa, factentailment} cannot be compared due to the lack of common benchmarks. More critically, while these approaches differ in the way they model factuality, they all consider factuality as a binary concept, labeling summaries of any length as factual or non-factual. They do not provide any fine-grained understanding of the factual errors made by different systems that could serve as an actionable feedback on a system's limitations. The binary factuality of a text can be difficult to determine. \citet{falke-etal-2019-ranking} show relatively low crowd--expert agreement, indicating the presence of subjectivity in the annotation process. Moreover, not all factual errors are equally important and the number of errors can have a significant impact on the perceived factuality of a text. This suggests that non-factuality should be modeled as a multi-dimensional construct and not a label. \begin{figure*} \centering \includegraphics[width=\linewidth]{plots/Factuality_Benchmark_Diagrams_Frank_Simple.pdf} \caption{We propose a linguistically grounded typology of factual errors. We select crowd workers to annotate summaries from two datasets according to this typology achieving near perfect agreement with experts. We collect FRANK{}, the resulting dataset, to benchmark factuality metrics and state-of-art summarization systems. } \label{fig:methodology} \end{figure*} In this work, we propose a linguistically motivated typology of factual errors for fine-grained analysis of factuality in summarization systems (\Sref{sec:typology}). Our typology is theoretically grounded in frame semantics \citep{fillmore1976frame, palmer2005proposition} and linguistic discourse theory \citep{brown1983discourse}. It provides several benefits. First, we find that decomposing the concept of factuality in (relatively) well-defined and grounded categories makes the final binary decision more objective leading to near perfect agreement between crowd and expert annotators ($\kappa = 0.86$). Second, this approach provides some measure of the degree of non-factuality both in terms of the quantity and the category of factual violations that appear in the text. This typology also provides us with the means to categorize the types of errors made by summarization systems, helping us gain deeper insights than simply categorizing content as factual or hallucinated. We define an annotation protocol of factuality based on our typology and collect a dataset of human judgements over a diverse set of model generated summaries on the CNN/DM \citep{cnn-dm} and XSum \citep{xsum} datasets (\Sref{sec:humanevalstudy}). Through this dataset, we aim to both assess the factuality of summarization systems and benchmark recently proposed factuality metrics. In \Sref{sec:model_analysis} we discuss various state-of-art models and show a detailed analysis of the factual errors they make. Finally, in \Sref{sec:metric_evaluation} we evaluate multiple summarization metrics against our benchmark and show their strengths and weaknesses in detecting specific types of factual errors. \autoref{fig:methodology} shows an overview of this work. \begin{table*}[ht] \small \begin{tabular}{ l p{2.5cm} p{5.2cm} p{5.5cm} } \toprule & \textbf{Category} & \textbf{Description} & \textbf{Example} \\\midrule \textbf{{PredE}{}} & Relation Error & The predicate in the summary statement is inconsistent with the source article. & \textit{The Ebola vaccine \wrong{was rejected} by the FDA in 2019.} \\\hline \textbf{{EntE}{} } & Entity Error & The primary arguments (or their attributes) of the predicate are wrong. & \textit{The \wrong{COVID-19 vaccine} was approved by the FDA in 2019.} \\\hline \textbf{{CircE}{} } & Circumstance Error & The additional information (like location or time) specifying the circumstance around a predicate is wrong. & \textit{ The first vaccine for Ebola was approved by the FDA in \wrong{2014}.} \\\hline \textbf{{CorefE}{}} & Coreference Error & A pronoun/reference with wrong or non-existing antecedent. & \textit{The first vaccine for Ebola was approved in 2019. \wrong{They} say a vaccine for COVID-19 is unlikely to be ready this year.} \\\hline \textbf{{LinkE}{}} & Discourse Link Error & Error in how multiple statements are linked together in the discourse (for example temporal ordering/causal link). & \textit{To produce the vaccine, scientists have to show successful human trials, \wrong{then} sequence the DNA of the virus.} \\\hline \textbf{{OutE}{}} & Out of Article Error & The statement contains information not present in the source article. & \textit{\wrong{China} has already started clinical trials of the COVID-19 vaccine.} \\\hline \textbf{{GramE}{}} & Grammatical Error & The grammar of the sentence is so wrong that it becomes meaningless. & \textit{The Ebola vaccine \wrong{accepted have already started.}} \\\bottomrule \end{tabular} \caption{Typology of factual errors. Original text for the examples: \textit{The first vaccine for Ebola was approved by the FDA in 2019 in the US, five years after the initial outbreak in 2014. To produce the vaccine, scientists had to sequence the DNA of Ebola, then identify possible vaccines, and finally show successful clinical trials. Scientists say a vaccine for COVID-19 is unlikely to be ready this year, although clinical trials have already started. }} \label{tab:framework} \end{table*} \section{Typology of Factual Errors} \label{sec:typology} Previous studies of factuality in summarization only distinguish factual and hallucinated content \citep{sumcriticaleval, factentailment} and provide limited insights on the fine-grained types of factual errors. In the simplest case, factual errors appear within a single proposition. However, as summaries include several sentences, discourse markers describe relations across propositions. These cross-sentence links, such as causality or temporal ordering, can introduce inconsistencies with the article. Furthermore, information in the summary should be verifiable given the article. This understanding outlines different levels of linguistic structure where factual mistakes can arise in summaries: at the semantic frame level, at the discourse level, or because the content cannot be verified. Below we define a typology of factual errors further detailing these three levels. This typology is theoretically grounded in frame semantics \citep{fillmore1976frame, baker-etal-1998-berkeley, palmer2005proposition} and linguistic discourse analysis \citep{brown1983discourse}. Examples for each category are shown in \autoref{tab:framework}. \subsection{Semantic Frame Errors} A \emph{semantic frame} is a schematic representation of an event, relation, or state, which consists of a predicate and a list of participants, called frame elements \cite{baker-etal-1998-berkeley}. A semantic frame has both core and non-core frame elements (FE). Core frame elements are essential to the meaning of the frame, while non-core (e.g. location, time) provide additional descriptive information. Our first three categories capture factual errors in each of these components (frame, core and non-core FE) respectively. \paragraph{Predicate Error ({PredE}{}):} Category \emph{{PredE}{}} encompasses errors where the predicate in a summary statement is inconsistent with the source text. More generally, this represents cases where the frame from a summary statement does not align with what is expressed in the source text. \paragraph{Entity Error ({EntE}{}):} Category \emph{{EntE}{}} captures errors where the primary arguments (like entities) of the predicate are wrong or have the wrong attributes, although the relation was expressed in the original text. More generally, these account for cases where the core frame elements in a frame are wrong. This also captures directionality errors where the elements are interchanged (similar to agent-patient swap). \paragraph{Circumstance Error ({CircE}{}):} In additional to the core arguments, predicates can be further specified using additional information or attributes that describe the circumstance in which the arguments and predicates interact (e.g. location, time, manner, direction, modality). Category \emph{{CircE}{}} captures errors where one or more such attributes (non-core frame elements within a frame) are wrong. \subsection{Discourse Errors} The communicative intent of an author is also expressed through relations that hold between parts of the text. Factual errors in summarized text can often extend beyond a single semantic frame introducing erroneous links between discourse segments. Below we outline such categories of errors which are grounded in discourse analysis and rhetorical structure theory (RST) \citep{brown1983discourse,mann1988rhetorical}. RST is an elaborate system for annotating coherence relations in discourse. Some examples of such relations include: ``Elaboration'', ``Background'', ``Motivation'', and ``Volitional Cause''. Here we depart from semantic frame terminology as its rooting in a single frame does not allow us to represent such errors. \paragraph{Coreference Error ({CorefE}):} Category \emph{{CorefE}} accounts for errors where pronouns and other types of references to previously mentioned entities either are incorrect or have no clear antecedents, making them ambiguous. \paragraph{Discourse Link Error ({LinkE}):} Category \emph{{LinkE}} encompasses errors involving a discourse link between different statements. These include errors of incorrect temporal ordering or incorrect discourse links (e.g. RST relations, discourse connectors) between statements. \subsection{Content Verifiability Errors} Often statements in a summary cannot be verified against the source text due to difficulty in aligning them to the source. Below we outline two categories of errors for such cases. \paragraph{Out of Article Error ({OutE}):} Since summaries of a document should only contain information that can be deduced from the original text, we include a category for such errors \emph{{OutE}} (prior work refers to this as extrinsic hallucinations \citep{factentailment}). \paragraph{Grammatical Error ({GramE}):} We use \emph{{GramE}} to categorize statements that are not well formed. When grammatical mistakes make the meaning of a statement incomprehensible or ambiguous, it cannot be verified against the source and is thus considered trivially wrong. Minor grammatical errors are acceptable. Finally, for completeness in our annotation exercise, we add two additional categories \textbf{Others ({OthE}{})} for factually errors that do not correspond to any of the above categories and \textbf{Not an Error ({NE}{})} for statements that do not contain any errors. \section{Dataset Creation} \label{sec:humanevalstudy} Beyond theoretical grounding, we empirically verify our typology through large scale human annotations of five abstractive summarization models on the CNN/DM dataset and four on the XSum dataset. Through our dataset, we aim to have a broad coverage of different types of errors made by neural summarization systems, with human judgements on their fine-grained factuality errors. \paragraph{Annotation Data} For the annotation, we include model summaries from CNN/DM and XSum datasets as they present different characteristics. CNN/DM summaries are longer, with three sentences on average, while XSum has only single sentence summaries. Having longer summaries is crucial to identify discourse level errors. On the other hand, XSum summaries are more abstractive and include more factual errors on average \citep{factentailment}. For a diverse set of model summaries, we collect publicly available model outputs from different summarization models with differing factuality capabilities. For the CNN/DM dataset, we use model outputs from a LSTM Seq-to-Seq model (S2S) \citep{rush}, a Pointer-Generator Network (PGN) model \citep{pgn}, a Bottom-Up Summarization (BUS) model \citep{bus}, a Bert based Extractive-Abstractive model (BertSum) \citep{liu-lapata-2019-text} and a jointly pretrained transformer based encoder-decoder model BART \citep{bart}. For the XSum dataset, we collect model outputs from a Topic-Aware CNN Model \citep{xsum}, a Pointer-Generator Network (PGN) model, a randomly initialized (TransS2S) \citep{vaswani2017attention} and one initialized with Bert-Base (BertS2S) \citep{bert}.\footnote{As we use publicly available model outputs, the summaries across different datasets are from different models owing to their availability.} Details of the models used are provided in \Sref{sec:model-details}. \begin{figure*}[ht!] \centering \includegraphics[width=\linewidth]{plots/factcategory.pdf} \caption{Proportion of summaries with factual errors based on collected annotations, with breakdown of the categories of errors within. Full specification of categories of errors in \autoref{tab:framework}.} \label{fig:factcategory} \end{figure*} \paragraph{Annotation Collection} Using the above model generated summaries, we collect human annotations from three independent annotators for 250 articles from each dataset (with a total of 1250 model outputs on CNN/DM and 1000 on XSum). We annotate each sentence of a summary to break the judgement of factuality into smaller units. We present sentences in the context of the entire summary to identify discourse errors spanning multiple sentences. Annotations are a two step process: for each sentence in the summary, the annotator first selects whether the sentence is factual, and if marked not factual, identifies the category of each error based on our typology. \footnote{We experimented with Likert scale evaluation of full summaries in a pilot study. Such an annotation would not provide precise information about where in the summary an error appears and also resulted in lower agreement. Hence, we opted for sentence level judgements.} A sentence can be annotated with more than one category of errors to account for multiple errors within a sentence. We conduct the annotation task on the Amazon Mechanical Turk (MTurk) platform. To achieve high quality crowd-sourced annotations, we build an intuitive interface\footnote{We make the interface available for future human annotations that follow our typology} which combines: \begin{compactenum} \item \textbf{Clear Instructions: }We explain the annotation scheme without assuming linguistic knowledge and give several examples for each category. \item \textbf{Training and Evaluation: }We setup training tutorials for first time users to train and provide feedback on the task. We also setup a qualification test which tests their understanding of our annotation scheme and require annotators to obtain >85\% score to qualify. Further, we continuously evaluate annotators during the task against artificially generated factual errors to ensure continued high quality. \item \textbf{Fair Pay and Bonus: } All workers are paid 50\% more than the average American minimum wage. We offer bonuses for scores of 60\% or above on the continuous evaluation, and for completing sets of 10 annotations. \end{compactenum} Further details on our interface are added in \Sref{sec:annotation_details} \paragraph{Inter-Annotator Agreement: } We report inter-annotator agreement in terms of Fleiss Kappa $\kappa$ \citep{fleiss1971measuring}. Following \citet{feqa}, we report the percentage $p$ of annotators that agree with the majority class. Each datapoint in our dataset corresponds to a sentence in a summary. We compute agreement on all 4942 annotated sentences. On the annotation of whether a sentence is factual or not we obtain $\kappa=0.58$, with $p=91\%$ of annotators agreeing with the majority class. As a comparison, \citet{feqa} reports $p=76.7\%$ average agreement. When all three annotators agree that a sentence is not factual, we obtain $\kappa=0.39$ with $p=73.9\%$ of annotators agreeing with the majority class on the eight category annotation (seven categories of errors and ``other'') which indicate a moderate agreement. \paragraph{Agreement with Domain Expert: } We measure agreement between the majority class of the three annotators and one expert annotator on 201 datapoints (10 summaries from CNN/DM and 10 summaries from XSum). We find a Cohen Kappa of $\kappa = 0.86$ indicating nearly perfect agreement. Previous work found agreement of $\kappa = 0.65$ between three crowd annotators and expert annotations of factuality \citep{falke-etal-2019-ranking}. Even with more than nine workers, they report agreement with expert annotations of at most $\kappa = 0.74$. This improvement validates the robustness of our annotation interface and protocol which achieves higher agreement with fewer workers. \section{Summarization Model Analysis} \label{sec:model_analysis} We evaluate the performance of different summarization models in terms of factuality. \autoref{fig:factcategory} visualizes the percentage of summaries with factual errors for each category model and dataset, with a breakdown of proportion of different error types within each. A summary is considered incorrect if it contains at least one sentence with a factual error. A sentence contains a factual error if the majority of annotators indicate the presence of an error (here we do not consider annotations where all three annotators disagree on the category). \paragraph{How factual are generated summaries across different datasets? }From our annotations, we observe that 60\% of the summaries that were annotated contain at least one factual error. From \autoref{fig:factcategory}, we see that the XSum dataset has more factually incorrect model summaries (92\%) than CNN/DM (43\%). It poses more significant challenges in terms of factuality as all models produce > 80\% summaries with factual errors, with the best model (BertS2S) producing 83\% wrong summaries. On the CNN/DM dataset, while state-of-the-art pretrained models like BERTSum and BART have better factuality numbers, the percentage of factually incorrect summaries is still high (23\% for BERTSum and 27\% for BART). The proportion of errors across different categories vary widely between the two datasets. For the CNN/DM dataset, the most frequent classes of errors are Entity Error ({EntE}{}) and Coreference Error ({CorefE}{}). For the XSum dataset they are Out of Article Error ({OutE}{}) and Entity Error ({EntE}{}). Note that there are no discourse errors ({CorefE}{}, {LinkE}{}) in the XSum dataset because the data only contains single sentence summaries. Additionally, we observe that {OthE}{} makes up a very small percentage ($\sim$ 1\%) of errors overall showing that our typology is \emph{complete} with most errors being mapped to one of our existing categories. \paragraph{How factual are generated summaries across different models? } From \autoref{fig:factcategory}, we observe that LSTM based models like S2S and BUS generate many incorrect summaries. Interestingly, PGN on CNN/DM has fewer summaries with factual errors (26\%) compared to S2S (74\%) and BUS (62\%) potentially due to the extractive nature of CNN/DM and the copy based objective in PGN. PGN has been previously shown to produce highly extractive summaries on CNN/DM copying large portions of text (often entire sentences) \citep{bus, balachandran2020structsum}. On the more abstractive dataset XSum, PGN produces $> 96\%$ factually incorrect summaries. We also observe that large-scale pretrained models improve factuality on both datasets, as also noted by \citet{feqa}, with more significant gains on CNN/DM. On CNN/DM, BERTSum and BART display half the error rate of BUS. In contrast, on XSum, BertS2S improves over non-pretrained models by $\sim$ 10\% only, showing that XSum poses a significant challenge for factuality even in pretrained models. Different models also exhibit different distributions in the error categories. LSTM based models have higher proportion of Grammatical Errors ({GramE}{}) while transformer and CNN based models have a lower proportion. For pretrained transformer models, we observe that the improved error-rate on the CNN/DM dataset can be attributed to improvements at the frame level ({PredE}{}, {EntE}{}, {CircE}{}) while the discourse level errors still remain a challenge. Errors {CorefE}{}, {LinkE}{} account for a higher proportion of errors in BERTSum and BART compared to the other models. \begin{table*} \centering \resizebox{\linewidth}{!}{ \begin{tabular}{l| c c c c | c c c c | c c c c } \toprule & \multicolumn{4}{c|}{All data} & \multicolumn{4}{c|}{CNN/DM} & \multicolumn{4}{c}{XSum} \\ \midrule \multirow{2}{*}{Metrics} & \multicolumn{2}{c}{Pearson} & \multicolumn{2}{c|}{Spearman} & \multicolumn{2}{c}{Pearson} & \multicolumn{2}{c|}{Spearman} & \multicolumn{2}{c}{Pearson} & \multicolumn{2}{c}{Spearman} \\ &$\rho$ & p-val & $r$ & p-val &$\rho$ & p-val & $r$ & p-val &$\rho$ & p-val & $r$ & p-val \\ \midrule BLEU & 0.10 & 0.00 & 0.07 & 0.00 & 0.08 & 0.01 & 0.08 & 0.01 & 0.14 & 0.00 & 0.20 & 0.00\\ METEOR & 0.14 & 0.00 & 0.11 & 0.00 & 0.12 & 0.00 & 0.10 & 0.00 & 0.15 & 0.00 & 0.10 & 0.00\\ Rouge-1 & 0.14 & 0.00 & 0.10 & 0.00 & 0.12 & 0.00 & 0.10 & 0.00 & 0.15 & 0.00 & 0.09 & 0.01\\ Rouge-2 & 0.12 & 0.00 & 0.08 & 0.00 & 0.08 & 0.00 & 0.07 & 0.01 & 0.17 & 0.00 & 0.14 & 0.00\\ Rouge-L & 0.13 & 0.00 & 0.09 & 0.00 & 0.11 & 0.00 & 0.09 & 0.00 & 0.16 & 0.00 & 0.10 & 0.00\\ \midrule OpenIE & 0.11 & 0.00 & 0.02 & 0.36 & 0.16 & 0.00 & 0.15 & 0.00 & 0.00 & 0.93 & -0.45 & 0.00\\ BERTS P & \textbf{0.27} & 0.00 & 0.24 & 0.00 & 0.35 & 0.00 & 0.29 & 0.00 & \textbf{0.18} & 0.00 & 0.09 & 0.00\\ BERTS R & 0.14 & 0.00 & 0.13 & 0.00 & 0.21 & 0.00 & 0.17 & 0.00 & 0.07 & 0.03 & 0.03 & 0.38\\ BERTS F1 & 0.24 & 0.00 & 0.21 & 0.00 & 0.32 & 0.00 & 0.26 & 0.00 & 0.15 & 0.00 & 0.06 & 0.05\\ FEQA & 0.00 & 0.83 & 0.01 & 0.60 & -0.01 & 0.76 & -0.01 & 0.72 & 0.02 & 0.45 & 0.07 & 0.04\\ QAGS & 0.06 & 0.00 & 0.08 & 0.00 & 0.13 & 0.00 & 0.09 & 0.00 & -0.02 & 0.48 & 0.01 & 0.65\\ DAE & 0.16 & 0.00 & 0.14 & 0.00 & 0.25 & 0.00 & 0.24 & 0.00 & 0.04 & 0.16 &\textbf{ 0.28} & 0.00\\ FactCC & 0.20 & 0.00 &\textbf{ 0.30} & 0.00 & \textbf{0.36} & 0.00 & \textbf{0.33} & 0.00 & 0.07 & 0.02 & 0.25 & 0.00\\ \bottomrule \end{tabular} } \caption{Partial Pearson correlation and Spearman rank correlation coefficients and p-values between human judgements and metrics scores. Comparisons should be made along with the pairwise Williams test found in \autoref{tab:metric-metric-results}.} \label{tab:results} \end{table*} \section{Factuality Metric Evaluation} \label{sec:metric_evaluation} We propose the FRANK{} dataset resulting from the human annotation study as a common benchmark to assess different factuality metrics. We provide an evaluation protocol of factuality metrics, which controls for dataset biases, and a fine grained analysis of the strengths of each metric. \subsection{Benchmark} The FRANK{} benchmark provides a diverse dataset for evaluating various metrics on their ability to capture factual errors. Notably, our benchmark has \emph{factual error diversity}, as it covers all types of errors described in the typology in \Sref{sec:typology}, and \emph{data diversity} as it combines 2250 summaries from different systems and datasets. Our annotations go beyond binary labels of factuality on a summary by providing fine-grained category annotations for every sentence. This allows us to determine how well each metric can capture each type of error. Furthermore, through averaging of sentence level judgements, we can also obtain a factuality scores (0 to 1 range) for a summary. To measure the degree that automated metrics capture a certain characteristic, we compute their correlation with human judgements and report Pearson correlation and Spearman rank correlation along with their p-values. We evaluate different classes of metrics against the FRANK{} benchmark. We select four general summarization metrics. ROUGE, BLEU, and Meteor are n-gram based metrics and computed with respect to the reference summary. BERTScore \citep{bertscore} computes BERT \citep{bert} contextual embeddings on summary and source article and measures distances between matched embeddings. We select five metrics focused on factuality. As \citet{goodrich19}, we use a simple OpenIE \citep{openie} baseline. This involves extracting OpenIE triples and matching them through sentence embeddings \citep{sentence-bert}. FactCC \citep{factcc} and DAE \citep{goyal2020evaluating} are entailment based metrics. FactCC operates with sentences as claims, while DAE uses dependency level entailment. FEQA \citep{feqa} and QAGS \citep{factqa} are two question answering and generation metrics (QGA). More details on the differences between these metrics is in \Sref{sec:metrics_appendix}. \begin{figure} \centering \includegraphics[width=\columnwidth]{plots/partial_correlation_plot_model_name.pdf} \caption{Correlation between metrics and human judgement on subsets of data. The $x$ and $y$ axis represent the human judgement the metric scores respectively. The red line is a linear regression fitted on full data. Each dotted line is a linear regression fitted on a model-dataset subset. Each colored point has coordinates equal to average factuality judgement, and metric score for its corresponding partition.} \label{fig:partial_correlations} \vspace{-3mm} \end{figure} \subsection{Controlling for Dataset Biases} Since our benchmark contains diverse summaries from different datasets and models, dataset biases can hamper accurate reporting. In \autoref{fig:partial_correlations}, we visually show correlations between two factuality metrics (FEQA and FactCC) and human judgement on the entire data and on partitions of the data. For both metrics, we notice that the slope (an unscaled measure of correlation) of the line fitted through the entire data (red line) is significantly larger. In FEQA, the dotted lines (fitted on subsets of the data of each model and dataset) are almost horizontal. This likely indicates the presence of a confounding variable associated with the properties of each system and dataset. This can lead to false measures of high correlation if not accounted for. To address this, we suggest to control for confounding variables using partial correlations. We include details on partial correlations in the Appendix. In this case, both the system and the dataset are taken to be confounding variables. \subsection{Results} In \autoref{tab:results}, we report the partial Pearson correlation and Spearman rank correlation coefficients with human judgements for each metric, along with their $p$-values indicating statistical significance. \paragraph{How do different metrics correlate with human judgements? } From \autoref{tab:results} we observe that all metrics exhibit low correlations with human judgements of factuality. The best metrics overall are FactCC with 0.20 Pearson and 0.30 Spearman correlation and BERTScore P with 0.27 Pearson and 0.35 Spearman correlation. Interestingly, we observe that general summarization metrics BLEU, Rouge, and METEOR, and the OpenIE baseline have statistically significant correlations with factuality, close to FactCC ($\rho=0.14$ for Rouge-1 and METEOR versus $\rho=0.20$ for FactCC). The entailment metrics (FactCC and DAE) and contextual embedding method (BERTScore) have the highest correlations and are statistically significant. The two QGA metrics have lower overall correlation. FEQA's correlation is not statistically significant. QAGS has low, but significant correlation of $\rho=0.06$. \begin{figure*} \centering \includegraphics[width=\linewidth]{plots/correlation_model_data_analysis.pdf} \caption{Partial Pearson correlation on different partitions of the data. Entailment metrics have highest correlation on pretrained models in the CNN/DM dataset. Their performance degrades significantly on XSum.} \label{fig:correlation_model_data_analysis} \end{figure*} \paragraph{How well do different metrics capture errors in different datasets?} In \autoref{fig:correlation_model_data_analysis}, we observe that entailment metrics have significantly higher partial Pearson correlation on the CNN/DM dataset than XSum where their correlation is reduced by a factor of four. QAGS and the OpenIE baseline have similar behavior. This suggests that these metrics capture the error types from CNN/DM better that those from XSum. Specifically, XSum has uniquely high Out of Article ({OutE}{}) errors which they might not capture well. This also highlights the importance of data diversity in building and benchmarking factuality metrics to avoid overfitting to certain types of errors. \paragraph{How well do different metrics capture errors from pretrained and non-pretrained models?} On the CNN/DM dataset we observe that entailment metrics and QAGS perform significantly better on non-pretrained models. This indicates that the artificial factual errors on which entailment metrics are trained on are closest to the mistakes that non-pretrained models make. This also suggests that the errors made by pretrained models might be more difficult to capture by these metrics. These trends are less clear on the XSum dataset which we again attribute to high Out of Article ({OutE}{}) errors in the pretrained and non-pretrained models (ref \autoref{fig:factcategory}) \begin{figure} \centering \includegraphics[width=\columnwidth]{plots/correlation_category_analysis.pdf} \caption{Variation in partial Pearson correlation when omitting error types. Higher variation indicates greater influence of an error type in the overall correlation.} \label{fig:correlation_category_analysis} \vspace{-3mm} \end{figure} \subsection{Error Analysis} \autoref{fig:correlation_model_data_analysis} shows partial Pearson correlation on six subsets of the data. To understand capabilities of metrics across the broad categories of errors (semantic frame errors, discourse errors, and content verifiability errors) we perform an ablation study. For each category, we compute the variation in partial correlation with errors from that category omitted. In \autoref{fig:correlation_category_analysis}, we visualize the influence of a given type of error using the variation for each metric and category. A higher positive bar indicates that the error type was a significant contributer to the overall correlation (or metric highly correlates with error) causing the correlation without it to drop. \paragraph{General Summarization metrics} Unsurprisingly, we observe that Rouge L is best correlated with content verifiability errors (which contains Out of Article Errors) as n-gram matches detect them. Rouge L has negative correlation with semantic frame errors and low correlation with discourse level errors indicating that n-gram matching fails to capture them. We observe that OpenIE is more correlated with semantic frame errors. The metric matches entities and verifies the predicate that relates them and hence is able to capture semantic frame errors. We observe that BERTScore's high correlation is primarily due to its ability to capture content verifiability errors while it has low correlation on semantic frame errors. \paragraph{QGA metrics} Both QGA metrics have negative correlation with discourse errors suggesting that QGA metrics are not able to capture coreference errors or discourse link errors potentially due to the entity oriented questions in their training data. FEQA additionally is also negatively correlated with semantic frame errors and has low positive correlation with content verifiability errors. In contrast QAGS is best correlated with semantic frame errors. \paragraph{Entailment metrics} Both entailment metrics correlate well with semantic frame and content verifiability errors. DAE has the highest correlation of all metrics with discourse errors suggesting that entailment at the dependency level can help model discourse errors ({CorefE}{} and {LinkE}{}). FactCC is nearly uncorrelated in this category, indicating that artificially generated factual errors need to go beyond simple pronoun swaps to train models to capture discourse errors. FactCC is best at capturing semantic frame among all metrics evaluated. \section{Related Work} \label{sec:related_work} \citet{sumcriticaleval} and \citet{fabbri2020summeval} find that standard n-gram based metrics have low correlation with human judgements of factuality. Motivated by this, several automated metrics falling in two paradigms were proposed to improve the evaluation of factuality. \paragraph{Entailment Classification} \citet{goodrich19,factcc,factentailment, goyal2020evaluating} model factuality as entailment classification breaking down the summary into smaller units, such as sentences, which are verified against the original article. However, modeling factuality as a classification task requires supervision on factual and hallucinated data. FactCC \citep{factcc} is trained on the CNN/DM dataset augmented with four types of artificial mistakes as supervision. \paragraph{Question Generation and Answering (QGA)} FEQA \citep{feqa} and QAGS \citep{factqa} are two metrics which reduce factuality evaluation to question generation and answering. These methods use a question generation model to obtain questions from the output summary and a question answering model to answer them, separately using the article and the output summary. \paragraph{Prior Efforts on Factuality Annotations of Summaries} \citet{fabbri2020summeval} and \citet{factentailment} have collected annotations on the CNN/DM and XSum dataset respectively. In this work we cover both datasets to ensure greater data diversity. Other efforts \citep{factcc, factqa, feqa} were smaller in scale \citet{feqa} and \citet{factcc} annotated 200 and 503 sentences while \citet{factqa} annotated 470 summaries (we collect judgements on 2250 summaries). Crucially, all previous efforts portray factuality as a binary label without variations in degree or type of factual errors. \section{Conclusion} In this work we provide a linguistically grounded typology of factual errors which we use to collect FRANK{}, a dataset of human annotations of 2250 summaries covering both CNN/DM and XSum datasets. We use FRANK{} to assess the factuality of summarization systems and benchmark recently proposed factuality metrics highlighting the types of errors they can capture. With the FRANK{} benchmark we have started moving away from a summary-level binary understanding of factuality. \section{Ethical Considerations} We have collected crowd annotations using the Amazon Mechanical Turk platform. Workers were paid 50\% more than the average American minimum wage and offered additional bonuses as an incentive to maintain high quality work. No information about the workers will be released and worker IDs will be anonymized. \section*{Acknowledgements} The authors are grateful to the anonymous reviewers for their feedback, and to Anjalie Field, Rishabh Joshi, Alissa Ostapenko, Dheeraj Rajagopal, Evangelia Spiliopoulou, Shuly Wintner, and the members of the Tsvetshop group for their invaluable feedback and support in various stages of the project. This material is based upon work supported by the DARPA CMO under Contract No.~HR001120C0124, and in part by the National Science Foundation under Grants No.~IIS2040926 and No.~IIS2007960. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily state or reflect those of the United States Government or any agency thereof.

8533e8c776de5e82112a4c0e7bf93b479286e38c

\section{Introduction} Electromagnetic (EM) waves are widely used in plasma applications, including magnetic confinement fusion~\cite{Stix92,Freidberg10,Wesson11} and inertial confinement fusion~\cite{Lindl04,Craxton15}. Accurately modeling how EM waves propagate in plasma is therefore of upmost importance. Full-wave modeling, that is, directly solving Maxwell's equations with appropriate source terms, can be computationally expensive. Instead, the geometrical-optics (GO) approximation is often used to quickly calculate the wave amplitude along the GO rays that illuminate the region of interest~\cite{Tracy14, Kravtsov90}. The obtained amplitude profile can then be used as a source term in calculations of macroscopic plasma equilibrium~\cite{Prater08,Poli18a}. Design studies for EM-wave systems are often performed in this manner~\cite{Poli18a,Poli15,Poli16,Lopez18a}. Unfortunately, GO solutions develop singularities at caustics such as cutoffs or focal points~\cite{Kravtsov93,Berry80b}. This is an issue for applications in which caustics play a central role, such as initializing spherical tokamak plasmas~\cite{Peng86,Peng00,Ono15a} via electron cyclotron resonance heating~\cite{Erckmann94,Prater04} (where the time-evolution of caustic surfaces directly defines the window of operation~\cite{Lopez18a}), or driving current in overdense plasmas via mode conversion to the electron Bernstein wave~\cite{Ram00,Shiraiwa06,Uchijima15,Seltzman17,Lopez18b,Laqua07,Preinhaelter73,Hansen85,Mjolhus84,Laqua03,Shevchenko07} (where the field structure near the EM wave cutoffs must be precisely resolved to obtain accurate mode-conversion efficiency estimates). Reduced modeling of these processes requires a more advanced machinery than traditional GO. In response to this need, a new reduced theory called metaplectic GO (MGO) has been recently developed that leads to solutions that are finite at caustics~\cite{Lopez20a,Lopez21a}. By default, MGO yields an integral representation of the wavefield, which can be approximated analytically to some extent but in general must be evaluated numerically. Unfortunately, the integrands in MGO are highly oscillatory, so standard integration methods are insufficient~\cite{Deano17}. Special numerical algorithms tailored to MGO are needed. As part of ongoing work on MGO algorithms~\cite{Lopez20a,Lopez21a,Lopez19a,Lopez21b}, here we present a quadrature rule for calculating MGO integrals based on numerical steepest-descent integration~\cite{Deano09}. This algorithm emerges naturally from the MGO framework in that MGO integrals always contain saddlepoints that correspond to the ray contributions to the wavefield. We benchmark our algorithm on a class of examples in which the MGO integral contains a single isolated saddlepoint of various degeneracy, physically representing a wavefield either far from a caustic or at the critical point of a cuspoid-type caustic. We then apply our algorithm to the classic problem of an EM-wave reflecting off an isolated cutoff as governed by Airy's equation~\cite{Stix92,Tracy14}. We show that the numerical MGO solution agrees with the exact result amazingly well, much better than the analytical approximation to the MGO integral that was previously derived~\cite{Lopez20a}. This paper is organized as follows. In \Sec{sec:background} the basic machinery of MGO is summarized, and the various types of caustics one expects to encounter are briefly surveyed. Section~\ref{sec:algorithm} constitutes the bulk of our paper, first introducing steepest-descent integration and Gaussian quadrature, then proceeding to derive our new quadrature rule. Benchmarking examples are provided in \Sec{sec:examples}, and \Sec{sec:concl} summarizes our main results. \section{Metaplectic geometrical optics and caustics} \label{sec:background} \subsection{A brief overview} Here we provide a brief overview of MGO; for more details, see \Refs{Lopez20a,Lopez21a}. Let $\psi(\Vect{x})$ be a scalar stationary wavefield in a plasma described by an $N$-dimensional ($N$-D) Euclidean coordinate system. (Generalizations to arbitrary metric are discussed in \Ref{Dodin19}). Neglecting nonlinear effects, the governing wave equation for $\psi(\Vect{x})$ is most generally written in the following integral form: \begin{equation} \int \mathrm{d} \Vect{x}' \, D(\Vect{x}, \Vect{x}') \psi(\Vect{x}') = 0 , \end{equation} \noindent where the integral kernel $D(\Vect{x}, \Vect{x}')$ is determined from the linear dielectric response of the plasma $\epsilon(\Vect{x}, \Vect{x}')$ in a known manner~\cite{Tracy14,Stix92}. For example, transverse EM waves have $D(\Vect{x}, \Vect{x}')$ given by \begin{equation} D(\Vect{x}, \Vect{x}') = - \nabla^2 \delta(\Vect{x} - \Vect{x}') - \epsilon(\Vect{x}, \Vect{x}') , \end{equation} \noindent where $\nabla^2$ is the Laplacian operator with respect to $\Vect{x}$. In the traditional GO limit, when $\psi(\Vect{x})$ is highly oscillatory, adopting the eikonal partition \begin{equation} \psi(\Vect{x}) = \phi(\Vect{x}) e^{ i \theta(\Vect{x}) } \end{equation} \noindent ultimately leads to the local dispersion relation that governs the phase function $\theta(\Vect{x})$, \begin{subequations} \begin{equation} \Symb{D}[\Vect{x}, \nabla \theta(\Vect{x})] = 0 , \label{eq:dZERO} \end{equation} \noindent and the transport equation that governs the envelope function $\phi(\Vect{x})$, \begin{equation} \Vect{v}(\Vect{x})^\intercal \pd{\Vect{x}} \log \phi(\Vect{x}) = - \frac{1}{2} \pd{\Vect{x}} \cdot \Vect{v}(\Vect{x}) . \label{eq:GOenv} \end{equation} \end{subequations} \noindent Here, we have defined the local group velocity as \begin{equation} \Vect{v}(\Vect{x}) \doteq \pd{\Vect{k}} \Symb{D}\left[\Vect{x}, \nabla \theta(\Vect{x}) \right] , \end{equation} \noindent and we have introduced the dispersion function $\Symb{D}(\Vect{x}, \Vect{k})$, obtained from $D(\Vect{x}, \Vect{x}')$ using the Wigner transform% ~\footnote{Strictly speaking, the Wigner transform is a mapping between Hilbert-space operators and phase-space functions. In \Eqs{eq:wigner} and \eq{eq:wigFUNC}, we choose to represent the abstract operators $\hat{D}$ and $\ket{\psi} \bra{\psi}$ explicitly by their configuration-space ($\Vect{x}$-space) matrix elements for convenience.}% \begin{equation} \Symb{D}(\Vect{x}, \Vect{k}) \doteq \int \mathrm{d} \Vect{s} \, e^{i\Vect{k}^\intercal \Vect{s}} \, D \left( \Vect{x} - \frac{\Vect{s}}{2} , \Vect{x} + \frac{\Vect{s}}{2} \right) . \label{eq:wigner} \end{equation} \noindent (Vectors are interpreted as row vectors unless explicitly transposed via $^\intercal$, so $\Vect{k}^\intercal \Vect{s} = \Vect{k} \cdot \Vect{s} $. Also, the symbol $\doteq$ denotes definitions.) Note that the Wigner transform of the two-point correlation function $\psi(\Vect{x})\psi^*(\Vect{x}')$, {i.e., } \begin{equation} W_\psi(\Vect{x}, \Vect{k}) \doteq \int \frac{\mathrm{d} \Vect{s}}{(2\pi)^N} \, e^{i\Vect{k}^\intercal \Vect{s}} \, \psi \left( \Vect{x} - \frac{\Vect{s}}{2} \right) \psi^* \left( \Vect{x} + \frac{\Vect{s}}{2} \right) , \label{eq:wigFUNC} \end{equation} \noindent acts as a phase-space (quasi-)distribution function for the field intensity, satisfying~\cite{Case08} \begin{subequations} \begin{align} |\psi(\Vect{x})|^2 &= \int \mathrm{d} \Vect{k} \, W_\psi(\Vect{x}, \Vect{k}) , \\ |\fourier{\psi}(\Vect{k})|^2 &= \int \mathrm{d} \Vect{x} \, W_\psi(\Vect{x}, \Vect{k}), \end{align} \end{subequations} \noindent where $\fourier{\psi}$ is the Fourier transform of $\psi$. The local dispersion relation \eq{eq:dZERO} is commonly solved via the ray equations \begin{equation} \pd{\xi} \Vect{x} = \pd{\Vect{k}} \Symb{D}(\Vect{x}, \Vect{k}) , \quad \pd{\xi} \Vect{k} = - \pd{\Vect{x}} \Symb{D}(\Vect{x}, \Vect{k}) , \label{eq:rays} \end{equation} \noindent where the evolution of $\Vect{k}$ along a ray (considered as a vector field over $\Vect{x}$) is constrained by \begin{subequations}% \label{eq:rayCONSTR} \begin{equation} \Vect{k} = \nabla \theta(\Vect{x}) , \end{equation} \noindent and the initial conditions must satisfy \begin{equation} \Symb{D}(\Vect{x}, \Vect{k}) = 0 . \end{equation} \end{subequations} \noindent The rays encoded by \Eq{eq:rays} naturally reside in the $2N$-D ray phase space $(\Vect{x}, \Vect{k})$ with the wavevector $\Vect{k}$ serving as the ray momentum. The constraints \eq{eq:rayCONSTR} define an $N$-D manifold in this phase space called the `dispersion manifold' on which $\psi$ is asymptotically confined. More specifically, if we parameterize this manifold as the zero set of $N$ independent functions~\cite[pp.~467--468]{Tracy14} \begin{equation} \Vect{\Symb{M}}(\Vect{x}, \Vect{k}) \doteq (\Symb{M}_1(\Vect{x}, \Vect{k}), \ldots \Symb{M}_N(\Vect{x}, \Vect{k})) , \quad \Symb{M}_1 \equiv \Symb{D} , \end{equation} \noindent then $W_\psi(\Vect{x}, \Vect{k}) \sim \delta[\Vect{\Symb{M}}(\Vect{x},\Vect{k})]$ in the GO limit~\cite{Berry77a,Berry77b}. Correspondingly, $|\phi(\Vect{x})|^2$ diverges in the GO limit (\Fig{fig:wigner}) at locations where the projection of the dispersion manifold onto $\Vect{x}$-space is singular, {i.e., } where% ~\footnote{Note that $\pd{\Vect{k}} \Vect{\Symb{M}}\left[\Vect{x}, \nabla \theta(\Vect{x}) \right] = \pd{\Vect{\tau}} \Vect{x}(\Vect{\tau}) $ when $\Vect{\tau}$ is generated by $\Vect{\Symb{M}}$, {i.e., } $\pd{\Vect{k}} \Symb{M}_j\left[\Vect{x}, \nabla \theta(\Vect{x}) \right] = \pd{\tau_j} \Vect{x}(\Vect{\tau})$.}% \begin{equation} \det \pd{\Vect{k}} \Vect{\Symb{M}}\left[\Vect{x}, \nabla \theta(\Vect{x}) \right] = 0, \end{equation} \noindent or equivalently, where \begin{equation} \det \partial^2_{\Vect{x} \Vect{x}} \theta(\Vect{x}) \to \infty . \end{equation} \noindent These locations are called `caustics'~\cite{Kravtsov93}, and they typically occur where distinct branches of the dispersion manifold coalesce. Notice, though, that the ray equations \eq{eq:rays} in these regions remain well-defined. \begin{figure*} \centering \begin{overpic}[width = \linewidth,trim={2mm 3mm 3mm 2mm},clip]{collage.pdf} \put(32,16){\textbf{\large(a)}} \put(62,16){\textbf{\large(b)}} \put(92,16){\textbf{\large(c)}} \put(32,10){\textbf{\large(d)}} \put(62,10){\textbf{\large(e)}} \put(92,10){\textbf{\large(f)}} \end{overpic} \caption{Top row -- the phase-space distribution of the wave intensity in the phase space $(x, k)$ for a wave near a fold-type Airy caustic (see \Sec{sec:exAIRY}): \textbf{(a)} The exact Wigner function $W_\psi \propto \airyA[ \Symb{D}(x,k)]$; \textbf{(b)} GO solution $W_\psi \sim \delta\left[ \Symb{D}(x,k) \right]$; \textbf{(c)} GO solution in the phase space rotated by $\pi/2$. Bottom row -- the corresponding wave fields: \textbf{(d)} the exact field $|\psi(x)|^2$; \textbf{(e)} the field intensity $|\phi(x)|^2 \doteq \langle|\psi(x)|^2 \rangle $ in the GO approximation (where the angular brackets denote averaging over the local wavelength); \textbf{(f)} the field intensity in the spectral representation, $|\fourier{\phi}(k)|^2 \doteq \langle|\fourier{\psi}(k)|^2 \rangle$. The dispersion manifold $\Symb{D}(x,k) = 0$ is shown as the black dashed line. Clearly, $|\phi(x)|^2$ diverges at $x\to 0$, but $|\fourier{\phi}(k)|^2$ is well-behaved everywhere.} \label{fig:wigner} \end{figure*} The singularity of $\phi$ at caustics signifies that the traditional GO approximation fails in regions where the projection of the dispersion manifold onto $\Vect{x}$-space is ill-behaved. However, working in $\Vect{x}$-space is not a necessity. One can instead formulate GO on more general phase-space planes specifically chosen to avoid singular projections of the dispersion manifold as the wave propagates. This is the main idea of the MGO method as proposed in \Refs{Lopez20a, Lopez21a} (see \Fig{fig:wigner}). To develop this idea in more detail, let us introduce an $N$-D coordinate system $\Vect{\tau}$ such that the dispersion manifold can be parameterized as $(\Vect{x}(\Vect{\tau}), \Vect{k}(\Vect{\tau}) )$. We choose $\tau_1 = \xi$, the longitudinal coordinate along the ray trajectories \eq{eq:rays}. The remaining $\tau_2, \ldots, \tau_N$, which can be chosen as the coordinates generated by $\Symb{M}_2, \ldots, \Symb{M}_N$, parameterize the initial conditions to \Eqs{eq:rays} along the dispersion manifold, {e.g., } the remaining spatial coordinates. Let us consider a specific point $\Vect{\tau} = \Vect{t}$ on the dispersion manifold. The optimal rotation to avoid projection singularities is the one that aligns the tangent plane of the dispersion manifold at $\Vect{t}$, denoted $\Vect{X}_\Vect{t}$-space, with $\Vect{x}$-space. This is accomplished by performing the following linear coordinate transformation of the phase space: \begin{equation} \begin{pmatrix} \Vect{X}_\Vect{t} \\ \Vect{K}_\Vect{t} \end{pmatrix} = \Mat{S}_\Vect{t} \begin{pmatrix} \Vect{x} \\ \Vect{k} \end{pmatrix} , \quad \Mat{S}_\Vect{t} \doteq \begin{pmatrix} \Mat{A}_\Vect{t} & \Mat{B}_\Vect{t} \\ \Mat{C}_\Vect{t} & \Mat{D}_\Vect{t} \end{pmatrix} , \label{eq:rotPHASE} \end{equation} \noindent where $\Mat{S}_\Vect{t}$ is the $2N \times 2N$ unitary symplectic matrix that rotates phase space to align $\Vect{X}_\Vect{t}$-space with $\Vect{x}$-space. (The matrices $\Mat{A}_\Vect{t}$, $\Mat{B}_\Vect{t}$, $\Mat{C}_\Vect{t}$, and $\Mat{D}_\Vect{t}$ are each $N \times N$.) An explicit construction of $\Mat{S}_\Vect{t}$ using ray trajectories near $\Vect{t}$ is provided by the `symplectic Gram--Schmidt' algorithm of \Ref{Lopez20a}. Ultimately, repeating these `optimal' rotations for all points on the dispersion manifold, synthesizing the resulting GO solutions, and transforming them back to the original phase space yields the MGO solution~\cite{Lopez20a} \begin{equation} \psi(\Vect{x}) = \sum_{\Vect{t} \in \Vect{\tau}(\Vect{x})} \mc{N}_\Vect{t}(\Vect{x}) \, \Upsilon_\Vect{t}(\Vect{x}) , \label{eq:MGO} \end{equation} \noindent where $\Vect{\tau}(\Vect{x})$ is the function inverse of $\Vect{x}(\Vect{\tau})$; accordingly, the sum is taken over all branches of $\Vect{k}(\Vect{x}) \doteq \Vect{k}[\Vect{\tau}(\Vect{x})]$. In \Eq{eq:MGO}, we have introduced the integral function \begin{align} \Upsilon_\Vect{t}(\Vect{x}) &\doteq \int_{\cont{0}} \mathrm{d} \Vect{\varepsilon} \, \Psi_\Vect{t}\left[ \Vect{\varepsilon} + \Vect{X}_\Vect{t}(\Vect{t}) \right] \exp\left( - \frac{i}{2} \Vect{\varepsilon}^\intercal \Mat{D}_\Vect{t} \Mat{B}_\Vect{t}^{-1} \Vect{\varepsilon} \right) \nonumber\\ &\hspace{15mm}\times \exp\left\{ i \Vect{\varepsilon}^\intercal \Mat{B}_\Vect{t}^{-\intercal} \left[ \Vect{x} - \Mat{D}_\Vect{t}^\intercal \Vect{X}_\Vect{t}(\Vect{t}) \right] \right\} , \label{eq:upsilon} \end{align} \noindent where $\Psi_\Vect{t}(\Vect{X}_\Vect{t})$ is the GO solution in the rotated phase space \eq{eq:rotPHASE}. The integration is performed along the steepest-descent contour that passes through the saddlepoint $\Vect{\varepsilon} = \Vect{0}$ (\Sec{sec:steepDESC}), which importantly means that $\Upsilon_\Vect{t}(\Vect{x})$ is not generally a unitary mapping of $\Psi_\Vect{t}$ unless $\cont{0}$ can be deformed to lie along the real axis% ~\footnote{The fact that $\Upsilon_\Vect{t}(\Vect{x})$ is not generally unitary does not greatly diminish the accuracy of MGO, since the multiple contributing $\Upsilon_\Vect{t}(\Vect{x})$ are summed in such a manner to keep the solution finite [\Eq{eq:MGO}].}. % Note that \Eq{eq:upsilon} requires $\Mat{B}_\Vect{t}$ to be invertible; this is done for simplicity, and the generalization to arbitrary $\Mat{B}_\Vect{t}$ is provided in \Ref{Lopez21a}. Also, the prefactor $\mc{N}_\Vect{t}(\Vect{x})$ in \Eq{eq:MGO} can be simply evolved along the ray trajectories using the formulas provided in \Ref{Lopez20a}. Efficiently computing $\Upsilon_\Vect{t}(\Vect{x})$ is comparatively less understood, and shall be the focus of the remainder of this work. \subsection{Caustics as optical catastrophes} \label{sec:caustic} \begin{table*} \centering \begin{tabular}{| c | c | c | c | c |} \multicolumn{1}{c}{\textbf{Name}} & \multicolumn{1}{c}{$\alpha$} & \multicolumn{1}{c}{$m$} & \multicolumn{1}{c}{$M$} & \multicolumn{1}{c}{$f_\alpha(\Vect{\kappa}, \Vect{y})$} \\ \hline No caustic & $A_1$ & $0$ & $1$ & $\kappa_1^2$ \\ Fold & $A_2$ & $1$ & $1$ & $\kappa_1^3 + y_1 \kappa_1$ \\ Cusp & $A_3$ & $2$ & $1$ & $\kappa_1^4 + y_2 \kappa_1^2 + y_1 \kappa_1$ \\ Swallowtail & $A_4$ & $3$ & $1$ & $\kappa_1^5 + y_3 \kappa_1^3 + y_2 \kappa_1^2 + y_1 \kappa_1$ \\ Hyperbolic umbilic & $D_4^+$ & $3$ & $2$ & $\kappa_1^3 + \kappa_2^3 + y_3 \kappa_1 \kappa_2 + y_2 \kappa_2 + y_1 \kappa_1$ \\[1mm] Elliptic umbilic & $D_4^-$ & $3$ & $2$ & $\kappa_1^3 - 3 \kappa_1 \kappa_2^2 + y_3(\kappa_1^2 + \kappa_2^2) + y_2 \kappa_2 + y_1 \kappa_1$ \\[1mm] \hline \end{tabular} \caption{A complete list of the normal-form generators $f_\alpha(\Vect{\kappa}, \Vect{y})$ for caustics with codimension $m \le 3$~\cite{Berry80b, Arnold83, Kravtsov93}. The interference patterns that correspond to these caustics can be viewed in their entirety when the number of spatial dimensions $N = 3$. For each caustic, $\alpha$ is the Arnold label~\cite{Arnold83}, $m$ is the codimension, and $M$ is the corank.} \label{tab:caustic} \end{table*} The behavior of $\Upsilon_\Vect{t}(\Vect{x})$ near caustics can be broadly understood using catastrophe theory~\cite{Berry80b,Kravtsov93, Poston96}. This theory provides a classification system for caustics based on their codimension, that is, the minimum number of spatial dimensions in which they can be observed. For example, the simplest type of caustic is the fold caustic, which occurs when a wave encounters a cutoff. The fold caustic has codimension $1$, so it can be observed in $N$-D systems with $N \ge 1$. On the other hand, the cusp caustic, which occurs at a focal point, has codimension $2$ and can thus only be observed for $N \ge 2$. This supports the intuition that cutoffs are well-described by $1$-D models like Airy's equation, but foci are inherently $2$-D. There are three main advantages to using the catastrophe classification system to study caustics: \textbf{(i)} Only `structurally stable' caustics that are robust under small perturbations are included. These are the caustics that are most physically relevant, since a structurally unstable caustic will be destroyed by any imperfections in the experimental setup (which are of course unavoidable). For example, an EM wave propagating in an unmagnetized cold plasma with a linear density profile $n(x)$ will have a cutoff at some $x_c$. If the density is perturbed from $n(x)$ to $\tilde{n}(x)$ by some global motion of the plasma, the cutoff location will shift from $x_c$ to $\tilde{x}_c$, but will generally not disappear; hence, the cutoff (fold caustic) is `structurally stable'. \textbf{(ii)} There are only a finite number of distinct caustic types that are stable in a given number of dimensions. For example, only six different caustics can occur in $3$-D (including `no caustic'; see Table~\ref{tab:caustic}). \textbf{(iii)} General properties of a given caustic type can be determined by studying a single member in detail, often chosen to be the `simplest' member (the so-called `normal-form generator' of the caustic class; see below). These three results from catastrophe theory greatly reduce the work required to validate any new method in catastrophe optics; indeed, a new method for modeling caustics need only be tested on three different caustics to be fully viable in $2$-D, or on six different caustics for $3$-D. In our case, a numerical quadrature rule for $\Upsilon_\Vect{t}(\Vect{x})$ can be validated on the standard integrals of catastrophe theory, which are integral representations for caustic wavefields and take the general form \begin{equation} I_\alpha(\Vect{y}) \doteq \int \mathrm{d} \Vect{\kappa} \, \exp\left[ i f_\alpha(\Vect{\kappa}, \Vect{y}) \right] , \label{eq:Ialpha} \end{equation} \noindent where $\Vect{y}$ is an $m$-D collection of `external' (or `control') variables, $\Vect{\kappa}$ is an $M$-D collection of `internal' (or `state') variables, and $\alpha$ labels the type of caustic. The integers $m \le N$ and $M \le N$ are the `codimension' and `corank' of the caustic, respectively, and the function $f_\alpha(\Vect{\kappa}, \Vect{y})$ is the normal-form generator for a type-$\alpha$ caustic. For example, the fold caustic, also called the $A_2$ caustic in Arnold's nomenclature~\cite{Arnold83}, has $m = 1$, $M = 1$, and \begin{equation} f_{A_2}(\kappa_1, y_1) = \kappa_1^3 + y_1 \kappa_1 . \end{equation} \noindent The corresponding $I_{A_2}(y_1)$ is proportional to the Airy function $\airyA(y_1/ \sqrt[3]{3})$~\cite[pp.~194--213]{Olver10a}. See Table~\ref{tab:caustic} for more examples. Note that if $M < N$ such that only a subset of the integration variables of $\Upsilon_\Vect{t}$ are included in the standard integral $I_\alpha$, then by the splitting lemma of catastrophe theory~\cite{Berry80b,Poston96}, the remaining $N - M$ integrals contained in $\Upsilon_\Vect{t}$ are decoupled and involve phase functions that are quadratic at most, and thereby trivially integrated. For practical problems, $I_\alpha(\Vect{y})$ typically represents only the local behavior of a given type-$\alpha$ caustic. The global behavior can sometimes be modeled using the method of `uniform approximation'~\cite{Olver10a,Chester57,Ludwig66}. However, this method relies on \textbf{(i)} the caustic type being known beforehand (which is fine for interpretive but not predictive simulations) and \textbf{(ii)} only a single caustic being present. Indeed, the elementary catastrophes mentioned here often combine to form `caustic networks', an example being an EM wave focused on a cutoff producing a fold-cusp network. It might be possible to infer basic properties of such caustic networks from the constituent members, but complete understanding can only be achieved by considering the caustic network as a whole, which is very difficult to do analytically. Hence, a robust numerical scheme for computing catastrophe integrals that does not assume any specific caustic structure is needed. \section{Gauss--Freud quadrature for steepest-descent integration} \label{sec:algorithm} \subsection{Steepest-descent method} \label{sec:steepDESC} As seen from \Eq{eq:Ialpha} and Table~\ref{tab:caustic}, we generally expect $I_\alpha(\Vect{y})$ to involve a highly oscillatory integrand when $\Vect{y}$ and $\Vect{\kappa}$ are both real. These rapid oscillations would make the direct evaluation of $\Upsilon_\Vect{t}(\Vect{x})$ analytically and numerically challenging, if not for the fact that $\Upsilon_\Vect{t}(\Vect{x})$ is evaluated along the steepest-descent contour $\cont{0}$ [\Eq{eq:upsilon}]. Along $\cont{0}$, oscillatory terms become exponentially decaying terms, reinstating the viability of standard numerical integration methods like Gaussian quadrature (\Sec{sec:gaussQUAD}). However, this simplification is contingent on the ability to determine $\cont{0}$ for arbitrary wavefields. Let us therefore characterize the steepest-descent contours of the standard forms $I_\alpha(\Vect{y})$. For simplicity, we restrict attention to $1$-D integrals, {i.e., } $M = 1$ in \Eq{eq:Ialpha}. For integrals of the form \begin{equation} I(\Vect{y}) = \int \mathrm{d} \kappa \, g(\kappa, \Vect{y}) \exp[i f(\kappa, \Vect{y}) ] \label{eq:SDMint} \end{equation} \noindent [where we have generalized \Eq{eq:Ialpha} to include a slowly varying amplitude $g(\kappa, \Vect{y})$], the steepest-descent contours at fixed $\Vect{y}$ are by definition the streamlines of $\nabla \Im(f)$ in the complex $\kappa$ plane $\mbb{C}^1$, where $\nabla \doteq (\pd{\Re(\kappa)}, \pd{\Im(\kappa)} )$. (Here, $\Re$ and $\Im$ denote the real and imaginary parts of a complex function.) When $f$ is analytic in $\kappa$, a more useful definition arises from the Cauchy--Riemann relation~\cite{Rudin87equation} \begin{equation} i \pd{\Re(\kappa)} f(\kappa, \Vect{y}) = \pd{\Im(\kappa)} f(\kappa, \Vect{y}) , \label{eq:cauchy1} \end{equation} \noindent which implies that $\nabla \Re(f)$ and $\nabla \Im(f)$ are orthogonal, {i.e., } \begin{equation} \nabla \Re(f) \cdot \nabla \Im(f) = 0 . \label{eq:cauchy2} \end{equation} \noindent Therefore, the streamlines of $\nabla \Im(f)$ that pass through a given point $\kappa_0$ are also the set of points in $\mbb{C}^1$ that satisfy the implicit equation \begin{equation} \Re\left[ f(\kappa, \Vect{y}) \right] = \Re\left[ f(\kappa_0, \Vect{y}) \right]. \end{equation} The steepest-descent contours are almost-everywhere smooth curves, although non-differentiable kinks can occur at special points where \Eq{eq:cauchy2} is indeterminate, {i.e., } where $\nabla \Re\left(f \right) = \nabla \Im\left(f \right) = \Vect{0}$% ~\footnote{By \Eq{eq:cauchy1}, $\nabla \Re\left(f \right)$ and $\nabla \Im\left(f \right)$ always vanish simultaneously.}. % At these points [which are saddlepoints per \Eq{eq:cauchy1}% ~\footnote{This follows by differentiating \Eq{eq:cauchy1} to show that the Hessian matrices for $\Re(f)$ and $\Im(f)$ are symmetric and traceless, thereby possessing two real eigenvalues of opposite sign.}% ], the direction of $\nabla \Im\left(f \right)$ generally changes abruptly, producing the aforementioned kinks that require special parameterization. As shown later, such parameterization can be done by treating the kink as two independent curves that intersect at a finite angle. Although saddlepoints are `rare' in that they occur at isolated points in $\mbb{C}^1$, they are often of primary interest due to their prominent role in asymptotic wave theory. More specifically, each saddlepoint of $f$ encodes the contribution to $I(\Vect{y})$ from a single corresponding GO ray. Consequently, a saddlepoint $\kappa_s(\Vect{y})$ will be real when $\Vect{y}$ is in the lit region of a caustic, but may be complex when $\Vect{y}$ is in the shadow region. Also, a caustic occurs at the specific values of $\Vect{y}$, denoted $\Vect{y}_c$, such that multiple saddlepoints coalesce and consequently, \begin{equation} \pd{\kappa}^2 f[\kappa_s(\Vect{y}_c), \Vect{y}_c] = 0 . \label{eq:caustic} \end{equation} Having now characterized the general behavior of steepest-descent contours, in the following subsections, we shall briefly overview the Gaussian quadrature method~\cite{Press07gauss}, and then show how it can be used to accurately compute $\Upsilon_\Vect{t}(\Vect{x})$ along $\cont{0}$. \subsection{Gaussian quadrature for numerical integration } \label{sec:gaussQUAD} Suppose we wish to compute the integral of some real-valued function $h(\kappa)$ over the real interval $(a,b)$, with both $a$ and $b$ allowed to be infinite. Suppose further that $h(\kappa)$ can be partitioned as \begin{equation} h(\kappa) = \omega(\kappa) r(\kappa) , \label{eq:GQdecomp} \end{equation} \noindent with $\omega(\kappa)$ positive-definite on $(a,b)$ and $r(\kappa)$ a polynomial of degree $2n - 1$. Then, the following $n$-point quadrature formula holds: \begin{equation} \int_a^b \mathrm{d} \kappa \, h(\kappa) \equiv \int_a^b \mathrm{d} \kappa \, \omega(\kappa) r(\kappa) = \sum_{j = 1}^n w_j r(\kappa_j) , \label{eq:gqEXACT} \end{equation} \noindent where the quadrature weights $\{ w_j \}$ and nodes $\{ \kappa_j \}$ are determined as follows. Let us introduce the inner product \begin{equation} \langle h_1, h_2 \rangle \doteq \int_a^b \mathrm{d} \kappa \, \omega(\kappa) h_1(\kappa) h_2(\kappa) . \label{eq:innerPROD} \end{equation} \noindent Let us also introduce the family of real-valued polynomials $\{p_\ell(\kappa)\}$ (with $\ell$ the polynomial degree) that are orthogonal with respect to \Eq{eq:innerPROD}, that is, \begin{equation} \langle p_\ell, p_m \rangle = \eta_\ell \, \delta_{\ell m} , \quad \eta_\ell \doteq \langle p_\ell, p_{\ell} \rangle , \end{equation} \noindent where $\delta_{\ell m}$ is the Kronecker delta. By performing polynomial division of $r(\kappa)$ by $p_n(\kappa)$ and Lagrange interpolation of the residual, it can be shown~\cite[pp.~135--137]{Gil07} that the quadrature weights $\{w_j \}$ are determined by the formula \begin{equation} w_j = \left\langle \prod_{\substack{\ell = 1 \\ j \neq \ell}}^n \frac{\kappa - \kappa_\ell}{\kappa_j - \kappa_\ell}, 1 \right\rangle = \frac{\langle p_n(\kappa), (\kappa - \kappa_j)^{-1} \rangle}{p'_n(\kappa_j)} , \label{eq:GQweights} \end{equation} \noindent and the quadrature nodes are the $n$ zeros of $p_n(\kappa)$, {i.e., } \begin{equation} \{\kappa_j\} = \{ \kappa ~ | ~ p_n(\kappa) = 0 \} . \label{eq:GQnodes} \end{equation} If $h(\kappa)$ cannot be decomposed as \Eq{eq:GQdecomp} with polynomial $r(\kappa)$, the corresponding integral can still be approximately computed as \begin{equation} \int_a^b \mathrm{d} \kappa \, h(\kappa) \approx \sum_{j = 1}^n w_j \frac{h(\kappa_j)}{\omega(\kappa_j)} . \label{eq:gq} \end{equation} \noindent Equation \eq{eq:gq} defines the Gaussian quadrature method of numerical integration. The error in using \Eq{eq:gq} depends on how `close' $h(\kappa)/\omega(\kappa)$ is to being a $2n-1$ degree polynomial, as determined by the maximum value of $\pd{\kappa}^{2n}(h/\omega)$ over $(a,b)$. Explicitly,~\cite{Suli03equation} \begin{equation} \left| \int_a^b \mathrm{d} \kappa \, h(\kappa) - \sum_{j = 1}^n w_j \frac{h(\kappa_j)}{\omega(\kappa_j)} \right| \le \frac{\eta_n}{(2n)!} \max_{\zeta \in (a,b)} \left| \pd{\kappa}^{2n} \frac{h(\zeta)}{\omega(\zeta)} \right| . \label{eq:gqERROR} \end{equation} \noindent Note that the right-hand side vanishes when $h(\kappa)/\omega(\kappa)$ is a $2n-1$ degree polynomial, as desired. Also note that \Eq{eq:gq} is still valid when $h(\kappa)$ is complex-valued. Common choices for $\{p_\ell(\kappa) \}$ are the rescaled Legendre polynomials for integrals over finite $(a,b)$ with $\omega(\kappa) = 1$, or the Hermite polynomials for integrals with $(a,b) = (-\infty, +\infty)$ and $\omega(\kappa) = \exp(-\kappa^2)$. For our purposes, though, we will find it more convenient to use the less-common Freud polynomials (\App{sec:FreudQUAD}), as we shall now explain. \subsection{Gauss--Freud quadrature} Let us now develop the appropriate Gaussian quadrature rule for MGO. Along the steepest-descent contour that passes through a given saddlepoint at $\kappa = \kappa_0$, denoted $\cont{0}$, \Eq{eq:SDMint} takes the general form \begin{align} I(\Vect{y}_0) = &\exp \left\{ i \Re[f(\kappa_0, \Vect{y}_0)] \vphantom{\frac{}{}} \right\} \nonumber\\ &\times \int_{\cont{0}} \mathrm{d} \kappa \, g(\kappa, \Vect{y}_0) \exp \left\{ - \Im \left[ f(\kappa, \Vect{y}_0) \vphantom{\frac{}{}} \right] \right\} , \end{align} \noindent or equivalently, \begin{align} I(\Vect{y}_0) = &\exp \left\{ i \Re[f(\kappa_0, \Vect{y}_0)] \vphantom{\frac{}{}} \right\} \nonumber\\ &\times \int_{-\infty}^\infty \mathrm{d} l \, \kappa'(l) g[\kappa(l), \Vect{y}_0] \exp \left[ - F(l,\Vect{y}_0) \right] , \label{eq:steepI} \end{align} \noindent where we have introduced $\kappa(l)$ as a $1$-D parameterization of $\cont{0}$ with $\kappa(0) = \kappa_0$, we have defined \begin{equation} F(l, \Vect{y}_0) \doteq \Im \left\{ f[\kappa(l), \Vect{y}_0] \vphantom{\frac{}{}} \right\} , \end{equation} \noindent and we have set $\Vect{y} = \Vect{y}_0$ to emphasize that $\Vect{y}$ should be considered a \textit{fixed parameter} for the integration over $\kappa$% ~\footnote{Specifically, $\Vect{y}$ is related to the physical location of the wavefield $\psi(\Vect{x})$ via the ray map $\Vect{x}(\Vect{\tau})$ along with the local coordinate transformation $\Vect{\tau}(\Vect{y})$ needed to place $\Upsilon_\Vect{t}(\Vect{x})$ into standard form.}. % Note that $\kappa_0$ being a saddlepoint implies \begin{equation} \nabla \Im \left[ f(\kappa_0, \Vect{y}_0) \vphantom{\frac{}{}} \right] = \Vect{0} , \quad \pd{l} F(0,\Vect{y}_0) = 0 , \label{eq:saddleI} \end{equation} \noindent and $\mc{C}_0$ being a steepest-descent contour implies \begin{equation} F(l, \Vect{y}_0) \ge F(0, \Vect{y}_0) . \label{eq:Fdecrease} \end{equation} Suppose first that $\kappa_0$ is a non-degenerate saddlepoint, which typically occurs when $\Vect{y}_0$ does not coincide with a caustic. This means that \begin{equation} \pd{l}^2 F(0, \Vect{y}_0) > 0 . \label{eq:nonDEGEN} \end{equation} \noindent Hence, $F(l, \Vect{y}_0)$ is well-approximated around $l = 0$ as \begin{equation} F(l, \Vect{y}_0) \approx F(0, \Vect{y}_0) + \pd{l}^2 F(0, \Vect{y}_0) \, l^2 . \label{eq:taylorF} \end{equation} \noindent However, this approximation has obvious issues when $\kappa_0$ is a degenerate saddlepoint, which occurs when $\Vect{y}_0$ coincides with a caustic. For this case, although by \Eq{eq:caustic}, \begin{equation} \pd{l}^2 F(0, \Vect{y}_0) = 0 , \label{eq:degen} \end{equation} \noindent a quadratic function can still be fit to $F(l, \Vect{y}_0)$ as \begin{equation} F(l, \Vect{y}_0) \approx F(0, \Vect{y}_0) + s(\Vect{y}_0) \, l^2 , \label{eq:secantF} \end{equation} \noindent provided the scaling factor $s(\Vect{y}_0)$ is chosen appropriately; we choose to use the finite-difference formula \begin{subequations} \begin{align} \label{eq:sDEF} s(\Vect{y}_0) &= \left\{ \begin{array}{lr} s_-(\Vect{y}_0), & l \le 0 \\ s_+(\Vect{y}_0), & l > 0 \end{array} \right. , \\ \label{eq:sPM} s_\pm(\Vect{y}_0) &\doteq \frac{F(l_\pm, \Vect{y}_0) - F(0, \Vect{y}_0)}{ | l_\pm |^2 } , \end{align} \end{subequations} \noindent where $l_\pm$ satisfy the threshold condition \begin{equation} F(l_\pm, \Vect{y}_0) - F(0, \Vect{y}_0) \ge C_\pm \label{eq:kappaL} \end{equation} \noindent with $C_\pm$ arbitrary constants. (We use $C_\pm = 1$ for simplicity, but we found varying $C_\pm$ over the range $[0.5, 5]$ produced only $O(10^{-5})$ differences.) Note that we have allowed the possibility for different scaling factors on either side of $l = 0$ in case $\cont{0}$ has a kink at $\kappa_0$ (see \Fig{fig:EXcontour}). Importantly, \Eq{eq:Fdecrease} implies that $s > 0$. Also note that \Eq{eq:secantF} reduces to \Eq{eq:taylorF} in the limit $C_\pm \to 0$ when $\kappa_0$ is non-degenerate. As a final simplification, let us adopt a piecewise linear approximation to $\cont{0}$ such that% ~\footnote{The mapping $\kappa(l)$ being nonlinear is not a problem for Gaussian quadrature \textit{per se}, but it necessitates the inclusion of expensive root-finding steps into the Gaussian quadrature algorithm~\cite{Deano09}}% \begin{equation} \kappa(l) \approx \kappa_0 + |l| \times \left\{ \begin{array}{lr} \exp(i \sigma_-), & l \le 0 \\ \exp(i \sigma_+), & l > 0 \end{array} \right. \label{eq:unionC0} \end{equation} \noindent for suitable rotation angles $\sigma_\pm$, which are allowed to be different in case $\cont{0}$ has a kink at $\kappa_0$. It would be natural to choose $\sigma_\pm$ such that \Eq{eq:unionC0} is a tangent-line approximation to $\cont{0}$ at $\kappa_0$; however, this choice cannot be applied to degenerate saddlepoints. Instead, we shall allow \Eq{eq:unionC0} to generally describe the secant-line approximations to $\cont{0}$ that underlie \Eq{eq:sPM}, namely, \begin{align} \sigma_\pm &= \text{arg} \left[ \kappa(l_\pm) - \kappa_0 \right] \nonumber\\ &= \text{sign} \left\{ \frac{ \Im\left[ \kappa(l_\pm) - \kappa_0 \right] }{ \| \kappa(l_\pm) - \kappa_0 \| } \right\} \cos^{-1} \left\{ \frac{ \Re\left[ \kappa(l_\pm) - \kappa_0 \right] }{ \| \kappa(l_\pm) - \kappa_0 \| } \right\} \label{eq:sigmaDEF} . \end{align} \noindent Here, we take the convention that $\text{sign}(0) = 1$; hence \Eq{eq:sigmaDEF} restricts $\sigma_\pm$ to lie on the interval $(-\pi, \pi]$. \begin{figure*} \centering \includegraphics[width=0.32\linewidth,trim={16mm 4mm 26mm 4mm},clip]{EXcontour_a2.pdf} \includegraphics[width=0.32\linewidth,trim={16mm 4mm 26mm 4mm},clip]{EXcontour_a3.pdf} \includegraphics[width=0.32\linewidth,trim={16mm 4mm 26mm 4mm},clip]{EXcontour_a4.pdf} \caption{Steepest-descent contours (orange) for the integrand phase function $f(\kappa) \doteq \kappa^a$ of \Eq{eq:EXgauss} for various values of the parameter $a$, which characterizes the saddlepoint degeneracy. The background color represents the magnitude of the integrand $-\Im(f)$, with green corresponding to larger values and blue corresponding to smaller values. The order $n = 5$ quadrature nodes are shown as white dots, while the points $\kappa(l_\pm)$ [\Eq{eq:kappaL}] that are used to determine the rotation angles $\sigma_\pm$ [\Eq{eq:sigmaDEF}] are shown as black dots. Unused steepest-descent contours are shown as dashed orange lines. As can be seen, the steepest-descent contour has a kink when $a$ is odd, which requires $\sigma_+$ and $\sigma_-$ to be calculated separately.} \label{fig:EXcontour} \end{figure*} Inserting \Eqs{eq:secantF}, \eq{eq:sDEF}, and \eq{eq:unionC0} into \Eq{eq:steepI} yields \begin{widetext} \begin{align} \frac{ I(\Vect{y}_0) }{ \exp[i f(\kappa_0,\Vect{y}_0)] } &\approx \int_0^\infty \mathrm{d} l \left\{ g[\kappa_0 + l \exp(i \sigma_+), \Vect{y}_0] \exp[i \sigma_+ - s_+(\Vect{y}_0) l^2] - g[\kappa_0 + l \exp(i \sigma_-), \Vect{y}_0] \exp[i \sigma_- - s_-(\Vect{y}_0) l^2] \vphantom{\frac{}{}} \right\} \nonumber\\ &= \int_0^\infty \mathrm{d} l \left\{ g \left[ \kappa_0 + \frac{ l \exp(i \sigma_+) }{ \sqrt{s_+(\Vect{y}_0)} } , \Vect{y}_0 \right] \frac{ \exp(i \sigma_+) }{ \sqrt{s_+(\Vect{y}_0)} } - g \left[ \kappa_0 + \frac{ l \exp(i \sigma_-) }{ \sqrt{s_-(\Vect{y}_0)} } , \Vect{y}_0 \right] \frac{ \exp(i \sigma_-) }{ \sqrt{s_-(\Vect{y}_0)} } \right\} \exp(- l^2) . \end{align} \noindent Hence, an appropriate Gaussian quadrature rule for MGO is based on the inner product \begin{equation} \langle h_1, h_2 \rangle = \int_0^\infty \mathrm{d} l \, h_1(l) h_2(l) \exp(-l^2) , \label{eq:FreudIP} \end{equation} \noindent for which Freud polynomials are orthogonal (\App{sec:FreudQUAD}). Using \Eq{eq:gq}, we obtain the quadrature rule for MGO: \begin{equation} I(\Vect{y}_0) \approx \sum_{j = 1}^n w_j \exp(l_j^2) \left\{ h\left[ \kappa_0 + \frac{l_j \exp(i \sigma_+)}{\sqrt{s_+(\Vect{y}_0)}}, \Vect{y}_0 \right] \frac{\exp(i \sigma_+)}{\sqrt{s_+(\Vect{y}_0)}} - h\left[ \kappa_0 + \frac{l_j \exp(i \sigma_-)}{\sqrt{s_-(\Vect{y}_0)}}, \Vect{y}_0 \right] \frac{\exp(i \sigma_-)}{\sqrt{s_-(\Vect{y}_0)}} \right\} , \label{eq:MGOquad} \end{equation} \end{widetext} \noindent where $h(\kappa) = g(\kappa, \Vect{y}_0) \exp[i f(\kappa, \Vect{y}_0) ]$, and $\{l_j \}$ are the quadrature nodes. [We use $\{ l_j \}$ rather than $\{ \kappa_j \}$ to be consistent with the notation of \Eq{eq:FreudIP}.] Since Gauss--Freud quadrature is somewhat uncommon, a table of the corresponding $\{w_j\}$ and $\{l_j\}$ for various values of $n$ is also provided in \App{sec:FreudQUAD}. \subsection{Angle memory feedback for MGO simulations} \label{sec:feedback} Let us now allow $\Vect{y}$ to vary in \Eq{eq:MGOquad} (rather than being fixed at some $\Vect{y} = \Vect{y}_0$), as will occur when MGO is used to simulate a propagating wave. The steepest-descent topology for $I(\Vect{y})$ will be different for each new value of $\Vect{y}$, with correspondingly new values of $\sigma_\pm$. Repeatedly searching for $\cont{0}$ to compute $\sigma_\pm$ via \Eq{eq:sigmaDEF} can be computationally expensive, and merely identifying the correct $\cont{0}$ can be difficult in situations where multiple valid steepest-descent lines exist, as occurs at caustics. Fortunately, the steepest-descent topology of $I(\Vect{y})$ typically evolves smoothly with $\Vect{y}$, which means successive calculations of $\sigma_\pm$ will be correlated. We use this fact to construct a `memory feedback' algorithm to both speed up the time required to calculate the steepest-descent topology of $I(\Vect{y})$ and to correctly identify $\cont{0}$ at caustics. First, let us initialize the MGO simulation far from a caustic such that $\cont{0}$ is sufficiently simple: we expect the initial angles $\sigma_\pm^{(0)}$ to be approximately given as \begin{equation} \sigma_\pm^{(0)} \approx -\frac{\pi}{4} - \frac{\text{arg}[\pd{\kappa}^2f(\kappa_0, \Vect{y}_0)]}{2} \pm \frac{\pi}{2} \end{equation} \noindent restricted to the interval $(-\pi, \pi]$. By starting the search for the exact $\cont{0}$ near this value of $\sigma_\pm^{(0)}$, the search time can be reduced. As the simulation progresses, $\cont{0}$ will evolve smoothly; at each new point $\Vect{y}_j$, the search-time for $\cont{0}$ can be reduced by initializing the search near the previously calculated $\sigma_\pm^{(j-1)}$ corresponding to the previous position $\Vect{y}_{j-1}$. Moreover, by restricting the search to \textit{only} consider angles near $\sigma_\pm^{(j-1)}$, {i.e., } restricting $|\sigma_\pm^{(j)} - \sigma_\pm^{(j-i)}| \le \Delta $ for some threshold $\Delta$ (we choose $\Delta = 0.01$), the correct $\cont{0}$ will naturally be identified by analytic continuation, even at caustics. \section{Benchmarking results} \label{sec:examples} \subsection{Isolated saddlepoint} As a first benchmarking of our numerical steepest-descent algorithm \eq{eq:MGOquad}, let us consider the numerical evaluation of the following family of integrals: \begin{align} I(a,b) \doteq \int_{-\infty}^\infty \mathrm{d} \kappa \, \kappa^b \exp\left(i \kappa^a\right) , \quad a \ge 2, \, b \ge 0 , \label{eq:EXgauss} \end{align} \noindent whose exact solution is given by \begin{equation} I_\textrm{ex}(a,b) = \frac{2}{a} \, \Gamma\left( \frac{2 \chi}{\pi} \right) \times \left\{ \begin{array}{lr} \exp\left( i \chi \right), & \bar{a} = 0, \, \bar{b} = 0\\[1mm] 0, & \bar{a} = 0, \, \bar{b} = 1\\[1mm] \cos \chi, & \bar{a} = 1, \, \bar{b} = 0\\[1mm] i \sin \chi, & \bar{a} = 1, \, \bar{b} = 1 \end{array} \right. , \end{equation} \noindent where we have defined \begin{equation} \chi \doteq \frac{1+b}{2a} \pi , \quad \bar{a} \doteq \textrm{mod}_2(a) , \quad \bar{b} \doteq \textrm{mod}_2(b) . \end{equation} \noindent The family $I(a,b)$ also corresponds to the $A_{a-1}$ `cuspoid' caustic family (Table~\ref{tab:caustic}) evaluated at $\Vect{y} = 0$; as such, the integrand of $I(a,b)$ has an isolated saddlepoint at $\kappa = 0$ whose degeneracy is controlled by the value of $a$, with $a = 2$ being non-degenerate. To evaluate $I(a,b)$ via \Eq{eq:MGOquad} the scaling factors $s_\pm$ and rotation angles $\sigma_\pm$ are needed; these are given respectively as $s_\pm = 1$ and \begin{equation} \sigma_+ = \frac{\pi}{2a} , \quad \sigma_- = \left\{ \begin{array}{lr} \sigma_+ - \pi, & \bar{a} = 0 \\ \pi - \sigma_+, & \bar{a} = 1 \end{array} \right. . \label{eq:EXangles} \end{equation} \noindent In particular, \Eq{eq:EXangles} implies that the steepest-descent contour has a kink at $\kappa = 0$ when $a$ is odd, which necessitates our partitioning of \Eq{eq:MGOquad} into incoming and outgoing branches. This feature is also shown in \Fig{fig:EXcontour}. Figure~\ref{fig:EXerr} shows the error that results from evaluating $I(a,b)$ via \Eq{eq:MGOquad} for quadrature order $n \le 10$. The quadrature weights and nodes used have precision $10^{-15}$ and are listed explicitly in Table~\ref{tab:GFnodes}. Note that our quadrature rule was developed to evaluate $I(a,b)$ exactly when $a = 2$ and $b \in [0, 2n - 1]$, and indeed, we observe that the error for these values of $a$ and $b$ remains on the order of the node/weight precision until $n = 6$, beyond which the error is slightly larger than expected. However, this increased error is not due to issues with our quadrature rule \textit{per se}, but rather due to the round-off error that unavoidably accumulates when subtracting large numbers. This conclusion is corroborated by the fact that the increased error is isolated to the cases when $b$ is even and $I(a,b)$ should be identically zero in exact arithmetic by (anti-) symmetry. When $a > 2$, our quadrature rule achieves a respectable accuracy of $10^{-4}$ even at the relatively low quadrature order of $n = 10$, demonstrating the utility of \Eq{eq:MGOquad} at caustics and regular points alike. \begin{figure} \centering \includegraphics[width=\linewidth,trim={2mm 6mm 4mm 4mm},clip]{EXerr.pdf} \caption{Comparison of the error in computing \Eq{eq:EXgauss} using the quadrature rule of \Eq{eq:MGOquad} for various values of $a$ and $b$. The error metric used is the relative error when $I(a, b)$ [Eq. (48)] is nonzero and the absolute error otherwise. The shaded gray region marks the range of error over the entire range $b \in [0, 2n-1]$ for which our quadrature rule is expected to be exact, while the dashed black lines bound the region obtained when only even values of $b$ are considered. The precision of the quadrature nodes and weights used is $10^{-15}$. } \label{fig:EXerr} \end{figure} \subsection{EM wave in unmagnetized plasma slab with linear density profile} \label{sec:exAIRY} \begin{figure*} \centering \includegraphics[width=0.24\linewidth,trim={18mm 4mm 32mm 4mm},clip]{AIRYcont_p2.pdf} \includegraphics[width=0.24\linewidth,trim={18mm 4mm 32mm 4mm},clip]{AIRYcont_p04.pdf} \includegraphics[width=0.24\linewidth,trim={18mm 4mm 32mm 4mm},clip]{AIRYcont_p01.pdf} \includegraphics[width=0.24\linewidth,trim={18mm 4mm 32mm 4mm},clip]{AIRYcont_p001.pdf} \vspace{2mm} \includegraphics[width=0.24\linewidth,trim={18mm 4mm 32mm 4mm},clip]{AIRYcont_p-2.pdf} \includegraphics[width=0.24\linewidth,trim={18mm 4mm 32mm 4mm},clip]{AIRYcont_p-04.pdf} \includegraphics[width=0.24\linewidth,trim={18mm 4mm 32mm 4mm},clip]{AIRYcont_p-01.pdf} \includegraphics[width=0.24\linewidth,trim={18mm 4mm 32mm 4mm},clip]{AIRYcont_p-001.pdf} \caption{Same as \Fig{fig:EXcontour} for the phase function $f(\epsilon, p)$ [\Eq{eq:fAIRY}] at various values of $p$. The white dots correspond here to the $n = 10$ quadrature nodes. The steepest-descent contours evolve smoothly with $p$ and ultimately coalesce into a fold-type $A_2$ caustic at $p = 0$.} \label{fig:MGOcontour} \end{figure*} As a more realistic example, let us consider the MGO description of an EM wave propagating in a stationary unmagnetized plasma slab with a linearly varying density profile. Suppose that the EM wave and all subsequently induced fluctuations have time dependence of the form $\exp(-i \Omega t)$, where $\Omega$ is the wave frequency. Then, after defining $x$ as the direction of inhomogeneity, the electric field of the EM wave can be shown to satisfy~\cite[p.~344]{Stix92} \begin{equation} \pd{x}^2 E(x) + \frac{\Omega^2}{c^2}\left[1 - \frac{n(x)}{n_c} \right] E(x) = 0 , \label{eq:waveEQ} \end{equation} \noindent where $c$ is the speed of light in vacuum and $n_c$ is the cutoff density. Let us assume \begin{equation} n(x) = n_c \left(1 + \frac{x}{L_n} \right) , \end{equation} \noindent where $L_n$ is some constant length scale. Then, \Eq{eq:waveEQ} takes the form \begin{equation} \pd{q}^2 E(q) - q E(q) = 0 , \label{eq:eqAIRY} \end{equation} \noindent where we have introduced the re-scaled spatial variable \begin{equation} q \doteq x \left( \frac{\Omega^2}{c^2 L_n} \right)^{1/3} . \end{equation} Equation \eq{eq:eqAIRY} is known as Airy's equation, and contains a fold-type $A_2$ caustic at the cutoff location $q = 0$. Assuming that $E(q \to \infty) = 0$, the exact solution is given by the Airy function \begin{equation} E_\textrm{ex}(q) = \airyA(q) , \label{eq:exactAIRY} \end{equation} \noindent (where the overall constant is set to unity for simplicity) while the MGO solution \eq{eq:MGO} to \Eq{eq:eqAIRY} can be written in the underdense region $q \le 0$ as~\cite{Lopez20a} \begin{align} E_\textrm{MGO}(q) =& \Upsilon\left(|q|^{1/2} \right) \exp\left(- i \frac{2}{3} |q|^{3/2} \right) \nonumber\\ &+ \Upsilon\left(-|q|^{1/2} \right) \exp\left(i \frac{2}{3} |q|^{3/2} \right) . \label{eq:mgoAIRY} \end{align} \noindent The integral function $\Upsilon$ in \Eq{eq:mgoAIRY} has the form \begin{equation} \Upsilon(p) \doteq \frac{1}{2\pi} \int_{\cont{0}} \mathrm{d} \epsilon \, \frac{ \vartheta(p) \exp\left[ i f(\epsilon, p) \right] }{ \left[ \vartheta^4(p) - 8 \vartheta(p) p \epsilon \right]^{1/4} } , \label{eq:upsilonAIRY} \end{equation} \noindent where the phase function $f$ is given as \begin{align} f(\epsilon,p) &\doteq \frac{\vartheta^6(p) - \left[\vartheta^4(p) - 8 \vartheta(p) p \epsilon \right]^{3/2}}{96p^3} \nonumber\\ &\hspace{4mm}- \frac{\vartheta^3(p)}{8p^2} \epsilon + \frac{\vartheta^2(p)}{4p} \epsilon^2 , \label{eq:fAIRY} \end{align} \noindent and we have defined $\vartheta(p) \doteq \sqrt{1 + 4 p^2}$. When \Eq{eq:upsilonAIRY} is evaluated using the stationary-phase approximation, the standard GO approximation for \Eq{eq:eqAIRY} is obtained: \begin{equation} E_\textrm{GO}(q) = \pi^{-1/2} |q|^{-1/4} \sin\left(\frac{2}{3} |q|^{3/2} + \frac{\pi}{4} \right) . \label{eq:goAIRY} \end{equation} \noindent Clearly, the GO solution diverges at the caustic $q = 0$. Conversely, if \Eq{eq:fAIRY} is expanded to cubic order in $\epsilon$, then \Eq{eq:upsilonAIRY} can be evaluated along the steepest-descent contour to yield the approximate MGO solution~\cite{Lopez20a} \begin{align} E_\textrm{approx}(q) &= \sqrt{1 - 4 q} \, \airyA\left[- \varrho^2(q) \right] \cos[ \varpi(q)] \nonumber\\ &\hspace{5mm}- \sqrt{1 - 4 q} \, \airyB\left[ - \varrho^2(q) \right] \sin[ \varpi(q) ] \, , \label{eq:approxAIRY} \end{align} \noindent where we have defined \begin{align} \varrho(q) &\doteq (1 - 4q) \sqrt{|q|} , \quad \varpi(q) \doteq \frac{2}{3} \varrho^3(q) - \frac{2}{3}|q|^{3/2} . \end{align} \begin{figure} \centering \begin{overpic}[width=\linewidth,trim={4mm 5mm 4mm 4mm},clip]{MGOairy.pdf} \put(5,12){\textbf{\large(a)}} \end{overpic} \vspace{2mm} \hspace{-3mm}\begin{overpic}[width=0.95\linewidth,trim={4mm 22mm 4mm 36mm},clip]{Airy_ORDERscan.pdf} \put(5,12){\textbf{\large(b)}} \end{overpic} \caption{\textbf{(a)} Comparison of the numerical MGO solution (orange) with the analytically approximated MGO solution \eq{eq:approxAIRY}~\cite{Lopez20a} (dashed pink), the standard GO solution \eq{eq:goAIRY} (dashed gray), and the exact solution \eq{eq:exactAIRY} (black) for Airy's equation \eq{eq:eqAIRY}. The numerical MGO solution was obtained by applying the quadrature rule of \Eq{eq:MGOquad} with order $n = 10$ to \Eqs{eq:mgoAIRY} and \eq{eq:upsilonAIRY}. The numerical MGO solution displays remarkable agreement with the exact solution compared with the analytical approximations, even near the fold-type caustic at $q = 0$. \textbf{(b)} Error of the numerical MGO solution with respect to a scan over the quadrature order $n$. Note that the `pseudo error' is defined as the relative error with respect to the $n = 10$ solution used in \textbf{(a)}.} \label{fig:MGOairy} \end{figure} Here, we evaluate \Eq{eq:upsilonAIRY} numerically via \Eq{eq:MGOquad} over the range $q \in [-8, 0]$ using the angle memory feedback algorithm described in \Sec{sec:feedback}. Figure~\ref{fig:MGOcontour} shows the smooth evolution of steepest-descent curves obtained with the memory feedback algorithm, while \Fig{fig:MGOairy} compares the resultant numerical MGO solution with the exact solution \eq{eq:exactAIRY} and the two analytical approximations of \Eqs{eq:goAIRY} and \eq{eq:approxAIRY}. As \Fig{fig:MGOairy} shows, both the numerical MGO solution and the analytically approximated MGO solution remain finite at the caustic $q = 0$, whereas the GO solution diverges. However, the analytical approximation overestimates the peak intensity width near the caustic. Conversely, the numerical MGO solution agrees remarkably well with the exact solution everywhere, even though a relatively low quadrature order of $n = 10$ was used. Moreover, although the relative error with respect to the exact solution does not decrease much after quadrature order $n = 2$, the `pseudo error' (defined as the relative error between the numerical MGO solution for a given $n$ compared with the reference solution $n = 10$) continues to decrease with increasing $n$. This suggests that the numerical MGO algorithm quickly converges to the residual intrinsic error of the MGO theory, at least for this specific example. \section{Conclusions} \label{sec:concl} Metaplectic geometrical optics is a recently proposed formalism for modeling wave propagation in general linear media that avoids the usual singularities at caustics. MGO is therefore a promising alternative to the traditional GO approximation underlying ray-tracing codes. However, MGO yields solutions in the form of highly oscillatory integrals, which cannot be easily calculated using standard numerical methods. Here, we present a new algorithm for taking such integrals numerically that is based on the steepest-descent method combined with Gauss--Freud quadrature. We first validate our algorithm on isolated saddlepoints of various degeneracy to demonstrate the expected $2n-1$ polynomial accuracy of an $n$-point Gaussian quadrature formula. We then use our algorithm to simulate an EM wave propagating into an unmagnetized plasma that has a fold-type caustic at the critical cutoff density. The numerical solution agrees remarkably well with the exact solution and significantly improves upon the analytically approximated MGO solution that was previously obtained in \Ref{Lopez20a}. This encouraging result provides strong evidence that MGO can be suitable for practical applications. \section*{Acknowledgments} The authors thank Laura Xin Zhang for invaluable coding advice. This work was supported by the U.S.~DOE through Contract No.~DE-AC02-09CH11466 and through funding for the Summer Undergraduate Laboratory Internship (SULI) program.

2800b4573ccbbc6c45c37f91378ece177cf2237f